diff --git a/.github/labeler.yml b/.github/labeler.yml
index d8ada150c5..08cfd7e0bc 100644
--- a/.github/labeler.yml
+++ b/.github/labeler.yml
@@ -89,7 +89,10 @@ nix:
 embedding:
     - changed-files:
         - any-glob-to-any-file: examples/embedding/
-
+jinja parser:
+    - changed-files:
+        - any-glob-to-any-file:
+            - common/jinja/**
 Ascend NPU:
     - changed-files:
         - any-glob-to-any-file:
diff --git a/.github/workflows/build-cache.yml b/.github/workflows/build-cache.yml
index 6a22e41c3b..3de0be9fad 100644
--- a/.github/workflows/build-cache.yml
+++ b/.github/workflows/build-cache.yml
@@ -16,7 +16,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Get latest Vulkan SDK version
         id: vulkan_sdk_version
@@ -24,7 +24,7 @@ jobs:
           echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
 
       - name: Setup Cache
-        uses: actions/cache@v4
+        uses: actions/cache@v5
         id: cache-sdk
         with:
           path: ./vulkan_sdk
@@ -47,10 +47,10 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Setup Cache
-        uses: actions/cache@v4
+        uses: actions/cache@v5
         id: cache-toolchain
         with:
           path: ./spacemit_toolchain
@@ -73,10 +73,10 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Setup Cache
-        uses: actions/cache@v4
+        uses: actions/cache@v5
         id: cache-rocm
         with:
           path: C:\Program Files\AMD\ROCm
diff --git a/.github/workflows/build-cmake-pkg.yml b/.github/workflows/build-cmake-pkg.yml
index 510352a5cc..259efa43c8 100644
--- a/.github/workflows/build-cmake-pkg.yml
+++ b/.github/workflows/build-cmake-pkg.yml
@@ -7,7 +7,7 @@ jobs:
   linux:
     runs-on: ubuntu-24.04
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
diff --git a/.github/workflows/build-linux-cross.yml b/.github/workflows/build-linux-cross.yml
index 4d3b687a51..8b6ebaf4a3 100644
--- a/.github/workflows/build-linux-cross.yml
+++ b/.github/workflows/build-linux-cross.yml
@@ -8,7 +8,7 @@ jobs:
   #   runs-on: ubuntu-24.04
 
   #   steps:
-  #     - uses: actions/checkout@v4
+  #     - uses: actions/checkout@v6
   #     - name: Setup Riscv
   #       run: |
   #         sudo dpkg --add-architecture riscv64
@@ -52,7 +52,7 @@ jobs:
   #   runs-on: ubuntu-24.04
 
   #   steps:
-  #     - uses: actions/checkout@v4
+  #     - uses: actions/checkout@v6
   #     - name: Setup Riscv
   #       run: |
   #         sudo dpkg --add-architecture riscv64
@@ -99,7 +99,7 @@ jobs:
   #   runs-on: ubuntu-24.04
 
   #   steps:
-  #     - uses: actions/checkout@v4
+  #     - uses: actions/checkout@v6
   #     - name: Setup Arm64
   #       run: |
   #         sudo dpkg --add-architecture arm64
@@ -146,7 +146,7 @@ jobs:
     container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
       - name: Setup LoongArch
         run: |
           rm -f /etc/apt/sources.list.d/*
@@ -201,7 +201,7 @@ jobs:
     container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
       - name: Setup LoongArch
         run: |
           rm -f /etc/apt/sources.list.d/*
@@ -262,10 +262,10 @@ jobs:
       SPACEMIT_IME_TOOLCHAIN_VERSION: "1.1.2"
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
 
       - name: Use SpacemiT Toolchain Cache
-        uses: actions/cache@v4
+        uses: actions/cache@v5
         id: cache-toolchain
         with:
           path: ./spacemit_toolchain
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 187c861437..f738edefc4 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -21,7 +21,8 @@ on:
       '**/*.m',
       '**/*.metal',
       '**/*.comp',
-      '**/*.glsl'
+      '**/*.glsl',
+      '**/*.wgsl'
     ]
 
   pull_request:
@@ -42,7 +43,8 @@ on:
       '**/*.m',
       '**/*.metal',
       '**/*.comp',
-      '**/*.glsl'
+      '**/*.glsl',
+      '**/*.wgsl'
     ]
 
 concurrency:
@@ -63,7 +65,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -99,7 +101,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -135,7 +137,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -189,7 +191,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -269,7 +271,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -317,7 +319,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Dependencies
         id: depends
@@ -347,7 +349,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       # - name: ccache
       #   uses: ggml-org/ccache-action@v1.2.16
@@ -380,7 +382,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -414,7 +416,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -436,7 +438,7 @@ jobs:
           echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
 
       - name: Use Vulkan SDK Cache
-        uses: actions/cache@v4
+        uses: actions/cache@v5
         id: cache-sdk
         with:
           path: ./vulkan_sdk
@@ -472,7 +474,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -494,7 +496,7 @@ jobs:
           echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
 
       - name: Use Vulkan SDK Cache
-        uses: actions/cache@v4
+        uses: actions/cache@v5
         id: cache-sdk
         with:
           path: ./vulkan_sdk
@@ -543,7 +545,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -585,7 +587,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Dependencies
         id: depends
@@ -616,7 +618,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Dependencies
         id: depends
@@ -644,7 +646,7 @@ jobs:
     continue-on-error: true
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
 
       - name: add oneAPI to apt
         shell: bash
@@ -668,7 +670,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -693,7 +695,7 @@ jobs:
     continue-on-error: true
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
 
       - name: add oneAPI to apt
         shell: bash
@@ -717,7 +719,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -749,7 +751,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -781,7 +783,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -813,7 +815,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Build
         id: cmake_build
@@ -843,7 +845,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -853,7 +855,7 @@ jobs:
           save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Download xcframework artifact
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@v7
         with:
           name: llama-xcframework
           path: build-apple/llama.xcframework/
@@ -885,7 +887,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -954,7 +956,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1053,7 +1055,7 @@ jobs:
     steps:
         - name: Clone
           id: checkout
-          uses: actions/checkout@v4
+          uses: actions/checkout@v6
 
         - name: Install dependencies
           env:
@@ -1092,7 +1094,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Install ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1145,7 +1147,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1177,7 +1179,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Grab rocWMMA package
         id: grab_rocwmma
@@ -1187,7 +1189,7 @@ jobs:
           7z x data.tar
 
       - name: Use ROCm Installation Cache
-        uses: actions/cache@v4
+        uses: actions/cache@v5
         id: cache-rocm
         with:
           path: C:\Program Files\AMD\ROCm
@@ -1239,7 +1241,7 @@ jobs:
 
     steps:
       - name: Checkout code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Setup Xcode
         uses: maxim-lobanov/setup-xcode@v1
@@ -1269,7 +1271,7 @@ jobs:
           ./build-xcframework.sh
 
       - name: Upload xcframework artifact
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           name: llama-xcframework
           path: build-apple/llama.xcframework/
@@ -1285,7 +1287,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       # Disabled due to size (400MB) and always 0 cache hits
       # - name: ccache
@@ -1295,7 +1297,7 @@ jobs:
       #     evict-old-files: 1d
 
       - name: Set up JDK
-        uses: actions/setup-java@v3
+        uses: actions/setup-java@v5
         with:
           java-version: 17
           distribution: zulu
@@ -1327,7 +1329,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Install OpenCL Headers and Libs
         id: install_opencl
@@ -1371,7 +1373,7 @@ jobs:
         id: update_presets
         if: ${{ matrix.build == 'arm64-snapdragon' }}
         run: |
-          cp docs/backend/hexagon/CMakeUserPresets.json .
+          cp docs/backend/snapdragon/CMakeUserPresets.json .
 
       - name: Build
         id: ndk_build
@@ -1402,7 +1404,7 @@ jobs:
     runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
     steps:
       - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -1460,7 +1462,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1486,7 +1488,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1512,7 +1514,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1538,7 +1540,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1564,7 +1566,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -1590,7 +1592,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Test
         id: ggml-ci
@@ -1604,7 +1606,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Test
         id: ggml-ci
@@ -1618,7 +1620,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Test
         id: ggml-ci
@@ -1632,7 +1634,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Test
         id: ggml-ci
@@ -1645,7 +1647,7 @@ jobs:
   #   steps:
   #     - name: Clone
   #       id: checkout
-  #       uses: actions/checkout@v4
+  #       uses: actions/checkout@v6
 
   #     - name: Test
   #       id: ggml-ci
@@ -1659,7 +1661,7 @@ jobs:
   #   steps:
   #     - name: Clone
   #       id: checkout
-  #       uses: actions/checkout@v4
+  #       uses: actions/checkout@v6
 
   #     - name: Test
   #       id: ggml-ci
@@ -1673,7 +1675,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Test
         id: ggml-ci
@@ -1686,7 +1688,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Dawn Dependency
         id: dawn-depends
@@ -1714,7 +1716,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Test
         id: ggml-ci
@@ -1728,7 +1730,7 @@ jobs:
      steps:
        - name: Clone
          id: checkout
-         uses: actions/checkout@v4
+         uses: actions/checkout@v6
 
        - name: ccache
          uses: ggml-org/ccache-action@v1.2.16
@@ -1773,7 +1775,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Check environment
         run: |
@@ -1875,7 +1877,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Setup ccache
         run: |
@@ -1969,7 +1971,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Setup ccache
         run: |
@@ -2043,7 +2045,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Setup ccache
         run: |
@@ -2089,7 +2091,7 @@ jobs:
      steps:
        - name: Clone
          id: checkout
-         uses: actions/checkout@v4
+         uses: actions/checkout@v6
 
        - name: Dependencies
          id: depends
diff --git a/.github/workflows/check-vendor.yml b/.github/workflows/check-vendor.yml
index 7b3016079c..1671ed7b8b 100644
--- a/.github/workflows/check-vendor.yml
+++ b/.github/workflows/check-vendor.yml
@@ -19,16 +19,16 @@ on:
 
 jobs:
   check-vendor:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-slim
 
     steps:
       - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
       - name: Setup Python
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v6
         with:
           python-version: '3.x'
 
diff --git a/.github/workflows/close-issue.yml b/.github/workflows/close-issue.yml
index cbfc4990db..ec3df08b2d 100644
--- a/.github/workflows/close-issue.yml
+++ b/.github/workflows/close-issue.yml
@@ -10,12 +10,12 @@ permissions:
 
 jobs:
   close-issues:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-slim
     permissions:
       issues: write
       pull-requests: write
     steps:
-      - uses: actions/stale@v5
+      - uses: actions/stale@v10
         with:
           exempt-issue-labels: "refactoring,help wanted,good first issue,research 🔬,bug,roadmap"
           days-before-issue-stale: 30
diff --git a/.github/workflows/copilot-setup-steps.yml b/.github/workflows/copilot-setup-steps.yml
index 5f733e684e..fc3cec5ea1 100644
--- a/.github/workflows/copilot-setup-steps.yml
+++ b/.github/workflows/copilot-setup-steps.yml
@@ -26,7 +26,7 @@ jobs:
     # If you do not check out your code, Copilot will do this for you.
     steps:
       - name: Checkout code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -45,7 +45,7 @@ jobs:
           sudo chmod +x /usr/local/bin/git-clang-format
 
       - name: Set up Python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: '3.11'
 
diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml
index d9fe0686d3..8062177ba5 100644
--- a/.github/workflows/docker.yml
+++ b/.github/workflows/docker.yml
@@ -49,7 +49,7 @@ jobs:
           - { tag: "rocm",   dockerfile: ".devops/rocm.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
     steps:
       - name: Check out the repo
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0 # preserve git history, so we can determine the build number
 
@@ -63,7 +63,7 @@ jobs:
         uses: docker/setup-buildx-action@v3
 
       - name: Log in to Docker Hub
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
         with:
           registry: ghcr.io
           username: ${{ github.repository_owner }}
@@ -208,7 +208,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
diff --git a/.github/workflows/editorconfig.yml b/.github/workflows/editorconfig.yml
index f02b7c2194..702dc89f5b 100644
--- a/.github/workflows/editorconfig.yml
+++ b/.github/workflows/editorconfig.yml
@@ -20,9 +20,9 @@ concurrency:
 
 jobs:
   editorconfig:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-slim
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v6
       - uses: editorconfig-checker/action-editorconfig-checker@v2
         with:
           version: v3.0.3
diff --git a/.github/workflows/gguf-publish.yml b/.github/workflows/gguf-publish.yml
index 3ca4d30581..0e95766459 100644
--- a/.github/workflows/gguf-publish.yml
+++ b/.github/workflows/gguf-publish.yml
@@ -21,12 +21,12 @@ on:
 jobs:
   deploy:
 
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-slim
 
     steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@v6
     - name: Set up Python
-      uses: actions/setup-python@v5
+      uses: actions/setup-python@v6
       with:
         python-version: '3.9.x'
     - name: Install dependencies
diff --git a/.github/workflows/labeler.yml b/.github/workflows/labeler.yml
index 0b0f300aa4..eab20c6881 100644
--- a/.github/workflows/labeler.yml
+++ b/.github/workflows/labeler.yml
@@ -7,11 +7,11 @@ jobs:
     permissions:
       contents: read
       pull-requests: write
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-slim
     steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@v6
       with:
         repository: "ggml-org/llama.cpp"
-    - uses: actions/labeler@v5
+    - uses: actions/labeler@v6
       with:
         configuration-path: '.github/labeler.yml'
diff --git a/.github/workflows/pre-tokenizer-hashes.yml b/.github/workflows/pre-tokenizer-hashes.yml
index dff998e239..7126b62b69 100644
--- a/.github/workflows/pre-tokenizer-hashes.yml
+++ b/.github/workflows/pre-tokenizer-hashes.yml
@@ -12,14 +12,14 @@ on:
 
 jobs:
     pre-tokenizer-hashes:
-        runs-on: ubuntu-latest
+        runs-on: ubuntu-slim
 
         steps:
         - name: Checkout repository
-          uses: actions/checkout@v4
+          uses: actions/checkout@v6
 
         - name: Set up Python
-          uses: actions/setup-python@v5
+          uses: actions/setup-python@v6
           with:
               python-version: '3.11'
 
diff --git a/.github/workflows/python-check-requirements.yml b/.github/workflows/python-check-requirements.yml
index 46e80aecd0..1219b87459 100644
--- a/.github/workflows/python-check-requirements.yml
+++ b/.github/workflows/python-check-requirements.yml
@@ -20,13 +20,13 @@ concurrency:
 
 jobs:
   python-check-requirements:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-slim
     name: check-requirements
     steps:
       - name: Check out source repository
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
       - name: Set up Python environment
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: "3.11"
       - name: Run check-requirements.sh script
diff --git a/.github/workflows/python-lint.yml b/.github/workflows/python-lint.yml
index ddfdf73b8f..8d1dd7a7d5 100644
--- a/.github/workflows/python-lint.yml
+++ b/.github/workflows/python-lint.yml
@@ -15,13 +15,13 @@ concurrency:
 
 jobs:
   flake8-lint:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-slim
     name: Lint
     steps:
       - name: Check out source repository
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
       - name: Set up Python environment
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: "3.11"
       - name: flake8 Lint
diff --git a/.github/workflows/python-type-check.yml b/.github/workflows/python-type-check.yml
index 373bb60102..e801a9f42e 100644
--- a/.github/workflows/python-type-check.yml
+++ b/.github/workflows/python-type-check.yml
@@ -24,14 +24,12 @@ jobs:
     name: pyright type-check
     steps:
       - name: Check out source repository
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
       - name: Set up Python environment
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: "3.11"
-      - name: Install Python dependencies
-        # TODO: use a venv
-        run: pip install -r requirements/requirements-all.txt
+          pip-install: -r requirements/requirements-all.txt
       - name: Type-check with Pyright
         uses: jakebailey/pyright-action@v2
         with:
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index d8b3b95df0..1914c08489 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -27,7 +27,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -63,7 +63,7 @@ jobs:
           tar -czvf llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.tar.gz -s ",./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.tar.gz
           name: llama-bin-macos-arm64.tar.gz
@@ -74,7 +74,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -111,7 +111,7 @@ jobs:
           tar -czvf llama-${{ steps.tag.outputs.name }}-bin-macos-x64.tar.gz -s ",./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.tar.gz
           name: llama-bin-macos-x64.tar.gz
@@ -133,7 +133,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -173,7 +173,7 @@ jobs:
           tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.tar.gz
           name: llama-bin-ubuntu-${{ matrix.build }}.tar.gz
@@ -184,7 +184,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -226,7 +226,7 @@ jobs:
           tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz
           name: llama-bin-ubuntu-vulkan-x64.tar.gz
@@ -242,7 +242,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -278,7 +278,7 @@ jobs:
           7z a -snl llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-bin-win-cpu-${{ matrix.arch }}.zip
           name: llama-bin-win-cpu-${{ matrix.arch }}.zip
@@ -305,7 +305,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -360,7 +360,7 @@ jobs:
           7z a -snl llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip .\build\bin\Release\${{ matrix.target }}.dll
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
           name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
@@ -375,7 +375,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Install ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -416,7 +416,7 @@ jobs:
           7z a -snl llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip .\build\bin\Release\ggml-cuda.dll
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
           name: llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
@@ -431,7 +431,7 @@ jobs:
           7z a cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip $dst\*
 
       - name: Upload Cuda runtime
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
           name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
@@ -451,7 +451,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
@@ -511,7 +511,7 @@ jobs:
           7z a -snl llama-bin-win-sycl-x64.zip ./build/bin/*
 
       - name: Upload the release package
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-bin-win-sycl-x64.zip
           name: llama-bin-win-sycl-x64.zip
@@ -531,7 +531,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
 
       - name: Grab rocWMMA package
         id: grab_rocwmma
@@ -542,7 +542,7 @@ jobs:
 
       - name: Cache ROCm Installation
         id: cache-rocm
-        uses: actions/cache@v4
+        uses: actions/cache@v5
         with:
           path: C:\Program Files\AMD\ROCm
           key: rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
@@ -617,7 +617,7 @@ jobs:
           7z a -snl llama-bin-win-hip-${{ matrix.name }}-x64.zip .\build\bin\*
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-bin-win-hip-${{ matrix.name }}-x64.zip
           name: llama-bin-win-hip-${{ matrix.name }}-x64.zip
@@ -627,7 +627,7 @@ jobs:
 
     steps:
       - name: Checkout code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -672,7 +672,7 @@ jobs:
           zip -r -y llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-${{ steps.tag.outputs.name }}-xcframework.zip
           name: llama-${{ steps.tag.outputs.name }}-xcframework.zip
@@ -703,7 +703,7 @@ jobs:
     runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
     steps:
       - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -763,7 +763,7 @@ jobs:
           tar -czvf llama-${{ steps.tag.outputs.name }}-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}${{ matrix.use_acl_graph == 'on' && '-aclgraph' || '' }}.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
 
       - name: Upload artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v6
         with:
           path: llama-${{ steps.tag.outputs.name }}-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}${{ matrix.use_acl_graph == 'on' && '-aclgraph' || '' }}.tar.gz
           name: llama-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}${{ matrix.use_acl_graph == 'on' && '-aclgraph' || '' }}.tar.gz
@@ -794,7 +794,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
 
@@ -804,7 +804,7 @@ jobs:
 
       - name: Download artifacts
         id: download-artifact
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@v7
         with:
           path: ./artifact
           merge-multiple: true
@@ -887,7 +887,7 @@ jobs:
 
       - name: Upload release
         id: upload_release
-        uses: actions/github-script@v3
+        uses: actions/github-script@v8
         with:
           github-token: ${{secrets.GITHUB_TOKEN}}
           script: |
@@ -897,7 +897,7 @@ jobs:
             for (let file of await fs.readdirSync('./release')) {
               if (path.extname(file) === '.zip' || file.endsWith('.tar.gz')) {
                 console.log('uploadReleaseAsset', file);
-                await github.repos.uploadReleaseAsset({
+                await github.rest.repos.uploadReleaseAsset({
                   owner: context.repo.owner,
                   repo: context.repo.repo,
                   release_id: release_id,
diff --git a/.github/workflows/server-webui.yml b/.github/workflows/server-webui.yml
index 318003c5cc..6d1b617371 100644
--- a/.github/workflows/server-webui.yml
+++ b/.github/workflows/server-webui.yml
@@ -37,14 +37,14 @@ jobs:
     continue-on-error: true
     steps:
       - name: Checkout code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
       - name: Setup Node.js
         id: node
-        uses: actions/setup-node@v4
+        uses: actions/setup-node@v6
         with:
           node-version: "22"
           cache: "npm"
@@ -131,14 +131,14 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
       - name: Python setup
         id: setup_python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: '3.11'
 
@@ -148,7 +148,7 @@ jobs:
           pip install -r tools/server/tests/requirements.txt
 
       - name: Setup Node.js for WebUI
-        uses: actions/setup-node@v4
+        uses: actions/setup-node@v6
         with:
           node-version: "22"
           cache: "npm"
diff --git a/.github/workflows/server.yml b/.github/workflows/server.yml
index ab7c520e11..9f1ef48c82 100644
--- a/.github/workflows/server.yml
+++ b/.github/workflows/server.yml
@@ -64,7 +64,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
@@ -72,12 +72,12 @@ jobs:
       - name: Build
         id: cmake_build
         run: |
-          cmake -B build -DLLAMA_BUILD_BORINGSSL=ON
+          cmake -B build -DLLAMA_BUILD_BORINGSSL=ON -DGGML_SCHED_NO_REALLOC=ON
           cmake --build build --config ${{ matrix.build_type }} -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
 
       - name: Python setup
         id: setup_python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: '3.11'
 
@@ -100,7 +100,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
         with:
           fetch-depth: 0
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
@@ -108,12 +108,12 @@ jobs:
       - name: Build
         id: cmake_build
         run: |
-          cmake -B build -DLLAMA_BUILD_BORINGSSL=ON
+          cmake -B build -DLLAMA_BUILD_BORINGSSL=ON -DGGML_SCHED_NO_REALLOC=ON
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
 
       - name: Python setup
         id: setup_python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v6
         with:
           python-version: '3.11'
 
diff --git a/.github/workflows/update-ops-docs.yml b/.github/workflows/update-ops-docs.yml
index d5e264b34f..2ab06eb981 100644
--- a/.github/workflows/update-ops-docs.yml
+++ b/.github/workflows/update-ops-docs.yml
@@ -14,14 +14,14 @@ on:
 
 jobs:
     update-ops-docs:
-        runs-on: ubuntu-latest
+        runs-on: ubuntu-slim
 
         steps:
         - name: Checkout repository
-          uses: actions/checkout@v4
+          uses: actions/checkout@v6
 
         - name: Set up Python
-          uses: actions/setup-python@v5
+          uses: actions/setup-python@v6
           with:
               python-version: '3.x'
 
diff --git a/.github/workflows/winget.yml b/.github/workflows/winget.yml
index d3d9be23ce..2047c276f8 100644
--- a/.github/workflows/winget.yml
+++ b/.github/workflows/winget.yml
@@ -21,23 +21,24 @@ jobs:
 
       - name: Find latest release
         id: find_latest_release
-        uses: actions/github-script@v6
+        uses: actions/github-script@v8
         with:
           script: |
             const { data: releases } = await github.rest.repos.listReleases({
               owner: context.repo.owner,
               repo: context.repo.repo,
             });
-            console.log("Latest release:", releases[0].tag_name);
-            return releases[0].tag_name;
+            const { tag_name: version, assets: assets } = releases.find(({assets}) => assets.find(asset => asset.name.includes('win-vulkan')));
+            const { browser_download_url: asset_url } = assets.find(asset => asset.name.includes('win-vulkan'));
+            console.log("Latest release:", version);
+            core.setOutput('VERSION', version);
+            core.setOutput('ASSETURL', asset_url);
 
       - name: Update manifest
-        env:
-          VERSION: ${{ steps.find_latest_release.outputs.result }}
         run: |
           echo "Updating manifest..."
-          komac update --version ${{ env.VERSION }} \
-            --urls "https://github.com/ggml-org/llama.cpp/releases/download/${{ env.VERSION }}/llama-${{ env.VERSION }}-bin-win-vulkan-x64.zip" \
+          komac update --version ${{ steps.find_latest_release.outputs.VERSION }} \
+            --urls "${{ steps.find_latest_release.outputs.ASSETURL }}" \
             --token ${{ secrets.WINGET_GITHUB_TOKEN }} \
             --submit \
             ggml.llamacpp
diff --git a/CODEOWNERS b/CODEOWNERS
index 750096d9a1..e573a3d2e6 100644
--- a/CODEOWNERS
+++ b/CODEOWNERS
@@ -15,8 +15,10 @@
 /common/common.*                        @ggerganov
 /common/console.*                       @ggerganov
 /common/http.*                          @angt
+/common/jinja/                          @ngxson @CISC @aldehir
 /common/llguidance.*                    @ggerganov
 /common/log.*                           @ggerganov
+/common/ngram-map.*                     @srogmann
 /common/peg-parser.*                    @aldehir
 /common/sampling.*                      @ggerganov
 /common/speculative.*                   @ggerganov
@@ -66,6 +68,7 @@
 /ggml/src/ggml-rpc/                     @rgerganov
 /ggml/src/ggml-threading.*              @ggerganov
 /ggml/src/ggml-vulkan/                  @0cc4m
+/ggml/src/ggml-virtgpu/                 @kpouget
 /ggml/src/ggml-webgpu/                  @reeselevine
 /ggml/src/ggml-zdnn/                    @taronaeo @Andreas-Krebbel @AlekseiNikiforovIBM
 /ggml/src/ggml.c                        @ggerganov
diff --git a/README.md b/README.md
index 91a8f25d1c..dac020ad37 100644
--- a/README.md
+++ b/README.md
@@ -132,6 +132,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
 - [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat)
 - [x] [Bielik-11B-v2.3](https://huggingface.co/collections/speakleash/bielik-11b-v23-66ee813238d9b526a072408a)
+- [x] [RWKV-7](https://huggingface.co/collections/shoumenchougou/rwkv7-gxx-gguf)
 - [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM)
 - [x] [QRWKV-6](https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1)
 - [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct)
@@ -212,6 +213,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [llama.vim](https://github.com/ggml-org/llama.vim) (MIT)
 - [LARS](https://github.com/abgulati/LARS) (AGPL)
 - [Llama Assistant](https://github.com/vietanhdev/llama-assistant) (GPL)
+- [LlamaLib](https://github.com/undreamai/LlamaLib) (Apache-2.0)
 - [LLMFarm](https://github.com/guinmoon/LLMFarm?tab=readme-ov-file) (MIT)
 - [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT)
 - [LMStudio](https://lmstudio.ai/) (proprietary)
diff --git a/ci/run.sh b/ci/run.sh
index 6ca6ea5669..dfcf959661 100755
--- a/ci/run.sh
+++ b/ci/run.sh
@@ -254,7 +254,7 @@ function gg_run_ctest_release {
     (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     if [ -z ${GG_BUILD_LOW_PERF} ]; then
-        (time ctest --output-on-failure -L main ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+        (time ctest --output-on-failure -L 'main|python' ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     else
         (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     fi
diff --git a/common/CMakeLists.txt b/common/CMakeLists.txt
index ae02c0bd77..295ae9ea25 100644
--- a/common/CMakeLists.txt
+++ b/common/CMakeLists.txt
@@ -73,6 +73,10 @@ add_library(${TARGET} STATIC
     log.h
     ngram-cache.cpp
     ngram-cache.h
+    ngram-map.cpp
+    ngram-map.h
+    ngram-mod.cpp
+    ngram-mod.h
     peg-parser.cpp
     peg-parser.h
     preset.cpp
diff --git a/common/arg.cpp b/common/arg.cpp
index 163c9b71b0..5fbc9022c0 100644
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -6,6 +6,7 @@
 #include "json-schema-to-grammar.h"
 #include "log.h"
 #include "sampling.h"
+#include "speculative.h"
 #include "preset.h"
 
 // fix problem with std::min and std::max
@@ -579,14 +580,14 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
             params.mmproj = res.mmproj;
         }
         // only download mmproj if the current example is using it
-        for (auto & ex : mmproj_examples) {
+        for (const auto & ex : mmproj_examples) {
             if (ctx_arg.ex == ex) {
                 common_params_handle_model(params.mmproj,    params.hf_token, params.offline);
                 break;
             }
         }
-        common_params_handle_model(params.speculative.model, params.hf_token, params.offline);
-        common_params_handle_model(params.vocoder.model,     params.hf_token, params.offline);
+        common_params_handle_model(params.speculative.mparams_dft, params.hf_token, params.offline);
+        common_params_handle_model(params.vocoder.model,           params.hf_token, params.offline);
     }
 
     // model is required (except for server)
@@ -1216,21 +1217,25 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-lcs", "--lookup-cache-static"}, "FNAME",
         "path to static lookup cache to use for lookup decoding (not updated by generation)",
         [](common_params & params, const std::string & value) {
-            params.lookup_cache_static = value;
+            params.speculative.lookup_cache_static = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_LOOKUP}));
+    ).set_examples({LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-lcd", "--lookup-cache-dynamic"}, "FNAME",
         "path to dynamic lookup cache to use for lookup decoding (updated by generation)",
         [](common_params & params, const std::string & value) {
-            params.lookup_cache_dynamic = value;
+            params.speculative.lookup_cache_dynamic = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_LOOKUP}));
+    ).set_examples({LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-c", "--ctx-size"}, "N",
         string_format("size of the prompt context (default: %d, 0 = loaded from model)", params.n_ctx),
         [](common_params & params, int value) {
             params.n_ctx = value;
+            if (value == 0) {
+                // disable context reduction in llama_params_fit if the user explicitly requests the full context size:
+                params.fit_params_min_ctx = UINT32_MAX;
+            }
         }
     ).set_env("LLAMA_ARG_CTX_SIZE"));
     add_opt(common_arg(
@@ -1291,11 +1296,12 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_ARG_CACHE_RAM").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
     add_opt(common_arg(
         {"-kvu", "--kv-unified"},
+        {"-no-kvu", "--no-kv-unified"},
         "use single unified KV buffer shared across all sequences (default: enabled if number of slots is auto)",
-        [](common_params & params) {
-            params.kv_unified = true;
+        [](common_params & params, bool value) {
+            params.kv_unified = value;
         }
-    ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_BATCHED}));
+    ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_BATCHED, LLAMA_EXAMPLE_BENCH}));
     add_opt(common_arg(
         {"--context-shift"},
         {"--no-context-shift"},
@@ -1573,7 +1579,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--temp"}, "N",
-        string_format("temperature (default: %.1f)", (double)params.sampling.temp),
+        string_format("temperature (default: %.2f)", (double)params.sampling.temp),
         [](common_params & params, const std::string & value) {
             params.sampling.temp = std::stof(value);
             params.sampling.temp = std::max(params.sampling.temp, 0.0f);
@@ -1590,7 +1596,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam().set_env("LLAMA_ARG_TOP_K"));
     add_opt(common_arg(
         {"--top-p"}, "N",
-        string_format("top-p sampling (default: %.1f, 1.0 = disabled)", (double)params.sampling.top_p),
+        string_format("top-p sampling (default: %.2f, 1.0 = disabled)", (double)params.sampling.top_p),
         [](common_params & params, const std::string & value) {
             params.sampling.top_p = std::stof(value);
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P;
@@ -1598,7 +1604,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--min-p"}, "N",
-        string_format("min-p sampling (default: %.1f, 0.0 = disabled)", (double)params.sampling.min_p),
+        string_format("min-p sampling (default: %.2f, 0.0 = disabled)", (double)params.sampling.min_p),
         [](common_params & params, const std::string & value) {
             params.sampling.min_p = std::stof(value);
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P;
@@ -1606,14 +1612,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--top-nsigma"}, "N",
-        string_format("top-n-sigma sampling (default: %.1f, -1.0 = disabled)", params.sampling.top_n_sigma),
+        string_format("top-n-sigma sampling (default: %.2f, -1.0 = disabled)", params.sampling.top_n_sigma),
         [](common_params & params, const std::string & value) {
             params.sampling.top_n_sigma = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--xtc-probability"}, "N",
-        string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sampling.xtc_probability),
+        string_format("xtc probability (default: %.2f, 0.0 = disabled)", (double)params.sampling.xtc_probability),
         [](common_params & params, const std::string & value) {
             params.sampling.xtc_probability = std::stof(value);
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY;
@@ -1621,7 +1627,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--xtc-threshold"}, "N",
-        string_format("xtc threshold (default: %.1f, 1.0 = disabled)", (double)params.sampling.xtc_threshold),
+        string_format("xtc threshold (default: %.2f, 1.0 = disabled)", (double)params.sampling.xtc_threshold),
         [](common_params & params, const std::string & value) {
             params.sampling.xtc_threshold = std::stof(value);
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD;
@@ -1629,7 +1635,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--typical"}, "N",
-        string_format("locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)", (double)params.sampling.typ_p),
+        string_format("locally typical sampling, parameter p (default: %.2f, 1.0 = disabled)", (double)params.sampling.typ_p),
         [](common_params & params, const std::string & value) {
             params.sampling.typ_p = std::stof(value);
         }
@@ -1648,7 +1654,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--repeat-penalty"}, "N",
-        string_format("penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)", (double)params.sampling.penalty_repeat),
+        string_format("penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)", (double)params.sampling.penalty_repeat),
         [](common_params & params, const std::string & value) {
             params.sampling.penalty_repeat = std::stof(value);
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT;
@@ -1656,21 +1662,21 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--presence-penalty"}, "N",
-        string_format("repeat alpha presence penalty (default: %.1f, 0.0 = disabled)", (double)params.sampling.penalty_present),
+        string_format("repeat alpha presence penalty (default: %.2f, 0.0 = disabled)", (double)params.sampling.penalty_present),
         [](common_params & params, const std::string & value) {
             params.sampling.penalty_present = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--frequency-penalty"}, "N",
-        string_format("repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)", (double)params.sampling.penalty_freq),
+        string_format("repeat alpha frequency penalty (default: %.2f, 0.0 = disabled)", (double)params.sampling.penalty_freq),
         [](common_params & params, const std::string & value) {
             params.sampling.penalty_freq = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dry-multiplier"}, "N",
-        string_format("set DRY sampling multiplier (default: %.1f, 0.0 = disabled)", (double)params.sampling.dry_multiplier),
+        string_format("set DRY sampling multiplier (default: %.2f, 0.0 = disabled)", (double)params.sampling.dry_multiplier),
         [](common_params & params, const std::string & value) {
             params.sampling.dry_multiplier = std::stof(value);
         }
@@ -1751,14 +1757,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--dynatemp-range"}, "N",
-        string_format("dynamic temperature range (default: %.1f, 0.0 = disabled)", (double)params.sampling.dynatemp_range),
+        string_format("dynamic temperature range (default: %.2f, 0.0 = disabled)", (double)params.sampling.dynatemp_range),
         [](common_params & params, const std::string & value) {
             params.sampling.dynatemp_range = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
         {"--dynatemp-exp"}, "N",
-        string_format("dynamic temperature exponent (default: %.1f)", (double)params.sampling.dynatemp_exponent),
+        string_format("dynamic temperature exponent (default: %.2f)", (double)params.sampling.dynatemp_exponent),
         [](common_params & params, const std::string & value) {
             params.sampling.dynatemp_exponent = std::stof(value);
         }
@@ -1774,7 +1780,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--mirostat-lr"}, "N",
-        string_format("Mirostat learning rate, parameter eta (default: %.1f)", (double)params.sampling.mirostat_eta),
+        string_format("Mirostat learning rate, parameter eta (default: %.2f)", (double)params.sampling.mirostat_eta),
         [](common_params & params, const std::string & value) {
             params.sampling.mirostat_eta = std::stof(value);
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA;
@@ -1782,7 +1788,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--mirostat-ent"}, "N",
-        string_format("Mirostat target entropy, parameter tau (default: %.1f)", (double)params.sampling.mirostat_tau),
+        string_format("Mirostat target entropy, parameter tau (default: %.2f)", (double)params.sampling.mirostat_tau),
         [](common_params & params, const std::string & value) {
             params.sampling.mirostat_tau = std::stof(value);
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU;
@@ -1916,28 +1922,28 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_ARG_YARN_ORIG_CTX"));
     add_opt(common_arg(
         {"--yarn-ext-factor"}, "N",
-        string_format("YaRN: extrapolation mix factor (default: %.1f, 0.0 = full interpolation)", (double)params.yarn_ext_factor),
+        string_format("YaRN: extrapolation mix factor (default: %.2f, 0.0 = full interpolation)", (double)params.yarn_ext_factor),
         [](common_params & params, const std::string & value) {
             params.yarn_ext_factor = std::stof(value);
         }
     ).set_env("LLAMA_ARG_YARN_EXT_FACTOR"));
     add_opt(common_arg(
         {"--yarn-attn-factor"}, "N",
-        string_format("YaRN: scale sqrt(t) or attention magnitude (default: %.1f)", (double)params.yarn_attn_factor),
+        string_format("YaRN: scale sqrt(t) or attention magnitude (default: %.2f)", (double)params.yarn_attn_factor),
         [](common_params & params, const std::string & value) {
             params.yarn_attn_factor = std::stof(value);
         }
     ).set_env("LLAMA_ARG_YARN_ATTN_FACTOR"));
     add_opt(common_arg(
         {"--yarn-beta-slow"}, "N",
-        string_format("YaRN: high correction dim or alpha (default: %.1f)", (double)params.yarn_beta_slow),
+        string_format("YaRN: high correction dim or alpha (default: %.2f)", (double)params.yarn_beta_slow),
         [](common_params & params, const std::string & value) {
             params.yarn_beta_slow = std::stof(value);
         }
     ).set_env("LLAMA_ARG_YARN_BETA_SLOW"));
     add_opt(common_arg(
         {"--yarn-beta-fast"}, "N",
-        string_format("YaRN: low correction dim or beta (default: %.1f)", (double)params.yarn_beta_fast),
+        string_format("YaRN: low correction dim or beta (default: %.2f)", (double)params.yarn_beta_fast),
         [](common_params & params, const std::string & value) {
             params.yarn_beta_fast = std::stof(value);
         }
@@ -2194,18 +2200,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     add_opt(common_arg(
         {"--mmap"},
         {"--no-mmap"},
-        string_format("whether to memory-map model. Explicitly enabling mmap disables direct-io. (if mmap disabled, slower load but may reduce pageouts if not using mlock) (default: %s)", params.use_mmap ? "enabled" : "disabled"),
+        string_format("whether to memory-map model. (if mmap disabled, slower load but may reduce pageouts if not using mlock) (default: %s)", params.use_mmap ? "enabled" : "disabled"),
         [](common_params & params, bool value) {
             params.use_mmap = value;
-            if (value) {
-                params.use_direct_io = false;  // disable direct io when mmap is explicitly enabled
-            }
         }
     ).set_env("LLAMA_ARG_MMAP"));
     add_opt(common_arg(
         {"-dio", "--direct-io"},
         {"-ndio", "--no-direct-io"},
-        string_format("use DirectIO if available. Takes precedence over --mmap (default: %s)", params.use_direct_io ? "enabled" : "disabled"),
+        string_format("use DirectIO if available. (default: %s)", params.use_direct_io ? "enabled" : "disabled"),
         [](common_params & params, bool value) {
             params.use_direct_io = value;
         }
@@ -2561,7 +2564,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-hfd", "-hfrd", "--hf-repo-draft"}, "<user>/<model>[:quant]",
         "Same as --hf-repo, but for the draft model (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.speculative.model.hf_repo = value;
+            params.speculative.mparams_dft.hf_repo = value;
         }
     ).set_env("LLAMA_ARG_HFD_REPO"));
     add_opt(common_arg(
@@ -3331,14 +3334,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_DRAFT_MIN"));
     add_opt(common_arg(
         {"--draft-p-split"}, "P",
-        string_format("speculative decoding split probability (default: %.1f)", (double)params.speculative.p_split),
+        string_format("speculative decoding split probability (default: %.2f)", (double)params.speculative.p_split),
         [](common_params & params, const std::string & value) {
             params.speculative.p_split = std::stof(value);
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}).set_env("LLAMA_ARG_DRAFT_P_SPLIT"));
     add_opt(common_arg(
         {"--draft-p-min"}, "P",
-        string_format("minimum speculative decoding probability (greedy) (default: %.1f)", (double)params.speculative.p_min),
+        string_format("minimum speculative decoding probability (greedy) (default: %.2f)", (double)params.speculative.p_min),
         [](common_params & params, const std::string & value) {
             params.speculative.p_min = std::stof(value);
         }
@@ -3382,7 +3385,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-md", "--model-draft"}, "FNAME",
         "draft model for speculative decoding (default: unused)",
         [](common_params & params, const std::string & value) {
-            params.speculative.model.path = value;
+            params.speculative.mparams_dft.path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_MODEL_DRAFT"));
     add_opt(common_arg(
@@ -3392,6 +3395,68 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.replacements.push_back({ tgt, dft });
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
+    add_opt(common_arg(
+        {"--spec-type"}, "[none|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]",
+        string_format("type of speculative decoding to use when no draft model is provided (default: %s)\n",
+            common_speculative_type_to_str(params.speculative.type).c_str()),
+        [](common_params & params, const std::string & value) {
+            if (value == "none") {
+                params.speculative.type = COMMON_SPECULATIVE_TYPE_NONE;
+            } else if (value == "ngram-cache") {
+                params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_CACHE;
+            } else if (value == "ngram-simple") {
+                params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE;
+            } else if (value == "ngram-map-k") {
+                params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K;
+            } else if (value == "ngram-map-k4v") {
+                params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V;
+            } else if (value == "ngram-mod") {
+                params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MOD;
+            } else {
+                throw std::invalid_argument("unknown speculative decoding type without draft model");
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--spec-ngram-size-n"}, "N",
+        string_format("ngram size N for ngram-simple/ngram-map speculative decoding, length of lookup n-gram (default: %d)", params.speculative.ngram_size_n),
+        [](common_params & params, int value) {
+            if (value < 1 || value > 1024) {
+                throw std::invalid_argument("ngram size N must be between 1 and 1024 inclusive");
+            }
+            params.speculative.ngram_size_n = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--spec-ngram-size-m"}, "N",
+        string_format("ngram size M for ngram-simple/ngram-map speculative decoding, length of draft m-gram (default: %d)", params.speculative.ngram_size_m),
+        [](common_params & params, int value) {
+            if (value < 1 || value > 1024) {
+                throw std::invalid_argument("ngram size M must be between 1 and 1024 inclusive");
+            }
+            params.speculative.ngram_size_m = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--spec-ngram-check-rate"}, "N",
+        string_format("ngram check rate for ngram-simple/ngram-map speculative decoding (default: %d)", params.speculative.ngram_check_rate),
+        [](common_params & params, int value) {
+            if (value < 1) {
+                throw std::invalid_argument("ngram check rate must be at least 1");
+            }
+            params.speculative.ngram_check_rate = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--spec-ngram-min-hits"}, "N",
+        string_format("minimum hits for ngram-map speculative decoding (default: %d)", params.speculative.ngram_min_hits),
+        [](common_params & params, int value) {
+            if (value < 1) {
+                throw std::invalid_argument("ngram min hits must be at least 1");
+            }
+            params.speculative.ngram_min_hits = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-ctkd", "--cache-type-k-draft"}, "TYPE",
         string_format(
@@ -3618,8 +3683,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.model.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
             params.model.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
-            params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
-            params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
+            params.speculative.mparams_dft.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
+            params.speculative.mparams_dft.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
             params.port = 8012;
             params.n_ubatch = 1024;
             params.n_batch = 1024;
@@ -3634,8 +3699,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.model.hf_repo = "ggml-org/Qwen2.5-Coder-14B-Q8_0-GGUF";
             params.model.hf_file = "qwen2.5-coder-14b-q8_0.gguf";
-            params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
-            params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
+            params.speculative.mparams_dft.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
+            params.speculative.mparams_dft.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
             params.port = 8012;
             params.n_ubatch = 1024;
             params.n_batch = 1024;
diff --git a/common/chat-parser.cpp b/common/chat-parser.cpp
index 2f073512e0..29819e48d3 100644
--- a/common/chat-parser.cpp
+++ b/common/chat-parser.cpp
@@ -129,7 +129,7 @@ static void parse_json_tool_calls(
     }
 }
 
-common_chat_msg_parser::common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_syntax & syntax)
+common_chat_msg_parser::common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_parser_params & syntax)
     : input_(input), is_partial_(is_partial), syntax_(syntax)
 {
     result_.role = "assistant";
@@ -1611,7 +1611,7 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
     builder.finish();
 }
 
-common_chat_msg common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax) {
+common_chat_msg common_chat_parse(const std::string & input, bool is_partial, const common_chat_parser_params & syntax) {
     if (syntax.format == COMMON_CHAT_FORMAT_PEG_SIMPLE ||
         syntax.format == COMMON_CHAT_FORMAT_PEG_NATIVE ||
         syntax.format == COMMON_CHAT_FORMAT_PEG_CONSTRUCTED) {
@@ -1630,12 +1630,12 @@ common_chat_msg common_chat_parse(const std::string & input, bool is_partial, co
     }
     auto msg = builder.result();
     if (!is_partial) {
-        LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat<json>({msg}).at(0).dump().c_str());
+        LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat({msg}).at(0).dump().c_str());
     }
     return msg;
 }
 
-common_chat_msg common_chat_peg_parse(const common_peg_arena & parser, const std::string & input, bool is_partial, const common_chat_syntax & syntax) {
+common_chat_msg common_chat_peg_parse(const common_peg_arena & parser, const std::string & input, bool is_partial, const common_chat_parser_params & syntax) {
     if (parser.empty()) {
         throw std::runtime_error("Failed to parse due to missing parser definition.");
     }
@@ -1663,7 +1663,7 @@ common_chat_msg common_chat_peg_parse(const common_peg_arena & parser, const std
         mapper.from_ast(ctx.ast, result);
     }
     if (!is_partial) {
-        LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat<json>({msg}).at(0).dump().c_str());
+        LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat({msg}).at(0).dump().c_str());
     }
     return msg;
 }
diff --git a/common/chat-parser.h b/common/chat-parser.h
index 78c4b74c2d..3ed9c30a2b 100644
--- a/common/chat-parser.h
+++ b/common/chat-parser.h
@@ -5,7 +5,7 @@
 #include "json-partial.h"
 #include "regex-partial.h"
 
-#include <nlohmann/json.hpp>
+#include <nlohmann/json_fwd.hpp>
 
 #include <optional>
 #include <string>
@@ -19,20 +19,20 @@ class common_chat_msg_partial_exception : public std::runtime_error {
 class common_chat_msg_parser {
     std::string input_;
     bool is_partial_;
-    common_chat_syntax syntax_;
+    common_chat_parser_params syntax_; // TODO: rename to params
     std::string healing_marker_;
 
     size_t pos_ = 0;
     common_chat_msg result_;
 
   public:
-    common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
+    common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_parser_params & syntax);
     const std::string & input() const { return input_; }
     size_t pos() const { return pos_; }
     const std::string & healing_marker() const { return healing_marker_; }
     const bool & is_partial() const { return is_partial_; }
     const common_chat_msg & result() const { return result_; }
-    const common_chat_syntax & syntax() const { return syntax_; }
+    const common_chat_parser_params & syntax() const { return syntax_; }
 
     void move_to(size_t pos) {
         if (pos > input_.size()) {
diff --git a/common/chat.cpp b/common/chat.cpp
index 28721ac7da..2bf4632669 100644
--- a/common/chat.cpp
+++ b/common/chat.cpp
@@ -7,9 +7,6 @@
 #include "log.h"
 #include "regex-partial.h"
 
-// #include <minja/chat-template.hpp>
-// #include <minja/minja.hpp>
-
 #include "jinja/parser.h"
 #include "jinja/value.h"
 #include "jinja/runtime.h"
@@ -56,39 +53,73 @@ static bool has_content_or_tool_calls(const common_chat_msg & msg) {
     return !msg.content.empty() || !msg.tool_calls.empty();
 }
 
-template <>
-json common_chat_msg::to_json_oaicompat() const
-{
-    json message {
-        {"role", "assistant"},
-    };
-    if (!reasoning_content.empty()) {
-        message["reasoning_content"] = reasoning_content;
+json common_chat_msg::to_json_oaicompat(bool concat_typed_text) const {
+    if (!content.empty() && !content_parts.empty()) {
+        throw std::runtime_error("Cannot specify both content and content_parts");
     }
-    if (content.empty() && !tool_calls.empty()) {
-        message["content"] = json();
+    json jmsg {
+        {"role", role},
+    };
+    if (!content.empty()) {
+        jmsg["content"] = content;
+    } else if (!content_parts.empty()) {
+        if (concat_typed_text) {
+            std::string text;
+            for (const auto & part : content_parts) {
+                if (part.type != "text") {
+                    LOG_WRN("Ignoring content part type: %s\n", part.type.c_str());
+                    continue;
+                }
+                if (!text.empty()) {
+                    text += '\n';
+                }
+                text += part.text;
+            }
+            jmsg["content"] = text;
+        } else {
+            auto & parts = jmsg["content"] = json::array();
+            for (const auto & part : content_parts) {
+                parts.push_back({
+                    {"type", part.type},
+                    {"text", part.text},
+                });
+            }
+        }
     } else {
-        message["content"] = content;
+        jmsg["content"] = "";
+    }
+    if (!reasoning_content.empty()) {
+        jmsg["reasoning_content"] = reasoning_content;
+    }
+    if (!tool_name.empty()) {
+        jmsg["name"] = tool_name;
+    }
+    if (!tool_call_id.empty()) {
+        jmsg["tool_call_id"] = tool_call_id;
     }
     if (!tool_calls.empty()) {
-        auto arr = json::array();
-        for (const auto & tc : tool_calls) {
-            arr.push_back({
+        jmsg["tool_calls"] = json::array();
+        auto & jtool_calls = jmsg["tool_calls"];
+        for (const auto & tool_call : tool_calls) {
+            json tc {
                 {"type", "function"},
                 {"function", {
-                    {"name", tc.name},
-                    {"arguments", tc.arguments},
+                    {"name", tool_call.name},
+                    {"arguments", tool_call.arguments},
                 }},
-                {"id", tc.id},
-                // // Some templates generate and require an id (sometimes in a very specific format, e.g. Mistral Nemo).
-                // // We only generate a random id for the ones that don't generate one by themselves
-                // // (they also won't get to see it as their template likely doesn't use it, so it's all for the client)
-                // {"id", tc.id.empty() ? gen_tool_call_id() : tc.id},
-            });
+            };
+            if (!tool_call.id.empty()) {
+                tc["id"] = tool_call.id;
+            }
+            // Some templates generate and require an id (sometimes in a very specific format, e.g. Mistral Nemo).
+            // We only generate a random id for the ones that don't generate one by themselves
+            // (they also won't get to see it as their template likely doesn't use it, so it's all for the client)
+            // {"id", tc.id.empty() ? gen_tool_call_id() : tc.id},
+            jtool_calls.push_back(tc);
         }
-        message["tool_calls"] = arr;
     }
-    return message;
+
+    return jmsg;
 }
 
 std::vector<common_chat_msg_diff> common_chat_msg_diff::compute_diffs(const common_chat_msg & msg_prv, const common_chat_msg & msg_new) {
@@ -256,7 +287,6 @@ bool common_chat_templates_support_enable_thinking(const common_chat_templates *
     return rendered_no_thinking.prompt != rendered_with_thinking.prompt;
 }
 
-template <>
 std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messages) {
     std::vector<common_chat_msg> msgs;
 
@@ -350,80 +380,15 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
     return msgs;
 }
 
-template <>
 json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text) {
     json messages = json::array();
     for (const auto & msg : msgs) {
-        if (!msg.content.empty() && !msg.content_parts.empty()) {
-            throw std::runtime_error("Cannot specify both content and content_parts");
-        }
-        json jmsg {
-            {"role", msg.role},
-        };
-        if (!msg.content.empty()) {
-            jmsg["content"] = msg.content;
-        } else if (!msg.content_parts.empty()) {
-            if (concat_typed_text) {
-                std::string text;
-                for (const auto & part : msg.content_parts) {
-                    if (part.type != "text") {
-                        LOG_WRN("Ignoring content part type: %s\n", part.type.c_str());
-                        continue;
-                    }
-                    if (!text.empty()) {
-                        text += '\n';
-                    }
-                    text += part.text;
-                }
-                jmsg["content"] = text;
-            } else {
-                auto & parts = jmsg["content"] = json::array();
-                for (const auto & part : msg.content_parts) {
-                    parts.push_back({
-                        {"type", part.type},
-                        {"text", part.text},
-                    });
-                }
-            }
-        } else {
-            jmsg["content"] = "";
-        }
-        if (!msg.reasoning_content.empty()) {
-            jmsg["reasoning_content"] = msg.reasoning_content;
-        }
-        if (!msg.tool_name.empty()) {
-            jmsg["name"] = msg.tool_name;
-        }
-        if (!msg.tool_call_id.empty()) {
-            jmsg["tool_call_id"] = msg.tool_call_id;
-        }
-        if (!msg.tool_calls.empty()) {
-            auto & tool_calls = jmsg["tool_calls"] = json::array();
-            for (const auto & tool_call : msg.tool_calls) {
-                json tc {
-                    {"type", "function"},
-                    {"function", {
-                        {"name", tool_call.name},
-                        {"arguments", tool_call.arguments},
-                    }},
-                };
-                if (!tool_call.id.empty()) {
-                    tc["id"] = tool_call.id;
-                }
-                tool_calls.push_back(tc);
-            }
-        }
+        json jmsg = msg.to_json_oaicompat(concat_typed_text);
         messages.push_back(jmsg);
     }
     return messages;
 }
 
-template <>
-std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const std::string & messages) {
-    return common_chat_msgs_parse_oaicompat(json::parse(messages));
-}
-
-template <>
 std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & tools) {
     std::vector<common_chat_tool> result;
 
@@ -459,12 +424,6 @@ std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & too
     return result;
 }
 
-template <>
-std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const std::string & tools) {
-    return common_chat_tools_parse_oaicompat(json::parse(tools));
-}
-
-template <>
 json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools) {
     if (tools.empty()) {
         return json();
@@ -484,7 +443,7 @@ json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & t
     return result;
 }
 
-template <> json common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff) {
+json common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff) {
     json delta = json::object();
     if (!diff.reasoning_content_delta.empty()) {
         delta["reasoning_content"] = diff.reasoning_content_delta;
@@ -601,18 +560,18 @@ bool common_chat_templates_was_explicit(const struct common_chat_templates * tmp
     return tmpls->has_explicit_template;
 }
 
-const char * common_chat_templates_source(const struct common_chat_templates * tmpls, const char * variant) {
-    if (variant != nullptr) {
-        if (strcmp(variant, "tool_use") == 0) {
+std::string common_chat_templates_source(const struct common_chat_templates * tmpls, const std::string & variant) {
+    if (!variant.empty()) {
+        if (variant == "tool_use") {
             if (tmpls->template_tool_use) {
-                return tmpls->template_tool_use->source().c_str();
+                return tmpls->template_tool_use->source();
             }
-            return nullptr;
+            return "";
         } else {
-            LOG_DBG("%s: unknown template variant: %s\n", __func__, variant);
+            LOG_DBG("%s: unknown template variant: %s\n", __func__, variant.c_str());
         }
     }
-    return tmpls->template_default->source().c_str();
+    return tmpls->template_default->source();
 }
 
 common_chat_templates_ptr common_chat_templates_init(
@@ -812,10 +771,12 @@ static std::string apply(
 
     nlohmann::ordered_json inp = nlohmann::ordered_json{
         {"messages", messages_override.has_value() ? *messages_override : inputs.messages},
-        {"tools", tools_override.has_value() ? *tools_override : inputs.tools},
         {"bos_token", tmpl.bos_token()},
         {"eos_token", tmpl.eos_token()},
     };
+    if (tools_override.has_value() || !inputs.tools.empty()) {
+        inp["tools"] = tools_override.has_value() ? *tools_override : inputs.tools;
+    }
     if (inputs.extra_context.is_object()) {
         // TODO: do we need to merge, or replacing is fine?
         for (const auto & [k, v] : inputs.extra_context.items()) {
@@ -831,9 +792,6 @@ static std::string apply(
     if (inputs.add_generation_prompt) {
         inp["add_generation_prompt"] = true;
     }
-    if (inp["tools"].is_null()) {
-        inp["tools"] = json::array();
-    }
 
     jinja::global_from_json(ctx, inp, inputs.mark_input);
 
@@ -2260,12 +2218,11 @@ static common_chat_params common_chat_params_init_glm_4_5(const common_chat_temp
 static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct templates_params & inputs) {
     LOG_DBG("%s\n", __func__);
     common_chat_params data;
-    const std::optional<json> tools_override = json();
     const std::optional<json> additional_context = json {
         {"datetime", format_time(inputs.now, "%b %d %Y %H:%M:%S GMT")},
         {"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))},
     };
-    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, tools_override, additional_context);
+    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override =*/ std::nullopt, additional_context);
     if (inputs.tools.is_array() && !inputs.tools.empty()) {
         data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
         data.grammar = build_grammar([&](const common_grammar_builder & builder) {
@@ -2614,20 +2571,165 @@ static common_chat_params common_chat_params_init_granite(const common_chat_temp
 static common_chat_params common_chat_params_init_solar_open(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
 
-    // TODO: Reasoning effort
-    json additional_context = {};
+    // Copy `reasoning_content` to `reasoning`
+    auto adjusted_messages = json::array();
+    for (const auto & msg : inputs.messages) {
+        if (msg.contains("reasoning_content") && msg.at("reasoning_content").is_string()) {
+            auto adjusted_message = msg;
+            adjusted_message["reasoning"] = msg.at("reasoning_content");
+            adjusted_message.erase("reasoning_content");
+            adjusted_messages.push_back(adjusted_message);
+        } else {
+            adjusted_messages.push_back(msg);
+        }
+    }
 
-    data.prompt = apply(tmpl, inputs, std::nullopt, std::nullopt, additional_context);
-    data.format = COMMON_CHAT_FORMAT_SOLAR_OPEN;
+    auto has_tools = inputs.tools.is_array() && !inputs.tools.empty();
+    auto include_grammar = true;
 
+    auto prompt = apply(tmpl, inputs, /* messages_override= */ adjusted_messages);
+
+    // Check if we need to replace the flush token with end token during inference and without generation prompt.
+    if (inputs.is_inference && !inputs.add_generation_prompt) {
+        static constexpr std::string_view return_token = "<|flush|>";
+        static constexpr std::string_view end_token    = "<|end|>";
+        if (size_t pos = prompt.rfind(return_token); pos != std::string::npos) {
+            prompt.replace(pos, return_token.length(), end_token);
+        }
+    }
+
+    data.prompt = prompt;
+    data.format = COMMON_CHAT_FORMAT_PEG_NATIVE;
     data.preserved_tokens = {
         "<|think|>",
         "<|content|>",
         "<|begin|>",
         "<|end|>",
+        "<|tool_calls|>",
+        "<|tool_call:begin|>",
+        "<|tool_call:end|>",
+        "<|tool_call:name|>",
+        "<|tool_call:args|>",
     };
 
-    // TODO: Tool calling
+    auto parser = build_chat_peg_native_parser([&](common_chat_peg_native_builder & p) {
+        auto lit_think = p.atomic(p.literal("<|think|>"));
+        auto lit_assistant_begin = p.atomic(p.literal("<|begin|>assistant"));
+        auto lit_content = p.atomic(p.literal("<|content|>"));
+        auto lit_end = p.atomic(p.literal("<|end|>"));
+        auto parser_until_end = p.until("<|end|>");
+
+        // reasoning <- "<|think|>" (!"<|end|>" .)*
+        auto parser_reasoning = p.rule("reasoning", lit_think + p.reasoning(parser_until_end));
+
+        // content <- "<|content|>" (!"<|end|>" .)*
+        auto parser_content = p.rule("content", lit_content + p.content(parser_until_end));
+
+        // wrap_choice(items) <- item-choice wrapped*
+        // item-choice        <- items[0] / ... / items[n]
+        // wrapped            <- "<|end|><|begin|>assistant" item-choice
+        auto wrap_choice = [&](const std::vector<common_peg_parser> & items) {
+            auto choice = p.choice(items);
+            return choice + p.zero_or_more(lit_end + lit_assistant_begin + choice);
+        };
+
+        // wrap_seq(items) <- item[0] "<|end|><|begin|>assistant" item[1] ...
+        auto wrap_seq = [&](const std::vector<common_peg_parser> & items) {
+            auto seq = p.sequence();
+            for (auto i = 0u; i < items.size(); i++) {
+                if (i == 0) {
+                    seq += items[i];
+                    continue;
+                }
+                seq += lit_end + lit_assistant_begin + items[i];
+            }
+            return seq;
+        };
+
+        // Response format parser
+        if (inputs.json_schema.is_object() && !inputs.json_schema.empty()) {
+            auto parser_response_format = lit_content + p.content(p.schema(p.json(), "response-format", inputs.json_schema));
+            return p.choice({
+                wrap_seq({parser_reasoning, parser_response_format}),
+                wrap_seq({parser_response_format})
+            });
+        }
+
+        auto lit_tool_call_begin = p.literal("<|tool_call:begin|>");
+        auto lit_tool_call_name = p.literal("<|tool_call:name|>");
+        auto lit_tool_call_args = p.literal("<|tool_call:args|>");
+        auto lit_tool_call_end = p.literal("<|tool_call:end|>");
+
+        // Tool call parser
+        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
+            auto parser_tool_call = p.choice();
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                std::string name = function.at("name");
+                const auto & schema = function.at("parameters");
+
+                // tool(name, schema) <- name "<|tool_call:args|>" schema
+                parser_tool_call |= p.rule("tool-" + name,
+                    p.atomic(p.tool_name(p.literal(name)) + lit_tool_call_args)
+                    + p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", schema)));
+            });
+
+            auto min_calls = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
+            auto max_calls = inputs.parallel_tool_calls ? -1 : 1;
+
+            // tool-calls  <- "<|tool_calls|>" tool-call+
+            // tool-call   <- "<|tool_call:begin|> call-id "<|tool_call:name|>" &([^<]+ "<|tool_call:args|>") tool-choice "<|tool_call:end|>"
+            // call-id     <- [a-zA-Z0-9_-]+
+            // tool-choice <- tool(t[0].name, t[0].schema) / ... / tool(t[n].name, t[n].schema)
+            auto parser_tool_calls = p.trigger_rule("tool-calls",
+                p.atomic(p.literal("<|tool_calls|>"))
+                + p.repeat(
+                    p.tool_open(
+                        lit_tool_call_begin
+                        + p.tool_id(p.chars("[a-zA-Z0-9_-]", 1, -1))
+                        + lit_tool_call_name
+                        + p.peek(p.chars("[^<]", 1, -1) + lit_tool_call_args))
+                    + parser_tool_call
+                    + p.tool_close(lit_tool_call_end),
+                /* min = */ 1,
+                /* max = */ max_calls));
+
+            if (min_calls == 1) {
+                // If required, then try any combination of the reasoning, content, and tool call
+                return p.choice({
+                    wrap_seq({parser_reasoning, parser_content, parser_tool_calls}),
+                    wrap_seq({parser_reasoning, parser_tool_calls}),
+                    wrap_seq({parser_content, parser_tool_calls}),
+                    wrap_seq({parser_tool_calls})
+                });
+            }
+
+            return wrap_choice({parser_reasoning, parser_content, parser_tool_calls});
+        }
+
+        // Content only parser
+        include_grammar = false;
+        return wrap_choice({parser_reasoning, parser_content});
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto schema = function.at("parameters");
+                builder.resolve_refs(schema);
+            });
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            {COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|tool_calls|>"}
+        };
+    }
 
     return data;
 }
@@ -2691,6 +2793,51 @@ static common_chat_params common_chat_params_init_exaone_moe(const common_chat_t
     return data;
 }
 
+static common_chat_params common_chat_params_init_translate_gemma(const common_chat_template & tmpl, const struct templates_params & inputs) {
+    common_chat_params data;
+
+    // This template does not support tools or reasoning
+    // we just need to transform the messages into the correct schema
+
+    templates_params inputs_new = inputs;
+    json & messages = inputs_new.messages;
+
+    // default to chat_template_kwargs, or en-GB if not specified
+    std::string default_src_lang = inputs.extra_context.value("source_lang_code", "en-GB");
+    std::string default_tgt_lang = inputs.extra_context.value("target_lang_code", "en-GB");
+
+    GGML_ASSERT(messages.is_array());
+    for (auto & message : messages) {
+        if (message.contains("role") && message["role"].get<std::string>() != "user") {
+            continue;
+        }
+        if (!message.contains("content")) {
+            message["content"] = json::array();
+        }
+        if (message.contains("content") && !message["content"].is_array()) {
+            auto content_str = message["content"].get<std::string>();
+            // default to en-GB if not specified (to make common_chat_format_example works)
+            auto src_lang = message.contains("source_lang_code")
+                        ? message["source_lang_code"].get<std::string>() : default_src_lang;
+            auto tgt_lang = message.contains("target_lang_code")
+                        ? message["target_lang_code"].get<std::string>() : default_tgt_lang;
+            message["content"] = json::array({
+                json{
+                    {"type", "text"},
+                    {"text", content_str},
+                    {"source_lang_code", src_lang},
+                    {"target_lang_code", tgt_lang},
+                }
+            });
+        }
+    }
+
+    data.prompt = apply(tmpl, inputs_new, std::nullopt, std::nullopt);
+    data.format = COMMON_CHAT_FORMAT_GENERIC;
+
+    return data;
+}
+
 static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
     data.prompt = apply(tmpl, inputs);
@@ -2867,13 +3014,13 @@ static common_chat_params common_chat_templates_apply_jinja(
     const struct common_chat_templates_inputs & inputs)
 {
     templates_params params;
-    params.tools = common_chat_tools_to_json_oaicompat<json>(inputs.tools);
+    params.tools = common_chat_tools_to_json_oaicompat(inputs.tools);
     const auto & tmpl = params.tools.is_array() && tmpls->template_tool_use
         ? *tmpls->template_tool_use
         : *tmpls->template_default;
     const auto & src = tmpl.source();
     const auto & caps = tmpl.original_caps();
-    params.messages = common_chat_msgs_to_json_oaicompat<json>(inputs.messages, /* concat_text= */ !tmpl.original_caps().requires_typed_content);
+    params.messages = common_chat_msgs_to_json_oaicompat(inputs.messages, /* concat_text= */ !tmpl.original_caps().requires_typed_content);
     params.add_generation_prompt = inputs.add_generation_prompt;
     params.tool_choice = inputs.tool_choice;
     params.reasoning_format = inputs.reasoning_format;
@@ -2943,6 +3090,10 @@ static common_chat_params common_chat_templates_apply_jinja(
         src.find("<arg_value>") != std::string::npos &&
         params.json_schema.is_null()) {
         workaround::func_args_not_string(params.messages);
+        if (!params.extra_context.contains("clear_thinking")) {
+            // by default, do not clear reasoning_content (added since GLM-4.7)
+            params.extra_context["clear_thinking"] = false;
+        }
         return common_chat_params_init_glm_4_5(tmpl, params);
     }
 
@@ -3035,6 +3186,13 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_apriel_1_5(tmpl, params);
     }
 
+    // Solar Open
+    if (src.find("<|tool_response:begin|>") != std::string::npos &&
+        src.find("<|tool_response:name|>") != std::string::npos &&
+        src.find("<|tool_response:result|>") != std::string::npos) {
+        return common_chat_params_init_solar_open(tmpl, params);
+    }
+
     // Use generic handler when mixing tools + JSON schema.
     // TODO: support that mix in handlers below.
     if ((params.tools.is_array() && params.json_schema.is_object())) {
@@ -3082,6 +3240,12 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_solar_open(tmpl, params);
     }
 
+    // TranslateGemma
+    if (src.find("[source_lang_code]") != std::string::npos &&
+        src.find("[target_lang_code]") != std::string::npos) {
+        return common_chat_params_init_translate_gemma(tmpl, params);
+    }
+
     // Plain handler (no tools)
     if (params.tools.is_null() || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
         return common_chat_params_init_without_tools(tmpl, params);
@@ -3174,3 +3338,9 @@ common_chat_params common_chat_templates_apply(
         ? common_chat_templates_apply_jinja(tmpls, inputs)
         : common_chat_templates_apply_legacy(tmpls, inputs);
 }
+
+std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_templates * chat_templates) {
+    GGML_ASSERT(chat_templates != nullptr);
+    GGML_ASSERT(chat_templates->template_default != nullptr);
+    return chat_templates->template_default->caps.to_map();
+}
diff --git a/common/chat.h b/common/chat.h
index 454085e90e..24aa4aab5c 100644
--- a/common/chat.h
+++ b/common/chat.h
@@ -10,6 +10,8 @@
 #include <vector>
 #include <map>
 
+#include <nlohmann/json_fwd.hpp>
+
 struct common_chat_templates;
 
 struct common_chat_tool_call {
@@ -26,6 +28,11 @@ struct common_chat_msg_content_part {
     std::string type;
     std::string text;
 
+    // TODO @ngxson : no known chat templates support reasoning_content in content parts yet
+    //                this can be useful for models with interleaved thinking (like Kimi-K2)
+    //                if you see any templates explicitly support this, please ping me
+    // std::string reasoning_content;
+
     bool operator==(const common_chat_msg_content_part & other) const {
         return type == other.type && text == other.text;
     }
@@ -40,7 +47,7 @@ struct common_chat_msg {
     std::string tool_name;
     std::string tool_call_id;
 
-    template <class T> T to_json_oaicompat() const;
+    nlohmann::ordered_json to_json_oaicompat(bool concat_typed_text = false) const;
 
     bool empty() const {
         return content.empty() && content_parts.empty() && tool_calls.empty() && reasoning_content.empty() && tool_name.empty() && tool_call_id.empty();
@@ -145,7 +152,7 @@ struct common_chat_templates_inputs {
     std::vector<common_chat_tool> tools;
     common_chat_tool_choice tool_choice = COMMON_CHAT_TOOL_CHOICE_AUTO;
     bool parallel_tool_calls = false;
-    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_NONE;
+    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_NONE; // TODO: refactor this to "bool enable_thinking"
     bool enable_thinking = true;
     std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
     std::map<std::string, std::string> chat_template_kwargs;
@@ -165,14 +172,21 @@ struct common_chat_params {
     std::string                         parser;
 };
 
-struct common_chat_syntax {
+// per-message parsing syntax
+// should be derived from common_chat_params
+struct common_chat_parser_params {
     common_chat_format       format                = COMMON_CHAT_FORMAT_CONTENT_ONLY;
-    common_reasoning_format  reasoning_format      = COMMON_REASONING_FORMAT_NONE;
+    common_reasoning_format  reasoning_format      = COMMON_REASONING_FORMAT_NONE; // TODO: refactor this to "bool parse_reasoning"
     // Whether reasoning_content should be inlined in the content (e.g. for reasoning_format=deepseek in stream mode)
     bool                     reasoning_in_content  = false;
     bool                     thinking_forced_open  = false;
     bool                     parse_tool_calls      = true;
     common_peg_arena         parser                = {};
+    common_chat_parser_params() = default;
+    common_chat_parser_params(const common_chat_params & chat_params) {
+        format               = chat_params.format;
+        thinking_forced_open = chat_params.thinking_forced_open;
+    }
 };
 
 // Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
@@ -191,7 +205,7 @@ common_chat_templates_ptr common_chat_templates_init(
                                            const std::string & eos_token_override = "");
 
 bool         common_chat_templates_was_explicit(const struct common_chat_templates * tmpls);
-const char * common_chat_templates_source(const struct common_chat_templates * tmpls, const char * variant = nullptr);
+std::string  common_chat_templates_source(const struct common_chat_templates * tmpls, const std::string & variant = "");
 
 
 struct common_chat_params      common_chat_templates_apply(
@@ -213,23 +227,25 @@ std::string common_chat_format_example(
     const std::map<std::string, std::string> & chat_template_kwargs);
 
 const char*               common_chat_format_name(common_chat_format format);
-const char*               common_reasoning_format_name(common_reasoning_format format);
-common_reasoning_format   common_reasoning_format_from_name(const std::string & format);
-common_chat_msg           common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
-common_chat_msg           common_chat_peg_parse(const common_peg_arena & parser, const std::string & input, bool is_partial, const common_chat_syntax & syntax);
+common_chat_msg           common_chat_parse(const std::string & input, bool is_partial, const common_chat_parser_params & syntax);
+common_chat_msg           common_chat_peg_parse(const common_peg_arena & parser, const std::string & input, bool is_partial, const common_chat_parser_params & syntax);
+
+// used by arg and server
+const char *             common_reasoning_format_name(common_reasoning_format format);
+common_reasoning_format  common_reasoning_format_from_name(const std::string & format);
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice);
 
 bool common_chat_templates_support_enable_thinking(const common_chat_templates * chat_templates);
 
 // Parses a JSON array of messages in OpenAI's chat completion API format.
-// T can be std::string containing JSON or nlohmann::ordered_json
-template <class T> std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const T & messages);
-template <class T> T common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
+std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const nlohmann::ordered_json & messages);
+nlohmann::ordered_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
 
-// Parses a JSON array of tools in OpenAI's chat completion tool call API format.
-// T can be std::string containing JSON or nlohmann::ordered_json
-template <class T> std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const T & tools);
-template <class T> T common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
+std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);
+nlohmann::ordered_json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
 
-template <class T> T common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff);
+nlohmann::ordered_json common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff);
+
+// get template caps, useful for reporting to server /props endpoint
+std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_templates * chat_templates);
diff --git a/common/common.cpp b/common/common.cpp
index 26250abb6c..3aa396127c 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -1097,7 +1097,10 @@ common_init_result::common_init_result(common_params & params) :
     if (params.fit_params) {
         LOG_INF("%s: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on\n", __func__);
         llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
-            params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target.data(), params.fit_params_min_ctx,
+            params.tensor_split,
+            params.tensor_buft_overrides.data(),
+            params.fit_params_target.data(),
+            params.fit_params_min_ctx,
             params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
     }
 
@@ -1208,10 +1211,6 @@ std::vector<llama_adapter_lora_ptr> & common_init_result::lora() {
     return pimpl->lora;
 }
 
-void common_init_result::free_context() {
-    pimpl->context.reset();
-}
-
 common_init_result_ptr common_init_from_params(common_params & params) {
     common_init_result_ptr res(new common_init_result(params));
 
diff --git a/common/common.h b/common/common.h
index b9566df62c..398ebb0960 100644
--- a/common/common.h
+++ b/common/common.h
@@ -57,6 +57,8 @@ extern const char * LLAMA_COMMIT;
 extern const char * LLAMA_COMPILER;
 extern const char * LLAMA_BUILD_TARGET;
 
+const static std::string build_info("b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT);
+
 struct common_control_vector_load_info;
 
 //
@@ -162,6 +164,17 @@ enum common_params_sampling_config : uint64_t {
     COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA    = 1 << 11,
 };
 
+enum common_speculative_type {
+    COMMON_SPECULATIVE_TYPE_NONE,          // no speculative decoding
+    COMMON_SPECULATIVE_TYPE_DRAFT,         // draft model
+    COMMON_SPECULATIVE_TYPE_EAGLE3,        // eagle draft model
+    COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE,  // simple self-speculative decoding
+    COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K,   // self-speculative decoding with n-gram keys only
+    COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, // self-speculative decoding with n-gram keys and 4 m-gram values
+    COMMON_SPECULATIVE_TYPE_NGRAM_MOD,
+    COMMON_SPECULATIVE_TYPE_NGRAM_CACHE,   // self-speculative decoding with 3-level n-gram cache
+    COMMON_SPECULATIVE_TYPE_COUNT          // number of types, unknown type
+};
 
 // sampling parameters
 struct common_params_sampling {
@@ -240,17 +253,40 @@ struct common_params_model {
     std::string name        = ""; // in format <user>/<model>[:<tag>] (tag is optional)     // NOLINT
 };
 
-struct common_params_speculative {
-    std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
+struct common_ngram_mod;
 
-    int32_t n_ctx        =     0; // draft context size
-    int32_t n_max        =    16; // maximum number of tokens to draft during speculative decoding
-    int32_t n_min        =     0; // minimum number of draft tokens to use for speculative decoding
-    int32_t n_gpu_layers =    -1; // number of layers to store in VRAM for the draft model (-1 - use default)
-    float   p_split      =  0.1f; // speculative decoding split probability
-    float   p_min        = 0.75f; // minimum speculative decoding probability (greedy)
-    std::vector<std::pair<std::string, std::string>> replacements; // main to speculative model replacements
-    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
+struct common_params_speculative {
+    common_speculative_type type = COMMON_SPECULATIVE_TYPE_NONE; // type of speculative decoding
+
+    // general-purpose speculative decoding parameters
+
+    int32_t n_max   = 16; // maximum number of tokens to draft during speculative decoding
+    int32_t n_min   = 0; // minimum number of draft tokens to use for speculative decoding
+    float   p_split = 0.1f; // speculative decoding split probability
+    float   p_min   = 0.75f; // minimum speculative decoding probability (greedy)
+
+    // ngram-based speculative decoding
+
+    uint16_t ngram_size_n     = 12; // ngram size for lookup
+    uint16_t ngram_size_m     = 48; // mgram size for speculative tokens
+    uint16_t ngram_check_rate =  1; // check rate for ngram lookup
+    uint16_t ngram_min_hits   =  1; // minimum hits at ngram/mgram lookup for mgram to be proposed
+
+    std::shared_ptr<common_ngram_mod> ngram_mod;
+
+    std::string lookup_cache_static;  // path of static ngram cache file for lookup decoding           // NOLINT
+    std::string lookup_cache_dynamic; // path of dynamic ngram cache file for lookup decoding          // NOLINT
+
+    // draft-model speculative decoding
+
+    struct common_params_model mparams_dft;
+
+    llama_model * model_dft = nullptr; // a llama_model that can be shared by multiple speculative contexts
+
+    llama_context_params cparams_dft; // these are the parameters for the draft llama_context
+
+    int32_t n_ctx        = 0;  // draft context size
+    int32_t n_gpu_layers = -1; // number of layers to store in VRAM for the draft model (-1 - use default)
 
     ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
     ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
@@ -258,7 +294,14 @@ struct common_params_speculative {
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
 
-    struct common_params_model model;
+    std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
+
+    std::vector<std::pair<std::string, std::string>> replacements; // main to speculative model replacements
+    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
+
+    bool has_dft() const {
+        return !mparams_dft.path.empty() || !mparams_dft.hf_repo.empty();
+    }
 };
 
 struct common_params_vocoder {
@@ -284,6 +327,7 @@ struct common_params_diffusion {
 };
 
 // reasoning API response format (not to be confused as chat template's reasoning format)
+// only used by server
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
     COMMON_REASONING_FORMAT_AUTO,            // Same as deepseek, using `message.reasoning_content`
@@ -375,8 +419,6 @@ struct common_params {
     std::string path_prompt_cache    = ""; // path to file for saving/loading prompt eval state             // NOLINT
     std::string input_prefix         = ""; // string to prefix user inputs with                             // NOLINT
     std::string input_suffix         = ""; // string to suffix user inputs with                             // NOLINT
-    std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding           // NOLINT
-    std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding          // NOLINT
     std::string logits_file          = ""; // file for saving *all* logits                                  // NOLINT
 
     // llama-debug specific options
@@ -435,7 +477,7 @@ struct common_params {
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool use_mmap          = true;  // enable mmap to use filesystem cache
-    bool use_direct_io     = true;  // read from disk without buffering for faster model loading
+    bool use_direct_io     = false; // read from disk without buffering
     bool use_mlock         = false; // use mlock to keep model in memory
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
@@ -572,10 +614,6 @@ struct common_params {
     // return false from callback to abort model loading or true to continue
     llama_progress_callback load_progress_callback = NULL;
     void *                  load_progress_callback_user_data = NULL;
-
-    bool has_speculative() const {
-        return !speculative.model.path.empty() || !speculative.model.hf_repo.empty();
-    }
 };
 
 // call once at the start of a program if it uses libcommon
@@ -711,8 +749,6 @@ struct common_init_result {
 
     std::vector<llama_adapter_lora_ptr> & lora();
 
-    void free_context();
-
 private:
     struct impl;
     std::unique_ptr<impl> pimpl;
diff --git a/common/download.cpp b/common/download.cpp
index a37780421a..57f29a23ba 100644
--- a/common/download.cpp
+++ b/common/download.cpp
@@ -314,23 +314,26 @@ static bool common_pull_file(httplib::Client & cli,
 
 // download one single file from remote URL to local path
 // returns status code or -1 on error
-static int common_download_file_single_online(const std::string & url,
-                                               const std::string & path,
-                                               const std::string & bearer_token,
-                                               const common_header_list & custom_headers) {
+static int common_download_file_single_online(const std::string        & url,
+                                              const std::string        & path,
+                                              const std::string        & bearer_token,
+                                              const common_header_list & custom_headers) {
     static const int max_attempts        = 3;
     static const int retry_delay_seconds = 2;
 
     auto [cli, parts] = common_http_client(url);
 
-    httplib::Headers default_headers = {{"User-Agent", "llama-cpp"}};
-    if (!bearer_token.empty()) {
-        default_headers.insert({"Authorization", "Bearer " + bearer_token});
-    }
+    httplib::Headers headers;
     for (const auto & h : custom_headers) {
-        default_headers.emplace(h.first, h.second);
+        headers.emplace(h.first, h.second);
     }
-    cli.set_default_headers(default_headers);
+    if (headers.find("User-Agent") == headers.end()) {
+        headers.emplace("User-Agent", "llama-cpp/" + build_info);
+    }
+    if (!bearer_token.empty()) {
+        headers.emplace("Authorization", "Bearer " + bearer_token);
+    }
+    cli.set_default_headers(headers);
 
     const bool file_exists = std::filesystem::exists(path);
 
@@ -437,10 +440,12 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string
                                                              const common_remote_params & params) {
     auto [cli, parts] = common_http_client(url);
 
-    httplib::Headers headers = {{"User-Agent", "llama-cpp"}};
-
-    for (const auto & header : params.headers) {
-        headers.emplace(header.first, header.second);
+    httplib::Headers headers;
+    for (const auto & h : params.headers) {
+        headers.emplace(h.first, h.second);
+    }
+    if (headers.find("User-Agent") == headers.end()) {
+        headers.emplace("User-Agent", "llama-cpp/" + build_info);
     }
 
     if (params.timeout > 0) {
diff --git a/common/http.h b/common/http.h
index 8e29787dcc..e8ed56f952 100644
--- a/common/http.h
+++ b/common/http.h
@@ -57,6 +57,17 @@ static std::pair<httplib::Client, common_http_url> common_http_client(const std:
         throw std::runtime_error("error: invalid URL format");
     }
 
+#ifndef CPPHTTPLIB_OPENSSL_SUPPORT
+    if (parts.scheme == "https") {
+        throw std::runtime_error(
+            "HTTPS is not supported. Please rebuild with one of:\n"
+            "  -DLLAMA_BUILD_BORINGSSL=ON\n"
+            "  -DLLAMA_BUILD_LIBRESSL=ON\n"
+            "  -DLLAMA_OPENSSL=ON (default, requires OpenSSL dev files installed)"
+        );
+    }
+#endif
+
     httplib::Client cli(parts.scheme + "://" + parts.host);
 
     if (!parts.user.empty()) {
diff --git a/common/jinja/caps.cpp b/common/jinja/caps.cpp
index 61deccd1f5..f27490f1fb 100644
--- a/common/jinja/caps.cpp
+++ b/common/jinja/caps.cpp
@@ -61,14 +61,23 @@ static void caps_print_stats(value & v, const std::string & path) {
                 ops.c_str());
 }
 
+std::map<std::string, bool> caps::to_map() const {
+    return {
+        {"requires_typed_content", requires_typed_content},
+        {"supports_tools", supports_tools},
+        {"supports_tool_calls", supports_tool_calls},
+        {"supports_parallel_tool_calls", supports_parallel_tool_calls},
+        {"supports_system_role", supports_system_role},
+        {"supports_preserve_reasoning", supports_preserve_reasoning},
+    };
+}
+
 std::string caps::to_string() const {
     std::ostringstream ss;
     ss << "Caps(\n";
-    ss << "  requires_typed_content=" << requires_typed_content << "\n";
-    ss << "  supports_tools=" << supports_tools << "\n";
-    ss << "  supports_tool_calls=" << supports_tool_calls << "\n";
-    ss << "  supports_parallel_tool_calls=" << supports_parallel_tool_calls << "\n";
-    ss << "  supports_system_role=" << supports_system_role << "\n";
+    for (const auto & [key, value] : to_map()) {
+        ss << "  " << key << "=" << (value ? "true" : "false") << "\n";
+    }
     ss << ")";
     return ss.str();
 }
@@ -229,6 +238,40 @@ caps caps_get(jinja::program & prog) {
         }
     );
 
+    // case: preserve reasoning content in chat history
+    caps_try_execute(
+        prog,
+        [&]() {
+            // messages
+            return json::array({
+                {
+                    {"role", "user"},
+                    {"content", "User message"}
+                },
+                {
+                    {"role", "assistant"},
+                    {"content", "Assistant message"},
+                    {"reasoning_content", "Reasoning content"}
+                },
+                {
+                    {"role", "user"},
+                    {"content", "User message"}
+                },
+            });
+        },
+        [&]() {
+            // tools
+            return json::array();
+        },
+        [&](bool, value & messages, value &) {
+            auto & content = messages->at(1)->at("reasoning_content");
+            caps_print_stats(content, "messages[1].reasoning_content");
+            if (content->stats.used) {
+                result.supports_preserve_reasoning = true;
+            }
+        }
+    );
+
     JJ_DEBUG("%s\n", result.to_string().c_str());
 
     return result;
diff --git a/common/jinja/caps.h b/common/jinja/caps.h
index deb2df180f..77df117baa 100644
--- a/common/jinja/caps.h
+++ b/common/jinja/caps.h
@@ -3,6 +3,7 @@
 #include "runtime.h"
 
 #include <string>
+#include <map>
 
 namespace jinja {
 
@@ -11,14 +12,17 @@ struct caps {
     bool supports_tool_calls = true;
     bool supports_system_role = true;
     bool supports_parallel_tool_calls = true;
+    bool supports_preserve_reasoning = false; // support assistant message with reasoning_content
 
     bool requires_typed_content = false; // default: use string content
 
+    // for reporting on server
+    std::map<std::string, bool> to_map() const;
+
     // for debugging
     std::string to_string() const;
 };
 
 caps caps_get(jinja::program & prog);
-void debug_print_caps(const caps & c);
 
 } // namespace jinja
diff --git a/common/jinja/lexer.cpp b/common/jinja/lexer.cpp
index 85eaa1a76b..598982c2fe 100644
--- a/common/jinja/lexer.cpp
+++ b/common/jinja/lexer.cpp
@@ -91,6 +91,16 @@ lexer_result lexer::tokenize(const std::string & source) {
         return str;
     };
 
+    auto consume_numeric = [&]() -> std::string {
+        std::string num = consume_while(is_integer);
+        if (pos < src.size() && src[pos] == '.' && pos + 1 < src.size() && is_integer(src[pos + 1])) {
+            ++pos; // Consume '.'
+            std::string frac = consume_while(is_integer);
+            num += "." + frac;
+        }
+        return num;
+    };
+
     auto next_pos_is = [&](std::initializer_list<char> chars, size_t n = 1) -> bool {
         if (pos + n >= src.size()) return false;
         for (char c : chars) {
@@ -258,7 +268,7 @@ lexer_result lexer::tokenize(const std::string & source) {
                     ++pos; // Consume the operator
 
                     // Check for numbers following the unary operator
-                    std::string num = consume_while(is_integer);
+                    std::string num = consume_numeric();
                     std::string value = std::string(1, ch) + num;
                     token::type t = num.empty() ? token::unary_operator : token::numeric_literal;
                     // JJ_DEBUG("consumed unary operator or numeric literal: '%s'", value.c_str());
@@ -307,12 +317,7 @@ lexer_result lexer::tokenize(const std::string & source) {
         // Numbers
         if (is_integer(ch)) {
             start_pos = pos;
-            std::string num = consume_while(is_integer);
-            if (pos < src.size() && src[pos] == '.' && pos + 1 < src.size() && is_integer(src[pos + 1])) {
-                ++pos; // Consume '.'
-                std::string frac = consume_while(is_integer);
-                num += "." + frac;
-            }
+            std::string num = consume_numeric();
             // JJ_DEBUG("consumed numeric literal: '%s'", num.c_str());
             tokens.push_back({token::numeric_literal, num, start_pos});
             continue;
diff --git a/common/jinja/runtime.cpp b/common/jinja/runtime.cpp
index ba07f7a6d9..f234d9284f 100644
--- a/common/jinja/runtime.cpp
+++ b/common/jinja/runtime.cpp
@@ -44,6 +44,12 @@ static std::string get_line_col(const std::string & source, size_t pos) {
     return "line " + std::to_string(line) + ", column " + std::to_string(col);
 }
 
+static void ensure_key_type_allowed(const value & val) {
+    if (!val->is_hashable()) {
+        throw std::runtime_error("Type: " + val->type() + " is not allowed as object key");
+    }
+}
+
 // execute with error handling
 value statement::execute(context & ctx) {
     try {
@@ -95,20 +101,10 @@ value identifier::execute_impl(context & ctx) {
 value object_literal::execute_impl(context & ctx) {
     auto obj = mk_val<value_object>();
     for (const auto & pair : val) {
-        value key_val = pair.first->execute(ctx);
-        if (!is_val<value_string>(key_val) && !is_val<value_int>(key_val)) {
-            throw std::runtime_error("Object literal: keys must be string or int values, got " + key_val->type());
-        }
-        std::string key = key_val->as_string().str();
+        value key = pair.first->execute(ctx);
         value val = pair.second->execute(ctx);
-        JJ_DEBUG("Object literal: setting key '%s' with value type %s", key.c_str(), val->type().c_str());
+        JJ_DEBUG("Object literal: setting key '%s' with value type %s", key->as_string().str().c_str(), val->type().c_str());
         obj->insert(key, val);
-
-        if (is_val<value_int>(key_val)) {
-            obj->val_obj.is_key_numeric = true;
-        } else if (obj->val_obj.is_key_numeric) {
-            throw std::runtime_error("Object literal: cannot mix numeric and non-numeric keys");
-        }
     }
     return obj;
 }
@@ -127,9 +123,9 @@ value binary_expression::execute_impl(context & ctx) {
     value right_val = right->execute(ctx);
     JJ_DEBUG("Executing binary expression %s '%s' %s", left_val->type().c_str(), op.value.c_str(), right_val->type().c_str());
     if (op.value == "==") {
-        return mk_val<value_bool>(value_compare(left_val, right_val, value_compare_op::eq));
+        return mk_val<value_bool>(*left_val == *right_val);
     } else if (op.value == "!=") {
-        return mk_val<value_bool>(!value_compare(left_val, right_val, value_compare_op::eq));
+        return mk_val<value_bool>(!(*left_val == *right_val));
     }
 
     auto workaround_concat_null_with_str = [&](value & res) -> bool {
@@ -230,7 +226,7 @@ value binary_expression::execute_impl(context & ctx) {
         auto & arr = right_val->as_array();
         bool member = false;
         for (const auto & item : arr) {
-            if (value_compare(left_val, item, value_compare_op::eq)) {
+            if (*left_val == *item) {
                 member = true;
                 break;
             }
@@ -265,11 +261,9 @@ value binary_expression::execute_impl(context & ctx) {
         }
     }
 
-    // String in object
-    if (is_val<value_string>(left_val) && is_val<value_object>(right_val)) {
-        auto key = left_val->as_string().str();
-        auto & obj = right_val->as_object();
-        bool has_key = obj.find(key) != obj.end();
+    // Value key in object
+    if (is_val<value_object>(right_val)) {
+        bool has_key = right_val->has_key(left_val);
         if (op.value == "in") {
             return mk_val<value_bool>(has_key);
         } else if (op.value == "not in") {
@@ -464,16 +458,10 @@ value for_statement::execute_impl(context & ctx) {
     std::vector<value> items;
     if (is_val<value_object>(iterable_val)) {
         JJ_DEBUG("%s", "For loop over object keys");
-        auto & obj = iterable_val->as_object();
+        auto & obj = iterable_val->as_ordered_object();
         for (auto & p : obj) {
-            auto tuple = mk_val<value_array>();
-            if (iterable_val->val_obj.is_key_numeric) {
-                tuple->push_back(mk_val<value_int>(std::stoll(p.first)));
-            } else {
-                tuple->push_back(mk_val<value_string>(p.first));
-            }
-            tuple->push_back(p.second);
-            items.push_back(tuple);
+            auto tuple = mk_val<value_tuple>(p);
+            items.push_back(std::move(tuple));
         }
         if (ctx.is_get_stats) {
             iterable_val->stats.used = true;
@@ -560,6 +548,7 @@ value for_statement::execute_impl(context & ctx) {
     for (size_t i = 0; i < filtered_items.size(); i++) {
         JJ_DEBUG("For loop iteration %zu/%zu", i + 1, filtered_items.size());
         value_object loop_obj = mk_val<value_object>();
+        loop_obj->has_builtins = false; // loop object has no builtins
         loop_obj->insert("index", mk_val<value_int>(i + 1));
         loop_obj->insert("index0", mk_val<value_int>(i));
         loop_obj->insert("revindex", mk_val<value_int>(filtered_items.size() - i));
@@ -602,11 +591,13 @@ value set_statement::execute_impl(context & ctx) {
     auto rhs = val ? val->execute(ctx) : exec_statements(body, ctx);
 
     if (is_stmt<identifier>(assignee)) {
+        // case: {% set my_var = value %}
         auto var_name = cast_stmt<identifier>(assignee)->val;
         JJ_DEBUG("Setting global variable '%s' with value type %s", var_name.c_str(), rhs->type().c_str());
         ctx.set_val(var_name, rhs);
 
     } else if (is_stmt<tuple_literal>(assignee)) {
+        // case: {% set a, b = value %}
         auto tuple = cast_stmt<tuple_literal>(assignee);
         if (!is_val<value_array>(rhs)) {
             throw std::runtime_error("Cannot unpack non-iterable type in set: " + rhs->type());
@@ -625,6 +616,7 @@ value set_statement::execute_impl(context & ctx) {
         }
 
     } else if (is_stmt<member_expression>(assignee)) {
+        // case: {% set ns.my_var = value %}
         auto member = cast_stmt<member_expression>(assignee);
         if (member->computed) {
             throw std::runtime_error("Cannot assign to computed member");
@@ -717,6 +709,7 @@ value member_expression::execute_impl(context & ctx) {
 
     value property;
     if (this->computed) {
+        // syntax: obj[expr]
         JJ_DEBUG("Member expression, computing property type %s", this->property->type().c_str());
 
         int64_t arr_size = 0;
@@ -745,32 +738,43 @@ value member_expression::execute_impl(context & ctx) {
             property = this->property->execute(ctx);
         }
     } else {
+        // syntax: obj.prop
         if (!is_stmt<identifier>(this->property)) {
-            throw std::runtime_error("Non-computed member property must be an identifier");
+            throw std::runtime_error("Static member property must be an identifier");
         }
         property = mk_val<value_string>(cast_stmt<identifier>(this->property)->val);
+        std::string prop = property->as_string().str();
+        JJ_DEBUG("Member expression, object type %s, static property '%s'", object->type().c_str(), prop.c_str());
+
+        // behavior of jinja2: obj having prop as a built-in function AND 'prop', as an object key,
+        // then obj.prop returns the built-in function, not the property value.
+        // while obj['prop'] returns the property value.
+        // example: {"obj": {"items": 123}} -> obj.items is the built-in function, obj['items'] is 123
+
+        value val = try_builtin_func(ctx, prop, object, true);
+        if (!is_val<value_undefined>(val)) {
+            return val;
+        }
+        // else, fallthrough to normal property access below
     }
 
     JJ_DEBUG("Member expression on object type %s, property type %s", object->type().c_str(), property->type().c_str());
+    ensure_key_type_allowed(property);
 
     value val = mk_val<value_undefined>("object_property");
 
     if (is_val<value_undefined>(object)) {
         JJ_DEBUG("%s", "Accessing property on undefined object, returning undefined");
         return val;
+
     } else if (is_val<value_object>(object)) {
-        if (!is_val<value_string>(property)) {
-            throw std::runtime_error("Cannot access object with non-string: got " + property->type());
-        }
         auto key = property->as_string().str();
-        auto & obj = object->as_object();
-        auto it = obj.find(key);
-        if (it != obj.end()) {
-            val = it->second;
-        } else {
+        val = object->at(property, val);
+        if (is_val<value_undefined>(val)) {
             val = try_builtin_func(ctx, key, object, true);
         }
         JJ_DEBUG("Accessed property '%s' value, got type: %s", key.c_str(), val->type().c_str());
+
     } else if (is_val<value_array>(object) || is_val<value_string>(object)) {
         if (is_val<value_int>(property)) {
             int64_t index = property->as_int();
@@ -793,7 +797,8 @@ value member_expression::execute_impl(context & ctx) {
         } else if (is_val<value_string>(property)) {
             auto key = property->as_string().str();
             JJ_DEBUG("Accessing %s built-in '%s'", is_val<value_array>(object) ? "array" : "string", key.c_str());
-            val = try_builtin_func(ctx, key, object);
+            val = try_builtin_func(ctx, key, object, true);
+
         } else {
             throw std::runtime_error("Cannot access property with non-string/non-number: got " + property->type());
         }
@@ -802,7 +807,7 @@ value member_expression::execute_impl(context & ctx) {
             throw std::runtime_error("Cannot access property with non-string: got " + property->type());
         }
         auto key = property->as_string().str();
-        val = try_builtin_func(ctx, key, object);
+        val = try_builtin_func(ctx, key, object, true);
     }
 
     if (ctx.is_get_stats && val && object && property) {
diff --git a/common/jinja/runtime.h b/common/jinja/runtime.h
index 1e7c63b85c..17a6dff5aa 100644
--- a/common/jinja/runtime.h
+++ b/common/jinja/runtime.h
@@ -56,6 +56,7 @@ struct context {
     // src is optional, used for error reporting
     context(std::string src = "") : src(std::make_shared<std::string>(std::move(src))) {
         env = mk_val<value_object>();
+        env->has_builtins = false; // context object has no builtins
         env->insert("true",  mk_val<value_bool>(true));
         env->insert("True",  mk_val<value_bool>(true));
         env->insert("false", mk_val<value_bool>(false));
@@ -68,7 +69,7 @@ struct context {
 
     context(const context & parent) : context() {
         // inherit variables (for example, when entering a new scope)
-        auto & pvar = parent.env->as_object();
+        auto & pvar = parent.env->as_ordered_object();
         for (const auto & pair : pvar) {
             set_val(pair.first, pair.second);
         }
@@ -78,18 +79,18 @@ struct context {
     }
 
     value get_val(const std::string & name) {
-        auto it = env->val_obj.unordered.find(name);
-        if (it != env->val_obj.unordered.end()) {
-            return it->second;
-        } else {
-            return mk_val<value_undefined>(name);
-        }
+        value default_val = mk_val<value_undefined>(name);
+        return env->at(name, default_val);
     }
 
     void set_val(const std::string & name, const value & val) {
         env->insert(name, val);
     }
 
+    void set_val(const value & name, const value & val) {
+        env->insert(name, val);
+    }
+
     void print_vars() const {
         printf("Context Variables:\n%s\n", value_to_json(env, 2).c_str());
     }
@@ -265,7 +266,7 @@ struct comment_statement : public statement {
 struct member_expression : public expression {
     statement_ptr object;
     statement_ptr property;
-    bool computed;
+    bool computed; // true if obj[expr] and false if obj.prop
 
     member_expression(statement_ptr && object, statement_ptr && property, bool computed)
         : object(std::move(object)), property(std::move(property)), computed(computed) {
@@ -343,9 +344,19 @@ struct array_literal : public expression {
     }
 };
 
-struct tuple_literal : public array_literal {
-    explicit tuple_literal(statements && val) : array_literal(std::move(val)) {}
+struct tuple_literal : public expression {
+    statements val;
+    explicit tuple_literal(statements && val) : val(std::move(val)) {
+        for (const auto& item : this->val) chk_type<expression>(item);
+    }
     std::string type() const override { return "TupleLiteral"; }
+    value execute_impl(context & ctx) override {
+        auto arr = mk_val<value_array>();
+        for (const auto & item_stmt : val) {
+            arr->push_back(item_stmt->execute(ctx));
+        }
+        return mk_val<value_tuple>(std::move(arr->as_array()));
+    }
 };
 
 struct object_literal : public expression {
diff --git a/common/jinja/string.cpp b/common/jinja/string.cpp
index 21ebde39e3..8087e15b35 100644
--- a/common/jinja/string.cpp
+++ b/common/jinja/string.cpp
@@ -61,6 +61,12 @@ size_t string::length() const {
     return len;
 }
 
+void string::hash_update(hasher & hash) const noexcept {
+    for (const auto & part : parts) {
+        hash.update(part.val.data(), part.val.length());
+    }
+}
+
 bool string::all_parts_are_input() const {
     for (const auto & part : parts) {
         if (!part.is_input) {
diff --git a/common/jinja/string.h b/common/jinja/string.h
index 78457f9e41..c4963000ad 100644
--- a/common/jinja/string.h
+++ b/common/jinja/string.h
@@ -4,6 +4,8 @@
 #include <string>
 #include <vector>
 
+#include "utils.h"
+
 namespace jinja {
 
 // allow differentiate between user input strings and template strings
@@ -37,6 +39,7 @@ struct string {
 
     std::string str() const;
     size_t length() const;
+    void hash_update(hasher & hash) const noexcept;
     bool all_parts_are_input() const;
     bool is_uppercase() const;
     bool is_lowercase() const;
diff --git a/common/jinja/utils.h b/common/jinja/utils.h
index 1e9f2a12a1..de6947fc28 100644
--- a/common/jinja/utils.h
+++ b/common/jinja/utils.h
@@ -3,6 +3,8 @@
 #include <string>
 #include <sstream>
 #include <algorithm>
+#include <cstdint>
+#include <cstring>
 
 namespace jinja {
 
@@ -46,4 +48,102 @@ static std::string fmt_error_with_source(const std::string & tag, const std::str
     return oss.str();
 }
 
+// Note: this is a simple hasher, not cryptographically secure, just for hash table usage
+struct hasher {
+    static constexpr auto size_t_digits = sizeof(size_t) * 8;
+    static constexpr size_t prime = size_t_digits == 64 ? 0x100000001b3 : 0x01000193;
+    static constexpr size_t seed = size_t_digits == 64 ? 0xcbf29ce484222325 : 0x811c9dc5;
+    static constexpr auto block_size = sizeof(size_t); // in bytes; allowing the compiler to vectorize the computation
+
+    static_assert(size_t_digits == 64 || size_t_digits == 32);
+    static_assert(block_size == 8 || block_size == 4);
+
+    uint8_t buffer[block_size];
+    size_t idx = 0; // current index in buffer
+    size_t state = seed;
+
+    hasher() = default;
+    hasher(const std::type_info & type_inf) noexcept {
+        const auto type_hash = type_inf.hash_code();
+        update(&type_hash, sizeof(type_hash));
+    }
+
+    // Properties:
+    //   - update is not associative: update(a).update(b) != update(b).update(a)
+    //   - update(a ~ b) == update(a).update(b) with ~ as concatenation operator --> useful for streaming
+    //   - update("", 0) --> state unchanged with empty input
+    hasher& update(void const * bytes, size_t len) noexcept {
+        const uint8_t * c = static_cast<uint8_t const *>(bytes);
+        if (len == 0) {
+            return *this;
+        }
+        size_t processed = 0;
+
+        // first, fill the existing buffer if it's partial
+        if (idx > 0) {
+            size_t to_fill = block_size - idx;
+            if (to_fill > len) {
+                to_fill = len;
+            }
+            std::memcpy(buffer + idx, c, to_fill);
+            idx += to_fill;
+            processed += to_fill;
+            if (idx == block_size) {
+                update_block(buffer);
+                idx = 0;
+            }
+        }
+
+        // process full blocks from the remaining input
+        for (; processed + block_size <= len; processed += block_size) {
+            update_block(c + processed);
+        }
+
+        // buffer any remaining bytes
+        size_t remaining = len - processed;
+        if (remaining > 0) {
+            std::memcpy(buffer, c + processed, remaining);
+            idx = remaining;
+        }
+        return *this;
+    }
+
+    // convenience function for testing only
+    hasher& update(const std::string & s) noexcept {
+        return update(s.data(), s.size());
+    }
+
+    // finalize and get the hash value
+    // note: after calling digest, the hasher state is modified, do not call update() again
+    size_t digest() noexcept {
+        // if there are remaining bytes in buffer, fill the rest with zeros and process
+        if (idx > 0) {
+            for (size_t i = idx; i < block_size; ++i) {
+                buffer[i] = 0;
+            }
+            update_block(buffer);
+            idx = 0;
+        }
+
+        return state;
+    }
+
+private:
+    // IMPORTANT: block must have at least block_size bytes
+    void update_block(const uint8_t * block) noexcept {
+        size_t blk = static_cast<uint32_t>(block[0])
+                    | (static_cast<uint32_t>(block[1]) << 8)
+                    | (static_cast<uint32_t>(block[2]) << 16)
+                    | (static_cast<uint32_t>(block[3]) << 24);
+        if constexpr (block_size == 8) {
+            blk = blk | (static_cast<uint64_t>(block[4]) << 32)
+                      | (static_cast<uint64_t>(block[5]) << 40)
+                      | (static_cast<uint64_t>(block[6]) << 48)
+                      | (static_cast<uint64_t>(block[7]) << 56);
+        }
+        state ^= blk;
+        state *= prime;
+    }
+};
+
 } // namespace jinja
diff --git a/common/jinja/value.cpp b/common/jinja/value.cpp
index 0ae9d1c565..f254ae9251 100644
--- a/common/jinja/value.cpp
+++ b/common/jinja/value.cpp
@@ -114,6 +114,18 @@ static T slice(const T & array, int64_t start, int64_t stop, int64_t step = 1) {
     return result;
 }
 
+template<typename T>
+static value empty_value_fn(const func_args &) {
+    if constexpr (std::is_same_v<T, value_int>) {
+        return mk_val<T>(0);
+    } else if constexpr (std::is_same_v<T, value_float>) {
+        return mk_val<T>(0.0);
+    } else if constexpr (std::is_same_v<T, value_bool>) {
+        return mk_val<T>(false);
+    } else {
+        return mk_val<T>();
+    }
+}
 template<typename T>
 static value test_type_fn(const func_args & args) {
     args.ensure_count(1);
@@ -128,6 +140,13 @@ static value test_type_fn(const func_args & args) {
     JJ_DEBUG("test_type_fn: type=%s or %s result=%d", typeid(T).name(), typeid(U).name(), is_type ? 1 : 0);
     return mk_val<value_bool>(is_type);
 }
+template<typename T, typename U, typename V>
+static value test_type_fn(const func_args & args) {
+    args.ensure_count(1);
+    bool is_type = is_val<T>(args.get_pos(0)) || is_val<U>(args.get_pos(0)) || is_val<V>(args.get_pos(0));
+    JJ_DEBUG("test_type_fn: type=%s, %s or %s result=%d", typeid(T).name(), typeid(U).name(), typeid(V).name(), is_type ? 1 : 0);
+    return mk_val<value_bool>(is_type);
+}
 template<value_compare_op op>
 static value test_compare_fn(const func_args & args) {
     args.ensure_count(2, 2);
@@ -163,7 +182,7 @@ static value selectattr(const func_args & args) {
     args.ensure_vals<value_array, value_string, value_string, value_string>(true, true, false, false);
 
     auto arr = args.get_pos(0)->as_array();
-    auto attr_name = args.get_pos(1)->as_string().str();
+    auto attribute = args.get_pos(1);
     auto out = mk_val<value_array>();
     value val_default = mk_val<value_undefined>();
 
@@ -173,7 +192,7 @@ static value selectattr(const func_args & args) {
             if (!is_val<value_object>(item)) {
                 throw raised_exception("selectattr: item is not an object");
             }
-            value attr_val = item->at(attr_name, val_default);
+            value attr_val = item->at(attribute, val_default);
             bool is_selected = attr_val->as_bool();
             if constexpr (is_reject) is_selected = !is_selected;
             if (is_selected) out->push_back(item);
@@ -217,7 +236,7 @@ static value selectattr(const func_args & args) {
             if (!is_val<value_object>(item)) {
                 throw raised_exception("selectattr: item is not an object");
             }
-            value attr_val = item->at(attr_name, val_default);
+            value attr_val = item->at(attribute, val_default);
             func_args test_args(args.ctx);
             test_args.push_back(attr_val); // attribute value
             test_args.push_back(extra_arg); // extra argument
@@ -347,8 +366,8 @@ const func_builtins & global_builtins() {
         {"test_is_integer", test_type_fn<value_int>},
         {"test_is_float", test_type_fn<value_float>},
         {"test_is_number", test_type_fn<value_int, value_float>},
-        {"test_is_iterable", test_type_fn<value_array, value_string>},
-        {"test_is_sequence", test_type_fn<value_array, value_string>},
+        {"test_is_iterable", test_type_fn<value_array, value_string, value_undefined>},
+        {"test_is_sequence", test_type_fn<value_array, value_string, value_undefined>},
         {"test_is_mapping", test_type_fn<value_object>},
         {"test_is_lower", [](const func_args & args) -> value {
             args.ensure_vals<value_string>();
@@ -698,6 +717,7 @@ const func_builtins & value_bool_t::get_builtins() const {
             bool val = args.get_pos(0)->as_bool();
             return mk_val<value_string>(val ? "True" : "False");
         }},
+        {"tojson", tojson},
     };
     return builtins;
 }
@@ -740,6 +760,7 @@ const func_builtins & value_array_t::get_builtins() const {
             args.ensure_count(1, 4);
             args.ensure_vals<value_array, value_int, value_int, value_int>(true, true, false, false);
 
+            auto val  = args.get_pos(0);
             auto arg0 = args.get_pos(1);
             auto arg1 = args.get_pos(2, mk_val<value_undefined>());
             auto arg2 = args.get_pos(3, mk_val<value_undefined>());
@@ -761,10 +782,8 @@ const func_builtins & value_array_t::get_builtins() const {
             if (step == 0) {
                 throw raised_exception("slice step cannot be zero");
             }
-            auto arr = slice(args.get_pos(0)->as_array(), start, stop, step);
-            auto res = mk_val<value_array>();
-            res->val_arr = std::move(arr);
-            return res;
+            auto arr = slice(val->as_array(), start, stop, step);
+            return is_val<value_tuple>(val) ? mk_val<value_tuple>(std::move(arr)) : mk_val<value_array>(std::move(arr));
         }},
         {"selectattr", selectattr<false>},
         {"select", selectattr<false>},
@@ -775,19 +794,29 @@ const func_builtins & value_array_t::get_builtins() const {
             if (!is_val<value_array>(args.get_pos(0))) {
                 throw raised_exception("join() first argument must be an array");
             }
-            value val_delim     = args.get_kwarg_or_pos("d",         1);
-            value val_attribute = args.get_kwarg_or_pos("attribute", 2);
-            if (!val_attribute->is_undefined()) {
-                throw not_implemented_exception("array attribute join not implemented");
-            }
+            value val_delim = args.get_kwarg_or_pos("d",         1);
+            value attribute = args.get_kwarg_or_pos("attribute", 2);
             const auto & arr = args.get_pos(0)->as_array();
-            std::string delim = is_val<value_string>(val_delim) ? val_delim->as_string().str() : "";
+            const bool attr_is_int = is_val<value_int>(attribute);
+            if (!attribute->is_undefined() && !is_val<value_string>(attribute) && !attr_is_int) {
+                throw raised_exception("join() attribute must be string or integer");
+            }
+            const int64_t attr_int = attr_is_int ? attribute->as_int() : 0;
+            const std::string delim = val_delim->is_undefined() ? "" : val_delim->as_string().str();
             std::string result;
             for (size_t i = 0; i < arr.size(); ++i) {
-                if (!is_val<value_string>(arr[i]) && !is_val<value_int>(arr[i]) && !is_val<value_float>(arr[i])) {
+                value val_arr = arr[i];
+                if (!attribute->is_undefined()) {
+                    if (attr_is_int && is_val<value_array>(val_arr)) {
+                        val_arr = val_arr->at(attr_int);
+                    } else if (!attr_is_int && is_val<value_object>(val_arr)) {
+                        val_arr = val_arr->at(attribute);
+                    }
+                }
+                if (!is_val<value_string>(val_arr) && !is_val<value_int>(val_arr) && !is_val<value_float>(val_arr)) {
                     throw raised_exception("join() can only join arrays of strings or numerics");
                 }
-                result += arr[i]->as_string().str();
+                result += val_arr->as_string().str();
                 if (i < arr.size() - 1) {
                     result += delim;
                 }
@@ -796,35 +825,37 @@ const func_builtins & value_array_t::get_builtins() const {
         }},
         {"string", [](const func_args & args) -> value {
             args.ensure_vals<value_array>();
-            auto str = mk_val<value_string>();
-            gather_string_parts_recursive(args.get_pos(0), str);
-            return str;
+            return mk_val<value_string>(args.get_pos(0)->as_string());
         }},
         {"tojson", tojson},
         {"map", [](const func_args & args) -> value {
-            args.ensure_count(2, 3);
+            args.ensure_count(2);
             if (!is_val<value_array>(args.get_pos(0))) {
                 throw raised_exception("map: first argument must be an array");
             }
-            value attribute = args.get_kwarg_or_pos("attribute", 1);
-            if (is_val<value_int>(attribute)) {
-                throw not_implemented_exception("map: integer attribute not implemented");
+            if (!is_val<value_kwarg>(args.get_args().at(1))) {
+                throw not_implemented_exception("map: filter-mapping not implemented");
             }
-            if (!is_val<value_string>(attribute)) {
+            value val       = args.get_pos(0);
+            value attribute = args.get_kwarg_or_pos("attribute", 1);
+            const bool attr_is_int = is_val<value_int>(attribute);
+            if (!is_val<value_string>(attribute) && !attr_is_int) {
                 throw raised_exception("map: attribute must be string or integer");
             }
-            std::string attr_name = attribute->as_string().str();
+            const int64_t attr_int = attr_is_int ? attribute->as_int() : 0;
             value default_val = args.get_kwarg("default", mk_val<value_undefined>());
             auto out = mk_val<value_array>();
-            auto arr = args.get_pos(0)->as_array();
+            auto arr = val->as_array();
             for (const auto & item : arr) {
-                if (!is_val<value_object>(item)) {
-                    throw raised_exception("map: item is not an object");
+                value attr_val;
+                if (attr_is_int) {
+                    attr_val = is_val<value_array>(item) ? item->at(attr_int, default_val) : default_val;
+                } else {
+                    attr_val = is_val<value_object>(item) ? item->at(attribute, default_val) : default_val;
                 }
-                value attr_val = item->at(attr_name, default_val);
                 out->push_back(attr_val);
             }
-            return out;
+            return is_val<value_tuple>(val) ? mk_val<value_tuple>(std::move(out->as_array())) : out;
         }},
         {"append", [](const func_args & args) -> value {
             args.ensure_count(2);
@@ -847,37 +878,44 @@ const func_builtins & value_array_t::get_builtins() const {
             return arr_editable->pop_at(index);
         }},
         {"sort", [](const func_args & args) -> value {
-            args.ensure_count(1, 3);
+            args.ensure_count(1, 4);
             if (!is_val<value_array>(args.get_pos(0))) {
                 throw raised_exception("sort: first argument must be an array");
             }
-            bool reverse = args.get_kwarg("reverse", mk_val<value_undefined>())->as_bool();
-            value attribute = args.get_kwarg("attribute", mk_val<value_undefined>());
-            std::string attr = attribute->is_undefined() ? "" : attribute->as_string().str();
-            std::vector<value> arr = cast_val<value_array>(args.get_pos(0))->as_array(); // copy
+            value val         = args.get_pos(0);
+            value val_reverse = args.get_kwarg_or_pos("reverse",        1);
+            value val_case    = args.get_kwarg_or_pos("case_sensitive", 2);
+            value attribute   = args.get_kwarg_or_pos("attribute",      3);
+            // FIXME: sorting is currently always case sensitive
+            //const bool case_sensitive = val_case->as_bool(); // undefined == false
+            const bool reverse = val_reverse->as_bool(); // undefined == false
+            const bool attr_is_int = is_val<value_int>(attribute);
+            const int64_t attr_int = attr_is_int ? attribute->as_int() : 0;
+            std::vector<value> arr = val->as_array(); // copy
             std::sort(arr.begin(), arr.end(),[&](const value & a, const value & b) {
                 value val_a = a;
                 value val_b = b;
                 if (!attribute->is_undefined()) {
-                    if (!is_val<value_object>(a) || !is_val<value_object>(b)) {
-                        throw raised_exception("sort: items are not objects");
+                    if (attr_is_int && is_val<value_array>(a) && is_val<value_array>(b)) {
+                        val_a = a->at(attr_int);
+                        val_b = b->at(attr_int);
+                    } else if (!attr_is_int && is_val<value_object>(a) && is_val<value_object>(b)) {
+                        val_a = a->at(attribute);
+                        val_b = b->at(attribute);
+                    } else {
+                        throw raised_exception("sort: unsupported object attribute comparison between " + a->type() + " and " + b->type());
                     }
-                    val_a = attr.empty() ? a : a->at(attr);
-                    val_b = attr.empty() ? b : b->at(attr);
-                }
-                if (reverse) {
-                    return value_compare(val_a, val_b, value_compare_op::gt);
-                } else {
-                    return !value_compare(val_a, val_b, value_compare_op::gt);
                 }
+                return value_compare(val_a, val_b, reverse ? value_compare_op::gt : value_compare_op::lt);
             });
-            return mk_val<value_array>(arr);
+            return is_val<value_tuple>(val) ? mk_val<value_tuple>(std::move(arr)) : mk_val<value_array>(std::move(arr));
         }},
         {"reverse", [](const func_args & args) -> value {
             args.ensure_vals<value_array>();
-            std::vector<value> arr = cast_val<value_array>(args.get_pos(0))->as_array(); // copy
+            value val = args.get_pos(0);
+            std::vector<value> arr = val->as_array(); // copy
             std::reverse(arr.begin(), arr.end());
-            return mk_val<value_array>(arr);
+            return is_val<value_tuple>(val) ? mk_val<value_tuple>(std::move(arr)) : mk_val<value_array>(std::move(arr));
         }},
         {"unique", [](const func_args &) -> value {
             throw not_implemented_exception("Array unique builtin not implemented");
@@ -888,6 +926,11 @@ const func_builtins & value_array_t::get_builtins() const {
 
 
 const func_builtins & value_object_t::get_builtins() const {
+    if (!has_builtins) {
+        static const func_builtins no_builtins = {};
+        return no_builtins;
+    }
+
     static const func_builtins builtins = {
         // {"default", default_value}, // cause issue with gpt-oss
         {"get", [](const func_args & args) -> value {
@@ -902,27 +945,22 @@ const func_builtins & value_object_t::get_builtins() const {
             if (args.count() == 3) {
                 default_val = args.get_pos(2);
             }
-            const auto & obj = args.get_pos(0)->as_object();
-            std::string key = args.get_pos(1)->as_string().str();
-            auto it = obj.find(key);
-            if (it != obj.end()) {
-                return it->second;
-            } else {
-                return default_val;
-            }
+            const value obj = args.get_pos(0);
+            const value key = args.get_pos(1);
+            return obj->at(key, default_val);
         }},
         {"keys", [](const func_args & args) -> value {
             args.ensure_vals<value_object>();
-            const auto & obj = args.get_pos(0)->as_object();
+            const auto & obj = args.get_pos(0)->as_ordered_object();
             auto result = mk_val<value_array>();
             for (const auto & pair : obj) {
-                result->push_back(mk_val<value_string>(pair.first));
+                result->push_back(pair.first);
             }
             return result;
         }},
         {"values", [](const func_args & args) -> value {
             args.ensure_vals<value_object>();
-            const auto & obj = args.get_pos(0)->as_object();
+            const auto & obj = args.get_pos(0)->as_ordered_object();
             auto result = mk_val<value_array>();
             for (const auto & pair : obj) {
                 result->push_back(pair.second);
@@ -931,21 +969,22 @@ const func_builtins & value_object_t::get_builtins() const {
         }},
         {"items", [](const func_args & args) -> value {
             args.ensure_vals<value_object>();
-            const auto & obj = args.get_pos(0)->as_object();
+            const auto & obj = args.get_pos(0)->as_ordered_object();
             auto result = mk_val<value_array>();
             for (const auto & pair : obj) {
-                auto item = mk_val<value_array>();
-                item->push_back(mk_val<value_string>(pair.first));
-                item->push_back(pair.second);
+                auto item = mk_val<value_tuple>(pair);
                 result->push_back(std::move(item));
             }
             return result;
         }},
         {"tojson", tojson},
-        {"string", tojson},
+        {"string", [](const func_args & args) -> value {
+            args.ensure_vals<value_object>();
+            return mk_val<value_string>(args.get_pos(0)->as_string());
+        }},
         {"length", [](const func_args & args) -> value {
             args.ensure_vals<value_object>();
-            const auto & obj = args.get_pos(0)->as_object();
+            const auto & obj = args.get_pos(0)->as_ordered_object();
             return mk_val<value_int>(static_cast<int64_t>(obj.size()));
         }},
         {"tojson", [](const func_args & args) -> value {
@@ -958,21 +997,18 @@ const func_builtins & value_object_t::get_builtins() const {
             value val_case    = args.get_kwarg_or_pos("case_sensitive", 1);
             value val_by      = args.get_kwarg_or_pos("by",             2);
             value val_reverse = args.get_kwarg_or_pos("reverse",        3);
-            // FIXME: sorting is case sensitive
+            // FIXME: sorting is currently always case sensitive
             //const bool case_sensitive = val_case->as_bool(); // undefined == false
             const bool reverse = val_reverse->as_bool(); // undefined == false
-            if (!val_by->is_undefined()) {
-                throw not_implemented_exception("dictsort by key not implemented");
-            }
-            if (reverse) {
-                throw not_implemented_exception("dictsort reverse not implemented");
-            }
-            value_t::map obj = val_input->val_obj; // copy
-            std::sort(obj.ordered.begin(), obj.ordered.end(), [&](const auto & a, const auto & b) {
-                return a.first < b.first;
+            const bool by_value = is_val<value_string>(val_by) && val_by->as_string().str() == "value" ? true : false;
+            auto result = mk_val<value_object>(val_input); // copy
+            std::sort(result->val_obj.begin(), result->val_obj.end(), [&](const auto & a, const auto & b) {
+                if (by_value) {
+                    return value_compare(a.second, b.second, reverse ? value_compare_op::gt : value_compare_op::lt);
+                } else {
+                    return value_compare(a.first, b.first, reverse ? value_compare_op::gt : value_compare_op::lt);
+                }
             });
-            auto result = mk_val<value_object>();
-            result->val_obj = std::move(obj);
             return result;
         }},
         {"join", [](const func_args &) -> value {
@@ -986,6 +1022,22 @@ const func_builtins & value_none_t::get_builtins() const {
     static const func_builtins builtins = {
         {"default", default_value},
         {"tojson", tojson},
+        {"string", [](const func_args &) -> value {
+            return mk_val<value_string>("None");
+        }},
+        {"safe", [](const func_args &) -> value {
+            return mk_val<value_string>("None");
+        }},
+        {"strip", [](const func_args &) -> value {
+            return mk_val<value_string>("None");
+        }},
+        {"items", empty_value_fn<value_array>},
+        {"map", empty_value_fn<value_array>},
+        {"reject", empty_value_fn<value_array>},
+        {"rejectattr", empty_value_fn<value_array>},
+        {"select", empty_value_fn<value_array>},
+        {"selectattr", empty_value_fn<value_array>},
+        {"unique", empty_value_fn<value_array>},
     };
     return builtins;
 }
@@ -994,10 +1046,33 @@ const func_builtins & value_none_t::get_builtins() const {
 const func_builtins & value_undefined_t::get_builtins() const {
     static const func_builtins builtins = {
         {"default", default_value},
-        {"tojson", [](const func_args & args) -> value {
-            args.ensure_vals<value_undefined>();
-            return mk_val<value_string>("null");
-        }},
+        {"capitalize", empty_value_fn<value_string>},
+        {"first", empty_value_fn<value_undefined>},
+        {"items", empty_value_fn<value_array>},
+        {"join", empty_value_fn<value_string>},
+        {"last", empty_value_fn<value_undefined>},
+        {"length", empty_value_fn<value_int>},
+        {"list", empty_value_fn<value_array>},
+        {"lower", empty_value_fn<value_string>},
+        {"map", empty_value_fn<value_array>},
+        {"max", empty_value_fn<value_undefined>},
+        {"min", empty_value_fn<value_undefined>},
+        {"reject", empty_value_fn<value_array>},
+        {"rejectattr", empty_value_fn<value_array>},
+        {"replace", empty_value_fn<value_string>},
+        {"reverse", empty_value_fn<value_array>},
+        {"safe", empty_value_fn<value_string>},
+        {"select", empty_value_fn<value_array>},
+        {"selectattr", empty_value_fn<value_array>},
+        {"sort", empty_value_fn<value_array>},
+        {"string", empty_value_fn<value_string>},
+        {"strip", empty_value_fn<value_string>},
+        {"sum", empty_value_fn<value_int>},
+        {"title", empty_value_fn<value_string>},
+        {"truncate", empty_value_fn<value_string>},
+        {"unique", empty_value_fn<value_array>},
+        {"upper", empty_value_fn<value_string>},
+        {"wordcount", empty_value_fn<value_int>},
     };
     return builtins;
 }
@@ -1114,6 +1189,8 @@ void global_from_json(context & ctx, const nlohmann::ordered_json & json_obj, bo
     }
 }
 
+// recursively convert value to JSON string
+// TODO: avoid circular references
 static void value_to_json_internal(std::ostringstream & oss, const value & val, int curr_lvl, int indent, const std::string_view item_sep, const std::string_view key_sep) {
     auto indent_str = [indent, curr_lvl]() -> std::string {
         return (indent > 0) ? std::string(curr_lvl * indent, ' ') : "";
@@ -1169,14 +1246,15 @@ static void value_to_json_internal(std::ostringstream & oss, const value & val,
         }
         oss << "]";
     } else if (is_val<value_object>(val)) {
-        const auto & obj = val->val_obj.ordered; // IMPORTANT: need to keep exact order
+        const auto & obj = val->as_ordered_object(); // IMPORTANT: need to keep exact order
         oss << "{";
         if (!obj.empty()) {
             oss << newline();
             size_t i = 0;
             for (const auto & pair : obj) {
                 oss << indent_str() << (indent > 0 ? std::string(indent, ' ') : "");
-                oss << "\"" << pair.first << "\"" << key_sep;
+                value_to_json_internal(oss, mk_val<value_string>(pair.first->as_string().str()), curr_lvl + 1, indent, item_sep, key_sep);
+                oss << key_sep;
                 value_to_json_internal(oss, pair.second, curr_lvl + 1, indent, item_sep, key_sep);
                 if (i < obj.size() - 1) {
                     oss << item_sep;
@@ -1199,4 +1277,19 @@ std::string value_to_json(const value & val, int indent, const std::string_view
     return oss.str();
 }
 
+// TODO: avoid circular references
+std::string value_to_string_repr(const value & val) {
+    if (is_val<value_string>(val)) {
+        const std::string val_str = val->as_string().str();
+
+        if (val_str.find('\'') != std::string::npos) {
+            return value_to_json(val);
+        } else {
+            return "'" + val_str + "'";
+        }
+    } else {
+        return val->as_repr();
+    }
+}
+
 } // namespace jinja
diff --git a/common/jinja/value.h b/common/jinja/value.h
index 05e7d1e41a..1c04760a08 100644
--- a/common/jinja/value.h
+++ b/common/jinja/value.h
@@ -1,8 +1,10 @@
 #pragma once
 
 #include "string.h"
+#include "utils.h"
 
 #include <algorithm>
+#include <cmath>
 #include <cstdint>
 #include <functional>
 #include <map>
@@ -10,6 +12,7 @@
 #include <set>
 #include <sstream>
 #include <string>
+#include <unordered_map>
 #include <vector>
 
 namespace jinja {
@@ -93,7 +96,8 @@ void global_from_json(context & ctx, const T_JSON & json_obj, bool mark_input);
 
 struct func_args; // function argument values
 
-using func_handler = std::function<value(const func_args &)>;
+using func_hptr = value(const func_args &);
+using func_handler = std::function<func_hptr>;
 using func_builtins = std::map<std::string, func_handler>;
 
 enum value_compare_op { eq, ge, gt, lt, ne };
@@ -103,28 +107,9 @@ struct value_t {
     int64_t val_int;
     double val_flt;
     string val_str;
-    bool val_bool;
 
     std::vector<value> val_arr;
-
-    struct map {
-        // once set to true, all keys must be numeric
-        // caveat: we only allow either all numeric keys or all non-numeric keys
-        // for now, this only applied to for_statement in case of iterating over object keys/items
-        bool is_key_numeric = false;
-        std::map<std::string, value> unordered;
-        std::vector<std::pair<std::string, value>> ordered;
-        void insert(const std::string & key, const value & val) {
-            if (unordered.find(key) != unordered.end()) {
-                // if key exists, remove from ordered list
-                ordered.erase(std::remove_if(ordered.begin(), ordered.end(),
-                    [&](const std::pair<std::string, value> & p) { return p.first == key; }),
-                    ordered.end());
-            }
-            unordered[key] = val;
-            ordered.push_back({key, val});
-        }
-    } val_obj;
+    std::vector<std::pair<value, value>> val_obj;
 
     func_handler val_func;
 
@@ -139,6 +124,7 @@ struct value_t {
     value_t(const value_t &) = default;
     virtual ~value_t() = default;
 
+    // Note: only for debugging and error reporting purposes
     virtual std::string type() const { return ""; }
 
     virtual int64_t as_int() const { throw std::runtime_error(type() + " is not an int value"); }
@@ -146,7 +132,7 @@ struct value_t {
     virtual string as_string() const { throw std::runtime_error(type() + " is not a string value"); }
     virtual bool as_bool() const { throw std::runtime_error(type() + " is not a bool value"); }
     virtual const std::vector<value> & as_array() const { throw std::runtime_error(type() + " is not an array value"); }
-    virtual const std::map<std::string, value> & as_object() const { throw std::runtime_error(type() + " is not an object value"); }
+    virtual const std::vector<std::pair<value, value>> & as_ordered_object() const { throw std::runtime_error(type() + " is not an object value"); }
     virtual value invoke(const func_args &) const { throw std::runtime_error(type() + " is not a function value"); }
     virtual bool is_none() const { return false; }
     virtual bool is_undefined() const { return false; }
@@ -154,28 +140,66 @@ struct value_t {
         throw std::runtime_error("No builtins available for type " + type());
     }
 
-    virtual value & at(const std::string & key, value & default_val) {
-        auto it = val_obj.unordered.find(key);
-        if (it == val_obj.unordered.end()) {
-            return default_val;
-        }
-        return val_obj.unordered.at(key);
-    }
-    virtual value & at(const std::string & key) {
-        auto it = val_obj.unordered.find(key);
-        if (it == val_obj.unordered.end()) {
-            throw std::runtime_error("Key '" + key + "' not found in value of type " + type());
-        }
-        return val_obj.unordered.at(key);
-    }
-    virtual value & at(size_t index) {
-        if (index >= val_arr.size()) {
-            throw std::runtime_error("Index " + std::to_string(index) + " out of bounds for array of size " + std::to_string(val_arr.size()));
-        }
-        return val_arr[index];
-    }
+    virtual bool has_key(const value &) { throw std::runtime_error(type() + " is not an object value"); }
+    virtual void insert(const value & /* key */, const value & /* val */) { throw std::runtime_error(type() + " is not an object value"); }
+    virtual value & at(const value & /* key */, value & /* default_val */) { throw std::runtime_error(type() + " is not an object value"); }
+    virtual value & at(const value & /* key */) { throw std::runtime_error(type() + " is not an object value"); }
+    virtual value & at(const std::string & /* key */, value & /* default_val */) { throw std::runtime_error(type() + " is not an object value"); }
+    virtual value & at(const std::string & /* key */) { throw std::runtime_error(type() + " is not an object value"); }
+    virtual value & at(int64_t /* idx */, value & /* default_val */) { throw std::runtime_error(type() + " is not an array value"); }
+    virtual value & at(int64_t /* idx */) { throw std::runtime_error(type() + " is not an array value"); }
 
+    virtual bool is_numeric() const { return false; }
+    virtual bool is_hashable() const { return false; }
+    virtual bool is_immutable() const { return true; }
+    virtual hasher unique_hash() const noexcept = 0;
+    // TODO: C++20 <=> operator
+    // NOTE: We are treating == as equivalent (for normal comparisons) and != as strict nonequal (for strict (is) comparisons)
+    virtual bool operator==(const value_t & other) const { return equivalent(other); }
+    virtual bool operator!=(const value_t & other) const { return nonequal(other); }
+
+    // Note: only for debugging purposes
     virtual std::string as_repr() const { return as_string().str(); }
+
+protected:
+    virtual bool equivalent(const value_t &) const = 0;
+    virtual bool nonequal(const value_t & other) const { return !equivalent(other); }
+};
+
+//
+// utils
+//
+
+const func_builtins & global_builtins();
+
+std::string value_to_json(const value & val, int indent = -1, const std::string_view item_sep = ", ", const std::string_view key_sep = ": ");
+
+// Note: only used for debugging purposes
+std::string value_to_string_repr(const value & val);
+
+struct not_implemented_exception : public std::runtime_error {
+    not_implemented_exception(const std::string & msg) : std::runtime_error("NotImplemented: " + msg) {}
+};
+
+struct value_hasher {
+    size_t operator()(const value & val) const noexcept {
+        return val->unique_hash().digest();
+    }
+};
+
+struct value_equivalence {
+    bool operator()(const value & lhs, const value & rhs) const {
+        return *lhs == *rhs;
+    }
+    bool operator()(const std::pair<value, value> & lhs, const std::pair<value, value> & rhs) const {
+        return *(lhs.first) == *(rhs.first) && *(lhs.second) == *(rhs.second);
+    }
+};
+
+struct value_equality {
+    bool operator()(const value & lhs, const value & rhs) const {
+        return !(*lhs != *rhs);
+    }
 };
 
 //
@@ -183,28 +207,77 @@ struct value_t {
 //
 
 struct value_int_t : public value_t {
-    value_int_t(int64_t v) { val_int = v; }
+    value_int_t(int64_t v) {
+        val_int = v;
+        val_flt = static_cast<double>(v);
+        if (static_cast<int64_t>(val_flt) != v) {
+            val_flt = v < 0 ? -INFINITY : INFINITY;
+        }
+    }
     virtual std::string type() const override { return "Integer"; }
     virtual int64_t as_int() const override { return val_int; }
-    virtual double as_float() const override { return static_cast<double>(val_int); }
+    virtual double as_float() const override { return val_flt; }
     virtual string as_string() const override { return std::to_string(val_int); }
+    virtual bool as_bool() const override {
+        return val_int != 0;
+    }
     virtual const func_builtins & get_builtins() const override;
+    virtual bool is_numeric() const override { return true; }
+    virtual bool is_hashable() const override { return true; }
+    virtual hasher unique_hash() const noexcept override {
+        return hasher(typeid(*this))
+            .update(&val_int, sizeof(val_int))
+            .update(&val_flt, sizeof(val_flt));
+    }
+protected:
+    virtual bool equivalent(const value_t & other) const override {
+        return other.is_numeric() && val_int == other.val_int && val_flt == other.val_flt;
+    }
+    virtual bool nonequal(const value_t & other) const override {
+        return !(typeid(*this) == typeid(other) && val_int == other.val_int);
+    }
 };
 using value_int = std::shared_ptr<value_int_t>;
 
 
 struct value_float_t : public value_t {
-    value_float_t(double v) { val_flt = v; }
+    value val;
+    value_float_t(double v) {
+        val_flt = v;
+        val_int = std::isfinite(v) ? static_cast<int64_t>(v) : 0;
+        val = mk_val<value_int>(val_int);
+    }
     virtual std::string type() const override { return "Float"; }
     virtual double as_float() const override { return val_flt; }
-    virtual int64_t as_int() const override { return static_cast<int64_t>(val_flt); }
+    virtual int64_t as_int() const override { return val_int; }
     virtual string as_string() const override {
         std::string out = std::to_string(val_flt);
         out.erase(out.find_last_not_of('0') + 1, std::string::npos); // remove trailing zeros
         if (out.back() == '.') out.push_back('0'); // leave one zero if no decimals
         return out;
     }
+    virtual bool as_bool() const override {
+        return val_flt != 0.0;
+    }
     virtual const func_builtins & get_builtins() const override;
+    virtual bool is_numeric() const override { return true; }
+    virtual bool is_hashable() const override { return true; }
+    virtual hasher unique_hash() const noexcept override {
+        if (static_cast<double>(val_int) == val_flt) {
+            return val->unique_hash();
+        } else {
+            return hasher(typeid(*this))
+                .update(&val_int, sizeof(val_int))
+                .update(&val_flt, sizeof(val_flt));
+        }
+    }
+protected:
+    virtual bool equivalent(const value_t & other) const override {
+        return other.is_numeric() && val_int == other.val_int && val_flt == other.val_flt;
+    }
+    virtual bool nonequal(const value_t & other) const override {
+        return !(typeid(*this) == typeid(other) && val_flt == other.val_flt);
+    }
 };
 using value_float = std::shared_ptr<value_float_t>;
 
@@ -226,19 +299,49 @@ struct value_string_t : public value_t {
         return val_str.length() > 0;
     }
     virtual const func_builtins & get_builtins() const override;
+    virtual bool is_hashable() const override { return true; }
+    virtual hasher unique_hash() const noexcept override {
+        const auto type_hash = typeid(*this).hash_code();
+        auto hash = hasher();
+        hash.update(&type_hash, sizeof(type_hash));
+        val_str.hash_update(hash);
+        return hash;
+    }
     void mark_input() {
         val_str.mark_input();
     }
+protected:
+    virtual bool equivalent(const value_t & other) const override {
+        return typeid(*this) == typeid(other) && val_str.str() == other.val_str.str();
+    }
 };
 using value_string = std::shared_ptr<value_string_t>;
 
 
 struct value_bool_t : public value_t {
-    value_bool_t(bool v) { val_bool = v; }
+    value val;
+    value_bool_t(bool v) {
+        val_int = static_cast<int64_t>(v);
+        val_flt = static_cast<double>(v);
+        val = mk_val<value_int>(val_int);
+    }
     virtual std::string type() const override { return "Boolean"; }
-    virtual bool as_bool() const override { return val_bool; }
-    virtual string as_string() const override { return std::string(val_bool ? "True" : "False"); }
+    virtual int64_t as_int() const override { return val_int; }
+    virtual bool as_bool() const override { return val_int; }
+    virtual string as_string() const override { return std::string(val_int ? "True" : "False"); }
     virtual const func_builtins & get_builtins() const override;
+    virtual bool is_numeric() const override { return true; }
+    virtual bool is_hashable() const override { return true; }
+    virtual hasher unique_hash() const noexcept override {
+        return val->unique_hash();
+    }
+protected:
+    virtual bool equivalent(const value_t & other) const override {
+        return other.is_numeric() && val_int == other.val_int && val_flt == other.val_flt;
+    }
+    virtual bool nonequal(const value_t & other) const override {
+        return !(typeid(*this) == typeid(other) && val_int == other.val_int);
+    }
 };
 using value_bool = std::shared_ptr<value_bool_t>;
 
@@ -248,13 +351,34 @@ struct value_array_t : public value_t {
     value_array_t(value & v) {
         val_arr = v->val_arr;
     }
+    value_array_t(std::vector<value> && arr) {
+        val_arr = arr;
+    }
     value_array_t(const std::vector<value> & arr) {
         val_arr = arr;
     }
-    void reverse() { std::reverse(val_arr.begin(), val_arr.end()); }
-    void push_back(const value & val) { val_arr.push_back(val); }
-    void push_back(value && val) { val_arr.push_back(std::move(val)); }
+    void reverse() {
+        if (is_immutable()) {
+            throw std::runtime_error("Attempting to modify immutable type");
+        }
+        std::reverse(val_arr.begin(), val_arr.end());
+    }
+    void push_back(const value & val) {
+        if (is_immutable()) {
+            throw std::runtime_error("Attempting to modify immutable type");
+        }
+        val_arr.push_back(val);
+    }
+    void push_back(value && val) {
+        if (is_immutable()) {
+            throw std::runtime_error("Attempting to modify immutable type");
+        }
+        val_arr.push_back(std::move(val));
+    }
     value pop_at(int64_t index) {
+        if (is_immutable()) {
+            throw std::runtime_error("Attempting to modify immutable type");
+        }
         if (index < 0) {
             index = static_cast<int64_t>(val_arr.size()) + index;
         }
@@ -266,61 +390,228 @@ struct value_array_t : public value_t {
         return val;
     }
     virtual std::string type() const override { return "Array"; }
+    virtual bool is_immutable() const override { return false; }
     virtual const std::vector<value> & as_array() const override { return val_arr; }
     virtual string as_string() const override {
+        const bool immutable = is_immutable();
         std::ostringstream ss;
-        ss << "[";
+        ss << (immutable ? "(" : "[");
         for (size_t i = 0; i < val_arr.size(); i++) {
             if (i > 0) ss << ", ";
-            ss << val_arr.at(i)->as_repr();
+            value val = val_arr.at(i);
+            ss << value_to_string_repr(val);
         }
-        ss << "]";
+        if (immutable && val_arr.size() == 1) {
+            ss << ",";
+        }
+        ss << (immutable ? ")" : "]");
         return ss.str();
     }
     virtual bool as_bool() const override {
         return !val_arr.empty();
     }
+    virtual value & at(int64_t index, value & default_val) override {
+        if (index < 0) {
+            index += val_arr.size();
+        }
+        if (index < 0 || static_cast<size_t>(index) >= val_arr.size()) {
+            return default_val;
+        }
+        return val_arr[index];
+    }
+    virtual value & at(int64_t index) override {
+        if (index < 0) {
+            index += val_arr.size();
+        }
+        if (index < 0 || static_cast<size_t>(index) >= val_arr.size()) {
+            throw std::runtime_error("Index " + std::to_string(index) + " out of bounds for array of size " + std::to_string(val_arr.size()));
+        }
+        return val_arr[index];
+    }
     virtual const func_builtins & get_builtins() const override;
+    virtual bool is_hashable() const override {
+        if (std::all_of(val_arr.begin(), val_arr.end(), [&](auto & val) -> bool {
+            return val->is_immutable() && val->is_hashable();
+        })) {
+            return true;
+        }
+        return false;
+    }
+    virtual hasher unique_hash() const noexcept override {
+        auto hash = hasher(typeid(*this));
+        for (const auto & val : val_arr) {
+            // must use digest to prevent problems from "concatenation" property of hasher
+            // for ex. hash of [ "ab", "c" ] should be different from [ "a", "bc" ]
+            const size_t val_hash = val->unique_hash().digest();
+            hash.update(&val_hash, sizeof(size_t));
+        }
+        return hash;
+    }
+protected:
+    virtual bool equivalent(const value_t & other) const override {
+        return typeid(*this) == typeid(other) && is_hashable() && other.is_hashable() && std::equal(val_arr.begin(), val_arr.end(), other.val_arr.begin(), value_equivalence());
+    }
 };
 using value_array = std::shared_ptr<value_array_t>;
 
 
+struct value_tuple_t : public value_array_t {
+    value_tuple_t(value & v) {
+        val_arr = v->val_arr;
+    }
+    value_tuple_t(std::vector<value> && arr) {
+        val_arr = arr;
+    }
+    value_tuple_t(const std::vector<value> & arr) {
+        val_arr = arr;
+    }
+    value_tuple_t(const std::pair<value, value> & pair) {
+        val_arr.push_back(pair.first);
+        val_arr.push_back(pair.second);
+    }
+    virtual std::string type() const override { return "Tuple"; }
+    virtual bool is_immutable() const override { return true; }
+};
+using value_tuple = std::shared_ptr<value_tuple_t>;
+
+
 struct value_object_t : public value_t {
+    std::unordered_map<value, value, value_hasher, value_equivalence> unordered;
+    bool has_builtins = true; // context and loop objects do not have builtins
     value_object_t() = default;
     value_object_t(value & v) {
         val_obj = v->val_obj;
+        for (const auto & pair : val_obj) {
+            unordered[pair.first] = pair.second;
+        }
     }
-    value_object_t(const std::map<std::string, value> & obj) {
+    value_object_t(const std::map<value, value> & obj) {
         for (const auto & pair : obj) {
-            val_obj.insert(pair.first, pair.second);
+            insert(pair.first, pair.second);
+        }
+    }
+    value_object_t(const std::vector<std::pair<value, value>> & obj) {
+        for (const auto & pair : obj) {
+            insert(pair.first, pair.second);
         }
     }
     void insert(const std::string & key, const value & val) {
-        val_obj.insert(key, val);
+        insert(mk_val<value_string>(key), val);
     }
     virtual std::string type() const override { return "Object"; }
-    virtual const std::map<std::string, value> & as_object() const override { return val_obj.unordered; }
+    virtual bool is_immutable() const override { return false; }
+    virtual const std::vector<std::pair<value, value>> & as_ordered_object() const override { return val_obj; }
+    virtual string as_string() const override {
+        std::ostringstream ss;
+        ss << "{";
+        for (size_t i = 0; i < val_obj.size(); i++) {
+            if (i > 0) ss << ", ";
+            auto & [key, val] = val_obj.at(i);
+            ss << value_to_string_repr(key) << ": " << value_to_string_repr(val);
+        }
+        ss << "}";
+        return ss.str();
+    }
     virtual bool as_bool() const override {
-        return !val_obj.unordered.empty();
+        return !unordered.empty();
+    }
+    virtual bool has_key(const value & key) override {
+        if (!key->is_immutable() || !key->is_hashable()) {
+            throw std::runtime_error("Object key of unhashable type: " + key->type());
+        }
+        return unordered.find(key) != unordered.end();
+    }
+    virtual void insert(const value & key, const value & val) override {
+        bool replaced = false;
+        if (is_immutable()) {
+            throw std::runtime_error("Attempting to modify immutable type");
+        }
+        if (has_key(key)) {
+            // if key exists, replace value in ordered list instead of appending
+            for (auto & pair : val_obj) {
+                if (*(pair.first) == *key) {
+                    pair.second = val;
+                    replaced = true;
+                    break;
+                }
+            }
+        }
+        unordered[key] = val;
+        if (!replaced) {
+            val_obj.push_back({key, val});
+        }
+    }
+    virtual value & at(const value & key, value & default_val) override {
+        if (!has_key(key)) {
+            return default_val;
+        }
+        return unordered.at(key);
+    }
+    virtual value & at(const value & key) override {
+        if (!has_key(key)) {
+            throw std::runtime_error("Key '" + key->as_string().str() + "' not found in value of type " + type());
+        }
+        return unordered.at(key);
+    }
+    virtual value & at(const std::string & key, value & default_val) override {
+        value key_val = mk_val<value_string>(key);
+        return at(key_val, default_val);
+    }
+    virtual value & at(const std::string & key) override {
+        value key_val = mk_val<value_string>(key);
+        return at(key_val);
     }
     virtual const func_builtins & get_builtins() const override;
+    virtual bool is_hashable() const override {
+        if (std::all_of(val_obj.begin(), val_obj.end(), [&](auto & pair) -> bool {
+            const auto & val = pair.second;
+            return val->is_immutable() && val->is_hashable();
+        })) {
+            return true;
+        }
+        return false;
+    }
+    virtual hasher unique_hash() const noexcept override {
+        auto hash = hasher(typeid(*this));
+        for (const auto & [key, val] : val_obj) {
+            // must use digest to prevent problems from "concatenation" property of hasher
+            // for ex. hash of key="ab", value="c" should be different from key="a", value="bc"
+            const size_t key_hash = key->unique_hash().digest();
+            const size_t val_hash = val->unique_hash().digest();
+            hash.update(&key_hash, sizeof(key_hash));
+            hash.update(&val_hash, sizeof(val_hash));
+        }
+        return hash;
+    }
+protected:
+    virtual bool equivalent(const value_t & other) const override {
+        return typeid(*this) == typeid(other) && is_hashable() && other.is_hashable() && std::equal(val_obj.begin(), val_obj.end(), other.val_obj.begin(), value_equivalence());
+    }
 };
 using value_object = std::shared_ptr<value_object_t>;
 
 //
-// null and undefined types
+// none and undefined types
 //
 
 struct value_none_t : public value_t {
     virtual std::string type() const override { return "None"; }
     virtual bool is_none() const override { return true; }
     virtual bool as_bool() const override { return false; }
+    virtual string as_string() const override { return string(type()); }
     virtual std::string as_repr() const override { return type(); }
     virtual const func_builtins & get_builtins() const override;
+    virtual bool is_hashable() const override { return true; }
+    virtual hasher unique_hash() const noexcept override {
+        return hasher(typeid(*this));
+    }
+protected:
+    virtual bool equivalent(const value_t & other) const override {
+        return typeid(*this) == typeid(other);
+    }
 };
 using value_none = std::shared_ptr<value_none_t>;
 
-
 struct value_undefined_t : public value_t {
     std::string hint; // for debugging, to indicate where undefined came from
     value_undefined_t(const std::string & h = "") : hint(h) {}
@@ -329,6 +620,13 @@ struct value_undefined_t : public value_t {
     virtual bool as_bool() const override { return false; }
     virtual std::string as_repr() const override { return type(); }
     virtual const func_builtins & get_builtins() const override;
+    virtual hasher unique_hash() const noexcept override {
+        return hasher(typeid(*this));
+    }
+protected:
+    virtual bool equivalent(const value_t & other) const override {
+        return is_undefined() == other.is_undefined();
+    }
 };
 using value_undefined = std::shared_ptr<value_undefined_t>;
 
@@ -409,7 +707,23 @@ struct value_func_t : public value_t {
         return val_func(new_args);
     }
     virtual std::string type() const override { return "Function"; }
-    virtual std::string as_repr() const override { return type(); }
+    virtual std::string as_repr() const override { return type() + "<" + name + ">(" + (arg0 ? arg0->as_repr() : "") + ")"; }
+    virtual bool is_hashable() const override { return false; }
+    virtual hasher unique_hash() const noexcept override {
+        // Note: this is unused for now, we don't support function as object keys
+        // use function pointer as unique identifier
+        const auto target = val_func.target<func_hptr>();
+        return hasher(typeid(*this)).update(&target, sizeof(target));
+    }
+protected:
+    virtual bool equivalent(const value_t & other) const override {
+        // Note: this is unused for now, we don't support function as object keys
+        // compare function pointers
+        // (val_func == other.val_func does not work as std::function::operator== is only used for nullptr check)
+        const auto target_this  = this->val_func.target<func_hptr>();
+        const auto target_other = other.val_func.target<func_hptr>();
+        return typeid(*this) == typeid(other) && target_this == target_other;
+    }
 };
 using value_func = std::shared_ptr<value_func_t>;
 
@@ -420,18 +734,21 @@ struct value_kwarg_t : public value_t {
     value_kwarg_t(const std::string & k, const value & v) : key(k), val(v) {}
     virtual std::string type() const override { return "KwArg"; }
     virtual std::string as_repr() const override { return type(); }
+    virtual bool is_hashable() const override { return true; }
+    virtual hasher unique_hash() const noexcept override {
+        const auto type_hash = typeid(*this).hash_code();
+        auto hash = val->unique_hash();
+        hash.update(&type_hash, sizeof(type_hash))
+            .update(key.data(), key.size());
+        return hash;
+    }
+protected:
+    virtual bool equivalent(const value_t & other) const override {
+        const value_kwarg_t & other_val = static_cast<const value_kwarg_t &>(other);
+        return typeid(*this) == typeid(other) && key == other_val.key && val == other_val.val;
+    }
 };
 using value_kwarg = std::shared_ptr<value_kwarg_t>;
 
 
-// utils
-
-const func_builtins & global_builtins();
-std::string value_to_json(const value & val, int indent = -1, const std::string_view item_sep = ", ", const std::string_view key_sep = ": ");
-
-struct not_implemented_exception : public std::runtime_error {
-    not_implemented_exception(const std::string & msg) : std::runtime_error("NotImplemented: " + msg) {}
-};
-
-
 } // namespace jinja
diff --git a/common/json-partial.h b/common/json-partial.h
index f63356dc48..be51aabfbf 100644
--- a/common/json-partial.h
+++ b/common/json-partial.h
@@ -1,5 +1,6 @@
 #pragma once
 
+// TODO: use json_fwd.hpp when possible
 #include <nlohmann/json.hpp>
 
 // Healing marker (empty if the JSON was fully parsed / wasn't healed).
diff --git a/common/ngram-cache.cpp b/common/ngram-cache.cpp
index d1a4d84c40..dce54b3647 100644
--- a/common/ngram-cache.cpp
+++ b/common/ngram-cache.cpp
@@ -192,12 +192,12 @@ void common_ngram_cache_draft(
             break;
         }
 
-        LOG(" - draft candidate: token=%d\n", drafted_token);
+        LOG_DBG(" - draft candidate: token=%d\n", drafted_token);
         draft.push_back(drafted_token);
     }
 }
 
-void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & filename) {
+void common_ngram_cache_save(common_ngram_cache & ngram_cache, const std::string & filename) {
     std::ofstream file_out(filename, std::ios::binary);
     for (std::pair<common_ngram, common_ngram_cache_part> item : ngram_cache) {
         const common_ngram      ngram        = item.first;
@@ -217,10 +217,9 @@ void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & fil
             file_out.write(reinterpret_cast<const char *>(&count), sizeof(int32_t));
         }
     }
-
 }
 
-common_ngram_cache common_ngram_cache_load(std::string & filename) {
+common_ngram_cache common_ngram_cache_load(const std::string & filename) {
     std::ifstream hashmap_file(filename, std::ios::binary);
     if (!hashmap_file) {
         throw std::ifstream::failure("Unable to open file " + filename);
diff --git a/common/ngram-cache.h b/common/ngram-cache.h
index dfe012abe4..6e7cfea966 100644
--- a/common/ngram-cache.h
+++ b/common/ngram-cache.h
@@ -88,12 +88,12 @@ void common_ngram_cache_draft(
 // Save an ngram cache to a file.
 // ngram_cache: the ngram cache to save.
 // filename:    the path under which to save the ngram cache.
-void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & filename);
+void common_ngram_cache_save(common_ngram_cache & ngram_cache, const std::string & filename);
 
 // Load an ngram cache saved with common_ngram_cache_save.
 // filename: the path from which to load the ngram cache.
 // returns:  an ngram cache containing the information saved to filename.
-common_ngram_cache common_ngram_cache_load(std::string & filename);
+common_ngram_cache common_ngram_cache_load(const std::string & filename);
 
 // Merge two ngram caches.
 // ngram_cache_target: the ngram cache to which to add the information from ngram_cache_add.
diff --git a/common/ngram-map.cpp b/common/ngram-map.cpp
new file mode 100644
index 0000000000..84fd761367
--- /dev/null
+++ b/common/ngram-map.cpp
@@ -0,0 +1,362 @@
+#include "common.h"
+#include "log.h"
+#include "ngram-map.h"
+
+#include <cinttypes>
+#include <cstdint>
+#include <cstdio>
+#include <sstream>
+
+// Print the values of a sublist of `llama_tokens & inp` to a string in the form [v0, v1, v2, ...].
+static std::string common_tokens_to_str(const llama_tokens & inp, size_t start, size_t length) {
+    std::ostringstream oss;
+    oss << '[';
+    for (size_t i = 0; i < length; ++i) {
+        if (i > 0) {
+            oss << ", ";
+        }
+        oss << inp[start + i];
+    }
+    oss << ']';
+    return oss.str();
+}
+
+
+// n-gram simple
+//
+
+/**
+ * Perform speculative generation using the model's own token history.
+ * Searches for a matching pattern in the token history and returns draft tokens.
+ *
+ * @param state     Current state of this implementation
+ * @param tokens    Token history to search in
+ * @param sampled   Last sampled token
+ * @return Vector of draft tokens, empty if no matching pattern is found
+ */
+llama_tokens common_ngram_simple_draft(
+        common_ngram_simple_state & state,
+        const llama_tokens & tokens, llama_token sampled) {
+
+    // Simple implementation of self-speculative decoding without a draft model.
+    //
+    const size_t cur_len = tokens.size();
+    // Only check every check_rate tokens to save compute
+    // i.e., perform check if (cur_len - idx_last_check) >= check_rate
+    if (state.idx_last_check + state.config.check_rate > cur_len) {
+        llama_tokens draft_tokens;
+        return draft_tokens;
+    }
+
+    size_t n_draft_min = state.config.size_ngram; // size of n-gram to lookup in token history
+    size_t n_draft_max = state.config.size_mgram; // the m-gram following the found n-gram is used for draft
+
+    // vector for tokens we want to verify.
+    // return empty vector if there is no match.
+    llama_tokens draft_tokens;
+
+    // We need at least n_draft_min + n_draft_max + 1 tokens.
+    if (cur_len <= static_cast<size_t>(n_draft_min + n_draft_max + 1)) {
+        return draft_tokens;
+    }
+
+    // pattern search
+    llama_tokens pattern;
+    pattern.reserve(n_draft_min);
+    for (size_t j = cur_len - n_draft_min + 1; j < cur_len; ++j) {
+        pattern.push_back(tokens[j]);
+    }
+    pattern.push_back(sampled); // add the last token to the pattern
+
+    // We do a search in the token history.
+    state.idx_last_check = cur_len;
+
+    size_t match_pos = 0; // we ignore position 0, position 0 == no match
+                          // search backwards, but skip the current match (we are currently there)
+    for (size_t j = cur_len - n_draft_min - 1; j > 0; --j) {
+        bool match = true;
+        for (size_t k = 0; k < pattern.size(); ++k) {
+            if (tokens[j + k] != pattern[k]) {
+                match = false;
+                break;
+            }
+        }
+        if (match) {
+            match_pos = j;
+            break;
+        }
+    }
+    if (match_pos == 0) {
+        return draft_tokens;
+    }
+
+    const size_t copy_max = std::min(
+            n_draft_max,
+            cur_len - (match_pos + n_draft_min)
+            );
+    if (copy_max < n_draft_min) {
+        return draft_tokens;
+    }
+    LOG_DBG("%s: #tokens = %zu: found matching pattern at pos %zu, length %zu, draft length %zu\n",
+            __func__, cur_len,
+            match_pos, pattern.size(), copy_max);
+
+    draft_tokens.reserve(copy_max);
+    for (size_t j = 0; j < copy_max; ++j) {
+        draft_tokens.push_back(tokens[match_pos + n_draft_min + j]);
+    }
+    return draft_tokens;
+}
+
+
+// n-gram map
+//
+
+// maximum number of counted values of a ngram map value.
+#define COMMON_NGRAM_MAX_VALUE_COUNT 16380
+
+void common_ngram_map_draft(common_ngram_map & map,
+        const llama_tokens & inp, llama_token sampled,
+        llama_tokens & draft) {
+    // reset last key and value.
+    map.last_draft_created   = false;
+    map.last_draft_key_idx   = 0;
+    map.last_draft_value_idx = 0;
+
+    const size_t cur_len = inp.size();
+    const uint16_t n = map.size_key;
+    const uint16_t m = map.size_value;
+    if (cur_len < static_cast<size_t>(2 * n + m)) {
+        return;
+    }
+
+    // Only check every check_rate tokens to save compute
+    // i.e., perform check if (cur_len - idx_last_check) >= check_rate
+    if (map.idx_last_check + map.check_rate > cur_len) {
+        return;
+    }
+    map.idx_last_check = cur_len;
+
+    // search pattern, the key n-gram
+    std::vector<llama_token> key_tokens;
+    key_tokens.reserve(n);
+    for (size_t j = cur_len - n + 1; j < cur_len; ++j) {
+        key_tokens.push_back(inp[j]);
+    }
+    key_tokens.push_back(sampled);
+
+    // search for the key in the map
+    size_t match_pos = 0;
+    for (size_t j = cur_len - n - m - 1; j > 0; --j) {
+        bool match = true;
+        for (size_t k = 0; k < n; ++k) {
+            if (inp[j + k] != key_tokens[k]) {
+                match = false;
+                break;
+            }
+        }
+        if (match) {
+           match_pos = j;
+           break;
+        }
+    }
+    if (match_pos > 0) {
+        LOG_INF("%s: cur_len = %zu, n = %d, m = %d, sz_tkns = %zu, sampled = %d, match_pos = %zu\n", __func__,
+            cur_len, n, m, key_tokens.size(), sampled, match_pos);
+    }
+
+    if (match_pos == 0) {
+        return;
+    }
+
+    // We have a match, now we look for the statistics of the key.
+    size_t key_offset = map.keys.size(); // offset in the map
+    // We iterate through the std::vector<common_ngram_map_key> map->keys.
+    for (size_t i = 0; i < map.keys.size(); ++i) {
+        bool match = true;
+        for (size_t j = 0; j < n; ++j) {
+            if (inp[map.keys[i].key_idx + j] != key_tokens[j]) {
+                match = false;
+                break;
+            }
+        }
+        if (match) {
+            key_offset = i;
+            break;
+        }
+    }
+    if (key_offset == map.keys.size()) {
+        // We create a new key-entry, it will get offset key_offset.
+        common_ngram_map_key new_key;
+        new_key.key_idx = match_pos;
+        new_key.stat_idx = 0;
+        new_key.key_num = 0;
+        for (int i = 0; i < COMMON_NGRAM_MAX_VALUES; ++i) {
+            new_key.values[i].value_num = 0;
+            new_key.values[i].n_accepted = m;
+        }
+        map.keys.push_back(new_key);
+    }
+
+    // our key n-gram:
+    common_ngram_map_key & curr_key = map.keys[key_offset];
+
+    // update number of key hits
+    curr_key.key_num = (uint16_t) std::min((int) map.keys[key_offset].key_num + 1,
+            (int) COMMON_NGRAM_MAX_VALUE_COUNT);
+
+    if (map.key_only) {
+        // simple mode:
+        // Fill in the draft with the m tokens following the key.
+        // We work with value values[0] only.
+        int n_draft_tokens = std::min((int) m, (int) curr_key.values[0].n_accepted);
+
+        for (int i = 0; i < n_draft_tokens; ++i) {
+            draft.push_back(inp[match_pos + n + i]);
+        }
+
+        LOG_INF("%s: key_offset = %zu, key_num = %d, draft.size = %zu\n", __func__,
+                key_offset, curr_key.key_num, draft.size());
+
+        map.last_draft_created   = false;
+        map.last_draft_key_idx   = key_offset;
+        map.last_draft_value_idx = 0; // value 0 is used for simple mode
+        return;
+    }
+
+    if (curr_key.key_num < map.min_hits) {
+        // not enough hits to consider this a good draft
+        LOG_DBG("%s: key_offset = %zu, key_num = %d, min_hits = %d, no draft\n", __func__,
+                key_offset, curr_key.key_num, map.min_hits);
+        return;
+    }
+
+    // complex mode: examine the different m-grams after this key n-gram.
+    //
+
+    // determine all (max COMMON_NGRAM_MAX_VALUES) m-grams after the key n-gram.
+    for (size_t i = curr_key.stat_idx; i <= match_pos; ++i) {
+        // begins the key n-gram at index i?
+        bool match_key = true;
+        for (size_t k = 0; k < n; ++k) {
+            if (inp[i + k] != key_tokens[k]) {
+                match_key = false;
+                break;
+            }
+        }
+        if (!match_key) {
+            continue;
+        }
+
+        // Do we haven a existing value m-gram or a new one after the key at index i?
+        size_t idx_begin_value_key = i + n;
+        int idx_value = -1;
+        for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) {
+            size_t idx_begin_value_v = curr_key.values[v].value_idx;
+            if (idx_begin_value_v == 0) {
+                // We found an empty value slot => we found a new value m-gram after the key n-gram.
+                curr_key.values[v].value_idx = idx_begin_value_key;
+                curr_key.values[v].value_num = 0;
+                curr_key.values[v].n_accepted = m;
+                idx_value = v;
+                break;
+            }
+            bool match = true;
+            for (size_t j = 0; j < m; ++j) {
+                if (inp[idx_begin_value_key + j] != inp[idx_begin_value_v + j]) {
+                    match = false;
+                    break;
+                }
+            }
+            if (match) {
+                // We found an existing value m-gram after the key n-gram.
+                idx_value = v;
+                break;
+            }
+        }
+        if (idx_value >= 0) {
+            // We found a value m-gram of the key n-gram.
+            curr_key.values[idx_value].value_num = (uint16_t) std::min((int) curr_key.values[idx_value].value_num + 1,
+                    (int) COMMON_NGRAM_MAX_VALUE_COUNT);
+        }
+    }
+    // the statistics are updated up to match_pos.
+    curr_key.stat_idx = match_pos;
+
+    // Do we have a value we could use for the draft?
+    uint16_t max_occur = 0;
+    int slot_max = 0;
+    for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) {
+        uint16_t curr_occur = curr_key.values[v].value_num;
+        if (curr_occur > max_occur) {
+            max_occur = curr_occur;
+            slot_max = v;
+        }
+    }
+    // What is sum of the other occurences?
+    uint32_t sum_occur = 0;
+    for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) {
+        if (v == slot_max) {
+            continue;
+        }
+        uint16_t curr_occur = curr_key.values[v].value_num;
+        sum_occur += curr_occur;
+    }
+
+    LOG_INF("%s: key_offset = %zu, max_occur = %d, sum_occur = %d, slot_max = %d [%zu/%d, %zu/%d, %zu/%d, %zu/%d]\n", __func__,
+            key_offset,
+            max_occur, sum_occur, slot_max,
+            curr_key.values[0].value_idx, curr_key.values[0].value_num,
+            curr_key.values[1].value_idx, curr_key.values[1].value_num,
+            curr_key.values[2].value_idx, curr_key.values[2].value_num,
+            curr_key.values[3].value_idx, curr_key.values[3].value_num
+        );
+    // Print the tokens of the four values (if idx != 0), use LOG_INF
+    for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) {
+        if (curr_key.values[v].value_idx != 0) {
+            LOG_INF("%s: value[%d] = %s\n", __func__, v, common_tokens_to_str(inp, curr_key.values[v].value_idx, m).c_str());
+        }
+    }
+
+    if (sum_occur > 0 && max_occur < 3 * sum_occur) {
+        // The most frequent value is not much more frequent than the other values.
+        // We do not use the draft.
+        return;
+    }
+
+    // We use the most frequent value values[slot_max] for the draft.
+    // Fill in the draft with the m tokens following the key.
+    int n_draft_tokens = std::min((int) m, (int) curr_key.values[slot_max].n_accepted);
+
+    for (int i = 0; i < n_draft_tokens; ++i) {
+        draft.push_back(inp[match_pos + n + i]);
+    }
+
+    LOG_INF("%s: key_offset = %zu, slot_max = %d, key_num = %d, draft.size = %zu\n", __func__,
+            key_offset, slot_max,
+            curr_key.key_num, draft.size());
+
+    map.last_draft_created   = true;
+    map.last_draft_key_idx   = key_offset;
+    map.last_draft_value_idx = slot_max; // value used for draft generation.
+}
+
+void common_ngram_map_accept(common_ngram_map & map, uint16_t n_accepted) {
+    if (!map.last_draft_created) {
+        return;
+    }
+
+    // find the key and its chosen value.
+    const size_t key_idx = map.last_draft_key_idx;
+    const size_t val_idx = map.last_draft_value_idx;
+
+    // find key corresponding to key_idx.
+    common_ngram_map_key & curr_key = map.keys[key_idx];
+    // find value corresponding to val_idx.
+    struct common_ngram_map_value & curr_value = curr_key.values[val_idx]; // value used for draft generation.
+
+    // update the value statistics
+    LOG_INF("common_ngram_map_send_accepted: n_accepted = %d, prev value_num = %d\n",
+            n_accepted, curr_value.n_accepted);
+    curr_value.n_accepted = n_accepted;
+}
diff --git a/common/ngram-map.h b/common/ngram-map.h
new file mode 100644
index 0000000000..b365034ac5
--- /dev/null
+++ b/common/ngram-map.h
@@ -0,0 +1,106 @@
+#pragma once
+//
+// common/ngram-map.h: structures used to manage a map from n-grams to a list of m-grams
+//
+// These structures are used to do a lookup of n-grams followed by m-grams in token history.
+//
+// There are two algorithms implemented:
+// 1. ngram_simple: lookup of n-grams followed by m-grams in token history.
+// 2. ngram_map: lookup of n-grams followed by m-grams in token history using a map.
+//    The map is a vector of key n-grams, and for each key n-gram there is a list of value m-grams.
+//
+
+#include "llama.h"
+#include "common.h"
+
+#include <vector>
+
+// n-gram simple
+//
+
+// config of n-gram simple.
+struct common_ngram_simple_config {
+    uint16_t   size_ngram;      // size of n-grams to lookup in self-mode
+    uint16_t   size_mgram;      // size of m-grams to draft in self-mode
+    uint16_t   check_rate;      // check for speculative decoding without draft model for each check_rate token
+};
+
+// current state (and config) of n-gram simple.
+struct common_ngram_simple_state {
+    common_ngram_simple_config config;
+
+    size_t idx_last_check = 0; // index of last check in context history (mutable)
+
+    common_ngram_simple_state(const common_ngram_simple_config & config)
+        : config(config) {}
+};
+
+// Searches for a n-gram in the history and checks whether a draft sequence should be generated.
+// state:              the ngram simple state to search in.
+// inp:                the tokens generated so far.
+// sampled:            the token that was just sampled.
+// draft:              vector to store the draft tokens, initially empty.
+llama_tokens common_ngram_simple_draft(
+        common_ngram_simple_state & state,
+        const llama_tokens & tokens, llama_token sampled);
+
+
+// n-gram map
+//
+
+// maximum number of m-gram values stored for each key n-gram.
+#define COMMON_NGRAM_MAX_VALUES 4
+
+// statistics of a m-gram after a known n-gram
+struct common_ngram_map_value {
+    size_t   value_idx = 0;  // index of value m-gram in token-history (0 if unused)
+    uint16_t value_num = 0;  // number of occurences of this value m-gram after the key n-gram (0 in an unused values-slot)
+    int16_t n_accepted = -1;  // number of accepted tokens at last draft (-1 if unused)
+};
+
+// statistics of a n-gram
+struct common_ngram_map_key {
+    size_t   key_idx;   // index of key n-gram in token-history
+    size_t   stat_idx;  // index of last token of stastistics computation (key_num, values)
+
+    uint16_t key_num;   // number of occurences of this key n-gram in token-history
+    common_ngram_map_value values[COMMON_NGRAM_MAX_VALUES]; // some known values after the key
+};
+
+// map from n-grams to following m-grams in token-history
+struct common_ngram_map {
+    uint16_t size_key;   // size of key n-grams
+    uint16_t size_value; // size of value m-grams
+
+    bool key_only;       // true if only key n-grams are used, no values.
+
+    // first draft: vector only, no map.
+    std::vector<common_ngram_map_key> keys; // key n-grams which occur several times in token-history
+    uint16_t check_rate; // check for speculative decoding without draft model for each check_rate token
+    uint16_t min_hits;   // minimum number of key hits to consider a draft
+
+    common_ngram_map(uint16_t sz_key, uint16_t sz_value, bool only_keys,
+                     uint16_t check_rate, uint16_t min_hits)
+        : size_key(sz_key), size_value(sz_value), key_only(only_keys),
+          check_rate(check_rate), min_hits(min_hits) {}
+
+    bool     last_draft_created   = false; // true if a draft was created at last call.
+    size_t   last_draft_key_idx   = 0; // index of last key used for draft generation.
+    uint16_t last_draft_value_idx = 0; // index of last value used for draft generation.
+
+    size_t   idx_last_check       = 0; // index of last check in context history
+};
+
+
+// Searches for the n-gram in the history and checks whether a draft sequence should be generated.
+// map:                the ngram map to search in.
+// inp:                the tokens generated so far.
+// sampled:            the token that was just sampled.
+// draft:              vector to store the draft tokens, initially empty.
+void common_ngram_map_draft(
+    common_ngram_map & map,
+    const llama_tokens & inp, llama_token sampled,
+    llama_tokens & draft);
+
+// Update the statistics of a value after a draft was processed.
+void common_ngram_map_accept(common_ngram_map & map, uint16_t n_accepted);
diff --git a/common/ngram-mod.cpp b/common/ngram-mod.cpp
new file mode 100644
index 0000000000..76f7257f61
--- /dev/null
+++ b/common/ngram-mod.cpp
@@ -0,0 +1,60 @@
+#include "ngram-mod.h"
+
+//
+// common_ngram_mod
+//
+
+common_ngram_mod::common_ngram_mod(uint16_t n, size_t size) : n(n), used(0) {
+    entries.resize(size);
+
+    reset();
+}
+
+size_t common_ngram_mod::idx(const entry_t * tokens) const {
+    size_t res = 0;
+
+    for (size_t i = 0; i < n; ++i) {
+        res = res*6364136223846793005ULL + tokens[i];
+    }
+
+    res = res % entries.size();
+
+    return res;
+}
+
+void common_ngram_mod::add(const entry_t * tokens) {
+    const size_t i = idx(tokens);
+
+    if (entries[i] == EMPTY) {
+        used++;
+    }
+
+    entries[i] = tokens[n];
+}
+
+common_ngram_mod::entry_t common_ngram_mod::get(const entry_t * tokens) const {
+    const size_t i = idx(tokens);
+
+    return entries[i];
+}
+
+void common_ngram_mod::reset() {
+    std::fill(entries.begin(), entries.end(), EMPTY);
+    used = 0;
+}
+
+size_t common_ngram_mod::get_n() const {
+    return n;
+}
+
+size_t common_ngram_mod::get_used() const {
+    return used;
+}
+
+size_t common_ngram_mod::size() const {
+    return entries.size();
+}
+
+size_t common_ngram_mod::size_bytes() const {
+    return entries.size() * sizeof(entries[0]);
+}
diff --git a/common/ngram-mod.h b/common/ngram-mod.h
new file mode 100644
index 0000000000..7af92e9dde
--- /dev/null
+++ b/common/ngram-mod.h
@@ -0,0 +1,38 @@
+#pragma once
+
+#include <cstdint>
+#include <vector>
+#include <cstddef>
+
+//
+// common_ngram_mod
+// ref: https://github.com/ggml-org/llama.cpp/pull/19164
+//
+
+// basic n-gram hasher
+struct common_ngram_mod {
+    using entry_t = int32_t;
+
+    static constexpr entry_t EMPTY = -1;
+
+    common_ngram_mod(uint16_t n, size_t size);
+
+    size_t  idx(const entry_t * tokens) const;
+    void    add(const entry_t * tokens);
+    entry_t get(const entry_t * tokens) const; // return -1 if not found
+
+    void reset();
+
+    size_t get_n()    const;
+    size_t get_used() const;
+
+    size_t size()       const;
+    size_t size_bytes() const;
+
+private:
+    size_t n; // ngram size to hash
+
+    size_t used;
+
+    std::vector<entry_t> entries;
+};
diff --git a/common/speculative.cpp b/common/speculative.cpp
index 3e83b0964c..a1a3b51c13 100644
--- a/common/speculative.cpp
+++ b/common/speculative.cpp
@@ -1,99 +1,57 @@
 #include "speculative.h"
 
+#include "common.h"
 #include "ggml.h"
 #include "llama.h"
 #include "log.h"
-#include "common.h"
+#include "ngram-cache.h"
+#include "ngram-map.h"
+#include "ngram-mod.h"
 #include "sampling.h"
 
-#include <cstring>
 #include <algorithm>
+#include <cstring>
+#include <iomanip>
 #include <map>
 
 #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  128
 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
 
-struct common_speculative {
-    struct llama_context * ctx_tgt; // only used for retokenizing from ctx_dft
-    struct llama_context * ctx_dft;
-    struct common_sampler * smpl;
-
-    llama_batch batch;
-    llama_tokens prompt_dft;
-    bool vocab_dft_compatible = true; // whether retokenization is needed
-    std::map<std::string, std::string> tgt_dft_replacements = {};
+const std::vector<enum common_speculative_type> common_speculative_types = {
+    COMMON_SPECULATIVE_TYPE_NONE,
+    COMMON_SPECULATIVE_TYPE_DRAFT,
+    COMMON_SPECULATIVE_TYPE_EAGLE3,
+    COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE,
+    COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K,
+    COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V,
+    COMMON_SPECULATIVE_TYPE_NGRAM_MOD,
+    COMMON_SPECULATIVE_TYPE_NGRAM_CACHE
 };
 
-struct common_speculative * common_speculative_init(
-        struct llama_context * ctx_tgt,
-        struct llama_context * ctx_dft) {
-    auto * result = new common_speculative {
-        /* .ctx_tgt    = */ ctx_tgt,
-        /* .ctx_dft    = */ ctx_dft,
-        /* .smpl       = */ nullptr,
-        /* .batch      = */ llama_batch_init(llama_n_batch(ctx_dft), 0, 1),
-        /* .prompt_dft = */ {},
-        /* .vocab_dft_compatible = */ false,
-    };
+const std::map<std::string, enum common_speculative_type> common_speculative_type_from_name_map = {
+    {"none",          COMMON_SPECULATIVE_TYPE_NONE},
+    {"draft",         COMMON_SPECULATIVE_TYPE_DRAFT},
+    {"eagle3",        COMMON_SPECULATIVE_TYPE_EAGLE3},
+    {"ngram_simple",  COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE},
+    {"ngram_map_k",   COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K},
+    {"ngram_map_k4v", COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V},
+    {"ngram_mod",     COMMON_SPECULATIVE_TYPE_NGRAM_MOD},
+    {"ngram_cache",   COMMON_SPECULATIVE_TYPE_NGRAM_CACHE}
+};
 
-    // TODO: optimize or pass from outside?
-#if 0
-    {
-        common_params_sampling params;
-        params.no_perf = false;
+struct common_speculative_config {
+    common_speculative_type type;
+    common_params_speculative params;
 
-        params.top_k = 40;
-        params.top_p = 0.9;
+    common_speculative_config(common_speculative_type t,
+            const common_params_speculative & p = common_params_speculative{}) : type(t), params(p) {}
+};
 
-        params.samplers = {
-            COMMON_SAMPLER_TYPE_TOP_K,
-            COMMON_SAMPLER_TYPE_TOP_P,
-            COMMON_SAMPLER_TYPE_INFILL,
-        };
-
-        result->smpl = common_sampler_init(llama_get_model(ctx_dft), params);
-    }
-#else
-    {
-        common_params_sampling params;
-        params.no_perf = false;
-
-        params.top_k = 10;
-
-        params.samplers = {
-            COMMON_SAMPLER_TYPE_TOP_K,
-        };
-
-        result->smpl = common_sampler_init(llama_get_model(ctx_dft), params);
-    }
-#endif
-
-    result->vocab_dft_compatible = common_speculative_are_compatible(ctx_tgt, ctx_dft);
-    LOG_DBG("vocab_dft_compatible = %d\n", result->vocab_dft_compatible);
-
-    return result;
-}
-
-void common_speculative_free(struct common_speculative * spec) {
-    if (spec == nullptr) {
-        return;
-    }
-
-    common_sampler_free(spec->smpl);
-
-    llama_batch_free(spec->batch);
-
-    delete spec;
-}
-
-bool common_speculative_are_compatible(
-    const struct llama_context * ctx_tgt,
-    const struct llama_context * ctx_dft) {
-    const struct llama_model * model_tgt = llama_get_model(ctx_tgt);
-    const struct llama_model * model_dft = llama_get_model(ctx_dft);
-
-    const struct llama_vocab * vocab_tgt = llama_model_get_vocab(model_tgt);
-    const struct llama_vocab * vocab_dft = llama_model_get_vocab(model_dft);
+static bool common_speculative_are_compatible(
+    const llama_model * model_tgt,
+    const llama_model * model_dft) {
+    const llama_vocab * vocab_tgt = llama_model_get_vocab(model_tgt);
+    const llama_vocab * vocab_dft = llama_model_get_vocab(model_dft);
 
     const bool vocab_type_tgt = llama_vocab_type(vocab_tgt);
     LOG_DBG("%s: vocab_type tgt: %d\n", __func__, vocab_type_tgt);
@@ -134,11 +92,12 @@ bool common_speculative_are_compatible(
         for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
             const char * token_text_tgt = llama_vocab_get_text(vocab_tgt, i);
             const char * token_text_dft = llama_vocab_get_text(vocab_dft, i);
+
             if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
                 LOG_DBG("%s: draft model vocab must match target model to use speculation but ", __func__);
                 LOG_DBG("token %d content differs - target '%s', draft '%s'\n", i,
-                        common_token_to_piece(ctx_tgt, i).c_str(),
-                        common_token_to_piece(ctx_dft, i).c_str());
+                        common_token_to_piece(vocab_tgt, i).c_str(),
+                        common_token_to_piece(vocab_dft, i).c_str());
                 return false;
             }
         }
@@ -147,215 +106,932 @@ bool common_speculative_are_compatible(
     return true;
 }
 
-void common_speculative_add_replacement_tgt_dft(
-        struct common_speculative * spec,
-        const char *source, const char *dest) {
-    spec->tgt_dft_replacements[source] = dest;
+// state of an implementation of speculative decoding
+//
+// each implementation has a unique type and a state that is implementation-specific
+// in a subclass of common_speculative_state
+struct common_speculative_state {
+    const enum common_speculative_type type;
+
+    // TODO: rename to n_call_draft, n_gen_drafts, n_acc_drafts, n_gen_tokens, n_acc_tokens
+    // TODO: add n_call_begin, n_call_accept
+    size_t drafts_call_count       = 0; // number of times this implementation was called.
+    size_t drafts_generated_count  = 0; // number of times a draft or part was generated by this implementation.
+    size_t drafts_accepted_count   = 0; // number of times a draft or part was accepted by the target model.
+    size_t drafts_generated_tokens = 0; // number of tokens generated by this implementation.
+    size_t drafts_accepted_tokens  = 0; // number of tokens accepted by the target model.
+
+    // TODO: track performance of most recent calls
+    const bool gen_perf = true; // whether to generate performance stats.
+
+    // TODO: rename to t_draft_us
+    // TODO: add t_begin_us, t_accept_us
+    int64_t gen_duration_us = 0; // total time spent in this implementation in microseconds.
+
+    common_speculative_state(enum common_speculative_type type) : type(type) {}
+
+    virtual ~common_speculative_state() = default;
+
+    virtual void begin(const llama_tokens & prompt) = 0;
+
+    virtual void draft(
+            const common_params_speculative & params,
+            const llama_tokens & prompt_tgt,
+            llama_token id_last,
+            llama_tokens & result) = 0;
+
+    virtual void accept(uint16_t n_accepted) = 0;
+};
+
+struct common_speculative_state_draft : public common_speculative_state {
+    llama_context * ctx_tgt; // only used for retokenizing from ctx_dft
+    llama_context * ctx_dft;
+
+    common_sampler * smpl;
+
+    llama_batch  batch;
+    llama_tokens prompt_dft;
+
+    bool vocab_cmpt = true; // whether retokenization is needed
+    std::unordered_map<std::string, std::string> vocab_map;
+
+    common_speculative_state_draft(
+            enum common_speculative_type type,
+            llama_context * ctx_tgt,
+            llama_context * ctx_dft,
+            const std::vector<std::pair<std::string, std::string>> & replacements)
+        : common_speculative_state(type)
+        , ctx_tgt(ctx_tgt)
+        , ctx_dft(ctx_dft)
+    {
+        batch = llama_batch_init(llama_n_batch(ctx_dft), 0, 1);
+        smpl = nullptr;
+
+        // TODO: optimize or pass from outside?
+        // {
+        //     common_params_sampling params;
+        //     params.no_perf = false;
+        //
+        //     params.top_k = 40;
+        //     params.top_p = 0.9;
+        //
+        //     params.samplers = {
+        //         COMMON_SAMPLER_TYPE_TOP_K,
+        //         COMMON_SAMPLER_TYPE_TOP_P,
+        //         COMMON_SAMPLER_TYPE_INFILL,
+        //     };
+        //
+        //     result->smpl = common_sampler_init(llama_get_model(ctx_dft), params);
+        // }
+        {
+            common_params_sampling params;
+            params.no_perf = false;
+            params.top_k = 10;
+            params.samplers = {
+                COMMON_SAMPLER_TYPE_TOP_K,
+            };
+
+            smpl = common_sampler_init(llama_get_model(ctx_dft), params);
+        }
+
+        vocab_cmpt = common_speculative_are_compatible(llama_get_model(ctx_tgt), llama_get_model(ctx_dft));
+        LOG_DBG("vocab_cmpt = %d\n", vocab_cmpt);
+
+        if (!vocab_cmpt) {
+            LOG_WRN("the target and draft vocabs are not compatible - tokens will be translated between the two\n");
+
+            for (const auto & pair : replacements) {
+                vocab_map[pair.first] = pair.second;
+            }
+        }
+    }
+
+    ~common_speculative_state_draft() override {
+        llama_perf_context_print(ctx_dft);
+
+        llama_free(ctx_dft);
+
+        common_sampler_free(smpl);
+
+        llama_batch_free(batch);
+    }
+
+    void begin(const llama_tokens & prompt) override {
+        GGML_UNUSED(prompt);
+    }
+
+    void draft(
+            const common_params_speculative & params,
+            const llama_tokens & prompt_tgt,
+            llama_token id_last,
+            llama_tokens & result) override {
+        auto * spec = this;
+
+        auto & batch      = spec->batch;
+        auto & ctx_tgt    = spec->ctx_tgt;
+        auto & ctx_dft    = spec->ctx_dft;
+        auto & smpl       = spec->smpl;
+        auto & prompt_dft = spec->prompt_dft;
+
+        auto * mem_dft = llama_get_memory(ctx_dft);
+
+        int reuse_i = 0;
+        int reuse_n = 0;
+
+        const int n_ctx = llama_n_ctx(ctx_dft) - params.n_max;
+
+        llama_tokens prompt_cnv;
+        if (!spec->vocab_cmpt) {
+            std::string text;
+
+            text = common_detokenize(ctx_tgt, prompt_tgt, true);
+            text = replace_to_dft(text);
+
+            LOG_DBG("%s: main->draft detokenized string: '%s'\n", __func__, text.c_str());
+
+            prompt_cnv = common_tokenize(ctx_dft, text, false, true);
+
+            // convert id_last to draft vocab. llama_detokenize is called directly to avoid an allocation
+            const auto * model_tgt = llama_get_model(ctx_tgt);
+            const auto * vocab_tgt = llama_model_get_vocab(model_tgt);
+
+            int32_t n_chars = llama_detokenize(vocab_tgt, &id_last, 1, nullptr, 0, false, false);
+            GGML_ASSERT(n_chars < 0 && "failed to detokenize id_last");
+
+            text.resize(-n_chars);
+            llama_detokenize(vocab_tgt, &id_last, 1, text.data(), text.size(), false, false);
+            text = replace_to_dft(text);
+
+            LOG_DBG("main->draft detokenized id_last(%d): '%s'\n", id_last, text.c_str());
+            id_last = common_tokenize(ctx_dft, text, false, true)[0];
+        }
+
+        const llama_tokens & prompt_cur = spec->vocab_cmpt ? prompt_tgt : prompt_cnv;
+
+        const int i_start = std::max<int>(0, (int) prompt_cur.size() - n_ctx);
+
+        // reuse as much as possible from the old draft context
+        // ideally, the draft context should be as big as the target context and we will always reuse the entire prompt
+        for (int i = 0; i < (int) prompt_dft.size(); ++i) {
+            int cur = 0;
+            while (i_start + cur < (int) prompt_cur.size() &&
+                    i       + cur < (int) prompt_dft.size() &&
+                    prompt_cur[i_start + cur] == prompt_dft[i + cur]) {
+                cur++;
+            }
+
+            if ((cur >= 256 || n_ctx >= (int) prompt_cur.size()) && cur > reuse_n) {
+                reuse_i = i;
+                reuse_n = cur;
+            }
+        }
+
+        LOG_DBG("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt_dft.size());
+
+        result.clear();
+        result.reserve(params.n_max);
+
+        if (reuse_n == 0) {
+            llama_memory_clear(mem_dft, false);
+            prompt_dft.clear();
+        } else {
+            // this happens when a previous draft has been discarded (for example, due to being too small), but the
+            // target model agreed with it. in this case, we simply pass back the previous results to save compute
+            if (reuse_i + reuse_n < (int) prompt_dft.size() && prompt_dft[reuse_i + reuse_n] == id_last) {
+                for (int i = reuse_i + reuse_n + 1; i < (int) prompt_dft.size(); ++i) {
+                    result.push_back(prompt_dft[i]);
+
+                    if (params.n_max <= (int) result.size()) {
+                        break;
+                    }
+                }
+
+                return;
+            }
+
+            if (reuse_i > 0) {
+                llama_memory_seq_rm (mem_dft, 0, 0, reuse_i);
+                llama_memory_seq_add(mem_dft, 0, reuse_i, -1, -reuse_i);
+
+                prompt_dft.erase(prompt_dft.begin(), prompt_dft.begin() + reuse_i);
+            }
+
+            if (reuse_n < (int) prompt_dft.size()) {
+                llama_memory_seq_rm (mem_dft, 0, reuse_n, -1);
+                prompt_dft.erase(prompt_dft.begin() + reuse_n, prompt_dft.end());
+            }
+        }
+
+        // prepare a batch to evaluate any new tokens in the prompt
+        common_batch_clear(batch);
+
+        for (size_t i = i_start + reuse_n; i < prompt_cur.size(); ++i) {
+            //LOG_DBG("i = %d, i_start = %d, reuse_n = %d, i - i_start = %d, id = %6d\n", i, i_start, reuse_n, i - i_start, prompt_cur[i]);
+            common_batch_add(batch, prompt_cur[i], i - i_start, { 0 }, false);
+
+            prompt_dft.push_back(prompt_cur[i]);
+        }
+
+        // we should rarely end-up here during normal decoding
+        if (batch.n_tokens > 0) {
+            //LOG_DBG("%s: draft prompt batch: %s\n", __func__, string_from(ctx, batch).c_str());
+
+            llama_decode(ctx_dft, batch);
+        }
+
+        const llama_pos n_past = prompt_dft.size();
+
+        LOG_DBG("%s: n_past = %d\n", __func__, n_past);
+
+        common_batch_clear(batch);
+        common_batch_add  (batch, id_last, n_past, { 0 }, true);
+
+        prompt_dft.push_back(id_last);
+
+        LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx_dft, prompt_dft).c_str());
+
+        llama_decode(ctx_dft, batch);
+
+        common_sampler_reset(smpl);
+
+        // sample n_draft tokens from the draft model
+        for (int i = 0; i < params.n_max; ++i) {
+            common_batch_clear(batch);
+
+            common_sampler_sample(smpl, ctx_dft, 0, true);
+
+            const auto * cur_p = common_sampler_get_candidates(smpl, true);
+
+            for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
+                LOG_DBG(" - draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",
+                        k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
+            }
+
+            // add drafted token for each sequence
+            const llama_token id = cur_p->data[0].id;
+
+            common_sampler_accept(smpl, id, true);
+
+            result.push_back(id);
+
+            if (params.n_max <= (int) result.size()) {
+                break;
+            }
+
+            // only collect very high-confidence draft tokens
+            if (cur_p->data[0].p < params.p_min) {
+                break;
+            }
+
+            common_batch_add(batch, id, n_past + i + 1, { 0 }, true);
+
+            // evaluate the drafted tokens on the draft model
+            llama_decode(ctx_dft, batch);
+
+            prompt_dft.push_back(id);
+        }
+
+        if (!spec->vocab_cmpt) {
+            std::string detokenized = common_detokenize(ctx_dft, result, true);
+            detokenized = replace_to_tgt(detokenized);
+            LOG_DBG("draft->main detokenized string: '%s'\n", detokenized.c_str());
+            result = common_tokenize(ctx_tgt, detokenized, false, true);
+            if (result.size() > (size_t)params.n_max) {
+                result.resize(params.n_max);
+            }
+        }
+    }
+
+    void accept(uint16_t n_accepted) override {
+        // noop
+        GGML_UNUSED(n_accepted);
+    }
+
+    std::string replace_to_dft(const std::string & input) const {
+        std::string result = input;
+
+        for (const auto & pair : this->vocab_map) {
+            size_t pos = result.find(pair.first);
+            while (pos != std::string::npos) {
+                result.replace(pos, pair.first.length(), pair.second);
+                pos = result.find(pair.first, pos + pair.second.length());
+            }
+        }
+
+        return result;
+    }
+
+    std::string replace_to_tgt(const std::string & input) const {
+        std::string result = input;
+
+        for (const auto & pair : this->vocab_map) {
+            size_t pos = result.find(pair.second);
+            while (pos != std::string::npos) {
+                result.replace(pos, pair.second.length(), pair.first);
+                pos = result.find(pair.second, pos + pair.first.length());
+            }
+        }
+
+        return result;
+    }
+};
+
+struct common_speculative_state_eagle3 : public common_speculative_state {
+    common_speculative_state_eagle3(enum common_speculative_type type) : common_speculative_state(type) {}
+
+    void begin(const llama_tokens & prompt) override {
+        GGML_UNUSED(prompt);
+    }
+
+    void draft(
+            const common_params_speculative & params,
+            const llama_tokens & prompt_tgt,
+            llama_token id_last,
+            llama_tokens & draft_tokens) override {
+        // TODO: implement
+        GGML_UNUSED(params);
+        GGML_UNUSED(prompt_tgt);
+        GGML_UNUSED(id_last);
+        GGML_UNUSED(draft_tokens);
+    }
+
+    void accept(uint16_t n_accepted) override {
+        // noop
+        GGML_UNUSED(n_accepted);
+    }
+};
+
+// state of self-speculation (simple implementation, not ngram-map)
+struct common_speculative_state_ngram_simple : public common_speculative_state {
+    common_ngram_simple_state state;
+
+    common_speculative_state_ngram_simple(
+            enum common_speculative_type type,
+            common_ngram_simple_state state)
+        : common_speculative_state(type), state(state) {}
+
+    void begin(const llama_tokens & prompt) override {
+        GGML_UNUSED(prompt);
+    }
+
+    void draft(
+            const common_params_speculative & params,
+            const llama_tokens & prompt_tgt,
+            llama_token id_last,
+            llama_tokens & result) override {
+        result = common_ngram_simple_draft(state, prompt_tgt, id_last);
+        GGML_UNUSED(params);
+    }
+
+    void accept(uint16_t n_accepted) override {
+        // noop
+        GGML_UNUSED(n_accepted);
+    }
+};
+
+struct common_speculative_state_ngram_map_k : public common_speculative_state {
+    // draft ngram map for speculative decoding without draft model
+    common_ngram_map map;
+
+    common_speculative_state_ngram_map_k(
+            enum common_speculative_type type,
+            common_ngram_map map)
+        : common_speculative_state(type), map(std::move(map)) {}
+
+    void begin(const llama_tokens & prompt) override {
+        GGML_UNUSED(prompt);
+    }
+
+    void draft(
+            const common_params_speculative & params,
+            const llama_tokens & prompt_tgt,
+            llama_token id_last,
+            llama_tokens & result) override {
+        common_ngram_map_draft(map, prompt_tgt, id_last, result);
+        GGML_UNUSED(params);
+    }
+
+    void accept(uint16_t n_accepted) override {
+        common_ngram_map_accept(map, n_accepted);
+    }
+};
+
+struct common_speculative_state_ngram_mod : public common_speculative_state {
+    common_ngram_mod & mod;
+
+    // the last position in the prompt that was added to the ngram container
+    size_t i_last = 0;
+
+    // length of the last drafted n‑gram (number of tokens returned by draft)
+    size_t n_draft_last = 0;
+
+    // consecutive accept rounds with low acceptance fraction (< 0.5)
+    int n_low = 0;
+
+    // enable trace logging if LLAMA_TRACE is set
+    const bool verbose;
+
+    common_speculative_state_ngram_mod(enum common_speculative_type type, common_ngram_mod & mod)
+        : common_speculative_state(type), mod(mod), verbose(std::getenv("LLAMA_TRACE") != nullptr) {
+        static_assert(sizeof(llama_token) == sizeof(common_ngram_mod::entry_t));
+    }
+
+    void begin(const llama_tokens & prompt) override {
+        i_last = 0;
+
+        n_draft_last = 0;
+
+        const size_t n = mod.get_n();
+
+        if (prompt.size() < n) {
+            return;
+        }
+
+        for (size_t i = 0; i < prompt.size() - n; ++i) {
+            mod.add(prompt.data() + i);
+        }
+
+        i_last = prompt.size() - n;
+
+        const double f = (double)mod.get_used() / (double)mod.size();
+        LOG_INF("%s: ngram_mod occupancy = %zu/%zu (%.2f)\n", __func__, mod.get_used(), mod.size(), f);
+
+        constexpr double f_thold = 0.25;
+        if (f > f_thold) {
+            LOG_WRN("%s: ngram_mod occupancy %.2f exceeds threshold (%.2f) - resetting\n", __func__, f, f_thold);
+
+            mod.reset();
+        }
+    }
+
+    void draft(
+            const common_params_speculative & params,
+            const llama_tokens & prompt_tgt,
+            llama_token id_last,
+            llama_tokens & result) override {
+        GGML_UNUSED(params);
+
+        n_draft_last = 0;
+
+        const size_t cur_len = prompt_tgt.size();
+        if (cur_len < mod.get_n()) {
+            return;
+        }
+
+        const size_t n = mod.get_n();
+
+        // add new ngrams in chunks
+        if (i_last + 32 < cur_len) {
+            for (size_t i = i_last; i < cur_len - n; ++i) {
+                mod.add(prompt_tgt.data() + i);
+            }
+
+            i_last = cur_len - n;
+        }
+
+        result.resize(n + params.n_max);
+        for (size_t i = 0; i < n - 1; ++i) {
+            result[i] = prompt_tgt[cur_len - n + 1 + i];
+        }
+        result[n - 1] = id_last;
+
+        for (int i = 0; i < params.n_max; ++i) {
+            const llama_token token = mod.get(result.data() + i);
+            if (token == common_ngram_mod::EMPTY) {
+                if (i < params.n_min) {
+                    result.clear();
+                    return;
+                }
+
+                result.resize(n + i);
+                break;
+            }
+            result[n + i] = token;
+        }
+
+        // only return the m tokens that were drafted
+        for (size_t i = 0; n + i < result.size(); ++i) {
+            result[i] = result[n + i];
+        }
+        result.resize(result.size() - n);
+
+        // store length of drafted n‑gram for later acceptance analysis
+        n_draft_last = result.size();
+    }
+
+    void accept(uint16_t n_accepted) override {
+        if (verbose) {
+            LOG_INF("%s: accepted %d tokens from %zu drafted tokens\n", __func__, n_accepted, n_draft_last);
+        }
+
+        // compute acceptance fraction if we have a recorded draft length
+        if (n_draft_last > 0) {
+            const double f_acc = (double)n_accepted / (double)n_draft_last;
+            if (f_acc < 0.5) {
+                n_low++;
+                if (n_low >= 3) {
+                    LOG_WRN("%s: low acceptance streak (%d) – resetting ngram_mod\n", __func__, n_low);
+
+                    mod.reset();
+                    n_low = 0;
+                }
+            } else {
+                n_low = 0;
+            }
+        }
+    }
+};
+
+struct common_speculative_state_ngram_cache : public common_speculative_state {
+    uint16_t n_draft;
+    bool save_dynamic;
+    bool save_static;
+
+    common_ngram_cache ngram_cache_context;
+    common_ngram_cache ngram_cache_dynamic;
+    common_ngram_cache ngram_cache_static;
+
+    size_t cache_size = 0; // number of tokens in n-gram cache
+
+    common_speculative_state_ngram_cache(
+            const enum common_speculative_type type,
+            const std::string & path_static,
+            const std::string & path_dynamic,
+            uint16_t            n_draft,
+            bool                save_dynamic,
+            bool                save_static)
+        : common_speculative_state(type)
+        , n_draft(n_draft)
+        , save_dynamic(save_dynamic)
+        , save_static(save_static)
+    {
+        if (!path_static.empty()) {
+            try {
+                ngram_cache_static = common_ngram_cache_load(path_static);
+            } catch (...) {
+                LOG_ERR("failed to open static lookup cache: %s", path_static.c_str());
+                GGML_ABORT("Couldn't read static lookup cache");
+            }
+        }
+
+        if (!path_dynamic.empty()) {
+            try {
+                ngram_cache_dynamic = common_ngram_cache_load(path_dynamic);
+            } catch (...) {
+                LOG_ERR("failed to open dynamic lookup cache: %s", path_dynamic.c_str());
+                GGML_ABORT("Couldn't read dynamic lookup cache");
+            }
+        }
+    }
+
+    void begin(const llama_tokens & prompt) override {
+        GGML_UNUSED(prompt);
+    }
+
+    void draft(
+            const common_params_speculative & params,
+            const llama_tokens & prompt_tgt,
+            llama_token id_last,
+            llama_tokens & result) override {
+        GGML_UNUSED(params);
+
+        if (cache_size < prompt_tgt.size() + 1) {
+            llama_tokens tokens_new;
+            tokens_new.reserve(prompt_tgt.size() + 1 - cache_size);
+            for (size_t j = cache_size; j < prompt_tgt.size(); ++j) {
+                tokens_new.push_back(prompt_tgt[j]);
+            }
+            tokens_new.push_back(id_last); // add the last token
+
+            // Update context ngram cache with new prompt_tgt:
+            common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX,
+                    tokens_new, tokens_new.size(), false);
+            cache_size = prompt_tgt.size() + 1;
+        }
+
+        llama_tokens inp;
+        inp.reserve(prompt_tgt.size() + 1);
+        for (size_t j = 0; j < prompt_tgt.size(); ++j) {
+            inp.push_back(prompt_tgt[j]);
+        }
+        inp.push_back(id_last);
+
+        result.push_back(id_last);
+
+        common_ngram_cache_draft(inp, result, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX,
+                ngram_cache_context,
+                ngram_cache_dynamic,
+                ngram_cache_static);
+
+        if (result.size() > 0) {
+            // delete first token in result (which is the id_last token)
+            result.erase(result.begin());
+        }
+    }
+
+    void accept(uint16_t n_accepted) override {
+        // TODO: noop
+        GGML_UNUSED(n_accepted);
+    }
+};
+
+struct common_speculative {
+    std::vector<std::unique_ptr<common_speculative_state>> impls; // list of implementations to use and their states
+    common_speculative_state * curr_impl = nullptr; // current implementation in use (for stats)
+};
+
+static common_ngram_map get_common_ngram_map(const common_speculative_config & config) {
+    uint16_t size_key   = config.params.ngram_size_n;
+    uint16_t size_value = config.params.ngram_size_m;
+    bool     key_only   = (config.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K);
+    uint16_t check_rate = config.params.ngram_check_rate;
+    uint16_t min_hits   = config.params.ngram_min_hits;
+
+    return common_ngram_map(size_key, size_value, key_only, check_rate, min_hits);
 }
 
-static std::string replace_to_dft(
-        struct common_speculative * spec,
-        const std::string& input) {
-    std::string result = input;
-    for (const auto & pair : spec->tgt_dft_replacements) {
-        size_t pos = result.find(pair.first);
-        while (pos != std::string::npos) {
-            result.replace(pos, pair.first.length(), pair.second);
-            pos = result.find(pair.first, pos + pair.second.length());
+static common_speculative_state_ngram_cache create_state_ngram_cache(
+        const std::string & path_static, const std::string & path_dynamic,
+        const common_speculative_config & config) {
+    uint16_t n_draft = 8; // TODO get from config?
+
+    // TODO bool param in common/common.h to set save_static/save_dynamic?
+    bool save_static = false;
+    bool save_dynamic = false;
+
+    common_speculative_state_ngram_cache state(config.type, path_static, path_dynamic, n_draft, save_static, save_dynamic);
+
+    return state;
+}
+
+std::string common_speculative_type_name_str() {
+    std::string result;
+    for (size_t i = 0; i < common_speculative_types.size(); i++) {
+        if (i > 0) {
+            result += ", ";
         }
+        result += common_speculative_type_to_str(common_speculative_types[i]);
     }
     return result;
 }
 
-static std::string replace_to_tgt(
-        struct common_speculative * spec,
-        const std::string& input) {
-    std::string result = input;
-    for (const auto& pair : spec->tgt_dft_replacements) {
-        size_t pos = result.find(pair.second);
-        while (pos != std::string::npos) {
-            result.replace(pos, pair.second.length(), pair.first);
-            pos = result.find(pair.second, pos + pair.first.length());
-        }
+std::string common_speculative_type_to_str(enum common_speculative_type type) {
+    switch (type) {
+        case COMMON_SPECULATIVE_TYPE_NONE:          return "none";
+        case COMMON_SPECULATIVE_TYPE_DRAFT:         return "draft";
+        case COMMON_SPECULATIVE_TYPE_EAGLE3:        return "eagle3";
+        case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE:  return "ngram_simple";
+        case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K:   return "ngram_map_k";
+        case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V: return "ngram_map_k4v";
+        case COMMON_SPECULATIVE_TYPE_NGRAM_MOD:     return "ngram_mod";
+        case COMMON_SPECULATIVE_TYPE_NGRAM_CACHE:   return "ngram_cache";
+        default:                                    return "unknown";
     }
-    return result;
 }
 
-
-llama_tokens common_speculative_gen_draft(
-        struct common_speculative * spec,
-        struct common_speculative_params params,
-        const llama_tokens & prompt_tgt_main_model, // specified in target model vocab
-        llama_token id_last) {
-    auto & batch  = spec->batch;
-    auto & ctx_tgt = spec->ctx_tgt;
-    auto & ctx_dft = spec->ctx_dft;
-    auto & smpl   = spec->smpl;
-    auto & prompt_dft = spec->prompt_dft;
-
-    auto * mem_dft = llama_get_memory(ctx_dft);
-
-    int reuse_i = 0;
-    int reuse_n = 0;
-
-    const int n_ctx = llama_n_ctx(ctx_dft) - params.n_draft;
-
-    llama_tokens prompt_tgt_draft_model;
-    if (!spec->vocab_dft_compatible) {
-        std::string text;
-        text = common_detokenize(ctx_tgt, prompt_tgt_main_model, true);
-        text = replace_to_dft(spec, text);
-        LOG_DBG("%s: main->draft detokenized string: '%s'\n", __func__, text.c_str());
-        prompt_tgt_draft_model = common_tokenize(ctx_dft, text, false, true);
-
-        // convert id_last to draft vocab. llama_detokenize is called directly to avoid an allocation
-        const auto * model_tgt = llama_get_model(ctx_tgt);
-        const auto * vocab_tgt = llama_model_get_vocab(model_tgt);
-
-        int32_t n_chars = llama_detokenize(vocab_tgt, &id_last, 1, nullptr, 0, false, false);
-        GGML_ASSERT(n_chars < 0 && "failed to detokenize id_last");
-        text.resize(-n_chars);
-        llama_detokenize(vocab_tgt, &id_last, 1, text.data(), text.size(), false, false);
-        text = replace_to_dft(spec, text);
-
-        LOG_DBG("main->draft detokenized id_last(%d): '%s'\n", id_last, text.c_str());
-        id_last = common_tokenize(ctx_dft, text, false, true)[0];
+enum common_speculative_type common_speculative_type_from_name(const std::string & name) {
+    const auto it = common_speculative_type_from_name_map.find(name);
+    if (it == common_speculative_type_from_name_map.end()) {
+        return COMMON_SPECULATIVE_TYPE_COUNT;
     }
-    // prompt_tgt's tokens will always be compatible with ctx_dft
-    const llama_tokens &prompt_tgt =
-        spec->vocab_dft_compatible ? prompt_tgt_main_model : prompt_tgt_draft_model;
+    return it->second;
+}
 
-    const int i_start = std::max<int>(0, (int) prompt_tgt.size() - n_ctx);
-
-    // reuse as much as possible from the old draft context
-    // ideally, the draft context should be as big as the target context and we will always reuse the entire prompt
-    for (int i = 0; i < (int) prompt_dft.size(); ++i) {
-        int cur = 0;
-        while (i_start + cur < (int) prompt_tgt.size() &&
-               i       + cur < (int) prompt_dft.size() &&
-               prompt_tgt[i_start + cur] == prompt_dft[i + cur]) {
-            cur++;
-        }
-
-        if ((cur >= params.n_reuse || n_ctx >= (int) prompt_tgt.size()) && cur > reuse_n) {
-            reuse_i = i;
-            reuse_n = cur;
+// initialization of the speculative decoding system
+//
+common_speculative * common_speculative_init(
+        common_params_speculative & params,
+        llama_context             * ctx_tgt) {
+    llama_context * ctx_dft = nullptr;
+    if (params.model_dft) {
+        ctx_dft = llama_init_from_model(params.model_dft, params.cparams_dft);
+        if (ctx_dft == nullptr) {
+            LOG_ERR("%s", "failed to create draft context\n");
+            return nullptr;
         }
     }
 
-    LOG_DBG("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt_dft.size());
+    // Compute the implementations to use based on the config and their order of preference
+    std::vector<common_speculative_config> configs = {}; // list of speculative configs to try
+    {
+        bool has_draft = !params.mparams_dft.path.empty();
+        bool has_draft_eagle3 = false; // TODO PR-18039: if params.speculative.eagle3
 
-    llama_tokens result;
-    result.reserve(params.n_draft);
+        bool has_ngram_cache   = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_CACHE);
+        bool has_ngram_simple  = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE);
+        bool has_ngram_map_k   = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K);
+        bool has_ngram_map_k4v = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V);
+        bool has_ngram_mod     = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_MOD);
 
-    if (reuse_n == 0) {
-        llama_memory_clear(mem_dft, false);
-        prompt_dft.clear();
-    } else {
-        // this happens when a previous draft has been discarded (for example, due to being too small), but the
-        // target model agreed with it. in this case, we simply pass back the previous results to save compute
-        if (reuse_i + reuse_n < (int) prompt_dft.size() && prompt_dft[reuse_i + reuse_n] == id_last) {
-            for (int i = reuse_i + reuse_n + 1; i < (int) prompt_dft.size(); ++i) {
-                result.push_back(prompt_dft[i]);
+        // In a more complex implementation we could use the same implementation but with different parameters.
+        // This was initially used in PR-18471 but removed to simplify the code.
+        if (has_ngram_simple) {
+            // This implementation can guess a lot of tokens without any draft model.
+            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, params));
+        }
+        if (has_ngram_map_k) {
+            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K, params));
+        }
+        if (has_ngram_map_k4v) {
+            // This implementation can guess tokens with high acceptance rate but is more expensive.
+            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, params));
+        }
+        if (has_ngram_mod) {
+            // shared instance for all speculative decoding contexts
+            if (!params.ngram_mod) {
+                params.ngram_mod = std::make_shared<common_ngram_mod>(params.ngram_size_n, 4*1024*1024);
 
-                if (params.n_draft <= (int) result.size()) {
-                    break;
+                LOG_INF("%s: initialized ngram_mod with n=%d, size=%zu (%.3f MB)\n", __func__,
+                        params.ngram_size_n, params.ngram_mod->size(),
+                        (float)(params.ngram_mod->size_bytes())/1024/1024);
+
+                if (params.ngram_size_n < 16) {
+                    LOG_WRN("%s: ngram_mod n=%d is too small - poor quality is possible, see: https://github.com/ggml-org/llama.cpp/pull/19164\n", __func__, params.ngram_size_n);
                 }
             }
 
-            return result;
+            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MOD, params));
         }
-
-        if (reuse_i > 0) {
-            llama_memory_seq_rm (mem_dft, 0, 0, reuse_i);
-            llama_memory_seq_add(mem_dft, 0, reuse_i, -1, -reuse_i);
-
-            prompt_dft.erase(prompt_dft.begin(), prompt_dft.begin() + reuse_i);
+        if (has_ngram_cache) {
+            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, params));
         }
-
-        if (reuse_n < (int) prompt_dft.size()) {
-            llama_memory_seq_rm (mem_dft, 0, reuse_n, -1);
-            prompt_dft.erase(prompt_dft.begin() + reuse_n, prompt_dft.end());
+        if (has_draft) {
+            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT, params));
+        }
+        if (has_draft_eagle3) {
+            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_EAGLE3, params));
         }
     }
 
-    // prepare a batch to evaluate any new tokens in the prompt
-    common_batch_clear(batch);
+    std::vector<std::unique_ptr<common_speculative_state>> impls = {};
 
-    for (size_t i = i_start + reuse_n; i < prompt_tgt.size(); ++i) {
-        //LOG_DBG("i = %d, i_start = %d, reuse_n = %d, i - i_start = %d, id = %6d\n", i, i_start, reuse_n, i - i_start, prompt_tgt[i]);
-        common_batch_add(batch, prompt_tgt[i], i - i_start, { 0 }, false);
+    for (const common_speculative_config & config : configs) {
+        LOG_DBG("%s: adding implementation %s\n", __func__, common_speculative_type_to_str(config.type).c_str());
+        switch (config.type) {
+            case COMMON_SPECULATIVE_TYPE_NONE:
+                break;
+            case COMMON_SPECULATIVE_TYPE_DRAFT: {
+                impls.push_back(std::make_unique<common_speculative_state_draft>(config.type,
+                    /* .ctx_tgt      = */ ctx_tgt,
+                    /* .ctx_dft      = */ ctx_dft,
+                    /* .replacements = */ params.replacements
+                ));
+                break;
+            }
+            case COMMON_SPECULATIVE_TYPE_EAGLE3: {
+                impls.push_back(std::make_unique<common_speculative_state_eagle3>(config.type));
+                break;
+            }
+            case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE: {
+                common_ngram_map ngram_map = get_common_ngram_map(config);
 
-        prompt_dft.push_back(prompt_tgt[i]);
-    }
+                uint16_t ngram_size_key   = ngram_map.size_key;
+                uint16_t mgram_size_value = ngram_map.size_value;
+                uint16_t check_rate       = ngram_map.check_rate;
 
-    // we should rarely end-up here during normal decoding
-    if (batch.n_tokens > 0) {
-        //LOG_DBG("%s: draft prompt batch: %s\n", __func__, string_from(ctx, batch).c_str());
-
-        llama_decode(ctx_dft, batch);
-    }
-
-    const llama_pos n_past = prompt_dft.size();
-
-    LOG_DBG("%s: n_past = %d\n", __func__, n_past);
-
-    common_batch_clear(batch);
-    common_batch_add  (batch, id_last, n_past, { 0 }, true);
-
-    prompt_dft.push_back(id_last);
-
-    LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx_dft, prompt_dft).c_str());
-
-    llama_decode(ctx_dft, batch);
-
-    common_sampler_reset(smpl);
-
-    // sample n_draft tokens from the draft model
-    for (int i = 0; i < params.n_draft; ++i) {
-        common_batch_clear(batch);
-
-        common_sampler_sample(smpl, ctx_dft, 0, true);
-
-        const auto * cur_p = common_sampler_get_candidates(smpl, true);
-
-        for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
-            LOG_DBG(" - draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",
-                    k, i, cur_p->data[k].id, cur_p->data[k].p, common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
-        }
-
-        // add drafted token for each sequence
-        const llama_token id = cur_p->data[0].id;
-
-        common_sampler_accept(smpl, id, true);
-
-        result.push_back(id);
-
-        if (params.n_draft <= (int) result.size()) {
-            break;
-        }
-
-        // only collect very high-confidence draft tokens
-        if (cur_p->data[0].p < params.p_min) {
-            break;
-        }
-
-        common_batch_add(batch, id, n_past + i + 1, { 0 }, true);
-
-        // evaluate the drafted tokens on the draft model
-        llama_decode(ctx_dft, batch);
-
-        prompt_dft.push_back(id);
-    }
-
-    if (!spec->vocab_dft_compatible) {
-        std::string detokenized = common_detokenize(ctx_dft, result, true);
-        detokenized = replace_to_tgt(spec, detokenized);
-        LOG_DBG("draft->main detokenized string: '%s'\n", detokenized.c_str());
-        result = common_tokenize(ctx_tgt, detokenized, false, true);
-        if (result.size() > (size_t)params.n_draft) {
-            result.resize(params.n_draft);
+                auto config_simple = common_ngram_simple_config{
+                    /* .size_ngram      = */ ngram_size_key,
+                    /* .size_mgram      = */ mgram_size_value,
+                    /* .check_rate      = */ check_rate
+                };
+                auto state = std::make_unique<common_speculative_state_ngram_simple>(
+                    /* .type            = */ config.type,
+                    /* .state           = */ common_ngram_simple_state(config_simple)
+                );
+                impls.push_back(std::move(state));
+                break;
+            }
+            case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K:
+            case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V: {
+                impls.push_back(std::make_unique<common_speculative_state_ngram_map_k>(
+                    (config.type),
+                    get_common_ngram_map(config)
+                ));
+                break;
+            }
+            case COMMON_SPECULATIVE_TYPE_NGRAM_MOD: {
+                GGML_ASSERT(config.params.ngram_mod);
+                impls.push_back(std::make_unique<common_speculative_state_ngram_mod>(config.type, *config.params.ngram_mod));
+                break;
+            }
+            case COMMON_SPECULATIVE_TYPE_NGRAM_CACHE: {
+                auto state = create_state_ngram_cache(
+                        params.lookup_cache_static, params.lookup_cache_dynamic, config);
+                impls.push_back(std::make_unique<common_speculative_state_ngram_cache>(state));
+                break;
+            }
+            default:
+                break;
         }
     }
+
+    if (impls.empty()) {
+        LOG_WRN("%s", "no implementations specified for speculative decoding\n");
+        return nullptr;
+    }
+
+    auto * result = new common_speculative {
+        /* .impls = */ std::move(impls)
+    };
+
     return result;
 }
+
+void common_speculative_free(common_speculative * spec) {
+    if (spec == nullptr) {
+        return;
+    }
+
+    delete spec;
+}
+
+void common_speculative_begin(common_speculative * spec, const llama_tokens & prompt) {
+    if (spec == nullptr) {
+        return;
+    }
+
+    for (auto & impl : spec->impls) {
+        impl->begin(prompt);
+    }
+}
+
+llama_tokens common_speculative_draft(
+        common_speculative * spec,
+        const common_params_speculative & params,
+        const llama_tokens & prompt_tgt, // specified in target model vocab
+        llama_token id_last) {
+    llama_tokens result;
+
+    spec->curr_impl = nullptr; // reset current implementation
+
+    for (auto & impl : spec->impls) {
+        {
+            const int64_t t_start_us = impl->gen_perf ? ggml_time_us() : 0;
+
+            impl->draft(params, prompt_tgt, id_last, result);
+
+            const int64_t t_now_us = impl->gen_perf ? ggml_time_us() : 0;
+
+            impl->drafts_call_count++;
+            impl->gen_duration_us += t_now_us - t_start_us; // accumulate duration for this implementation
+        }
+
+        if (!result.empty()) {
+            LOG_DBG("%s: called impl %s, hist size = %zu, call_count = %zu, gen = %zu\n", __func__,
+                    common_speculative_type_to_str(impl.get()->type).c_str(), prompt_tgt.size(),
+                    impl.get()->drafts_call_count, result.size());
+
+            spec->curr_impl = impl.get(); // set current implementation for stats
+            impl->drafts_generated_count++;
+            impl->drafts_generated_tokens += result.size();
+
+            break; // We have a draft, so break out of the loop and return it.
+        }
+    }
+
+    return result;
+}
+
+void common_speculative_accept(common_speculative * spec, uint16_t n_accepted) {
+    if (n_accepted == 0) {
+        return;
+    }
+
+    common_speculative_state * impl = spec->curr_impl;
+
+    GGML_ASSERT(impl);
+
+    if (n_accepted > 0) {
+        impl->drafts_accepted_count++;
+        impl->drafts_accepted_tokens += n_accepted;
+    }
+
+    impl->accept(n_accepted);
+}
+
+void common_speculative_print_stats(const common_speculative * spec) {
+    if (spec == nullptr) {
+        return;
+    }
+
+    for (const auto & impl : spec->impls) {
+        std::string str_perf;
+        if (impl->gen_perf) {
+            std::ostringstream oss;
+            oss << std::fixed << std::setprecision(3) << impl->gen_duration_us / 1000.0;
+            str_perf = ", dur = " + oss.str() + " ms";
+        } else {
+            str_perf = "";
+        }
+
+        // TODO: report time for begin() and accept()
+        LOG_INF("statistics %s: #calls = %zu, #gen drafts = %zu, #acc drafts = %zu, #gen tokens = %zu, #acc tokens = %zu%s\n",
+                common_speculative_type_to_str(impl->type).c_str(),
+                impl->drafts_call_count,
+                impl->drafts_generated_count,
+                impl->drafts_accepted_count,
+                impl->drafts_generated_tokens,
+                impl->drafts_accepted_tokens,
+                str_perf.c_str());
+    }
+}
diff --git a/common/speculative.h b/common/speculative.h
index e69d7aaa1e..76fe6bb7bc 100644
--- a/common/speculative.h
+++ b/common/speculative.h
@@ -5,31 +5,33 @@
 
 struct common_speculative;
 
-struct common_speculative_params {
-    int n_draft = 16;  // max drafted tokens
-    int n_reuse = 256;
+// comma separated list of all types
+std::string common_speculative_type_name_str();
 
-    float p_min = 0.75f; // min probability required to accept a token in the draft
-};
+// convert string to type
+enum common_speculative_type common_speculative_type_from_name(const std::string & name);
 
-struct common_speculative * common_speculative_init(
-        struct llama_context * ctx_tgt,
-        struct llama_context * ctx_dft
-);
+// convert type to string
+std::string common_speculative_type_to_str(enum common_speculative_type type);
 
-void common_speculative_free(struct common_speculative * spec);
+common_speculative * common_speculative_init(
+        common_params_speculative & params,
+        llama_context             * ctx_tgt);
 
-bool common_speculative_are_compatible(
-        const struct llama_context * ctx_tgt,
-        const struct llama_context * ctx_dft);
+void common_speculative_free(common_speculative * spec);
 
-void common_speculative_add_replacement_tgt_dft(
-        struct common_speculative * spec,
-        const char *source, const char *dest);
+// optionally call once at the beginning of a new generation
+void common_speculative_begin(common_speculative * spec, const llama_tokens & prompt);
 
 // sample up to n_draft tokens and add them to the batch using the draft model
-llama_tokens common_speculative_gen_draft(
-               struct common_speculative * spec,
-        struct common_speculative_params   params,
-                      const llama_tokens & prompt,
-                             llama_token   id_last);
+llama_tokens common_speculative_draft(
+                     common_speculative * spec,
+        const common_params_speculative & params,
+                     const llama_tokens & prompt,
+                            llama_token   id_last);
+
+// informs the speculative decoder that n_accepted tokens were accepted by the target model
+void common_speculative_accept(common_speculative * spec, uint16_t n_accepted);
+
+// print statistics about the speculative decoding
+void common_speculative_print_stats(const common_speculative * spec);
diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index 464ecbaab9..eb43520f98 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -514,8 +514,7 @@ class ModelBase:
         raise NotImplementedError("set_gguf_parameters() must be implemented in subclasses")
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
+        del bid # unused
         return [(self.map_tensor_name(name), data_torch)]
 
     def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
@@ -1078,6 +1077,9 @@ class TextModel(ModelBase):
         if chkhsh == "b3d1dd861f1d4c5c0d2569ce36baf3f90fe8a102db3de50dd71ff860d91be3df":
             # ref: https://huggingface.co/aari1995/German_Semantic_V3
             res = "jina-v2-de"
+        if chkhsh == "cdf5f35325780597efd76153d4d1c16778f766173908894c04afc20108536267":
+            # ref: https://huggingface.co/zai-org/GLM-4.7-Flash
+            res = "glm4"
         if chkhsh == "0ef9807a4087ebef797fc749390439009c3b9eda9ad1a097abbe738f486c01e5":
             # ref: https://huggingface.co/meta-llama/Meta-Llama-3-8B
             res = "llama-bpe"
@@ -1978,13 +1980,9 @@ class GPTNeoXModel(TextModel):
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
         n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
 
-        tensors: list[tuple[str, Tensor]] = []
-
         if re.match(r"gpt_neox\.layers\.\d+\.attention\.query_key_value\.weight", name):
             # Map bloom-style qkv_linear to gpt-style qkv_linear
             # bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252  # noqa
@@ -2011,9 +2009,7 @@ class GPTNeoXModel(TextModel):
             )
             logger.info("re-format attention.linear_qkv.bias")
 
-        tensors.append((self.map_tensor_name(name), data_torch))
-
-        return tensors
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("BloomForCausalLM", "BloomModel")
@@ -2033,15 +2029,11 @@ class BloomModel(TextModel):
         self.gguf_writer.add_file_type(self.ftype)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
         n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
 
         name = re.sub(r'transformer\.', '', name)
 
-        tensors: list[tuple[str, Tensor]] = []
-
         if re.match(r"h\.\d+\.self_attention\.query_key_value\.weight", name):
             # Map bloom-style qkv_linear to gpt-style qkv_linear
             # bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252  # noqa
@@ -2068,9 +2060,7 @@ class BloomModel(TextModel):
             )
             logger.info("re-format attention.linear_qkv.bias")
 
-        tensors.append((self.map_tensor_name(name), data_torch))
-
-        return tensors
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("MPTForCausalLM")
@@ -2105,15 +2095,13 @@ class MPTModel(TextModel):
             self.gguf_writer.add_max_alibi_bias(0.0)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         if "scales" in name:
             new_name = self.map_tensor_name(name, try_suffixes=(".weight", ".bias", ".scales"))
             new_name = new_name.replace("scales", "act.scales")
         else:
             new_name = self.map_tensor_name(name, try_suffixes=(".weight", ".bias"))
 
-        return [(new_name, data_torch)]
+        yield from super().modify_tensors(data_torch, new_name, bid)
 
 
 @ModelBase.register("OrionForCausalLM")
@@ -2167,11 +2155,9 @@ class BaichuanModel(TextModel):
         head_count = self.hparams["num_attention_heads"]
         head_count_kv = self.hparams.get("num_key_value_heads", head_count)
 
-        tensors: list[tuple[str, Tensor]] = []
-
         if bid is not None and name == f"model.layers.{bid}.self_attn.W_pack.weight":
             logger.info(f"Unpacking and permuting layer {bid}")
-            tensors = [
+            yield from [
                 (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid),
                     self._reverse_hf_permute_part(data_torch, 0, head_count, head_count)),
                 (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid),
@@ -2180,9 +2166,7 @@ class BaichuanModel(TextModel):
                     self._reverse_hf_part(data_torch, 2)),
             ]
         else:
-            tensors = [(self.map_tensor_name(name), data_torch)]
-
-        return tensors
+            yield from self.modify_tensors(data_torch, self.map_tensor_name(name), bid)
 
     def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
         if n_kv_head is not None and n_head != n_kv_head:
@@ -2263,8 +2247,6 @@ class XverseModel(TextModel):
         self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         head_count = self.hparams["num_attention_heads"]
         head_count_kv = self.hparams.get("num_key_value_heads", head_count)
 
@@ -2274,7 +2256,7 @@ class XverseModel(TextModel):
         if name.endswith("k_proj.weight"):
             data_torch = self._reverse_hf_permute(data_torch, head_count, head_count_kv)
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
         if n_kv_head is not None and n_head != n_kv_head:
@@ -2311,8 +2293,6 @@ class FalconModel(TextModel):
         self.gguf_writer.add_file_type(self.ftype)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         # QKV tensor transform
         # The original query_key_value tensor contains n_head_kv "kv groups",
         # each consisting of n_head/n_head_kv query weights followed by one key
@@ -2334,7 +2314,7 @@ class FalconModel(TextModel):
             v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
             data_torch = torch.cat((q, k, v)).reshape_as(data_torch)
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("GPTBigCodeForCausalLM")
@@ -2396,22 +2376,20 @@ class RefactModel(TextModel):
         n_head_kv = 1
         head_dim = self.hparams["n_embd"] // n_head
 
-        tensors: list[tuple[str, Tensor]] = []
-
         if bid is not None:
             if name == f"transformer.h.{bid}.attn.kv.weight":
-                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), data_torch[:n_head_kv * head_dim]))
-                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), data_torch[n_head_kv * head_dim:]))
-            elif name == f"transformer.h.{bid}.attn.q.weight":
-                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), data_torch))
-            elif name == f"transformer.h.{bid}.mlp.gate_up_proj.weight":
-                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), data_torch[:ff_dim]))
-                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), data_torch[ff_dim:]))
+                yield from super().modify_tensors(data_torch[:n_head_kv * head_dim], self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), bid)
+                yield from super().modify_tensors(data_torch[n_head_kv * head_dim:], self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), bid)
+                return
+            if name == f"transformer.h.{bid}.attn.q.weight":
+                yield from super().modify_tensors(data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), bid)
+                return
+            if name == f"transformer.h.{bid}.mlp.gate_up_proj.weight":
+                yield from super().modify_tensors(data_torch[:ff_dim], self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), bid)
+                yield from super().modify_tensors(data_torch[ff_dim:], self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), bid)
+                return
 
-        if len(tensors) == 0:
-            tensors.append((self.map_tensor_name(name), data_torch))
-
-        return tensors
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM")
@@ -2458,7 +2436,7 @@ class StableLMModel(TextModel):
             if len(self._q_norms[bid]) >= n_head:
                 return self._stack_qk_norm(bid, n_head, self._q_norms[bid], "q_layernorm")
             else:
-                return []
+                return
 
         if name.find("k_layernorm.norms") != -1:
             assert bid is not None
@@ -2471,9 +2449,9 @@ class StableLMModel(TextModel):
             if len(self._k_norms[bid]) >= n_kv_head:
                 return self._stack_qk_norm(bid, n_kv_head, self._k_norms[bid], "k_layernorm")
             else:
-                return []
+                return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def _stack_qk_norm(self, bid: int, n_head: int, norms: dict[str, Tensor], layer_name: str = "q_layernorm"):
         datas: list[Tensor] = []
@@ -2485,9 +2463,8 @@ class StableLMModel(TextModel):
         data_torch = torch.stack(datas, dim=0)
 
         merged_name = f"model.layers.{bid}.self_attn.{layer_name}.weight"
-        new_name = self.map_tensor_name(merged_name)
 
-        return [(new_name, data_torch)]
+        yield from super().modify_tensors(data_torch, merged_name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -2613,7 +2590,7 @@ class LlamaModel(TextModel):
             )
 
         if is_multimodal_tensor:
-            return [] # skip vision tensors
+            return  # skip vision tensors
         elif self.hf_arch == "LlamaModel":
             name = "model." + name
         elif name.startswith("model.text_model"):
@@ -2639,8 +2616,6 @@ class LlamaModel(TextModel):
             self._experts[bid][name] = data_torch
 
             if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
                 # merge the experts into a single 3d tensor
                 for wid in ["w1", "w2", "w3"]:
                     datas: list[Tensor] = []
@@ -2654,14 +2629,12 @@ class LlamaModel(TextModel):
 
                     merged_name = f"layers.{bid}.feed_forward.experts.{wid}.weight"
 
-                    new_name = self.map_tensor_name(merged_name)
-
-                    tensors.append((new_name, data_torch))
-                return tensors
+                    yield from super().modify_tensors(data_torch, merged_name, bid)
+                return
             else:
-                return []
+                return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         if rope_params := self.rope_parameters.get("full_attention", self.rope_parameters):
@@ -2752,8 +2725,6 @@ class AfmoeModel(LlamaModel):
             self._experts[bid][name] = data_torch
 
             if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
                 # merge the experts into a single 3d tensor
                 for w_name in ["gate_proj", "up_proj", "down_proj"]:
                     datas: list[Tensor] = []
@@ -2765,17 +2736,16 @@ class AfmoeModel(LlamaModel):
 
                     data_torch = torch.stack(datas, dim=0)
                     merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
-                    new_name = self.map_tensor_name(merged_name)
-                    tensors.append((new_name, data_torch))
+                    yield from ModelBase.modify_tensors(self, data_torch, merged_name, bid)
 
-                return tensors
+                return
             else:
-                return []
+                return
 
         if name.endswith(".expert_bias"):
             name = name.replace(".expert_bias", ".expert_bias.bias")
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from ModelBase.modify_tensors(self, data_torch, name, bid)
 
 
 @ModelBase.register(
@@ -2832,7 +2802,6 @@ class LlavaVisionModel(MmprojModel):
                 self.gguf_writer.add_vision_spatial_merge_size(self.global_config["spatial_merge_size"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
         n_head = (
             self.hparams["num_attention_heads"] if not self.is_mistral_format else self.find_vparam(["num_attention_heads"])
         )
@@ -2853,7 +2822,8 @@ class LlavaVisionModel(MmprojModel):
                 data_torch = LlamaModel.permute(data_torch, n_head, n_head)
             if name.endswith(("k_proj.weight", "k_proj.bias")) and not self.is_mistral_format:
                 data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
-            return [(self.map_tensor_name(name), data_torch)]
+            yield from super().modify_tensors(data_torch, name, bid)
+            return
 
         embed_key = "embed_tokens.weight" if not self.is_mistral_format else "tok_embeddings.weight"
         if self.img_break_tok_id > 0 and embed_key in name:
@@ -2861,9 +2831,9 @@ class LlavaVisionModel(MmprojModel):
             # for pixtral model, we need to extract the [IMG_BREAK] token embedding
             img_break_embd = data_torch[self.img_break_tok_id]
             name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK]
-            return [(self.map_tensor_name(name), img_break_embd)]
+            yield from super().modify_tensors(img_break_embd, name, bid)
 
-        return [] # skip other tensors
+        return # skip other tensors
 
 
 @ModelBase.register("Idefics3ForConditionalGeneration", "SmolVLMForConditionalGeneration")
@@ -2894,13 +2864,12 @@ class SmolVLMModel(MmprojModel):
         return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
         is_vision_tensor = "vision_tower" in name or "vision_model" in name or "model.connector" in name
 
         if is_vision_tensor:
-            return [(self.map_tensor_name(name), data_torch)]
+            yield from super().modify_tensors(data_torch, name, bid)
 
-        return [] # skip other tensors
+        return # skip other tensors
 
 
 @ModelBase.register(
@@ -2939,18 +2908,17 @@ class Llama4Model(LlamaModel):
             name_gate = name.replace("gate_up_proj", "gate_proj.weight")
             dim_half = data_torch.shape[-1] // 2
             gate_proj_weight, up_proj_weight = data_torch.transpose(-1, -2).split(dim_half, dim=-2)
-            return [
-                (self.map_tensor_name(name_gate), gate_proj_weight),
-                (self.map_tensor_name(name_up), up_proj_weight)
-            ]
+            yield from super().modify_tensors(gate_proj_weight, name_gate, bid)
+            yield from super().modify_tensors(up_proj_weight, name_up, bid)
+            return
 
         if name.endswith("down_proj"):
             name += ".weight"
             data_torch = data_torch.transpose(-1, -2)
 
         if "multi_modal_projector" in name or "vision_model" in name:
-            return []
-        return super().modify_tensors(data_torch, name, bid)
+            return
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Llama4ForConditionalGeneration")
@@ -2964,19 +2932,21 @@ class Llama4VisionModel(MmprojModel):
         self.gguf_writer.add_vision_use_gelu(True)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid # unused
         if "multi_modal_projector" in name or "vision_model" in name:
             # process vision tensors
             if "positional_embedding_vlm" in name and ".weight" not in name:
                 name += ".weight"
             if "multi_modal_projector.linear_1" in name:
                 # despite the name with number postfix, this is a single fully connected layer
-                return [(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_MMPROJ_FC] + '.weight', data_torch)]
-            return [(self.map_tensor_name(name), data_torch)]
-        return []
+                yield (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_MMPROJ_FC] + '.weight', data_torch)
+            else:
+                yield from super().modify_tensors(data_torch, name, bid)
 
 
-@ModelBase.register("Mistral3ForConditionalGeneration")
+@ModelBase.register(
+    "Mistral3ForConditionalGeneration",
+    "Ministral3ForCausalLM",
+)
 class Mistral3Model(LlamaModel):
     model_arch = gguf.MODEL_ARCH.MISTRAL3
 
@@ -3002,9 +2972,9 @@ class Mistral3Model(LlamaModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
         name = name.replace("language_model.", "")
         if "multi_modal_projector" in name or "vision_tower" in name:
-            return []
+            return
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("DeciLMForCausalLM")
@@ -3143,7 +3113,7 @@ class DeciModel(TextModel):
             data_torch = DeciModel.permute(data_torch, n_head, n_head)
         if name.endswith(("k_proj.weight", "k_proj.bias")):
             data_torch = DeciModel.permute(data_torch, n_head, n_kv_head)
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         if rope_params := self.rope_parameters.get("full_attention", self.rope_parameters):
@@ -3217,7 +3187,7 @@ class BitnetModel(TextModel):
             # transform weight into 1/0/-1 (in fp32)
             data_torch = self.weight_quant(data_torch)
 
-        yield (new_name, data_torch)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("GrokForCausalLM", "Grok1ForCausalLM")
@@ -3273,11 +3243,11 @@ class GrokModel(TextModel):
     _cur_expert = ""
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        tensors: list[tuple[str, Tensor]] = []
+        deferred: list[tuple[Tensor, str, int | None]] = []
         is_expert = ".moe." in name or ".block_sparse_moe.experts." in name
 
         if not is_expert:
-            tensors.append((self.map_tensor_name(name), data_torch))
+            deferred.append((data_torch, name, bid))
 
         # process the experts separately
         if is_expert or self._cur_expert:
@@ -3292,11 +3262,11 @@ class GrokModel(TextModel):
             if name in self._experts[bid]:
                 self._cur_expert = name
                 self._experts[bid][name].append(data_torch)
-                return []
+                return
             elif is_expert:
                 self._cur_expert = name
                 self._experts[bid][name] = [data_torch]
-                return []
+                return
             else:
                 self._cur_expert = ""
 
@@ -3318,11 +3288,10 @@ class GrokModel(TextModel):
 
                         merged_name = f"transformer.decoder_layer.{bid}.moe.{wid[0]}.weight"
 
-                        new_name = self.map_tensor_name(merged_name)
+                        yield from super().modify_tensors(data_torch, merged_name, bid)
 
-                        yield (new_name, data_torch)
-
-        yield from tensors
+        for t in deferred:
+            yield from super().modify_tensors(*t)
 
 
 @ModelBase.register("DbrxForCausalLM")
@@ -3354,8 +3323,6 @@ class DbrxModel(TextModel):
         logger.info(f"gguf: file type = {self.ftype}")
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         n_expert = self.hparams["ffn_config"]["moe_num_experts"]
         n_ff = self.hparams["ffn_config"]["ffn_hidden_size"]
         n_embd = self.hparams["d_model"]
@@ -3386,7 +3353,7 @@ class DbrxModel(TextModel):
         # https://huggingface.co/databricks/dbrx-instruct/blob/main/model.safetensors.index.json#L15
         new_name = self.map_tensor_name(name if not experts else name + ".weight", try_suffixes=(".weight",))
 
-        return [(new_name, data_torch)]
+        yield from super().modify_tensors(data_torch, new_name, bid)
 
     def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
         del name, new_name, bid  # unused
@@ -3431,8 +3398,6 @@ class MiniCPMModel(TextModel):
         self._set_vocab_sentencepiece()
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         n_head = self.hparams["num_attention_heads"]
         n_kv_head = self.hparams.get("num_key_value_heads")
 
@@ -3442,7 +3407,7 @@ class MiniCPMModel(TextModel):
         if name.endswith(("k_proj.weight")):
             data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("MiniCPM3ForCausalLM")
@@ -3552,7 +3517,7 @@ class Qwen2Model(TextModel):
                 or name.startswith("vision_model") or name.startswith("audio_tower") \
                 or name.startswith("model.vision_tower") or name.startswith("model.multi_modal_projector"):
             # skip vision and audio tensors
-            return []
+            return
         yield from super().modify_tensors(data_torch, name, bid)
 
 
@@ -3749,23 +3714,20 @@ class Ernie4_5Model(TextModel):
             total_k_dim = num_kv_heads * head_dim
             total_v_dim = num_kv_heads * head_dim
             q_proj_weight, k_proj_weight, v_proj_weight = data_torch.split([total_q_dim, total_k_dim, total_v_dim], dim=0)
-            return [
-                (self.map_tensor_name(name_q), q_proj_weight),
-                (self.map_tensor_name(name_k), k_proj_weight),
-                (self.map_tensor_name(name_v), v_proj_weight)
-            ]
+            yield from super().modify_tensors(q_proj_weight, name_q, bid)
+            yield from super().modify_tensors(k_proj_weight, name_k, bid)
+            yield from super().modify_tensors(v_proj_weight, name_v, bid)
         # split the up_gate_proj into gate and up
         # up_gate_proj shape: [2 * intermediate_size, hidden_size]
-        if "up_gate_proj" in name:
+        elif "up_gate_proj" in name:
             name_up = name.replace("up_gate_proj.weight", "up_proj.weight")
             name_gate = name.replace("up_gate_proj.weight", "gate_proj.weight")
             dim_half = data_torch.shape[0] // 2
             gate_proj_weight, up_proj_weight = data_torch.split(dim_half, dim=0)
-            return [
-                (self.map_tensor_name(name_gate), gate_proj_weight),
-                (self.map_tensor_name(name_up), up_proj_weight)
-            ]
-        return [(self.map_tensor_name(name), data_torch)]
+            yield from super().modify_tensors(gate_proj_weight, name_gate, bid)
+            yield from super().modify_tensors(up_proj_weight, name_up, bid)
+        else:
+            yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Ernie4_5_MoeForCausalLM")
@@ -3798,20 +3760,20 @@ class Ernie4_5MoeModel(Ernie4_5Model):
         # skip Multi-Token Prediction (MTP) layers (again, same as DeepseekV2)
         match = re.match(r"model.mtp_block.(\d+)", name)
         if match:
-            return []
+            return
 
         # skip all other MTP tensors for now
         match = re.match(r"model.mtp_emb_norm.(\d+)", name)
         if match:
-            return []
+            return
 
         match = re.match(r"model.mtp_hidden_norm.(\d+)", name)
         if match:
-            return []
+            return
 
         match = re.match(r"model.mtp_linear_proj.(\d+)", name)
         if match:
-            return []
+            return
 
         # process the experts separately
         if name.find("mlp.experts") != -1:
@@ -3824,8 +3786,6 @@ class Ernie4_5MoeModel(Ernie4_5Model):
             self._experts[bid][name] = data_torch
 
             if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
                 # merge the experts into a single 3d tensor
                 for w_name in ["gate_proj", "up_proj", "down_proj"]:
                     datas: list[Tensor] = []
@@ -3837,13 +3797,9 @@ class Ernie4_5MoeModel(Ernie4_5Model):
 
                     data_torch = torch.stack(datas, dim=0)
                     merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
-                    new_name = self.map_tensor_name(merged_name)
-                    tensors.append((new_name, data_torch))
-
-                return tensors
-            else:
-                return []
-        return [(self.map_tensor_name(name), data_torch)]
+                    yield from super().modify_tensors(data_torch, merged_name, bid)
+        else:
+            yield from ModelBase.modify_tensors(self, data_torch, name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -3874,14 +3830,13 @@ class Qwen2VLModel(TextModel):
             self._set_vocab_gpt2()
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
         if name.startswith("thinker."):
             name = name.replace("thinker.", "")
         if name.startswith("visual") or name.startswith("audio") or \
                 name.startswith("talker") or name.startswith("token2wav"):
             # skip multimodal tensors
-            return []
-        return [(self.map_tensor_name(name), data_torch)]
+            return
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Qwen2VLModel", "Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration")
@@ -3930,7 +3885,6 @@ class Qwen2VLVisionModel(MmprojModel):
         return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
         if name.startswith("visual."):
             # process visual tensors
             # split QKV tensors if needed
@@ -3944,23 +3898,18 @@ class Qwen2VLVisionModel(MmprojModel):
                 wq = data_torch[:c]
                 wk = data_torch[c: c * 2]
                 wv = data_torch[c * 2:]
-                return [
-                    (self.map_tensor_name(name.replace("qkv", "q")), wq),
-                    (self.map_tensor_name(name.replace("qkv", "k")), wk),
-                    (self.map_tensor_name(name.replace("qkv", "v")), wv),
-                ]
+                yield from super().modify_tensors(wq, name.replace("qkv", "q"), bid)
+                yield from super().modify_tensors(wk, name.replace("qkv", "k"), bid)
+                yield from super().modify_tensors(wv, name.replace("qkv", "v"), bid)
             elif 'patch_embed.proj.weight' in name:
                 # split Conv3D into Conv2Ds
                 c1, c2, kt, kh, kw = data_torch.shape
                 del c1, c2, kh, kw  # unused
                 assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
-                return [
-                    (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight"  , data_torch[:, :, 0, ...]),
-                    (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight.1", data_torch[:, :, 1, ...]),
-                ]
+                yield (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight"  , data_torch[:, :, 0, ...])
+                yield (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight.1", data_torch[:, :, 1, ...])
             else:
-                return [(self.map_tensor_name(name), data_torch)]
-        return [] # skip other tensors
+                yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Qwen2_5OmniModel")
@@ -4016,10 +3965,8 @@ class Qwen25OmniModel(Qwen2VLVisionModel):
             if "audio_bos_eos_token" in name:
                 # this tensor is left unused in transformers code
                 # https://github.com/huggingface/transformers/blob/6e3063422c4b1c014aa60c32b9254fd2902f0f28/src/transformers/models/qwen2_5_omni/modular_qwen2_5_omni.py#L1809
-                return []
-            return [(self.map_tensor_name(name), data_torch)]
-
-        return super().modify_tensors(data_torch, name, bid)
+                return
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("InternVisionModel")
@@ -4066,7 +4013,6 @@ class InternVisionModel(MmprojModel):
         return name
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
         vision_prefix = ['vision_model', 'mlp', 'model.vision_tower', 'model.multi_modal_projector']
         # deal with intern-s1 special case
         name = self._mapping_interns1_name(name)
@@ -4088,13 +4034,11 @@ class InternVisionModel(MmprojModel):
                 wq = data_torch[:c]
                 wk = data_torch[c: c * 2]
                 wv = data_torch[c * 2:]
-                return [
-                    (self.map_tensor_name(name.replace("attn.qkv", "self_attn.q_proj")), wq),
-                    (self.map_tensor_name(name.replace("attn.qkv", "self_attn.k_proj")), wk),
-                    (self.map_tensor_name(name.replace("attn.qkv", "self_attn.v_proj")), wv),
-                ]
-            return [(self.map_tensor_name(name), data_torch)]
-        return [] # skip other tensors
+                yield from super().modify_tensors(wq, name.replace("attn.qkv", "self_attn.q_proj"), bid)
+                yield from super().modify_tensors(wk, name.replace("attn.qkv", "self_attn.k_proj"), bid)
+                yield from super().modify_tensors(wv, name.replace("attn.qkv", "self_attn.v_proj"), bid)
+            else:
+                yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("WavTokenizerDec")
@@ -4102,18 +4046,16 @@ class WavTokenizerDecModel(TextModel):
     model_arch = gguf.MODEL_ARCH.WAVTOKENIZER_DEC
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         if \
                 name.endswith("codebook.cluster_size") or \
                 name.endswith("codebook.embed_avg") or \
                 name.endswith("codebook.inited"):
             logger.debug(f"Skipping {name!r}")
-            return []
+            return
 
         logger.info(f"{self.map_tensor_name(name)} -> {data_torch.shape}")
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def set_vocab(self):
         self._set_vocab_none()
@@ -4168,7 +4110,8 @@ class Qwen2MoeModel(TextModel):
             # Need PyTorch: (128, 2048, 768) [reversed of GGML]
             # So: permute(0, 2, 1): (128, 768, 2048) -> (128, 2048, 768)
             permuted = data_torch.permute(0, 2, 1).contiguous()
-            return [(self.map_tensor_name(mapped), permuted)]
+            yield from super().modify_tensors(permuted, mapped, bid)
+            return
 
         if name.endswith("mlp.experts.gate_up_proj") or name.endswith("mlp.experts.gate_up_proj.weight"):
             if data_torch.ndim < 3 or data_torch.shape[-1] % 2 != 0:
@@ -4186,14 +4129,13 @@ class Qwen2MoeModel(TextModel):
             mapped_up = f"{base}.up_proj.weight"
             perm_gate = gate.permute(0, 2, 1).contiguous()
             perm_up = up.permute(0, 2, 1).contiguous()
-            return [
-                (self.map_tensor_name(mapped_gate), perm_gate),
-                (self.map_tensor_name(mapped_up), perm_up),
-            ]
+            yield from super().modify_tensors(perm_gate, mapped_gate, bid)
+            yield from super().modify_tensors(perm_up, mapped_up, bid)
+            return
 
         if name.startswith("mlp") or name.startswith("vision_model") or name.startswith("model.vision_tower") or name.startswith("model.multi_modal_projector") or name.startswith("model.visual"):
             # skip visual tensors
-            return []
+            return
         if name.find("experts") != -1:
             n_experts = self.hparams["num_experts"]
             assert bid is not None
@@ -4204,8 +4146,6 @@ class Qwen2MoeModel(TextModel):
             self._experts[bid][name] = data_torch
 
             if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
                 # merge the experts into a single 3d tensor
                 for w_name in ["down_proj", "gate_proj", "up_proj"]:
                     datas: list[Tensor] = []
@@ -4219,14 +4159,12 @@ class Qwen2MoeModel(TextModel):
 
                     merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
 
-                    new_name = self.map_tensor_name(merged_name)
-
-                    tensors.append((new_name, data_torch))
-                return tensors
+                    yield from super().modify_tensors(data_torch, merged_name, bid)
+                return
             else:
-                return []
+                return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -4306,7 +4244,7 @@ class Qwen3Model(Qwen2Model):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if "model.vision_" in name:
             # skip multimodal tensors
-            return []
+            return
 
         if self.is_rerank:
             is_tied_head = self.is_tied_embeddings and "embed_tokens" in name
@@ -4316,13 +4254,12 @@ class Qwen3Model(Qwen2Model):
                     gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.CLS_OUT] + ".weight",
                     self._get_cls_out_tensor(data_torch),
                 )
+                yield cls_out_head
                 if is_tied_head:
-                    embed = (self.map_tensor_name(name), data_torch)
-                    return [cls_out_head, embed]
-                if is_real_head:
-                    return [cls_out_head]
+                    yield from super().modify_tensors(data_torch, name, bid)
+                return
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Qwen3MoeForCausalLM")
@@ -4360,7 +4297,7 @@ class Qwen3NextModel(Qwen2MoeModel):
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if name.startswith("mtp"):
-            return [] # ignore MTP layers for now
+            return  # ignore MTP layers for now
         if name.endswith(".A_log"):
             data_torch = -torch.exp(data_torch)
         elif name.endswith(".dt_bias"):
@@ -4462,7 +4399,7 @@ class Qwen3VLVisionModel(MmprojModel):
         assert self.hparams_vision is not None
         # Skip text model tensors - they go in the text model file
         if name.startswith("model.language_model.") or name.startswith("lm_head."):
-            return []
+            return
 
         if name.startswith("model.visual."):
             name = name.replace("model.visual.", "visual.", 1)
@@ -4487,7 +4424,8 @@ class Qwen3VLVisionModel(MmprojModel):
                 raise ValueError(f"Unexpected deepstack tensor: {name}")
 
             new_name = self.format_tensor_name(tensor_type, idx, suffix=f".{suffix}")
-            return [(new_name, data_torch)]
+            yield from super().modify_tensors(data_torch, new_name, bid)
+            return
 
         if name.startswith("visual.merger."):
             suffix = name.split(".", 2)[2]
@@ -4507,7 +4445,8 @@ class Qwen3VLVisionModel(MmprojModel):
                 new_name = self.format_tensor_name(gguf.MODEL_TENSOR.V_POST_NORM, suffix=f".{suffix.split('.', 1)[1]}")
             else:
                 raise ValueError(f"Unexpected merger tensor: {name}")
-            return [(new_name, data_torch)]
+            yield (new_name, data_torch)
+            return
 
         if name == "visual.patch_embed.proj.weight":
             # split Conv3D into Conv2Ds along temporal dimension
@@ -4515,20 +4454,21 @@ class Qwen3VLVisionModel(MmprojModel):
             del c1, c2
             if kt != 2:
                 raise ValueError("Current implementation only supports temporal_patch_size of 2")
-            return [
-                (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight", data_torch[:, :, 0, ...]),
-                (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight.1", data_torch[:, :, 1, ...]),
-            ]
+            yield (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight", data_torch[:, :, 0, ...])
+            yield (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight.1", data_torch[:, :, 1, ...])
+            return
 
         if name == "visual.patch_embed.proj.bias":
             # Include the bias - it's used by the C++ code
-            return [(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".bias", data_torch)]
+            yield (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".bias", data_torch)
+            return
 
         if name.startswith("visual."):
-            return [(self.map_tensor_name(name), data_torch)]
+            yield from super().modify_tensors(data_torch, name, bid)
+            return
 
         # Fall back to parent class for other tensors
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Glm4vForConditionalGeneration", "Glm4vMoeForConditionalGeneration")
@@ -4551,8 +4491,9 @@ class Glm4VVisionModel(Qwen3VLVisionModel):
         if name.startswith("model.visual."):
             name = name.replace("model.visual.", "visual.")
         if name.startswith("visual.merger."):
-            return [(self.map_tensor_name(name), data_torch)]
-        return super().modify_tensors(data_torch, name, bid)
+            yield from ModelBase.modify_tensors(self, data_torch, name, bid)
+            return
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Qwen3VLForConditionalGeneration")
@@ -4570,9 +4511,9 @@ class Qwen3VLTextModel(Qwen3Model):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # Skip vision tensors - they go in the mmproj file
         if name.startswith("model.visual."):
-            return []
+            return
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Qwen3VLMoeForConditionalGeneration")
@@ -4588,9 +4529,9 @@ class Qwen3VLMoeTextModel(Qwen3MoeModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # Skip vision tensors - they go in the mmproj file
         if name.startswith("model.visual."):
-            return []
+            return
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("GPT2LMHeadModel")
@@ -4607,22 +4548,17 @@ class GPT2Model(TextModel):
         self.gguf_writer.add_file_type(self.ftype)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
-        tensors: list[tuple[str, Tensor]] = []
-
         # we don't need these
         if name.endswith((".attn.bias", ".attn.masked_bias")):
-            return tensors
+            yield from super().modify_tensors(data_torch, name, bid)
+            return
 
         if name.endswith((".c_attn.weight", ".c_proj.weight", ".c_fc.weight", ".c_proj.weight")):
             data_torch = data_torch.transpose(1, 0)
 
         new_name = self.map_tensor_name(name)
 
-        tensors.append((new_name, data_torch))
-
-        return tensors
+        yield from super().modify_tensors(data_torch, new_name, bid)
 
 
 @ModelBase.register("PhiForCausalLM")
@@ -4846,8 +4782,6 @@ class PhiMoeModel(Phi3MiniModel):
             self._experts[bid][name] = data_torch
 
             if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
                 # merge the experts into a single 3d tensor
                 for w_name in ["w1", "w2", "w3"]:
                     datas: list[Tensor] = []
@@ -4861,14 +4795,12 @@ class PhiMoeModel(Phi3MiniModel):
 
                     merged_name = f"model.layers.{bid}.block_sparse_moe.experts.{w_name}.weight"
 
-                    new_name = self.map_tensor_name(merged_name)
-
-                    tensors.append((new_name, data_torch))
-                return tensors
+                    yield from super().modify_tensors(data_torch, merged_name, bid)
+                return
             else:
-                return []
+                return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -4914,8 +4846,6 @@ class PlamoModel(TextModel):
         return data_torch
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         new_name = self.map_tensor_name(name)
 
         # shuffle for broadcasting of gqa in ggml_mul_mat
@@ -4924,7 +4854,7 @@ class PlamoModel(TextModel):
         elif new_name.endswith("attn_output.weight"):
             data_torch = self.shuffle_attn_output_weight(data_torch)
 
-        return [(new_name, data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Plamo2ForCausalLM", "PLaMo2ForCausalLM")
@@ -4985,8 +4915,6 @@ class Plamo2Model(TextModel):
         self.gguf_writer.add_file_type(self.ftype)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         if name.endswith(".A_log"):
             data_torch = -torch.exp(data_torch)
         elif name.endswith(".dt_bias"):
@@ -5015,9 +4943,7 @@ class Plamo2Model(TextModel):
         elif name.endswith(".norm.weight"):
             data_torch += 1.0
 
-        new_name = self.map_tensor_name(name)
-
-        return [(new_name, data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Plamo3ForCausalLM", "PLaMo3ForCausalLM")
@@ -5066,7 +4992,7 @@ class Plamo3Model(TextModel):
         elif name.endswith(".norm.weight"):
             data_torch = data_torch + 1.0
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("CodeShellForCausalLM")
@@ -5228,7 +5154,7 @@ class InternLM2Model(TextModel):
         name = name.replace("language_model.", "") # InternVL
         if name.startswith("mlp") or name.startswith("vision_model"):
             # skip visual tensors
-            return []
+            return
 
         if bid is not None and f"model.layers.{bid}.attention.wqkv" in name:
             qkv = data_torch
@@ -5241,13 +5167,11 @@ class InternLM2Model(TextModel):
             k = LlamaModel.permute(k.reshape((-1, k.shape[-1])), num_heads, num_kv_heads)
             v = v.reshape((-1, v.shape[-1]))
 
-            return [
-                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), q),
-                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), k),
-                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), v),
-            ]
+            yield from super().modify_tensors(q, self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), bid)
+            yield from super().modify_tensors(k, self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), bid)
+            yield from super().modify_tensors(v, self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), bid)
         else:
-            return [(self.map_tensor_name(name), data_torch)]
+            yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("InternLM3ForCausalLM")
@@ -5299,12 +5223,12 @@ class InternLM3Model(TextModel):
         name = name.replace("language_model.", "") # InternVL
         if name.startswith("mlp") or name.startswith("vision_model"):
             # skip visual tensors
-            return []
+            return
         if name.endswith(("q_proj.weight", "q_proj.bias")):
             data_torch = LlamaModel.permute(data_torch, n_head, n_head)
         if name.endswith(("k_proj.weight", "k_proj.bias")):
             data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("BertModel", "BertForMaskedLM", "CamembertModel", "BertForSequenceClassification")
@@ -5359,8 +5283,6 @@ class BertModel(TextModel):
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         if name.startswith("bert."):
             name = name[5:]
 
@@ -5372,13 +5294,13 @@ class BertModel(TextModel):
 
         # we are only using BERT for embeddings so we don't need the pooling layer
         if name in ("embeddings.position_ids", "pooler.dense.weight", "pooler.dense.bias"):
-            return [] # we don't need these
+            return # we don't need these
 
         if name.startswith("cls.predictions"):
-            return []
+            return
 
         if name.startswith("cls.seq_relationship"):
-            return []
+            return
 
         if self.cls_out_labels:
             # For BertForSequenceClassification (direct projection layer)
@@ -5388,7 +5310,7 @@ class BertModel(TextModel):
             if name == "classifier.bias":
                 name = "classifier.out_proj.bias"
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def _xlmroberta_tokenizer_init(self) -> None:
         # we need the pad_token_id to know how to chop down position_embd matrix
@@ -5543,9 +5465,9 @@ class DistilBertModel(BertModel):
 
         # These layers act as MLM head, so we don't need them
         if name.startswith("vocab_"):
-            return []
+            return
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("RobertaModel", "RobertaForSequenceClassification")
@@ -5588,7 +5510,7 @@ class RobertaModel(BertModel):
             if self._position_offset is not None:
                 data_torch = data_torch[self._position_offset:,:]
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("NomicBertModel")
@@ -5641,7 +5563,7 @@ class NomicBertModel(BertModel):
     def modify_tensors(self, data_torch: torch.Tensor, name: str, bid: int | None) -> Iterable[tuple[str, torch.Tensor]]:
         # If the tensor is an experts bias tensor, skip it by returning an empty list.
         if "mlp.experts.bias" in name:
-            return []  # Explicitly return an empty list.
+            return # Explicitly return.
 
         if "mlp.experts.mlp.w1" in name:
             data_torch = data_torch.view(self.hparams["num_experts"], self.hparams["n_inner"], self.hparams["n_embd"])
@@ -5652,7 +5574,7 @@ class NomicBertModel(BertModel):
             data_torch = data_torch.transpose(1, 2)
             name += ".weight"
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
@@ -5692,12 +5614,12 @@ class NeoBert(BertModel):
 
     def modify_tensors(self, data_torch, name, bid):
         if name.startswith("decoder."):
-            return []
+            return
 
         if name.startswith("model."):
             name = name[6:]
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
@@ -5754,7 +5676,7 @@ class XLMRobertaModel(BertModel):
 
         if name.endswith(".0.lora_A") or name.endswith(".0.lora_B"):
             if name.startswith("pooler.dense"):
-                return []
+                return
 
             num_loras = data_torch.size(0)
             assert num_loras == len(self._lora_names)
@@ -5770,9 +5692,9 @@ class XLMRobertaModel(BertModel):
                     new_name = new_name[:-1] + ("a" if new_name[-1:] == "b" else "b")
                 lora_writer.add_tensor(new_name, data.float().numpy(), raw_dtype=gguf.GGMLQuantizationType.F32)
 
-            return []
+            return
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
@@ -5831,19 +5753,17 @@ class GemmaModel(TextModel):
         self.gguf_writer.add_file_type(self.ftype)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         # lm_head is not used in llama.cpp, while autoawq will include this tensor in model
         # To prevent errors, skip loading lm_head.weight.
         if name == "lm_head.weight":
             logger.debug(f"Skipping get tensor {name!r} in safetensors so that convert can end normally.")
-            return []
+            return
 
         # ref: https://github.com/huggingface/transformers/blob/fc37f38915372c15992b540dfcbbe00a916d4fc6/src/transformers/models/gemma/modeling_gemma.py#L89
         if name.endswith("norm.weight"):
             data_torch = data_torch + 1
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Gemma2ForCausalLM")
@@ -5877,19 +5797,17 @@ class Gemma2Model(TextModel):
         self.gguf_writer.add_sliding_window(self.hparams["sliding_window"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         # lm_head is not used in llama.cpp, while autoawq will include this tensor in model
         # To prevent errors, skip loading lm_head.weight.
         if name == "lm_head.weight":
             logger.debug(f"Skipping get tensor {name!r} in safetensors so that convert can end normally.")
-            return []
+            return
 
         # ref: https://github.com/huggingface/transformers/blob/fc37f38915372c15992b540dfcbbe00a916d4fc6/src/transformers/models/gemma/modeling_gemma.py#L89
         if name.endswith("norm.weight"):
             data_torch = data_torch + 1
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Gemma3ForCausalLM", "Gemma3ForConditionalGeneration")
@@ -5924,14 +5842,12 @@ class Gemma3Model(TextModel):
         self.gguf_writer.add_head_count_kv(hparams.get("num_key_value_heads", 4))
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         if "language_model." in name:
             name = name.replace("language_model.", "")
 
         elif name.startswith("multi_modal_projector.") or name.startswith("vision_tower.") \
                 or name.startswith("multimodal_projector.") or name.startswith("vision_model."):
-            return [] # skip vision tensors
+            return # skip vision tensors
 
         # remove OOV (out-of-vocabulary) rows in token_embd
         if "embed_tokens.weight" in name:
@@ -5947,7 +5863,7 @@ class Gemma3Model(TextModel):
         if name.endswith("norm.weight"):
             data_torch = data_torch + self.norm_shift
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Gemma3TextModel")
@@ -6053,10 +5969,8 @@ class Gemma3VisionModel(MmprojModel):
         return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         if "vision_model.head." in name:
-            return [] # skip redundant tensors for tinygemma3
+            return # skip redundant tensors for tinygemma3
 
         if name.startswith("multi_modal_projector.") or name.startswith("vision_tower.") \
                 or name.startswith("multimodal_projector.") or name.startswith("vision_model."):
@@ -6070,9 +5984,9 @@ class Gemma3VisionModel(MmprojModel):
                 logger.info(f"Correcting norm value for '{name}'")
                 data_torch = data_torch + 1
 
-            return [(self.map_tensor_name(name), data_torch)]
+            yield from super().modify_tensors(data_torch, name, bid)
 
-        return [] # skip other tensors
+        return # skip other tensors
 
 
 class ConformerAudioModel(MmprojModel):
@@ -6097,7 +6011,7 @@ class ConformerAudioModel(MmprojModel):
             self._batch_norm_tensors[bid][name] = data_torch
 
             if len(self._batch_norm_tensors[bid]) < 5:
-                return []
+                return
 
             weight = self._batch_norm_tensors[bid][f"conformer.layers.{bid}.conv.batch_norm.weight"]
             bias = self._batch_norm_tensors[bid][f"conformer.layers.{bid}.conv.batch_norm.bias"]
@@ -6107,10 +6021,9 @@ class ConformerAudioModel(MmprojModel):
 
             a = weight / torch.sqrt(running_var + eps)
             b = bias - running_mean * a
-            return [
-                (self.map_tensor_name(f"conformer.layers.{bid}.conv.batch_norm.weight"), a),
-                (self.map_tensor_name(f"conformer.layers.{bid}.conv.batch_norm.bias"), b),
-            ]
+            yield from super().modify_tensors(a, f"conformer.layers.{bid}.conv.batch_norm.weight", bid)
+            yield from super().modify_tensors(b, f"conformer.layers.{bid}.conv.batch_norm.bias", bid)
+            return
 
         # reshape conv weights
         if name.startswith("conformer.pre_encode.conv.") and name.endswith(".bias"):
@@ -6122,7 +6035,7 @@ class ConformerAudioModel(MmprojModel):
             assert data_torch.shape[2] == 1
             data_torch = data_torch.reshape(data_torch.shape[0], data_torch.shape[1])
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Gemma3nForConditionalGeneration")
@@ -6221,18 +6134,19 @@ class Gemma3nVisionAudioModel(ConformerAudioModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if (ConformerAudioModel.is_audio_tensor(name)):
             name = name.replace("model.audio_tower.conformer.", "conformer.layers.")
-            return super().modify_tensors(data_torch, name, bid)
+            yield from super().modify_tensors(data_torch, name, bid)
 
         # Gemma3n uses
         # - model.embed_vision.* for projection layers
         # - model.vision_tower.* for vision encoder
         # Skip non-vision tensors
         if not (name.startswith("model.embed_vision.") or name.startswith("model.vision_tower.")):
-            return []
+            return
 
         if name.startswith("model.vision_tower.timm_model.blocks."):
             # Double-indexed block tensors through custom logic
-            new_name = self.custom_map(name)
+            yield (self.custom_map(name), data_torch)
+            return
         else:
             # Route non-repeating (conv_stem, msfa, embedding, etc.) and un-catched through tensor_mapping.py
             new_name = self.map_tensor_name(name)
@@ -6240,7 +6154,7 @@ class Gemma3nVisionAudioModel(ConformerAudioModel):
         if new_name.endswith("conv_stem.conv.bias") or new_name.endswith("layer_scale.gamma"):
             data_torch = data_torch.unsqueeze(0).unsqueeze(-1).unsqueeze(-1) # [1, C, 1, 1]
 
-        return [(new_name, data_torch)]
+        yield from ModelBase.modify_tensors(self, data_torch, new_name, bid)
 
 
 @ModelBase.register("Gemma3nForCausalLM", "Gemma3nForConditionalGeneration")
@@ -6318,7 +6232,7 @@ class Gemma3NModel(Gemma3Model):
         # TODO: implement self.prediction_coefs.weight.clamp_(...)
 
         if "language_model." not in name:
-            return [] # skip non-language model tensors
+            return # skip non-language model tensors
 
         # Pad token embeddings for vision/audio special tokens (262144-262399)
         if "embed_tokens.weight" in name or "embed_tokens_per_layer" in name:
@@ -6340,7 +6254,8 @@ class Gemma3NModel(Gemma3Model):
 
             # Continue with normal processing
             name = name.replace("language_model.", "")
-            return [(self.map_tensor_name(name), data_torch)]
+            yield from ModelBase.modify_tensors(self, data_torch, name, bid)
+            return
 
         if "altup_unembed_projections" in name:
             data_torch = data_torch.to(device="cpu")
@@ -6356,9 +6271,10 @@ class Gemma3NModel(Gemma3Model):
                 raise ValueError(f"Unknown name: {name}")
             out = self._stack_matrices(self._altup_unembd)
             if out is not None:
-                return [(self.map_tensor_name("model.altup_unembed_projections.weight"), out)]
+                yield from ModelBase.modify_tensors(self, out, "model.altup_unembed_projections.weight", bid)
+                return
             else:
-                return []
+                return
 
         if "altup_projections" in name:
             data_torch = data_torch.to(device="cpu")
@@ -6372,11 +6288,12 @@ class Gemma3NModel(Gemma3Model):
                 raise ValueError(f"Unknown name: {name}")
             out = self._stack_matrices(self._altup_proj)
             if out is not None:
-                return [(self.map_tensor_name("model.altup_projections.weight"), out)]
+                yield from ModelBase.modify_tensors(self, out, "model.altup_projections.weight", bid)
+                return
             else:
-                return []
+                return
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Starcoder2ForCausalLM")
@@ -6759,11 +6676,11 @@ class MambaModel(TextModel):
         if self._tok_embd is not None and new_name == output_name:
             if torch.equal(self._tok_embd, data_torch):
                 logger.debug(f"{output_name} is equivalent to {tok_embd_name}, omitting")
-                return []
+                return
         elif new_name == tok_embd_name:
             self._tok_embd = data_torch
 
-        return [(new_name, data_torch)]
+        yield from super().modify_tensors(data_torch, new_name, bid)
 
 
 @ModelBase.register("Mamba2ForCausalLM")
@@ -7019,8 +6936,6 @@ class OlmoModel(TextModel):
     # Same as super class, but permuting q_proj, k_proj
     # Copied from: LlamaModel
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         n_head = self.hparams["num_attention_heads"]
         n_kv_head = self.hparams.get("num_key_value_heads")
 
@@ -7029,7 +6944,7 @@ class OlmoModel(TextModel):
         if name.endswith("k_proj.weight"):
             data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("SeedOssForCausalLM")
@@ -7085,8 +7000,6 @@ class OlmoeModel(TextModel):
             self._experts[bid][name] = data_torch
 
             if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
                 # merge the experts into a single 3d tensor
                 for w_name in ["down_proj", "gate_proj", "up_proj"]:
                     datas: list[Tensor] = []
@@ -7100,14 +7013,12 @@ class OlmoeModel(TextModel):
 
                     merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
 
-                    new_name = self.map_tensor_name(merged_name)
-
-                    tensors.append((new_name, data_torch))
-                return tensors
+                    yield from super().modify_tensors(data_torch, merged_name, bid)
+                return
             else:
-                return []
+                return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     # Copied from: Qwen2MoeModel
     def prepare_tensors(self):
@@ -7330,8 +7241,6 @@ class ArcticModel(TextModel):
             self._experts[bid][name] = data_torch
 
             if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
                 # merge the experts into a single 3d tensor
                 for wid in ["w1", "w2", "w3"]:
                     datas: list[Tensor] = []
@@ -7345,14 +7254,12 @@ class ArcticModel(TextModel):
 
                     merged_name = f"layers.{bid}.feed_forward.experts.{wid}.weight"
 
-                    new_name = self.map_tensor_name(merged_name)
-
-                    tensors.append((new_name, data_torch))
-                return tensors
+                    yield from super().modify_tensors(data_torch, merged_name, bid)
+                return
             else:
-                return []
+                return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -7419,8 +7326,6 @@ class DeepseekModel(TextModel):
             self._experts[bid][name] = data_torch
 
             if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
                 # merge the experts into a single 3d tensor
                 for w_name in ["down_proj", "gate_proj", "up_proj"]:
                     datas: list[Tensor] = []
@@ -7434,14 +7339,12 @@ class DeepseekModel(TextModel):
 
                     merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
 
-                    new_name = self.map_tensor_name(merged_name)
-
-                    tensors.append((new_name, data_torch))
-                return tensors
+                    yield from super().modify_tensors(data_torch, merged_name, bid)
+                return
             else:
-                return []
+                return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -7458,7 +7361,7 @@ class DeepseekModel(TextModel):
     "DeepseekV3ForCausalLM",
     "KimiVLForConditionalGeneration",
     "YoutuForCausalLM",
-    "YoutuVLForConditionalGeneration"
+    "YoutuVLForConditionalGeneration",
 )
 class DeepseekV2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
@@ -7577,9 +7480,9 @@ class DeepseekV2Model(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # skip vision tensors and remove "language_model." for Kimi-VL
         if "vision_tower" in name or "multi_modal_projector" in name:
-            return []
+            return
         if name.startswith("siglip2.") or name.startswith("merger."):
-            return []
+            return
         if name.startswith("language_model."):
             name = name.replace("language_model.", "")
 
@@ -7587,7 +7490,7 @@ class DeepseekV2Model(TextModel):
         if self.hparams.get("tie_word_embeddings", False):
             if name == "lm_head.weight" or name == "model.lm_head.weight":
                 logger.info("Skipping tied output layer 'lm_head.weight' (will use token_embd.weight)")
-                return []
+                return
 
         # rename e_score_correction_bias tensors
         if name.endswith("e_score_correction_bias"):
@@ -7597,7 +7500,7 @@ class DeepseekV2Model(TextModel):
         block_count = self.hparams["num_hidden_layers"]
         match = re.match(r"model.layers.(\d+)", name)
         if match and int(match.group(1)) >= block_count:
-            return []
+            return
 
         # process the experts separately
         if name.find("mlp.experts") != -1:
@@ -7610,8 +7513,6 @@ class DeepseekV2Model(TextModel):
             self._experts[bid][name] = data_torch
 
             if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
                 # merge the experts into a single 3d tensor
                 for w_name in ["down_proj", "gate_proj", "up_proj"]:
                     datas: list[Tensor] = []
@@ -7625,12 +7526,10 @@ class DeepseekV2Model(TextModel):
 
                     merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
 
-                    new_name = self.map_tensor_name(merged_name)
-
-                    tensors.append((new_name, data_torch))
-                return tensors
+                    yield from super().modify_tensors(data_torch, merged_name, bid)
+                return
             else:
-                return []
+                return
 
         # note: MLA with the absorption optimization, needs these two split and k_b_proj transposed
         if name.endswith("kv_b_proj.weight"):
@@ -7647,12 +7546,11 @@ class DeepseekV2Model(TextModel):
             k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1)
             k_b = k_b.transpose(1, 2)
 
-            return [
-                (self.map_tensor_name(name_kb), k_b),
-                (self.map_tensor_name(name_vb), v_b)
-            ]
+            yield from super().modify_tensors(k_b, name_kb, bid)
+            yield from super().modify_tensors(v_b, name_vb, bid)
+            return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -7694,9 +7592,8 @@ class MiniMaxM2Model(TextModel):
 
             # not enough expert weights to merge
             if len(expert_cache) < n_experts * len(expert_weights):
-                return []
+                return
 
-            tensors: list[tuple[str, Tensor]] = []
             for w_name in expert_weights:
                 datas: list[Tensor] = []
 
@@ -7708,12 +7605,12 @@ class MiniMaxM2Model(TextModel):
                 data_torch = torch.stack(datas, dim=0)
                 merged_name = f"model.layers.{bid}.block_sparse_moe.experts.{w_name}.weight"
                 new_name = self.map_tensor_name(merged_name)
-                tensors.append((new_name, data_torch))
+                yield from super().modify_tensors(data_torch, new_name, bid)
 
             del self._experts_cache[bid]
-            return tensors
+            return
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("MiMoV2FlashForCausalLM")
@@ -7755,7 +7652,7 @@ class MimoV2Model(TextModel):
 
         # TODO: mimo v2 does not indicate the number of next-token-prediction layers, therefore we cannot do the same way as GLM4_MOE
         if "model.mtp." in name:
-            return []
+            return
 
         # process the experts separately
         if name.find("mlp.experts") != -1:
@@ -7768,8 +7665,6 @@ class MimoV2Model(TextModel):
             self._experts[bid][name] = data_torch
 
             if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
                 # merge the experts into a single 3d tensor
                 for w_name in ["gate_proj", "up_proj", "down_proj"]:
                     datas: list[Tensor] = []
@@ -7781,13 +7676,12 @@ class MimoV2Model(TextModel):
 
                     data_torch = torch.stack(datas, dim=0)
                     merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
-                    new_name = self.map_tensor_name(merged_name)
-                    tensors.append((new_name, data_torch))
 
-                return tensors
+                    yield from super().modify_tensors(data_torch, merged_name, bid)
+                return
             else:
-                return []
-        return [(self.map_tensor_name(name), data_torch)]
+                return
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -7831,8 +7725,8 @@ class PanguEmbeddedModel(TextModel):
         if name == "lm_head.weight":
             if self.hparams.get("tie_word_embeddings", False):
                 logger.info("Skipping tied output layer 'lm_head.weight'")
-                return []
-        return [(self.map_tensor_name(name), data_torch)]
+                return
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Dots1ForCausalLM")
@@ -7854,8 +7748,9 @@ class Dots1Model(Qwen2MoeModel):
         if name.endswith("e_score_correction_bias"):
             name = name.replace("e_score_correction_bias", "e_score_correction.bias")
         if "shared_experts" in name:
-            return [(self.map_tensor_name(name), data_torch)]
-        return super().modify_tensors(data_torch, name, bid)
+            yield from ModelBase.modify_tensors(self, data_torch, name, bid)
+        else:
+            yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("PLMForCausalLM")
@@ -7874,9 +7769,6 @@ class PLMModel(TextModel):
         self.gguf_writer.add_value_length(hparams["v_head_dim"])
         self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
 
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        return [(self.map_tensor_name(name), data_torch)]
-
     def prepare_tensors(self):
         super().prepare_tensors()
 
@@ -8007,8 +7899,6 @@ class T5Model(TextModel):
         self.gguf_writer.add_file_type(self.ftype)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         # T5 based models contain shared token embeddings tensors saved randomly as either "encoder.embed_tokens.weight",
         # "decoder.embed_tokens.weight" or "shared.weight" tensor. In some models there are even multiple of them stored
         # in the safetensors files. We use the first tensor from these three as the token embeddings for both encoder
@@ -8019,9 +7909,9 @@ class T5Model(TextModel):
                 self.shared_token_embeddings_found = True
             else:
                 logger.debug(f"Skipping shared tensor {name!r} in safetensors so that convert can end normally.")
-                return []
+                return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("T5EncoderModel")
@@ -8143,8 +8033,6 @@ class T5EncoderModel(TextModel):
         self.gguf_writer.add_file_type(self.ftype)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         # T5 based models contain shared token embeddings tensors saved randomly as either "encoder.embed_tokens.weight",
         # "decoder.embed_tokens.weight" or "shared.weight" tensor. In some models there are even multiple of them stored
         # in the safetensors files. We use the first tensor from these three as the token embeddings for both encoder
@@ -8155,9 +8043,9 @@ class T5EncoderModel(TextModel):
                 self.shared_token_embeddings_found = True
             else:
                 logger.debug(f"Skipping shared tensor {name!r} in safetensors so that convert can end normally.")
-                return []
+                return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("JAISLMHeadModel")
@@ -8205,13 +8093,9 @@ class JaisModel(TextModel):
         self.gguf_writer.add_file_type(self.ftype)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
-        tensors: list[tuple[str, Tensor]] = []
-
         # we don't need these
         if name.endswith((".attn.bias")):
-            return tensors
+            return
 
         if name.endswith(("relative_pe.slopes")):
             # Calculate max ALiBi bias (this is the inverse of the ALiBi calculation)
@@ -8222,7 +8106,7 @@ class JaisModel(TextModel):
             first_val = float(data_torch[0].item())
             self.max_alibi_bias = -round(math.log2(first_val) * n_head_closest_log2)
 
-            return tensors
+            return
 
         if name.endswith((".c_attn.weight", ".c_proj.weight", ".c_fc.weight", ".c_fc2.weight")):
             data_torch = data_torch.transpose(1, 0)
@@ -8230,13 +8114,11 @@ class JaisModel(TextModel):
         new_name = self.map_tensor_name(name)
 
         if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
-            tensors.append((new_name, data_torch * self.embeddings_scale))
+            yield from super().modify_tensors(data_torch * self.embeddings_scale, new_name, bid)
         elif new_name == self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT):
-            tensors.append((new_name, data_torch * self.width_scale))
+            yield from super().modify_tensors(data_torch * self.width_scale, new_name, bid)
         else:
-            tensors.append((new_name, data_torch))
-
-        return tensors
+            yield from super().modify_tensors(data_torch, new_name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -8302,7 +8184,7 @@ class Glm4Model(TextModel):
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if name.startswith("model.visual."): # ignore visual part of Glm4v
-            return []
+            return
         elif name.startswith("model.language_model."):
             name = name.replace("language_model.", "") # for Glm4v
         if self.use_mrope:
@@ -8315,7 +8197,7 @@ class Glm4Model(TextModel):
                 data_torch = Glm4Model.normal_to_neox(data_torch, n_head, n_head, head_dim, self.partial_rotary_factor)
             if name.endswith(("k_proj.weight", "k_proj.bias")):
                 data_torch = Glm4Model.normal_to_neox(data_torch, n_head, n_kv_head, head_dim, self.partial_rotary_factor)
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Glm4MoeForCausalLM", "Glm4vMoeForConditionalGeneration")
@@ -8390,13 +8272,14 @@ class Glm4MoeModel(TextModel):
         self, data_torch: Tensor, name: str, bid: int | None
     ) -> Iterable[tuple[str, Tensor]]:
         if name.startswith("model.visual."):  # ignore visual part
-            return []
+            return
         elif name.startswith("model.language_model."):
             name = name.replace("language_model.", "")  # for multimodal variants
 
         # Handle main token embedding (but not layer-specific NextN embeddings)
         if name == "model.embed_tokens.weight" and ".layers." not in name:
-            return [(self.map_tensor_name("token_embd.weight"), data_torch)]
+            yield from super().modify_tensors(data_torch, "token_embd.weight", bid)
+            return
 
         # Handle routed experts
         if name.find("mlp.experts") != -1:
@@ -8409,8 +8292,6 @@ class Glm4MoeModel(TextModel):
             self._experts[bid][name] = data_torch
 
             if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
                 # merge the experts into a single 3d tensor
                 for w_name in ["down_proj", "gate_proj", "up_proj"]:
                     datas: list[Tensor] = []
@@ -8424,18 +8305,15 @@ class Glm4MoeModel(TextModel):
 
                     merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
 
-                    new_name = self.map_tensor_name(merged_name)
-                    tensors.append((new_name, data_torch))
-                return tensors
+                    yield from super().modify_tensors(data_torch, merged_name, bid)
+                return
             else:
-                return []
+                return
 
         if name.endswith("e_score_correction_bias"):
             name = name.replace("e_score_correction_bias", "e_score_correction.bias")
 
-        new_name = self.map_tensor_name(name)
-
-        return [(new_name, data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -8446,6 +8324,32 @@ class Glm4MoeModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("Glm4MoeLiteForCausalLM")
+class Glm4MoeLiteModel(DeepseekV2Model):
+    model_arch = gguf.MODEL_ARCH.DEEPSEEK2
+
+    # copied from Glm4MoeModel
+    def set_vocab(self):
+        from transformers import AutoTokenizer
+
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        # Special tokens
+        # Note: Using <|endoftext|> (151329) for eot causes endless generation
+        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["[gMASK]"])  # 151331
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # 151336
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"]) # 151329
+        special_vocab._set_special_token("eom", tokenizer.get_added_vocab()["<|observation|>"])  # 151338
+
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+
 @ModelBase.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
 class ChatGLMModel(TextModel):
     model_arch = gguf.MODEL_ARCH.CHATGLM
@@ -8592,13 +8496,11 @@ class ChatGLMModel(TextModel):
         self.gguf_writer.add_rope_freq_base(rope_freq)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         if name.endswith(".rotary_pos_emb.inv_freq") or name.startswith("model.vision."):
-            return []
+            return
 
         name = name.removeprefix("transformer.")
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("NemotronForCausalLM")
@@ -8639,7 +8541,7 @@ class NemotronModel(TextModel):
         if name.endswith("norm.weight"):
             data_torch = data_torch + 1
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("ExaoneForCausalLM")
@@ -8795,11 +8697,9 @@ class ExaoneMoEModel(Exaone4Model):
                 new_name = remapper[_n.stem] + _n.suffix
 
                 # set shared weights for all NextN/MTP layers
-                tensors = []
                 for bid in range(self.hparams['num_hidden_layers'], self.block_count):
-                    new_name = new_name.format(bid=bid)
-                    tensors.append((self.map_tensor_name(new_name), data_torch))
-                return tensors
+                    yield from super().modify_tensors(data_torch, new_name.format(bid=bid), bid)
+                return
 
         if name.endswith("e_score_correction_bias"):
             name = name.replace("e_score_correction_bias", "e_score_correction.bias")
@@ -8814,8 +8714,6 @@ class ExaoneMoEModel(Exaone4Model):
             self._experts[bid][name] = data_torch
 
             if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
                 # merge the experts into a single 3d tensor
                 for w_name in ["down_proj", "gate_proj", "up_proj"]:
                     datas: list[Tensor] = []
@@ -8831,12 +8729,12 @@ class ExaoneMoEModel(Exaone4Model):
 
                     new_name = self.map_tensor_name(merged_name)
 
-                    tensors.append((new_name, data_torch))
-                return tensors
+                    yield from super().modify_tensors(data_torch, new_name, bid)
+                return
             else:
-                return []
+                return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -8906,10 +8804,9 @@ class GraniteMoeModel(GraniteModel):
             ffn_dim = self.hparams["intermediate_size"]
             assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * intermediate_size"
             gate, up = data_torch.split(ffn_dim, dim=-2)
-            return [
-                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), gate),
-                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), up),
-            ]
+            yield from ModelBase.modify_tensors(self, gate, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), bid)
+            yield from ModelBase.modify_tensors(self, up, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), bid)
+            return
 
         has_experts = bool(self.hparams.get('num_local_experts'))
 
@@ -8918,21 +8815,18 @@ class GraniteMoeModel(GraniteModel):
             assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * shared_intermediate_size"
             gate, up = data_torch.split(ffn_dim, dim=-2)
             if has_experts:
-                return [
-                    (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), gate),
-                    (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), up),
-                ]
-            return [
-                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), gate),
-                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), up),
-            ]
+                yield from ModelBase.modify_tensors(self, gate,self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), bid)
+                yield from ModelBase.modify_tensors(self, up, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), bid)
+                return
+            yield from ModelBase.modify_tensors(self, gate, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), bid)
+            yield from ModelBase.modify_tensors(self, up, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), bid)
+            return
 
         if not has_experts and name.endswith("shared_mlp.output_linear.weight"):
-            return [
-                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN, bid), data_torch)
-            ]
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN, bid), bid)
+            return
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("GraniteMoeHybridForCausalLM", "BambaForCausalLM")
@@ -9019,14 +8913,17 @@ class GraniteHybridModel(Mamba2Model, GraniteMoeModel):
             name.endswith("block_sparse_moe.input_linear.weight")
             or "shared_mlp" in name
         ):
-            return GraniteMoeModel.modify_tensors(self, data_torch, name, bid)
+            yield from GraniteMoeModel.modify_tensors(self, data_torch, name, bid)
+            return
 
         # Determine whether this is a mamba layer or an attention layer
         if bid in self._ssm_layers:
-            return Mamba2Model.modify_tensors(self, data_torch, name, bid)
+            yield from Mamba2Model.modify_tensors(self, data_torch, name, bid)
+            return
         elif bid in self._attn_layers:
-            return GraniteMoeModel.modify_tensors(self, data_torch, name, bid)
-        return [(self.map_tensor_name(name), data_torch)]
+            yield from GraniteMoeModel.modify_tensors(self, data_torch, name, bid)
+            return
+        yield from ModelBase.modify_tensors(self, data_torch, name, bid)
 
     def set_gguf_parameters(self):
         """This method merges params from both parents and some that are
@@ -9158,34 +9055,34 @@ class NemotronHModel(GraniteHybridModel):
         if self.is_moe and bid is not None:
             if name.endswith("mixer.gate.e_score_correction_bias"):
                 new_name = name.replace("e_score_correction_bias", "e_score_correction.bias")
-                mapped_name = self.map_tensor_name(new_name)
-                return [(mapped_name, data_torch)]
+                yield from ModelBase.modify_tensors(self, data_torch, new_name, bid)
+                return
 
             if name.endswith("mixer.dt_bias"):
                 new_name = name.replace("dt_bias", "dt.bias")
-                mapped_name = self.map_tensor_name(new_name)
-                return [(mapped_name, data_torch)]
+                yield from ModelBase.modify_tensors(self, data_torch, new_name, bid)
+                return
 
             if name.endswith("mixer.conv1d.weight"):
                 squeezed_data = data_torch.squeeze()
-                mapped_name = self.map_tensor_name(name)
-                return [(mapped_name, squeezed_data)]
+                yield from ModelBase.modify_tensors(self, squeezed_data, name, bid)
+                return
 
             if name.endswith("mixer.A_log"):
                 transformed_data = -torch.exp(data_torch)
                 reshaped_data = transformed_data.squeeze().reshape(-1, 1)
-                mapped_name = self.map_tensor_name(name)
-                return [(mapped_name, reshaped_data)]
+                yield from ModelBase.modify_tensors(self, reshaped_data, name, bid)
+                return
 
             if name.endswith("mixer.D"):
                 reshaped_data = data_torch.squeeze().reshape(-1, 1)
-                mapped_name = self.map_tensor_name(name)
-                return [(mapped_name, reshaped_data)]
+                yield from ModelBase.modify_tensors(self, reshaped_data, name, bid)
+                return
 
             if name.endswith("mixer.norm.weight"):
-                reshaped_data = data_torch.reshape(8, 512)
-                mapped_name = self.map_tensor_name(name)
-                return [(mapped_name, reshaped_data)]
+                reshaped_data = data_torch.reshape(self.n_group, -1)
+                yield from ModelBase.modify_tensors(self, reshaped_data, name, bid)
+                return
 
             if name.find("mixer.experts") != -1:
                 n_experts = self.hparams["n_routed_experts"]
@@ -9198,7 +9095,6 @@ class NemotronHModel(GraniteHybridModel):
 
                 if len(self._experts[bid]) >= n_experts * 2:
                     # merge the experts into a single tensor
-                    tensors: list[tuple[str, Tensor]] = []
                     for w_name in ["down_proj", "up_proj"]:
                         datas: list[Tensor] = []
 
@@ -9209,14 +9105,13 @@ class NemotronHModel(GraniteHybridModel):
 
                         data_torch = torch.stack(datas, dim=0)
                         merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
-                        new_name = self.map_tensor_name(merged_name)
-                        tensors.append((new_name, data_torch))
 
-                    return tensors
+                        yield from ModelBase.modify_tensors(self, data_torch, merged_name, bid)
+                    return
                 else:
-                    return []
+                    return
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -9275,21 +9170,19 @@ class BailingMoeModel(TextModel):
         output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
 
         if name.endswith("attention.dense.weight"):
-            return [(self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_OUT, bid), data_torch)]
+            yield from super().modify_tensors(data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_OUT, bid), bid)
+            return
         elif name.endswith("query_key_value.weight"):
             q, k, v = data_torch.split([n_head * head_dim, n_kv_head * head_dim, n_kv_head * head_dim], dim=-2)
 
-            return [
-                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), BailingMoeModel.permute(q, n_head, n_head)),
-                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), BailingMoeModel.permute(k, n_head, n_kv_head)),
-                (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), v)
-            ]
+            yield from super().modify_tensors(BailingMoeModel.permute(q, n_head, n_head), self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), bid)
+            yield from super().modify_tensors(BailingMoeModel.permute(k, n_head, n_kv_head), self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), bid)
+            yield from super().modify_tensors(v,self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), bid)
+            return
         elif name.find("mlp.experts") != -1:
             n_experts = self.hparams["num_experts"]
             assert bid is not None
 
-            tensors: list[tuple[str, Tensor]] = []
-
             if self._experts is None:
                 self._experts = [{} for _ in range(self.block_count)]
 
@@ -9311,9 +9204,9 @@ class BailingMoeModel(TextModel):
 
                     new_name = self.map_tensor_name(merged_name)
 
-                    tensors.append((new_name, data_torch))
+                    yield from super().modify_tensors(data_torch, new_name, bid)
 
-            return tensors
+            return
 
         new_name = self.map_tensor_name(name)
 
@@ -9321,7 +9214,7 @@ class BailingMoeModel(TextModel):
             data_torch = data_torch.float()
             data_torch /= torch.norm(data_torch, p=2, dim=0, keepdim=True) + 1e-7
 
-        return [(new_name, data_torch)]
+        yield from super().modify_tensors(data_torch, new_name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -9372,8 +9265,6 @@ class BailingMoeV2Model(TextModel):
             n_experts = self.hparams["num_experts"]
             assert bid is not None
 
-            tensors: list[tuple[str, Tensor]] = []
-
             if self._experts is None:
                 self._experts = [{} for _ in range(self.block_count)]
 
@@ -9393,16 +9284,13 @@ class BailingMoeV2Model(TextModel):
 
                     merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
 
-                    new_name = self.map_tensor_name(merged_name)
-
-                    tensors.append((new_name, data_torch))
-
-            return tensors
+                    yield from super().modify_tensors(data_torch, merged_name, bid)
+            return
 
         if name.endswith(".expert_bias"):
             name = name.replace(".expert_bias", ".expert_bias.bias")
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -9438,7 +9326,7 @@ class GroveMoeModel(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if name.endswith(".expert_bias"):
             # FIXME?: Unused https://huggingface.co/inclusionAI/GroveMoE-Inst/blob/c4c69e5970d18907b5e6ddccdfd55176fe292df1/modeling_grove_moe.py#L303
-            return []
+            return
 
         # process the experts separately
         if name.find("chunk_experts") != -1:
@@ -9451,8 +9339,6 @@ class GroveMoeModel(TextModel):
             self._chunk_experts[bid][name] = data_torch
 
             if len(self._chunk_experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
                 # merge the experts into a single 3d tensor
                 for w_name in ["down_proj", "gate_proj", "up_proj"]:
                     datas: list[Tensor] = []
@@ -9466,12 +9352,10 @@ class GroveMoeModel(TextModel):
 
                     merged_name = f"model.layers.{bid}.mlp.chunk_experts.{w_name}.weight"
 
-                    new_name = self.map_tensor_name(merged_name)
-
-                    tensors.append((new_name, data_torch))
-                return tensors
+                    yield from super().modify_tensors(data_torch, merged_name, bid)
+                return
             else:
-                return []
+                return
         elif name.find("experts") != -1:
             n_experts = self.hparams["num_experts"]
             assert bid is not None
@@ -9482,8 +9366,6 @@ class GroveMoeModel(TextModel):
             self._experts[bid][name] = data_torch
 
             if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
                 # merge the experts into a single 3d tensor
                 for w_name in ["down_proj", "gate_proj", "up_proj"]:
                     datas: list[Tensor] = []
@@ -9497,14 +9379,12 @@ class GroveMoeModel(TextModel):
 
                     merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
 
-                    new_name = self.map_tensor_name(merged_name)
-
-                    tensors.append((new_name, data_torch))
-                return tensors
+                    yield from super().modify_tensors(data_torch, merged_name, bid)
+                return
             else:
-                return []
+                return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -9538,7 +9418,7 @@ class ChameleonModel(TextModel):
         # ignore image tokenizer for now
         # TODO: remove this once image support is implemented for Chameleon
         if name.startswith("model.vqmodel"):
-            return []
+            return
 
         n_head = self.hparams["num_attention_heads"]
         n_kv_head = self.hparams.get("num_key_value_heads")
@@ -9553,7 +9433,7 @@ class ChameleonModel(TextModel):
         if name.endswith(("k_norm.weight", "k_norm.bias")):
             data_torch = ChameleonModel._reverse_hf_permute(data_torch, n_kv_head, hidden_dim)
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     # see: https://github.com/huggingface/transformers/blob/72fb02c47dbbe1999ae105319f24631cad6e2e00/src/transformers/models/chameleon/convert_chameleon_weights_to_hf.py#L176-L203
     @staticmethod
@@ -9598,11 +9478,9 @@ class GlmASRWhisperEncoderModel(MmprojModel):
         return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         if name.startswith("model.") or name.startswith("lm_head."):
             # skip language model tensors
-            return []
+            return
 
         if name.startswith("audio_encoder.whisper."):
             name = name.replace("audio_encoder.whisper.","audio_tower.")
@@ -9610,7 +9488,9 @@ class GlmASRWhisperEncoderModel(MmprojModel):
             name = name.replace("audio_encoder.", "audio_encoder.adapting.")
 
         if name.startswith("audio_encoder.audio_bos_eos_token."):
-            return [(self.map_tensor_name("model.vision.boi"), data_torch[0]), (self.map_tensor_name("model.vision.eoi"), data_torch[1])]
+            yield from super().modify_tensors(data_torch[0], "model.vision.boi", bid)
+            yield from super().modify_tensors(data_torch[1], "model.vision.eoi", bid)
+            return
 
         if name.startswith("audio_encoder.adapting."):
             name = name.replace("audio_encoder.adapting.","audio.multi_modal_projector.")
@@ -9621,13 +9501,13 @@ class GlmASRWhisperEncoderModel(MmprojModel):
             if ".2." in name:
                 name = name.replace(".2.", ".linear_2.")
             if ".proj." in name:
-                return []
+                return
 
         if "conv1.bias" in name or "conv2.bias" in name:
             # transpose conv1 and conv2 bias
             data_torch = data_torch.unsqueeze(-1)
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("Qwen2AudioForConditionalGeneration")
@@ -9654,11 +9534,9 @@ class WhisperEncoderModel(MmprojModel):
         return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         if name.startswith("language_model."):
             # skip language model tensors
-            return []
+            return
 
         # prevent clash naming with vision tensors
         if name.startswith("multi_modal_projector"):
@@ -9668,7 +9546,7 @@ class WhisperEncoderModel(MmprojModel):
             # transpose conv1 and conv2 bias
             data_torch = data_torch.unsqueeze(-1)
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("UltravoxModel")
@@ -9912,7 +9790,7 @@ class HunYuanMoEModel(TextModel):
         if name == "lm_head.weight":
             if self.hparams.get("tie_word_embeddings", False):
                 logger.info("Skipping tied output layer 'lm_head.weight'")
-                return []
+                return
 
         if name.find("mlp.experts") != -1:
             n_experts = self.hparams["num_experts"]
@@ -9925,7 +9803,6 @@ class HunYuanMoEModel(TextModel):
 
             if len(self._experts[bid]) >= n_experts * 3:
                 # merge the experts into a single 3d tensor
-                tensors: list[tuple[str, Tensor]] = []
                 for w_name in ["down_proj", "gate_proj", "up_proj"]:
                     datas: list[Tensor] = []
 
@@ -9936,14 +9813,13 @@ class HunYuanMoEModel(TextModel):
 
                     data_torch = torch.stack(datas, dim=0)
                     merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
-                    new_name = self.map_tensor_name(merged_name)
-                    tensors.append((new_name, data_torch))
 
-                return tensors
+                    yield from super().modify_tensors(data_torch, merged_name, bid)
+                return
             else:
-                return []
+                return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -9988,8 +9864,6 @@ class LLaDAMoEModel(TextModel):
             self._experts[bid][name] = data_torch
 
             if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
                 # merge the experts into a single 3d tensor
                 for w_name in ["down_proj", "gate_proj", "up_proj"]:
                     datas: list[Tensor] = []
@@ -10003,14 +9877,12 @@ class LLaDAMoEModel(TextModel):
 
                     merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
 
-                    new_name = self.map_tensor_name(merged_name)
-
-                    tensors.append((new_name, data_torch))
-                return tensors
+                    yield from super().modify_tensors(data_torch, merged_name, bid)
+                return
             else:
-                return []
+                return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     # Copied from: Qwen2MoeModel
     def prepare_tensors(self):
@@ -10109,9 +9981,9 @@ class HunYuanModel(TextModel):
         if name == "lm_head.weight":
             if self.hparams.get("tie_word_embeddings", False):
                 logger.info("Skipping tied output layer 'lm_head.weight'")
-                return []
+                return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("SmolLM3ForCausalLM")
@@ -10191,8 +10063,6 @@ class GptOssModel(TextModel):
         return []
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         if "sinks" in name:
             name += ".weight"
 
@@ -10206,7 +10076,7 @@ class GptOssModel(TextModel):
                 data_torch = data_torch.transpose(-1, -2)
             else:
                 # otherwise, it should already be repacked to ggml MXFP4 format
-                return []
+                return
 
         # split the gate_up into gate and up
         if "gate_up_proj" in name:
@@ -10214,25 +10084,18 @@ class GptOssModel(TextModel):
                 name_up = name.replace("gate_up_proj_bias", "up_proj.bias")
                 name_gate = name.replace("gate_up_proj_bias", "gate_proj.bias")
                 gate_proj_bias, up_proj_bias = data_torch[..., ::2], data_torch[..., 1::2]
-                return [
-                    (self.map_tensor_name(name_gate), gate_proj_bias),
-                    (self.map_tensor_name(name_up), up_proj_bias)
-                ]
+                yield from super().modify_tensors(gate_proj_bias, name_gate, bid)
+                yield from super().modify_tensors(up_proj_bias, name_up, bid)
             elif "_blocks" not in name and "_scales" not in name:
                 logger.warning(f"{name} is not in MXFP4, performance may be degraded")
                 name_up = name.replace("gate_up_proj", "up_proj.weight")
                 name_gate = name.replace("gate_up_proj", "gate_proj.weight")
                 data_torch = data_torch.transpose(-1, -2)
                 gate_proj_weight, up_proj_weight = data_torch[:, ::2, :], data_torch[:, 1::2, :]
-                return [
-                    (self.map_tensor_name(name_gate), gate_proj_weight),
-                    (self.map_tensor_name(name_up), up_proj_weight)
-                ]
-            else:
-                # otherwise, it should already be repacked to ggml MXFP4 format
-                return []
-
-        return [(self.map_tensor_name(name), data_torch)]
+                yield from super().modify_tensors(gate_proj_weight, name_gate, bid)
+                yield from super().modify_tensors(up_proj_weight, name_up, bid)
+        else:
+            yield from super().modify_tensors(data_torch, name, bid)
 
     def set_vocab(self):
         self._set_vocab_gpt2()
@@ -10280,7 +10143,7 @@ class LFM2Model(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if self._is_vision_tensor(name) or ConformerAudioModel.is_audio_tensor(name):
             # skip multimodal tensors
-            return []
+            return
 
         name = name.replace("language_model.", "") # vision
         name = name.replace("lfm.", "model.")      # audio
@@ -10289,7 +10152,7 @@ class LFM2Model(TextModel):
         if 'conv.conv' in name:
             data_torch = data_torch.squeeze(1)
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def _is_vision_tensor(self, name: str) -> bool:
         return "vision_tower" in name or "multi_modal_projector" in name
@@ -10304,7 +10167,7 @@ class LFM2ColBertModel(LFM2Model):
         if not name.startswith(self.dense_tensor_name):
             name = "model." + name
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         # dense tensor is stored in a separate safetensors file
@@ -10359,9 +10222,8 @@ class LFM2MoeModel(TextModel):
 
             # not enough expert weights to merge
             if len(expert_cache) < n_experts * len(expert_weights):
-                return []
+                return
 
-            tensors: list[tuple[str, Tensor]] = []
             for w_name in expert_weights:
                 datas: list[Tensor] = []
 
@@ -10372,13 +10234,13 @@ class LFM2MoeModel(TextModel):
 
                 data_torch = torch.stack(datas, dim=0)
                 merged_name = f"layers.{bid}.feed_forward.experts.{w_name}.weight"
-                new_name = self.map_tensor_name(merged_name)
-                tensors.append((new_name, data_torch))
+
+                yield from super().modify_tensors(data_torch, merged_name, bid)
 
             del self._experts_cache[bid]
-            return tensors
+            return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -10404,7 +10266,6 @@ class LFM2VLModel(MmprojModel):
         self.gguf_writer.add_vision_block_count(self.find_vparam(self.n_block_keys) - vision_feature_layers_to_drop)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
         is_vision_tensor = "vision_tower" in name or "multi_modal_projector" in name
 
         if is_vision_tensor:
@@ -10415,9 +10276,10 @@ class LFM2VLModel(MmprojModel):
             if "patch_embedding.weight" in name:
                 data_torch = data_torch.view(data_torch.shape[0], 16, 16, 3).permute(0, 3, 1, 2)
 
-            return [(self.map_tensor_name(name), data_torch)]
+            yield from super().modify_tensors(data_torch, name, bid)
+            return
 
-        return [] # skip other tensors
+        return # skip other tensors
 
 
 @ModelBase.register("Lfm2AudioForConditionalGeneration")
@@ -10442,17 +10304,17 @@ class LFM2AudioModel(ConformerAudioModel):
     def modify_tensors(self, data_torch, name, bid):
         # skip language model tensors
         if name.startswith("lfm."):
-            return []
+            return
 
         # for training only
         if any(p in name for p in ["audio_loss_weight"]):
-            return []
+            return
 
         # for audio output
         if any(p in name for p in ["codebook_offsets", "depth_embeddings", "depth_linear", "depthformer"]):
-            return []
+            return
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("SmallThinkerForCausalLM")
@@ -10497,8 +10359,6 @@ class SmallThinkerModel(TextModel):
             self._experts[bid][name] = data_torch
 
             if len(self._experts[bid]) >= n_experts * 3:
-                tensors: list[tuple[str, Tensor]] = []
-
                 # merge the experts into a single 3d tensor
                 for w_name in ["down", "gate", "up"]:
                     datas: list[Tensor] = []
@@ -10512,14 +10372,12 @@ class SmallThinkerModel(TextModel):
 
                     merged_name = f"model.layers.{bid}.block_sparse_moe.experts.{w_name}.weight"
 
-                    new_name = self.map_tensor_name(merged_name)
-
-                    tensors.append((new_name, data_torch))
-                return tensors
+                    yield from super().modify_tensors(data_torch, merged_name, bid)
+                return
             else:
-                return []
+                return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
     def prepare_tensors(self):
         super().prepare_tensors()
@@ -10552,12 +10410,12 @@ class ModernBertModel(BertModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # these layers act as MLM head, so we don't need them
         if name.startswith("decoder."):
-            return []
+            return
 
         if name.startswith("model."):
             name = name[6:]
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("ApertusForCausalLM")
@@ -10577,24 +10435,24 @@ class ApertusModel(LlamaModel):
             self._alpha_n[bid] = data_torch.to("cpu").float().item()
             if (len(self._alpha_n) == n_layers):
                 self.gguf_writer.add_xielu_alpha_n([self._alpha_n[k] for k in sorted(self._alpha_n)])
-            return []
+            return
         if name.endswith(".act_fn.alpha_p"):
             self._alpha_p[bid] = data_torch.to("cpu").float().item()
             if (len(self._alpha_p) == n_layers):
                 self.gguf_writer.add_xielu_alpha_p([self._alpha_p[k] for k in sorted(self._alpha_p)])
-            return []
+            return
         if name.endswith(".act_fn.beta"):
             self._beta[bid] = data_torch.to("cpu").float().item()
             if (len(self._beta) == n_layers):
                 self.gguf_writer.add_xielu_beta([self._beta[k] for k in sorted(self._beta)])
-            return []
+            return
         if name.endswith(".act_fn.eps"):
             self._eps[bid] = data_torch.to("cpu").float().item()
             if (len(self._eps) == n_layers):
                 self.gguf_writer.add_xielu_eps([self._eps[k] for k in sorted(self._eps)])
-            return []
+            return
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 class MistralModel(LlamaModel):
@@ -10757,7 +10615,7 @@ class MistralMoeModel(DeepseekV2Model):
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
         if name.startswith("vision_") or name.startswith("patch_merger.") or "mm_projector" in name:
-            return []
+            return
 
         # rename certain tensors so that we can reuse DeepseekV2Model modify_tensors logic
         if name.endswith(".qscale_act"):
@@ -10773,7 +10631,7 @@ class MistralMoeModel(DeepseekV2Model):
             name = name.replace(".w3.", ".up_proj.")
             name = "model." + name
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 class PixtralModel(LlavaVisionModel):
@@ -10818,7 +10676,7 @@ class LightOnOCRVisionModel(LlavaVisionModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
         name = name.replace("model.vision_encoder.", "vision_tower.")
         name = name.replace("model.vision_projection.", "multi_modal_projector.")
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("KimiVLForConditionalGeneration")
@@ -10838,24 +10696,20 @@ class KimiVLModel(MmprojModel):
         self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("layer_norm_eps", 1e-5))
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
         is_vision_tensor = "vision_tower" in name or "multi_modal_projector" in name
 
         if is_vision_tensor:
             if "pos_emb.weight" in name:
                 data_torch = data_torch.view(data_torch.shape[0] * data_torch.shape[1], data_torch.shape[2])
-            elif "wqkv" in name:
+
+            if "wqkv" in name:
                 split_dim = 0 if "weight" in name else -1
                 wq, wk, wv = data_torch.chunk(3, dim=split_dim)
-                return [
-                    (self.map_tensor_name(name.replace("wqkv", "wq")), wq),
-                    (self.map_tensor_name(name.replace("wqkv", "wk")), wk),
-                    (self.map_tensor_name(name.replace("wqkv", "wv")), wv)
-                ]
-
-            return [(self.map_tensor_name(name), data_torch)]
-
-        return [] # skip other tensors
+                yield from super().modify_tensors(wq, name.replace("wqkv", "wq"), bid)
+                yield from super().modify_tensors(wk, name.replace("wqkv", "wk"), bid)
+                yield from super().modify_tensors(wv, name.replace("wqkv", "wv"), bid)
+            else:
+                yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("CogVLMForCausalLM")
@@ -10867,12 +10721,10 @@ class CogVLMVisionModel(MmprojModel):
         self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.COGVLM)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         if not name.startswith("model.vision."):
-            return []
+            return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("CogVLMForCausalLM")
@@ -10880,13 +10732,11 @@ class CogVLMModel(LlamaModel):
     model_arch = gguf.MODEL_ARCH.COGVLM
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         # block vision tensors
         if name.startswith("model.vision."):
-            return []
+            return
 
-        return [(self.map_tensor_name(name), data_torch)]
+        yield from ModelBase.modify_tensors(self, data_torch, name, bid)
 
 
 @ModelBase.register("JanusForConditionalGeneration")
@@ -10904,14 +10754,14 @@ class JanusProModel(LlamaModel):
             'model.generation_head.',
         )
         if name.startswith(skip_prefixes):
-            return []
+            return
 
         if name.startswith('model.language_model.'):
             name = name.replace('model.language_model.', 'model.')
         elif name.startswith('language_model.'):
             name = name.replace('language_model.', '')
 
-        return super().modify_tensors(data_torch, name, bid)
+        yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("JanusForConditionalGeneration")
@@ -10964,11 +10814,9 @@ class JanusProVisionModel(MmprojModel):
         return [(tensor_name, data_torch)]
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         # Skip language model tensors as they will be handled by `JanusProModel`
         if name.startswith(('model.language_model.', 'language_model.')):
-            return []
+            return
 
         # Skip generation-related components
         skip_generation_prefixes = (
@@ -10982,17 +10830,19 @@ class JanusProVisionModel(MmprojModel):
             'generation_head.',
         )
         if name.startswith(skip_generation_prefixes):
-            return []
+            return
 
         # Handle aligner tensors
         if name.startswith(('model.aligner.', 'aligner.')):
-            return list(self._map_aligner_tensor(data_torch, name))
+            yield from self._map_aligner_tensor(data_torch, name)
+            return
 
         # Handle vision tensors
         if name.startswith(('model.vision_model.', 'vision_model.')):
-            return [(self.map_tensor_name(name), data_torch)]
+            yield from super().modify_tensors(data_torch, name, bid)
+            return
 
-        return []
+        return
 
 
 @ModelBase.register("YoutuVLForConditionalGeneration")
@@ -11031,21 +10881,18 @@ class YoutuVLVisionModel(MmprojModel):
         self.gguf_writer.add_vision_wa_layer_indexes(layers=fullatt_block_indexes)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        del bid  # unused
-
         # Skip language model tensors
         skip_prefixes = ('lm_head.', 'model.layers.', 'model.embed_tokens.', 'model.norm.')
         if name.startswith(skip_prefixes):
-            return []
+            return
 
         # Try to map the tensor using TensorNameMap (handles vision encoder and projector)
         try:
-            new_name = self.map_tensor_name(name)
-            return [(new_name, data_torch)]
+            yield from super().modify_tensors(data_torch, name, bid)
         except ValueError:
             # If mapping fails, log warning and skip
             logger.warning(f"Cannot map tensor: {name}")
-            return []
+            return
 
 
 @ModelBase.register("SolarOpenForCausalLM")
diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py
index aa9843ea17..2811f7f884 100755
--- a/convert_hf_to_gguf_update.py
+++ b/convert_hf_to_gguf_update.py
@@ -170,6 +170,7 @@ pre_computed_hashes = [
     {"name": "grok-2",    "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/alvarobartt/grok-2-tokenizer", "chkhsh": "66b8d4e19ab16c3bfd89bce5d785fb7e0155e8648708a1f42077cb9fe002c273"},
     # jina-v2-de variants
     {"name": "jina-v2-de", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/aari1995/German_Semantic_V3", "chkhsh": "b3d1dd861f1d4c5c0d2569ce36baf3f90fe8a102db3de50dd71ff860d91be3df"},
+    {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/zai-org/GLM-4.7-Flash", "chkhsh": "cdf5f35325780597efd76153d4d1c16778f766173908894c04afc20108536267"},
 ]
 
 
diff --git a/docs/backend/OPENCL.md b/docs/backend/OPENCL.md
index 0561a74c47..044ac606ba 100644
--- a/docs/backend/OPENCL.md
+++ b/docs/backend/OPENCL.md
@@ -8,6 +8,7 @@
 - [CMake Options](#cmake-options)
 - [Android](#android)
 - [Windows 11 Arm64](#windows-11-arm64)
+- [Linux](#Linux)
 - [Known Issue](#known-issues)
 - [TODO](#todo)
 
diff --git a/docs/backend/SYCL.md b/docs/backend/SYCL.md
index bcb3ce6743..c0a422b3dc 100644
--- a/docs/backend/SYCL.md
+++ b/docs/backend/SYCL.md
@@ -35,9 +35,9 @@ The following releases are verified and recommended:
 
 |Commit ID|Tag|Release|Verified  Platform| Update date|
 |-|-|-|-|-|
-|24e86cae7219b0f3ede1d5abdf5bf3ad515cccb8|b5377 |[llama-b5377-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b5377/llama-b5377-bin-win-sycl-x64.zip) |ArcB580/Linux/oneAPI 2025.1<br>LNL Arc GPU/Windows 11/oneAPI 2025.1.1|2025-05-15|
-|3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19|
-|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1||
+|24e86cae7219b0f3ede1d5abdf5bf3ad515cccb8|b5377 |[llama-b5377-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b5377/llama-b5377-bin-win-sycl-x64.zip) |Arc B580/Linux/oneAPI 2025.1<br>LNL Arc GPU/Windows 11/oneAPI 2025.1.1|2025-05-15|
+|3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc A770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19|
+|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc A770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1||
 
 
 ## News
@@ -51,7 +51,7 @@ The following releases are verified and recommended:
     |-|-|-|-|
     |PVC 1550|39|73|+87%|
     |Flex 170|39|50|+28%|
-    |Arc770|42|55|+30%|
+    |Arc A770|42|55|+30%|
     |MTL|13|16|+23%|
     |ARL-H|14|17|+21%|
 
@@ -62,7 +62,7 @@ The following releases are verified and recommended:
   - Use oneDNN as the default GEMM library, improve the compatibility for new Intel GPUs.
 
 - 2024.5
-  - Performance is increased: 34 -> 37 tokens/s of llama-2-7b.Q4_0 on Arc770.
+  - Performance is increased: 34 -> 37 tokens/s of llama-2-7b.Q4_0 on Arc A770.
   - Arch Linux is verified successfully.
 
 - 2024.4
@@ -111,7 +111,8 @@ On older Intel GPUs, you may try [OpenCL](/docs/backend/OPENCL.md) although the
 |-------------------------------|---------|---------------------------------------|
 | Intel Data Center Max Series  | Support | Max 1550, 1100                        |
 | Intel Data Center Flex Series | Support | Flex 170                              |
-| Intel Arc Series              | Support | Arc 770, 730M, Arc A750, B580         |
+| Intel Arc A-Series              | Support | Arc A770, Arc A730M, Arc A750         |
+| Intel Arc B-Series              | Support | Arc B580         |
 | Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake, Arrow Lake, Lunar Lake |
 | Intel iGPU                    | Support | iGPU in 13700k, 13400, i5-1250P, i7-1260P, i7-1165G7  |
 
diff --git a/docs/backend/hexagon/CMakeUserPresets.json b/docs/backend/snapdragon/CMakeUserPresets.json
similarity index 70%
rename from docs/backend/hexagon/CMakeUserPresets.json
rename to docs/backend/snapdragon/CMakeUserPresets.json
index 1f2676c0bc..4cf473d05f 100644
--- a/docs/backend/hexagon/CMakeUserPresets.json
+++ b/docs/backend/snapdragon/CMakeUserPresets.json
@@ -1,5 +1,10 @@
 {
-  "version": 4,
+  "version": 5,
+  "cmakeMinimumRequired": {
+      "major": 3,
+      "minor": 28,
+      "patch": 0
+  },
   "configurePresets": [
     {
         "name": "arm64-android-snapdragon",
@@ -16,7 +21,9 @@
             "CMAKE_CXX_FLAGS_RELEASE":        "-O3 -DNDEBUG",
             "CMAKE_C_FLAGS_RELWITHDEBINFO":   "-O3 -DNDEBUG -g",
             "CMAKE_CXX_FLAGS_RELWITHDEBINFO": "-O3 -DNDEBUG -g",
-            "HEXAGON_SDK_ROOT": "$env{HEXAGON_SDK_ROOT}",
+            "CMAKE_PREFIX_PATH":  "$env{OPENCL_SDK_ROOT}",
+            "HEXAGON_SDK_ROOT":   "$env{HEXAGON_SDK_ROOT}",
+            "HEXAGON_TOOLS_ROOT": "$env{HEXAGON_TOOLS_ROOT}",
             "PREBUILT_LIB_DIR": "android_aarch64",
             "GGML_OPENMP":      "OFF",
             "GGML_LLAMAFILE":   "OFF",
@@ -31,7 +38,15 @@
         "name": "arm64-windows-snapdragon",
         "inherits": [ "base", "arm64-windows-llvm" ],
         "cacheVariables": {
-            "HEXAGON_SDK_ROOT": "$env{HEXAGON_SDK_ROOT}",
+            "CMAKE_C_FLAGS":   "-march=armv8.7a+fp16 -fvectorize -ffp-model=fast -flto -D_GNU_SOURCE",
+            "CMAKE_CXX_FLAGS": "-march=armv8.7a+fp16 -fvectorize -ffp-model=fast -flto -D_GNU_SOURCE",
+            "CMAKE_C_FLAGS_RELEASE":          "-O3 -DNDEBUG",
+            "CMAKE_CXX_FLAGS_RELEASE":        "-O3 -DNDEBUG",
+            "CMAKE_C_FLAGS_RELWITHDEBINFO":   "-O3 -DNDEBUG -g",
+            "CMAKE_CXX_FLAGS_RELWITHDEBINFO": "-O3 -DNDEBUG -g",
+            "CMAKE_PREFIX_PATH":  "$env{OPENCL_SDK_ROOT}",
+            "HEXAGON_SDK_ROOT":   "$env{HEXAGON_SDK_ROOT}",
+            "HEXAGON_TOOLS_ROOT": "$env{HEXAGON_TOOLS_ROOT}",
             "PREBUILT_LIB_DIR": "windows_aarch64",
             "GGML_OPENMP":      "OFF",
             "GGML_LLAMAFILE":   "OFF",
diff --git a/docs/backend/hexagon/README.md b/docs/backend/snapdragon/README.md
similarity index 84%
rename from docs/backend/hexagon/README.md
rename to docs/backend/snapdragon/README.md
index 3befdf7225..8e1f37b206 100644
--- a/docs/backend/hexagon/README.md
+++ b/docs/backend/snapdragon/README.md
@@ -1,6 +1,8 @@
-# Snapdragon-based Android devices
+# Snapdragon-based devices
 
-## How to Build
+## Setup
+
+### Android
 
 The easiest way to build llama.cpp for a Snapdragon-based Android device is using the toolchain Docker image (see github.com/snapdragon-toolchain).
 This image includes Android NDK, OpenCL SDK, Hexagon SDK, CMake, etc.
@@ -12,7 +14,24 @@ This method works on Linux, macOS, and Windows. macOS and Windows users should i
 [d]/> cd /workspace
 ```
 
-The rest of the Android build process assumes that you're running inside the toolchain container.
+Note: The rest of the **Android** build process assumes that you're running inside the toolchain container.
+
+### Windows On Snapdragon
+
+Native Windows 11 arm64 builds has the following tools dependencies:
+- MS Visual Studio 2026 (Community Edition or Pro)
+  - MSVC arm64 standard and runtime libraries
+  - UCRT and Driver Kit
+- LLVM core libraries and Clang compiler (winget)
+- CMake, Git, Python (winget)
+- Hexagon SDK Community Edition 6.4 or later (see windows.md)
+- OpenCL SDK 2.3 or later (see windows.md)
+
+Note: The rest of the **Windows** build process assumes that you're running natively in Powershell.
+Adapt below build commands accordingly.
+
+## How to Build
+
 Let's build llama.cpp with CPU, OpenCL, and Hexagon backends via CMake presets:
 
 ```
@@ -49,24 +68,26 @@ Preset CMake variables:
 To generate an installable "package" simply use cmake --install:
 
 ```
-[d]/workspace> cmake --install build-snapdragon --prefix pkg-adb/llama.cpp
+[d]/workspace> cmake --install build-snapdragon --prefix pkg-snapdragon/llama.cpp
 -- Install configuration: "Release"
--- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-cpu.so
--- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-opencl.so
--- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-hexagon.so
--- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-htp-v73.so
--- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-htp-v75.so
--- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-htp-v79.so
--- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml-htp-v81.so
--- Installing: /workspace/pkg-adb/llama.cpp/lib/libggml.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-cpu.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-opencl.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-hexagon.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-htp-v73.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-htp-v75.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-htp-v79.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-htp-v81.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml.so
 ...
--- Installing: /workspace/pkg-adb/llama.cpp/bin/llama-bench
--- Installing: /workspace/pkg-adb/llama.cpp/bin/llama-cli
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/bin/llama-bench
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/bin/llama-cli
 ...
 ```
 
 ## How to Install
 
+### Android
+
 For this step, your device needs to be configured for on-device development.
 Please see https://developer.android.com/studio/debug/dev-options for details.
 
@@ -74,10 +95,10 @@ Once ADB is enabled, use `adb push` to install `pkg-snapdragon` on the device.
 **Note that the toolchain Docker image doesn't have ADB and doesn't set up the ADB bridge. Please use native ADB on the host.**
 
 ```
-~/src/llama.cpp$ adb push pkg-adb/llama.cpp /data/local/tmp/
-pkg-adb/llama.cpp/bin/: 67 files pushed, 0 skipped. 190.2 MB/s (919095042 bytes in 4.607s)
-pkg-adb/llama.cpp/include/: 19 files pushed, 0 skipped. 20.5 MB/s (255173 bytes in 0.012s)
-pkg-adb/llama.cpp/lib/: 16 files pushed, 0 skipped. 144.4 MB/s (43801382 bytes in 0.289s)
+~/src/llama.cpp$ adb push pkg-snapdragon/llama.cpp /data/local/tmp/
+pkg-snapdragon/llama.cpp/bin/: 67 files pushed, 0 skipped. 190.2 MB/s (919095042 bytes in 4.607s)
+pkg-snapdragon/llama.cpp/include/: 19 files pushed, 0 skipped. 20.5 MB/s (255173 bytes in 0.012s)
+pkg-snapdragon/llama.cpp/lib/: 16 files pushed, 0 skipped. 144.4 MB/s (43801382 bytes in 0.289s)
 102 files pushed, 0 skipped. 186.9 MB/s (963151597 bytes in 4.914s)
 ```
 
@@ -92,6 +113,11 @@ At this point, you should also install some models:
 Llama-3.2-1B-Instruct-Q4_0.gguf: 1 file pushed, 0 skipped. 38.3 MB/s (773025920 bytes in 19.250s)
 ```
 
+### Windows
+
+All artifacts are already installed in the `pkg-snapdragon` folder.
+To run, adapt below instructions to use Powershell scrits in `scripts/snapdragon/windows`.
+
 ## How to Run
 
 The easiest way to run llama.cpp cli tools is using provided wrapper scripts that properly set up all required environment variables.
diff --git a/docs/backend/hexagon/developer.md b/docs/backend/snapdragon/developer.md
similarity index 100%
rename from docs/backend/hexagon/developer.md
rename to docs/backend/snapdragon/developer.md
diff --git a/docs/backend/snapdragon/windows.md b/docs/backend/snapdragon/windows.md
new file mode 100644
index 0000000000..710ad8fdf4
--- /dev/null
+++ b/docs/backend/snapdragon/windows.md
@@ -0,0 +1,161 @@
+## Overview
+
+The document covers procedures for installing the latest GPU and NPU drivers, and OpenCL and Hexagon SDKs.
+
+
+In order to use Hexagon NPU on Snapdragon Windows devices the underlying HTP Ops libraries (e.g libggml-htp-v73.so)
+must be included in the .cat file digitally signed with a trusted certificate.
+
+This document covers details on how to generate personal certificate files (.pfx) and how to configure the system
+to allow for test signatures (aka test-signing).
+
+## Install the latest Adreno OpenCL SDK
+
+Either use the trimmed down version (optimized for CI) from
+
+    https://github.com/snapdragon-toolchain/opencl-sdk/releases/download/v2.3.2/adreno-opencl-sdk-v2.3.2-arm64-wos.tar.xz
+
+Or download the complete official version from
+
+    https://softwarecenter.qualcomm.com/catalog/item/Adreno_OpenCL_SDK?version=2.3.2
+
+Unzip/untar the archive into
+```
+c:\Qualcomm\OpenCL_SDK\2.3.2
+```
+
+## Install the latest Hexagon SDK Community Edition
+
+Either use the trimmed down version (optimized for CI) from
+
+    https://github.com/snapdragon-toolchain/hexagon-sdk/releases/download/v6.4.0.2/hexagon-sdk-v6.4.0.2-arm64-wos.tar.xz
+
+Or download the complete official version from
+
+    https://softwarecenter.qualcomm.com/catalog/item/Hexagon_SDK?version=6.4.0.2
+
+Unzip/untar the archive into
+```
+c:\Qualcomm\Hexagon_SDK\6.4.0.2
+```
+
+## Install the latest Adreno GPU driver
+
+Download the driver from
+
+    https://softwarecenter.qualcomm.com/catalog/item/Windows_Graphics_Driver
+
+After the automated installation and reboot please make sure that the GPU device shows up in the `Device Manager` (under 'Display Adapters`)
+
+## Install the latest Qualcomm NPU driver
+
+Download the driver from
+
+    https://softwarecenter.qualcomm.com/catalog/item/Qualcomm_HND
+
+After the automated installation and reboot please make sure that the Hexagon NPU device shows up in the `Device Manager` (under `Neural Processors`).
+
+If the device is not available you can try installing all components (`qcnspmcdm8380`, `qcnspmcdm8380_ext`) manually.
+The components are extracted into
+```
+c:\QCDrivers\qcnspmcdm...
+```
+
+## Enable NPU driver test signatures
+
+Please note that the following steps are required only for the Hexagon NPU.
+Adreno GPU backend does not require test signatures.
+
+### Enable testsigning
+
+Use `bcdedit` to enable test-signing
+```
+> bcdedit /set TESTSIGNING ON
+```
+(Secure Boot may need to be disabled for this to work)
+
+Make sure test-signing is enabled after reboot
+```
+> bcdedit /enum
+...
+testsigning             Yes
+...
+```
+For additional details see Microsoft guide at
+
+   https://learn.microsoft.com/en-us/windows-hardware/drivers/install/the-testsigning-boot-configuration-option
+
+### Create personal certificate
+
+The tools required for this procedure are available as part of Windows SDK and Windows Driver Kit which should be
+installed as part of the MS Visual Studio.
+They are typically located at
+```
+c:\Program Files (x86)\Windows Kits\10\bin\10.0.26100.0
+```
+(replace 10.0.26100.0 with correct version).
+
+To create personal self-signed certificate run the following commands (either from cmd or power-shell):
+```
+> cd c:\Users\MyUser
+> mkdir Certs
+> cd Certs
+> makecert -r -pe -ss PrivateCertStore -n CN=GGML.HTP.v1 -eku 1.3.6.1.5.5.7.3.3 -sv ggml-htp-v1.pvk ggml-htp-v1.cer
+> pvk2pfx.exe -pvk ggml-htp-v1.pvk -spc ggml-htp-v1.cer -pfx ggml-htp-v1.pfx
+```
+(replace `MyUser` with your username).
+
+Add this certificate to `Trusted Root Certification Authorities` and `Trusted Publishers` stores.
+This can be done using `certlm` Certificate Manager tool.
+Right click on the certificate store, select `All Tasks -> Import` and follow the prompts to import the certificate from the
+PFX file you created above.
+
+For additional details see Microsoft guide at
+
+    https://learn.microsoft.com/en-us/windows-hardware/drivers/install/introduction-to-test-signing
+
+Make sure to save the PFX file, you will need it for the build procedures.
+Please note that the same certificate can be used for signing any number of builds.
+
+## Build Hexagon backend with signed HTP ops libraries
+
+The overall Hexagon backend build procedure for Windows on Snapdragon is the same as for other platforms.
+However, additional settings are required for generating and signing HTP Ops libraries.
+```
+> $env:OPENCL_SDK_ROOT="C:\Qualcomm\OpenCL_SDK\2.3.2"
+> $env:HEXAGON_SDK_ROOT="C:\Qualcomm\Hexagon_SDK\6.4.0.2"
+> $env:HEXAGON_TOOLS_ROOT="C:\Qualcomm\Hexagon_SDK\6.4.0.2\tools\HEXAGON_Tools\19.0.04"
+> $env:HEXAGON_HTP_CERT="c:\Users\MyUsers\Certs\ggml-htp-v1.pfx"
+> $env:WINDOWS_SDK_BIN="C:\Program Files (x86)\Windows Kits\10\bin\10.0.26100.0\arm64"
+
+> cmake --preset arm64-windows-snapdragon -B build-wos
+...
+> cmake --install build-wos --prefix pkg-snapdragon
+```
+
+Once the build is complete HTP ops libraries will be installed like this
+```
+> dir pkg-snapdragon/lib
+...
+-a----         1/22/2026   6:01 PM         187656 libggml-htp-v73.so
+-a----         1/22/2026   6:01 PM         191752 libggml-htp-v75.so
+-a----         1/22/2026   6:01 PM         187656 libggml-htp-v79.so
+-a----         1/22/2026   6:01 PM         187656 libggml-htp-v81.so
+-a----         1/22/2026   6:01 PM           4139 libggml-htp.cat
+```
+
+The .cat file, the signature and proper certicate installation can be verified with
+
+```
+> signtool.exe verify /v /pa .\pkg-snapdragon\lib\libggml-htp.cat
+Verifying: .\pkg-snapdragon\lib\libggml-htp.cat
+
+Signature Index: 0 (Primary Signature)
+Hash of file (sha256): 9820C664DA59D5EAE31DBB664127FCDAEF59CDC31502496BC567544EC2F401CF
+
+Signing Certificate Chain:
+        Issued to: GGML.HTP.v1
+...
+Successfully verified: .\pkg-snapdragon\lib\libggml-htp.cat
+...
+```
diff --git a/docs/build.md b/docs/build.md
index fce9361b2d..3a43f2a45a 100644
--- a/docs/build.md
+++ b/docs/build.md
@@ -144,7 +144,7 @@ We also have a [guide](./backend/CUDA-FEDORA.md) for setting up CUDA toolkit in
 - ***Necessary*** for users of [Atomic Desktops for Fedora](https://fedoraproject.org/atomic-desktops/); such as: [Silverblue](https://fedoraproject.org/atomic-desktops/silverblue/) and [Kinoite](https://fedoraproject.org/atomic-desktops/kinoite/).
   - (there are no supported CUDA packages for these systems)
 - ***Necessary*** for users that have a host that is not a: [Supported Nvidia CUDA Release Platform](https://developer.nvidia.com/cuda-downloads).
-  - (for example, you may have [Fedora 42 Beta](https://fedoramagazine.org/announcing-fedora-linux-42-beta/) as your your host operating system)
+  - (for example, you may have [Fedora 42 Beta](https://fedoramagazine.org/announcing-fedora-linux-42-beta/) as your host operating system)
 - ***Convenient*** For those running [Fedora Workstation](https://fedoraproject.org/workstation/) or [Fedora KDE Plasma Desktop](https://fedoraproject.org/spins/kde), and want to keep their host system clean.
 - *Optionally* toolbox packages are available: [Arch Linux](https://archlinux.org/), [Red Hat Enterprise Linux >= 8.5](https://www.redhat.com/en/technologies/linux-platforms/enterprise-linux), or [Ubuntu](https://ubuntu.com/download)
 
@@ -248,6 +248,14 @@ You may set the [cuda environmental variables](https://docs.nvidia.com/cuda/cuda
 CUDA_VISIBLE_DEVICES="-0" ./build/bin/llama-server --model /srv/models/llama.gguf
 ```
 
+#### CUDA_SCALE_LAUNCH_QUEUES
+
+The environment variable [`CUDA_SCALE_LAUNCH_QUEUES`](https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/environment-variables.html#cuda-scale-launch-queues) controls the size of CUDA's command buffer, which determines how many GPU operations can be queued before the CPU must wait for the GPU to catch up. A larger buffer reduces CPU-side stalls and allows more work to be queued on a GPU.
+
+**Default behavior:** llama.cpp automatically sets `CUDA_SCALE_LAUNCH_QUEUES=4x`, which increases the CUDA command buffer to 4 times its default size. This optimization is particularly beneficial for **Multi-GPU setups with pipeline parallelism**, where it significantly improves prompt processing throughput by allowing more operations to be enqueued across GPUs.
+
+See PR [#19042](https://github.com/ggml-org/llama.cpp/pull/19042) for performance benchmarks and technical details.
+
 ### Unified Memory
 
 The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. In Windows this setting is available in the NVIDIA control panel as `System Memory Fallback`.
@@ -487,6 +495,37 @@ Finally, after finishing your build, you should be able to do something like thi
 # ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
 ```
 
+### For Mac users:
+
+Generally, follow LunarG's [Getting Started with the MacOS Vulkan SDK](https://vulkan.lunarg.com/doc/sdk/latest/mac/getting_started.html) guide for installation and setup of the Vulkan SDK. There are two options of Vulkan drivers on macOS, both of which implement translation layers to map Vulkan to Metal. They can be hot-swapped by setting the `VK_ICD_FILENAMES` environment variable to point to the respective ICD JSON file.
+
+Check the box for "KosmicKrisp" during the LunarG Vulkan SDK installation.
+
+Set environment variable for the LunarG Vulkan SDK after installation (and optionally add to your shell profile for persistence):
+```bash
+source /path/to/vulkan-sdk/setup-env.sh
+```
+
+#### Using MoltenVK
+
+MoltenVK is the default Vulkan driver installed with the LunarG Vulkan SDK on macOS, so you can use the above environment variable settings as is.
+
+#### Using KosmicKrisp
+
+Override the environment variable for KosmicKrisp:
+```bash
+export VK_ICD_FILENAMES=$VULKAN_SDK/share/vulkan/icd.d/libkosmickrisp_icd.json
+export VK_DRIVER_FILES=$VULKAN_SDK/share/vulkan/icd.d/libkosmickrisp_icd.json
+```
+
+#### Build
+
+This is the only step different from [above](#common-steps) instructions.
+```bash
+cmake -B build -DGGML_VULKAN=1 -DGGML_METAL=OFF
+cmake --build build --config Release
+```
+
 ## CANN
 This provides NPU acceleration using the AI cores of your Ascend NPU. And [CANN](https://www.hiascend.com/en/software/cann) is a hierarchical APIs to help you to quickly build AI applications and service based on Ascend NPU.
 
diff --git a/docs/ops.md b/docs/ops.md
index c066ab5a85..2c7c60dcca 100644
--- a/docs/ops.md
+++ b/docs/ops.md
@@ -97,7 +97,7 @@ Legend:
 |                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
-|                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ |
+|                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
 |                        SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
@@ -114,7 +114,7 @@ Legend:
 |                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            TOP_K | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ |
-|                              TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                              TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ |
 |                            XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
diff --git a/docs/ops/SYCL.csv b/docs/ops/SYCL.csv
index 91b442bde8..c1d22e65d4 100644
--- a/docs/ops/SYCL.csv
+++ b/docs/ops/SYCL.csv
@@ -29,8 +29,8 @@
 "SYCL0","EXP","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","SYCL"
 "SYCL0","EXPM1","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","SYCL"
 "SYCL0","EXPM1","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","SYCL"
-"SYCL0","SOFTPLUS","type=f16,ne_a=[128,2,2,2],v=0","support","0","no","SYCL"
-"SYCL0","SOFTPLUS","type=f16,ne_a=[5,7,11,13],v=0","support","0","no","SYCL"
+"SYCL0","SOFTPLUS","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","SYCL"
+"SYCL0","SOFTPLUS","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","SYCL"
 "SYCL0","GELU_ERF","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","SYCL"
 "SYCL0","GELU_ERF","type=f16,ne_a=[5,7,11,13],v=0","support","1","yes","SYCL"
 "SYCL0","FLOOR","type=f16,ne_a=[128,2,2,2],v=0","support","1","yes","SYCL"
@@ -71,8 +71,8 @@
 "SYCL0","EXP","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","EXPM1","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
 "SYCL0","EXPM1","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
-"SYCL0","SOFTPLUS","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
-"SYCL0","SOFTPLUS","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
+"SYCL0","SOFTPLUS","type=f16,ne_a=[128,2,2,2],v=1","support","1","yes","SYCL"
+"SYCL0","SOFTPLUS","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","GELU_ERF","type=f16,ne_a=[128,2,2,2],v=1","support","1","yes","SYCL"
 "SYCL0","GELU_ERF","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","FLOOR","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
@@ -113,8 +113,8 @@
 "SYCL0","EXP","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","SYCL"
 "SYCL0","EXPM1","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","SYCL"
 "SYCL0","EXPM1","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","SYCL"
-"SYCL0","SOFTPLUS","type=f32,ne_a=[128,2,2,2],v=0","support","0","no","SYCL"
-"SYCL0","SOFTPLUS","type=f32,ne_a=[5,7,11,13],v=0","support","0","no","SYCL"
+"SYCL0","SOFTPLUS","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","SYCL"
+"SYCL0","SOFTPLUS","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","SYCL"
 "SYCL0","GELU_ERF","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","SYCL"
 "SYCL0","GELU_ERF","type=f32,ne_a=[5,7,11,13],v=0","support","1","yes","SYCL"
 "SYCL0","FLOOR","type=f32,ne_a=[128,2,2,2],v=0","support","1","yes","SYCL"
@@ -155,8 +155,8 @@
 "SYCL0","EXP","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","EXPM1","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
 "SYCL0","EXPM1","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
-"SYCL0","SOFTPLUS","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
-"SYCL0","SOFTPLUS","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
+"SYCL0","SOFTPLUS","type=f32,ne_a=[128,2,2,2],v=1","support","1","yes","SYCL"
+"SYCL0","SOFTPLUS","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","GELU_ERF","type=f32,ne_a=[128,2,2,2],v=1","support","1","yes","SYCL"
 "SYCL0","GELU_ERF","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","FLOOR","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
@@ -878,6 +878,54 @@
 "SYCL0","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=2,p0=0,p1=1","support","1","yes","SYCL"
 "SYCL0","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=2,p0=1,p1=0","support","1","yes","SYCL"
 "SYCL0","POOL_2D","pool_type=max,type_input=f32,ne_input=[10,10,3,1],k0=3,k1=3,s0=2,s1=2,p0=1,p1=1","support","1","yes","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[10,3,2,1],k0=1,s0=1,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[11,1,3,2],k0=1,s0=1,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[128,2,1,3],k0=1,s0=1,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[10,3,2,1],k0=1,s0=1,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[11,1,3,2],k0=1,s0=1,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[128,2,1,3],k0=1,s0=1,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[10,3,2,1],k0=1,s0=2,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[11,1,3,2],k0=1,s0=2,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[128,2,1,3],k0=1,s0=2,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[10,3,2,1],k0=1,s0=2,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[11,1,3,2],k0=1,s0=2,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[128,2,1,3],k0=1,s0=2,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[10,3,2,1],k0=3,s0=1,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[11,1,3,2],k0=3,s0=1,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[128,2,1,3],k0=3,s0=1,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[10,3,2,1],k0=3,s0=1,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[11,1,3,2],k0=3,s0=1,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[128,2,1,3],k0=3,s0=1,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[10,3,2,1],k0=3,s0=2,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[11,1,3,2],k0=3,s0=2,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[128,2,1,3],k0=3,s0=2,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[10,3,2,1],k0=3,s0=2,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[11,1,3,2],k0=3,s0=2,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=avg,type_input=f32,ne_input=[128,2,1,3],k0=3,s0=2,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[10,3,2,1],k0=1,s0=1,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[11,1,3,2],k0=1,s0=1,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[128,2,1,3],k0=1,s0=1,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[10,3,2,1],k0=1,s0=1,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[11,1,3,2],k0=1,s0=1,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[128,2,1,3],k0=1,s0=1,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[10,3,2,1],k0=1,s0=2,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[11,1,3,2],k0=1,s0=2,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[128,2,1,3],k0=1,s0=2,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[10,3,2,1],k0=1,s0=2,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[11,1,3,2],k0=1,s0=2,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[128,2,1,3],k0=1,s0=2,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[10,3,2,1],k0=3,s0=1,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[11,1,3,2],k0=3,s0=1,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[128,2,1,3],k0=3,s0=1,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[10,3,2,1],k0=3,s0=1,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[11,1,3,2],k0=3,s0=1,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[128,2,1,3],k0=3,s0=1,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[10,3,2,1],k0=3,s0=2,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[11,1,3,2],k0=3,s0=2,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[128,2,1,3],k0=3,s0=2,p0=0","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[10,3,2,1],k0=3,s0=2,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[11,1,3,2],k0=3,s0=2,p0=1","support","0","no","SYCL"
+"SYCL0","POOL_1D","pool_type=max,type_input=f32,ne_input=[128,2,1,3],k0=3,s0=2,p0=1","support","0","no","SYCL"
 "SYCL0","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[3000,128,1,1],ne_kernel=[3,128,1280,1],s0=1,s1=0,p0=1,p1=0,d0=1,d1=0,is_2D=0","support","1","yes","SYCL"
 "SYCL0","IM2COL","type_input=f32,type_kernel=f16,dst_type=f32,ne_input=[3000,128,1,1],ne_kernel=[3,128,1280,1],s0=1,s1=0,p0=1,p1=0,d0=1,d1=0,is_2D=0","support","1","yes","SYCL"
 "SYCL0","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[3000,128,1,1],ne_kernel=[3,128,1280,1],s0=1,s1=0,p0=1,p1=0,d0=1,d1=0,is_2D=0","support","1","yes","SYCL"
@@ -965,6 +1013,7 @@
 "SYCL0","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,1,2560],ne_kernel=[3,3,1,2560],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1","support","1","yes","SYCL"
 "SYCL0","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,2,2560],ne_kernel=[3,3,2,2560],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1","support","1","yes","SYCL"
 "SYCL0","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[5,5,1,32],ne_kernel=[3,4,1,32],s0=1,s1=1,p0=0,p1=0,d0=1,d1=1,is_2D=1","support","1","yes","SYCL"
+"SYCL0","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[2,2,1536,729],ne_kernel=[2,2,1536,4096],s0=1,s1=1,p0=0,p1=0,d0=1,d1=1,is_2D=1","support","1","yes","SYCL"
 "SYCL0","IM2COL_3D","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[10,10,10,9],ne_kernel=[3,3,3,1],IC=3,s0=1,s1=1,s2=1,p0=1,p1=1,p2=1,d0=1,d1=1,d2=1,v=0","support","0","no","SYCL"
 "SYCL0","IM2COL_3D","type_input=f32,type_kernel=f16,dst_type=f32,ne_input=[10,10,10,9],ne_kernel=[3,3,3,1],IC=3,s0=1,s1=1,s2=1,p0=1,p1=1,p2=1,d0=1,d1=1,d2=1,v=0","support","0","no","SYCL"
 "SYCL0","IM2COL_3D","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[10,10,10,9],ne_kernel=[3,3,3,1],IC=3,s0=1,s1=1,s2=1,p0=1,p1=1,p2=1,d0=1,d1=1,d2=1,v=0","support","0","no","SYCL"
@@ -5696,35 +5745,58 @@
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000000,inplace=0","support","1","yes","SYCL"
 "SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000000","support","1","yes","SYCL"
 "SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL"
+"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=0,eps=0.000000","support","1","yes","SYCL"
+"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=0,eps=0.000000,inplace=0","support","1","yes","SYCL"
+"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=1,eps=0.000000","support","1","yes","SYCL"
+"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=1,eps=0.000000,inplace=0","support","1","yes","SYCL"
+"SYCL0","RMS_NORM_BACK","type=f32,ne=[1025,5,4,3],eps=0.000000","support","1","yes","SYCL"
+"SYCL0","L2_NORM","type=f32,ne=[1025,5,4,3]","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000001","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000001,inplace=0","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000001","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000001,inplace=0","support","1","yes","SYCL"
 "SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000001","support","1","yes","SYCL"
 "SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL"
+"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=0,eps=0.000001","support","1","yes","SYCL"
+"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=0,eps=0.000001,inplace=0","support","1","yes","SYCL"
+"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=1,eps=0.000001","support","1","yes","SYCL"
+"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=1,eps=0.000001,inplace=0","support","1","yes","SYCL"
+"SYCL0","RMS_NORM_BACK","type=f32,ne=[1025,5,4,3],eps=0.000001","support","1","yes","SYCL"
+"SYCL0","L2_NORM","type=f32,ne=[1025,5,4,3]","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000100","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000100,inplace=0","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000100","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000100,inplace=0","support","1","yes","SYCL"
 "SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000100","support","1","yes","SYCL"
 "SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL"
+"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=0,eps=0.000100","support","1","yes","SYCL"
+"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=0,eps=0.000100,inplace=0","support","1","yes","SYCL"
+"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=1,eps=0.000100","support","1","yes","SYCL"
+"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=1,eps=0.000100,inplace=0","support","1","yes","SYCL"
+"SYCL0","RMS_NORM_BACK","type=f32,ne=[1025,5,4,3],eps=0.000100","support","1","yes","SYCL"
+"SYCL0","L2_NORM","type=f32,ne=[1025,5,4,3]","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.100000","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.100000,inplace=0","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.100000","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.100000,inplace=0","support","1","yes","SYCL"
 "SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.100000","support","1","yes","SYCL"
 "SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL"
+"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=0,eps=0.100000","support","1","yes","SYCL"
+"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=0,eps=0.100000,inplace=0","support","1","yes","SYCL"
+"SYCL0","NORM","type=f32,ne=[1025,5,4,3],v=1,eps=0.100000","support","1","yes","SYCL"
+"SYCL0","RMS_NORM","type=f32,ne=[1025,5,4,3],v=1,eps=0.100000,inplace=0","support","1","yes","SYCL"
+"SYCL0","RMS_NORM_BACK","type=f32,ne=[1025,5,4,3],eps=0.100000","support","1","yes","SYCL"
+"SYCL0","L2_NORM","type=f32,ne=[1025,5,4,3]","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000001,inplace=1","support","1","yes","SYCL"
-"SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL"
-"SYCL0","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","SYCL"
-"SYCL0","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","SYCL"
-"SYCL0","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[3,1024,1,1]","support","1","yes","SYCL"
-"SYCL0","SSM_CONV","type=f32,ne_a=[4,1536,1,1],ne_b=[3,1536,1,1]","support","1","yes","SYCL"
-"SYCL0","SSM_CONV","type=f32,ne_a=[8,1536,1,1],ne_b=[3,1536,1,1]","support","1","yes","SYCL"
-"SYCL0","SSM_CONV","type=f32,ne_a=[4,1536,4,1],ne_b=[3,1536,1,1]","support","1","yes","SYCL"
-"SYCL0","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[3,2048,1,1]","support","1","yes","SYCL"
-"SYCL0","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[3,2048,1,1]","support","1","yes","SYCL"
-"SYCL0","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[3,2048,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[3,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[6,1024,1,1],ne_b=[3,1024,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[3,1024,4,1],ne_b=[3,1024,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[3,1536,1,1],ne_b=[3,1536,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[6,1536,1,1],ne_b=[3,1536,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[3,1536,4,1],ne_b=[3,1536,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[3,2048,1,1],ne_b=[3,2048,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[6,2048,1,1],ne_b=[3,2048,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[3,2048,4,1],ne_b=[3,2048,1,1]","support","1","yes","SYCL"
 "SYCL0","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[4,1024,1,1]","support","1","yes","SYCL"
 "SYCL0","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[4,1024,1,1]","support","1","yes","SYCL"
 "SYCL0","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[4,1024,1,1]","support","1","yes","SYCL"
@@ -5734,6 +5806,15 @@
 "SYCL0","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[4,2048,1,1]","support","1","yes","SYCL"
 "SYCL0","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[4,2048,1,1]","support","1","yes","SYCL"
 "SYCL0","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[4,2048,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[9,1024,1,1],ne_b=[9,1024,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[18,1024,1,1],ne_b=[9,1024,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[9,1024,4,1],ne_b=[9,1024,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[9,1536,1,1],ne_b=[9,1536,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[18,1536,1,1],ne_b=[9,1536,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[9,1536,4,1],ne_b=[9,1536,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[9,2048,1,1],ne_b=[9,2048,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[18,2048,1,1],ne_b=[9,2048,1,1]","support","1","yes","SYCL"
+"SYCL0","SSM_CONV","type=f32,ne_a=[9,2048,4,1],ne_b=[9,2048,1,1]","support","1","yes","SYCL"
 "SYCL0","SSM_SCAN","type=f32,d_state=16,head_dim=1,n_head=1024,n_group=1,n_seq_tokens=32,n_seqs=4","support","0","no","SYCL"
 "SYCL0","SSM_SCAN","type=f32,d_state=128,head_dim=64,n_head=16,n_group=2,n_seq_tokens=32,n_seqs=4","support","0","no","SYCL"
 "SYCL0","SSM_SCAN","type=f32,d_state=256,head_dim=64,n_head=8,n_group=2,n_seq_tokens=32,n_seqs=4","support","0","no","SYCL"
@@ -6593,6 +6674,30 @@
 "SYCL0","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=67,bs=[1,1],nr=[4,1],per=[0,2,1,3],k_v=0,o=1","support","1","yes","SYCL"
 "SYCL0","MUL_MAT","type_a=f32,type_b=f32,m=64,n=77,k=77,bs=[12,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
 "SYCL0","MUL_MAT","type_a=q4_0,type_b=f32,m=576,n=512,k=576,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=q4_0,type_b=f32,m=1,n=2048,k=8192,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=f32,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=f16,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=bf16,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","SYCL"
+"SYCL0","MUL_MAT","type_a=q4_0,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=q4_1,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=q5_0,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=q5_1,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=q8_0,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=mxfp4,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=q2_K,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=q3_K,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=q4_K,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=q5_K,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=q6_K,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=iq2_xs,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=iq2_s,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=iq3_xxs,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=iq1_s,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=iq1_m,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=iq4_nl,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=iq3_s,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
+"SYCL0","MUL_MAT","type_a=iq4_xs,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","SYCL"
 "SYCL0","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=128,bs=[1,1],nr=[1,1],per=[0,2,1,3],k_v=0,o=1","support","1","yes","SYCL"
 "SYCL0","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1056,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=2112,o=1","support","1","yes","SYCL"
 "SYCL0","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=128,bs=[1,1],nr=[1,1],per=[0,2,1,3],k_v=0,o=1","support","0","no","SYCL"
@@ -8917,6 +9022,11 @@
 "SYCL0","SOFT_MAX","type=f32,ne=[32,2,32,1],mask=1,sinks=0,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=0.000000,inplace=0","support","1","yes","SYCL"
 "SYCL0","SOFT_MAX","type=f32,ne=[32,2,32,1],mask=1,sinks=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=8.000000,inplace=0","support","1","yes","SYCL"
 "SYCL0","SOFT_MAX","type=f32,ne=[32,2,32,1],mask=1,sinks=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=8.000000,inplace=0","support","1","yes","SYCL"
+"SYCL0","SOFT_MAX","type=f32,ne=[200001,2,3,1],mask=1,sinks=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=8.000000,inplace=0","support","1","yes","SYCL"
+"SYCL0","SOFT_MAX","type=f32,ne=[200001,2,3,1],mask=1,sinks=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=8.000000,inplace=0","support","1","yes","SYCL"
+"SYCL0","SOFT_MAX","type=f32,ne=[200000,1,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0","support","1","yes","SYCL"
+"SYCL0","SOFT_MAX","type=f32,ne=[200000,4,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0","support","1","yes","SYCL"
+"SYCL0","SOFT_MAX","type=f32,ne=[643251,3,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0","support","1","yes","SYCL"
 "SYCL0","SOFT_MAX_BACK","type=f32,ne=[16,16,1,1],scale=1.000000,max_bias=0.000000","support","1","yes","SYCL"
 "SYCL0","SOFT_MAX_BACK","type=f32,ne=[15,15,1,1],scale=1.000000,max_bias=0.000000","support","1","yes","SYCL"
 "SYCL0","SOFT_MAX_BACK","type=f32,ne=[16,16,2,3],scale=1.000000,max_bias=0.000000","support","1","yes","SYCL"
@@ -8969,6 +9079,7 @@
 "SYCL0","ROPE","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
+"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
@@ -8978,6 +9089,7 @@
 "SYCL0","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
+"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
@@ -8988,11 +9100,13 @@
 "SYCL0","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
+"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
+"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
@@ -9002,6 +9116,7 @@
 "SYCL0","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
+"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
@@ -9012,11 +9127,13 @@
 "SYCL0","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
+"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
+"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
@@ -9026,6 +9143,7 @@
 "SYCL0","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
+"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
@@ -9036,11 +9154,13 @@
 "SYCL0","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
+"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
+"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
@@ -9050,6 +9170,7 @@
 "SYCL0","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
+"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
@@ -9060,6 +9181,7 @@
 "SYCL0","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
+"SYCL0","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","1","yes","SYCL"
@@ -9185,6 +9307,7 @@
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
+"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
@@ -9194,6 +9317,7 @@
 "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
+"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
@@ -9204,11 +9328,13 @@
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
+"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
+"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
@@ -9218,6 +9344,7 @@
 "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
+"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
@@ -9228,11 +9355,13 @@
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
+"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
+"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
@@ -9242,6 +9371,7 @@
 "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
+"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
@@ -9252,11 +9382,13 @@
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
+"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
+"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
@@ -9266,6 +9398,7 @@
 "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
+"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
@@ -9276,6 +9409,7 @@
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
+"SYCL0","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
 "SYCL0","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","SYCL"
@@ -9850,26 +9984,26 @@
 "SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=15,ties=0","support","0","no","SYCL"
 "SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=0","support","1","yes","SYCL"
 "SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=1","support","1","yes","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=nearest,flags=none","support","1","yes","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=nearest,flags=none","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=nearest","support","1","yes","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=nearest","support","1","yes","SYCL"
 "SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear,transpose=0","support","0","no","SYCL"
 "SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear,transpose=1","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear,flags=none","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear,flags=none","support","0","no","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear","support","0","no","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear","support","0","no","SYCL"
 "SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bicubic,transpose=0","support","0","no","SYCL"
 "SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bicubic,transpose=1","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic,flags=none","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bicubic,flags=none","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=513,transpose=0","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=513,transpose=1","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear,flags=none","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear,flags=none","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear,flags=align_corners","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bilinear,flags=align_corners","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bilinear,flags=align_corners","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic,flags=align_corners","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bicubic,flags=align_corners","support","0","no","SYCL"
-"SYCL0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bicubic,flags=align_corners","support","0","no","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic","support","0","no","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bicubic","support","0","no","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear|antialias,transpose=0","support","0","no","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear|antialias,transpose=1","support","0","no","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear|antialias","support","0","no","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear|antialias","support","0","no","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear|align_corners","support","0","no","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bilinear|align_corners","support","0","no","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bilinear|align_corners","support","0","no","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic|align_corners","support","0","no","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bicubic|align_corners","support","0","no","SYCL"
+"SYCL0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bicubic|align_corners","support","0","no","SYCL"
 "SYCL0","SUM","type=f32,ne=[10,5,4,3]","support","1","yes","SYCL"
 "SYCL0","SUM_ROWS","type=f32,ne=[10,5,4,3],permute=0,slice=0","support","1","yes","SYCL"
 "SYCL0","SUM","type=f32,ne=[11,5,6,3],permute=[0,2,1,3]","support","0","no","SYCL"
@@ -9892,8 +10026,9 @@
 "SYCL0","GROUP_NORM","type=f32,ne=[64,64,320,1],num_groups=32,eps=0.000001","support","1","yes","SYCL"
 "SYCL0","GROUP_NORM","type=f32,ne=[9,9,1280,1],num_groups=32,eps=0.000001","support","1","yes","SYCL"
 "SYCL0","ACC","type=f32,ne_a=[256,17,1,1],ne_b=[256,16,1,1]","support","1","yes","SYCL"
-"SYCL0","PAD","type=f32,ne_a=[512,512,1,1],pad_0=1,pad_1=1","support","1","yes","SYCL"
-"SYCL0","PAD","type=f32,ne_a=[512,512,3,1],lp0=1,rp0=1,lp1=1,rp1=1,lp2=1,rp2=1,lp3=1,rp3=1,v=0","support","1","yes","SYCL"
+"SYCL0","PAD","type=f32,ne_a=[512,512,1,1],pad_0=1,pad_1=1,circular=0","support","1","yes","SYCL"
+"SYCL0","PAD","type=f32,ne_a=[33,17,2,1],pad_0=4,pad_1=3,circular=1","support","0","no","SYCL"
+"SYCL0","PAD","type=f32,ne_a=[512,512,3,1],lp0=1,rp0=1,lp1=1,rp1=1,lp2=1,rp2=1,lp3=1,rp3=1,v=0,circular=0","support","1","yes","SYCL"
 "SYCL0","PAD_REFLECT_1D","type=f32,ne_a=[512,34,2,1],pad_0=10,pad_1=9","support","1","yes","SYCL"
 "SYCL0","PAD_REFLECT_1D","type=f32,ne_a=[3000,384,4,1],pad_0=10,pad_1=9","support","1","yes","SYCL"
 "SYCL0","ROLL","shift0=3,shift1=-2,shift3=1,shift4=-1","support","1","yes","SYCL"
@@ -9915,28 +10050,51 @@
 "SYCL0","CUMSUM","type=f32,ne=[2048,5,4,3]","support","0","no","SYCL"
 "SYCL0","CUMSUM","type=f32,ne=[242004,1,1,1]","support","0","no","SYCL"
 "SYCL0","CUMSUM","type=f32,ne=[375960,1,1,1]","support","0","no","SYCL"
+"SYCL0","CUMSUM","type=f32,ne=[20481,4,1,1]","support","0","no","SYCL"
 "SYCL0","XIELU","type=f32,ne=[10,5,4,3]","support","0","no","SYCL"
-"SYCL0","TRI","type=f32,ne=[10,10,4,3],tri_type=3","support","0","no","SYCL"
-"SYCL0","TRI","type=f32,ne=[10,10,4,3],tri_type=2","support","0","no","SYCL"
-"SYCL0","TRI","type=f32,ne=[10,10,4,3],tri_type=1","support","0","no","SYCL"
-"SYCL0","TRI","type=f32,ne=[10,10,4,3],tri_type=0","support","0","no","SYCL"
+"SYCL0","TRI","type=f32,ne=[10,10,4,3],tri_type=3","support","1","yes","SYCL"
+"SYCL0","TRI","type=f32,ne=[10,10,4,3],tri_type=2","support","1","yes","SYCL"
+"SYCL0","TRI","type=f32,ne=[10,10,4,3],tri_type=1","support","1","yes","SYCL"
+"SYCL0","TRI","type=f32,ne=[10,10,4,3],tri_type=0","support","1","yes","SYCL"
 "SYCL0","FILL","type=f32,ne=[10,10,4,3],c=0.000000","support","0","no","SYCL"
 "SYCL0","FILL","type=f32,ne=[303,207,11,3],c=2.000000","support","0","no","SYCL"
 "SYCL0","FILL","type=f32,ne=[800,600,4,4],c=-152.000000","support","0","no","SYCL"
 "SYCL0","FILL","type=f32,ne=[2048,512,2,2],c=3.500000","support","0","no","SYCL"
+"SYCL0","DIAG","type=f32,ne=[10,1,4,3]","support","0","no","SYCL"
+"SYCL0","DIAG","type=f32,ne=[79,1,19,13]","support","0","no","SYCL"
+"SYCL0","DIAG","type=f32,ne=[256,1,8,16]","support","0","no","SYCL"
 "SYCL0","SOLVE_TRI","type=f32,ne_lhs=[10,10,4,3],ne_rhs=[3,10,4,3]","support","0","no","SYCL"
 "SYCL0","SOLVE_TRI","type=f32,ne_lhs=[11,11,1,1],ne_rhs=[5,11,1,1]","support","0","no","SYCL"
 "SYCL0","SOLVE_TRI","type=f32,ne_lhs=[17,17,2,4],ne_rhs=[9,17,2,4]","support","0","no","SYCL"
 "SYCL0","SOLVE_TRI","type=f32,ne_lhs=[30,30,7,1],ne_rhs=[8,30,7,1]","support","0","no","SYCL"
 "SYCL0","SOLVE_TRI","type=f32,ne_lhs=[42,42,5,2],ne_rhs=[10,42,5,2]","support","0","no","SYCL"
 "SYCL0","SOLVE_TRI","type=f32,ne_lhs=[64,64,2,2],ne_rhs=[10,64,2,2]","support","0","no","SYCL"
+"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[64,64,2,2],ne_rhs=[64,64,2,2]","support","0","no","SYCL"
+"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[79,79,5,3],ne_rhs=[417,79,5,3]","support","0","no","SYCL"
+"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[128,128,4,2],ne_rhs=[32,128,4,2]","support","0","no","SYCL"
+"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[80,80,2,8],ne_rhs=[80,80,2,8]","support","0","no","SYCL"
+"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[80,80,2,8],ne_rhs=[79,80,2,8]","support","0","no","SYCL"
+"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[80,80,2,8],ne_rhs=[81,80,2,8]","support","0","no","SYCL"
+"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[80,80,8,8],ne_rhs=[80,80,8,8]","support","0","no","SYCL"
+"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[80,80,8,8],ne_rhs=[79,80,8,8]","support","0","no","SYCL"
+"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[80,80,8,8],ne_rhs=[81,80,8,8]","support","0","no","SYCL"
+"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[84,84,4,4],ne_rhs=[32,84,4,4]","support","0","no","SYCL"
+"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[95,95,8,8],ne_rhs=[40,95,8,8]","support","0","no","SYCL"
 "SYCL0","SOLVE_TRI","type=f32,ne_lhs=[100,100,4,4],ne_rhs=[41,100,4,4]","support","0","no","SYCL"
 "SYCL0","SOLVE_TRI","type=f32,ne_lhs=[128,128,4,4],ne_rhs=[31,128,4,4]","support","0","no","SYCL"
-"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[64,64,4,4],ne_rhs=[300,64,4,4]","support","0","no","SYCL"
-"SYCL0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=0","support","1","yes","SYCL"
-"SYCL0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=0","support","1","yes","SYCL"
-"SYCL0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=1","support","0","no","SYCL"
-"SYCL0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=1","support","0","no","SYCL"
+"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[128,128,4,4],ne_rhs=[32,128,4,4]","support","0","no","SYCL"
+"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[128,128,3,4],ne_rhs=[32,128,3,4]","support","0","no","SYCL"
+"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[128,128,4,1],ne_rhs=[32,128,4,1]","support","0","no","SYCL"
+"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[64,64,4,4],ne_rhs=[200,64,4,4]","support","0","no","SYCL"
+"SYCL0","SOLVE_TRI","type=f32,ne_lhs=[64,64,4,4],ne_rhs=[384,64,4,4]","support","0","no","SYCL"
+"SYCL0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=0,circular=0","support","1","yes","SYCL"
+"SYCL0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=0,circular=0","support","1","yes","SYCL"
+"SYCL0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=0,circular=1","support","0","no","SYCL"
+"SYCL0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=0,circular=1","support","0","no","SYCL"
+"SYCL0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=1,circular=0","support","0","no","SYCL"
+"SYCL0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=1,circular=0","support","0","no","SYCL"
+"SYCL0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=1,circular=1","support","0","no","SYCL"
+"SYCL0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=1,circular=1","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=40,hsv=40,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=40,hsv=40,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=40,hsv=40,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
@@ -14097,86 +14255,86 @@
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
@@ -14337,46 +14495,46 @@
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
@@ -14537,46 +14695,46 @@
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
@@ -14737,46 +14895,46 @@
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=1,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
@@ -15017,86 +15175,86 @@
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,2,1,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,2,1,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
@@ -15257,46 +15415,46 @@
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
@@ -15457,46 +15615,46 @@
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
@@ -15657,46 +15815,46 @@
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=1,sinks=0,max_bias=8.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
@@ -15857,46 +16015,46 @@
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
@@ -16057,46 +16215,46 @@
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=1,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
@@ -16257,46 +16415,46 @@
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
@@ -16457,46 +16615,46 @@
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[4,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
-"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[16,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=1,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=3,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=32,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=f32,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q8_0,permute=[0,1,2,3]","support","0","no","SYCL"
+"SYCL0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=4,nr23=[12,1],kv=512,nb=35,mask=0,sinks=0,max_bias=0.000000,logit_softcap=10.000000,prec=def,type_KV=q4_0,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","SYCL"
 "SYCL0","FLASH_ATTN_EXT","hsk=192,hsv=128,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","SYCL"
diff --git a/docs/speculative.md b/docs/speculative.md
new file mode 100644
index 0000000000..8281eaa2d3
--- /dev/null
+++ b/docs/speculative.md
@@ -0,0 +1,120 @@
+# Speculative Decoding
+
+llama.cpp supports speculative decoding, a technique that can significantly accelerate token generation by predicting multiple tokens ahead of the main model.
+
+[Speculative decoding](https://en.wikipedia.org/wiki/Transformer_(deep_learning)#Speculative_decoding) leverages the fact that computing n tokens in a batch (as in prompt processing) is more efficient than computing n sequentially (as in response generation). By generating draft tokens quickly and then verifying them with the target model in a single batch, this approach can achieve substantial speedups when the draft predictions are frequently correct.
+
+## Implementations
+
+The `llama-server` application supports several implementations of speculative decoding:
+
+### Draft Model (`draft`)
+
+A much smaller model (called the _draft model_) generates drafts.
+A draft model is the most used approach in speculative decoding.
+
+### n-gram Cache (`ngram-cache`)
+
+An n-gram is a sequence of n tokens. The n-gram cache implementation maintains statistics about short n-gram sequences.
+A draft is computed using probabilities derived from these statistics. External statistics can also be loaded from files for improved accuracy.
+
+See:
+
+- #5479, #6828, #6848
+
+### n-gram Map (`ngram-simple`, `ngram-map-*`)
+
+These implementations search the token history for patterns and use matching sequences as draft candidates.
+They require no additional model but rely on patterns that have already appeared in the generated text.
+An example to use this approach can be the rewriting of source code by a LLM.
+
+#### n-gram Map (`ngram-simple`)
+
+This implementation looks for the last n-gram in history that matches the current n-gram and creates a draft using the m tokens following the matched n-gram. It is the simplest self-speculative approach with minimal overhead.
+
+#### n-gram Map Key (`ngram-map-k`)
+
+This implementation looks for the current n-gram of size n (called the _key_) in the token history. If the key n-gram is followed by the same m tokens (called the _mgram_) multiple times, it creates a draft using these m tokens. This approach requires a minimum number of occurrences (argument `--spec-ngram-min-hits`) before generating drafts.
+
+The number of accepted tokens is stored for each used n-gram.
+
+#### n-gram Map Key-4-Values (`ngram-map-k4v`)
+
+This experimental implementation looks for the current n-gram of size n (called the _key_) in the token history. For each key, up to four _values_ (n-grams of size m, called _mgrams_) are tracked. An internal statistic counts the occurrences of each mgram after the key n-gram. If one mgram is significantly more frequent than the others, it is used as the draft.
+
+The number of accepted tokens is stored for each used n-gram.
+
+**Example:** Server options to be used if there are a lot of longer repetitions.
+```bash
+llama-server [...] --spec-type ngram-map-k4v --spec-ngram-size-n 8 --spec-ngram-size-m 8 --spec-ngram-min-hits 2
+```
+
+
+## Command-Line Options
+
+If a draft model is combined with a draftless decoding the draftless decoding has higher precedence.
+
+```
+--spec-type [none|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v]
+                                        type of speculative decoding to use when no draft model is provided
+                                        (default: none)
+--spec-ngram-size-n N                   ngram size N for ngram-simple/ngram-map speculative decoding, length
+                                        of lookup n-gram (default: 12)
+--spec-ngram-size-m N                   ngram size M for ngram-simple/ngram-map speculative decoding, length
+                                        of draft m-gram (default: 48)
+--spec-ngram-check-rate N               ngram check rate for ngram-simple/ngram-map speculative decoding
+                                        (default: 1)
+--spec-ngram-min-hits N                 minimum hits for ngram-map speculative decoding (default: 1)
+```
+
+### `--spec-type TYPE`
+
+Specifies a type of speculative decoding without draft model.
+
+| Type | Description |
+|------|-------------|
+| `none` | No speculative decoding (default) |
+| `ngram-cache` | Use n-gram cache lookup |
+| `ngram-simple` | Use simple n-gram pattern matching |
+| `ngram-map-k` | Use n-gram pattern matching with n-gram-keys |
+| `ngram-map-k4v` | Use n-gram pattern matching with n-gram-keys and up to four m-gram values (experimental) |
+
+**Example:** Server-instance used to refactor source code.
+```bash
+./llama-server [...] --spec-type ngram-simple
+```
+
+### `--spec-ngram-size-n N`
+
+Sets the size N of the lookup n-gram for n-gram map based speculative decoding.
+The n-gram size N determines how many tokens in a row to look back when searching for matching patterns.
+
+### `--spec-ngram-size-m M`
+
+Sets the size M of the draft m-gram for n-gram map based speculative decoding.
+The m-gram size determines how many tokens to draft when a match is found.
+Larger values can provide more speedup but may reduce acceptance rate.
+
+### `--spec-ngram-check-rate R`
+
+This option aims at performance if the n-gram lookup in history is to costly. A lookup will be executed at every R tokens (default is 1, every token).
+
+### `--spec-ngram-min-hits H`
+
+This option defines how often a key has to appear in the token history to be used as a draft (default is 1).
+
+## Statistics
+Each speculative decoding implementation prints statistics.
+
+```
+draft acceptance rate = 0.57576 (  171 accepted /   297 generated)
+statistics ngram_simple: #calls = 15, #gen drafts = 5, #acc drafts = 5, #gen tokens = 187, #acc tokens = 73
+statistics draft: #calls = 10, #gen drafts = 10, #acc drafts = 10, #gen tokens = 110, #acc tokens = 98
+```
+
+- `#calls`: number of calls of this implementations
+- `#gen drafts`: number of drafts generated by this implementation
+- `#acc drafts`: number of drafts accepted (partially) by the main model
+- `#gen tokens`: number of tokens generated by this implementation (including rejected tokens)
+- `#acc tokens`: number of tokens accepted by the main model
+
diff --git a/examples/lookahead/lookahead.cpp b/examples/lookahead/lookahead.cpp
index f54cfdd77f..aa6efa62b3 100644
--- a/examples/lookahead/lookahead.cpp
+++ b/examples/lookahead/lookahead.cpp
@@ -50,6 +50,12 @@ int main(int argc, char ** argv) {
     const int N = 5;  // n-gram size
     const int G = 15; // max verification n-grams
 
+    // lookahead requires W + G + 1 sequences for parallel Jacobi decoding
+    params.n_parallel = W + G + 1;
+
+    // unified KV cache is required for coupled sequences in batch splitting
+    params.kv_unified = true;
+
     // init llama.cpp
     llama_backend_init();
     llama_numa_init(params.numa);
@@ -115,7 +121,7 @@ int main(int argc, char ** argv) {
     // seq_id == 0           : the current input token
     // seq_id [1, W]         : tokens from the past N - 1 Jacobi iterations
     // seq_id [W + 1, W + G] : verification n-grams
-    llama_batch batch = llama_batch_init(params.n_ctx, 0, W + G + 1);
+    llama_batch batch = llama_batch_init(llama_n_ctx(ctx), 0, W + G + 1);
 
     // target model sampling context
     struct common_sampler * smpl = common_sampler_init(model, params.sampling);
diff --git a/examples/lookup/lookup-create.cpp b/examples/lookup/lookup-create.cpp
index bb94a8fe06..f7b6ea1b19 100644
--- a/examples/lookup/lookup-create.cpp
+++ b/examples/lookup/lookup-create.cpp
@@ -32,9 +32,9 @@ int main(int argc, char ** argv){
 
     common_ngram_cache ngram_cache;
     common_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true);
-    fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str());
+    fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.speculative.lookup_cache_static.c_str());
 
-    common_ngram_cache_save(ngram_cache, params.lookup_cache_static);
+    common_ngram_cache_save(ngram_cache, params.speculative.lookup_cache_static);
 
     return 0;
 }
diff --git a/examples/lookup/lookup-stats.cpp b/examples/lookup/lookup-stats.cpp
index 135f6fcab9..ae28b2e6e8 100644
--- a/examples/lookup/lookup-stats.cpp
+++ b/examples/lookup/lookup-stats.cpp
@@ -46,18 +46,18 @@ int main(int argc, char ** argv){
     {
         const int64_t t_start_draft_us = ggml_time_us();
 
-        if (!params.lookup_cache_static.empty()) {
+        if (!params.speculative.lookup_cache_static.empty()) {
             try {
-                ngram_cache_static = common_ngram_cache_load(params.lookup_cache_static);
+                ngram_cache_static = common_ngram_cache_load(params.speculative.lookup_cache_static);
             } catch (std::ifstream::failure const &) {
-                LOG_ERR("failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
+                LOG_ERR("failed to open static lookup cache: %s", params.speculative.lookup_cache_static.c_str());
                 exit(1);
             }
         }
 
-        if (!params.lookup_cache_dynamic.empty()) {
+        if (!params.speculative.lookup_cache_dynamic.empty()) {
             try {
-                ngram_cache_dynamic = common_ngram_cache_load(params.lookup_cache_dynamic);
+                ngram_cache_dynamic = common_ngram_cache_load(params.speculative.lookup_cache_dynamic);
             } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
         }
 
diff --git a/examples/lookup/lookup.cpp b/examples/lookup/lookup.cpp
index 27f159940a..c7552ddde1 100644
--- a/examples/lookup/lookup.cpp
+++ b/examples/lookup/lookup.cpp
@@ -51,18 +51,18 @@ int main(int argc, char ** argv){
         const int64_t t_start_draft_us = ggml_time_us();
         common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, inp.size(), false);
 
-        if (!params.lookup_cache_static.empty()) {
+        if (!params.speculative.lookup_cache_static.empty()) {
             try {
-                ngram_cache_static = common_ngram_cache_load(params.lookup_cache_static);
+                ngram_cache_static = common_ngram_cache_load(params.speculative.lookup_cache_static);
             } catch (std::ifstream::failure const &) {
-                LOG_ERR("failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
+                LOG_ERR("failed to open static lookup cache: %s", params.speculative.lookup_cache_static.c_str());
                 exit(1);
             }
         }
 
-        if (!params.lookup_cache_dynamic.empty()) {
+        if (!params.speculative.lookup_cache_dynamic.empty()) {
             try {
-                ngram_cache_dynamic = common_ngram_cache_load(params.lookup_cache_dynamic);
+                ngram_cache_dynamic = common_ngram_cache_load(params.speculative.lookup_cache_dynamic);
             } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
         }
 
@@ -106,7 +106,7 @@ int main(int argc, char ** argv){
 
     std::vector<llama_token> draft;
 
-    llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, 1);
+    llama_batch batch_tgt = llama_batch_init(llama_n_ctx(ctx), 0, 1);
 
     const auto t_dec_start = ggml_time_us();
 
@@ -210,7 +210,7 @@ int main(int argc, char ** argv){
 
     // Update dynamic ngram cache with context ngram cache and save it to disk:
     common_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
-    common_ngram_cache_save(ngram_cache_dynamic, params.lookup_cache_dynamic);
+    common_ngram_cache_save(ngram_cache_dynamic, params.speculative.lookup_cache_dynamic);
 
     LOG("\n\n");
 
diff --git a/examples/model-conversion/scripts/causal/run-converted-model-embeddings-logits.sh b/examples/model-conversion/scripts/causal/run-converted-model-embeddings-logits.sh
index 3cce3fc94d..1b5ff8611b 100755
--- a/examples/model-conversion/scripts/causal/run-converted-model-embeddings-logits.sh
+++ b/examples/model-conversion/scripts/causal/run-converted-model-embeddings-logits.sh
@@ -4,6 +4,7 @@ set -e
 
 # First try command line argument, then environment variable, then file
 CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+BUILD_DIR="${2:-"$BUILD_DIR"}"
 
 # Final check if we have a model path
 if [ -z "$CONVERTED_MODEL" ]; then
@@ -13,6 +14,10 @@ if [ -z "$CONVERTED_MODEL" ]; then
     exit 1
 fi
 
-cmake --build ../../build --target llama-debug -j8
+if [ -z "$BUILD_DIR" ]; then
+    BUILD_DIR="../../build"
+fi
 
-../../build/bin/llama-debug -m $CONVERTED_MODEL --embedding -p "Hello world today" --save-logits
+cmake --build ${BUILD_DIR} --target llama-debug -j8
+
+${BUILD_DIR}/bin/llama-debug -m $CONVERTED_MODEL --embedding -p "Hello world today" --save-logits
diff --git a/examples/model-conversion/scripts/causal/run-converted-model.sh b/examples/model-conversion/scripts/causal/run-converted-model.sh
index b6c3d38662..b684804e02 100755
--- a/examples/model-conversion/scripts/causal/run-converted-model.sh
+++ b/examples/model-conversion/scripts/causal/run-converted-model.sh
@@ -5,11 +5,16 @@ set -e
 # First try command line argument, then environment variable, then file
 CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
 MODEL_TESTING_PROMPT="${2:-"$MODEL_TESTING_PROMPT"}"
+BUILD_DIR="${3:-"$BUILD_DIR"}"
 
-if [ -z "$MODEL_TESTING_PROMPT"]; then
+if [ -z "$MODEL_TESTING_PROMPT" ]; then
     MODEL_TESTING_PROMPT="Hello, my name is"
 fi
 
+if [ -z "$BUILD_DIR" ]; then
+    BUILD_DIR="../../build"
+fi
+
 # Final check if we have a model path
 if [ -z "$CONVERTED_MODEL" ]; then
     echo "Error: Model path must be provided either as:" >&2
@@ -21,6 +26,6 @@ fi
 echo $CONVERTED_MODEL
 echo $MODEL_TESTING_PROMPT
 
-cmake --build ../../build --target llama-debug -j8
+cmake --build ${BUILD_DIR} --target llama-debug -j8
 
-../../build/bin/llama-debug -m "$CONVERTED_MODEL" -p "$MODEL_TESTING_PROMPT" --save-logits
+${BUILD_DIR}/bin/llama-debug -m "$CONVERTED_MODEL" -p "$MODEL_TESTING_PROMPT" --save-logits
diff --git a/examples/model-conversion/scripts/embedding/run-converted-model.sh b/examples/model-conversion/scripts/embedding/run-converted-model.sh
index 84625cec3d..ba8a3afae6 100755
--- a/examples/model-conversion/scripts/embedding/run-converted-model.sh
+++ b/examples/model-conversion/scripts/embedding/run-converted-model.sh
@@ -28,6 +28,7 @@ done
 
 # First try command line argument, then environment variable
 CONVERTED_MODEL="${CONVERTED_MODEL:-"$CONVERTED_EMBEDDING_MODEL"}"
+BUILD_DIR="${BUILD_DIR:-"../../build"}"
 
 # Final check if we have a model path
 if [ -z "$CONVERTED_MODEL" ]; then
@@ -50,5 +51,5 @@ fi
 
 echo $CONVERTED_MODEL
 
-cmake --build ../../build --target llama-debug -j8
-../../build/bin/llama-debug -m "$CONVERTED_MODEL" --embedding -p "$PROMPT" --save-logits --embd-normalize $EMBD_NORMALIZE
+cmake --build ${BUILD_DIR} --target llama-debug -j8
+${BUILD_DIR}/bin/llama-debug -m "$CONVERTED_MODEL" --embedding -p "$PROMPT" --save-logits --embd-normalize $EMBD_NORMALIZE
diff --git a/examples/model-conversion/scripts/utils/perplexity-gen.sh b/examples/model-conversion/scripts/utils/perplexity-gen.sh
index 4885acbae2..ef4b650fda 100755
--- a/examples/model-conversion/scripts/utils/perplexity-gen.sh
+++ b/examples/model-conversion/scripts/utils/perplexity-gen.sh
@@ -3,6 +3,7 @@
 set -e
 
 CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+BUILD_DIR="${2:-"$BUILD_DIR"}"
 
 # Final check if we have a model path
 if [ -z "$CONVERTED_MODEL" ]; then
@@ -25,9 +26,13 @@ mkdir -p ppl
 OUTPUTFILE="ppl/$(basename $CONVERTED_MODEL).kld"
 echo "Model: $CONVERTED_MODEL"
 
-cmake --build ../../build --target llama-perplexity -j8
+if [ -z "$BUILD_DIR" ]; then
+    BUILD_DIR="../../build"
+fi
 
-../.././build/bin/llama-perplexity -m $CONVERTED_MODEL \
+cmake --build $BUILD_DIR --target llama-perplexity -j8
+
+${BUILD_DIR}/bin/llama-perplexity -m $CONVERTED_MODEL \
     -f ppl/wikitext-2-raw/wiki.test.raw \
     --kl-divergence-base $OUTPUTFILE
 
diff --git a/examples/model-conversion/scripts/utils/perplexity-run-simple.sh b/examples/model-conversion/scripts/utils/perplexity-run-simple.sh
index a2545436a5..20ee9653a9 100755
--- a/examples/model-conversion/scripts/utils/perplexity-run-simple.sh
+++ b/examples/model-conversion/scripts/utils/perplexity-run-simple.sh
@@ -3,6 +3,7 @@
 set -e
 
 QUANTIZED_MODEL="${1:-"$QUANTIZED_MODEL"}"
+BUILD_DIR="${2:-"$BUILD_DIR"}"
 
 if [ -z "$QUANTIZED_MODEL" ]; then
     echo "Error: Model path must be provided either as:" >&2
@@ -20,8 +21,12 @@ if [ ! -d "ppl/wikitext-2-raw" ]; then
     popd
 fi
 
-cmake --build ../../build --target llama-perplexity -j8
+if [ -z "$BUILD_DIR" ]; then
+    BUILD_DIR="../../build"
+fi
 
-../.././build/bin/llama-perplexity -m $QUANTIZED_MODEL -f ppl/wikitext-2-raw/wiki.test.raw
+cmake --build $BUILD_DIR --target llama-perplexity -j8
+
+${BUILD_DIR}/bin/llama-perplexity -m $QUANTIZED_MODEL -f ppl/wikitext-2-raw/wiki.test.raw
 
 
diff --git a/examples/model-conversion/scripts/utils/perplexity-run.sh b/examples/model-conversion/scripts/utils/perplexity-run.sh
index 68b38e6628..c11f32c65f 100755
--- a/examples/model-conversion/scripts/utils/perplexity-run.sh
+++ b/examples/model-conversion/scripts/utils/perplexity-run.sh
@@ -3,7 +3,8 @@
 set -e
 
 QUANTIZED_MODEL="${1:-"$QUANTIZED_MODEL"}"
-LOGITS_FILE="${1:-"$LOGITS_FILE"}"
+LOGITS_FILE="${2:-"$LOGITS_FILE"}"
+BUILD_DIR="${3:-"$BUILD_DIR"}"
 
 if [ -z "$QUANTIZED_MODEL" ]; then
     echo "Error: Model path must be provided either as:" >&2
@@ -18,11 +19,15 @@ if [ ! -f ${LOGITS_FILE} ]; then
     exit 1
 fi
 
+if [ -z "$BUILD_DIR" ]; then
+    BUILD_DIR="../../build"
+fi
+
 echo "Model: $QUANTIZED_MODEL"
 echo "Data file: $LOGITS_FILE"
 
-cmake --build ../../build --target llama-perplexity -j8
+cmake --build $BUILD_DIR --target llama-perplexity -j8
 
-../.././build/bin/llama-perplexity -m $QUANTIZED_MODEL \
+${BUILD_DIR}/bin/llama-perplexity -m $QUANTIZED_MODEL \
     --kl-divergence-base $LOGITS_FILE \
     --kl-divergence
diff --git a/examples/model-conversion/scripts/utils/quantize.sh b/examples/model-conversion/scripts/utils/quantize.sh
index c25c5c21f3..4c21a1345a 100755
--- a/examples/model-conversion/scripts/utils/quantize.sh
+++ b/examples/model-conversion/scripts/utils/quantize.sh
@@ -6,6 +6,7 @@ CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
 QUANTIZED_TYPE="${2:-"$QUANTIZED_TYPE"}"
 TOKEN_EMBD_TYPE="${3:-"${TOKEN_EMBD_TYPE}"}"
 OUTPUT_TYPE="${4:-"${OUTPUT_TYPE}"}"
+BUILD_DIR="${5:-"$BUILD_DIR"}"
 QUANTIZED_MODEL=$CONVERTED_MODEL
 
 # Final check if we have a model path
@@ -33,12 +34,16 @@ else
     exit 1
 fi
 
-cmake --build ../../build --target llama-quantize -j8
+if [ -z "$BUILD_DIR" ]; then
+    BUILD_DIR="../../build"
+fi
+
+cmake --build $BUILD_DIR --target llama-quantize -j8
 
 echo $TOKEN_EMBD_TYPE
 echo $OUTPUT_TYPE
 
-CMD_ARGS=("../../build/bin/llama-quantize")
+CMD_ARGS=("${BUILD_DIR}/bin/llama-quantize")
 [[ -n "$TOKEN_EMBD_TYPE" ]] && CMD_ARGS+=("--token-embedding-type" "$TOKEN_EMBD_TYPE")
 [[ -n "$OUTPUT_TYPE" ]]     && CMD_ARGS+=("--output-tensor-type" "$OUTPUT_TYPE")
 CMD_ARGS+=("$CONVERTED_MODEL" "$QUANTIZED_MODEL" "$QUANTIZED_TYPE")
diff --git a/examples/model-conversion/scripts/utils/run-embedding-server.sh b/examples/model-conversion/scripts/utils/run-embedding-server.sh
index d30b765964..9f5fc2cf70 100755
--- a/examples/model-conversion/scripts/utils/run-embedding-server.sh
+++ b/examples/model-conversion/scripts/utils/run-embedding-server.sh
@@ -4,6 +4,7 @@ set -e
 #
 # First try command line argument, then environment variable, then file
 CONVERTED_MODEL="${1:-"$CONVERTED_MODEL"}"
+BUILD_DIR="${2:-"$BUILD_DIR"}"
 
 # Final check if we have a model path
 if [ -z "$CONVERTED_MODEL" ]; then
@@ -13,10 +14,14 @@ if [ -z "$CONVERTED_MODEL" ]; then
     exit 1
 fi
 
+if [ -z "$BUILD_DIR" ]; then
+    BUILD_DIR="../../build"
+fi
+
 echo $CONVERTED_MODEL
 
-cmake --build ../../build --target llama-server
+cmake --build $BUILD_DIR --target llama-server
 
-../../build/bin/llama-server -m $CONVERTED_MODEL \
+${BUILD_DIR}/bin/llama-server -m $CONVERTED_MODEL \
     --embedding \
     --pooling none
diff --git a/examples/speculative-simple/speculative-simple.cpp b/examples/speculative-simple/speculative-simple.cpp
index 8141052a22..d8b1f5a480 100644
--- a/examples/speculative-simple/speculative-simple.cpp
+++ b/examples/speculative-simple/speculative-simple.cpp
@@ -24,7 +24,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    if (params.speculative.model.path.empty()) {
+    if (params.speculative.mparams_dft.path.empty()) {
         LOG_ERR("%s: --model-draft is required\n", __func__);
         return 1;
     }
@@ -34,10 +34,8 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     llama_model * model_tgt = NULL;
-    //llama_model * model_dft = NULL;
 
     llama_context * ctx_tgt = NULL;
-    llama_context * ctx_dft = NULL;
 
     // load the target model
     auto llama_init_tgt = common_init_from_params(params);
@@ -48,26 +46,38 @@ int main(int argc, char ** argv) {
     const llama_vocab * vocab = llama_model_get_vocab(model_tgt);
 
     // load the draft model
-    params.devices      = params.speculative.devices;
-    params.model        = params.speculative.model;
-    params.n_ctx        = params.speculative.n_ctx;
-    params.n_batch      = params.speculative.n_ctx > 0 ? params.speculative.n_ctx : params.n_batch;
-    params.n_gpu_layers = params.speculative.n_gpu_layers;
+    llama_model_ptr model_dft;
 
-    if (params.speculative.cpuparams.n_threads > 0) {
-        params.cpuparams.n_threads = params.speculative.cpuparams.n_threads;
-    }
+    // TODO: simplify this logic
+    {
+        const auto & params_spec = params.speculative;
 
-    params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
-    params.tensor_buft_overrides     = params.speculative.tensor_buft_overrides;
+        auto params_dft = params;
 
-    auto llama_init_dft = common_init_from_params(params);
+        params_dft.n_parallel   = 1;
+        params_dft.n_ctx        = params_spec.n_ctx;
+        params_dft.n_batch      = llama_n_ctx_seq(ctx_tgt);
+        params_dft.devices      = params_spec.devices;
+        params_dft.model        = params_spec.mparams_dft;
+        params_dft.n_gpu_layers = params_spec.n_gpu_layers;
 
-    //model_dft = llama_init_dft->model();
-    ctx_dft   = llama_init_dft->context();
+        if (params_spec.cpuparams.n_threads > 0) {
+            params_dft.cpuparams.n_threads       = params.speculative.cpuparams.n_threads;
+            params_dft.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
+        }
 
-    if (!common_speculative_are_compatible(ctx_tgt, ctx_dft)) {
-        LOG_INF("the draft model '%s' is not compatible with the target model '%s'. tokens will be translated between the draft and target models.\n", params.speculative.model.path.c_str(), params.model.path.c_str());
+        params_dft.tensor_buft_overrides = params.speculative.tensor_buft_overrides;
+
+        auto mparams_dft = common_model_params_to_llama(params_dft);
+
+        model_dft.reset(llama_model_load_from_file(params_dft.model.path.c_str(), mparams_dft));
+        if (model_dft == nullptr) {
+            LOG_ERR("failed to load draft model, '%s'\n", params_dft.model.path.c_str());
+            return 1;
+        }
+
+        params.speculative.model_dft = model_dft.get();
+        params.speculative.cparams_dft = common_context_params_to_llama(params_dft);
     }
 
     // Tokenize the prompt
@@ -92,12 +102,6 @@ int main(int argc, char ** argv) {
         LOG("%s", common_token_to_piece(ctx_tgt, id).c_str());
     }
 
-    // how many tokens to draft each time
-    int n_draft     = params.speculative.n_max;
-    int n_draft_min = params.speculative.n_min;
-
-    float p_min = params.speculative.p_min;
-
     int n_predict = 0;
     int n_drafted = 0;
     int n_accept  = 0;
@@ -127,15 +131,11 @@ int main(int argc, char ** argv) {
     int n_past = inp.size() - 1;
 
     // init the speculator
-    struct common_speculative_params params_spec;
-    params_spec.n_draft = n_draft;
-    params_spec.n_reuse = llama_n_ctx(ctx_dft) - n_draft;
-    params_spec.p_min   = p_min;
+    const auto & params_spec = params.speculative;
 
-    struct common_speculative * spec = common_speculative_init(ctx_tgt, ctx_dft);
-    for (auto &pair : params.speculative.replacements) {
-        common_speculative_add_replacement_tgt_dft(spec, pair.first.c_str(), pair.second.c_str());
-    }
+    struct common_speculative * spec = common_speculative_init(params.speculative, ctx_tgt);
+
+    common_speculative_begin(spec, prompt_tgt);
 
     llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, 1);
 
@@ -151,7 +151,7 @@ int main(int argc, char ** argv) {
         // offloaded to a remote device. it doesn't even have to be based on an LLM. instead, it can provide tokens
         // from a cache or lookup tables.
         //
-        llama_tokens draft = common_speculative_gen_draft(spec, params_spec, prompt_tgt, id_last);
+        llama_tokens draft = common_speculative_draft(spec, params_spec, prompt_tgt, id_last);
 
         //LOG_DBG("draft: %s\n", string_from(ctx_dft, draft).c_str());
 
@@ -162,7 +162,7 @@ int main(int argc, char ** argv) {
         // evaluate the target model on [id_last, draft0, draft1, ..., draftN-1]
         {
             // do not waste time on small drafts
-            if (draft.size() < (size_t) n_draft_min) {
+            if (draft.size() < (size_t) params_spec.n_min) {
                 draft.clear();
             }
 
@@ -240,7 +240,7 @@ int main(int argc, char ** argv) {
     LOG_INF("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict  / ((t_dec_end - t_dec_start) / 1e6f));
 
     LOG_INF("\n");
-    LOG_INF("n_draft   = %d\n", n_draft);
+    LOG_INF("n_draft   = %d\n", params_spec.n_max);
     LOG_INF("n_predict = %d\n", n_predict);
     LOG_INF("n_drafted = %d\n", n_drafted);
     LOG_INF("n_accept  = %d\n", n_accept);
@@ -249,8 +249,6 @@ int main(int argc, char ** argv) {
     LOG_INF("\n");
     LOG_INF("draft:\n\n");
 
-    llama_perf_context_print(ctx_dft);
-
     LOG_INF("\n");
     LOG_INF("target:\n\n");
     common_perf_print(ctx_tgt, smpl);
diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp
index 89d3249431..3e5cf5f46b 100644
--- a/examples/speculative/speculative.cpp
+++ b/examples/speculative/speculative.cpp
@@ -46,7 +46,7 @@ int main(int argc, char ** argv) {
 
     common_init();
 
-    if (params.speculative.model.path.empty()) {
+    if (params.speculative.mparams_dft.path.empty()) {
         LOG_ERR("%s: --model-draft is required\n", __func__);
         return 1;
     }
@@ -78,7 +78,7 @@ int main(int argc, char ** argv) {
 
     // load the draft model
     params.devices = params.speculative.devices;
-    params.model = params.speculative.model;
+    params.model = params.speculative.mparams_dft;
     params.n_gpu_layers = params.speculative.n_gpu_layers;
     if (params.speculative.cpuparams.n_threads > 0) {
         params.cpuparams.n_threads = params.speculative.cpuparams.n_threads;
diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index 0176ca1ce9..b0b8e57898 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -228,6 +228,8 @@ option(GGML_WEBGPU_CPU_PROFILE              "ggml: enable WebGPU profiling (CPU)
 option(GGML_WEBGPU_GPU_PROFILE              "ggml: enable WebGPU profiling (GPU)"             OFF)
 option(GGML_WEBGPU_JSPI                     "ggml: use JSPI for WebGPU"                       ON)
 option(GGML_ZDNN                            "ggml: use zDNN"                                  OFF)
+option(GGML_VIRTGPU                         "ggml: use the VirtGPU/Virglrenderer API Remoting frontend"     OFF)
+option(GGML_VIRTGPU_BACKEND                 "ggml: build the VirtGPU/Virglrenderer API Remoting backend"    OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
 option(GGML_METAL_SHADER_DEBUG              "ggml: compile Metal with -fno-fast-math"         OFF)
@@ -320,6 +322,7 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-opt.h
     include/ggml-metal.h
     include/ggml-rpc.h
+    include/ggml-virtgpu.h
     include/ggml-sycl.h
     include/ggml-vulkan.h
     include/ggml-webgpu.h
diff --git a/ggml/include/ggml-virtgpu.h b/ggml/include/ggml-virtgpu.h
new file mode 100644
index 0000000000..1cb4bd7a03
--- /dev/null
+++ b/ggml/include/ggml-virtgpu.h
@@ -0,0 +1,16 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_REMOTING_FRONTEND_NAME "RemotingFrontend"
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_virtgpu_reg();
+
+#ifdef  __cplusplus
+}
+#endif
diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index b69583dd3f..1988d16dc4 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -630,10 +630,11 @@ extern "C" {
 
     // this tensor...
     enum ggml_tensor_flag {
-        GGML_TENSOR_FLAG_INPUT  =  1, // ...is an input for the GGML compute graph
-        GGML_TENSOR_FLAG_OUTPUT =  2, // ...is an output for the GGML compute graph
-        GGML_TENSOR_FLAG_PARAM  =  4, // ...contains trainable parameters
-        GGML_TENSOR_FLAG_LOSS   =  8, // ...defines loss for numerical optimization (multiple loss tensors add up)
+        GGML_TENSOR_FLAG_INPUT   =  1, // ...is an input for the GGML compute graph
+        GGML_TENSOR_FLAG_OUTPUT  =  2, // ...is an output for the GGML compute graph
+        GGML_TENSOR_FLAG_PARAM   =  4, // ...contains trainable parameters
+        GGML_TENSOR_FLAG_LOSS    =  8, // ...defines loss for numerical optimization (multiple loss tensors add up)
+        GGML_TENSOR_FLAG_COMPUTE = 16, // ...must be computed
     };
 
     enum ggml_tri_type {
@@ -2577,11 +2578,42 @@ extern "C" {
         struct ggml_tensor *  grad,
         struct ggml_tensor *  sgd_params); // alpha, weight decay
 
+    // build forward mutiple tensors and select one of them for computing
+    // this is useful for creating graphs that have constant topology but compute different things based on the input
+    // ref: https://github.com/ggml-org/llama.cpp/pull/18550
     //
-    // automatic differentiation
+    // nodes:
+    //   | - build forward into the graph but do not compute
+    //   c - build forward into the graph and compute
     //
+    //    |  |  ...  c  ...  |
+    //    |  |  ...  c  ...  |
+    //    |  |  ...  c  ...  |
+    //   [0  1  ... idx ...  n-1]        <-- ggml_build_forward_select(..., n, idx)
+    //               c
+    //               c
+    //
+    // example:
+    //   struct ggml_tensor * curs[3];
+    //
+    //   curs[0]  = compute0(...);
+    //   curs[1]  = compute1(...);
+    //   curs[2]  = compute2(...);
+    //
+    //   int idx = select_branch(some_input);
+    //
+    //   struct ggml_tensor * out = ggml_build_forward_select(cgraph, curs, 3, idx);
+    //
+    GGML_API struct ggml_tensor * ggml_build_forward_select(
+            struct ggml_cgraph  * cgraph,
+            struct ggml_tensor ** tensors,
+            int                   n_tensors,
+            int                   idx);
+
+    GGML_API void ggml_build_forward_expand(
+            struct ggml_cgraph * cgraph,
+            struct ggml_tensor * tensor);
 
-    GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
     GGML_API void ggml_build_backward_expand(
         struct ggml_context *  ctx,        // context for gradient computation
         struct ggml_cgraph  *  cgraph,
@@ -2613,7 +2645,7 @@ extern "C" {
     GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph);
 
     // dump the graph into a file using the dot format
-    GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename);
+    GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * cgraph, const char * filename);
 
     // TODO these functions were sandwiched in the old optimization interface, is there a better place for them?
     typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data);
diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
index 6192a87046..265023733e 100644
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -222,6 +222,7 @@ if (GGML_SCHED_NO_REALLOC)
 endif()
 
 add_library(ggml
+            ggml-backend-dl.cpp
             ggml-backend-reg.cpp)
 add_library(ggml::ggml ALIAS ggml)
 
@@ -451,6 +452,7 @@ ggml_add_backend(HIP)
 ggml_add_backend(METAL)
 ggml_add_backend(MUSA)
 ggml_add_backend(RPC)
+ggml_add_backend(VirtGPU)
 ggml_add_backend(SYCL)
 ggml_add_backend(Vulkan)
 ggml_add_backend(WebGPU)
diff --git a/ggml/src/ggml-backend-dl.cpp b/ggml/src/ggml-backend-dl.cpp
new file mode 100644
index 0000000000..a65cf00905
--- /dev/null
+++ b/ggml/src/ggml-backend-dl.cpp
@@ -0,0 +1,48 @@
+#include "ggml-backend-dl.h"
+
+#ifdef _WIN32
+
+dl_handle * dl_load_library(const fs::path & path) {
+    // suppress error dialogs for missing DLLs
+    DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
+    SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
+
+    HMODULE handle = LoadLibraryW(path.wstring().c_str());
+
+    SetErrorMode(old_mode);
+
+    return handle;
+}
+
+void * dl_get_sym(dl_handle * handle, const char * name) {
+    DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
+    SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
+
+    void * p = (void *) GetProcAddress(handle, name);
+
+    SetErrorMode(old_mode);
+
+    return p;
+}
+
+const char * dl_error() {
+    return "";
+}
+
+#else
+
+dl_handle * dl_load_library(const fs::path & path) {
+    dl_handle * handle = dlopen(path.string().c_str(), RTLD_NOW | RTLD_LOCAL);
+    return handle;
+}
+
+void * dl_get_sym(dl_handle * handle, const char * name) {
+    return dlsym(handle, name);
+}
+
+const char * dl_error() {
+    const char *rslt = dlerror();
+    return rslt != nullptr ? rslt : "";
+}
+
+#endif
diff --git a/ggml/src/ggml-backend-dl.h b/ggml/src/ggml-backend-dl.h
new file mode 100644
index 0000000000..f74b7c9489
--- /dev/null
+++ b/ggml/src/ggml-backend-dl.h
@@ -0,0 +1,45 @@
+#pragma once
+
+#ifdef _WIN32
+#   define WIN32_LEAN_AND_MEAN
+#   ifndef NOMINMAX
+#       define NOMINMAX
+#   endif
+#   include <windows.h>
+#   include <winevt.h>
+#else
+#    include <dlfcn.h>
+#    include <unistd.h>
+#endif
+#include <filesystem>
+
+namespace fs = std::filesystem;
+
+#ifdef _WIN32
+
+using dl_handle = std::remove_pointer_t<HMODULE>;
+
+struct dl_handle_deleter {
+    void operator()(HMODULE handle) {
+        FreeLibrary(handle);
+    }
+};
+
+#else
+
+using dl_handle = void;
+
+struct dl_handle_deleter {
+    void operator()(void * handle) {
+        dlclose(handle);
+    }
+};
+
+#endif
+
+using dl_handle_ptr = std::unique_ptr<dl_handle, dl_handle_deleter>;
+
+dl_handle * dl_load_library(const fs::path & path);
+void * dl_get_sym(dl_handle * handle, const char * name);
+const char * dl_error();
+
diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
index 4181a714ad..8a693f84af 100644
--- a/ggml/src/ggml-backend-reg.cpp
+++ b/ggml/src/ggml-backend-reg.cpp
@@ -1,5 +1,6 @@
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"
+#include "ggml-backend-dl.h"
 #include "ggml-impl.h"
 #include <algorithm>
 #include <cstring>
@@ -69,6 +70,10 @@
 #include "ggml-rpc.h"
 #endif
 
+#ifdef GGML_USE_VIRTGPU_FRONTEND
+#include "ggml-virtgpu.h"
+#endif
+
 #ifdef GGML_USE_CANN
 #include "ggml-cann.h"
 #endif
@@ -77,105 +82,23 @@
 #include "ggml-zendnn.h"
 #endif
 
-// disable C++17 deprecation warning for std::codecvt_utf8
-#if defined(__clang__)
-#    pragma clang diagnostic push
-#    pragma clang diagnostic ignored "-Wdeprecated-declarations"
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic push
-#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
-#endif
-
 namespace fs = std::filesystem;
 
 static std::string path_str(const fs::path & path) {
-    std::string u8path;
     try {
 #if defined(__cpp_lib_char8_t)
         // C++20 and later: u8string() returns std::u8string
-        std::u8string u8str = path.u8string();
-        u8path = std::string(reinterpret_cast<const char*>(u8str.c_str()));
+        const std::u8string u8str = path.u8string();
+        return std::string(reinterpret_cast<const char *>(u8str.data()), u8str.size());
 #else
         // C++17: u8string() returns std::string
-        u8path = path.u8string();
+        return path.u8string();
 #endif
     } catch (...) {
+        return std::string();
     }
-    return u8path;
 }
 
-#if defined(__clang__)
-#    pragma clang diagnostic pop
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic pop
-#endif
-
-#ifdef _WIN32
-
-using dl_handle = std::remove_pointer_t<HMODULE>;
-
-struct dl_handle_deleter {
-    void operator()(HMODULE handle) {
-        FreeLibrary(handle);
-    }
-};
-
-static dl_handle * dl_load_library(const fs::path & path) {
-    // suppress error dialogs for missing DLLs
-    DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
-    SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
-
-    HMODULE handle = LoadLibraryW(path.wstring().c_str());
-
-    SetErrorMode(old_mode);
-
-    return handle;
-}
-
-static void * dl_get_sym(dl_handle * handle, const char * name) {
-    DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
-    SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
-
-    void * p = (void *) GetProcAddress(handle, name);
-
-    SetErrorMode(old_mode);
-
-    return p;
-}
-
-static const char * dl_error() {
-    return "";
-}
-
-#else
-
-using dl_handle = void;
-
-struct dl_handle_deleter {
-    void operator()(void * handle) {
-        dlclose(handle);
-    }
-};
-
-static void * dl_load_library(const fs::path & path) {
-    dl_handle * handle = dlopen(path.string().c_str(), RTLD_NOW | RTLD_LOCAL);
-
-    return handle;
-}
-
-static void * dl_get_sym(dl_handle * handle, const char * name) {
-    return dlsym(handle, name);
-}
-
-static const char * dl_error() {
-    const char *rslt = dlerror();
-    return rslt != nullptr ? rslt : "";
-}
-
-#endif
-
-using dl_handle_ptr = std::unique_ptr<dl_handle, dl_handle_deleter>;
-
 struct ggml_backend_reg_entry {
     ggml_backend_reg_t reg;
     dl_handle_ptr handle;
@@ -196,7 +119,12 @@ struct ggml_backend_registry {
         register_backend(ggml_backend_sycl_reg());
 #endif
 #ifdef GGML_USE_VULKAN
+    // Add runtime disable check
+    if (getenv("GGML_DISABLE_VULKAN") == nullptr) {
         register_backend(ggml_backend_vk_reg());
+    } else {
+        GGML_LOG_DEBUG("Vulkan backend disabled by GGML_DISABLE_VULKAN environment variable\n");
+    }
 #endif
 #ifdef GGML_USE_WEBGPU
         register_backend(ggml_backend_webgpu_reg());
@@ -204,6 +132,10 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_ZDNN
         register_backend(ggml_backend_zdnn_reg());
 #endif
+#ifdef GGML_USE_VIRTGPU_FRONTEND
+        register_backend(ggml_backend_virtgpu_reg());
+#endif
+
 #ifdef GGML_USE_OPENCL
         register_backend(ggml_backend_opencl_reg());
 #endif
@@ -620,6 +552,7 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
     ggml_backend_load_best("rpc", silent, dir_path);
     ggml_backend_load_best("sycl", silent, dir_path);
     ggml_backend_load_best("vulkan", silent, dir_path);
+    ggml_backend_load_best("virtgpu", silent, dir_path);
     ggml_backend_load_best("opencl", silent, dir_path);
     ggml_backend_load_best("hexagon", silent, dir_path);
     ggml_backend_load_best("musa", silent, dir_path);
diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index 1b59924b8c..354876574a 100644
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -874,9 +874,9 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
         }
         if (sched->debug > 1) {
             ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
-            GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s] use=%d:", i, ggml_op_name(node->op), node->name,
+            GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s] use=%d,c=%d:", i, ggml_op_name(node->op), node->name,
                 fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node),
-                graph->use_counts[ggml_hash_find(&graph->visited_hash_set, node)]);
+                graph->use_counts[ggml_hash_find(&graph->visited_hash_set, node)], node->flags & GGML_TENSOR_FLAG_COMPUTE ? 1 : 0);
             for (int j = 0; j < GGML_MAX_SRC; j++) {
                 struct ggml_tensor * src = node->src[j];
                 if (src == NULL) {
@@ -1922,6 +1922,7 @@ static struct ggml_tensor * graph_copy_dup_tensor(struct ggml_hash_set hash_set,
         dst->view_offs = src->view_offs;
     }
     dst->op = src->op;
+    dst->flags = src->flags;
     memcpy(dst->op_params, src->op_params, sizeof(dst->op_params));
     ggml_set_name(dst, src->name);
 
diff --git a/ggml/src/ggml-blas/ggml-blas.cpp b/ggml/src/ggml-blas/ggml-blas.cpp
index 84956cbb9c..2e9ddf2240 100644
--- a/ggml/src/ggml-blas/ggml-blas.cpp
+++ b/ggml/src/ggml-blas/ggml-blas.cpp
@@ -226,6 +226,10 @@ static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend,
     for (int i = 0; i < cgraph->n_nodes; i++) {
         struct ggml_tensor * node = cgraph->nodes[i];
 
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
+
         switch (node->op) {
             case GGML_OP_MUL_MAT:
                 ggml_backend_blas_mul_mat(ctx, node);
diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
index eba83327f1..42c6c67a40 100644
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -2146,6 +2146,10 @@ static void evaluate_and_capture_cann_graph(ggml_backend_cann_context * cann_ctx
                 continue;
             }
 
+            if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+                continue;
+            }
+
             bool ok = ggml_cann_compute_forward(*cann_ctx, node);
             if (!ok) {
                 GGML_LOG_ERROR("%s: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
diff --git a/ggml/src/ggml-cpu/arch-fallback.h b/ggml/src/ggml-cpu/arch-fallback.h
index 3f8946ac70..427c1146e4 100644
--- a/ggml/src/ggml-cpu/arch-fallback.h
+++ b/ggml/src/ggml-cpu/arch-fallback.h
@@ -1,3 +1,4 @@
+
 #pragma once
 
 // Rename `_generic` functions if no native implementation is available.
@@ -38,9 +39,11 @@
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
@@ -48,9 +51,11 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#    define ggml_gemm_q6_K_8x8_q8_K_generic   ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
@@ -70,12 +75,16 @@
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
 #define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
@@ -94,9 +103,11 @@
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
@@ -104,9 +115,11 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
@@ -126,9 +139,11 @@
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
@@ -136,9 +151,11 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
@@ -165,18 +182,22 @@
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
 #define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
@@ -202,9 +223,11 @@
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
@@ -212,9 +235,11 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
@@ -242,9 +267,11 @@
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
-#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
@@ -252,9 +279,11 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
-#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
diff --git a/ggml/src/ggml-cpu/arch/arm/repack.cpp b/ggml/src/ggml-cpu/arch/arm/repack.cpp
index b61220a189..99bb70274c 100644
--- a/ggml/src/ggml-cpu/arch/arm/repack.cpp
+++ b/ggml/src/ggml-cpu/arch/arm/repack.cpp
@@ -25,9 +25,8 @@
 #define UNUSED GGML_UNUSED
 
 #if defined(__aarch64__) && defined(__ARM_NEON) && (defined(__ARM_FEATURE_MATMUL_INT8) || defined(__ARM_FEATURE_DOTPROD))
-static inline void decode_q4_Kx8_scales_mins(const uint8_t * scales_in,
-                                             int16x8_t *     out_mins,
-                                             int8_t *        out_scales) {
+// Helper for decoding scales and mins of Q4_K and Q5_K block formats
+static inline void decode_q_Kx8_6bit_scales(const uint8_t * scales_in, int16x8_t * out_mins, int8_t * out_scales) {
     constexpr uint32_t kmask1 = 0x3f3f3f3f;
     constexpr uint32_t kmask2 = 0x0f0f0f0f;
     constexpr uint32_t kmask3 = 0x03030303;
@@ -561,7 +560,7 @@ void ggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                 for (int i = 0; i < 2; i++) {
                     int8_t    aux_q4sb[8];
                     const int offset = sb * 24 + i * 12;
-                    decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
+                    decode_q_Kx8_6bit_scales(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
                     q4sb_scales[i] = vmovl_s8(vld1_s8(aux_q4sb));
                 }
 
@@ -701,7 +700,7 @@ void ggml_gemv_q4_K_8x8_q8_K(int                        n,
                 for (int i = 0; i < 2; i++) {
                     int8_t    aux_q4sb[8];
                     const int offset = sb * 24 + i * 12;
-                    decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
+                    decode_q_Kx8_6bit_scales(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
                     q4sb_scales[i] = vmovl_s8(vld1_s8(aux_q4sb));
                 }
 
@@ -786,6 +785,495 @@ void ggml_gemv_q4_K_8x8_q8_K(int                        n,
     ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 }
 
+void ggml_gemv_q5_K_8x8_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    col_pairs = ncols_interleaved / 2;
+    const uint8x16_t m4b       = vdupq_n_u8(0x0f);
+    const uint8x16_t mone      = vdupq_n_u8(1);
+    const uint8x16_t mtwo      = vdupq_n_u8(2);
+
+    // 1x8 tile = 2 x 4
+    float32x4_t acc_f32[ncols_interleaved / 4];
+
+    const block_q8_K * GGML_RESTRICT q8_ptr = (const block_q8_K *) vy;
+
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q5_Kx8 * GGML_RESTRICT q5_ptr = (const block_q5_Kx8 *) vx + (x * nb);
+
+        for (int i = 0; i < ncols_interleaved / 4; i++) {
+            acc_f32[i] = vdupq_n_f32(0);
+        }
+
+        for (int b = 0; b < nb; b++) {
+            float32x4_t q5_d_0     = vcvt_f32_f16(vld1_f16((const __fp16 *) q5_ptr[b].d));      // d0 d1 d2 d3
+            float32x4_t q5_d_1     = vcvt_f32_f16(vld1_f16((const __fp16 *) q5_ptr[b].d + 4));  // d4 d5 d6 d7
+            float32x4_t q8_d       = vdupq_n_f32(q8_ptr[b].d);
+            float32x4_t sb_scale_0 = vmulq_f32(q5_d_0, q8_d);
+            float32x4_t sb_scale_1 = vmulq_f32(q5_d_1, q8_d);
+            float32x4_t q5_dmin_0  = vcvt_f32_f16(vld1_f16((const __fp16 *) q5_ptr[b].dmin));      // dmin 0..3
+            float32x4_t q5_dmin_1  = vcvt_f32_f16(vld1_f16((const __fp16 *) q5_ptr[b].dmin + 4));  // dmin 4..7
+            float32x4_t sb_min_0   = vmulq_f32(q5_dmin_0, q8_d);
+            float32x4_t sb_min_1   = vmulq_f32(q5_dmin_1, q8_d);
+
+            // 2 sb each iteration
+            int32x4_t acc_lo[col_pairs];
+            int32x4_t acc_hi[col_pairs];
+
+            // Each bsum is 16 elements, pairwise add leaves us with the 8 bsums of the entire block
+            const int16x8_t bsums = vpaddq_s16(vld1q_s16(q8_ptr[b].bsums), vld1q_s16(q8_ptr[b].bsums + 8));
+            int16_t         bsums_arr[8];
+            vst1q_s16(bsums_arr, bsums);
+
+            // Load qh once per block and shift after each subblock
+            const uint8_t * qh_base = q5_ptr[b].qh;
+            uint8x16_t      qh[col_pairs][4];
+            for (int cp = 0; cp < col_pairs; cp++) {
+                qh[cp][0] = vld1q_u8(qh_base + 16 * cp);
+                qh[cp][1] = vld1q_u8(qh_base + 16 * cp + 64);
+                qh[cp][2] = vld1q_u8(qh_base + 16 * cp + 128);
+                qh[cp][3] = vld1q_u8(qh_base + 16 * cp + 192);
+            }
+
+            for (int sb = 0; sb < QK_K / 64; sb++) {
+                for (int i = 0; i < col_pairs; i++) {
+                    acc_lo[i] = vdupq_n_s32(0);
+                    acc_hi[i] = vdupq_n_s32(0);
+                }
+                // Need scales for the low and high nibbles
+                // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                int16x8_t q5sb_mins[2];  // int16 as its needed for bias_acc later
+                int16x8_t q5sb_scales[2];
+                for (int i = 0; i < 2; i++) {
+                    int8_t    aux_q5sb[8];
+                    const int offset = sb * 24 + i * 12;
+                    decode_q_Kx8_6bit_scales(&q5_ptr[b].scales[offset], &q5sb_mins[i], aux_q5sb);
+                    q5sb_scales[i] = vmovl_s8(vld1_s8(aux_q5sb));
+                }
+
+                const uint8_t * qs_base = q5_ptr[b].qs + sb * QK_K;
+
+                // Load the 64 quants from q8K duplicated to use vecdots with the interleaved columns
+                const int8_t * q8_base = q8_ptr[b].qs + sb * 64;
+                int8x16_t      q8_qs[8];
+                for (int i = 0; i < 8; i++) {
+                    q8_qs[i] = (int8x16_t) vld1q_dup_s64((const int64_t *) (q8_base + i * 8));
+                }
+
+                // Q5s column pair loop unrolled
+                {
+                    // Cols 01
+                    uint8x16_t qs_0 = vld1q_u8(qs_base);
+                    uint8x16_t qs_1 = vld1q_u8(qs_base + 64);
+                    uint8x16_t qs_2 = vld1q_u8(qs_base + 128);
+                    uint8x16_t qs_3 = vld1q_u8(qs_base + 192);
+
+                    uint8x16_t hbit_lo_0 = vandq_u8(qh[0][0], mone);
+                    uint8x16_t hbit_lo_1 = vandq_u8(qh[0][1], mone);
+                    uint8x16_t hbit_lo_2 = vandq_u8(qh[0][2], mone);
+                    uint8x16_t hbit_lo_3 = vandq_u8(qh[0][3], mone);
+                    uint8x16_t hbit_hi_0 = vshlq_n_u8(vandq_u8(qh[0][0], mtwo), 3);
+                    uint8x16_t hbit_hi_1 = vshlq_n_u8(vandq_u8(qh[0][1], mtwo), 3);
+                    uint8x16_t hbit_hi_2 = vshlq_n_u8(vandq_u8(qh[0][2], mtwo), 3);
+                    uint8x16_t hbit_hi_3 = vshlq_n_u8(vandq_u8(qh[0][3], mtwo), 3);
+
+                    qh[0][0] = vshrq_n_u8(qh[0][0], 2);
+                    qh[0][1] = vshrq_n_u8(qh[0][1], 2);
+                    qh[0][2] = vshrq_n_u8(qh[0][2], 2);
+                    qh[0][3] = vshrq_n_u8(qh[0][3], 2);
+
+                    acc_lo[0] = ggml_vdotq_s32(
+                        acc_lo[0], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_0, m4b), hbit_lo_0, 4)), q8_qs[0]);
+                    acc_lo[0] = ggml_vdotq_s32(
+                        acc_lo[0], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_1, m4b), hbit_lo_1, 4)), q8_qs[1]);
+                    acc_lo[0] = ggml_vdotq_s32(
+                        acc_lo[0], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_2, m4b), hbit_lo_2, 4)), q8_qs[2]);
+                    acc_lo[0] = ggml_vdotq_s32(
+                        acc_lo[0], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_3, m4b), hbit_lo_3, 4)), q8_qs[3]);
+                    acc_hi[0] = ggml_vdotq_s32(acc_hi[0], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_0, 4), hbit_hi_0)),
+                                               q8_qs[4]);
+                    acc_hi[0] = ggml_vdotq_s32(acc_hi[0], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_1, 4), hbit_hi_1)),
+                                               q8_qs[5]);
+                    acc_hi[0] = ggml_vdotq_s32(acc_hi[0], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_2, 4), hbit_hi_2)),
+                                               q8_qs[6]);
+                    acc_hi[0] = ggml_vdotq_s32(acc_hi[0], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_3, 4), hbit_hi_3)),
+                                               q8_qs[7]);
+
+                    // Cols 23
+                    qs_0 = vld1q_u8(qs_base + 16);
+                    qs_1 = vld1q_u8(qs_base + 80);
+                    qs_2 = vld1q_u8(qs_base + 144);
+                    qs_3 = vld1q_u8(qs_base + 208);
+
+                    hbit_lo_0 = vandq_u8(qh[1][0], mone);
+                    hbit_lo_1 = vandq_u8(qh[1][1], mone);
+                    hbit_lo_2 = vandq_u8(qh[1][2], mone);
+                    hbit_lo_3 = vandq_u8(qh[1][3], mone);
+                    hbit_hi_0 = vshlq_n_u8(vandq_u8(qh[1][0], mtwo), 3);
+                    hbit_hi_1 = vshlq_n_u8(vandq_u8(qh[1][1], mtwo), 3);
+                    hbit_hi_2 = vshlq_n_u8(vandq_u8(qh[1][2], mtwo), 3);
+                    hbit_hi_3 = vshlq_n_u8(vandq_u8(qh[1][3], mtwo), 3);
+
+                    qh[1][0] = vshrq_n_u8(qh[1][0], 2);
+                    qh[1][1] = vshrq_n_u8(qh[1][1], 2);
+                    qh[1][2] = vshrq_n_u8(qh[1][2], 2);
+                    qh[1][3] = vshrq_n_u8(qh[1][3], 2);
+
+                    acc_lo[1] = ggml_vdotq_s32(
+                        acc_lo[1], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_0, m4b), hbit_lo_0, 4)), q8_qs[0]);
+                    acc_lo[1] = ggml_vdotq_s32(
+                        acc_lo[1], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_1, m4b), hbit_lo_1, 4)), q8_qs[1]);
+                    acc_lo[1] = ggml_vdotq_s32(
+                        acc_lo[1], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_2, m4b), hbit_lo_2, 4)), q8_qs[2]);
+                    acc_lo[1] = ggml_vdotq_s32(
+                        acc_lo[1], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_3, m4b), hbit_lo_3, 4)), q8_qs[3]);
+                    acc_hi[1] = ggml_vdotq_s32(acc_hi[1], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_0, 4), hbit_hi_0)),
+                                               q8_qs[4]);
+                    acc_hi[1] = ggml_vdotq_s32(acc_hi[1], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_1, 4), hbit_hi_1)),
+                                               q8_qs[5]);
+                    acc_hi[1] = ggml_vdotq_s32(acc_hi[1], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_2, 4), hbit_hi_2)),
+                                               q8_qs[6]);
+                    acc_hi[1] = ggml_vdotq_s32(acc_hi[1], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_3, 4), hbit_hi_3)),
+                                               q8_qs[7]);
+
+                    // Cols 45
+                    qs_0 = vld1q_u8(qs_base + 32);
+                    qs_1 = vld1q_u8(qs_base + 96);
+                    qs_2 = vld1q_u8(qs_base + 160);
+                    qs_3 = vld1q_u8(qs_base + 224);
+
+                    hbit_lo_0 = vandq_u8(qh[2][0], mone);
+                    hbit_lo_1 = vandq_u8(qh[2][1], mone);
+                    hbit_lo_2 = vandq_u8(qh[2][2], mone);
+                    hbit_lo_3 = vandq_u8(qh[2][3], mone);
+                    hbit_hi_0 = vshlq_n_u8(vandq_u8(qh[2][0], mtwo), 3);
+                    hbit_hi_1 = vshlq_n_u8(vandq_u8(qh[2][1], mtwo), 3);
+                    hbit_hi_2 = vshlq_n_u8(vandq_u8(qh[2][2], mtwo), 3);
+                    hbit_hi_3 = vshlq_n_u8(vandq_u8(qh[2][3], mtwo), 3);
+
+                    qh[2][0] = vshrq_n_u8(qh[2][0], 2);
+                    qh[2][1] = vshrq_n_u8(qh[2][1], 2);
+                    qh[2][2] = vshrq_n_u8(qh[2][2], 2);
+                    qh[2][3] = vshrq_n_u8(qh[2][3], 2);
+
+                    acc_lo[2] = ggml_vdotq_s32(
+                        acc_lo[2], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_0, m4b), hbit_lo_0, 4)), q8_qs[0]);
+                    acc_lo[2] = ggml_vdotq_s32(
+                        acc_lo[2], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_1, m4b), hbit_lo_1, 4)), q8_qs[1]);
+                    acc_lo[2] = ggml_vdotq_s32(
+                        acc_lo[2], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_2, m4b), hbit_lo_2, 4)), q8_qs[2]);
+                    acc_lo[2] = ggml_vdotq_s32(
+                        acc_lo[2], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_3, m4b), hbit_lo_3, 4)), q8_qs[3]);
+                    acc_hi[2] = ggml_vdotq_s32(acc_hi[2], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_0, 4), hbit_hi_0)),
+                                               q8_qs[4]);
+                    acc_hi[2] = ggml_vdotq_s32(acc_hi[2], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_1, 4), hbit_hi_1)),
+                                               q8_qs[5]);
+                    acc_hi[2] = ggml_vdotq_s32(acc_hi[2], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_2, 4), hbit_hi_2)),
+                                               q8_qs[6]);
+                    acc_hi[2] = ggml_vdotq_s32(acc_hi[2], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_3, 4), hbit_hi_3)),
+                                               q8_qs[7]);
+
+                    // Cols 45
+                    qs_0 = vld1q_u8(qs_base + 48);
+                    qs_1 = vld1q_u8(qs_base + 112);
+                    qs_2 = vld1q_u8(qs_base + 176);
+                    qs_3 = vld1q_u8(qs_base + 240);
+
+                    hbit_lo_0 = vandq_u8(qh[3][0], mone);
+                    hbit_lo_1 = vandq_u8(qh[3][1], mone);
+                    hbit_lo_2 = vandq_u8(qh[3][2], mone);
+                    hbit_lo_3 = vandq_u8(qh[3][3], mone);
+                    hbit_hi_0 = vshlq_n_u8(vandq_u8(qh[3][0], mtwo), 3);
+                    hbit_hi_1 = vshlq_n_u8(vandq_u8(qh[3][1], mtwo), 3);
+                    hbit_hi_2 = vshlq_n_u8(vandq_u8(qh[3][2], mtwo), 3);
+                    hbit_hi_3 = vshlq_n_u8(vandq_u8(qh[3][3], mtwo), 3);
+
+                    qh[3][0] = vshrq_n_u8(qh[3][0], 2);
+                    qh[3][1] = vshrq_n_u8(qh[3][1], 2);
+                    qh[3][2] = vshrq_n_u8(qh[3][2], 2);
+                    qh[3][3] = vshrq_n_u8(qh[3][3], 2);
+
+                    acc_lo[3] = ggml_vdotq_s32(
+                        acc_lo[3], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_0, m4b), hbit_lo_0, 4)), q8_qs[0]);
+                    acc_lo[3] = ggml_vdotq_s32(
+                        acc_lo[3], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_1, m4b), hbit_lo_1, 4)), q8_qs[1]);
+                    acc_lo[3] = ggml_vdotq_s32(
+                        acc_lo[3], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_2, m4b), hbit_lo_2, 4)), q8_qs[2]);
+                    acc_lo[3] = ggml_vdotq_s32(
+                        acc_lo[3], vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_3, m4b), hbit_lo_3, 4)), q8_qs[3]);
+                    acc_hi[3] = ggml_vdotq_s32(acc_hi[3], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_0, 4), hbit_hi_0)),
+                                               q8_qs[4]);
+                    acc_hi[3] = ggml_vdotq_s32(acc_hi[3], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_1, 4), hbit_hi_1)),
+                                               q8_qs[5]);
+                    acc_hi[3] = ggml_vdotq_s32(acc_hi[3], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_2, 4), hbit_hi_2)),
+                                               q8_qs[6]);
+                    acc_hi[3] = ggml_vdotq_s32(acc_hi[3], vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_3, 4), hbit_hi_3)),
+                                               q8_qs[7]);
+                }
+
+                // Prepare bsum vectors for bias computation
+                // Each pair of subblocks share the same bsums
+                int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[2 * sb + 0]);
+                int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[2 * sb + 1]);
+
+                // Iterates over a pair of column pairs (4 columns) to use a single 128 register
+                // p = 0 -> 0123  p2 -> 4567
+                for (int i = 0, p = 0; p < col_pairs; i++, p += 2) {
+                    int16x4_t   group_scales_lo = p == 0 ? vget_low_s16(q5sb_scales[0]) : vget_high_s16(q5sb_scales[0]);
+                    int16x4_t   group_scales_hi = p == 0 ? vget_low_s16(q5sb_scales[1]) : vget_high_s16(q5sb_scales[1]);
+                    int16x4_t   group_mins_lo   = p == 0 ? vget_low_s16(q5sb_mins[0]) : vget_high_s16(q5sb_mins[0]);
+                    int16x4_t   group_mins_hi   = p == 0 ? vget_low_s16(q5sb_mins[1]) : vget_high_s16(q5sb_mins[1]);
+                    float32x4_t sb_scale        = p == 0 ? sb_scale_0 : sb_scale_1;
+                    float32x4_t sb_min          = p == 0 ? sb_min_0 : sb_min_1;
+
+                    // 0123 or 4567
+                    float32x4_t sumf_0 =
+                        vcvtq_f32_s32(vmulq_s32(vmovl_s16(group_scales_lo), vpaddq_s32(acc_lo[p], acc_lo[p + 1])));
+                    acc_f32[i] = vfmaq_f32(acc_f32[i], sb_scale, sumf_0);
+
+                    float32x4_t sumf_1 =
+                        vcvtq_f32_s32(vmulq_s32(vmovl_s16(group_scales_hi), vpaddq_s32(acc_hi[p], acc_hi[p + 1])));
+                    acc_f32[i] = vfmaq_f32(acc_f32[i], sb_scale, sumf_1);
+
+                    // FUSED BIAS: Compute and subtract bias immediately
+                    // bias = (bsums_lo * mins_lo + bsums_hi * mins_hi) * sb_min
+                    int32x4_t bias       = vmull_s16(bsums_vec_lo, group_mins_lo);
+                    bias                 = vmlal_s16(bias, bsums_vec_hi, group_mins_hi);
+                    float32x4_t bias_f32 = vcvtq_f32_s32(bias);
+                    acc_f32[i]           = vmlsq_f32(acc_f32[i], sb_min, bias_f32);
+                }
+            }  // for sb
+        }  // for b
+
+        int base = x * ncols_interleaved;
+        vst1q_f32(s + base, acc_f32[0]);
+        vst1q_f32(s + base + 4, acc_f32[1]);
+    }  // for x
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemv_q5_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
+void ggml_gemv_q6_K_8x8_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    col_pairs = ncols_interleaved / 2;
+    const uint8x16_t m4b       = vdupq_n_u8(0x0f);
+    const uint8x16_t mask_lo   = vdupq_n_u8(0x03);
+    const uint8x16_t mask_hi   = vdupq_n_u8(0x30);
+
+    // 1x8 tile = 2 x 4
+    float32x4_t acc_f32[2];
+
+    const block_q8_K * GGML_RESTRICT q8_ptr = (const block_q8_K *) vy;
+
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q6_Kx8 * GGML_RESTRICT q6_ptr = (const block_q6_Kx8 *) vx + (x * nb);
+
+        acc_f32[0] = vdupq_n_f32(0);
+        acc_f32[1] = vdupq_n_f32(0);
+
+        for (int b = 0; b < nb; b++) {
+            float32x4_t q6_d_0     = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d));      // d0 d1 d2 d3
+            float32x4_t q6_d_1     = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d + 4));  // d4 d5 d6 d7
+            float32x4_t q8_d       = vdupq_n_f32(q8_ptr[b].d);
+            float32x4_t sb_scale_0 = vmulq_f32(q6_d_0, q8_d);
+            float32x4_t sb_scale_1 = vmulq_f32(q6_d_1, q8_d);
+
+            int32x2_t acc[col_pairs];
+            for (int i = 0; i < col_pairs; i++) {
+                acc[i] = vdup_n_s32(0);
+            }
+
+            // Load all 16 scales once and widen to int16 (Q6_K has 16 scales per block)
+            // Reused for bias and dequantization later
+            int16_t q6_scales[16 * 8];
+            for (int i = 0; i < 16; i++) {
+                int16x8_t scales = vmovl_s8(vld1_s8(q6_ptr[b].scales + i * 8));
+                vst1q_s16(q6_scales + i * 8, scales);
+            }
+
+            // Compute bias per column using q8 bsums and preloaded scales to skip the -32 shift
+            int32x4_t bias_lo = vdupq_n_s32(0);
+            int32x4_t bias_hi = vdupq_n_s32(0);
+
+            // Load bsums in chunks of 4 to process with vectorized operations
+            for (int i = 0; i < 16; i += 4) {
+                int16x4_t bsums_vec   = vld1_s16(q8_ptr[b].bsums + i);
+                int16x4_t scales_lo_0 = vld1_s16(q6_scales + (i + 0) * 8);
+                int16x4_t scales_hi_0 = vld1_s16(q6_scales + (i + 0) * 8 + 4);
+                int16x4_t scales_lo_1 = vld1_s16(q6_scales + (i + 1) * 8);
+                int16x4_t scales_hi_1 = vld1_s16(q6_scales + (i + 1) * 8 + 4);
+                int16x4_t scales_lo_2 = vld1_s16(q6_scales + (i + 2) * 8);
+                int16x4_t scales_hi_2 = vld1_s16(q6_scales + (i + 2) * 8 + 4);
+                int16x4_t scales_lo_3 = vld1_s16(q6_scales + (i + 3) * 8);
+                int16x4_t scales_hi_3 = vld1_s16(q6_scales + (i + 3) * 8 + 4);
+
+                bias_lo = vmlal_lane_s16(bias_lo, scales_lo_0, bsums_vec, 0);
+                bias_hi = vmlal_lane_s16(bias_hi, scales_hi_0, bsums_vec, 0);
+                bias_lo = vmlal_lane_s16(bias_lo, scales_lo_1, bsums_vec, 1);
+                bias_hi = vmlal_lane_s16(bias_hi, scales_hi_1, bsums_vec, 1);
+                bias_lo = vmlal_lane_s16(bias_lo, scales_lo_2, bsums_vec, 2);
+                bias_hi = vmlal_lane_s16(bias_hi, scales_hi_2, bsums_vec, 2);
+                bias_lo = vmlal_lane_s16(bias_lo, scales_lo_3, bsums_vec, 3);
+                bias_hi = vmlal_lane_s16(bias_hi, scales_hi_3, bsums_vec, 3);
+            }
+            bias_lo = vshlq_n_s32(bias_lo, 5);
+            bias_hi = vshlq_n_s32(bias_hi, 5);
+
+            // Process two 128-value halves per superblock
+            for (int half = 0; half < 2; half++) {
+                const uint8_t * ql_base = q6_ptr[b].ql + half * 512;
+                const uint8_t * qh_base = q6_ptr[b].qh + half * 256;
+
+                // A subblock (sb) is a set of weights that share the scale
+                // Since q6_K scales are per 16 elements
+                // num sbs -> 256 elements / (16 elements/scale * 2 elements/byte * 2 halves)
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+                    const int8_t * q8_base_l = q8_ptr[b].qs + half * 128 + sb * 16;
+                    const int8_t * q8_base_h = q8_base_l + 64;
+
+                    // Load and duplicate q8 values (each register covers two interleaved columns of q6)
+                    int8x16_t q8_l[2];
+                    int8x16_t q8_h[2];
+                    for (int i = 0; i < 2; i++) {
+                        q8_l[i] = (int8x16_t) vld1q_dup_s64((const int64_t *) (q8_base_l + i * 8));
+                        q8_h[i] = (int8x16_t) vld1q_dup_s64((const int64_t *) (q8_base_h + i * 8));
+                    }
+
+                    // TODO: Test other qh repack patterns to reduce loads
+                    const int ql_off_base = sb * QK_K / 2;
+                    const int qh_off_base = ql_off_base & 255;  // wraps after 256 bytes
+
+                    // Load 4 vectors at once (64 bytes each for ql_0, ql_1, qh_0, qh_1)
+                    ggml_uint8x16x4_t q6_ql_0 = ggml_vld1q_u8_x4(ql_base + ql_off_base);
+                    ggml_uint8x16x4_t q6_ql_1 = ggml_vld1q_u8_x4(ql_base + ql_off_base + 64);
+                    ggml_uint8x16x4_t q6_qh_0 = ggml_vld1q_u8_x4(qh_base + qh_off_base);
+                    ggml_uint8x16x4_t q6_qh_1 = ggml_vld1q_u8_x4(qh_base + qh_off_base + 64);
+
+                    // Adjust qh for subblocks 2 and 3 (shift right by 2)
+                    if (sb > 1) {
+                        q6_qh_0.val[0] = vshrq_n_u8(q6_qh_0.val[0], 2);
+                        q6_qh_0.val[1] = vshrq_n_u8(q6_qh_0.val[1], 2);
+                        q6_qh_0.val[2] = vshrq_n_u8(q6_qh_0.val[2], 2);
+                        q6_qh_0.val[3] = vshrq_n_u8(q6_qh_0.val[3], 2);
+                        q6_qh_1.val[0] = vshrq_n_u8(q6_qh_1.val[0], 2);
+                        q6_qh_1.val[1] = vshrq_n_u8(q6_qh_1.val[1], 2);
+                        q6_qh_1.val[2] = vshrq_n_u8(q6_qh_1.val[2], 2);
+                        q6_qh_1.val[3] = vshrq_n_u8(q6_qh_1.val[3], 2);
+                    }
+
+                    // Process column pairs (0-1, 2-3, 4-5, 6-7)
+                    for (int cp = 0; cp < col_pairs; cp++) {
+                        const uint8x16_t q6_qs_cp_0_l = q6_ql_0.val[cp];
+                        const uint8x16_t q6_qs_cp_1_l = q6_ql_1.val[cp];
+                        const uint8x16_t q6_qs_cp_0_h = q6_qh_0.val[cp];
+                        const uint8x16_t q6_qs_cp_1_h = q6_qh_1.val[cp];
+
+                        // Extract high 2 bits for upper nibble reconstruction
+                        const uint8x16_t q6_qs_cp_0_hh = vandq_u8(q6_qs_cp_0_h, mask_hi);
+                        const uint8x16_t q6_qs_cp_1_hh = vandq_u8(q6_qs_cp_1_h, mask_hi);
+
+                        // q6 = (low4 | high2<<4), without -32 bias (handled via bsums)
+                        const int8x16_t q6_l0 = vreinterpretq_s8_u8(
+                            vsliq_n_u8(vandq_u8(q6_qs_cp_0_l, m4b), vandq_u8(q6_qs_cp_0_h, mask_lo), 4));
+                        const int8x16_t q6_l1 = vreinterpretq_s8_u8(
+                            vsliq_n_u8(vandq_u8(q6_qs_cp_1_l, m4b), vandq_u8(q6_qs_cp_1_h, mask_lo), 4));
+                        const int8x16_t q6_h0 =
+                            vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_qs_cp_0_l, 4), q6_qs_cp_0_hh));
+                        const int8x16_t q6_h1 =
+                            vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_qs_cp_1_l, 4), q6_qs_cp_1_hh));
+
+                        int32x4_t sb_acc_l = vdupq_n_s32(0);
+                        sb_acc_l           = vdotq_s32(sb_acc_l, q6_l0, q8_l[0]);
+                        sb_acc_l           = vdotq_s32(sb_acc_l, q6_l1, q8_l[1]);
+
+                        int32x4_t sb_acc_h = vdupq_n_s32(0);
+                        sb_acc_h           = vdotq_s32(sb_acc_h, q6_h0, q8_h[0]);
+                        sb_acc_h           = vdotq_s32(sb_acc_h, q6_h1, q8_h[1]);
+
+                        // Pairwise add to get per-column sums: [col0, col1]
+                        int32x2_t sum_l = vpadd_s32(vget_low_s32(sb_acc_l), vget_high_s32(sb_acc_l));
+                        int32x2_t sum_h = vpadd_s32(vget_low_s32(sb_acc_h), vget_high_s32(sb_acc_h));
+
+                        const int scale_idx_l = half * 8 + sb;
+                        const int scale_idx_h = half * 8 + sb + 4;
+
+                        // Access scales using array indexing (scales are interleaved by column)
+                        const int32x2_t scale_vec_l = { (int32_t) q6_scales[scale_idx_l * 8 + cp * 2],
+                                                        (int32_t) q6_scales[scale_idx_l * 8 + cp * 2 + 1] };
+                        const int32x2_t scale_vec_h = { (int32_t) q6_scales[scale_idx_h * 8 + cp * 2],
+                                                        (int32_t) q6_scales[scale_idx_h * 8 + cp * 2 + 1] };
+
+                        // Accumulate scaled results
+                        acc[cp] = vmla_s32(acc[cp], sum_l, scale_vec_l);
+                        acc[cp] = vmla_s32(acc[cp], sum_h, scale_vec_h);
+                    }
+                }
+            }  // for half
+
+            // Bias correction
+            acc[0] = vsub_s32(acc[0], vget_low_s32(bias_lo));
+            acc[1] = vsub_s32(acc[1], vget_high_s32(bias_lo));
+            acc[2] = vsub_s32(acc[2], vget_low_s32(bias_hi));
+            acc[3] = vsub_s32(acc[3], vget_high_s32(bias_hi));
+
+            // Apply superblock scale (no mins for q6_K)
+            // acc[cp] has [c0, c1]
+            float32x2_t w_01 = vmul_f32(vcvt_f32_s32(acc[0]), vget_low_f32(sb_scale_0));
+            float32x2_t w_23 = vmul_f32(vcvt_f32_s32(acc[1]), vget_high_f32(sb_scale_0));
+            float32x2_t w_45 = vmul_f32(vcvt_f32_s32(acc[2]), vget_low_f32(sb_scale_1));
+            float32x2_t w_67 = vmul_f32(vcvt_f32_s32(acc[3]), vget_high_f32(sb_scale_1));
+
+            acc_f32[0] = vaddq_f32(acc_f32[0], vcombine_f32(w_01, w_23));
+            acc_f32[1] = vaddq_f32(acc_f32[1], vcombine_f32(w_45, w_67));
+        }  // for b
+
+        int base = x * ncols_interleaved;
+        vst1q_f32(s + base, acc_f32[0]);
+        vst1q_f32(s + base + 4, acc_f32[1]);
+    }  // for x
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemv_q6_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemv_q8_0_4x4_q8_0(int                        n,
                              float * GGML_RESTRICT      s,
                              size_t                     bs,
@@ -2431,7 +2919,7 @@ void ggml_gemm_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                     for (int i = 0; i < 2; i++) {
                         int8_t    aux_q4sb[8];
                         const int offset = sb * 24 + i * 12;
-                        decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
+                        decode_q_Kx8_6bit_scales(&q4_ptr[b].scales[offset], &q4sb_mins[i], aux_q4sb);
                         q4sb_scales[i] = vmovl_s8(vld1_s8(aux_q4sb));
                     }
 
@@ -2595,7 +3083,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int                        n,
                     int16x8_t q4sb_mins[2];  // int16 as its needed for bias_acc later
                     for (int i = 0; i < 2; i++) {
                         const int offset = sb * 24 + i * 12;
-                        decode_q4_Kx8_scales_mins(&q4_ptr[b].scales[offset], &q4sb_mins[i], q4sb_scales[i]);
+                        decode_q_Kx8_6bit_scales(&q4_ptr[b].scales[offset], &q4sb_mins[i], q4sb_scales[i]);
                     }
 
                     // q8_ptr[b].qs has interleaved Q8 rows (01, 23)
@@ -2660,16 +3148,17 @@ void ggml_gemm_q4_K_8x8_q8_K(int                        n,
 
                         // Scales[i] corresponds to column i
                         const int scale_offset = cp * 2;
-                        for (int blk = 0; blk < 2; blk++) {
-                            const int32x4_t block_scale = {
-                                (int32_t) q4sb_scales[blk][scale_offset],
-                                (int32_t) q4sb_scales[blk][scale_offset],
-                                (int32_t) q4sb_scales[blk][scale_offset + 1],
-                                (int32_t) q4sb_scales[blk][scale_offset + 1],
-                            };
-                            acc[cp]     = vmlaq_s32(acc[cp], sb_acc[blk], block_scale);
-                            acc[cp + 4] = vmlaq_s32(acc[cp + 4], sb_acc[blk + 2], block_scale);
-                        }
+                        const int32_t scale_00 = q4sb_scales[0][scale_offset];
+                        const int32_t scale_01 = q4sb_scales[0][scale_offset + 1];
+                        const int32_t scale_10 = q4sb_scales[1][scale_offset];
+                        const int32_t scale_11 = q4sb_scales[1][scale_offset + 1];
+                        const int32x4_t block_scale_0 = vcombine_s32(vdup_n_s32(scale_00), vdup_n_s32(scale_01));
+                        const int32x4_t block_scale_1 = vcombine_s32(vdup_n_s32(scale_10), vdup_n_s32(scale_11));
+
+                        acc[cp]     = vmlaq_s32(acc[cp], sb_acc[0], block_scale_0);
+                        acc[cp + 4] = vmlaq_s32(acc[cp + 4], sb_acc[2], block_scale_0);
+                        acc[cp]     = vmlaq_s32(acc[cp], sb_acc[1], block_scale_1);
+                        acc[cp + 4] = vmlaq_s32(acc[cp + 4], sb_acc[3], block_scale_1);
                     }
 
                     // Multiply Acc bsum + mins
@@ -2738,6 +3227,469 @@ void ggml_gemm_q4_K_8x8_q8_K(int                        n,
     ggml_gemm_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 }
 
+void ggml_gemm_q5_K_8x8_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    constexpr int    q8_k_blocklen = 4;
+    constexpr int    col_pairs     = ncols_interleaved / 2;
+    const uint8x16_t m4b           = vdupq_n_u8(0x0f);
+    const uint8x16_t mone          = vdupq_n_u8(1);
+    const uint8x16_t mtwo          = vdupq_n_u8(2);
+
+    // 8 accumulators: 2 row pairs × 4 col pairs
+    float32x4_t acc_f32[blocklen];
+
+    for (int y = 0; y < nr / q8_k_blocklen; y++) {
+        const block_q8_Kx4 * GGML_RESTRICT q8_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q5_Kx8 * GGML_RESTRICT q5_ptr = (const block_q5_Kx8 *) vx + (x * nb);
+
+            for (int i = 0; i < blocklen; i++) {
+                acc_f32[i] = vdupq_n_f32(0);
+            }
+
+            for (int b = 0; b < nb; b++) {
+                // bsums pairs belongs to the same q8_k subblock
+                const int16x8_t bsums[4]{
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 0), vld1q_s16(q8_ptr[b].bsums + 16 * 0 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 1), vld1q_s16(q8_ptr[b].bsums + 16 * 1 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 2), vld1q_s16(q8_ptr[b].bsums + 16 * 2 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 3), vld1q_s16(q8_ptr[b].bsums + 16 * 3 + 8)),
+                };
+                int16_t bsums_arr[4][8];
+                for (int q8_row = 0; q8_row < 4; q8_row++) {
+                    vst1q_s16(bsums_arr[q8_row], bsums[q8_row]);
+                }
+
+                int32x4_t sb_acc[4];    // Aux accumulators to store subblock (partial) results
+                int32x4_t acc[8];       // rows 01 stored in [0][1][2][3] rows 23 stored in [4][5][6][7]
+                int32x4_t bias_acc[8];  // interleaved bias_acc: [0]->r0 0123, [1]->r0 4567, [2]->r1 0123 ...
+                for (int i = 0; i < 8; i++) {
+                    acc[i]      = vdupq_n_s32(0);
+                    bias_acc[i] = vdupq_n_s32(0);
+                }
+
+                // Load qh once per block and shift after each subblock
+                const uint8_t * qh_base = q5_ptr[b].qh;
+                uint8x16_t      qh[col_pairs][4];
+                for (int cp = 0; cp < col_pairs; cp++) {
+                    qh[cp][0] = vld1q_u8(qh_base + 16 * cp);
+                    qh[cp][1] = vld1q_u8(qh_base + 16 * cp + 64);
+                    qh[cp][2] = vld1q_u8(qh_base + 16 * cp + 128);
+                    qh[cp][3] = vld1q_u8(qh_base + 16 * cp + 192);
+                }
+
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+                    // Need scales for the low and high nibbles
+                    // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                    int8_t    q5sb_scales[2][8];
+                    int16x8_t q5sb_mins[2];  // int16 as its needed for bias_acc later
+                    for (int i = 0; i < 2; i++) {
+                        const int offset = sb * 24 + i * 12;
+                        decode_q_Kx8_6bit_scales(&q5_ptr[b].scales[offset], &q5sb_mins[i], q5sb_scales[i]);
+                    }
+
+                    // q8_ptr[b].qs has interleaved Q8 rows (01, 23)
+                    const int8_t * q8_base = q8_ptr[b].qs + sb * 256;
+
+                    int8x16_t q8_qs_01[8];
+                    int8x16_t q8_qs_23[8];
+
+                    // Load 32-byte per row pair, 1 subblock each time
+                    for (int i = 0; i < 8; i++) {
+                        const int offset = i * 32;  // 16 for row 01, 16 for row 23
+                        q8_qs_01[i]      = vld1q_s8(q8_base + offset);
+                        q8_qs_23[i]      = vld1q_s8(q8_base + offset + 16);
+                    }
+
+                    const int8x16_t q8s[2][8] = {
+                        { q8_qs_01[0], q8_qs_01[1], q8_qs_01[2], q8_qs_01[3], q8_qs_01[4], q8_qs_01[5], q8_qs_01[6],
+                         q8_qs_01[7] },
+                        { q8_qs_23[0], q8_qs_23[1], q8_qs_23[2], q8_qs_23[3], q8_qs_23[4], q8_qs_23[5], q8_qs_23[6],
+                         q8_qs_23[7] },
+                    };
+
+                    // Q5s columns iterated in pairs (01, 23, 45, 67)
+                    for (int cp = 0; cp < col_pairs; cp++) {
+                        for (int i = 0; i < 4; i++) {
+                            sb_acc[i] = vdupq_n_s32(0);
+                        }
+
+                        uint8x16_t qs_cp_0 = vld1q_u8(q5_ptr[b].qs + sb * QK_K + 16 * cp + 0);    // 0 .. 7 & 32..39
+                        uint8x16_t qs_cp_1 = vld1q_u8(q5_ptr[b].qs + sb * QK_K + 16 * cp + 64);   // 8 ..15 & 40..47
+                        uint8x16_t qs_cp_2 = vld1q_u8(q5_ptr[b].qs + sb * QK_K + 16 * cp + 128);  // 16..23 & 48..55
+                        uint8x16_t qs_cp_3 = vld1q_u8(q5_ptr[b].qs + sb * QK_K + 16 * cp + 192);  // 24..31 & 56..63
+
+                        // This is the only part of the algorithm that differs with Q4_K
+                        // Extract High bits and pack into 5 bit weights
+                        uint8x16_t hbit_lo_0    = vandq_u8(qh[cp][0], mone);
+                        uint8x16_t hbit_hi_0    = vshlq_n_u8(vandq_u8(qh[cp][0], mtwo), 3);
+                        qh[cp][0]               = vshrq_n_u8(qh[cp][0], 2);
+                        // Same as Q4_K, i8mm to dequantize the weights.
+                        const int8x16_t qs_lo_0 = vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_cp_0, m4b), hbit_lo_0, 4));
+                        int32x4_t       acc_0   = sb_acc[0];
+                        acc_0                   = vmmlaq_s32(acc_0, qs_lo_0, q8s[0][0]);
+                        int32x4_t acc_2         = sb_acc[2];
+                        acc_2                   = vmmlaq_s32(acc_2, qs_lo_0, q8s[1][0]);
+                        const int8x16_t qs_hi_0 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_cp_0, 4), hbit_hi_0));
+                        int32x4_t       acc_1   = sb_acc[1];
+                        acc_1                   = vmmlaq_s32(acc_1, qs_hi_0, q8s[0][4]);
+                        int32x4_t acc_3         = sb_acc[3];
+                        acc_3                   = vmmlaq_s32(acc_3, qs_hi_0, q8s[1][4]);
+
+                        // Repeat for the other 3 columns (8..15, 16..23, 24..31)
+                        uint8x16_t hbit_hi_1    = vshlq_n_u8(vandq_u8(qh[cp][1], mtwo), 3);
+                        uint8x16_t hbit_lo_1    = vandq_u8(qh[cp][1], mone);
+                        qh[cp][1]               = vshrq_n_u8(qh[cp][1], 2);
+                        const int8x16_t qs_lo_1 = vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_cp_1, m4b), hbit_lo_1, 4));
+                        acc_0                   = vmmlaq_s32(acc_0, qs_lo_1, q8s[0][1]);
+                        acc_2                   = vmmlaq_s32(acc_2, qs_lo_1, q8s[1][1]);
+                        const int8x16_t qs_hi_1 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_cp_1, 4), hbit_hi_1));
+                        acc_1                   = vmmlaq_s32(acc_1, qs_hi_1, q8s[0][5]);
+                        acc_3                   = vmmlaq_s32(acc_3, qs_hi_1, q8s[1][5]);
+
+                        uint8x16_t hbit_hi_2    = vshlq_n_u8(vandq_u8(qh[cp][2], mtwo), 3);
+                        uint8x16_t hbit_lo_2    = vandq_u8(qh[cp][2], mone);
+                        qh[cp][2]               = vshrq_n_u8(qh[cp][2], 2);
+                        const int8x16_t qs_lo_2 = vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_cp_2, m4b), hbit_lo_2, 4));
+                        acc_0                   = vmmlaq_s32(acc_0, qs_lo_2, q8s[0][2]);
+                        acc_2                   = vmmlaq_s32(acc_2, qs_lo_2, q8s[1][2]);
+                        const int8x16_t qs_hi_2 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_cp_2, 4), hbit_hi_2));
+                        acc_1                   = vmmlaq_s32(acc_1, qs_hi_2, q8s[0][6]);
+                        acc_3                   = vmmlaq_s32(acc_3, qs_hi_2, q8s[1][6]);
+
+                        uint8x16_t hbit_lo_3    = vandq_u8(qh[cp][3], mone);
+                        uint8x16_t hbit_hi_3    = vshlq_n_u8(vandq_u8(qh[cp][3], mtwo), 3);
+                        qh[cp][3]               = vshrq_n_u8(qh[cp][3], 2);
+                        const int8x16_t qs_lo_3 = vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(qs_cp_3, m4b), hbit_lo_3, 4));
+                        acc_0                   = vmmlaq_s32(acc_0, qs_lo_3, q8s[0][3]);
+                        sb_acc[0]               = acc_0;
+                        acc_2                   = vmmlaq_s32(acc_2, qs_lo_3, q8s[1][3]);
+                        sb_acc[2]               = acc_2;
+
+                        // Scales[i] corresponds to column i
+                        const int       scale_offset = cp * 2;
+                        const int32_t   s0           = q5sb_scales[0][scale_offset];
+                        const int32_t   s1           = q5sb_scales[0][scale_offset + 1];
+                        const int32x4_t block_scale  = vcombine_s32(vdup_n_s32(s0), vdup_n_s32(s1));
+                        acc[cp]                      = vmlaq_s32(acc[cp], sb_acc[0], block_scale);
+                        acc[cp + 4]                  = vmlaq_s32(acc[cp + 4], sb_acc[2], block_scale);
+
+                        const int8x16_t qs_hi_3 = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(qs_cp_3, 4), hbit_hi_3));
+                        acc_1                   = vmmlaq_s32(acc_1, qs_hi_3, q8s[0][7]);
+                        sb_acc[1]               = acc_1;
+                        acc_3                   = vmmlaq_s32(acc_3, qs_hi_3, q8s[1][7]);
+                        sb_acc[3]               = acc_3;
+
+                        const int32_t   s2           = q5sb_scales[1][scale_offset];
+                        const int32_t   s3           = q5sb_scales[1][scale_offset + 1];
+                        const int32x4_t block_scale2 = vcombine_s32(vdup_n_s32(s2), vdup_n_s32(s3));
+                        acc[cp]                      = vmlaq_s32(acc[cp], sb_acc[1], block_scale2);
+                        acc[cp + 4]                  = vmlaq_s32(acc[cp + 4], sb_acc[3], block_scale2);
+                    }
+
+                    // Multiply Acc bsum + mins
+                    for (int q8_row = 0; q8_row < 4; q8_row++) {
+                        // Each pair of subblocks share the same bsums
+                        // Load scalar bsum → broadcast to a vector (vdupq_n_s16(s)).
+                        int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[sb][q8_row * 2]);
+                        int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[sb][q8_row * 2 + 1]);
+
+                        bias_acc[2 * q8_row] =
+                            vmlal_s16(bias_acc[2 * q8_row], bsums_vec_lo, vget_low_s16(q5sb_mins[0]));
+                        bias_acc[2 * q8_row] =
+                            vmlal_s16(bias_acc[2 * q8_row], bsums_vec_hi, vget_low_s16(q5sb_mins[1]));
+                        bias_acc[2 * q8_row + 1] =
+                            vmlal_s16(bias_acc[2 * q8_row + 1], bsums_vec_lo, vget_high_s16(q5sb_mins[0]));
+                        bias_acc[2 * q8_row + 1] =
+                            vmlal_s16(bias_acc[2 * q8_row + 1], bsums_vec_hi, vget_high_s16(q5sb_mins[1]));
+                    }
+                }  // for sb
+
+                // Reorder of i8mm output with bias and output layout
+                for (int i = 0; i < 8; i++) {
+                    int32x2x2_t aux = vzip_s32(vget_low_s32(acc[i]), vget_high_s32(acc[i]));
+                    acc[i]          = vcombine_s32(aux.val[0], aux.val[1]);
+                }
+                int32x4_t reorder_acc[8] = {
+                    vcombine_s32(vget_low_s32(acc[0]), vget_low_s32(acc[1])),
+                    vcombine_s32(vget_low_s32(acc[2]), vget_low_s32(acc[3])),
+                    vcombine_s32(vget_high_s32(acc[0]), vget_high_s32(acc[1])),
+                    vcombine_s32(vget_high_s32(acc[2]), vget_high_s32(acc[3])),
+                    vcombine_s32(vget_low_s32(acc[4]), vget_low_s32(acc[5])),
+                    vcombine_s32(vget_low_s32(acc[6]), vget_low_s32(acc[7])),
+                    vcombine_s32(vget_high_s32(acc[4]), vget_high_s32(acc[5])),
+                    vcombine_s32(vget_high_s32(acc[6]), vget_high_s32(acc[7])),
+                };
+
+                for (int i = 0; i < q8_k_blocklen; i++) {
+                    for (int j = 0; j < 2; j++) {
+                        float32x4_t       q8_d    = vdupq_n_f32(q8_ptr[b].d[i]);
+                        float32x4_t       q5_dmin = vcvt_f32_f16(vld1_f16((const __fp16 *) (q5_ptr[b].dmin + j * 4)));
+                        const float32x4_t dmins   = vmulq_f32(q5_dmin, q8_d);
+
+                        float32x4_t       q5_d  = vcvt_f32_f16(vld1_f16((const __fp16 *) (q5_ptr[b].d + j * 4)));
+                        const float32x4_t scale = vmulq_f32(q5_d, q8_d);
+
+                        acc_f32[2 * i + j] = vmlsq_f32(acc_f32[2 * i + j], vcvtq_f32_s32(bias_acc[2 * i + j]), dmins);
+                        acc_f32[2 * i + j] =
+                            vmlaq_f32(acc_f32[2 * i + j], vcvtq_f32_s32(reorder_acc[2 * i + j]), scale);
+                    }
+                }
+            }  // for b
+
+            // With the previous reorder, the tile is already in the correct memory layout.
+            for (int i = 0; i < q8_k_blocklen; i++) {
+                int row = y * q8_k_blocklen + i;
+                for (int j = 0; j < 2; j++) {
+                    int col    = x * ncols_interleaved + j * 4;
+                    int offset = row * bs + col;
+                    vst1q_f32(s + offset, acc_f32[2 * i + j]);
+                }
+            }
+        }  // for x
+    }  // for y
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    ggml_gemm_q5_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
+void ggml_gemm_q6_K_8x8_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    constexpr int    q8_k_blocklen = 4;
+    const uint8x16_t m4b           = vdupq_n_u8(0x0f);
+    const uint8x16_t mask_lo       = vdupq_n_u8(0x03);
+    const uint8x16_t mask_hi       = vdupq_n_u8(0x30);
+    const int8x16_t  m32s          = vdupq_n_s8(32);
+
+    // 8 accumulators: 4 q8 rows × 2 col groups (0-3, 4-7)
+    float32x4_t acc_f32[blocklen];
+
+    for (int y = 0; y < nr / q8_k_blocklen; y++) {
+        const block_q8_Kx4 * GGML_RESTRICT q8_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q6_Kx8 * GGML_RESTRICT q6_ptr = (const block_q6_Kx8 *) vx + (x * nb);
+
+            for (int i = 0; i < blocklen; i++) {
+                acc_f32[i] = vdupq_n_f32(0);
+            }
+
+            for (int b = 0; b < nb; b++) {
+                int32x4_t acc[8];  // rows 01 stored in [0][1][2][3], rows 23 stored in [4][5][6][7]
+                for (int i = 0; i < 8; i++) {
+                    acc[i] = vdupq_n_s32(0);
+                }
+
+                // Q6_K has simple 8-bit scales, 16 per block (one per 16 values)
+                // Reused for bias and dequantization later
+                int16_t q6_scales[16 * 8];
+                for (int i = 0; i < 16; ++i) {
+                    int16x8_t s16 = vmovl_s8(vld1_s8(q6_ptr[b].scales + i * 8));
+                    vst1q_s16(q6_scales + i * 8, s16);
+                }
+
+                // Process two 128-value halves per superblock
+                for (int half = 0; half < 2; half++) {
+
+                    const uint8_t * ql_base = q6_ptr[b].ql + half * 512;
+                    const uint8_t * qh_base = q6_ptr[b].qh + half * 256;
+
+                    // A subblock (sb) is a set of weights that share the scale
+                    // Since q6_K scales are per 16 elements
+                    // num sbs -> 256 elements / (16 elements/scale * 2 elements/byte * 2 halves)
+                    for (int sb = 0; sb < QK_K / 64; sb++) {
+                        // Q6_K weight index increasing by 64 instead of 32 requires
+                        // loading various q8 memory regions
+                        const int8_t * q8_base_l = q8_ptr[b].qs + half * 512 + sb * 64;
+                        const int8_t * q8_base_h = q8_ptr[b].qs + half * 512 + 256 + sb * 64;
+
+                        int8x16_t q8_l_01[2];
+                        int8x16_t q8_l_23[2];
+                        for (int i = 0; i < 2; i++) {
+                            const int offset = i * 32;
+                            q8_l_01[i]       = vld1q_s8(q8_base_l + offset);       // 0..7 & 8..15 (r01)
+                            q8_l_23[i]       = vld1q_s8(q8_base_l + offset + 16);  // 0..7 & 8..15 (r23)
+                        }
+
+                        int8x16_t q8_h_01[2];
+                        int8x16_t q8_h_23[2];
+                        for (int i = 0; i < 2; i++) {
+                            const int offset = i * 32;
+                            q8_h_01[i]       = vld1q_s8(q8_base_h + offset);
+                            q8_h_23[i]       = vld1q_s8(q8_base_h + offset + 16);
+                        }
+
+                        const int ql_off_base = sb * QK_K / 2;
+
+                        uint8x16_t q6_ql_0[4];
+                        uint8x16_t q6_ql_1[4];
+                        for (int k = 0; k < 4; k++) {
+                            q6_ql_0[k] = vld1q_u8(ql_base + ql_off_base + 16 * k);
+                            q6_ql_1[k] = vld1q_u8(ql_base + ql_off_base + 64 + 16 * k);
+                        }
+
+                        const int  qh_off_base = (sb * QK_K / 2) & 255;  // wrap after 256 bytes
+                        uint8x16_t q6_qh_0[4];
+                        uint8x16_t q6_qh_1[4];
+                        for (int k = 0; k < 4; k++) {
+                            q6_qh_0[k] = vld1q_u8(qh_base + qh_off_base + 16 * k);
+                            q6_qh_1[k] = vld1q_u8(qh_base + qh_off_base + 64 + 16 * k);
+                        }
+
+                        // Adjust for the proper high bits (Sb 2 and 3)
+                        if (sb > 1) {
+                            for (int k = 0; k < 4; k++) {
+                                q6_qh_0[k] = vshrq_n_u8(q6_qh_0[k], 2);
+                                q6_qh_1[k] = vshrq_n_u8(q6_qh_1[k], 2);
+                            }
+                        }
+
+                        // Process column pairs (0-1, 2-3, 4-5, 6-7)
+                        for (int cp = 0; cp < ncols_interleaved / 2; cp++) {
+                            const uint8x16_t q6_qs_cp_0_l = q6_ql_0[cp];
+                            const uint8x16_t q6_qs_cp_1_l = q6_ql_1[cp];
+                            const uint8x16_t q6_qs_cp_0_h = q6_qh_0[cp];
+                            const uint8x16_t q6_qs_cp_1_h = q6_qh_1[cp];
+
+                            // Extract high 2 bits for upper nibble reconstruction
+                            const uint8x16_t q6_qs_cp_0_hh = vandq_u8(q6_qs_cp_0_h, mask_hi);
+                            const uint8x16_t q6_qs_cp_1_hh = vandq_u8(q6_qs_cp_1_h, mask_hi);
+
+                            // q6 = (low4 | high2<<4) - 32
+                            // Use vsliq_n_u8 to combine shift-left-insert in one instruction (like Q5_K)
+                            const int8x16_t q6_l0 = vsubq_s8(
+                                vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_qs_cp_0_l, m4b), vandq_u8(q6_qs_cp_0_h, mask_lo), 4)),
+                                m32s);
+                            const int8x16_t q6_l1 = vsubq_s8(
+                                vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_qs_cp_1_l, m4b), vandq_u8(q6_qs_cp_1_h, mask_lo), 4)),
+                                m32s);
+                            const int8x16_t q6_h0 = vsubq_s8(
+                                vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_qs_cp_0_l, 4), q6_qs_cp_0_hh)), m32s);
+                            const int8x16_t q6_h1 = vsubq_s8(
+                                vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_qs_cp_1_l, 4), q6_qs_cp_1_hh)), m32s);
+
+                            // row pair 0, base_l
+                            int32x4_t sb_acc_0l = vmmlaq_s32(vdupq_n_s32(0), q6_l0, q8_l_01[0]);
+                            sb_acc_0l           = vmmlaq_s32(sb_acc_0l, q6_l1, q8_l_01[1]);
+                            // row pair 0, base_h
+                            int32x4_t sb_acc_0h = vmmlaq_s32(vdupq_n_s32(0), q6_h0, q8_h_01[0]);
+                            sb_acc_0h           = vmmlaq_s32(sb_acc_0h, q6_h1, q8_h_01[1]);
+                            // row pair 1, base_l
+                            int32x4_t sb_acc_1l = vmmlaq_s32(vdupq_n_s32(0), q6_l0, q8_l_23[0]);
+                            sb_acc_1l           = vmmlaq_s32(sb_acc_1l, q6_l1, q8_l_23[1]);
+                            // row pair 1, base_h
+                            int32x4_t sb_acc_1h = vmmlaq_s32(vdupq_n_s32(0), q6_h0, q8_h_23[0]);
+                            sb_acc_1h           = vmmlaq_s32(sb_acc_1h, q6_h1, q8_h_23[1]);
+
+                            const int scale_idx_l = half * 8 + sb;
+                            const int scale_idx_h = half * 8 + sb + 4;
+
+                            const int32x4_t scale_vec_l = {
+                                q6_scales[scale_idx_l * 8 + cp * 2 + 0],
+                                q6_scales[scale_idx_l * 8 + cp * 2 + 0],
+                                q6_scales[scale_idx_l * 8 + cp * 2 + 1],
+                                q6_scales[scale_idx_l * 8 + cp * 2 + 1],
+                            };
+                            const int32x4_t scale_vec_h = {
+                                q6_scales[scale_idx_h * 8 + cp * 2 + 0],
+                                q6_scales[scale_idx_h * 8 + cp * 2 + 0],
+                                q6_scales[scale_idx_h * 8 + cp * 2 + 1],
+                                q6_scales[scale_idx_h * 8 + cp * 2 + 1],
+                            };
+
+                            acc[cp]     = vmlaq_s32(acc[cp], sb_acc_0l, scale_vec_l);
+                            acc[cp]     = vmlaq_s32(acc[cp], sb_acc_0h, scale_vec_h);
+                            acc[cp + 4] = vmlaq_s32(acc[cp + 4], sb_acc_1l, scale_vec_l);
+                            acc[cp + 4] = vmlaq_s32(acc[cp + 4], sb_acc_1h, scale_vec_h);
+                        }
+                    }
+                }  // for half
+
+                // Reorder i8mm output to match memory layout
+                for (int i = 0; i < 8; i++) {
+                    int32x2x2_t aux = vzip_s32(vget_low_s32(acc[i]), vget_high_s32(acc[i]));
+                    acc[i]          = vcombine_s32(aux.val[0], aux.val[1]);
+                }
+                int32x4_t reorder_acc[8] = {
+                    vcombine_s32(vget_low_s32(acc[0]), vget_low_s32(acc[1])),
+                    vcombine_s32(vget_low_s32(acc[2]), vget_low_s32(acc[3])),
+                    vcombine_s32(vget_high_s32(acc[0]), vget_high_s32(acc[1])),
+                    vcombine_s32(vget_high_s32(acc[2]), vget_high_s32(acc[3])),
+                    vcombine_s32(vget_low_s32(acc[4]), vget_low_s32(acc[5])),
+                    vcombine_s32(vget_low_s32(acc[6]), vget_low_s32(acc[7])),
+                    vcombine_s32(vget_high_s32(acc[4]), vget_high_s32(acc[5])),
+                    vcombine_s32(vget_high_s32(acc[6]), vget_high_s32(acc[7])),
+                };
+
+                // Apply superblock scale (no mins for q6_K)
+                for (int i = 0; i < q8_k_blocklen; i++) {
+                    for (int j = 0; j < 2; j++) {
+                        float32x4_t       q8_d  = vdupq_n_f32(q8_ptr[b].d[i]);
+                        float32x4_t       q6_d  = vcvt_f32_f16(vld1_f16((const __fp16 *) (q6_ptr[b].d + j * 4)));
+                        const float32x4_t scale = vmulq_f32(q6_d, q8_d);
+
+                        acc_f32[2 * i + j] =
+                            vmlaq_f32(acc_f32[2 * i + j], vcvtq_f32_s32(reorder_acc[2 * i + j]), scale);
+                    }
+                }
+            }  // for b
+
+            // Store results
+            for (int i = 0; i < q8_k_blocklen; i++) {
+                int row = y * q8_k_blocklen + i;
+                for (int j = 0; j < 2; j++) {
+                    int col    = x * ncols_interleaved + j * 4;
+                    int offset = row * bs + col;
+                    vst1q_f32(s + offset, acc_f32[2 * i + j]);
+                }
+            }
+        }  // for x
+    }  // for y
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    ggml_gemm_q6_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
 
 void ggml_gemm_q8_0_4x4_q8_0(int                        n,
                              float * GGML_RESTRICT      s,
diff --git a/ggml/src/ggml-cpu/common.h b/ggml/src/ggml-cpu/common.h
index 6adca5437f..1057b5bb15 100644
--- a/ggml/src/ggml-cpu/common.h
+++ b/ggml/src/ggml-cpu/common.h
@@ -6,6 +6,9 @@
 #include "ggml-impl.h"
 #include "simd-mappings.h"
 
+#define GGML_FA_TILE_Q  32
+#define GGML_FA_TILE_KV 16
+
 #ifdef __cplusplus
 
 #include <utility>
@@ -84,4 +87,9 @@ static std::pair<int64_t, int64_t> get_thread_range(const struct ggml_compute_pa
     return {ir0, ir1};
 }
 
+struct ggml_fa_tile_config {
+    static constexpr size_t Q  = GGML_FA_TILE_Q;
+    static constexpr size_t KV = GGML_FA_TILE_KV;
+};
+
 #endif
diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
index f7ba1fe317..b1de2ae871 100644
--- a/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -14,6 +14,7 @@
 #include "vec.h"
 #include "ops.h"
 #include "ggml.h"
+#include "common.h"
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <malloc.h> // using malloc.h with MSC/MINGW
@@ -2866,10 +2867,12 @@ struct ggml_cplan ggml_graph_plan(
                     } break;
                 case GGML_OP_FLASH_ATTN_EXT:
                     {
-                        const int64_t ne10 = node->src[1]->ne[0]; // DK
-                        const int64_t ne20 = node->src[2]->ne[0]; // DV
+                        const int64_t DK = node->src[1]->ne[0];
+                        const int64_t DV = node->src[2]->ne[0];
 
-                        cur = sizeof(float)*(1*ne10 + 2*ne20)*n_tasks; // 1x head size K + 2x head size V (per thread)
+                        // Tiled flash attention scratch (tile sizes defined in common.h)
+                        // Per-thread: Q_q + KQ + mask + VKQ32 + V32 + padding
+                        cur = sizeof(float)*(GGML_FA_TILE_Q*DK + 2*GGML_FA_TILE_Q*GGML_FA_TILE_KV + GGML_FA_TILE_Q*DV + GGML_FA_TILE_KV*DV)*n_tasks;
                     } break;
                 case GGML_OP_FLASH_ATTN_BACK:
                     {
@@ -2943,6 +2946,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
             continue;
         }
 
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
+
         ggml_compute_forward(&params, node);
 
         if (state->ith == 0 && cplan->abort_callback &&
diff --git a/ggml/src/ggml-cpu/llamafile/sgemm.cpp b/ggml/src/ggml-cpu/llamafile/sgemm.cpp
index 7dc36d4f8a..8f980c16b9 100644
--- a/ggml/src/ggml-cpu/llamafile/sgemm.cpp
+++ b/ggml/src/ggml-cpu/llamafile/sgemm.cpp
@@ -1797,10 +1797,27 @@ class tinyBLAS_Q0_AVX {
       } \
    } \
 
+template<typename T>
+struct mma_instr;
+
+template<>
+struct mma_instr<ggml_bf16_t> {
+    static inline void outer_product(acc_t *acc, vec_t a, vec_t b) {
+        __builtin_mma_xvbf16ger2pp(acc, a, b);
+    }
+};
+
+template<>
+struct mma_instr<ggml_fp16_t> {
+    static inline void outer_product(acc_t *acc, vec_t a, vec_t b) {
+        __builtin_mma_xvf16ger2pp(acc, a, b);
+    }
+};
+
 template <typename TA, typename TB, typename TC>
-class tinyBLAS_BF16_PPC {
+class tinyBLAS_HP16_PPC {
   public:
-    tinyBLAS_BF16_PPC(int64_t k,
+    tinyBLAS_HP16_PPC(int64_t k,
                 const TA *A, int64_t lda,
                 const TB *B, int64_t ldb,
                 TC *C, int64_t ldc,
@@ -2118,8 +2135,8 @@ class tinyBLAS_BF16_PPC {
             packNormal((A+(ii*lda)+l), lda, 4, 8, (uint8_t*)vec_A);
             packNormal((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B);
             for (int x = 0; x < 4; x++) {
-                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x], vec_B[x+4]);
+                mma_instr<TA>::outer_product(&acc_0, vec_A[x], vec_B[x]);
+                mma_instr<TA>::outer_product(&acc_1, vec_A[x], vec_B[x+4]);
             }
         }
         SAVE_ACC(&acc_0, ii, jj);
@@ -2135,8 +2152,8 @@ class tinyBLAS_BF16_PPC {
             packNormal((A+(ii*lda)+l), lda, 8, 8, (uint8_t*)vec_A);
             packNormal((B+(jj*ldb)+l), ldb, 8, 4, (uint8_t*)vec_B);
             for (int x = 0; x < 4; x++) {
-                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x+4], vec_B[x]);
+                mma_instr<TA>::outer_product(&acc_0, vec_A[x], vec_B[x]);
+                mma_instr<TA>::outer_product(&acc_1, vec_A[x], vec_B[x+4]);
             }
         }
         SAVE_ACC(&acc_0, ii, jj);
@@ -2155,10 +2172,10 @@ class tinyBLAS_BF16_PPC {
             packNormal(A+(ii*lda)+l, lda, 8, 8, (uint8_t*)vec_A);
             packNormal(B+(jj*ldb)+l, ldb, 8, 8, (uint8_t*)vec_B);
             for (int x = 0; x < 4; x++) {
-                __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvbf16ger2pp(&acc_1, (vec_t)vec_A[x], (vec_t)vec_B[x+4]);
-                __builtin_mma_xvbf16ger2pp(&acc_2, (vec_t)vec_A[x+4], (vec_t)vec_B[x]);
-                __builtin_mma_xvbf16ger2pp(&acc_3, (vec_t)vec_A[x+4], (vec_t)vec_B[x+4]);
+                mma_instr<TA>::outer_product(&acc_0, vec_A[x], vec_B[x]);
+                mma_instr<TA>::outer_product(&acc_1, vec_A[x], vec_B[x+4]);
+                mma_instr<TA>::outer_product(&acc_2, vec_A[x+4], vec_B[x]);
+                mma_instr<TA>::outer_product(&acc_3, vec_A[x+4], vec_B[x+4]);
             }
         }
 
@@ -2189,7 +2206,7 @@ class tinyBLAS_BF16_PPC {
                 packNormal(A+(ii*lda)+l, lda, RM, 4, (uint8_t*)vec_A);
                 packNormal(B+(jj*ldb)+l, ldb, RN, 4, (uint8_t*)vec_B);
                 for (int x = 0; x<2; x++) {
-                    __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
+                    mma_instr<TA>::outer_product(&acc_0, vec_A[x], vec_B[x]);
                 }
             }
             __builtin_mma_disassemble_acc(vec_C, &acc_0);
@@ -2224,8 +2241,8 @@ class tinyBLAS_BF16_PPC {
                 packNormal(A+(ii*lda)+l, lda, RM, 8, (uint8_t*)vec_A);
                 packNormal(B+(jj*ldb)+l, ldb, RN, 8, (uint8_t*)vec_B);
                 for (int x = 0; x<4; x++) {
-                    __builtin_mma_xvbf16ger2pp(&acc_0, vec_A[x], vec_B[x]);
-                    __builtin_mma_xvbf16ger2pp(&acc_1, vec_A[x], vec_B[x+4]);
+                    mma_instr<TA>::outer_product(&acc_0, vec_A[x], vec_B[x]);
+                    mma_instr<TA>::outer_product(&acc_1, vec_A[x], vec_B[x+4]);
                 }
             }
             __builtin_mma_disassemble_acc(vec_C, &acc_0);
@@ -3418,16 +3435,19 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             return tb.matmul(m, n);
         }
 #elif defined(__MMA__)
-        if ((k % 8))
-                return false;
-        if(Btype == GGML_TYPE_BF16) {
-           tinyBLAS_BF16_PPC<ggml_bf16_t, ggml_bf16_t, float> tb{ k,
-            (const ggml_bf16_t *)A, lda,
-            (const ggml_bf16_t *)B, ldb,
-            (float *)C, ldc,
-            params->ith, params->nth};
-        tb.matmul(m, n);
-        return true;
+        if (k % 8) {
+            return false;
+        }
+
+        if (Btype == GGML_TYPE_BF16) {
+            tinyBLAS_HP16_PPC<ggml_bf16_t, ggml_bf16_t, float> tb{ k,
+                (const ggml_bf16_t *)A, lda,
+                (const ggml_bf16_t *)B, ldb,
+                (float *)C, ldc,
+                params->ith, params->nth };
+
+            tb.matmul(m, n);
+            return true;
         }
 #elif defined(__riscv_zvfbfwma)
         #if LMUL == 1
@@ -3516,6 +3536,21 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
         #endif
             return tb.matmul(m, n);
         }
+#elif defined(__MMA__)
+        if (k % 8) {
+            return false;
+        }
+
+        if (Btype == GGML_TYPE_F16) {
+            tinyBLAS_HP16_PPC<ggml_fp16_t, ggml_fp16_t, float> tb{ k,
+                (const ggml_fp16_t *)A, lda,
+                (const ggml_fp16_t *)B, ldb,
+                (float *)C, ldc,
+                params->ith, params->nth };
+
+            tb.matmul(m, n);
+            return true;
+        }
 #endif
         return false;
     }
diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
index 387e2fe42c..48c8964361 100644
--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
@@ -8164,6 +8164,7 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
         // online softmax / attention
         // loop over n_kv and n_head_kv
         // ref: https://arxiv.org/pdf/2112.05682.pdf
+
         for (int64_t ic = 0; ic < nek1; ++ic) {
             const float mv = mp ? slope*GGML_CPU_FP16_TO_FP32(mp[ic]) : 0.0f;
             if (mv == -INFINITY) {
@@ -8271,6 +8272,280 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
     }
 }
 
+static void ggml_compute_forward_flash_attn_ext_tiled(
+        const ggml_compute_params * params,
+        ggml_tensor * dst,
+        int ir0, int ir1) {
+    const ggml_tensor * q     = dst->src[0];
+    const ggml_tensor * k     = dst->src[1];
+    const ggml_tensor * v     = dst->src[2];
+    const ggml_tensor * mask  = dst->src[3];
+    const ggml_tensor * sinks = dst->src[4];
+
+    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
+
+    const int64_t DK = nek0;
+    const int64_t DV = nev0;
+    const int64_t N  = neq1;
+
+    GGML_ASSERT(ne0 == DV);
+    GGML_ASSERT(ne2 == N);
+
+    // input tensor rows must be contiguous
+    GGML_ASSERT(nbq0 == ggml_type_size(q->type));
+    GGML_ASSERT(nbk0 == ggml_type_size(k->type));
+    GGML_ASSERT(nbv0 == ggml_type_size(v->type));
+
+    GGML_ASSERT(neq0 == DK);
+    GGML_ASSERT(nek0 == DK);
+    GGML_ASSERT(nev0 == DV);
+
+    GGML_ASSERT(neq1 == N);
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    GGML_ASSERT(k->type == v->type);
+    const ggml_type kv_type = k->type;
+
+    const auto * kv_type_traits_cpu = ggml_get_type_traits_cpu(kv_type);
+    const ggml_from_float_t kv_from_float = kv_type_traits_cpu->from_float;
+    const ggml_vec_dot_t    kv_vec_dot    = kv_type_traits_cpu->vec_dot;
+    const size_t kv_type_size = ggml_type_size(kv_type);
+
+    // broadcast factors
+    const int64_t rk2 = neq2/nek2;
+    const int64_t rk3 = neq3/nek3;
+
+    const int64_t rv2 = neq2/nev2;
+    const int64_t rv3 = neq3/nev3;
+
+    float scale         = 1.0f;
+    float max_bias      = 0.0f;
+    float logit_softcap = 0.0f;
+
+    memcpy(&scale,         (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias,      (float *) dst->op_params + 1, sizeof(float));
+    memcpy(&logit_softcap, (float *) dst->op_params + 2, sizeof(float));
+
+    if (logit_softcap != 0) {
+        scale /= logit_softcap;
+    }
+
+    const uint32_t n_head      = neq2;
+    const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    int ith = params->ith;
+
+    static constexpr int Q_TILE_SZ  = ggml_fa_tile_config::Q;
+    static constexpr int KV_TILE_SZ = ggml_fa_tile_config::KV;
+
+    GGML_ASSERT(nek1 % KV_TILE_SZ == 0 && "KV sequence length must be divisible by KV_TILE_SZ");
+
+    int ir = ir0;
+    while (ir < ir1) {
+        // q indices for the start of this tile
+        const int iq3 = ir/(neq2*neq1);
+        const int iq2 = (ir - iq3*neq2*neq1)/neq1;
+        const int iq1 = (ir - iq3*neq2*neq1 - iq2*neq1);
+
+        // Number of valid rows in this tile:
+        // - limited by tile size (Q_TILE_SZ)
+        // - limited by chunk boundary (ir1 - ir)
+        // - limited by head boundary (neq1 - iq1) to avoid crossing into next head
+        const int tile_rows = MIN(Q_TILE_SZ, MIN((int)(ir1 - ir), (int)(neq1 - iq1)));
+        GGML_ASSERT(tile_rows > 0);
+
+        const uint32_t h = iq2; // head index
+        const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f;
+
+        float S[Q_TILE_SZ];
+        float M[Q_TILE_SZ];
+
+        for (int i = 0 ; i < Q_TILE_SZ; ++i) {
+            S[i] = 0.;
+            M[i] = -INFINITY;
+        }
+
+        // Per-thread scratch layout:
+        // Q_q:    Q_TILE_SZ * DK (converted Q tile in KV type)
+        // KQ:     Q_TILE_SZ * KV_TILE_SZ (attention scores in float)
+        // mask:   Q_TILE_SZ * KV_TILE_SZ (mask in float)
+        // VKQ32:  Q_TILE_SZ * DV (FP32 output accumulator)
+        // V32:    KV_TILE_SZ * DV (F32 buffer for V tile - used for f166 conversion)
+        float * base  = (float *) params->wdata + ith*(Q_TILE_SZ*DK + 2*Q_TILE_SZ*KV_TILE_SZ + Q_TILE_SZ*DV + KV_TILE_SZ*DV + CACHE_LINE_SIZE_F32);
+
+        void  * Q_q    = base;
+        float * KQ     = (float *)((char *)base + Q_TILE_SZ * DK * sizeof(float));
+        float * mask32 = KQ + Q_TILE_SZ * KV_TILE_SZ;
+        float * VKQ32  = mask32 + Q_TILE_SZ * KV_TILE_SZ;
+        float * V32    = VKQ32 + Q_TILE_SZ * DV;  // F32 buffer for V tile
+
+        memset(VKQ32, 0, Q_TILE_SZ * DV * sizeof(float));
+        memset(mask32, 0, Q_TILE_SZ * KV_TILE_SZ * sizeof(float));
+
+        // k indices
+        const int ik3 = iq3 / rk3;
+        const int ik2 = iq2 / rk2;
+
+        // v indices
+        const int iv3 = iq3 / rv3;
+        const int iv2 = iq2 / rv2;
+
+        for (int tq = 0; tq < tile_rows; tq++) {
+            const float * pq = (const float *) ((char *) q->data + ((iq1 + tq)*nbq1 + iq2*nbq2 + iq3*nbq3));
+            kv_from_float(pq, (char *)Q_q + tq * DK * kv_type_size, DK);
+        }
+        // Zero-pad remaining rows
+        for (int tq = tile_rows; tq < Q_TILE_SZ; tq++) {
+            memset((char *)Q_q + tq * DK * kv_type_size, 0, DK * kv_type_size);
+        }
+
+        for (int64_t ic = 0; ic < nek1; ic += KV_TILE_SZ) {
+
+            // skip the tile entirely if all the masks are -inf
+            if (mask) {
+                bool can_skip = true;
+                for (int tq = 0; tq < tile_rows; tq++) {
+                    const ggml_fp16_t * mp_row = (const ggml_fp16_t *)((const char *) mask->data + (iq1 + tq)*mask->nb[1] + (iq2%mask->ne[2])*mask->nb[2] + (iq3%mask->ne[3])*mask->nb[3]);
+                    for (int tk = 0; tk < KV_TILE_SZ; tk++) {
+                        mask32[tq * KV_TILE_SZ + tk] = slope * GGML_CPU_FP16_TO_FP32(mp_row[ic + tk]);
+                        if (mask32[tq * KV_TILE_SZ + tk] != -INFINITY) {
+                            can_skip = false;
+                        }
+                    }
+                }
+
+                if (can_skip) {
+                    continue;
+                }
+            }
+
+            for (int tq = 0; tq < Q_TILE_SZ; tq++) {
+                const void * q_row = (const char *)Q_q + tq * DK * kv_type_size;
+                for (int tk = 0; tk < KV_TILE_SZ; tk++) {
+                    const void * k_row = (const char *) k->data + ((ic + tk)*nbk1 + ik2*nbk2 + ik3*nbk3);
+                    float s;
+                    kv_vec_dot(DK, &s, 0, k_row, 0, q_row, 0, 1);
+                    KQ[tq * KV_TILE_SZ + tk] = s * scale;
+                }
+            }
+
+            if (logit_softcap != 0.0f) {
+                ggml_vec_tanh_f32(Q_TILE_SZ * KV_TILE_SZ, KQ, KQ);
+                ggml_vec_scale_f32(Q_TILE_SZ * KV_TILE_SZ, KQ, logit_softcap);
+            }
+
+            if (mask) {
+                ggml_vec_add_f32(tile_rows * KV_TILE_SZ, KQ, KQ, mask32);
+            }
+
+            bool skip[Q_TILE_SZ] = {};
+
+            for (int tq = 0; tq < Q_TILE_SZ; tq++) {
+                float * kq_row = KQ + tq * KV_TILE_SZ;
+
+                float tile_max;
+                ggml_vec_max_f32(KV_TILE_SZ, &tile_max, kq_row);
+
+                if (tile_max == -INFINITY) {
+                    skip[tq] = true;
+                    continue;
+                }
+
+                const float Mold = M[tq];
+                const float Mnew = fmaxf(Mold, tile_max);
+
+                if (Mnew > Mold) {
+                    const float ms = expf(Mold - Mnew);
+                    ggml_vec_scale_f32(DV, VKQ32 + tq * DV, ms);
+                    S[tq] *= ms;
+                }
+                M[tq] = Mnew;
+
+
+                S[tq] += ggml_vec_soft_max_f32(KV_TILE_SZ, kq_row, kq_row, Mnew);
+            }
+
+            // Convert V tile to F32 first (if F16), then do MAD
+            // On x86, ggml_vec_mad_f16 internall converts F16<->F32 on every load/store, so pre-converting is faster.
+            // TODO: on ARM, native f16 should be faster
+            if (kv_type == GGML_TYPE_F16) {
+                for (int tk = 0; tk < KV_TILE_SZ; tk++) {
+                    const ggml_fp16_t * v_row = (const ggml_fp16_t *)((const char *) v->data + ((ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3));
+                    ggml_fp16_to_fp32_row(v_row, V32 + tk * DV, DV);
+                }
+                for (int tq = 0; tq < Q_TILE_SZ; tq++) {
+                    if (skip[tq]) continue;
+                    float * vkq_row = VKQ32 + tq * DV;
+                    for (int tk = 0; tk < KV_TILE_SZ; tk++) {
+                        const float p = KQ[tq * KV_TILE_SZ + tk];
+                        ggml_vec_mad_f32(DV, vkq_row, V32 + tk * DV, p);
+                    }
+                }
+            } else {
+                for (int tq = 0; tq < Q_TILE_SZ; tq++) {
+                    if (skip[tq]) continue;
+                    float * vkq_row = VKQ32 + tq * DV;
+                    for (int tk = 0; tk < KV_TILE_SZ; tk++) {
+                        const float p = KQ[tq * KV_TILE_SZ + tk];
+                        const float * v_row = (const float *)((const char *) v->data + ((ic + tk)*nbv1 + iv2*nbv2 + iv3*nbv3));
+                        ggml_vec_mad_f32(DV, vkq_row, v_row, p);
+                    }
+                }
+            }
+        }
+
+        // sinks (apply only to valid rows in the tile)
+        if (sinks) {
+            const float s = ((float *)((char *) sinks->data))[h];
+
+            for (int tq = 0; tq < tile_rows; tq++) {
+                float ms = 1.0f;
+                float vs = 1.0f;
+
+                if (s > M[tq]) {
+                    ms = expf(M[tq] - s);
+                    ggml_vec_scale_f32(DV, VKQ32 + tq * DV, ms);
+                } else {
+                    vs = expf(s - M[tq]);
+                }
+
+                S[tq] = S[tq] * ms + vs;
+            }
+        }
+
+        for (int tq = 0; tq < tile_rows; tq++) {
+            // V /= S
+            const float S_inv = S[tq] == 0.0f ? 0.0f : 1.0f / S[tq];
+            ggml_vec_scale_f32(DV, VKQ32 + tq * DV, S_inv);
+
+            // dst indices
+            const int i1 = iq1 + tq;
+            const int i2 = iq2;
+            const int i3 = iq3;
+
+            // permute(0, 2, 1, 3)
+            memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, VKQ32 + tq * DV, nb1);
+        }
+
+        ir += tile_rows;
+    }
+}
+
 static void ggml_compute_forward_flash_attn_ext_f16(
         const ggml_compute_params * params,
         ggml_tensor * dst) {
@@ -8343,6 +8618,15 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     // The number of elements in each chunk
     const int64_t dr = (nr + nchunk - 1) / nchunk;
 
+    static constexpr int64_t KV_TILE_SZ = ggml_fa_tile_config::KV;
+    static constexpr int64_t Q_TILE_SZ  = ggml_fa_tile_config::Q;
+    const bool kv_is_f32_or_f16 = (k->type == GGML_TYPE_F32 || k->type == GGML_TYPE_F16);
+    const bool use_tiled = (q->type == GGML_TYPE_F32 &&
+                            kv_is_f32_or_f16 &&
+                            k->type == v->type &&
+                            nek1 % KV_TILE_SZ == 0 &&
+                            neq1 >= Q_TILE_SZ);  // Only use tiled for batch >= tile size
+
     // The first chunk comes from our thread_id, the rest will get auto-assigned.
     int current_chunk = ith;
 
@@ -8350,7 +8634,11 @@ static void ggml_compute_forward_flash_attn_ext_f16(
         const int64_t ir0 = dr * current_chunk;
         const int64_t ir1 = MIN(ir0 + dr, nr);
 
-        ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1);
+        if (use_tiled) {
+            ggml_compute_forward_flash_attn_ext_tiled(params, dst, ir0, ir1);
+        } else {
+            ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1);
+        }
 
         current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
     }
diff --git a/ggml/src/ggml-cpu/repack.cpp b/ggml/src/ggml-cpu/repack.cpp
index fbf7ed9432..24e8ab4618 100644
--- a/ggml/src/ggml-cpu/repack.cpp
+++ b/ggml/src/ggml-cpu/repack.cpp
@@ -474,15 +474,8 @@ void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     assert (n % qk == 0);
     assert (nc % ncols_interleaved == 0);
 
-    UNUSED(s);
     UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
     UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
 
     float sumf[8];
     float sum_minf[8];
@@ -616,6 +609,191 @@ void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
 }
 
+void ggml_gemv_q5_K_8x8_q8_K_generic(int                        n,
+                                     float * GGML_RESTRICT      s,
+                                     size_t                     bs,
+                                     const void * GGML_RESTRICT vx,
+                                     const void * GGML_RESTRICT vy,
+                                     int                        nr,
+                                     int                        nc) {
+    const int             qk                = QK_K;
+    const int             nb                = n / qk;
+    const int             ncols_interleaved = 8;
+    const int             blocklen          = 8;
+    static const uint32_t kmask1            = 0x3f3f3f3f;
+    static const uint32_t kmask2            = 0x0f0f0f0f;
+    static const uint32_t kmask3            = 0x03030303;
+
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float    sumf[8];
+    float    sum_minf[8];
+    uint32_t utmp[32];
+    int      sumi1;
+    int      sumi2;
+    int      sumi;
+
+    const block_q8_K * a_ptr = (const block_q8_K *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q5_Kx8 * b_ptr = (const block_q5_Kx8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j]     = 0.0;
+            sum_minf[j] = 0.0;
+        }
+        for (int l = 0; l < nb; l++) {
+            for (int sb = 0; sb < 8; sb++) {
+                memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
+                utmp[sb * 4 + 3]      = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                utmp[sb * 4 + 1]      = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                utmp[sb * 4 + 2]      = uaux_0;
+                utmp[sb * 4 + 0] &= kmask1;
+            }
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                uint8_t * scales_0 = (uint8_t *) utmp + (k / 4) * 32;
+                uint8_t * scales_1 = (uint8_t *) utmp + (k / 4) * 32 + 16;
+
+                const int qh_shift = (k / 4) * 2;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi1 = 0;
+                    sumi2 = 0;
+                    sumi  = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int b_qs_offset = k * ncols_interleaved * blocklen + j * blocklen + i;
+
+                        const int qh_idx      = (k * 8 + i) % 32;
+                        const int qh_chunk    = qh_idx / 8;
+                        const int qh_pos      = qh_idx % 8;
+                        const int b_qh_offset = qh_chunk * 64 + j * 8 + qh_pos;
+
+                        const uint8_t qh_val = b_ptr[l].qh[b_qh_offset];
+                        const uint8_t h0     = (qh_val >> qh_shift) & 1;
+                        const uint8_t h1     = (qh_val >> (qh_shift + 1)) & 1;
+
+                        const int v0 = (int8_t) ((b_ptr[l].qs[b_qs_offset] & 0xF) | (h0 << 4));
+                        const int v1 = (int8_t) ((b_ptr[l].qs[b_qs_offset] >> 4) | (h1 << 4));
+
+                        const int q8_offset = (k >> 2) * 64 + (k % 4) * blocklen + i;
+
+                        sumi1 = (v0 * a_ptr[l].qs[q8_offset]);
+                        sumi2 = (v1 * a_ptr[l].qs[q8_offset + 32]);
+                        sumi1 = sumi1 * scales_0[j];
+                        sumi2 = sumi2 * scales_1[j];
+                        sumi += sumi1 + sumi2;
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                }
+            }
+            for (int sb = 0; sb < 8; sb++) {
+                uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) *
+                                   GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
+        }
+    }
+}
+
+
+void ggml_gemv_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    constexpr int qk = QK_K;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[8];
+
+    const block_q8_K * a_ptr = (const block_q8_K *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0f;
+        }
+
+        for (int l = 0; l < nb; l++) {
+
+
+            for (int k = 0; k < 16; k++) {
+                // k = 0.. 7 weights 0-63 low, 64-127 high
+                // k = 8..15 weights 128-191 low, 192-255 high
+                const int base_l = (k / 8) * 128 + (k % 8) * 8;
+                const int base_h = base_l + 64;
+
+                const int scale_idx_l = base_l / 16;
+                const int scale_idx_h = base_h / 16;
+
+                // Bit shift cycles 0,2,4,6 for each 32-value group within a 128-value half
+                const int qh_shift_l = ((base_l % 128) / 32) * 2;
+                const int qh_shift_h = ((base_h % 128) / 32) * 2;
+
+                // qh_half: offset to the correct 32-byte half (0 or 32)
+                const int qh_half_l = (base_l / 128) * 32;
+                const int qh_half_h = (base_h / 128) * 32;
+
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    // Interleaved scales
+                    const int8_t scale_l = b_ptr[l].scales[scale_idx_l * 8 + j];
+                    const int8_t scale_h = b_ptr[l].scales[scale_idx_h * 8 + j];
+
+                    int sumi_l = 0;
+                    int sumi_h = 0;
+
+                    for (int i = 0; i < blocklen; i++) {
+                        const int ql_pos = k * 64 + j * 8 + i;
+                        const int l_4    = b_ptr[l].ql[ql_pos] & 0xF;
+                        const int hi_4   = (b_ptr[l].ql[ql_pos] >> 4) & 0xF;
+
+                        // qh indexing with 8-byte interleaving (like q5_K)
+                        const int qh_byte_l   = qh_half_l + ((base_l + i) % 32);
+                        const int qh_chunk_l  = qh_byte_l / 8;
+                        const int qh_pos_l    = qh_byte_l % 8;
+                        const int qh_offset_l = qh_chunk_l * 64 + j * 8 + qh_pos_l;
+                        const int hi_2_l      = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
+
+                        const int qh_byte_h   = qh_half_h + ((base_h + i) % 32);
+                        const int qh_chunk_h  = qh_byte_h / 8;
+                        const int qh_pos_h    = qh_byte_h % 8;
+                        const int qh_offset_h = qh_chunk_h * 64 + j * 8 + qh_pos_h;
+                        const int hi_2_h      = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
+
+                        const int q_l = ((hi_2_l << 4) | l_4) - 32;
+                        const int q_h = ((hi_2_h << 4) | hi_4) - 32;
+
+                        const int8_t a_l = a_ptr[l].qs[base_l + i];
+                        const int8_t a_h = a_ptr[l].qs[base_h + i];
+
+                        sumi_l += q_l * a_l;
+                        sumi_h += q_h * a_h;
+                    }
+
+                    sumf[j] +=
+                        (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                }
+            }
+        }
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -1046,15 +1224,7 @@ void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     assert (nr % 4 == 0);
     assert (nc % ncols_interleaved == 0);
 
-    UNUSED(s);
     UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
-    UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
 
     float sumf[4][8];
     float sum_minf[4][8];
@@ -1212,6 +1382,213 @@ void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     }
 }
 
+void ggml_gemm_q5_K_8x8_q8_K_generic(int                        n,
+                                     float * GGML_RESTRICT      s,
+                                     size_t                     bs,
+                                     const void * GGML_RESTRICT vx,
+                                     const void * GGML_RESTRICT vy,
+                                     int                        nr,
+                                     int                        nc) {
+    const int qk                = QK_K;
+    const int nb                = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen          = 8;
+
+    constexpr uint32_t kmask1 = 0x3f3f3f3f;
+    constexpr uint32_t kmask2 = 0x0f0f0f0f;
+    constexpr uint32_t kmask3 = 0x03030303;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    float    sumf[4][8];
+    float    sum_minf[4][8];
+    uint32_t utmp[32];
+    int      sumi1;
+    int      sumi2;
+    int      sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q5_Kx8 * b_ptr = (const block_q5_Kx8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumf[m][j]     = 0.0;
+                    sum_minf[m][j] = 0.0;
+                }
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int sb = 0; sb < 8; sb++) {
+                    memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
+                    utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                    utmp[sb * 4 + 1]      = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                    utmp[sb * 4 + 2]      = uaux_0;
+                    utmp[sb * 4 + 0] &= kmask1;
+                }
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    uint8_t * scales_0 = (uint8_t *) utmp + (k / 4) * 32;
+                    uint8_t * scales_1 = (uint8_t *) utmp + (k / 4) * 32 + 16;
+
+                    const int qh_shift = (k / 4) * 2;
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi1 = 0;
+                            sumi2 = 0;
+                            sumi  = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int b_qs_offset = k * ncols_interleaved * blocklen + j * blocklen + i;
+
+                                const int qh_idx      = (k * 8 + i) % 32;
+                                const int qh_chunk    = qh_idx / 8;
+                                const int qh_pos      = qh_idx % 8;
+                                const int b_qh_offset = qh_chunk * 64 + j * 8 + qh_pos;
+
+                                const uint8_t qh_val = b_ptr[l].qh[b_qh_offset];
+                                const uint8_t h0     = (qh_val >> qh_shift) & 1;
+                                const uint8_t h1     = (qh_val >> (qh_shift + 1)) & 1;
+
+                                const int v0 = (int8_t) ((b_ptr[l].qs[b_qs_offset] & 0xF) | (h0 << 4));
+                                const int v1 = (int8_t) ((b_ptr[l].qs[b_qs_offset] >> 4) | (h1 << 4));
+
+                                const int q8_offset = (k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i;
+
+                                sumi1 = (v0 * a_ptr[l].qs[q8_offset]);
+                                sumi2 = (v1 * a_ptr[l].qs[q8_offset + 128]);
+                                sumi1 = sumi1 * scales_0[j];
+                                sumi2 = sumi2 * scales_1[j];
+                                sumi += sumi1 + sumi2;
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+                for (int sb = 0; sb < 8; sb++) {
+                    uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
+                    for (int m = 0; m < 4; m++) {
+                        const int16_t * bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) *
+                                              GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
+                }
+            }
+        }
+    }
+}
+
+void ggml_gemm_q6_K_8x8_q8_K_generic(int                        n,
+                                     float * GGML_RESTRICT      s,
+                                     size_t                     bs,
+                                     const void * GGML_RESTRICT vx,
+                                     const void * GGML_RESTRICT vy,
+                                     int                        nr,
+                                     int                        nc) {
+    const int qk                = QK_K;
+    const int nb                = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen          = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+
+    float sumf[4][8];
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb);
+
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumf[m][j] = 0.0f;
+                }
+            }
+
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < 16; k++) {
+                    // k = 0.. 7 weights 0-63 low, 64-127 high
+                    // k = 8..15 weights 128-191 low, 192-255 high
+                    const int base_l = (k / 8) * 128 + (k % 8) * 8;
+                    const int base_h = base_l + 64;
+
+                    const int scale_idx_l = base_l / 16;
+                    const int scale_idx_h = base_h / 16;
+
+                    // Bit shift cycles 0,2,4,6 for each 32-value group within a 128-value half
+                    const int qh_shift_l = ((base_l % 128) / 32) * 2;
+                    const int qh_shift_h = ((base_h % 128) / 32) * 2;
+
+                    // qh_half: offset to the correct 32-byte half (0 or 32)
+                    const int qh_half_l = (base_l / 128) * 32;
+                    const int qh_half_h = (base_h / 128) * 32;
+
+                    // Activation base indices for q8_Kx4 interleaved format
+                    // Layout: 128-value halves (k/8), then 8-value sub-blocks (k%8) with stride 32
+                    const int q8_base = (k / 8) * 512 + (k % 8) * 32;
+
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            // Interleaved scales
+                            const int8_t scale_l = b_ptr[l].scales[scale_idx_l * 8 + j];
+                            const int8_t scale_h = b_ptr[l].scales[scale_idx_h * 8 + j];
+
+                            int sumi_l = 0;
+                            int sumi_h = 0;
+
+                            for (int i = 0; i < blocklen; i++) {
+                                const int ql_pos = k * 64 + j * 8 + i;
+                                const int l_4    = b_ptr[l].ql[ql_pos] & 0xF;
+                                const int hi_4   = (b_ptr[l].ql[ql_pos] >> 4) & 0xF;
+
+                                const int qh_idx_l    = qh_half_l + ((base_l + i) % 32);
+                                const int qh_chunk_l  = qh_idx_l / 8;
+                                const int qh_pos_l    = qh_idx_l % 8;
+                                const int qh_offset_l = qh_chunk_l * 64 + j * 8 + qh_pos_l;
+                                const int hi_2_l      = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
+
+                                const int qh_idx_h    = qh_half_h + ((base_h + i) % 32);
+                                const int qh_chunk_h  = qh_idx_h / 8;
+                                const int qh_pos_h    = qh_idx_h % 8;
+                                const int qh_offset_h = qh_chunk_h * 64 + j * 8 + qh_pos_h;
+                                const int hi_2_h      = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
+
+                                const int q_l = ((hi_2_l << 4) | l_4) - 32;
+                                const int q_h = ((hi_2_h << 4) | hi_4) - 32;
+
+                                const int8_t q8_l = a_ptr[l].qs[q8_base + m * 8 + i];
+                                const int8_t q8_h = a_ptr[l].qs[q8_base + m * 8 + i + 256];
+
+                                sumi_l += q_l * q8_l;
+                                sumi_h += q_h * q8_h;
+                            }
+
+                            sumf[m][j] += (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) *
+                                          a_ptr[l].d[m];
+                        }
+                    }
+                }
+            }
+
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+                }
+            }
+        }
+    }
+}
 
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
@@ -1612,8 +1989,7 @@ static block_q2_Kx8 make_block_q2_Kx8(block_q2_K * in, unsigned int blck_size_in
     // Every 16 byte is packed such that it contains scales and mins for corresponding sub blocks from Q2_K structure
     // For eg - First 16 bytes contains 16 scales and 16 mins - each of first and second sub blocks from different Q2_K structures
 
-    for(int i = 0; i < 128; i++){
-
+    for (int i = 0; i < 128; i++) {
         // Index for selecting which q2k super block
         int src1 = (i % 16) / 2;
         // Index for selecting scale
@@ -1622,7 +1998,141 @@ static block_q2_Kx8 make_block_q2_Kx8(block_q2_K * in, unsigned int blck_size_in
         out.scales[i] = in[src1].scales[src2];
     }
     return out;
+}
 
+static block_q5_Kx8 make_block_q5_Kx8(block_q5_K * in, unsigned int blck_size_interleave) {
+    block_q5_Kx8 out;
+    //Delta(scale) and dmin values of the eight Q5_K structures are copied onto the output interleaved structure
+    for (int i = 0; i < 8; i++) {
+        out.d[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d;
+    }
+
+    for (int i = 0; i < 8; i++) {
+        out.dmin[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.dmin;
+    }
+
+    const int end = QK_K * 4 / blck_size_interleave;
+
+    // Interleave Q5_K quants by taking 8 bytes at a time
+    for (int i = 0; i < end; ++i) {
+        int src_id     = i % 8;
+        int src_offset = (i / 8) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+
+        uint64_t elems;
+        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+    }
+
+    // Repeat for low bits 8 bytes at a time as well, since
+    // the high bits are interleaved in Q5_K and the index is
+    // qh_idx = (qs_idx % 32);
+    // qh_val = qh[qh_idx] >> (qs_idx / 32);
+    for (int i = 0; i < end / 4; ++i) {
+        int src_id     = i % 8;
+        int src_offset = (i / 8) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+
+        uint64_t elems;
+        memcpy(&elems, &in[src_id].qh[src_offset], sizeof(uint64_t));
+        memcpy(&out.qh[dst_offset], &elems, sizeof(uint64_t));
+    }
+
+    // The below logic is copied over from Q4_K
+    // The point is to unpack all the scales and mins for each sub block every time we load 12 bytes.
+    // Currently the Q5_K structure has 8 scales and 8 mins packed in 12 bytes ( 6 bits for each value)
+    // The output Q5_Kx8 structure has 96 bytes
+    // Every 12 byte is packed such that it contains scales and mins for corresponding sub blocks from Q5_K structure
+    // For eg - First 12 bytes contains 8 scales and 8 mins - each of first sub block from different Q5_K structures
+    uint8_t s[8], m[8];
+
+    for (int i = 0; i < 4; i++) {
+        for (int j = 0; j < 8; j++) {
+            s[j] = in[j].scales[i] & 63;
+            m[j] = in[j].scales[i + 4] & 63;
+        }
+
+        out.scales[i * 12]      = (s[0] & 63) + ((s[4] & 48) << 2);
+        out.scales[i * 12 + 1]  = (s[1] & 63) + ((s[5] & 48) << 2);
+        out.scales[i * 12 + 2]  = (s[2] & 63) + ((s[6] & 48) << 2);
+        out.scales[i * 12 + 3]  = (s[3] & 63) + ((s[7] & 48) << 2);
+        out.scales[i * 12 + 4]  = (m[0] & 63) + ((m[4] & 48) << 2);
+        out.scales[i * 12 + 5]  = (m[1] & 63) + ((m[5] & 48) << 2);
+        out.scales[i * 12 + 6]  = (m[2] & 63) + ((m[6] & 48) << 2);
+        out.scales[i * 12 + 7]  = (m[3] & 63) + ((m[7] & 48) << 2);
+        out.scales[i * 12 + 8]  = (s[4] & 15) + ((m[4] & 15) << 4);
+        out.scales[i * 12 + 9]  = (s[5] & 15) + ((m[5] & 15) << 4);
+        out.scales[i * 12 + 10] = (s[6] & 15) + ((m[6] & 15) << 4);
+        out.scales[i * 12 + 11] = (s[7] & 15) + ((m[7] & 15) << 4);
+    }
+
+    for (int i = 0; i < 4; i++) {
+        for (int j = 0; j < 8; j++) {
+            s[j] = ((in[j].scales[i] & 192) >> 2) | (in[j].scales[i + 8] & 15);
+            m[j] = ((in[j].scales[i + 4] & 192) >> 2) | ((in[j].scales[i + 8] & 240) >> 4);
+        }
+
+        out.scales[i * 12 + 48] = (s[0] & 63) + ((s[4] & 48) << 2);
+        out.scales[i * 12 + 49] = (s[1] & 63) + ((s[5] & 48) << 2);
+        out.scales[i * 12 + 50] = (s[2] & 63) + ((s[6] & 48) << 2);
+        out.scales[i * 12 + 51] = (s[3] & 63) + ((s[7] & 48) << 2);
+        out.scales[i * 12 + 52] = (m[0] & 63) + ((m[4] & 48) << 2);
+        out.scales[i * 12 + 53] = (m[1] & 63) + ((m[5] & 48) << 2);
+        out.scales[i * 12 + 54] = (m[2] & 63) + ((m[6] & 48) << 2);
+        out.scales[i * 12 + 55] = (m[3] & 63) + ((m[7] & 48) << 2);
+        out.scales[i * 12 + 56] = (s[4] & 15) + ((m[4] & 15) << 4);
+        out.scales[i * 12 + 57] = (s[5] & 15) + ((m[5] & 15) << 4);
+        out.scales[i * 12 + 58] = (s[6] & 15) + ((m[6] & 15) << 4);
+        out.scales[i * 12 + 59] = (s[7] & 15) + ((m[7] & 15) << 4);
+    }
+
+    return out;
+}
+
+static block_q6_Kx8 make_block_q6_Kx8(block_q6_K * in, unsigned int blck_size_interleave) {
+    block_q6_Kx8  out;
+    constexpr int n_blocks = 8;  // Kx8
+    for (int i = 0; i < n_blocks; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end_ls = QK_K * 4 / blck_size_interleave;
+    // Interleave Q6_K quants by taking 8 bytes at a time
+    for (int i = 0; i < end_ls; ++i) {
+        int src_id     = i % n_blocks;
+        int src_offset = (i / n_blocks) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+
+        uint64_t elem_ls;
+        memcpy(&elem_ls, &in[src_id].ql[src_offset], sizeof(uint64_t));
+        memcpy(&out.ql[dst_offset], &elem_ls, sizeof(uint64_t));
+    }
+
+    // Interleave high bits using same 8-byte pattern as low bits
+    const int end_hs = end_ls / 2;
+    for (int i = 0; i < end_hs; ++i) {
+        int src_id     = i % n_blocks;
+        int src_offset = (i / n_blocks) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+
+        uint64_t elem_hs;
+        memcpy(&elem_hs, &in[src_id].qh[src_offset], sizeof(uint64_t));
+        memcpy(&out.qh[dst_offset], &elem_hs, sizeof(uint64_t));
+    }
+
+    // The below logic is designed so as to unpack and rearrange scales in Q6_K
+    // The output Q6_Kx8 structure interleaves the 8 bit scales in the same fashion as the quants
+    // Q6_K structure has an 8-bit scale per 16 elements -> 16 scales
+    // scales: [0 bl0 0 bl1 ... 0 bl7][1 bl0 ... 1 bl7] ... [15 bl0 ... 15 bl7]  (bl = block)
+    constexpr int n_scales = QK_K / 16;
+
+    for (int i = 0; i < n_blocks; i++) {
+        for (int j = 0; j < n_scales; j++) {
+            out.scales[j * n_blocks + i] = in[i].scales[j];
+        }
+    }
+
+    return out;
 }
 
 static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
@@ -1706,7 +2216,7 @@ static int repack_q2_K_to_q2_K_8_bl(struct ggml_tensor * t, int interleave_block
 
     for (int b = 0; b < nrow; b += nrows_interleaved) {
         for (int64_t x = 0; x < nblocks; x++) {
-            for (int i  = 0; i < nrows_interleaved; i++ ) {
+            for (int i = 0; i < nrows_interleaved; i++) {
                 dst_tmp[i] = src[x + i * nblocks];
             }
             *dst++ = make_block_q2_Kx8(dst_tmp, interleave_block);
@@ -1718,6 +2228,67 @@ static int repack_q2_K_to_q2_K_8_bl(struct ggml_tensor * t, int interleave_block
     GGML_UNUSED(data_size);
 }
 
+static int repack_q5_K_to_q5_K_8_bl(struct ggml_tensor *       t,
+                                    int                        interleave_block,
+                                    const void * GGML_RESTRICT data,
+                                    size_t                     data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q5_K);
+    GGML_ASSERT(interleave_block == 8);
+    constexpr int nrows_interleaved = 8;
+
+    block_q5_Kx8 *     dst = (block_q5_Kx8 *) t->data;
+    const block_q5_K * src = (const block_q5_K *) data;
+    block_q5_K         dst_tmp[8];
+    int                nrow    = ggml_nrows(t);
+    int                nblocks = t->ne[0] / QK_K;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q5_K));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_q5_Kx8(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+}
+
+static int repack_q6_K_to_q6_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q6_K);
+    GGML_ASSERT(interleave_block == 8);
+    constexpr int nrows_interleaved = 8;
+
+    block_q6_Kx8 * dst = (block_q6_Kx8 *)t->data;
+    const block_q6_K * src = (const block_q6_K *) data;
+    block_q6_K dst_tmp[8];
+    int nrow = ggml_nrows(t);
+    int nblocks = t->ne[0] / QK_K;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q6_K));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_q6_Kx8(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+}
+
 static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
     GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
     GGML_ASSERT(interleave_block == 8);
@@ -1936,6 +2507,14 @@ template <> int repack<block_q2_K, 8, 8>(struct ggml_tensor * t, const void * da
     return repack_q2_K_to_q2_K_8_bl(t, 8, data, data_size);
 }
 
+template <> int repack<block_q5_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q5_K_to_q5_K_8_bl(t, 8, data, data_size);
+}
+
+template <> int repack<block_q6_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q6_K_to_q6_K_8_bl(t, 8, data, data_size);
+}
+
 template <> int repack<block_iq4_nl, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
     return repack_iq4_nl_to_iq4_nl_4_bl(t, 4, data, data_size);
 }
@@ -1973,6 +2552,17 @@ template <> void gemv<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t
     ggml_gemv_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <>
+void gemv<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int          n,
+                                            float *      s,
+                                            size_t       bs,
+                                            const void * vx,
+                                            const void * vy,
+                                            int          nr,
+                                            int          nc) {
+    ggml_gemv_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemv<block_q4_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemv_q4_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
 }
@@ -1981,8 +2571,12 @@ template <> void gemv<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
     ggml_gemv_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
-template <> void gemv<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
-    ggml_gemv_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+template <> void gemv<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemv<block_q6_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q6_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
 template <> void gemv<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
@@ -2013,20 +2607,35 @@ template <> void gemm<block_q4_0, 8, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t
     ggml_gemm_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
-template <> void gemm<block_q4_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
-    ggml_gemm_q4_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+template <>
+void gemm<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int          n,
+                                            float *      s,
+                                            size_t       bs,
+                                            const void * vx,
+                                            const void * vy,
+                                            int          nr,
+                                            int          nc) {
+    ggml_gemm_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
-template <> void gemm<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
-    ggml_gemm_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
+template <> void gemm<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_q4_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q4_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
 template <> void gemm<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
     ggml_gemm_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
-template <> void gemm<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
-    ggml_gemm_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+template <> void gemm<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_q6_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q6_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
 
 template <> void gemm<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
@@ -2393,20 +3002,19 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             for (int ir1 = 0; ir1 < nr1; ir1++) {
                 struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1);
 
-                const int id = row_mapping.i1; // selected expert index
+                const int id = row_mapping.i1;  // selected expert index
 
                 const int64_t i11 = id % ne11;
-                const int64_t i12 = row_mapping.i2; // row index in src1
+                const int64_t i12 = row_mapping.i2;  // row index in src1
 
-                const int64_t i1 = id;  // selected expert index
-                const int64_t i2 = i12; // row
+                const int64_t i1 = id;               // selected expert index
+                const int64_t i2 = i12;              // row
 
                 const auto * src1_col = (const char *) wdata + (i11 * nbw1 + i12 * nbw2);
 
-                gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
-                        (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01,
-                        src0_cur + src0_cur_start * nb01,
-                        src1_col, 1, src0_cur_end - src0_cur_start);
+                gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(
+                    ne00, (float *) ((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01,
+                    src0_cur + src0_cur_start * nb01, src1_col, 1, src0_cur_end - src0_cur_start);
             }
         }
 #undef MMID_MATRIX_ROW
@@ -2422,7 +3030,6 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
 }  // namespace ggml::cpu::repack
 
 static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(const struct ggml_tensor * cur) {
-
     // instance for Q4
     static const ggml::cpu::repack::tensor_traits<block_q4_0, 4, 4, GGML_TYPE_Q8_0> q4_0_4x4_q8_0;
     static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 4, GGML_TYPE_Q8_0> q4_0_4x8_q8_0;
@@ -2432,6 +3039,12 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
     static const ggml::cpu::repack::tensor_traits<block_q4_K, 4, 8, GGML_TYPE_Q8_K> q4_K_8x4_q8_K;
     static const ggml::cpu::repack::tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
 
+    // instance for Q5_K
+    static const ggml::cpu::repack::tensor_traits<block_q5_K, 8, 8, GGML_TYPE_Q8_K> q5_K_8x8_q8_K;
+
+    // instance for Q6_K
+    static const ggml::cpu::repack::tensor_traits<block_q6_K, 8, 8, GGML_TYPE_Q8_K> q6_K_8x8_q8_K;
+
     // instance for Q2
     static const ggml::cpu::repack::tensor_traits<block_q2_K, 8, 8, GGML_TYPE_Q8_K> q2_K_8x8_q8_K;
 
@@ -2482,6 +3095,18 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
                 return &q2_K_8x8_q8_K;
             }
         }
+    } else if (cur->type == GGML_TYPE_Q5_K) {
+        if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q5_K_8x8_q8_K;
+            }
+        }
+    } else if (cur->type == GGML_TYPE_Q6_K) {
+        if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q6_K_8x8_q8_K;
+            }
+        }
     } else if (cur->type == GGML_TYPE_IQ4_NL) {
         if (ggml_cpu_has_avx2()) {
             if (cur->ne[1] % 8 == 0) {
diff --git a/ggml/src/ggml-cpu/repack.h b/ggml/src/ggml-cpu/repack.h
index af98e70344..855320eeeb 100644
--- a/ggml/src/ggml-cpu/repack.h
+++ b/ggml/src/ggml-cpu/repack.h
@@ -44,6 +44,7 @@ struct block_q4_Kx8 {
 };
 
 static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding");
+
 struct block_q2_Kx8 {
     ggml_half d[8];      // super-block scale for quantized scales
     ggml_half dmin[8];   // super-block scale for quantized mins
@@ -52,6 +53,28 @@ struct block_q2_Kx8 {
 };
 
 static_assert(sizeof(block_q2_Kx8) == sizeof(ggml_half) * 16 + QK_K/2 + QK_K * 2, "wrong q2_K block size/padding");
+
+struct block_q5_Kx8 {
+    ggml_half d[8];              // super-block scale for quantized scales
+    ggml_half dmin[8];           // super-block scale for quantized mins
+    uint8_t   scales[96];        // scales and mins, quantized with 6 bits
+    uint8_t   qh[QK_K * 8 / 8];  // high bits of 5-bit quants
+    uint8_t   qs[QK_K * 8 / 2];  // low bits of 5-bit quants (in groups of 4)
+};
+
+static_assert(sizeof(block_q5_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 5,
+              "wrong q5_K block size/padding");
+
+struct block_q6_Kx8 {
+    ggml_half d[8];
+    int8_t    scales[QK_K / 16 * 8];
+    uint8_t   ql[QK_K / 2 * 8];  // low bits of 6-bit quants (groups of 2)
+    uint8_t   qh[QK_K / 4 * 8];  // high bits of 6-bit quants (groups of 4)
+};
+
+static_assert(sizeof(block_q6_Kx8) == sizeof(ggml_half) * 8 + QK_K / 16 * 8 + 3 * QK_K / 4 * 8,
+              "wrong q6_K block size/padding");
+
 struct block_q8_Kx4 {
     float d[4];              // delta
     int8_t qs[QK_K * 4];     // quants
@@ -85,17 +108,21 @@ void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
 void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q8_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -111,17 +138,21 @@ void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GG
 void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q8_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
diff --git a/ggml/src/ggml-cuda/argsort.cu b/ggml/src/ggml-cuda/argsort.cu
index 57c8a99a28..4896669c32 100644
--- a/ggml/src/ggml-cuda/argsort.cu
+++ b/ggml/src/ggml-cuda/argsort.cu
@@ -2,6 +2,9 @@
 
 #ifdef GGML_CUDA_USE_CUB
 #    include <cub/cub.cuh>
+#    if (CCCL_MAJOR_VERSION >= 3 && CCCL_MINOR_VERSION >= 1)
+#        define STRIDED_ITERATOR_AVAILABLE
+#    endif
 using namespace cub;
 #endif  // GGML_CUDA_USE_CUB
 
@@ -14,12 +17,14 @@ static __global__ void init_indices(int * indices, const int ncols, const int nr
     }
 }
 
+#ifndef STRIDED_ITERATOR_AVAILABLE
 static __global__ void init_offsets(int * offsets, const int ncols, const int nrows) {
     const int idx = blockIdx.x * blockDim.x + threadIdx.x;
     if (idx <= nrows) {
         offsets[idx] = idx * ncols;
     }
 }
+#endif  // STRIDED_ITERATOR_AVAILABLE
 
 #ifdef GGML_CUDA_USE_CUB
 void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
@@ -31,19 +36,22 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
                               cudaStream_t     stream) {
     ggml_cuda_pool_alloc<int>   temp_indices_alloc(pool, ncols * nrows);
     ggml_cuda_pool_alloc<float> temp_keys_alloc(pool, ncols * nrows);
-    ggml_cuda_pool_alloc<int>   offsets_alloc(pool, nrows + 1);
 
     int *   temp_indices = temp_indices_alloc.get();
     float * temp_keys    = temp_keys_alloc.get();
-    int *   d_offsets    = offsets_alloc.get();
 
     static const int block_size = 256;
     const dim3 grid_size((ncols + block_size - 1) / block_size, nrows);
     init_indices<<<grid_size, block_size, 0, stream>>>(temp_indices, ncols, nrows);
 
-    const dim3 offset_grid((nrows + block_size - 1) / block_size);
-    init_offsets<<<offset_grid, block_size, 0, stream>>>(d_offsets, ncols, nrows);
-
+#ifdef STRIDED_ITERATOR_AVAILABLE
+    auto offset_iterator = cuda::make_strided_iterator(cuda::make_counting_iterator(0), ncols);
+#else
+    ggml_cuda_pool_alloc<int> offsets_alloc(pool, nrows + 1);
+    int *                     offset_iterator = offsets_alloc.get();
+    const dim3                offset_grid((nrows + block_size - 1) / block_size);
+    init_offsets<<<offset_grid, block_size, 0, stream>>>(offset_iterator, ncols, nrows);
+#endif
     CUDA_CHECK(cudaMemcpyAsync(temp_keys, x, ncols * nrows * sizeof(float), cudaMemcpyDeviceToDevice, stream));
 
     size_t temp_storage_bytes = 0;
@@ -57,7 +65,7 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
             DeviceSegmentedSort::SortPairs(nullptr, temp_storage_bytes, temp_keys, temp_keys,  // keys (in-place)
                                            temp_indices, dst,                                  // values (indices)
                                            ncols * nrows, nrows,  // num items, num segments
-                                           d_offsets, d_offsets + 1, stream);
+                                           offset_iterator, offset_iterator + 1, stream);
         }
     } else {
         if (nrows == 1) {
@@ -66,7 +74,8 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
                                                  ncols, 0, sizeof(float) * 8, stream);
         } else {
             DeviceSegmentedSort::SortPairsDescending(nullptr, temp_storage_bytes, temp_keys, temp_keys, temp_indices,
-                                                     dst, ncols * nrows, nrows, d_offsets, d_offsets + 1, stream);
+                                                     dst, ncols * nrows, nrows, offset_iterator, offset_iterator + 1,
+                                                     stream);
         }
     }
 
@@ -80,7 +89,7 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
                                        ncols, 0, sizeof(float) * 8, stream);
         } else {
             DeviceSegmentedSort::SortPairs(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys, temp_indices, dst,
-                                           ncols * nrows, nrows, d_offsets, d_offsets + 1, stream);
+                                           ncols * nrows, nrows, offset_iterator, offset_iterator + 1, stream);
         }
     } else {
         if (nrows == 1) {
@@ -89,8 +98,8 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
                                                  ncols, 0, sizeof(float) * 8, stream);
         } else {
             DeviceSegmentedSort::SortPairsDescending(d_temp_storage, temp_storage_bytes, temp_keys, temp_keys,
-                                                     temp_indices, dst, ncols * nrows, nrows, d_offsets, d_offsets + 1,
-                                                     stream);
+                                                     temp_indices, dst, ncols * nrows, nrows, offset_iterator,
+                                                     offset_iterator + 1, stream);
         }
     }
 }
diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index eaaf87612d..a3256d59dd 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -53,6 +53,7 @@
 // While BW spans CC 1000, 1100 & 1200, we are integrating Tensor Core instructions available to 1200 family, see
 // https://docs.nvidia.com/cutlass/media/docs/cpp/blackwell_functionality.html#blackwell-sm120-gemms
 #define GGML_CUDA_CC_BLACKWELL       1200
+#define GGML_CUDA_CC_DGX_SPARK       1210
 #define GGML_CUDA_CC_RUBIN           1300
 #define GGML_CUDA_CC_OFFSET_AMD      0x1000000
 #define GGML_CUDA_CC_OFFSET_MTHREADS 0x0100000
@@ -1121,14 +1122,18 @@ struct ggml_tensor_extra_gpu {
 #endif
 
 struct ggml_cuda_graph_node_properties {
-    void * node_address;
+    void * node_data;
     ggml_op node_op;
+    enum ggml_type node_type;
+    int32_t flags;
     int64_t ne[GGML_MAX_DIMS];
     size_t nb[GGML_MAX_DIMS];
-    void * src_address[GGML_MAX_SRC];
+    void * src_data[GGML_MAX_SRC];
     int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
 };
 
+static_assert(std::is_trivial<ggml_cuda_graph_node_properties>::value, "ggml_cuda_graph_node_properties must be trivial");
+
 struct ggml_cuda_graph {
 #ifdef USE_CUDA_GRAPH
     ~ggml_cuda_graph() {
@@ -1148,6 +1153,12 @@ struct ggml_cuda_graph {
     int number_consecutive_updates = 0;
     std::vector<ggml_cuda_graph_node_properties> props;
 
+    // these are extra tensors (inputs) that participate in the ggml graph but are not nodes
+    // they properties also have to match in order to be able to safely reuse a CUDA graph
+    // ref: https://github.com/ggml-org/llama.cpp/pull/18583
+    // ref: https://github.com/ggml-org/llama.cpp/pull/19165
+    std::vector<ggml_cuda_graph_node_properties> extra;
+
     void record_update(bool use_graph, bool update_required) {
         if (use_graph && update_required) {
             number_consecutive_updates++;
@@ -1326,10 +1337,44 @@ struct ggml_backend_cuda_context {
     cudaStream_t streams[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = { { nullptr } };
     cublasHandle_t cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
 
-    std::unique_ptr<ggml_cuda_graph> cuda_graph;
-
     int curr_stream_no = 0;
 
+#ifdef USE_CUDA_GRAPH
+    // Map from first_node_ptr to cuda_graph - allows multiple graphs per context
+    // when the computation is split across CPU/GPU (e.g., with --n-cpu-moe)
+    std::unordered_map<const void *, std::unique_ptr<ggml_cuda_graph>> cuda_graphs;
+
+    ggml_cuda_graph * cuda_graph(const void * first_node_ptr) {
+        auto it = cuda_graphs.find(first_node_ptr);
+        if (it == cuda_graphs.end()) {
+            cuda_graphs[first_node_ptr] = std::make_unique<ggml_cuda_graph>();
+            return cuda_graphs[first_node_ptr].get();
+        }
+        return it->second.get();
+    }
+
+    // Check if any CUDA graph is enabled for this context (used by kernels that need to know
+    // if graphs are in use without having access to the specific graph key)
+    bool any_cuda_graph_enabled() const {
+        for (const auto & [key, graph] : cuda_graphs) {
+            if (graph && graph->is_enabled()) {
+                return true;
+            }
+        }
+        return false;
+    }
+
+    // Check if any CUDA graph has an instance for this context
+    bool any_cuda_graph_has_instance() const {
+        for (const auto & [key, graph] : cuda_graphs) {
+            if (graph && graph->instance != nullptr) {
+                return true;
+            }
+        }
+        return false;
+    }
+#endif // USE_CUDA_GRAPH
+
     explicit ggml_backend_cuda_context(int device) :
         device(device),
         name(GGML_CUDA_NAME + std::to_string(device)) {
diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh
index 8468ba8488..b6a7460da8 100644
--- a/ggml/src/ggml-cuda/fattn-common.cuh
+++ b/ggml/src/ggml-cuda/fattn-common.cuh
@@ -629,8 +629,8 @@ static __global__ void flash_attn_mask_to_KV_max(
 template<int D, int ncols1, int ncols2> // D == head size
 __launch_bounds__(D, 1)
 static __global__ void flash_attn_stream_k_fixup(
-        float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne03, const int ne11,
-        const int nbatch_fa) {
+        float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne03,
+        const int ne11, const int ne12, const int nbatch_fa) {
     constexpr int ncols = ncols1*ncols2;
 
     const int bidx0 = blockIdx.x;
@@ -641,11 +641,14 @@ static __global__ void flash_attn_stream_k_fixup(
 
     const float * dst_fixup_data = ((const float *) dst_fixup) + gridDim.x*(2*2*ncols);
 
-    const int iter_k = (ne11 + (nbatch_fa - 1)) / nbatch_fa;
-    const int iter_j = (ne01 + (ncols1    - 1)) / ncols1;
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
 
-    const int kbc0      = int64_t(bidx0 + 0)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
-    const int kbc0_stop = int64_t(bidx0 + 1)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
+    const int iter_k     = (ne11      + (nbatch_fa - 1)) / nbatch_fa;
+    const int iter_j     = (ne01      + (ncols1    - 1)) / ncols1;
+    const int iter_z_gqa = (gqa_ratio + (ncols2    - 1)) / ncols2;
+
+    const int kbc0      = int64_t(bidx0 + 0)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x;
+    const int kbc0_stop = int64_t(bidx0 + 1)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x;
 
     const bool did_not_have_any_data   = kbc0 == kbc0_stop;
     const bool wrote_beginning_of_tile = kbc0 % iter_k == 0;
@@ -654,15 +657,19 @@ static __global__ void flash_attn_stream_k_fixup(
         return;
     }
 
-    const int sequence = kbc0 / (iter_k*iter_j*(ne02/ncols2));
-    const int head = (kbc0 - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j);
-    const int jt = (kbc0 - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*head) / iter_k; // j index of current tile.
+    // z_KV == K/V head index, zt_gqa = Q head start index per K/V head, jt = token position start index
+    const int sequence =  kbc0 /(iter_k*iter_j*iter_z_gqa*ne12);
+    const int z_KV     = (kbc0 - iter_k*iter_j*iter_z_gqa*ne12 * sequence)/(iter_k*iter_j*iter_z_gqa);
+    const int zt_gqa   = (kbc0 - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV)/(iter_k*iter_j);
+    const int jt       = (kbc0 - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV - iter_k*iter_j * zt_gqa) / iter_k;
 
-    if (jt*ncols1 + j >= ne01) {
+    const int zt_Q = z_KV*gqa_ratio + zt_gqa*ncols2; // Global Q head start index.
+
+    if (jt*ncols1 + j >= ne01 || zt_gqa*ncols2 + c >= gqa_ratio) {
         return;
     }
 
-    dst += sequence*ne02*ne01*D + jt*ne02*(ncols1*D) + head*(ncols2*D) + (j*ne02 + c)*D + tid;
+    dst += sequence*ne02*ne01*D + jt*ne02*(ncols1*D) + zt_Q*D + (j*ne02 + c)*D + tid;
 
     // Load the partial result that needs a fixup:
     float dst_val = 0.0f;
@@ -681,7 +688,7 @@ static __global__ void flash_attn_stream_k_fixup(
     int bidx = bidx0 - 1;
     int kbc_stop = kbc0;
     while(true) {
-        const int kbc = int64_t(bidx)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
+        const int kbc = int64_t(bidx)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x;
         if (kbc == kbc_stop) { // Did not have any data.
             bidx--;
             kbc_stop = kbc;
@@ -778,13 +785,11 @@ void launch_fattn(
 ) {
     constexpr int ncols = ncols1 * ncols2;
 
-    const bool is_mla = DV == 512; // TODO better parameterization
-
     const ggml_tensor * Q = dst->src[0];
     const ggml_tensor * K = dst->src[1];
     const ggml_tensor * V = dst->src[2];
 
-    GGML_ASSERT(V || is_mla);
+    const bool V_is_K_view = V->view_src && (V->view_src == K || (V->view_src == K->view_src && V->view_offs == K->view_offs));
 
     const ggml_tensor * mask  = dst->src[3];
     const ggml_tensor * sinks = dst->src[4];
@@ -794,9 +799,9 @@ void launch_fattn(
     GGML_ASSERT(Q->type == GGML_TYPE_F32);
     GGML_ASSERT(KQV->type == GGML_TYPE_F32);
 
-    GGML_ASSERT(      Q->nb[0] == ggml_element_size(Q));
-    GGML_ASSERT(      K->nb[0] == ggml_element_size(K));
-    GGML_ASSERT(!V || V->nb[0] == ggml_element_size(V));
+    GGML_ASSERT(Q->nb[0] == ggml_element_size(Q));
+    GGML_ASSERT(K->nb[0] == ggml_element_size(K));
+    GGML_ASSERT(V->nb[0] == ggml_element_size(V));
 
     GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16);
 
@@ -817,10 +822,10 @@ void launch_fattn(
     size_t nb12 = K->nb[2];
     size_t nb13 = K->nb[3];
 
-    const char * V_data = V ? (const char *) V->data : nullptr;
-    size_t nb21 = V ? V->nb[1] : nb11;
-    size_t nb22 = V ? V->nb[2] : nb12;
-    size_t nb23 = V ? V->nb[3] : nb13;
+    const char * V_data = (const char *) V->data;
+    size_t nb21 = V->nb[1];
+    size_t nb22 = V->nb[2];
+    size_t nb23 = V->nb[3];
 
     if (need_f16_K && K->type != GGML_TYPE_F16) {
         const size_t bs = ggml_blck_size(K->type);
@@ -849,36 +854,45 @@ void launch_fattn(
         K_data = (char *) K_f16.ptr;
     }
 
-    if (V && need_f16_V && V->type != GGML_TYPE_F16) {
-        const size_t bs = ggml_blck_size(V->type);
-        const size_t ts = ggml_type_size(V->type);
-
-        V_f16.alloc(ggml_nelements(V));
-        if (ggml_is_contiguously_allocated(V)) {
-            to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(V->type);
-            to_fp16(V_data, V_f16.ptr, ggml_nelements(V), main_stream);
-            V_data = (char *) V_f16.ptr;
-
-            nb21 = nb21*bs*sizeof(half)/ts;
-            nb22 = nb22*bs*sizeof(half)/ts;
-            nb23 = nb23*bs*sizeof(half)/ts;
+    if (need_f16_V && V->type != GGML_TYPE_F16) {
+        if (V_is_K_view) {
+            V_data = K_data;
+            nb21   = nb11;
+            nb22   = nb12;
+            nb23   = nb13;
         } else {
-            GGML_ASSERT(V->nb[0] == ts);
-            to_fp16_nc_cuda_t to_fp16 = ggml_get_to_fp16_nc_cuda(V->type);
-            const int64_t s01 = nb21 / ts;
-            const int64_t s02 = nb22 / ts;
-            const int64_t s03 = nb23 / ts;
-            to_fp16(V_data, V_f16.ptr, V->ne[0], V->ne[1], V->ne[2], V->ne[3], s01, s02, s03, main_stream);
+            const size_t bs = ggml_blck_size(V->type);
+            const size_t ts = ggml_type_size(V->type);
 
-            nb21 = V->ne[0] * sizeof(half);
-            nb22 = V->ne[1] * nb21;
-            nb23 = V->ne[2] * nb22;
+            V_f16.alloc(ggml_nelements(V));
+            if (ggml_is_contiguously_allocated(V)) {
+                to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(V->type);
+                to_fp16(V_data, V_f16.ptr, ggml_nelements(V), main_stream);
+                V_data = (char *) V_f16.ptr;
+
+                nb21 = nb21*bs*sizeof(half)/ts;
+                nb22 = nb22*bs*sizeof(half)/ts;
+                nb23 = nb23*bs*sizeof(half)/ts;
+            } else {
+                GGML_ASSERT(V->nb[0] == ts);
+                to_fp16_nc_cuda_t to_fp16 = ggml_get_to_fp16_nc_cuda(V->type);
+                const int64_t s01 = nb21 / ts;
+                const int64_t s02 = nb22 / ts;
+                const int64_t s03 = nb23 / ts;
+                to_fp16(V_data, V_f16.ptr, V->ne[0], V->ne[1], V->ne[2], V->ne[3], s01, s02, s03, main_stream);
+
+                nb21 = V->ne[0] * sizeof(half);
+                nb22 = V->ne[1] * nb21;
+                nb23 = V->ne[2] * nb22;
+            }
+            V_data = (char *) V_f16.ptr;
         }
-        V_data = (char *) V_f16.ptr;
     }
 
-    const int ntiles_x = ((Q->ne[1] + ncols1 - 1) / ncols1);
-    const int ntiles_total = ntiles_x * (Q->ne[2] / ncols2) * Q->ne[3];
+    const int ntiles_x     = ((Q->ne[1] + ncols1 - 1) / ncols1);
+    const int gqa_ratio    = Q->ne[2] / K->ne[2];
+    const int ntiles_z_gqa = ((gqa_ratio + ncols2 - 1) / ncols2);
+    const int ntiles_total = ntiles_x * ntiles_z_gqa * K->ne[2] * Q->ne[3];
 
     // Optional optimization where the mask is scanned to determine whether part of the calculation can be skipped.
     // Only worth the overhead if there is at lease one FATTN_KQ_STRIDE x FATTN_KQ_STRIDE square to be skipped or
@@ -953,7 +967,7 @@ void launch_fattn(
 
         blocks_num.x = ntiles_x;
         blocks_num.y = parallel_blocks;
-        blocks_num.z = (Q->ne[2]/ncols2)*Q->ne[3];
+        blocks_num.z = ntiles_z_gqa*K->ne[2]*Q->ne[3];
 
         if (parallel_blocks > 1) {
             dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
@@ -1007,7 +1021,7 @@ void launch_fattn(
 
             flash_attn_stream_k_fixup<DV, ncols1, ncols2>
                 <<<blocks_num_combine, block_dim_combine, 0, main_stream>>>
-                ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], Q->ne[3], K->ne[1], nbatch_fa);
+                ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], Q->ne[3], K->ne[1], K->ne[2], nbatch_fa);
         }
     } else if (parallel_blocks > 1) {
         const dim3 block_dim_combine(DV, 1, 1);
diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
index e53bbc0502..0b8ef90794 100644
--- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
@@ -400,7 +400,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(
 }
 
 template<int DKQ, int DV, int ncols1, int ncols2, int nwarps,
-    bool use_logit_softcap, bool mla, bool needs_fixup, bool is_fixup, bool last_iter, bool oob_check,
+    bool use_logit_softcap, bool V_is_K_view, bool needs_fixup, bool is_fixup, bool last_iter, bool oob_check,
     typename T_A_KQ, typename T_B_KQ, typename T_C_KQ, typename T_A_VKQ, typename T_B_VKQ, typename T_C_VKQ>
 static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         const float2 * const __restrict__ Q_f2,
@@ -432,7 +432,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     constexpr int  ncols           = ncols1 * ncols2;
     constexpr int  cols_per_warp   = T_B_KQ::I;
     constexpr int  cols_per_thread = get_cols_per_thread();
-    constexpr int  np              = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
+    constexpr int  np              = cols_per_warp > ncols ? nwarps : nwarps * cols_per_warp/ncols; // Number of parallel CUDA warps per Q column.
     constexpr int  nbatch_fa       = ggml_cuda_fattn_mma_get_nbatch_fa(DKQ, DV, ncols);
     constexpr int  nbatch_K2       = ggml_cuda_fattn_mma_get_nbatch_K2(DKQ, DV, ncols);
     constexpr int  nbatch_V2       = ggml_cuda_fattn_mma_get_nbatch_V2(DKQ, DV, ncols);
@@ -442,8 +442,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     constexpr int stride_tile_Q = DKQ/2     + 4;
     constexpr int stride_tile_K = nbatch_K2 + 4;
 
-    static_assert(!mla || nbatch_K2 >= nbatch_V2, "bad nbatch_K2, nbatch_V2 for MLA");
-    constexpr int stride_tile_V = mla ? stride_tile_K : nbatch_V2 + 4;
+    constexpr int stride_tile_V = V_is_K_view ? stride_tile_K : nbatch_V2 + 4;
 
     const int k_VKQ_0 = kb0 * nbatch_fa;
 #if defined(TURING_MMA_AVAILABLE)
@@ -456,7 +455,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
 
     if constexpr (nstages > 1) {
         static_assert(!oob_check, "OOB check incompatible with multi-stage pipeline");
-        static_assert(!mla, "multi-stage loading not implemented for MLA");
+        static_assert(!V_is_K_view, "K data reuse not implemented multi-stage loading");
         static_assert(nbatch_K2 == DKQ/2, "batching not implemented for multi stage loading");
         constexpr bool use_cp_async = true;
         cp_async_wait_all();
@@ -471,8 +470,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         }
     }
 
+    // For MLA K and V have the same data.
+    // Therefore, iterate over K in reverse and later re-use the data if possible.
 #pragma unroll
-    for (int k0_start = 0; k0_start < DKQ/2; k0_start += nbatch_K2) {
+    for (int k0_start = (DKQ/2-1) - (DKQ/2-1) % nbatch_K2; k0_start >= 0; k0_start -= nbatch_K2) {
         const int k0_stop = k0_start + nbatch_K2 < DKQ/2 ? k0_start + nbatch_K2 : DKQ/2;
         const int k0_diff = k0_stop - k0_start;
 
@@ -510,7 +511,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 }
             }
         } else {
-            static_assert(cols_per_warp != 8, "cols_per_warp == 8 not implemented");
 #pragma unroll
             for (int k_KQ_0 = k0_start; k_KQ_0 < k0_stop; k_KQ_0 += T_A_KQ::J) {
                 load_ldmatrix(Q_B[0], tile_Q + (threadIdx.y / np)*(T_B_KQ::I*stride_tile_Q) + k_KQ_0, stride_tile_Q);
@@ -522,14 +522,18 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                     T_A_KQ K_A;
                     load_ldmatrix(K_A, tile_K + i_KQ_0*stride_tile_K + (k_KQ_0 - k0_start), stride_tile_K);
 
-                    // Wide version of KQ_C is column-major
+                    if constexpr (cols_per_warp == 8) {
+                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
+                    } else {
+                        // Wide version of KQ_C is column-major
 #if defined(AMD_WMMA_AVAILABLE)
-                    // RDNA matrix C is column-major.
-                    mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
+                        // RDNA matrix C is column-major.
+                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
 #else
-                    // swap A and B for CUDA.
-                    mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
+                        // swap A and B for CUDA.
+                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
 #endif // defined(AMD_WMMA_AVAILABLE)
+                    }
                 }
             }
         }
@@ -773,6 +777,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     }
 
     if constexpr (nstages > 1) {
+        static_assert(!V_is_K_view, "K data reuse not implemented multi-stage loading");
         // Preload K tile for next iteration:
         constexpr bool use_cp_async = true;
         cp_async_wait_all();
@@ -788,10 +793,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     }
 
 
-    // For MLA K and V have the same data.
-    // Therefore, iterate over V in reverse and re-use the data if possible.
-    static_assert(!mla || nstages <= 1, "combination of MLA and multi-stage loading not implemented");
-    constexpr int reusable_cutoff = mla ? (DKQ - 1) - (DKQ - 1) % (2*nbatch_K2) - (DKQ - DV) : DV;
 #if defined(AMD_WMMA_AVAILABLE) && !defined(LDMATRIX_TRANS_AVAILABLE)
     T_A_VKQ A_identity;
     make_identity_mat(A_identity);
@@ -799,12 +800,13 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
 
     // Calculate VKQ tile, need to use logical rather than physical elements for i0 due to transposition of V:
 #pragma unroll
-    for (int i0_stop = DV; i0_stop > 0; i0_stop -= 2*nbatch_V2) {
-        const int i0_start = i0_stop - 2*nbatch_V2 > 0 ? i0_stop - 2*nbatch_V2 : 0;
-        const int i0_diff  = i0_stop - i0_start;
+    for (int i0_start = 0; i0_start < DV; i0_start += 2*nbatch_V2) {
+        static_assert(DV % (2*nbatch_V2) == 0, "bad loop size");
+        const int i0_stop = i0_start + 2*nbatch_V2;
+        const int i0_diff = i0_stop - i0_start;
 
         if constexpr (nstages <= 1) {
-            if (i0_start < reusable_cutoff) {
+            if (!V_is_K_view || i0_stop > 2*nbatch_K2) {
                 constexpr bool use_cp_async = nstages == 1;
                 flash_attn_ext_f16_load_tile<stride_tile_V, nwarps, nbatch_fa, use_cp_async, oob_check>
                     (V_h2 + int64_t(k_VKQ_0)*stride_V + i0_start/2, tile_V, i0_diff/2, stride_V, k_VKQ_sup);
@@ -814,7 +816,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 __syncthreads();
             }
         }
-        const half2 * tile_V_i = i0_start < reusable_cutoff ? tile_V : tile_V + (i0_start - reusable_cutoff)/2;
+        const half2 * tile_V_i = !V_is_K_view || i0_stop > 2*nbatch_K2 ? tile_V : tile_V + i0_start/2;
 
 #if defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         constexpr int i0_stride = cols_per_warp == 8 ? T_C_VKQ::I : 2*T_C_VKQ::J;
@@ -917,7 +919,7 @@ template<int ncols> struct mma_tile_sizes {
 };
 #endif // defined(TURING_MMA_AVAILABLE)
 
-template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, bool use_logit_softcap, bool mla, bool needs_fixup, bool is_fixup>
+template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, bool use_logit_softcap, bool V_is_K_view, bool needs_fixup, bool is_fixup>
 static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         const float2 * const __restrict__ Q_f2,
         const half2  * const __restrict__ K_h2,
@@ -931,6 +933,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         const float logit_softcap,
         const uint3 ne01,
         const int ne02,
+        const int gqa_ratio,
         const int ne11,
         const int stride_Q1,
         const int stride_Q2,
@@ -938,6 +941,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         const int stride_V,
         const int stride_mask,
         const int jt,
+        const int zt_gqa,
         const int kb0_start,
         const int kb0_stop) {
 #if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
@@ -953,7 +957,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
 
     constexpr int  cols_per_warp   = T_B_KQ::I;
     constexpr int  cols_per_thread = get_cols_per_thread();
-    constexpr int  np              = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
+    constexpr int  np              = cols_per_warp > ncols ? nwarps : nwarps * cols_per_warp/ncols; // Number of parallel CUDA warps per Q column.
     constexpr int  nbatch_fa       = ggml_cuda_fattn_mma_get_nbatch_fa     (DKQ, DV, ncols);
     constexpr int  nbatch_K2       = ggml_cuda_fattn_mma_get_nbatch_K2     (DKQ, DV, ncols);
     constexpr int  nbatch_V2       = ggml_cuda_fattn_mma_get_nbatch_V2     (DKQ, DV, ncols);
@@ -971,8 +975,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
     constexpr int stride_tile_Q = DKQ/2     + 4;
     constexpr int stride_tile_K = nbatch_K2 + 4;
 
-    static_assert(!mla || nbatch_K2 >= nbatch_V2, "bad nbatch_K2, nbatch_V2 for MLA");
-    constexpr int stride_tile_V = mla ? stride_tile_K : nbatch_V2 + 4;
+    constexpr int stride_tile_V = V_is_K_view ? stride_tile_K : nbatch_V2 + 4;
     constexpr int stride_tile_KV_max = stride_tile_K > stride_tile_V ? stride_tile_K : stride_tile_V;
 
     extern __shared__ half2 tile_Q[];
@@ -1021,7 +1024,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
             const int j = jc / ncols2;
             const int c = jc % ncols2;
 
-            if (jt*ncols1 + j < int(ne01.z)) {
+            if ((ncols1 == 1 || jt*ncols1 + j < int(ne01.z)) && (ncols2 == 1 || zt_gqa*ncols2 + c < gqa_ratio)) {
 #pragma unroll
                 for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) {
                     const int k = k0 + (stride_k == WARP_SIZE ? threadIdx.x : threadIdx.x % stride_k);
@@ -1076,7 +1079,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
             constexpr bool last_iter = false;
             constexpr int  k_VKQ_sup = nbatch_fa;
             flash_attn_ext_f16_iter
-                <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter, oob_check,
+                <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup, last_iter, oob_check,
                  T_A_KQ, T_B_KQ, T_C_KQ, T_A_VKQ, T_B_VKQ, T_C_VKQ>
                 (Q_f2, K_h2, V_h2, mask_h, dstk, dstk_fixup, scale, slope, logit_softcap,
                  ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C,
@@ -1085,7 +1088,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         constexpr bool last_iter = true;
         const     int  k_VKQ_sup = ne11 - kb0*nbatch_fa;
         flash_attn_ext_f16_iter
-            <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter, oob_check,
+            <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup, last_iter, oob_check,
               T_A_KQ, T_B_KQ, T_C_KQ, T_A_VKQ, T_B_VKQ, T_C_VKQ>
             (Q_f2, K_h2, V_h2, mask_h, dstk, dstk_fixup, scale, slope, logit_softcap,
              ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C,
@@ -1096,7 +1099,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
             constexpr bool last_iter = false;
             constexpr int  k_VKQ_sup = nbatch_fa;
             flash_attn_ext_f16_iter
-                <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter, oob_check,
+                <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup, last_iter, oob_check,
                  T_A_KQ, T_B_KQ, T_C_KQ, T_A_VKQ, T_B_VKQ, T_C_VKQ>
                 (Q_f2, K_h2, V_h2, mask_h, dstk, dstk_fixup, scale, slope, logit_softcap,
                  ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C,
@@ -1105,7 +1108,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         constexpr bool last_iter = true;
         constexpr int  k_VKQ_sup = nbatch_fa;
         flash_attn_ext_f16_iter
-            <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter, oob_check,
+            <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup, last_iter, oob_check,
              T_A_KQ, T_B_KQ, T_C_KQ, T_A_VKQ, T_B_VKQ, T_C_VKQ>
             (Q_f2, K_h2, V_h2, mask_h, dstk, dstk_fixup, scale, slope, logit_softcap,
              ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C,
@@ -1407,7 +1410,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
                     const int j_dst = jc_dst / ncols2;
                     const int c_dst = jc_dst % ncols2;
 
-                    if (!is_fixup && jt*ncols1 + j_dst >= int(ne01.z)) {
+                    if (!is_fixup && ((ncols1 > 1 && jt*ncols1 + j_dst >= int(ne01.z)) || (ncols2 > 1 && zt_gqa*ncols2 + c_dst >= gqa_ratio))) {
                         continue;
                     }
 
@@ -1446,14 +1449,14 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
     }
 #else
     GGML_UNUSED_VARS(Q_f2, K_h2, V_h2, mask_h, sinks_f, dstk, dstk_fixup,
-        scale, slope, logit_softcap, ne01, ne02,
+        scale, slope, logit_softcap, ne01, ne02, gqa_ratio,
         stride_Q1, stride_Q2, stride_K, stride_V, stride_mask,
         jt, kb0_start, kb0_stop);
     NO_DEVICE_CODE;
 #endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
 }
 
-template<int DKQ, int DV, int ncols1, int ncols2, bool use_logit_softcap, bool mla>
+template<int DKQ, int DV, int ncols1, int ncols2, bool use_logit_softcap, bool V_is_K_view>
 __launch_bounds__(ggml_cuda_fattn_mma_get_nthreads(DKQ, DV, ncols1*ncols2), ggml_cuda_fattn_mma_get_occupancy(DKQ, DV, ncols1*ncols2))
 static __global__ void flash_attn_ext_f16(
         const char * __restrict__ Q,
@@ -1484,6 +1487,13 @@ static __global__ void flash_attn_ext_f16(
         NO_DEVICE_CODE;
         return;
     }
+#ifdef VOLTA_MMA_AVAILABLE
+    if (ncols1*ncols2 < 32) {
+        NO_DEVICE_CODE;
+        return;
+    }
+#endif // VOLTA_MMA_AVAILABLE
+
 #if __CUDA_ARCH__ == GGML_CUDA_CC_TURING
     if (ncols1*ncols2 > 32) {
         NO_DEVICE_CODE;
@@ -1498,8 +1508,6 @@ static __global__ void flash_attn_ext_f16(
     }
 #endif // defined(AMD_WMMA_AVAILABLE)
 
-    static_assert(!mla || DKQ >= DV, "MLA needs DKQ >= DV");
-
     constexpr int ncols     = ncols1 * ncols2;
     constexpr int nbatch_fa = ggml_cuda_fattn_mma_get_nbatch_fa(DKQ, DV, ncols);
     constexpr int nthreads  = ggml_cuda_fattn_mma_get_nthreads(DKQ, DV, ncols);
@@ -1512,14 +1520,15 @@ static __global__ void flash_attn_ext_f16(
     const int stride_K    = nb11 / sizeof(half2);
     const int stride_mask = nb31 / sizeof(half);
 
-    const int stride_V = mla ? stride_K : nb21 / sizeof(half2);
+    const int stride_V = V_is_K_view ? stride_K : nb21 / sizeof(half2);
 
-    const int iter_k = (ne11   + (nbatch_fa - 1)) / nbatch_fa;
-    const int iter_j = (ne01.z + (ncols1    - 1)) / ncols1;
+    const int iter_k     = (ne11      + (nbatch_fa - 1)) / nbatch_fa;
+    const int iter_j     = (ne01.z    + (ncols1    - 1)) / ncols1;
+    const int iter_z_gqa = (gqa_ratio + (ncols2    - 1)) / ncols2;
 
     // kbc == k block continuous, current index in continuous ijk space.
-    int       kbc      = int64_t(blockIdx.x + 0)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
-    const int kbc_stop = int64_t(blockIdx.x + 1)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
+    int       kbc      = int64_t(blockIdx.x + 0)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x;
+    const int kbc_stop = int64_t(blockIdx.x + 1)*(iter_k*iter_j*iter_z_gqa*ne12*ne03) / gridDim.x;
 
     // If the seams of 2 CUDA blocks fall within an output tile their results need to be combined.
     // For this we need to track both the block that starts the tile (needs_fixup) and the block that finishes the tile (is_fixup).
@@ -1530,22 +1539,24 @@ static __global__ void flash_attn_ext_f16(
     int kb0_stop  = min(iter_k, kb0_start + kbc_stop - kbc);
 
     while (kbc < kbc_stop && kb0_stop == iter_k) {
-        const int sequence = kbc / (iter_k*iter_j*(ne02/ncols2));
-        const int zt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j); // head in units of ncols2
-        const int jt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*zt) / iter_k; // j index of current tile.
+        // z_KV == K/V head index, zt_gqa = Q head start index per K/V head, jt = token position start index
+        const int sequence =  kbc /(iter_k*iter_j*iter_z_gqa*ne12);
+        const int z_KV     = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence)/(iter_k*iter_j*iter_z_gqa);
+        const int zt_gqa   = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV)/(iter_k*iter_j);
+        const int jt       = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV - iter_k*iter_j * zt_gqa) / iter_k;
 
-        const int head0 = zt * ncols2;
+        const int zt_Q = z_KV*gqa_ratio + zt_gqa*ncols2; // Global Q head start index.
 
-        const float2 * Q_f2   = (const float2 *) (Q + nb03*sequence + nb02* head0);
-        const half2  * K_h2   = (const half2  *) (K + nb13*sequence + nb12*(head0 / gqa_ratio));
+        const float2 * Q_f2   = (const float2 *) (Q + nb03*sequence + nb02*zt_Q);
+        const half2  * K_h2   = (const half2  *) (K + nb13*sequence + nb12*z_KV);
         const half   * mask_h = ncols2 == 1 && !mask ? nullptr :
             (const half *) (mask + nb33*(sequence % ne33));
-        float2       * dstk   = ((float2 *) dst) + (sequence*ne01.z*ne02 + head0) * (DV/2);
+        float2       * dstk   = ((float2 *) dst) + (sequence*ne01.z*ne02 + zt_Q) * (DV/2);
 
-        const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb23*sequence + nb22*(head0 / gqa_ratio));
-        const float * sinks_f = sinks ? (const float *) sinks + head0 : nullptr;
+        const half2 * V_h2 = V_is_K_view ? K_h2 : (const half2 *) (V + nb23*sequence + nb22*z_KV);
+        const float * sinks_f = sinks ? (const float *) sinks + zt_Q : nullptr;
 
-        const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head0, n_head_log2, m0, m1) : 1.0f;
+        const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, zt_Q, n_head_log2, m0, m1) : 1.0f;
 
         if (KV_max) {
             kb0_stop = min(kb0_stop, KV_max[sequence*iter_j + jt] / nbatch_fa);
@@ -1553,14 +1564,14 @@ static __global__ void flash_attn_ext_f16(
         constexpr bool is_fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer.
         if (kb0_start == 0) {
             constexpr bool needs_fixup = false; // CUDA block is working on an entire tile.
-            flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup>
+            flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup>
                 (Q_f2, K_h2, V_h2, mask_h, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
-                 ne01, ne02, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start, kb0_stop);
+                 ne01, ne02, gqa_ratio, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, zt_gqa, kb0_start, kb0_stop);
         } else {
             constexpr bool needs_fixup = true; // CUDA block is missing the beginning of a tile.
-            flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup>
+            flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup>
                 (Q_f2, K_h2, V_h2, mask_h, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
-                 ne01, ne02, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start, kb0_stop);
+                 ne01, ne02, gqa_ratio, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, zt_gqa, kb0_start, kb0_stop);
         }
 
         kbc += iter_k;
@@ -1574,22 +1585,24 @@ static __global__ void flash_attn_ext_f16(
         return;
     }
 
-    const int sequence = kbc / (iter_k*iter_j*(ne02/ncols2));
-    const int zt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence) / (iter_k*iter_j); // head in units of ncols2
-    const int jt = (kbc - iter_k*iter_j*(ne02/ncols2)*sequence - iter_k*iter_j*zt) / iter_k; // j index of current tile.
+    // z_KV == K/V head index, zt_gqa = Q head start index per K/V head, jt = token position start index.
+    const int sequence =  kbc /(iter_k*iter_j*iter_z_gqa*ne12);
+    const int z_KV     = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence)/(iter_k*iter_j*iter_z_gqa);
+    const int zt_gqa   = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV)/(iter_k*iter_j);
+    const int jt       = (kbc - iter_k*iter_j*iter_z_gqa*ne12 * sequence - iter_k*iter_j*iter_z_gqa * z_KV - iter_k*iter_j * zt_gqa) / iter_k;
 
-    const int head0 = zt * ncols2;
+    const int zt_Q = z_KV*gqa_ratio + zt_gqa*ncols2; // Global Q head start index.
 
-    const float2 * Q_f2   = (const float2 *) (Q + nb03*sequence + nb02* head0);
-    const half2  * K_h2   = (const half2  *) (K + nb13*sequence + nb12*(head0 / gqa_ratio));
+    const float2 * Q_f2   = (const float2 *) (Q + nb03*sequence + nb02*zt_Q);
+    const half2  * K_h2   = (const half2  *) (K + nb13*sequence + nb12*z_KV);
     const half   * mask_h = ncols2 == 1 && !mask ? nullptr :
         (const half *) (mask + nb33*(sequence % ne33));
-    float2       * dstk   = ((float2 *) dst) + (sequence*ne01.z*ne02 + head0) * (DV/2);
+    float2       * dstk   = ((float2 *) dst) + (sequence*ne01.z*ne02 + zt_Q) * (DV/2);
 
-    const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb23*sequence + nb22*(head0 / gqa_ratio));
-    const float * sinks_f = sinks ? (const float *) sinks + head0 : nullptr;
+    const half2 * V_h2 = V_is_K_view ? K_h2 : (const half2 *) (V + nb23*sequence + nb22*z_KV);
+    const float * sinks_f = sinks ? (const float *) sinks + zt_Q : nullptr;
 
-    const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, head0, n_head_log2, m0, m1) : 1.0f;
+    const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, zt_Q, n_head_log2, m0, m1) : 1.0f;
 
     if (KV_max) {
         kb0_stop = min(kb0_stop, KV_max[sequence*iter_j + jt] / nbatch_fa);
@@ -1597,9 +1610,9 @@ static __global__ void flash_attn_ext_f16(
 
     constexpr bool is_fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks.
     constexpr bool needs_fixup = false;
-    flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup>
+    flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup>
         (Q_f2, K_h2, V_h2, mask_h, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
-         ne01, ne02, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start, kb0_stop);
+         ne01, ne02, gqa_ratio, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, zt_gqa, kb0_start, kb0_stop);
 #else
     GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
         max_bias, m0, m1, n_head_log2, logit_softcap,
@@ -1633,7 +1646,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
     const int cols_per_warp = std::min(ncols, get_cols_per_warp(cc));
     const int nwarps        = nthreads / WARP_SIZE;
 
-    constexpr bool mla = DKQ == 576;
+    constexpr bool V_is_K_view = DKQ == 576; // Guaranteed by the kernel selection logic in fattn.cu
 
     const size_t nbytes_shared_KV_1stage = nbatch_fa            * std::max(nbatch_K2 + 4,  nbatch_V2 + 4) * sizeof(half2);
     const size_t nbytes_shared_KV_2stage = nbatch_fa            *         (nbatch_K2 + 4 + nbatch_V2 + 4) * sizeof(half2);
@@ -1658,7 +1671,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
     fattn_kernel_t fattn_kernel;
     if (logit_softcap == 0.0f) {
         constexpr bool use_logit_softcap = false;
-        fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, mla>;
+        fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, V_is_K_view>;
 
 #if !defined(GGML_USE_MUSA)
         static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
@@ -1669,7 +1682,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
 #endif // !defined(GGML_USE_MUSA)
     } else {
         constexpr bool use_logit_softcap = true;
-        fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, mla>;
+        fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, V_is_K_view>;
 
 #if !defined(GGML_USE_MUSA)
         static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
@@ -1728,3 +1741,10 @@ DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 256,  64)
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 1, 16);
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 2, 16);
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 4, 16);
+
+// For GLM 4.7 Flash
+extern DECL_FATTN_MMA_F16_CASE(576, 512,  4,  4);
+extern DECL_FATTN_MMA_F16_CASE(576, 512,  8,  4);
+extern DECL_FATTN_MMA_F16_CASE(576, 512, 16,  4);
+extern DECL_FATTN_MMA_F16_CASE(576, 512,  1, 32);
+extern DECL_FATTN_MMA_F16_CASE(576, 512,  2, 32);
diff --git a/ggml/src/ggml-cuda/fattn-tile.cuh b/ggml/src/ggml-cuda/fattn-tile.cuh
index f055da8e2b..b6db582281 100644
--- a/ggml/src/ggml-cuda/fattn-tile.cuh
+++ b/ggml/src/ggml-cuda/fattn-tile.cuh
@@ -68,6 +68,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
 
     return 0;
@@ -122,6 +124,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  32,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  32,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  32,  64)
 
     return 0;
@@ -183,6 +187,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32, 128)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 512, 1, 128,  64)
 
@@ -245,6 +251,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 5,  32, 256)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 3,  64, 128)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 4,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 256, 2, 128,  64)
 
@@ -1187,6 +1195,10 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
             launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
             return;
         }
+        if (use_gqa_opt && gqa_ratio % 4 == 0) {
+            launch_fattn_tile_switch_ncols1<DKQ, DV, 4, use_logit_softcap>(ctx, dst);
+            return;
+        }
     }
 
     if constexpr (DV <= 256) {
diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
index 598cda7daa..721edd9994 100644
--- a/ggml/src/ggml-cuda/fattn.cu
+++ b/ggml/src/ggml-cuda/fattn.cu
@@ -18,9 +18,11 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_con
         }
     }
 
-    if ((turing_mma_available(cc) || amd_wmma_available(cc)) && Q->ne[1] <= 16/ncols2) {
-        ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 16/ncols2, ncols2>(ctx, dst);
-        return;
+    if constexpr (ncols2 <= 16) {
+        if ((turing_mma_available(cc) || amd_wmma_available(cc)) && Q->ne[1] <= 16/ncols2) {
+            ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 16/ncols2, ncols2>(ctx, dst);
+            return;
+        }
     }
 
     if (ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_TURING || amd_wmma_available(cc) || Q->ne[1] <= 32/ncols2) {
@@ -33,6 +35,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_con
 
 template <int DKQ, int DV>
 static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
     const ggml_tensor * KQV  = dst;
     const ggml_tensor * Q    = dst->src[0];
     const ggml_tensor * K    = dst->src[1];
@@ -46,7 +49,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_con
     //     are put into the template specialization without GQA optimizations.
     bool use_gqa_opt = mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
     for (const ggml_tensor * t : {Q, K, V, mask}) {
-        if (t == nullptr) {
+        if (t == nullptr || ggml_is_quantized(t->type)) {
             continue;
         }
         for (size_t i = 1; i < GGML_MAX_DIMS; ++i) {
@@ -60,17 +63,38 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_con
     GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
     const int gqa_ratio = Q->ne[2] / K->ne[2];
 
-    if (use_gqa_opt && gqa_ratio % 8 == 0) {
+    // On Volta the GQA optimizations aren't as impactful vs. minimizing wasted compute:
+    if (cc == GGML_CUDA_CC_VOLTA) {
+        if (use_gqa_opt && gqa_ratio % 8 == 0) {
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 8>(ctx, dst);
+            return;
+        }
+
+        if (use_gqa_opt && gqa_ratio % 4 == 0) {
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 4>(ctx, dst);
+            return;
+        }
+
+        if (use_gqa_opt && gqa_ratio % 2 == 0) {
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 2>(ctx, dst);
+            return;
+        }
+
+        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 1>(ctx, dst);
+        return;
+    }
+
+    if (use_gqa_opt && gqa_ratio > 4) {
         ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 8>(ctx, dst);
         return;
     }
 
-    if (use_gqa_opt && gqa_ratio % 4 == 0) {
+    if (use_gqa_opt && gqa_ratio > 2) {
         ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 4>(ctx, dst);
         return;
     }
 
-    if (use_gqa_opt && gqa_ratio % 2 == 0) {
+    if (use_gqa_opt && gqa_ratio > 1) {
         ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<DKQ, DV, 2>(ctx, dst);
         return;
     }
@@ -79,6 +103,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_con
 }
 
 static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
     const ggml_tensor * KQV  = dst;
     const ggml_tensor * Q    = dst->src[0];
     const ggml_tensor * K    = dst->src[1];
@@ -121,8 +146,50 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
 
             GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
             const int gqa_ratio = Q->ne[2] / K->ne[2];
-            GGML_ASSERT(gqa_ratio % 16 == 0);
-            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+            if (gqa_ratio == 20) { // GLM 4.7 Flash
+                if (cc >= GGML_CUDA_CC_DGX_SPARK) {
+                    if (Q->ne[1] <= 8) {
+                        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+                        break;
+                    }
+                    ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 4>(ctx, dst);
+                    break;
+                }
+                if (cc >= GGML_CUDA_CC_BLACKWELL) {
+                    if (Q->ne[1] <= 4 && K->ne[1] >= 65536) {
+                        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+                        break;
+                    }
+                    ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 4>(ctx, dst);
+                    break;
+                }
+                if (cc >= GGML_CUDA_CC_ADA_LOVELACE) {
+                    if (Q->ne[1] <= 4) {
+                        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+                        break;
+                    }
+                    ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 4>(ctx, dst);
+                    break;
+                }
+                if (cc >= GGML_CUDA_CC_TURING) {
+                    if (Q->ne[1] <= 4) {
+                        if (K->ne[1] <= 16384) {
+                            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+                            break;
+                        }
+                        ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 32>(ctx, dst);
+                        break;
+                    }
+                    ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 4>(ctx, dst);
+                    break;
+                }
+                // Volta:
+                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 4>(ctx, dst);
+            } else if (gqa_ratio % 16 == 0) {
+                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+            } else {
+                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512,  4>(ctx, dst);
+            }
         } break;
         default:
             GGML_ABORT("fatal error");
@@ -230,9 +297,9 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
 
     // The effective batch size for the kernel can be increased by gqa_ratio.
     // The kernel versions without this optimization are also used for ALiBi, if there is no mask, or if the KV cache is not padded,
-    bool gqa_opt_applies = gqa_ratio % 2 == 0 && mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
+    bool gqa_opt_applies = gqa_ratio >= 2 && mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
     for (const ggml_tensor * t : {Q, K, V, mask}) {
-        if (t == nullptr) {
+        if (t == nullptr || ggml_is_quantized(t->type)) {
             continue;
         }
         for (size_t i = 1; i < GGML_MAX_DIMS; ++i) {
@@ -262,7 +329,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
             if (V->ne[0] != 512) {
                 return BEST_FATTN_KERNEL_NONE;
             }
-            if (!gqa_opt_applies || gqa_ratio % 16 != 0) {
+            if (!gqa_opt_applies) {
                 return BEST_FATTN_KERNEL_NONE;
             }
             break;
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index ed1021469a..08383edb40 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -70,17 +70,18 @@
 #include <condition_variable>
 #include <cstddef>
 #include <cstdint>
-#include <float.h>
+#include <cfloat>
 #include <initializer_list>
 #include <limits>
 #include <map>
 #include <memory>
 #include <mutex>
-#include <stdarg.h>
-#include <stdio.h>
-#include <stdlib.h>
+#include <cstdarg>
+#include <cstdio>
+#include <cstdlib>
 #include <string>
 #include <vector>
+#include <unordered_set>
 
 static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
 
@@ -2916,21 +2917,27 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
 }
 
 static void ggml_cuda_graph_node_set_properties(ggml_cuda_graph_node_properties * props, ggml_tensor * node) {
-    props->node_address = node->data;
+    memset(props, 0, sizeof(ggml_cuda_graph_node_properties));
+    props->node_data = node->data;
     props->node_op = node->op;
+    props->node_type = node->type;
+    props->flags = node->flags;
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
         props->ne[i] = node->ne[i];
         props->nb[i] = node->nb[i];
     }
     for (int i = 0; i < GGML_MAX_SRC; i++) {
-        props->src_address[i] = node->src[i] ? node->src[i]->data : nullptr;
+        if (!node->src[i]) {
+            continue;
+        }
+
+        props->src_data[i] = node->src[i]->data;
     }
     memcpy(props->op_params, node->op_params, GGML_MAX_OP_PARAMS);
 }
 
 static bool ggml_cuda_graph_node_properties_match(ggml_tensor * node, ggml_cuda_graph_node_properties * props) {
-    if (node->data != props->node_address &&
-          node->op != GGML_OP_VIEW) {
+    if (node->data != props->node_data && node->op != GGML_OP_VIEW) {
         return false;
     }
 
@@ -2938,6 +2945,10 @@ static bool ggml_cuda_graph_node_properties_match(ggml_tensor * node, ggml_cuda_
         return false;
     }
 
+    if (node->type != props->node_type) {
+        return false;
+    }
+
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
         if (node->ne[i] != props->ne[i]) {
             return false;
@@ -2947,12 +2958,18 @@ static bool ggml_cuda_graph_node_properties_match(ggml_tensor * node, ggml_cuda_
         }
     }
 
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        if (node->src[i] &&
-            node->src[i]->data != props->src_address[i] &&
-            node->op != GGML_OP_VIEW
-        ) {
-            return false;
+    if (node->op != GGML_OP_VIEW) {
+        for (int i = 0; i < GGML_MAX_SRC; i++) {
+            if (!node->src[i]) {
+                if (props->src_data[i] != nullptr) {
+                    return false;
+                }
+                continue;
+            }
+
+            if (node->src[i]->data != props->src_data[i]) {
+                return false;
+            }
         }
     }
 
@@ -2961,59 +2978,89 @@ static bool ggml_cuda_graph_node_properties_match(ggml_tensor * node, ggml_cuda_
         return false;
     }
 
+    if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) != (props->flags & GGML_TENSOR_FLAG_COMPUTE)) {
+        return false;
+    }
+
     return true;
 }
 
-static bool ggml_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph) {
+static const void * ggml_cuda_graph_get_key(ggml_cgraph * cgraph) {
+    return cgraph->nodes[0];
+}
 
+static bool ggml_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph) {
     bool res = false;
 
-    if (cuda_ctx->cuda_graph->instance == nullptr) {
+    const void * graph_key = ggml_cuda_graph_get_key(cgraph);
+    ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
+
+    if (graph->instance == nullptr) {
         res = true;
     }
 
     // Check if the graph size has changed
-    if (cuda_ctx->cuda_graph->props.size() != (size_t)cgraph->n_nodes + cgraph->n_leafs) {
+    if (graph->props.size() != (size_t)cgraph->n_nodes) {
         res = true;
-        cuda_ctx->cuda_graph->props.resize(cgraph->n_nodes + cgraph->n_leafs);
+        graph->props.resize(cgraph->n_nodes);
     }
 
     // Loop over nodes in GGML graph to determine if CUDA graph update is required
     // and store properties to allow this comparison for the next token
+    std::unordered_set<ggml_tensor *> seen_node;
+    std::vector<ggml_tensor *> srcs_extra;
     for (int i = 0; i < cgraph->n_nodes; i++) {
         bool props_match = true;
+
+        seen_node.insert(cgraph->nodes[i]);
+
         if (!res) {
-            props_match = ggml_cuda_graph_node_properties_match(cgraph->nodes[i], &cuda_ctx->cuda_graph->props[i]);
+            props_match = ggml_cuda_graph_node_properties_match(cgraph->nodes[i], &graph->props[i]);
         }
         if (!props_match) {
             res = true;
         }
-        ggml_cuda_graph_node_set_properties(&cuda_ctx->cuda_graph->props[i], cgraph->nodes[i]);
+        ggml_cuda_graph_node_set_properties(&graph->props[i], cgraph->nodes[i]);
+
+        for (int src_idx = 0; src_idx < GGML_MAX_SRC; ++src_idx) {
+            ggml_tensor * src = cgraph->nodes[i]->src[src_idx];
+            if (src && seen_node.find(src) == seen_node.end()) {
+                srcs_extra.push_back(src);
+            }
+        }
     }
 
-    for (int i = 0; i < cgraph->n_leafs; i++) {
-        bool props_match= true;
+    if (graph->extra.size() != (size_t) srcs_extra.size()) {
+        res = true;
+        graph->extra.resize(srcs_extra.size());
+    }
+
+    for (size_t i = 0; i < srcs_extra.size(); ++i) {
+        bool props_match = true;
+
         if (!res) {
-            props_match = ggml_cuda_graph_node_properties_match(cgraph->leafs[i], &cuda_ctx->cuda_graph->props[cgraph->n_nodes + i]);
+            props_match = ggml_cuda_graph_node_properties_match(srcs_extra[i], &graph->extra[i]);
         }
+
         if (!props_match) {
             res = true;
         }
-        ggml_cuda_graph_node_set_properties(&cuda_ctx->cuda_graph->props[cgraph->n_nodes + i], cgraph->leafs[i]);
+        ggml_cuda_graph_node_set_properties(&graph->extra[i], srcs_extra[i]);
     }
 
     return res;
 }
 
-static void ggml_cuda_graph_update_executable(ggml_backend_cuda_context * cuda_ctx) {
+static void ggml_cuda_graph_update_executable(ggml_backend_cuda_context * cuda_ctx, const void * graph_key) {
+    ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
 
 #if CUDART_VERSION >= 12000
     cudaGraphExecUpdateResultInfo result_info;
-    cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &result_info);
+    cudaError_t stat = cudaGraphExecUpdate(graph->instance, graph->graph, &result_info);
 #else
     cudaGraphNode_t errorNode;
     cudaGraphExecUpdateResult result_info;
-    cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &errorNode, &result_info);
+    cudaError_t stat = cudaGraphExecUpdate(graph->instance, graph->graph, &errorNode, &result_info);
 #endif // CUDART_VERSION >= 12000
 
     if (stat == cudaErrorGraphExecUpdateFailure) {
@@ -3024,14 +3071,14 @@ static void ggml_cuda_graph_update_executable(ggml_backend_cuda_context * cuda_c
         // The pre-existing graph exec cannot be updated due to violated constraints
         // so instead clear error and re-instantiate
         (void)cudaGetLastError();
-        CUDA_CHECK(cudaGraphExecDestroy(cuda_ctx->cuda_graph->instance));
-        cuda_ctx->cuda_graph->instance = nullptr;
-        CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
+        CUDA_CHECK(cudaGraphExecDestroy(graph->instance));
+        graph->instance = nullptr;
+        CUDA_CHECK(cudaGraphInstantiate(&graph->instance, graph->graph, NULL, NULL, 0));
     } else {
         GGML_ASSERT(stat == cudaSuccess);
     }
 }
-#endif
+#endif // USE_CUDA_GRAPH
 
 static bool ggml_cuda_should_fuse_rope_set_rows(const ggml_tensor * rope,
                                                 const ggml_tensor * view,
@@ -3067,63 +3114,166 @@ static bool ggml_cuda_should_fuse_rope_set_rows(const ggml_tensor * rope,
     return true;
 }
 
-static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops, std::initializer_list<enum ggml_unary_op> unary_ops) {
+static bool ggml_cuda_topk_moe_fusion(const struct ggml_cgraph * cgraph, int node_idx, ggml_cuda_topk_moe_args & args) {
+    args.sigmoid         = false;
+    args.softmax         = false;
+    args.delayed_softmax = false;
+    args.prob_bias       = false;
+    args.norm            = false;
+
+    const int      n_nodes = cgraph->n_nodes;
+    ggml_tensor ** nodes   = cgraph->nodes;
+
+    if (nodes[node_idx]->op == GGML_OP_SOFT_MAX) {
+        args.softmax = true;
+    }
+
+    if (nodes[node_idx]->op == GGML_OP_UNARY) {
+        if (ggml_get_unary_op(nodes[node_idx]) != GGML_UNARY_OP_SIGMOID) {
+            return false;
+        }
+        args.sigmoid = true;
+    }
+
+    if (nodes[node_idx]->op == GGML_OP_ARGSORT) {
+        args.delayed_softmax = true;
+    }
+
+    node_idx++;
+
+    if (args.sigmoid || args.softmax) {
+        // SOFTMAX -> RESHAPE
+        if (node_idx >= n_nodes || nodes[node_idx]->op != GGML_OP_RESHAPE ||
+                nodes[node_idx]->src[0] != nodes[node_idx - 1]) {
+            return false;
+        }
+        ggml_tensor * probs_reshaped = nodes[node_idx];
+        node_idx++;
+
+        if (node_idx >= n_nodes) {
+            return false;
+        }
+
+        // src of bias add is the unreshaped probs (-2 instead of -1)
+        if (nodes[node_idx]->op == GGML_OP_ADD && nodes[node_idx]->src[0] == nodes[node_idx - 2]) {
+            args.prob_bias = true;
+            node_idx++;
+        }
+        // RESHAPE/ADD -> ARGSORT
+        if (node_idx >= n_nodes || nodes[node_idx]->op != GGML_OP_ARGSORT) {
+            return false;
+        }
+
+        if (args.prob_bias && nodes[node_idx]->src[0] != nodes[node_idx - 1]) {
+            return false;
+        } else if (!args.prob_bias && nodes[node_idx]->src[0] != nodes[node_idx - 2]) {
+            return false;
+        }
+
+        node_idx++;
+
+        // ARGSORT-> VIEW
+        if (node_idx >= n_nodes || nodes[node_idx]->op != GGML_OP_VIEW ||
+                nodes[node_idx]->src[0] != nodes[node_idx - 1]) {
+            return false;
+        }
+        node_idx++;
+
+        if (node_idx >= n_nodes || nodes[node_idx]->op != GGML_OP_GET_ROWS) {
+            return false;
+        }
+
+        // GET_ROWS
+        if (nodes[node_idx]->src[0] != probs_reshaped || nodes[node_idx]->src[1] != nodes[node_idx - 1]) {
+            return false;
+        }
+        node_idx++;
+    } else if (args.delayed_softmax) {
+        if (node_idx - 2 < 0) {
+            return false;
+        }
+        ggml_tensor * probs_reshaped = nodes[node_idx - 2];
+
+        // VIEW->ARGSORT
+        if (node_idx >= n_nodes || nodes[node_idx]->op != GGML_OP_VIEW ||
+            nodes[node_idx]->src[0] != nodes[node_idx - 1]) {
+            return false;
+        }
+        node_idx++;
+
+        // GET_ROWS
+        if (node_idx >= n_nodes || nodes[node_idx]->src[1] != nodes[node_idx - 1] ||
+                nodes[node_idx]->src[0] != probs_reshaped) {
+            return false;
+        }
+        node_idx++;
+
+        static const std::vector<ggml_op> remaining_ops = { GGML_OP_RESHAPE, GGML_OP_SOFT_MAX, GGML_OP_RESHAPE };
+
+        for (const ggml_op op : remaining_ops) {
+            if (node_idx >= n_nodes || nodes[node_idx]->op != op || nodes[node_idx]->src[0] != nodes[node_idx - 1]) {
+                return false;
+            }
+            node_idx++;
+        }
+    }
+
+    // At this point we can check for norm + scale. Everything is now at least valid till the norm
+    if (node_idx >= n_nodes) {
+        return true;
+    }
+
+    if (nodes[node_idx]->op == GGML_OP_RESHAPE) {
+        //check RESHAPE->SUM_ROWS->CLAMP->DIV->RESHAPE
+        static const std::vector<ggml_op> norm_ops = { GGML_OP_RESHAPE, GGML_OP_SUM_ROWS, GGML_OP_CLAMP };
+
+        args.norm = true;
+        for (const ggml_op op : norm_ops) {
+            if (nodes[node_idx]->op == op && nodes[node_idx]->src[0] == nodes[node_idx - 1]) {
+                node_idx++;
+            } else {
+                args.norm = false;
+                return true;
+            }
+        }
+
+        // DIV <- CLAMP, RESHAPE
+        if (nodes[node_idx]->op != GGML_OP_DIV || nodes[node_idx]->src[1] != nodes[node_idx - 1] ||
+            nodes[node_idx]->src[0] != nodes[node_idx - 3]) {
+            args.norm = false;
+            return true;
+        }
+        node_idx++;
+
+        if (nodes[node_idx]->op != GGML_OP_RESHAPE || nodes[node_idx]->src[0] != nodes[node_idx - 1]) {
+            args.norm = false;
+            return true;
+        }
+
+        node_idx++;
+    }
+
+    if (nodes[node_idx]->op == GGML_OP_SCALE && nodes[node_idx]->src[0] == nodes[node_idx - 1]) {
+        args.scale = true;
+    }
+
+    return true;
+}
+
+static bool ggml_cuda_can_fuse(const struct ggml_cgraph *                cgraph,
+                               int                                       node_idx,
+                               std::initializer_list<enum ggml_op>       ops,
+                               std::initializer_list<enum ggml_unary_op> unary_ops) {
 #ifndef NDEBUG
     const size_t num_unary = std::count(ops.begin(), ops.end(), GGML_OP_UNARY);
     GGML_ASSERT(unary_ops.size() == num_unary);
 #endif
 
-    //TODO: remove special case once ggml_can_fuse can handle empty nodes
-    std::initializer_list<enum ggml_op> topk_moe_ops =
-        ggml_cuda_topk_moe_ops(/*with_norm*/ false, /*delayed_softmax=*/false);
-    std::initializer_list<enum ggml_op> topk_moe_ops_with_norm =
-        ggml_cuda_topk_moe_ops(/*with_norm=*/true, /*delayed_softmax=*/false);
-    std::initializer_list<enum ggml_op> topk_moe_ops_delayed_softmax =
-        ggml_cuda_topk_moe_ops(/*with_norm=*/false, /*delayed_softmax=*/true);
-
     const auto is_equal = [](const std::initializer_list<enum ggml_op> & list1,
                              const std::initializer_list<enum ggml_op> & list2) {
         return std::equal(list1.begin(), list1.end(), list2.begin(), list2.end());
     };
 
-    if (is_equal(topk_moe_ops_with_norm, ops) &&
-        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 9 })) {
-        ggml_tensor * softmax = cgraph->nodes[node_idx];
-        ggml_tensor * weights = cgraph->nodes[node_idx + 9];
-        ggml_tensor * get_rows = cgraph->nodes[node_idx + 4];
-        ggml_tensor * argsort = cgraph->nodes[node_idx + 2];
-        int n_expert = cgraph->nodes[node_idx]->src[0]->ne[0];
-
-        if (ggml_cuda_should_use_topk_moe(softmax, weights, get_rows, argsort, nullptr, n_expert)) {
-            return true;
-        }
-    }
-
-    if (is_equal(topk_moe_ops, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 4 })) {
-        ggml_tensor * softmax = cgraph->nodes[node_idx];
-        ggml_tensor * weights = cgraph->nodes[node_idx + 4];
-        ggml_tensor * get_rows = cgraph->nodes[node_idx + 4];
-        ggml_tensor * argsort = cgraph->nodes[node_idx + 2];
-        int n_expert = cgraph->nodes[node_idx]->src[0]->ne[0];
-
-        if (ggml_cuda_should_use_topk_moe(softmax, weights, get_rows, argsort, nullptr, n_expert)) {
-            return true;
-        }
-    }
-
-    if (is_equal(topk_moe_ops_delayed_softmax, ops) &&
-        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 1, node_idx + 5 })) {
-        ggml_tensor * softmax = cgraph->nodes[node_idx + 4];
-        ggml_tensor * weights = cgraph->nodes[node_idx + 5];
-        ggml_tensor * get_rows = cgraph->nodes[node_idx + 2];
-        ggml_tensor * argsort = cgraph->nodes[node_idx + 0];
-        int n_expert = cgraph->nodes[node_idx]->src[0]->ne[0];
-
-        if (ggml_cuda_should_use_topk_moe(softmax, weights, get_rows, argsort, nullptr, n_expert)) {
-            return true;
-        }
-    }
-
     std::initializer_list<enum ggml_op> mul_mat_bias_glu_ops    = { GGML_OP_MUL_MAT,    GGML_OP_ADD,    GGML_OP_MUL_MAT,    GGML_OP_ADD,    GGML_OP_GLU };
     std::initializer_list<enum ggml_op> mul_mat_id_bias_glu_ops = { GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID, GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID, GGML_OP_GLU };
 
@@ -3236,7 +3386,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
     return false;
 }
 
-static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph, const bool use_cuda_graph, const bool cuda_graph_update_required) {
+static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph, const bool use_cuda_graph, const bool cuda_graph_update_required, const void * graph_key) {
     bool graph_evaluated_or_captured = false;
 
     // flag used to determine whether it is an integrated_gpu
@@ -3378,39 +3528,82 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
                     continue;
                 }
 
+                if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+                    continue;
+                }
 
                 // start of fusion operations
                 static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
                 if (!disable_fusion) {
+                    ggml_cuda_topk_moe_args args;
 
-                    if (ggml_cuda_can_fuse(cgraph, i, ggml_cuda_topk_moe_ops(/*with norm*/ true), {})) {
-                        ggml_tensor * weights          = cgraph->nodes[i + 9];
-                        ggml_tensor * selected_experts = cgraph->nodes[i + 3];
-                        ggml_tensor * clamp            = cgraph->nodes[i + 7];
-                        ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, selected_experts, /*with norm*/ true,
-                                              /*delayed softmax*/ false, clamp);
-                        i += 9;
-                        continue;
-                    }
+                    if (cgraph->nodes[i]->op == GGML_OP_UNARY || cgraph->nodes[i]->op == GGML_OP_SOFT_MAX ||
+                        cgraph->nodes[i]->op == GGML_OP_ARGSORT) {
+                        const bool can_fuse = ggml_cuda_topk_moe_fusion(cgraph, i, args);
 
-                    if (ggml_cuda_can_fuse(cgraph, i, ggml_cuda_topk_moe_ops(/*with norm*/ false), {})) {
-                        ggml_tensor * weights          = cgraph->nodes[i + 4];
-                        ggml_tensor * selected_experts = cgraph->nodes[i + 3];
-                        ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, selected_experts, /*with norm*/ false,
-                                              /*delayed softmax*/ false);
-                        i += 4;
-                        continue;
-                    }
+                        std::vector<ggml_op> ops;
 
-                    if (ggml_cuda_can_fuse(cgraph, i,
-                                           ggml_cuda_topk_moe_ops(/*with norm*/ false, /*delayed softmax*/ true), {})) {
-                        ggml_tensor * weights = cgraph->nodes[i + 5];
-                        ggml_tensor * ids     = cgraph->nodes[i + 1];
+                        if (can_fuse) {
+                            const ggml_tensor * logits  = node->src[0];
+                            ggml_tensor *       weights = nullptr;
+                            ggml_tensor *       ids     = nullptr;
+                            const ggml_tensor * bias    = nullptr;
+                            const ggml_tensor * clamp   = nullptr;
+                            const ggml_tensor * scale   = nullptr;
 
-                        ggml_cuda_op_topk_moe(*cuda_ctx, node->src[0], weights, ids, /*with norm*/ false,
-                                              /*delayed_softmax*/ true);
-                        i += 5;
-                        continue;
+                            if (!args.delayed_softmax) {
+                                ggml_op gating_op = args.sigmoid ? GGML_OP_UNARY : GGML_OP_SOFT_MAX;
+                                int     out_nodes[2];  // nodes which can't be elided
+
+                                if (args.prob_bias) {
+                                    bias = cgraph->nodes[i + 2]->src[1];
+                                    ops.insert(ops.end(), { gating_op, GGML_OP_RESHAPE, GGML_OP_ADD, GGML_OP_ARGSORT,
+                                                            GGML_OP_VIEW, GGML_OP_GET_ROWS });
+                                    out_nodes[0] = i + 4;
+                                    ids          = cgraph->nodes[i + 4];
+                                } else {
+                                    ops.insert(ops.end(), { gating_op, GGML_OP_RESHAPE, GGML_OP_ARGSORT, GGML_OP_VIEW,
+                                                            GGML_OP_GET_ROWS });
+                                    out_nodes[0] = i + 3;
+                                    ids          = cgraph->nodes[i + 3];
+                                }
+
+                                if (args.norm) {
+                                    ops.insert(ops.end(), { GGML_OP_RESHAPE, GGML_OP_SUM_ROWS, GGML_OP_CLAMP,
+                                                            GGML_OP_DIV, GGML_OP_RESHAPE });
+                                    clamp = cgraph->nodes[i + ops.size() - 3];
+                                }
+                                if (args.scale) {
+                                    ops.insert(ops.end(), { GGML_OP_SCALE });
+                                    scale = cgraph->nodes[i + ops.size() - 1];
+                                }
+
+                                weights      = cgraph->nodes[i + ops.size() - 1];
+                                out_nodes[1] = i + ops.size() - 1;
+
+                                if (ggml_can_fuse_subgraph(cgraph, i, ops.size(), ops.data(), out_nodes, 2) &&
+                                        ggml_cuda_should_use_topk_moe(node, logits, weights, ids)) {
+                                    ggml_cuda_op_topk_moe(*cuda_ctx, logits, weights, ids, clamp, scale, bias, args);
+                                    i += ops.size() - 1;
+                                    continue;
+                                }
+                            } else if (!args.norm && !args.prob_bias) {
+                                //special case gpt-oss, no norm, no bias.
+                                ops.insert(ops.end(), { GGML_OP_ARGSORT, GGML_OP_VIEW, GGML_OP_GET_ROWS,
+                                                        GGML_OP_RESHAPE, GGML_OP_SOFT_MAX, GGML_OP_RESHAPE });
+                                weights                     = cgraph->nodes[i + 5];
+                                ids                         = cgraph->nodes[i + 1];
+                                const ggml_tensor * softmax = cgraph->nodes[i + 4];
+
+                                int out_nodes[2] = { i + 1, i + 5 };
+                                if (ggml_can_fuse_subgraph(cgraph, i, ops.size(), ops.data(), out_nodes, 2) &&
+                                        ggml_cuda_should_use_topk_moe(softmax, logits, weights, ids)) {
+                                    ggml_cuda_op_topk_moe(*cuda_ctx, logits, weights, ids, clamp, scale, bias, args);
+                                    i += ops.size() - 1;
+                                    continue;
+                                }
+                            }
+                        }
                     }
 
                     if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, {})) {
@@ -3687,13 +3880,14 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
         }
 
 #ifdef USE_CUDA_GRAPH
+        ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
         if (use_cuda_graph && cuda_graph_update_required) { // End CUDA graph capture
-            if (cuda_ctx->cuda_graph->graph != nullptr) {
-                CUDA_CHECK(cudaGraphDestroy(cuda_ctx->cuda_graph->graph));
-                cuda_ctx->cuda_graph->graph = nullptr;
+            if (graph->graph != nullptr) {
+                CUDA_CHECK(cudaGraphDestroy(graph->graph));
+                graph->graph = nullptr;
             }
 
-            CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &cuda_ctx->cuda_graph->graph));
+            CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &graph->graph));
             graph_evaluated_or_captured = true; // CUDA graph has been captured
 
             std::lock_guard<std::mutex> lock(ggml_cuda_lock);
@@ -3706,43 +3900,38 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
     }
 
     if (use_cuda_graph) {
-        if (cuda_ctx->cuda_graph->instance == nullptr) { // Create executable graph from captured graph.
-            CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
+        ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
+        if (graph->instance == nullptr) { // Create executable graph from captured graph.
+            CUDA_CHECK(cudaGraphInstantiate(&graph->instance, graph->graph, NULL, NULL, 0));
         }
         if (cuda_graph_update_required) { // Update graph executable
-            ggml_cuda_graph_update_executable(cuda_ctx);
+            ggml_cuda_graph_update_executable(cuda_ctx, graph_key);
         }
         // Launch graph
-        CUDA_CHECK(cudaGraphLaunch(cuda_ctx->cuda_graph->instance, cuda_ctx->stream()));
+        CUDA_CHECK(cudaGraphLaunch(graph->instance, cuda_ctx->stream()));
 #else
+        GGML_UNUSED(graph_key);
         graph_evaluated_or_captured = true;
 #endif  // USE_CUDA_GRAPH
     }
 }
 
-static bool ggml_cuda_graph_set_enabled(ggml_backend_cuda_context * cuda_ctx) {
-
 #ifdef USE_CUDA_GRAPH
+static bool ggml_cuda_graph_set_enabled(ggml_backend_cuda_context * cuda_ctx, const void * graph_key) {
+    ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
 
-    if (cuda_ctx->cuda_graph == nullptr) {
-        cuda_ctx->cuda_graph.reset(new ggml_cuda_graph());
-    }
-
-    if (cuda_ctx->cuda_graph->graph == nullptr) {
+    if (graph->graph == nullptr) {
         if (ggml_cuda_info().devices[cuda_ctx->device].cc < GGML_CUDA_CC_AMPERE) {
-            if (!cuda_ctx->cuda_graph->disable_due_to_gpu_arch) {
+            if (!graph->disable_due_to_gpu_arch) {
                 GGML_LOG_DEBUG("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
             }
-            cuda_ctx->cuda_graph->disable_due_to_gpu_arch = true;
+            graph->disable_due_to_gpu_arch = true;
         }
     }
 
-    return cuda_ctx->cuda_graph->is_enabled();
-#else
-    GGML_UNUSED(cuda_ctx);
-    return false;
-#endif // USE_CUDA_GRAPH
+    return graph->is_enabled();
 }
+#endif // USE_CUDA_GRAPH
 
 static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
@@ -3751,15 +3940,19 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
 
     bool use_cuda_graph             = false;
     bool cuda_graph_update_required = false;
+    const void * graph_key = nullptr;
 
 #ifdef USE_CUDA_GRAPH
-    use_cuda_graph = ggml_cuda_graph_set_enabled(cuda_ctx);
+    graph_key = ggml_cuda_graph_get_key(cgraph);
 
-    if (cuda_ctx->cuda_graph->is_enabled()) {
+    use_cuda_graph = ggml_cuda_graph_set_enabled(cuda_ctx, graph_key);
+
+    ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
+    if (graph->is_enabled()) {
         cuda_graph_update_required = ggml_cuda_graph_update_required(cuda_ctx, cgraph);
         use_cuda_graph             = ggml_cuda_graph_check_compability(cgraph);
 
-        cuda_ctx->cuda_graph->record_update(use_cuda_graph, cuda_graph_update_required);
+        graph->record_update(use_cuda_graph, cuda_graph_update_required);
     }
 #endif // USE_CUDA_GRAPH
 
@@ -3773,7 +3966,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
         CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
     }
 
-    ggml_cuda_graph_evaluate_and_capture(cuda_ctx, cgraph, use_cuda_graph, cuda_graph_update_required);
+    ggml_cuda_graph_evaluate_and_capture(cuda_ctx, cgraph, use_cuda_graph, cuda_graph_update_required, graph_key);
 
     return GGML_STATUS_SUCCESS;
 }
@@ -3806,7 +3999,14 @@ static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_ev
 static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
 
-    const bool use_cuda_graph = ggml_cuda_graph_set_enabled(cuda_ctx);
+#ifdef USE_CUDA_GRAPH
+    const void * graph_key = ggml_cuda_graph_get_key(cgraph);
+    const bool use_cuda_graph = ggml_cuda_graph_set_enabled(cuda_ctx, graph_key);
+#else
+    const bool use_cuda_graph = false;
+    GGML_UNUSED(cuda_ctx);
+    GGML_UNUSED(cgraph);
+#endif
 
     static bool enable_graph_optimization = [] {
         const char * env     = getenv("GGML_CUDA_GRAPH_OPT");
@@ -4849,6 +5049,16 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
         static std::mutex mutex;
         std::lock_guard<std::mutex> lock(mutex);
         if (!initialized) {
+            // Set CUDA_SCALE_LAUNCH_QUEUES before any CUDA API call to improve multi-GPU pipeline parallelism performance
+            // PR: https://github.com/ggml-org/llama.cpp/pull/19042
+            if (getenv("CUDA_SCALE_LAUNCH_QUEUES") == nullptr) {
+#ifdef _WIN32
+                _putenv_s("CUDA_SCALE_LAUNCH_QUEUES", "4x");
+#else
+                setenv("CUDA_SCALE_LAUNCH_QUEUES", "4x", 0); // don't overwrite if already set
+#endif // _WIN32
+            }
+
             ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context;
             const int min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32;
 
diff --git a/ggml/src/ggml-cuda/mean.cu b/ggml/src/ggml-cuda/mean.cu
index 60542fc19d..49af538995 100644
--- a/ggml/src/ggml-cuda/mean.cu
+++ b/ggml/src/ggml-cuda/mean.cu
@@ -31,14 +31,15 @@ void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 #endif // USE_CUDA_GRAPH
     if ((nrows == 1) &&
 #ifdef USE_CUDA_GRAPH
-            // CUDA_GRAPHS_DISABLED
-            ((ncols > 65536) &&
-             ((ctx.cuda_graph->instance == nullptr) && (iscapturing == cudaStreamCaptureStatusNone) ||
-              ctx.cuda_graph->is_enabled())) ||
-        // CUDA_GRAPHS ENABLED
-        ((ncols > 32768) &&
-         !((ctx.cuda_graph->instance == nullptr) && (iscapturing == cudaStreamCaptureStatusNone) ||
-            ctx.cuda_graph->is_enabled()))) {
+            // Determine if CUDA graphs are effectively disabled for this context
+            // (no graph instance exists and we're not capturing, OR graphs are explicitly enabled)
+            (((ncols > 65536) &&
+              (((!ctx.any_cuda_graph_has_instance()) && (iscapturing == cudaStreamCaptureStatusNone)) ||
+               ctx.any_cuda_graph_enabled())) ||
+            // CUDA graphs are enabled - use lower threshold
+             ((ncols > 32768) &&
+              !(((!ctx.any_cuda_graph_has_instance()) && (iscapturing == cudaStreamCaptureStatusNone)) ||
+                ctx.any_cuda_graph_enabled())))) {
 #else
         (ncols > 65536)) {
 #endif // USE_CUDA_GRAPH
diff --git a/ggml/src/ggml-cuda/mma.cuh b/ggml/src/ggml-cuda/mma.cuh
index 42085d1002..dd45d6c78f 100644
--- a/ggml/src/ggml-cuda/mma.cuh
+++ b/ggml/src/ggml-cuda/mma.cuh
@@ -333,7 +333,33 @@ namespace ggml_cuda_mma {
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 16 && J == 8) {
-                return 4 * (threadIdx.x / 16) + l;
+                return ne * (threadIdx.x / 16) + l;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+#elif defined(AMD_MFMA_AVAILABLE)
+        static constexpr int ne = I * J / 64;
+        half2 x[ne] = {{0.0f, 0.0f}};
+
+        static constexpr __device__ bool supported() {
+            if (I == 16 && J == 8) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 16 && J == 8) {
+                return threadIdx.x % 16;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 16 && J == 8) {
+                return ne * (threadIdx.x / 16) + l;
             } else {
                 NO_DEVICE_CODE;
                 return -1;
@@ -391,7 +417,22 @@ namespace ggml_cuda_mma {
         static constexpr data_layout dl = DATA_LAYOUT_I_MAJOR;
 
 #if defined(AMD_WMMA_AVAILABLE)
-        static constexpr int ne = I * J / 32;
+        static constexpr int ne = tile<I_, J_, half2, DATA_LAYOUT_I_MAJOR>::ne;
+        nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
+
+        static constexpr __device__ bool supported() {
+            return tile<I_, J_, half2, DATA_LAYOUT_I_MAJOR>::supported();
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            return tile<I_, J_, half2, DATA_LAYOUT_I_MAJOR>::get_i(l);
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            return tile<I_, J_, half2, DATA_LAYOUT_I_MAJOR>::get_j(l);
+        }
+#elif defined(AMD_MFMA_AVAILABLE)
+        static constexpr int ne = tile<I_, J_, half2, DATA_LAYOUT_I_MAJOR>::ne;
         nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
 
         static constexpr __device__ bool supported() {
@@ -945,6 +986,32 @@ namespace ggml_cuda_mma {
 #endif // AMPERE_MMA_AVAILABLE
     }
 
+    template <data_layout dl_ab, data_layout dl_d>
+    static __device__ __forceinline__ void mma(
+            tile<16, 16, float, dl_d> & D, const tile<16, 8, float, dl_ab> & A, const tile<16, 8, float, dl_ab> & B) {
+#ifdef AMD_MFMA_AVAILABLE
+        using floatx4_t = __attribute__((ext_vector_type(4))) float;
+        floatx4_t& acc_frag = reinterpret_cast<floatx4_t&>(D.x[0]);
+#if defined(CDNA3)
+        using floatx2_t = __attribute__((ext_vector_type(2))) float;
+        const floatx2_t& a_frag = reinterpret_cast<const floatx2_t&>(A.x[0]);
+        const floatx2_t& b_frag = reinterpret_cast<const floatx2_t&>(B.x[0]);
+        acc_frag = __builtin_amdgcn_mfma_f32_16x16x8_xf32(a_frag, b_frag, acc_frag, 0, 0, 0);
+#elif defined(CDNA2) || defined(CDNA1)
+#pragma unroll
+        for (int i = 0; i < 2; ++i) {
+            acc_frag = __builtin_amdgcn_mfma_f32_16x16x4f32(A.x[i], B.x[i], acc_frag, 0, 0, 0);
+        }
+#else
+        GGML_UNUSED_VARS(D, A, B);
+        NO_DEVICE_CODE;
+#endif // defined(CDNA3)
+#else
+        GGML_UNUSED_VARS(D, A, B);
+        NO_DEVICE_CODE;
+#endif // AMD_MFMA_AVAILABLE
+    }
+
     static __device__ __forceinline__ void mma_block_scaled(tile<16, 8, float> &     D,
                                                             const tile<16, 8, int> & A,
                                                             const tile<8, 8, int> &  B,
@@ -1054,6 +1121,13 @@ namespace ggml_cuda_mma {
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
 #endif // RDNA4
+#elif defined(AMD_MFMA_AVAILABLE)
+        using halfx4_t = __attribute__((ext_vector_type(4))) _Float16;
+        using floatx4_t = __attribute__((ext_vector_type(4))) float;
+        floatx4_t& acc_frag = reinterpret_cast<floatx4_t&>(D.x[0]);
+        const halfx4_t& a_frag = reinterpret_cast<const halfx4_t&>(A.x[0]);
+        const halfx4_t& b_frag = reinterpret_cast<const halfx4_t&>(B.x[0]);
+        acc_frag = __builtin_amdgcn_mfma_f32_16x16x16f16(a_frag, b_frag, acc_frag, 0, 0, 0);
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
@@ -1081,11 +1155,31 @@ namespace ggml_cuda_mma {
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
-#endif // RDNA4
+#endif // defined(RDNA4)
+#elif defined(AMD_MFMA_AVAILABLE)
+        using floatx4_t = __attribute__((ext_vector_type(4))) float;
+        floatx4_t& acc_frag = reinterpret_cast<floatx4_t&>(D.x[0]);
+#if defined(CDNA3) || defined(CDNA2)
+        using bf16x4_t = __attribute__((ext_vector_type(4))) __bf16;
+        const bf16x4_t& a_frag = reinterpret_cast<const bf16x4_t&>(A.x[0]);
+        const bf16x4_t& b_frag = reinterpret_cast<const bf16x4_t&>(B.x[0]);
+        acc_frag = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(a_frag, b_frag, acc_frag, 0, 0, 0);
+#elif defined(CDNA1)
+#pragma unroll
+        for (int i = 0; i < 2; ++i) {
+            using bf16x2_t = __attribute__((ext_vector_type(2))) __bf16;
+            const bf16x2_t& a_frag = reinterpret_cast<const bf16x2_t&>(A.x[i]);
+            const bf16x2_t& b_frag = reinterpret_cast<const bf16x2_t&>(B.x[i]);
+            acc_frag = __builtin_amdgcn_mfma_f32_16x16x8bf16(a_frag, b_frag, acc_frag, 0, 0, 0);
+        }
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
-#endif // AMPERE_MMA_AVAILABLE
+#endif // defined(CDNA3) || defined(CDNA2)
+#else
+        GGML_UNUSED_VARS(D, A, B);
+        NO_DEVICE_CODE;
+#endif // defined(AMD_WMMA_AVAILABLE)
     }
 
     template <data_layout dl_d, data_layout dl_ab>
diff --git a/ggml/src/ggml-cuda/mmf.cu b/ggml/src/ggml-cuda/mmf.cu
index 6643f243b1..aad4c34aa6 100644
--- a/ggml/src/ggml-cuda/mmf.cu
+++ b/ggml/src/ggml-cuda/mmf.cu
@@ -2,6 +2,13 @@
 #include "mmf.cuh"
 #include "mmid.cuh"
 
+static __forceinline__ int mmf_get_rows_per_block(const int cc) {
+    if (GGML_CUDA_CC_IS_CDNA(cc)) {
+        return MMF_ROWS_PER_BLOCK_CDNA;
+    } else {
+        return MMF_ROWS_PER_BLOCK;
+    }
+}
 
 void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
     GGML_ASSERT(        src1->type == GGML_TYPE_F32);
@@ -89,28 +96,32 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
         ids_info_ptr = &ids_info;
     }
 
+    const int device    = ggml_cuda_get_device();
+    const int cc        = ggml_cuda_info().devices[device].cc;
+    const int rows_per_block = mmf_get_rows_per_block(cc);
+
     switch (src0->type) {
         case GGML_TYPE_F32: {
             const float * src0_d = (const float *) src0->data;
             constexpr int vals_per_T = 1;
-            mul_mat_f_switch_cols_per_block(
-                src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
+            mul_mat_f_switch_rows_per_block<float>(
+                rows_per_block, src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
                 ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
                 ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream(), ids_info_ptr);
         } break;
         case GGML_TYPE_F16: {
             const half2 * src0_d = (const half2 *) src0->data;
             constexpr int vals_per_T = 2;
-            mul_mat_f_switch_cols_per_block(
-                src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
+            mul_mat_f_switch_rows_per_block<half2>(
+                rows_per_block, src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
                 ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
                 ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream(), ids_info_ptr);
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat162 * src0_d = (const nv_bfloat162 *) src0->data;
             constexpr int vals_per_T = 2;
-            mul_mat_f_switch_cols_per_block(
-                src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
+            mul_mat_f_switch_rows_per_block<nv_bfloat162>(
+                rows_per_block, src0_d, src1_d, ids_d, dst_d, ne00/vals_per_T, ne01, ncols_dst, s01/vals_per_T, stride_col_y/vals_per_T, stride_col_dst,
                 ids_s0, ids_s1, ne02, nchannels_y, nchannels_dst, s02/vals_per_T, stride_channel_y, stride_channel_dst,
                 ne03, ne3, s03/vals_per_T, s13, s3, ctx.stream(), ids_info_ptr);
         } break;
@@ -140,7 +151,11 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
             return false;
         }
     }
-    if (src0_ne[1] % MMF_ROWS_PER_BLOCK != 0) {
+    if (src0_ne[1] % mmf_get_rows_per_block(cc) != 0) {
+        return false;
+    }
+
+    if (GGML_CUDA_CC_IS_CDNA3(cc) && type == GGML_TYPE_BF16) {
         return false;
     }
 
@@ -153,6 +168,11 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
     } else {
         if (GGML_CUDA_CC_IS_RDNA3_0(cc) && src1_ncols > 8) {
             return false;
+        } else if (GGML_CUDA_CC_IS_CDNA2(cc) && (type == GGML_TYPE_F16 || type == GGML_TYPE_BF16)) {
+            //TODO: truse CDNA2 as CDNA1, tune the perf when CDNA2 is available.
+            return false;
+        } else if (GGML_CUDA_CC_IS_CDNA1(cc) && (type == GGML_TYPE_F16 || type == GGML_TYPE_BF16)) {
+            return false;
         } else if (src1_ncols > 16) {
             return false;
         }
@@ -160,11 +180,11 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
 
     switch (type) {
         case GGML_TYPE_F32:
-            return ampere_mma_available(cc);
+            return ampere_mma_available(cc) || amd_mfma_available(cc);
         case GGML_TYPE_F16:
-            return volta_mma_available(cc) || turing_mma_available(cc) || amd_wmma_available(cc);
+            return volta_mma_available(cc) || turing_mma_available(cc) || amd_wmma_available(cc) || amd_mfma_available(cc);
         case GGML_TYPE_BF16:
-            return ampere_mma_available(cc) || amd_wmma_available(cc);
+            return ampere_mma_available(cc) || amd_wmma_available(cc) || amd_mfma_available(cc);
         default:
             return false;
     }
diff --git a/ggml/src/ggml-cuda/mmf.cuh b/ggml/src/ggml-cuda/mmf.cuh
index e36730948f..c2a8d54c95 100644
--- a/ggml/src/ggml-cuda/mmf.cuh
+++ b/ggml/src/ggml-cuda/mmf.cuh
@@ -7,6 +7,31 @@
 using namespace ggml_cuda_mma;
 
 #define MMF_ROWS_PER_BLOCK 32
+#define MMF_ROWS_PER_BLOCK_CDNA 64
+
+static __forceinline__ int64_t mmf_get_max_block_size(int cc) {
+    if (GGML_CUDA_CC_IS_CDNA(cc)) {
+        return 512;
+    } else {
+        return 256;
+    }
+}
+
+static __forceinline__ int mmf_get_padding(int cc) {
+    if (GGML_CUDA_CC_IS_CDNA(cc)) {
+        return 2;
+    } else {
+        return 4;
+    }
+}
+
+static constexpr __device__ int mmf_get_padding() {
+#if defined(AMD_MFMA_AVAILABLE)
+    return 2;
+#else
+    return 4;
+#endif // defined(AMD_MFMA_AVAILABLE)
+}
 
 struct mmf_ids_data {
     const int32_t * ids_src_compact = nullptr;
@@ -29,23 +54,25 @@ static __global__ void mul_mat_f(
         const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
         const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
 // TODO: handle this in a consistent and simpler way after AMD MFMA support has been added
-#if (!defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)) || defined(AMD_WMMA_AVAILABLE)
+#if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
 #if defined(AMD_WMMA_AVAILABLE)
-    // Special case for tf32, just dummy mma layout as wmma doesn't support it.
-    constexpr bool is_tf32 = std::is_same_v<T, float>;
-    constexpr int tile_B_I = is_tf32 ? 8 : 16;
-    constexpr int tile_C_J = is_tf32 ? 8 : 16;
-    constexpr data_layout ab_layout = is_tf32 ? DATA_LAYOUT_I_MAJOR : get_input_data_layout();
-    typedef tile<16,       8,        T,     ab_layout>           tile_A;
-    typedef tile<tile_B_I, 8,        T,     ab_layout>           tile_B;
-    typedef tile<16,       tile_C_J, float, DATA_LAYOUT_J_MAJOR> tile_C;
+    if constexpr (!(std::is_same_v<T, half2> || std::is_same_v<T, nv_bfloat162>) || rows_per_block != MMF_ROWS_PER_BLOCK) {NO_DEVICE_CODE;} else {
+    typedef tile<16, 8,  T,     get_input_data_layout()> tile_A;
+    typedef tile<16, 8,  T,     get_input_data_layout()> tile_B;
+    typedef tile<16, 16, float, DATA_LAYOUT_J_MAJOR>     tile_C;
+#elif defined(AMD_MFMA_AVAILABLE)
+    if constexpr (rows_per_block != MMF_ROWS_PER_BLOCK_CDNA) {NO_DEVICE_CODE;} else {
+    typedef tile<16, 8,  T,     DATA_LAYOUT_I_MAJOR> tile_A;
+    typedef tile<16, 8,  T,     DATA_LAYOUT_I_MAJOR> tile_B;
+    typedef tile<16, 16, float, DATA_LAYOUT_J_MAJOR> tile_C;
 #else
 #ifdef VOLTA_MMA_AVAILABLE
-    if constexpr (!std::is_same_v<T, half2>) {NO_DEVICE_CODE;} else {
+    if constexpr (!std::is_same_v<T, half2> || rows_per_block != MMF_ROWS_PER_BLOCK) {NO_DEVICE_CODE;} else {
     typedef tile<32, 4, T,     DATA_LAYOUT_I_MAJOR>          tile_A;
     typedef tile< 8, 4, T,     DATA_LAYOUT_I_MAJOR_MIRRORED> tile_B;
     typedef tile<32, 8, float, DATA_LAYOUT_I_MAJOR>          tile_C;
 #else
+    if constexpr (rows_per_block != MMF_ROWS_PER_BLOCK) {NO_DEVICE_CODE;} else {
     typedef tile<16, 8, T>     tile_A;
     typedef tile<8,  8, T>     tile_B;
     typedef tile<16, 8, float> tile_C;
@@ -57,7 +84,7 @@ static __global__ void mul_mat_f(
     }
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
-    constexpr int tile_k_padded = warp_size + 4;
+    constexpr int tile_k_padded = warp_size + mmf_get_padding();
     constexpr int ntA = rows_per_block / tile_A::I;
     constexpr int ntB = (cols_per_block + tile_B::I - 1) / tile_B::I;
 
@@ -198,7 +225,7 @@ static __global__ void mul_mat_f(
     }
 
     float * buf_iw = (float *) compute_base;
-    constexpr int kiw = nwarps*rows_per_block + 4;
+    constexpr int kiw = nwarps*rows_per_block + mmf_get_padding();
 
     if (nwarps > 1) {
         __syncthreads();
@@ -228,27 +255,34 @@ static __global__ void mul_mat_f(
             return;
         }
 
-        float sum = 0.0f;
-        static_assert(rows_per_block == warp_size, "need loop/check");
+        float sum[rows_per_block/warp_size] = {0.0f};
+        static_assert((rows_per_block % warp_size) == 0, "rows_per_block must be a multiple of warp_size.");
 #pragma unroll
         for (int i0 = 0; i0 < nwarps*rows_per_block; i0 += rows_per_block) {
-            const int i = i0 + threadIdx.x;
+#pragma unroll
+            for (int i1 = 0; i1 < sizeof(sum)/sizeof(sum[0]); ++i1) {
+                const int i = i0 + i1*warp_size + threadIdx.x;
 
-            sum += buf_iw[j*kiw + i];
+                sum[i1] += buf_iw[j*kiw + i];
+            }
         }
 
         if constexpr (!has_ids) {
-            dst[j*stride_col_dst + row0 + threadIdx.x] = sum;
+#pragma unroll
+            for (int i0 = 0; i0 < sizeof(sum)/sizeof(sum[0]); ++i0) {
+                dst[j*stride_col_dst + row0 + i0*warp_size + threadIdx.x] = sum[i0];
+            }
         } else {
             const int slot = (j < cols_per_block) ? slot_map[j] : -1;
             if (slot >= 0 && (col_base + j) < ncols_dst_total) {
-                dst[slot*stride_channel_dst + j*stride_col_dst + row0 + threadIdx.x] = sum;
+#pragma unroll
+                for (int i0 = 0; i0 < sizeof(sum)/sizeof(sum[0]); ++i0) {
+                    dst[slot*stride_channel_dst + j*stride_col_dst + row0 + i0*warp_size + threadIdx.x] = sum[i0];
+                }
             }
         }
     }
-#ifdef VOLTA_MMA_AVAILABLE
     }
-#endif //VOLTA_MMA_AVAILABLE
 #else
     GGML_UNUSED_VARS(x, y, ids, dst,
         ncols, ncols_dst_total, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
@@ -256,7 +290,7 @@ static __global__ void mul_mat_f(
         channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
         sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
     NO_DEVICE_CODE;
-#endif // (!defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)) || defined(AMD_WMMA_AVAILABLE)
+#endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
 }
 
 //This kernel is for larger batch sizes of mul_mat_id
@@ -271,23 +305,25 @@ static __global__ void mul_mat_f_ids(
         const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
         const uint3 sis1_fd, const uint3 nch_fd) {
 // TODO: handle this in a consistent and simpler way after AMD MFMA support has been added
-#if (!defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)) || defined(AMD_WMMA_AVAILABLE)
+#if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
 #if defined(AMD_WMMA_AVAILABLE)
-    // Special case for tf32, just dummy mma layout as wmma doesn't support it.
-    constexpr bool is_tf32 = std::is_same_v<T, float>;
-    constexpr int tile_B_I = is_tf32 ? 8 : 16;
-    constexpr int tile_C_J = is_tf32 ? 8 : 16;
-    constexpr data_layout ab_layout = is_tf32 ? DATA_LAYOUT_I_MAJOR : get_input_data_layout();
-    typedef tile<16,       8,        T,     ab_layout>           tile_A;
-    typedef tile<tile_B_I, 8,        T,     ab_layout>           tile_B;
-    typedef tile<16,       tile_C_J, float, DATA_LAYOUT_J_MAJOR> tile_C;
+    if constexpr (!(std::is_same_v<T, half2> || std::is_same_v<T, nv_bfloat162>) || rows_per_block != MMF_ROWS_PER_BLOCK) {NO_DEVICE_CODE;} else {
+    typedef tile<16, 8,  T,     get_input_data_layout()> tile_A;
+    typedef tile<16, 8,  T,     get_input_data_layout()> tile_B;
+    typedef tile<16, 16, float, DATA_LAYOUT_J_MAJOR>     tile_C;
+#elif defined(AMD_MFMA_AVAILABLE)
+    if constexpr (rows_per_block != MMF_ROWS_PER_BLOCK_CDNA) {NO_DEVICE_CODE;} else {
+    typedef tile<16, 8,  T,     DATA_LAYOUT_I_MAJOR> tile_A;
+    typedef tile<16, 8,  T,     DATA_LAYOUT_I_MAJOR> tile_B;
+    typedef tile<16, 16, float, DATA_LAYOUT_J_MAJOR> tile_C;
 #else
 #ifdef VOLTA_MMA_AVAILABLE
-    if constexpr (!std::is_same_v<T, half2>) {NO_DEVICE_CODE;} else {
+    if constexpr (!std::is_same_v<T, half2> || rows_per_block != MMF_ROWS_PER_BLOCK) {NO_DEVICE_CODE;} else {
     typedef tile<32, 4, T,     DATA_LAYOUT_I_MAJOR>          tile_A;
     typedef tile< 8, 4, T,     DATA_LAYOUT_I_MAJOR_MIRRORED> tile_B;
     typedef tile<32, 8, float, DATA_LAYOUT_I_MAJOR>          tile_C;
 #else
+    if constexpr (rows_per_block != MMF_ROWS_PER_BLOCK) {NO_DEVICE_CODE;} else {
     typedef tile<16, 8, T>     tile_A;
     typedef tile<8,  8, T>     tile_B;
     typedef tile<16, 8, float> tile_C;
@@ -300,7 +336,7 @@ static __global__ void mul_mat_f_ids(
 
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
-    constexpr int tile_k_padded = warp_size + 4;
+    constexpr int tile_k_padded = warp_size + mmf_get_padding();
     constexpr int ntA = rows_per_block / tile_A::I;
     constexpr int ntB = (cols_per_block + tile_B::I - 1) / tile_B::I;
 
@@ -467,7 +503,7 @@ static __global__ void mul_mat_f_ids(
     }
 
     float * buf_iw = (float *) compute_base;
-    constexpr int kiw = nwarps*rows_per_block + 4;
+    constexpr int kiw = nwarps*rows_per_block + mmf_get_padding();
 
     if (nwarps > 1) {
         __syncthreads();
@@ -497,13 +533,16 @@ static __global__ void mul_mat_f_ids(
             return;
         }
 
-        float sum = 0.0f;
-        static_assert(rows_per_block == warp_size, "need loop/check");
+        float sum[rows_per_block/warp_size] = {0.0f};
+        static_assert((rows_per_block % warp_size) == 0, "rows_per_block must be a multiple of warp_size.");
 #pragma unroll
         for (int i0 = 0; i0 < nwarps*rows_per_block; i0 += rows_per_block) {
-            const int i = i0 + threadIdx.x;
+#pragma unroll
+            for (int i1 = 0; i1 < sizeof(sum)/sizeof(sum[0]); ++i1) {
+                const int i = i0 + i1*warp_size + threadIdx.x;
 
-            sum += buf_iw[j*kiw + i];
+                sum[i1] += buf_iw[j * kiw + i];
+            }
         }
 
         const int global_j = col_base + j;
@@ -513,23 +552,24 @@ static __global__ void mul_mat_f_ids(
             const int token = (int) qrm.x;
             if (token < ncols_dst_total) {
                 const int slot = (int) qrm.y;
-                dst[slot*stride_channel_dst + token*stride_col_dst + row0 + threadIdx.x] = sum;
+#pragma unroll
+                for (int i0 = 0; i0 < sizeof(sum)/sizeof(sum[0]); ++i0) {
+                    dst[slot * stride_channel_dst + token * stride_col_dst + row0 + i0*warp_size + threadIdx.x] = sum[i0];
+                }
             }
         }
     }
-#ifdef VOLTA_MMA_AVAILABLE
     }
-#endif // VOLTA_MMA_AVAILABLE
 #else
     GGML_UNUSED_VARS(x, y, ids_src_compact, ids_dst_compact, expert_bounds, dst,
         ncols, ncols_dst_total, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
         channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
         sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, sis1_fd, nch_fd);
     NO_DEVICE_CODE;
-#endif // (!defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)) || defined(AMD_WMMA_AVAILABLE)
+#endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
 }
 
-template<typename T, int cols_per_block, int nwarps>
+template<typename T, int rows_per_block, int cols_per_block, int nwarps>
 static inline void mul_mat_f_switch_ids(
         const T * x, const float * y, const int32_t * ids, float * dst,
         const int64_t ncols_x, const int64_t ncols_dst, const int64_t nchannels_dst,
@@ -553,7 +593,7 @@ static inline void mul_mat_f_switch_ids(
         const uint3 sis1_fd = ids_data->sis1 > 0 ? init_fastdiv_values((uint32_t) ids_data->sis1) : make_uint3(0, 0, 1);
         const uint3 nch_fd  = init_fastdiv_values((uint32_t) nchannels_dst);
 
-        mul_mat_f_ids<T, MMF_ROWS_PER_BLOCK, cols_per_block, nwarps><<<block_nums_ids, block_dims, nbytes_shared_total, stream>>>
+        mul_mat_f_ids<T, rows_per_block, cols_per_block, nwarps><<<block_nums_ids, block_dims, nbytes_shared_total, stream>>>
             (x, y, ids_data->ids_src_compact, ids_data->ids_dst_compact, ids_data->expert_bounds_dev, dst,
             ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
             channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
@@ -564,19 +604,19 @@ static inline void mul_mat_f_switch_ids(
         dim3 block_nums_ids = block_nums;
         block_nums_ids.y *= col_tiles;
 
-        mul_mat_f<T, MMF_ROWS_PER_BLOCK, cols_per_block, nwarps, true><<<block_nums_ids, block_dims, nbytes_shared_total, stream>>>
+        mul_mat_f<T, rows_per_block, cols_per_block, nwarps, true><<<block_nums_ids, block_dims, nbytes_shared_total, stream>>>
             (x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
              stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
              sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
     } else {
-        mul_mat_f<T, MMF_ROWS_PER_BLOCK, cols_per_block, nwarps, false><<<block_nums, block_dims, nbytes_shared_total, stream>>>
+        mul_mat_f<T, rows_per_block, cols_per_block, nwarps, false><<<block_nums, block_dims, nbytes_shared_total, stream>>>
             (x, y, ids, dst, ncols_x, cols_per_block, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
              stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
              sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
     }
 }
 
-template <typename T, int cols_per_block>
+template <typename T, int rows_per_block, int cols_per_block>
 void mul_mat_f_cuda(
         const T * x, const float * y, const int32_t * ids, float * dst,
         const int64_t ncols_x, const int64_t nrows_x, const int64_t ncols_dst,
@@ -605,7 +645,7 @@ void mul_mat_f_cuda(
 
     int64_t nwarps_best     = 1;
     int64_t niter_best      = (ncols_x + warp_size*2 - 1) / (warp_size*2);
-    int64_t max_block_size  = 256;
+    int64_t max_block_size  = mmf_get_max_block_size(cc);
     for (int64_t nwarps = 2; nwarps <= max_block_size/warp_size; nwarps++) {
         const int64_t niter = (ncols_x + nwarps*warp_size*2 - 1) / (nwarps*warp_size*2);
         if (niter < niter_best) {
@@ -614,10 +654,9 @@ void mul_mat_f_cuda(
         }
     }
 
-    constexpr int rows_per_block = MMF_ROWS_PER_BLOCK;
-    const int nbytes_shared_iter = nwarps_best * (volta_mma_available(cc) ? tile_A_32::I : tile_A_16::I) * (warp_size + 4) * 4;
-    const int nbytes_cols_per_block_pad = amd_wmma_available(cc) ? tile_B_16::I : tile_B_8::I;
-    const int nbytes_shared_combine = GGML_PAD(cols_per_block, nbytes_cols_per_block_pad) * (nwarps_best*rows_per_block + 4) * 4;
+    const int nbytes_shared_iter = nwarps_best * (volta_mma_available(cc) ? tile_A_32::I : tile_A_16::I) * (warp_size + mmf_get_padding(cc)) * 4;
+    const int nbytes_cols_per_block_pad = (amd_wmma_available(cc) || amd_mfma_available(cc)) ? tile_B_16::I : tile_B_8::I;
+    const int nbytes_shared_combine = GGML_PAD(cols_per_block, nbytes_cols_per_block_pad) * (nwarps_best*rows_per_block + mmf_get_padding(cc)) * 4;
     const int nbytes_shared = std::max(nbytes_shared_iter, nbytes_shared_combine);
     const int nbytes_slotmap = ids ? GGML_PAD(cols_per_block, 16) * sizeof(int) : 0;
     const int nbytes_shared_total = nbytes_shared + nbytes_slotmap;
@@ -628,56 +667,56 @@ void mul_mat_f_cuda(
 
     switch (nwarps_best) {
         case 1: {
-            mul_mat_f_switch_ids<T, cols_per_block, 1>(
+            mul_mat_f_switch_ids<T, rows_per_block, cols_per_block, 1>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
                 ids_data);
         } break;
         case 2: {
-            mul_mat_f_switch_ids<T, cols_per_block, 2>(
+            mul_mat_f_switch_ids<T, rows_per_block, cols_per_block, 2>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
                 ids_data);
         } break;
         case 3: {
-            mul_mat_f_switch_ids<T, cols_per_block, 3>(
+            mul_mat_f_switch_ids<T, rows_per_block, cols_per_block, 3>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
                 ids_data);
         } break;
         case 4: {
-            mul_mat_f_switch_ids<T, cols_per_block, 4>(
+            mul_mat_f_switch_ids<T, rows_per_block, cols_per_block, 4>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
                 ids_data);
         } break;
         case 5: {
-            mul_mat_f_switch_ids<T, cols_per_block, 5>(
+            mul_mat_f_switch_ids<T, rows_per_block, cols_per_block, 5>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
                 ids_data);
         } break;
         case 6: {
-            mul_mat_f_switch_ids<T, cols_per_block, 6>(
+            mul_mat_f_switch_ids<T, rows_per_block, cols_per_block, 6>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
                 ids_data);
         } break;
         case 7: {
-            mul_mat_f_switch_ids<T, cols_per_block, 7>(
+            mul_mat_f_switch_ids<T, rows_per_block, cols_per_block, 7>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
                 ids_data);
         } break;
         case 8: {
-            mul_mat_f_switch_ids<T, cols_per_block, 8>(
+            mul_mat_f_switch_ids<T, rows_per_block, cols_per_block, 8>(
                 x, y, ids, dst, ncols_x, ncols_dst, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, block_nums, block_dims, nbytes_shared_total, stream,
@@ -691,7 +730,7 @@ void mul_mat_f_cuda(
     GGML_UNUSED_VARS(nchannels_y);
 }
 
-template <typename T>
+template <typename T, int rows_per_block>
 static void mul_mat_f_switch_cols_per_block(
         const T * x, const float * y, const int32_t * ids, float * dst,
         const int64_t ncols_x, const int64_t nrows_x, const int64_t ncols_dst,
@@ -708,82 +747,82 @@ static void mul_mat_f_switch_cols_per_block(
 
     switch (ncols_case) {
         case  1: {
-            mul_mat_f_cuda<T,  1>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  rows_per_block, 1>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  2: {
-            mul_mat_f_cuda<T,  2>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  rows_per_block, 2>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  3: {
-            mul_mat_f_cuda<T,  3>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  rows_per_block, 3>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  4: {
-            mul_mat_f_cuda<T,  4>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  rows_per_block, 4>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  5: {
-            mul_mat_f_cuda<T,  5>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  rows_per_block, 5>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y,  stride_sample_dst, stream, ids_data);
         } break;
         case  6: {
-            mul_mat_f_cuda<T,  6>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  rows_per_block, 6>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  7: {
-            mul_mat_f_cuda<T,  7>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  rows_per_block, 7>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  8: {
-            mul_mat_f_cuda<T,  8>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  rows_per_block, 8>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case  9: {
-            mul_mat_f_cuda<T,  9>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T,  rows_per_block, 9>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 10: {
-            mul_mat_f_cuda<T, 10>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T, rows_per_block, 10>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 11: {
-            mul_mat_f_cuda<T, 11>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T, rows_per_block, 11>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 12: {
-            mul_mat_f_cuda<T, 12>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T, rows_per_block, 12>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 13: {
-            mul_mat_f_cuda<T, 13>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T, rows_per_block, 13>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 14: {
-            mul_mat_f_cuda<T, 14>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T, rows_per_block, 14>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 15: {
-            mul_mat_f_cuda<T, 15>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T, rows_per_block, 15>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
         case 16: {
-            mul_mat_f_cuda<T, 16>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_f_cuda<T, rows_per_block, 16>(x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                 stride_col_id, stride_row_id, nchannels_x, nchannels_y,  nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
         } break;
@@ -793,8 +832,36 @@ static void mul_mat_f_switch_cols_per_block(
     }
 }
 
-#define DECL_MMF_CASE_HELPER(T, ncols_dst) \
-    template void mul_mat_f_cuda<T, ncols_dst>( \
+template <typename T>
+static void mul_mat_f_switch_rows_per_block(
+        const int rows_per_block, const T * x, const float * y, const int32_t * ids, float * dst,
+        const int64_t ncols_x, const int64_t nrows_x, const int64_t ncols_dst,
+        const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
+        const int64_t stride_col_id, const int stride_row_id,
+        const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
+        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
+        const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
+        cudaStream_t stream, const mmf_ids_data * ids_data) {
+    switch (rows_per_block) {
+        case MMF_ROWS_PER_BLOCK: {
+            mul_mat_f_switch_cols_per_block<T, MMF_ROWS_PER_BLOCK>(
+                x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
+        } break;
+        case MMF_ROWS_PER_BLOCK_CDNA: {
+            mul_mat_f_switch_cols_per_block<T, MMF_ROWS_PER_BLOCK_CDNA>(
+                x, y, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+                stride_col_id, stride_row_id, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream, ids_data);
+        } break;
+        default:
+            GGML_ABORT("unsupported rows_per_block: %i", rows_per_block);
+    }
+}
+
+#define DECL_MMF_CASE_HELPER(T, nrows_dst, ncols_dst) \
+    template void mul_mat_f_cuda<T, nrows_dst, ncols_dst>( \
         const T * x, const float * y, const int32_t * ids, float * dst, \
         const int64_t ncols_x, const int64_t nrows_x, int64_t ncols_dst_total, const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst, \
         const int64_t stride_col_id, const int64_t stride_row_id, \
@@ -803,16 +870,22 @@ static void mul_mat_f_switch_cols_per_block(
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst, \
         cudaStream_t stream, const mmf_ids_data * ids_data);
 
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_MUSA)
 #define DECL_MMF_CASE_EXTERN(ncols_dst) \
-    extern DECL_MMF_CASE_HELPER(float, ncols_dst) \
-    extern DECL_MMF_CASE_HELPER(half2, ncols_dst) \
-    extern DECL_MMF_CASE_HELPER(nv_bfloat162, ncols_dst)
+    extern DECL_MMF_CASE_HELPER(float, MMF_ROWS_PER_BLOCK, ncols_dst) \
+    extern DECL_MMF_CASE_HELPER(half2, MMF_ROWS_PER_BLOCK, ncols_dst) \
+    extern DECL_MMF_CASE_HELPER(nv_bfloat162, MMF_ROWS_PER_BLOCK, ncols_dst) \
+    extern DECL_MMF_CASE_HELPER(float, MMF_ROWS_PER_BLOCK_CDNA, ncols_dst) \
+    extern DECL_MMF_CASE_HELPER(half2, MMF_ROWS_PER_BLOCK_CDNA, ncols_dst) \
+    extern DECL_MMF_CASE_HELPER(nv_bfloat162, MMF_ROWS_PER_BLOCK_CDNA, ncols_dst)
 
 #define DECL_MMF_CASE(ncols_dst) \
-    DECL_MMF_CASE_HELPER(float, ncols_dst) \
-    DECL_MMF_CASE_HELPER(half2, ncols_dst) \
-    DECL_MMF_CASE_HELPER(nv_bfloat162, ncols_dst)
+    DECL_MMF_CASE_HELPER(float, MMF_ROWS_PER_BLOCK, ncols_dst) \
+    DECL_MMF_CASE_HELPER(half2, MMF_ROWS_PER_BLOCK, ncols_dst) \
+    DECL_MMF_CASE_HELPER(nv_bfloat162, MMF_ROWS_PER_BLOCK, ncols_dst) \
+    DECL_MMF_CASE_HELPER(float, MMF_ROWS_PER_BLOCK_CDNA, ncols_dst) \
+    DECL_MMF_CASE_HELPER(half2, MMF_ROWS_PER_BLOCK_CDNA, ncols_dst) \
+    DECL_MMF_CASE_HELPER(nv_bfloat162, MMF_ROWS_PER_BLOCK_CDNA, ncols_dst)
 
 DECL_MMF_CASE_EXTERN(1);
 DECL_MMF_CASE_EXTERN(2);
diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu
new file mode 100644
index 0000000000..1f554d81e5
--- /dev/null
+++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(576, 512, 1, 32);
diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu
index 2074e954a3..517993cb06 100644
--- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu
+++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu
@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 16, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 16, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 16, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 16, 4);
+DECL_FATTN_MMA_F16_CASE(576, 512, 16, 4);
diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu
new file mode 100644
index 0000000000..264751d65e
--- /dev/null
+++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-mma-f16.cuh"
+
+DECL_FATTN_MMA_F16_CASE(576, 512, 2, 32);
diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu
index 24c64cf000..97b19c67ad 100644
--- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu
+++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu
@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 2, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 2, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 2, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 2, 4);
+DECL_FATTN_MMA_F16_CASE(576, 512, 2, 4);
diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu
index 1ada657f19..989626dfa5 100644
--- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu
+++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu
@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 4, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 4, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 4, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 4, 4);
+DECL_FATTN_MMA_F16_CASE(576, 512, 4, 4);
diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu
index 86d4ffae27..173de7aac7 100644
--- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu
+++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu
@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 8, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 8, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 8, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 8, 4);
+DECL_FATTN_MMA_F16_CASE(576, 512, 8, 4);
diff --git a/ggml/src/ggml-cuda/template-instances/generate_cu_files.py b/ggml/src/ggml-cuda/template-instances/generate_cu_files.py
index a5602da02b..e382df1ae2 100755
--- a/ggml/src/ggml-cuda/template-instances/generate_cu_files.py
+++ b/ggml/src/ggml-cuda/template-instances/generate_cu_files.py
@@ -71,7 +71,7 @@ for type_k in TYPES_KV:
             f.write(SOURCE_FATTN_VEC.format(type_k=type_k, type_v=type_v))
 
 for ncols in [8, 16, 32, 64]:
-    for ncols2 in [1, 2, 4, 8, 16]:
+    for ncols2 in [1, 2, 4, 8, 16, 32]:
         if ncols2 > ncols:
             continue
         ncols1 = ncols // ncols2
@@ -83,9 +83,9 @@ for ncols in [8, 16, 32, 64]:
                     continue
                 if head_size_kq == 72:
                     continue
-                if head_size_kq != 576 and ncols2 == 16:
+                if head_size_kq != 576 and ncols2 in (16, 32):
                     continue
-                if head_size_kq == 576 and ncols2 != 16:
+                if head_size_kq == 576 and ncols2 not in (4, 16, 32):
                     continue
                 head_size_v = head_size_kq if head_size_kq != 576 else 512
                 f.write(SOURCE_FATTN_MMA_CASE.format(ncols1=ncols1, ncols2=ncols2, head_size_kq=head_size_kq, head_size_v=head_size_v))
diff --git a/ggml/src/ggml-cuda/top-k.cu b/ggml/src/ggml-cuda/top-k.cu
index 318ac38691..785a18389f 100644
--- a/ggml/src/ggml-cuda/top-k.cu
+++ b/ggml/src/ggml-cuda/top-k.cu
@@ -4,7 +4,6 @@
 #ifdef GGML_CUDA_USE_CUB
 #    include <cub/cub.cuh>
 #    if (CCCL_MAJOR_VERSION >= 3 && CCCL_MINOR_VERSION >= 2)
-#        include <cuda/iterator>
 #        define CUB_TOP_K_AVAILABLE
 using namespace cub;
 #    endif  // CCCL_MAJOR_VERSION >= 3 && CCCL_MINOR_VERSION >= 2
diff --git a/ggml/src/ggml-cuda/topk-moe.cu b/ggml/src/ggml-cuda/topk-moe.cu
index 48e569efa0..08a88990dd 100644
--- a/ggml/src/ggml-cuda/topk-moe.cu
+++ b/ggml/src/ggml-cuda/topk-moe.cu
@@ -5,6 +5,13 @@
 #include <cmath>
 #include <initializer_list>
 
+// Kernel config struct - passed by value to CUDA kernel
+struct topk_moe_config {
+    bool use_sigmoid;
+    bool with_norm;
+    bool delayed_softmax;
+};
+
 // Warp-local softmax used for both the pre-top-k logits and the post-top-k delayed path.
 template <int experts_per_thread, bool use_limit>
 __device__ void softmax_warp_inplace(float (&vals)[experts_per_thread], const int limit, const int lane) {
@@ -50,6 +57,16 @@ __device__ void softmax_warp_inplace(float (&vals)[experts_per_thread], const in
     }
 }
 
+template <int experts_per_thread, bool use_limit>
+__device__ void sigmoid_warp_inplace(float (&vals)[experts_per_thread], const int limit, const int lane) {
+#pragma unroll
+    for (int i = 0; i < experts_per_thread; i++) {
+        const int  idx    = lane + i * WARP_SIZE;
+        const bool active = !use_limit || (idx < limit);
+        vals[i]           = active ? 1.f / (1.f + expf(-vals[i])) : -INFINITY;
+    }
+}
+
 /*
     This kernel does the following:
     1. optionally softmax over the logits per token [n_experts, n_tokens]
@@ -59,13 +76,16 @@ __device__ void softmax_warp_inplace(float (&vals)[experts_per_thread], const in
 
     It is intended as fusion of softmax->top-k->get_rows pipeline for MoE models
 */
-template <int n_experts, bool with_norm, bool delayed_softmax = false>
-__launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float * logits,
-                                                                  float *       weights,
-                                                                  int32_t *     ids,
-                                                                  const int     n_rows,
-                                                                  const int     n_expert_used,
-                                                                  const float   clamp_val) {
+template <int n_experts, bool has_bias>
+__launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *         logits,
+                                                                  float *               weights,
+                                                                  int32_t *             ids,
+                                                                  float *               bias,
+                                                                  const int             n_rows,
+                                                                  const int             n_expert_used,
+                                                                  const float           clamp_val,
+                                                                  const float           scale_val,
+                                                                  const topk_moe_config config) {
     const int row = blockIdx.x * blockDim.y + threadIdx.y;
     if (row >= n_rows) {
         return;
@@ -79,14 +99,41 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
 
     float wt[experts_per_thread];
 
+    // Initialize all slots to -INFINITY
+#pragma unroll
+    for (int i = 0; i < experts_per_thread; i++) {
+        wt[i] = -INFINITY;
+    }
+
 #pragma unroll
     for (int i = 0; i < n_experts; i += WARP_SIZE) {
         const int expert  = i + threadIdx.x;
         wt[i / WARP_SIZE] = (n_experts % WARP_SIZE == 0 || expert < n_experts) ? logits[expert] : -INFINITY;
     }
 
-    if constexpr (!delayed_softmax) {
-        softmax_warp_inplace<experts_per_thread, false>(wt, n_experts, threadIdx.x);
+    if (!config.delayed_softmax) {
+        if (config.use_sigmoid) {
+           sigmoid_warp_inplace<experts_per_thread, false>(wt, n_experts, threadIdx.x);
+        } else {
+           softmax_warp_inplace<experts_per_thread, false>(wt, n_experts, threadIdx.x);
+        }
+    }
+
+    // selection_wt is only needed when bias is present (selection uses wt + bias)
+    // when no bias, we use wt directly for both selection and weight values
+    float selection_wt[has_bias ? experts_per_thread : 1];
+
+    if constexpr (has_bias) {
+#pragma unroll
+        for (int i = 0; i < experts_per_thread; i++) {
+            selection_wt[i] = -INFINITY;
+        }
+#pragma unroll
+        for (int i = 0; i < n_experts; i += WARP_SIZE) {
+            const int expert = i + threadIdx.x;
+            selection_wt[i / WARP_SIZE] =
+                (n_experts % WARP_SIZE == 0 || expert < n_experts) ? wt[i / WARP_SIZE] + bias[expert] : -INFINITY;
+        }
     }
 
     //at this point, each thread holds either a portion of the softmax distribution
@@ -106,22 +153,56 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
         float max_val    = wt[0];
         int   max_expert = threadIdx.x;
 
-#pragma unroll
-        for (int i = 1; i < experts_per_thread; i++) {
-            const int expert = threadIdx.x + i * WARP_SIZE;
-            if ((n_experts % WARP_SIZE == 0 || expert < n_experts) && wt[i] > max_val) {
-                max_val    = wt[i];
-                max_expert = expert;
-            }
-        }
+        if constexpr (has_bias) {
+            float max_val_s = selection_wt[0];
 
 #pragma unroll
-        for (int mask = WARP_SIZE / 2; mask > 0; mask /= 2) {
-            const float val    = __shfl_xor_sync(0xFFFFFFFF, max_val, mask, WARP_SIZE);
-            const int   expert = __shfl_xor_sync(0xFFFFFFFF, max_expert, mask, WARP_SIZE);
-            if (val > max_val || (val == max_val && expert < max_expert)) {
-                max_val    = val;
-                max_expert = expert;
+            for (int i = 1; i < experts_per_thread; i++) {
+                const int expert = threadIdx.x + i * WARP_SIZE;
+                if ((n_experts % WARP_SIZE == 0 || expert < n_experts) && selection_wt[i] > max_val_s) {
+                    max_val    = wt[i];
+                    max_val_s  = selection_wt[i];
+                    max_expert = expert;
+                }
+            }
+
+#pragma unroll
+            for (int mask = WARP_SIZE / 2; mask > 0; mask /= 2) {
+                const float val    = __shfl_xor_sync(0xFFFFFFFF, max_val, mask, WARP_SIZE);
+                const float val_s  = __shfl_xor_sync(0xFFFFFFFF, max_val_s, mask, WARP_SIZE);
+                const int   expert = __shfl_xor_sync(0xFFFFFFFF, max_expert, mask, WARP_SIZE);
+                if (val_s > max_val_s || (val_s == max_val_s && expert < max_expert)) {
+                    max_val    = val;
+                    max_val_s  = val_s;
+                    max_expert = expert;
+                }
+            }
+
+            if ((max_expert & (WARP_SIZE - 1)) == threadIdx.x) {
+                selection_wt[max_expert / WARP_SIZE] = -INFINITY;
+            }
+        } else {
+#pragma unroll
+            for (int i = 1; i < experts_per_thread; i++) {
+                const int expert = threadIdx.x + i * WARP_SIZE;
+                if ((n_experts % WARP_SIZE == 0 || expert < n_experts) && wt[i] > max_val) {
+                    max_val    = wt[i];
+                    max_expert = expert;
+                }
+            }
+
+#pragma unroll
+            for (int mask = WARP_SIZE / 2; mask > 0; mask /= 2) {
+                const float val    = __shfl_xor_sync(0xFFFFFFFF, max_val, mask, WARP_SIZE);
+                const int   expert = __shfl_xor_sync(0xFFFFFFFF, max_expert, mask, WARP_SIZE);
+                if (val > max_val || (val == max_val && expert < max_expert)) {
+                    max_val    = val;
+                    max_expert = expert;
+                }
+            }
+
+            if ((max_expert & (WARP_SIZE - 1)) == threadIdx.x) {
+                wt[max_expert / WARP_SIZE] = -INFINITY;
             }
         }
 
@@ -130,16 +211,14 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
         }
 
         if ((max_expert & (WARP_SIZE - 1)) == threadIdx.x) {
-            wt[max_expert / WARP_SIZE] = -INFINITY;
-
             ids[k] = max_expert;
-            if constexpr (with_norm) {
+            if (config.with_norm) {
                 wt_sum += max_val;
             }
         }
     }
 
-    if constexpr (with_norm) {
+    if (config.with_norm) {
         wt_sum              = warp_reduce_sum(wt_sum);
         wt_sum              = max(wt_sum, clamp_val);
         const float inv_sum = 1.0f / wt_sum;
@@ -149,7 +228,7 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
         }
     }
 
-    if constexpr (delayed_softmax) {
+    if (config.delayed_softmax) {
         softmax_warp_inplace<experts_per_thread, true>(output_weights, n_expert_used, threadIdx.x);
     }
 
@@ -157,25 +236,25 @@ __launch_bounds__(4 * WARP_SIZE, 1) __global__ void topk_moe_cuda(const float *
     for (int i = 0; i < experts_per_thread; i++) {
         const int idx = i * WARP_SIZE + threadIdx.x;
         if (idx < n_expert_used) {
-            weights[idx] = output_weights[i];
+            weights[idx] = output_weights[i] * scale_val;
         }
     }
-
-    if (!with_norm) {
-        GGML_UNUSED(clamp_val);
-    }
 }
 
-template <bool with_norm, bool delayed_softmax = false>
+template<bool has_bias>
 static void launch_topk_moe_cuda(ggml_backend_cuda_context & ctx,
                                  const float *               logits,
                                  float *                     weights,
                                  int32_t *                   ids,
+                                 float *                     bias,
                                  const int                   n_rows,
                                  const int                   n_expert,
                                  const int                   n_expert_used,
-                                 const float                 clamp_val) {
-    static_assert(!(with_norm && delayed_softmax), "delayed softmax is not supported with weight normalization");
+                                 const float                 clamp_val,
+                                 const float                 scale_val,
+                                 const topk_moe_config       config) {
+    GGML_ASSERT(!(config.with_norm && config.delayed_softmax) &&
+                "delayed softmax is not supported with weight normalization");
     const int    rows_per_block = 4;
     dim3         grid_dims((n_rows + rows_per_block - 1) / rows_per_block, 1, 1);
     dim3         block_dims(WARP_SIZE, rows_per_block, 1);
@@ -183,44 +262,48 @@ static void launch_topk_moe_cuda(ggml_backend_cuda_context & ctx,
 
     switch (n_expert) {
         case 1:
-            topk_moe_cuda<1, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
+            topk_moe_cuda<1, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
+                                                                   clamp_val, scale_val, config);
             break;
         case 2:
-            topk_moe_cuda<2, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
+            topk_moe_cuda<2, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
+                                                                   clamp_val, scale_val, config);
             break;
         case 4:
-            topk_moe_cuda<4, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
+            topk_moe_cuda<4, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
+                                                                   clamp_val, scale_val, config);
             break;
         case 8:
-            topk_moe_cuda<8, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
+            topk_moe_cuda<8, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
+                                                                   clamp_val, scale_val, config);
             break;
         case 16:
-            topk_moe_cuda<16, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
+            topk_moe_cuda<16, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
+                                                                    clamp_val, scale_val, config);
             break;
         case 32:
-            topk_moe_cuda<32, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
+            topk_moe_cuda<32, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
+                                                                    clamp_val, scale_val, config);
             break;
         case 64:
-            topk_moe_cuda<64, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
+            topk_moe_cuda<64, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
+                                                                    clamp_val, scale_val, config);
             break;
         case 128:
-            topk_moe_cuda<128, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
+            topk_moe_cuda<128, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
+                                                                     clamp_val, scale_val, config);
             break;
         case 256:
-            topk_moe_cuda<256, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
+            topk_moe_cuda<256, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
+                                                                     clamp_val, scale_val, config);
             break;
         case 512:
-            topk_moe_cuda<512, with_norm, delayed_softmax>
-                <<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, n_rows, n_expert_used, clamp_val);
+            topk_moe_cuda<512, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
+                                                                     clamp_val, scale_val, config);
+            break;
+        case 576:
+            topk_moe_cuda<576, has_bias><<<grid_dims, block_dims, 0, stream>>>(logits, weights, ids, bias, n_rows, n_expert_used,
+                                                                     clamp_val, scale_val, config);
             break;
         default:
             GGML_ASSERT(false && "fatal error");
@@ -228,13 +311,14 @@ static void launch_topk_moe_cuda(ggml_backend_cuda_context & ctx,
     }
 }
 
-void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx,
-                           const ggml_tensor *         logits,
-                           ggml_tensor *               weights,
-                           ggml_tensor *               ids,
-                           const bool                  with_norm,
-                           const bool                  delayed_softmax,
-                           ggml_tensor *               clamp) {
+void ggml_cuda_op_topk_moe(ggml_backend_cuda_context &     ctx,
+                           const ggml_tensor *             logits,
+                           ggml_tensor *                   weights,
+                           ggml_tensor *                   ids,
+                           const ggml_tensor *             clamp,
+                           const ggml_tensor *             scale,
+                           const ggml_tensor *             bias,
+                           const ggml_cuda_topk_moe_args & args) {
     GGML_ASSERT(logits->type == GGML_TYPE_F32);
     GGML_ASSERT(weights->type == GGML_TYPE_F32);
     GGML_ASSERT(ids->type == GGML_TYPE_I32);
@@ -245,107 +329,75 @@ void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx,
     const float * logits_d  = (const float *) logits->data;
     float *       weights_d = (float *) weights->data;
     int32_t *     ids_d     = (int32_t *) ids->data;
+    float *       bias_d    = bias ? (float *) bias->data : nullptr;
+
+    float scale_val = scale ? ggml_get_op_params_f32(scale, 0) : 1.0f;
 
     GGML_ASSERT(ids->nb[1] / ggml_type_size(ids->type) == (size_t) n_experts);
 
     const int n_expert_used = weights->ne[1];
 
+    const bool with_norm = clamp != nullptr;
+
     float clamp_val = -INFINITY;
-    if (with_norm) {
-        if (clamp) {
-            clamp_val = ggml_get_op_params_f32(clamp, 0);
-        }
-        launch_topk_moe_cuda<true>(ctx, logits_d, weights_d, ids_d, n_rows, n_experts, n_expert_used, clamp_val);
+    if (clamp) {
+        clamp_val = ggml_get_op_params_f32(clamp, 0);
+    }
+
+    topk_moe_config config;
+    config.use_sigmoid     = args.sigmoid;
+    config.with_norm       = with_norm;
+    config.delayed_softmax = args.delayed_softmax;
+
+    if (bias) {
+        launch_topk_moe_cuda<true>(ctx, logits_d, weights_d, ids_d, bias_d, n_rows, n_experts, n_expert_used, clamp_val,
+                             scale_val, config);
     } else {
-        GGML_ASSERT(clamp == nullptr);
-        if (delayed_softmax) {
-            launch_topk_moe_cuda<false, true>(ctx, logits_d, weights_d, ids_d, n_rows, n_experts, n_expert_used,
-                                              clamp_val);
-        } else {
-            launch_topk_moe_cuda<false, false>(ctx, logits_d, weights_d, ids_d, n_rows, n_experts, n_expert_used,
-                                               clamp_val);
-        }
+        launch_topk_moe_cuda<false>(ctx, logits_d, weights_d, ids_d, bias_d, n_rows, n_experts, n_expert_used, clamp_val,
+                             scale_val, config);
     }
 }
 
-bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax,
+bool ggml_cuda_should_use_topk_moe(const ggml_tensor * gating_op,
                                    const ggml_tensor * weights,
-                                   const ggml_tensor * get_rows,
-                                   const ggml_tensor * argsort,
-                                   const ggml_tensor * clamp,
-                                   int n_expert) {
-    ggml_tensor * probs = get_rows->src[0];
-    if (probs->op != GGML_OP_RESHAPE) {
-        return false;
-    }
-    probs = probs->src[0];
-    ggml_tensor * selection_probs = argsort->src[0];
-
-    if (probs != selection_probs) {
+                                   const ggml_tensor * logits,
+                                   const ggml_tensor * ids) {
+    const int n_expert = ids->nb[1] / ids->nb[0];
+    if (((n_expert & (n_expert - 1)) != 0 || n_expert > 512) && n_expert != 576) {
         return false;
     }
 
-    float scale    = 1.0f;
-    float max_bias = 0.0f;
-
-    memcpy(&scale, (const float *) softmax->op_params + 0, sizeof(float));
-    memcpy(&max_bias, (const float *) softmax->op_params + 1, sizeof(float));
-
-    if (!ggml_is_contiguous(softmax->src[0]) || !ggml_is_contiguous(weights)) {
+    if (!ggml_is_contiguous(weights) || !ggml_is_contiguous(logits)) {
         return false;
     }
 
-    if (scale != 1.0f || max_bias != 0.0f) {
-        return false;
-    }
+    if (gating_op->op == GGML_OP_SOFT_MAX) {
+        const ggml_tensor * softmax  = gating_op;
+        float               scale    = 1.0f;
+        float               max_bias = 0.0f;
 
-    // don't fuse when masks or sinks are present
-    if (softmax->src[1] || softmax->src[2]) {
-        return false;
-    }
+        memcpy(&scale, (const float *) softmax->op_params + 0, sizeof(float));
+        memcpy(&max_bias, (const float *) softmax->op_params + 1, sizeof(float));
 
-    // n_expert must be a power of 2
-    if ((n_expert & (n_expert - 1)) != 0 || n_expert > 512) {
-        return false;
-    }
-
-    if (clamp) {
-        if (clamp->op != GGML_OP_CLAMP) {
+        if (!ggml_is_contiguous(softmax->src[0])) {
             return false;
         }
-        float max_val = ggml_get_op_params_f32(clamp, 1);
 
-        if (max_val != INFINITY) {
+        if (scale != 1.0f || max_bias != 0.0f) {
+            return false;
+        }
+
+        // don't fuse when masks or sinks are present
+        if (softmax->src[1] || softmax->src[2]) {
+            return false;
+        }
+    } else if (gating_op->op == GGML_OP_UNARY) {
+        ggml_unary_op op = ggml_get_unary_op(gating_op);
+
+        if (op != GGML_UNARY_OP_SIGMOID) {
             return false;
         }
     }
 
-
     return true;
 }
-
-std::initializer_list<enum ggml_op> ggml_cuda_topk_moe_ops(bool norm, bool delayed_softmax) {
-    static std::initializer_list<enum ggml_op> norm_ops = { GGML_OP_SOFT_MAX, GGML_OP_RESHAPE,  GGML_OP_ARGSORT,
-                                                            GGML_OP_VIEW,     GGML_OP_GET_ROWS, GGML_OP_RESHAPE,
-                                                            GGML_OP_SUM_ROWS, GGML_OP_CLAMP,    GGML_OP_DIV,
-                                                            GGML_OP_RESHAPE };
-
-    static std::initializer_list<enum ggml_op> no_norm_ops = { GGML_OP_SOFT_MAX, GGML_OP_RESHAPE, GGML_OP_ARGSORT,
-                                                               GGML_OP_VIEW, GGML_OP_GET_ROWS };
-
-    static std::initializer_list<enum ggml_op> delayed_softmax_ops = { GGML_OP_ARGSORT,  GGML_OP_VIEW,
-                                                                       GGML_OP_GET_ROWS, GGML_OP_RESHAPE,
-                                                                       GGML_OP_SOFT_MAX, GGML_OP_RESHAPE };
-
-    GGML_ASSERT(!norm || !delayed_softmax);
-
-    if (delayed_softmax) {
-        return delayed_softmax_ops;
-    }
-
-    if (norm) {
-        return norm_ops;
-    }
-
-    return no_norm_ops;
-}
diff --git a/ggml/src/ggml-cuda/topk-moe.cuh b/ggml/src/ggml-cuda/topk-moe.cuh
index 6b6c13c587..243dc2f1c4 100644
--- a/ggml/src/ggml-cuda/topk-moe.cuh
+++ b/ggml/src/ggml-cuda/topk-moe.cuh
@@ -3,19 +3,25 @@
 
 #include <initializer_list>
 
-void ggml_cuda_op_topk_moe(ggml_backend_cuda_context & ctx,
-                           const ggml_tensor *         logits,
-                           ggml_tensor *               weights,
-                           ggml_tensor *               ids,
-                           const bool                  with_norm,
-                           const bool                  delayed_softmax = false,
-                           ggml_tensor *               weight_clamp    = nullptr);
+struct ggml_cuda_topk_moe_args {
+    bool sigmoid{};
+    bool softmax{};
+    bool delayed_softmax{};
+    bool prob_bias{};
+    bool norm{};
+    bool scale{};
+};
 
-bool ggml_cuda_should_use_topk_moe(const ggml_tensor * softmax,
+void ggml_cuda_op_topk_moe(ggml_backend_cuda_context &     ctx,
+                           const ggml_tensor *             logits,
+                           ggml_tensor *                   weights,
+                           ggml_tensor *                   ids,
+                           const ggml_tensor *             clamp,
+                           const ggml_tensor *             scale,
+                           const ggml_tensor *             bias,
+                           const ggml_cuda_topk_moe_args & args);
+
+bool ggml_cuda_should_use_topk_moe(const ggml_tensor * gating_op,
                                    const ggml_tensor * weights,
-                                   const ggml_tensor * get_rows,
-                                   const ggml_tensor * argsort,
-                                   const ggml_tensor * clamp,
-                                   int n_expert);
-
-std::initializer_list<enum ggml_op> ggml_cuda_topk_moe_ops(bool with_norm, bool delayed_softmax = false);
+                                   const ggml_tensor * logits,
+                                   const ggml_tensor * ids);
diff --git a/ggml/src/ggml-hexagon/CMakeLists.txt b/ggml/src/ggml-hexagon/CMakeLists.txt
index d58e287823..f3a583543c 100644
--- a/ggml/src/ggml-hexagon/CMakeLists.txt
+++ b/ggml/src/ggml-hexagon/CMakeLists.txt
@@ -1,7 +1,29 @@
+file(TO_CMAKE_PATH "${HEXAGON_SDK_ROOT}"   HEXAGON_SDK_ROOT)
+file(TO_CMAKE_PATH "${HEXAGON_TOOLS_ROOT}" HEXAGON_TOOLS_ROOT)
+
+if (NOT IS_DIRECTORY "${HEXAGON_SDK_ROOT}")
+    message(FATAL_ERROR "Make sure HEXAGON_SDK_ROOT point to the correct Hexagon SDK installation.")
+endif()
+
+if (NOT IS_DIRECTORY "${HEXAGON_TOOLS_ROOT}")
+    message("Try to read HEXAGON_TOOLS_ROOT from hexagon_sdk.json")
+    file(READ "${HEXAGON_SDK_ROOT}/hexagon_sdk.json" HEXAGON_SDK_CONFIG_PATH)
+    string(JSON HEXAGON_TOOLS_PATH GET ${HEXAGON_SDK_CONFIG_PATH} "root" "tools" "info" 0 "path")
+    message("Found HEXAGON_TOOLS_PATH: ${HEXAGON_TOOLS_PATH}")
+    set(HEXAGON_TOOLS_ROOT "${HEXAGON_SDK_ROOT}/${HEXAGON_TOOLS_PATH}")
+    file(TO_CMAKE_PATH "${HEXAGON_TOOLS_ROOT}" HEXAGON_TOOLS_ROOT)
+    if (NOT IS_DIRECTORY "${HEXAGON_TOOLS_ROOT}")
+        message(FATAL_ERROR "Make sure HEXAGON_TOOLS_ROOT point to the correct Hexagon SDK installation.")
+    endif()
+endif()
+
+message(STATUS "hexagon: using ${HEXAGON_SDK_ROOT} and ${HEXAGON_TOOLS_ROOT} for building libggml-htp skels")
+
 include(${HEXAGON_SDK_ROOT}/build/cmake/hexagon_fun.cmake)
 include(ExternalProject)
 
 option(GGML_HEXAGON_HTP_DEBUG "ggml-hexagon: enable HTP debug output" OFF)
+set(GGML_HEXAGON_HTP_CERT  "$ENV{HEXAGON_HTP_CERT}" CACHE PATH "ggml-hexagon: enable HTP library signing using certificate")
 set(GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE 128 CACHE STRING "ggml-hexagon: quantize group size (32, 64, or 128)")
 
 add_library(htp_iface OBJECT
@@ -25,56 +47,71 @@ else()
     target_link_options(htp_iface PUBLIC -ldl)
 endif()
 
-link_custom_library(htp_iface cdsprpc)
-link_custom_library(htp_iface rpcmem)
-
 set(TARGET_NAME ggml-hexagon)
 ggml_add_backend_library(${TARGET_NAME}
-    ggml-hexagon.cpp htp-utils.c htp-utils.h ../../include/ggml-hexagon.h)
+    ggml-hexagon.cpp
+    htp-drv.cpp
+    htp-drv.h
+    libdl.h
+    ../../include/ggml-hexagon.h)
 
 target_link_libraries(${TARGET_NAME} PRIVATE htp_iface)
 target_include_directories(${TARGET_NAME} PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/htp ${CMAKE_CURRENT_BINARY_DIR})
 
-# Build HTP bits
-set(HTP_CMAKE_ARGS
-    -DCMAKE_TOOLCHAIN_FILE=${CMAKE_CURRENT_SOURCE_DIR}/htp/cmake-toolchain.cmake
-    -DCMAKE_BUILD_TYPE=Release
-    -DCMAKE_INSTALL_LIBDIR=${CMAKE_CURRENT_BINARY_DIR}
-    -DHEXAGON_SDK_ROOT=$ENV{HEXAGON_SDK_ROOT}
-    -DHEXAGON_TOOLS_ROOT=$ENV{HEXAGON_TOOLS_ROOT}
-    -DHEXAGON_HTP_DEBUG=${GGML_HEXAGON_HTP_DEBUG}
-    -DGGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE=${GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE})
+# Build HTP skels
+set(HTP_SKELS)
+function(build_htp_skel V)
+    ExternalProject_Add(htp-${V}
+        SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
+        BUILD_BYPRODUCTS ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-${V}.so
+        CMAKE_ARGS
+            -DCMAKE_BUILD_TYPE=Release
+            -DCMAKE_TOOLCHAIN_FILE=${CMAKE_CURRENT_SOURCE_DIR}/htp/cmake-toolchain.cmake
+            -DCMAKE_INSTALL_LIBDIR=${CMAKE_CURRENT_BINARY_DIR}
+            -DHEXAGON_SDK_ROOT=${HEXAGON_SDK_ROOT}
+            -DHEXAGON_TOOLS_ROOT=${HEXAGON_TOOLS_ROOT}
+            -DHEXAGON_HTP_DEBUG=${GGML_HEXAGON_HTP_DEBUG}
+            -DGGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE=${GGML_HEXAGON_FP32_QUANTIZE_GROUP_SIZE}
+            -DDSP_VERSION=${V}
+            -DPREBUILT_LIB_DIR="toolv19_${V}")
+    list(APPEND HTP_SKELS ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-${V}.so)
+    set(HTP_SKELS ${HTP_SKELS} PARENT_SCOPE)
+endfunction()
 
-ExternalProject_Add(htp-v68
-    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
-    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v68 -DPREBUILT_LIB_DIR="toolv19_v68")
-
-ExternalProject_Add(htp-v69
-    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
-    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v69 -DPREBUILT_LIB_DIR="toolv19_v69")
-
-ExternalProject_Add(htp-v73
-    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
-    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v73 -DPREBUILT_LIB_DIR="toolv19_v73")
-
-ExternalProject_Add(htp-v75
-    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
-    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v75 -DPREBUILT_LIB_DIR="toolv19_v75")
-
-ExternalProject_Add(htp-v79
-    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
-    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v79 -DPREBUILT_LIB_DIR="toolv19_v79")
-
-ExternalProject_Add(htp-v81
-    SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/htp BUILD_ALWAYS ON
-    CMAKE_ARGS ${HTP_CMAKE_ARGS} -DDSP_VERSION=v81 -DPREBUILT_LIB_DIR="toolv19_v81")
+build_htp_skel(v68)
+build_htp_skel(v69)
+build_htp_skel(v73)
+build_htp_skel(v75)
+build_htp_skel(v79)
+build_htp_skel(v81)
 
 # Install Hexagon skels required at runtime
-install(FILES
-    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v68.so
-    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v69.so
-    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v73.so
-    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v75.so
-    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v79.so
-    ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp-v81.so
-    TYPE LIB)
+install(FILES ${HTP_SKELS} TYPE LIB)
+
+if (CMAKE_SYSTEM_NAME MATCHES Windows AND GGML_HEXAGON_HTP_CERT)
+    file(TO_CMAKE_PATH "$ENV{WINDOWS_SDK_BIN}/arm64"      WINSDK_BIN0_ARM64)
+    file(TO_CMAKE_PATH "$ENV{WINDOWS_SDK_BIN}/x86"        WINSDK_BIN0_X86)
+    file(TO_CMAKE_PATH "$ENV{WindowsSdkVerBinPath}/arm64" WINSDK_BIN1_ARM64)
+    file(TO_CMAKE_PATH "$ENV{WindowsSdkVerBinPath}/x86"   WINSDK_BIN1_X86)
+
+    set(WINSDK_PATHS ${WINSDK_BIN0_ARM64} ${WINSDK_BIN0_X86} ${WINSDK_BIN1_ARM64} ${WINSDK_BIN1_X86})
+
+    find_program(INF2CAT  NAMES inf2cat.exe  PATHS ${WINSDK_PATHS} REQUIRED)
+    find_program(SIGNTOOL NAMES signtool.exe PATHS ${WINSDK_PATHS} REQUIRED)
+
+    message(STATUS "hexagon: using ${GGML_HEXAGON_HTP_CERT} to sign libggml-htp skels")
+
+    set(LIBGGML_HTP_CAT ${CMAKE_CURRENT_BINARY_DIR}/libggml-htp.cat)
+    add_custom_target(libggml-htp-cat
+        BYPRODUCTS ${LIBGGML_HTP_CAT}
+        DEPENDS libggml-htp.inf ${HTP_SKELS}
+        COMMAND ${CMAKE_COMMAND} -E copy ${CMAKE_CURRENT_SOURCE_DIR}/libggml-htp.inf ${CMAKE_CURRENT_BINARY_DIR}
+        COMMAND ${INF2CAT} /driver:${CMAKE_CURRENT_BINARY_DIR} /os:10_25H2_ARM64
+        COMMAND ${SIGNTOOL} sign /fd sha256 /f ${GGML_HEXAGON_HTP_CERT} ${LIBGGML_HTP_CAT}
+        COMMENT "generating and signing libggml-htp.cat file"
+        VERBATIM
+    )
+
+    add_dependencies(${TARGET_NAME} libggml-htp-cat)
+    install(FILES ${LIBGGML_HTP_CAT} TYPE LIB)
+endif()
diff --git a/ggml/src/ggml-hexagon/ggml-hexagon.cpp b/ggml/src/ggml-hexagon/ggml-hexagon.cpp
index cf1eb994c3..4f0a1620fb 100644
--- a/ggml/src/ggml-hexagon/ggml-hexagon.cpp
+++ b/ggml/src/ggml-hexagon/ggml-hexagon.cpp
@@ -14,9 +14,6 @@
 
 #ifdef _WIN32
 #    include <sal.h>
-#    ifndef _WINDOWS
-#        define _WINDOWS
-#    endif
 #else
 #    include <semaphore.h>
 #    include <unistd.h>
@@ -25,8 +22,6 @@
 #pragma clang diagnostic ignored "-Wnested-anon-types"
 #pragma clang diagnostic ignored "-Wgnu-anonymous-struct"
 
-#include "htp-utils.h"
-
 #include <AEEStdErr.h>
 #include <dspqueue.h>
 #include <rpcmem.h>
@@ -40,6 +35,7 @@
 #include "op-desc.h"
 #include "htp-msg.h"
 #include "htp_iface.h"
+#include "htp-drv.h"
 
 static size_t opt_ndev         = 1;
 static size_t opt_nhvx         = 0; // use all
@@ -150,9 +146,9 @@ void ggml_hexagon_session::enqueue(struct htp_general_req &req, struct dspqueue_
                              0,                       // flags - the framework will autoset this
                              n_bufs,                  // number of buffers
                              bufs,                    // buffer references
-                             sizeof(req),
+                             sizeof(req),             // Message length
                              (const uint8_t *) &req,  // Message
-                             1000000                  // Timeout
+                             DSPQUEUE_TIMEOUT         // Timeout
     );
 
     if (err != 0) {
@@ -182,13 +178,13 @@ void ggml_hexagon_session::flush() {
 
         // Read response packet from queue
         int err = dspqueue_read(q, &flags,
-                                   HTP_MAX_PACKET_BUFFERS,  // Maximum number of buffer references
-                                   &n_bufs,                 // Number of buffer references
-                                   bufs,                    // Buffer references
-                                   sizeof(rsp),             // Max message length
-                                   &rsp_size,               // Message length
-                                   (uint8_t *) &rsp,
-                                   1000000);                // Timeout
+                                HTP_MAX_PACKET_BUFFERS,  // Maximum number of buffer references
+                                &n_bufs,                 // Number of buffer references
+                                bufs,                    // Buffer references
+                                sizeof(rsp),             // Max message length
+                                &rsp_size,               // Message length
+                                (uint8_t *) &rsp,        // Message
+                                DSPQUEUE_TIMEOUT);       // Timeout
 
         if (err == AEE_EEXPIRED) {
             // TODO: might need to bail out if the HTP is stuck on something
@@ -269,13 +265,7 @@ struct ggml_backend_hexagon_buffer_context {
     ggml_backend_hexagon_buffer_context(ggml_hexagon_session * sess, size_t size, bool repack) {
         size += 4 * 1024;  // extra page for padding
 
-        if (rpcmem_alloc2) {
-            this->base = (uint8_t *) rpcmem_alloc2(RPCMEM_HEAP_ID_SYSTEM, RPCMEM_DEFAULT_FLAGS | RPCMEM_HEAP_NOREG, size);
-        } else {
-            GGML_LOG_INFO("ggml-hex: %s rpcmem_alloc2 not found, falling back to rpcmem_alloc\n", sess->name.c_str());
-            this->base = (uint8_t *) rpcmem_alloc(RPCMEM_HEAP_ID_SYSTEM, RPCMEM_DEFAULT_FLAGS | RPCMEM_HEAP_NOREG, size);
-        }
-
+        this->base = (uint8_t *) rpcmem_alloc2(RPCMEM_HEAP_ID_SYSTEM, RPCMEM_DEFAULT_FLAGS | RPCMEM_HEAP_NOREG, size);
         if (!this->base) {
             GGML_LOG_ERROR("ggml-hex: %s failed to allocate buffer : size %zu\n", sess->name.c_str(), size);
             throw std::runtime_error("ggml-hex: rpcmem_alloc failed (see log for details)");
@@ -2461,12 +2451,12 @@ static void ggml_backend_hexagon_free(ggml_backend_t backend) {
 }
 
 static inline bool op_reuse_src1(const ggml_tensor * op1, const ggml_tensor * op0) {
-    return (op0 && op0->src[1] == op1->src[1] && ggml_is_quantized(op0->src[0]->type) && ggml_is_quantized(op1->src[1]->type));
+    return (op0 && op0->src[1] == op1->src[1] && ggml_is_quantized(op0->src[0]->type));
 }
 
 static inline bool is_compute_op(ggml_tensor *node)
 {
-    return !(ggml_op_is_empty(node->op) || ggml_is_empty(node));
+    return !ggml_op_is_empty(node->op) && !ggml_is_empty(node) && (node->flags & GGML_TENSOR_FLAG_COMPUTE);
 }
 
 // scan the graph and figure out last compute op index
@@ -2488,7 +2478,7 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
 
     const int last = last_compute_op(graph);
 
-    const struct ggml_tensor * prev_quant_op = nullptr;  // prev executed op with quantizer
+    const struct ggml_tensor * prev_op = nullptr;  // prev executed op
 
     for (int i = 0; i < graph->n_nodes; ++i) {
         ggml_tensor * node = graph->nodes[i];
@@ -2500,10 +2490,12 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
         uint32_t flags = 0;
 
         // skip quantizer if src1 is reused
-        if (op_reuse_src1(node, prev_quant_op)) {
+        if (op_reuse_src1(node, prev_op)) {
             flags |= HTP_OPFLAGS_SKIP_QUANTIZE;
         }
 
+        prev_op = node;
+
         // ask for early notification for the last Op
         if (i == last) {
             flags |= HTP_OPFLAGS_EARLY_WAKEUP;
@@ -2516,7 +2508,6 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
                 } else {
                     ggml_hexagon_dispatch_op<init_binary_req<false>>(sess, node, flags);
                 }
-                prev_quant_op = node;
                 break;
             case GGML_OP_MUL_MAT_ID:
                 if (ggml_is_quantized(node->src[0]->type)) {
@@ -2524,7 +2515,6 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
                 } else {
                     ggml_hexagon_dispatch_op<init_binary_id_req<false>>(sess, node, flags);
                 }
-                prev_quant_op = node;
                 break;
             case GGML_OP_MUL:
             case GGML_OP_ADD:
@@ -2666,7 +2656,7 @@ static std::vector<int> ggml_hexagon_graph_optimize_reorder(const std::vector<no
         }
 
         // that many nodes forward to search for stackable nodes that can reuse VTCM
-        constexpr int N_FORWARD = 8;
+        constexpr int N_FORWARD = 16;
 
         for (int i1 = i0 + 1; i1 < i0 + N_FORWARD && i1 < n; i1++) {
             if (used[i1]) {
@@ -3052,10 +3042,12 @@ ggml_hexagon_registry::ggml_hexagon_registry(ggml_backend_reg_t reg) {
         }
     }
 
+#if defined(__ANDROID__)
     if (opt_arch < 75) {
         opt_ndev = 1;
         GGML_LOG_WARN("ggml-hex: forcing ndev to 1 for SoCs archs lower than v75.\n");
     }
+#endif
 
     GGML_LOG_INFO("ggml-hex: Hexagon Arch version v%d\n", opt_arch);
 
@@ -3152,6 +3144,8 @@ static void ggml_hexagon_init(ggml_backend_reg * reg) {
         opt_arch = strtoul(str_arch, NULL, 0);
     }
 
+    opt_hostbuf = str_hostbuf ? atoi(str_hostbuf) : 1;
+
     reg->context = new ggml_hexagon_registry(reg);
 
     HEX_VERBOSE("ggml-hex: size-of-general-req %zu size-of-general-rsp %zu\n", sizeof(struct htp_general_req),
@@ -3176,6 +3170,11 @@ ggml_backend_reg_t ggml_backend_hexagon_reg(void) {
         static std::mutex           mutex;
         std::lock_guard<std::mutex> lock(mutex);
         if (!initialized) {
+            auto nErr = htpdrv_init();
+            if (nErr != AEE_SUCCESS) {
+                return NULL;
+            }
+
             ggml_hexagon_init(&reg);
         }
 
diff --git a/ggml/src/ggml-hexagon/htp-drv.cpp b/ggml/src/ggml-hexagon/htp-drv.cpp
new file mode 100644
index 0000000000..2530bb06d6
--- /dev/null
+++ b/ggml/src/ggml-hexagon/htp-drv.cpp
@@ -0,0 +1,418 @@
+// sample drv interface
+
+#pragma clang diagnostic ignored "-Wgnu-anonymous-struct"
+#pragma clang diagnostic ignored "-Wmissing-prototypes"
+#pragma clang diagnostic ignored "-Wsign-compare"
+
+#include <filesystem>
+#include <set>
+#include <sstream>
+#include <string>
+#ifdef _WIN32
+#   define WIN32_LEAN_AND_MEAN
+#   ifndef NOMINMAX
+#       define NOMINMAX
+#   endif
+#   include <windows.h>
+#   include <winevt.h>
+#else
+#    include <dlfcn.h>
+#    include <unistd.h>
+#endif
+#include "ggml-impl.h"
+#include "htp-drv.h"
+#include "libdl.h"
+
+#include <domain.h>
+
+//
+// Driver API types
+//
+
+typedef void * (*rpcmem_alloc_pfn_t)(int heapid, uint32_t flags, int size);
+typedef void * (*rpcmem_alloc2_pfn_t)(int heapid, uint32_t flags, size_t size);
+typedef void   (*rpcmem_free_pfn_t)(void * po);
+typedef int    (*rpcmem_to_fd_pfn_t)(void * po);
+
+typedef AEEResult (*dspqueue_create_pfn_t)(int                 domain,
+                                           uint32_t            flags,
+                                           uint32_t            req_queue_size,
+                                           uint32_t            resp_queue_size,
+                                           dspqueue_callback_t packet_callback,
+                                           dspqueue_callback_t error_callback,
+                                           void *              callback_context,
+                                           dspqueue_t *        queue);
+typedef AEEResult (*dspqueue_close_pfn_t)(dspqueue_t queue);
+typedef AEEResult (*dspqueue_export_pfn_t)(dspqueue_t queue, uint64_t *queue_id);
+typedef AEEResult (*dspqueue_write_pfn_t)(dspqueue_t queue, uint32_t flags,
+                                          uint32_t num_buffers,
+                                          struct dspqueue_buffer *buffers,
+                                          uint32_t message_length,
+                                          const uint8_t *message,
+                                          uint32_t timeout_us);
+typedef AEEResult (*dspqueue_read_pfn_t)(dspqueue_t queue, uint32_t *flags,
+                                         uint32_t max_buffers, uint32_t *num_buffers,
+                                         struct dspqueue_buffer *buffers,
+                                         uint32_t max_message_length,
+                                         uint32_t *message_length, uint8_t *message,
+                                         uint32_t timeout_us);
+
+typedef int (*fastrpc_mmap_pfn_t)(int domain, int fd, void *addr, int offset, size_t length, enum fastrpc_map_flags flags);
+typedef int (*fastrpc_munmap_pfn_t)(int domain, int fd, void *addr, size_t length);
+
+typedef int (*remote_handle64_open_pfn_t)(const char* name, remote_handle64 *ph);
+typedef int (*remote_handle64_invoke_pfn_t)(remote_handle64 h, uint32_t dwScalars, remote_arg *pra);
+typedef int (*remote_handle64_close_pfn_t)(remote_handle h);
+typedef int (*remote_handle_control_pfn_t)(uint32_t req, void* data, uint32_t datalen);
+typedef int (*remote_handle64_control_pfn_t)(remote_handle64 h, uint32_t req, void* data, uint32_t datalen);
+typedef int (*remote_session_control_pfn_t)(uint32_t req, void *data, uint32_t datalen);
+
+//
+// Driver API pfns
+//
+
+rpcmem_alloc_pfn_t  rpcmem_alloc_pfn  = nullptr;
+rpcmem_alloc2_pfn_t rpcmem_alloc2_pfn = nullptr;
+rpcmem_free_pfn_t   rpcmem_free_pfn   = nullptr;
+rpcmem_to_fd_pfn_t  rpcmem_to_fd_pfn  = nullptr;
+
+fastrpc_mmap_pfn_t   fastrpc_mmap_pfn   = nullptr;
+fastrpc_munmap_pfn_t fastrpc_munmap_pfn = nullptr;
+
+dspqueue_create_pfn_t dspqueue_create_pfn = nullptr;
+dspqueue_close_pfn_t  dspqueue_close_pfn  = nullptr;
+dspqueue_export_pfn_t dspqueue_export_pfn = nullptr;
+dspqueue_write_pfn_t  dspqueue_write_pfn  = nullptr;
+dspqueue_read_pfn_t   dspqueue_read_pfn   = nullptr;
+
+remote_handle64_open_pfn_t    remote_handle64_open_pfn    = nullptr;
+remote_handle64_invoke_pfn_t  remote_handle64_invoke_pfn  = nullptr;
+remote_handle64_close_pfn_t   remote_handle64_close_pfn   = nullptr;
+remote_handle_control_pfn_t   remote_handle_control_pfn   = nullptr;
+remote_handle64_control_pfn_t remote_handle64_control_pfn = nullptr;
+remote_session_control_pfn_t  remote_session_control_pfn  = nullptr;
+
+//
+// Driver API
+//
+
+void * rpcmem_alloc(int heapid, uint32_t flags, int size) {
+    return rpcmem_alloc_pfn(heapid, flags, size);
+}
+
+void * rpcmem_alloc2(int heapid, uint32_t flags, size_t size) {
+    if (rpcmem_alloc2_pfn) {
+        return rpcmem_alloc2_pfn(heapid, flags, size);
+    } else {
+        GGML_LOG_INFO("ggml-hex: rpcmem_alloc2 not found, falling back to rpcmem_alloc\n");
+        return rpcmem_alloc_pfn(heapid, flags, size);
+    }
+}
+
+void rpcmem_free(void * po) {
+    return rpcmem_free_pfn(po);
+}
+
+int rpcmem_to_fd(void * po) {
+    return rpcmem_to_fd_pfn(po);
+}
+
+HTPDRV_API int fastrpc_mmap(int domain, int fd, void * addr, int offset, size_t length, enum fastrpc_map_flags flags) {
+    return fastrpc_mmap_pfn(domain, fd, addr, offset, length, flags);
+}
+
+HTPDRV_API int fastrpc_munmap(int domain, int fd, void * addr, size_t length) {
+    return fastrpc_munmap_pfn(domain, fd, addr, length);
+}
+
+AEEResult dspqueue_create(int                 domain,
+                          uint32_t            flags,
+                          uint32_t            req_queue_size,
+                          uint32_t            resp_queue_size,
+                          dspqueue_callback_t packet_callback,
+                          dspqueue_callback_t error_callback,
+                          void *              callback_context,
+                          dspqueue_t *        queue) {
+    return dspqueue_create_pfn(domain, flags, req_queue_size, resp_queue_size, packet_callback, error_callback,
+                               callback_context, queue);
+}
+
+AEEResult dspqueue_close(dspqueue_t queue) {
+    return dspqueue_close_pfn(queue);
+}
+
+AEEResult dspqueue_export(dspqueue_t queue, uint64_t * queue_id) {
+    return dspqueue_export_pfn(queue, queue_id);
+}
+
+AEEResult dspqueue_write(dspqueue_t               queue,
+                         uint32_t                 flags,
+                         uint32_t                 num_buffers,
+                         struct dspqueue_buffer * buffers,
+                         uint32_t                 message_length,
+                         const uint8_t *          message,
+                         uint32_t                 timeout_us) {
+    return dspqueue_write_pfn(queue, flags, num_buffers, buffers, message_length, message, timeout_us);
+}
+
+AEEResult dspqueue_read(dspqueue_t               queue,
+                        uint32_t *               flags,
+                        uint32_t                 max_buffers,
+                        uint32_t *               num_buffers,
+                        struct dspqueue_buffer * buffers,
+                        uint32_t                 max_message_length,
+                        uint32_t *               message_length,
+                        uint8_t *                message,
+                        uint32_t                 timeout_us) {
+    return dspqueue_read_pfn(queue, flags, max_buffers, num_buffers, buffers, max_message_length, message_length,
+                             message, timeout_us);
+}
+
+HTPDRV_API int remote_handle64_open(const char * name, remote_handle64 * ph) {
+    return remote_handle64_open_pfn(name, ph);
+}
+
+HTPDRV_API int remote_handle64_invoke(remote_handle64 h, uint32_t dwScalars, remote_arg * pra) {
+    return remote_handle64_invoke_pfn(h, dwScalars, pra);
+}
+
+HTPDRV_API int remote_handle64_close(remote_handle64 h) {
+    return remote_handle64_close_pfn(h);
+}
+
+HTPDRV_API int remote_handle_control(uint32_t req, void * data, uint32_t datalen) {
+    return remote_handle_control_pfn(req, data, datalen);
+}
+
+HTPDRV_API int remote_handle64_control(remote_handle64 h, uint32_t req, void * data, uint32_t datalen) {
+    return remote_handle64_control_pfn(h, req, data, datalen);
+}
+
+HTPDRV_API int remote_session_control(uint32_t req, void * data, uint32_t datalen) {
+    return remote_session_control_pfn(req, data, datalen);
+}
+
+#ifdef _WIN32
+
+static std::string wstr_to_str(std::wstring_view wstr) {
+    std::string result;
+    if (wstr.empty()) {
+        return result;
+    }
+    auto bytes_needed = WideCharToMultiByte(CP_UTF8, WC_ERR_INVALID_CHARS,
+                                            wstr.data(), (int) wstr.size(),
+                                            nullptr, 0, nullptr, nullptr);
+    if (bytes_needed == 0) {
+        GGML_LOG_ERROR("ggml-hex: WideCharToMultiByte failed. Error %lu\n", GetLastError());
+        throw std::runtime_error("Invalid wstring input");
+    }
+
+    result.resize(bytes_needed, '\0');
+    int bytes_written = WideCharToMultiByte(CP_UTF8, WC_ERR_INVALID_CHARS,
+                                            wstr.data(), (int) wstr.size(),
+                                            result.data(), bytes_needed,
+                                            nullptr, nullptr);
+    if (bytes_written == 0) {
+        GGML_LOG_ERROR("ggml-hex: WideCharToMultiByte failed. Error %lu\n", GetLastError());
+        throw std::runtime_error("Wstring conversion failed");
+    }
+    return result;
+}
+
+static std::string get_driver_path() {
+    std::wstring serviceName = L"qcnspmcdm";
+    std::string result;
+
+    // Get a handle to the SCM database.
+    SC_HANDLE schSCManager = OpenSCManagerW(NULL, NULL, STANDARD_RIGHTS_READ);
+    if (nullptr == schSCManager) {
+        GGML_LOG_ERROR("ggml-hex: Failed to open SCManager. Error: %lu\n", GetLastError());
+        return result;
+    }
+
+    // Get a handle to the service.
+    SC_HANDLE schService = OpenServiceW(schSCManager,           // SCM database
+                                        serviceName.c_str(),    // name of service
+                                        SERVICE_QUERY_CONFIG);  // need query config access
+
+    if (nullptr == schService) {
+        GGML_LOG_ERROR("ggml-hex: Failed to open qcnspmcdm service. Error: %lu\n", GetLastError());
+        CloseServiceHandle(schSCManager);
+        return result;
+    }
+
+    // Store the size of buffer used as an output.
+    DWORD bufferSize;
+    if (!QueryServiceConfigW(schService, NULL, 0, &bufferSize) &&
+        (GetLastError() != ERROR_INSUFFICIENT_BUFFER)) {
+        GGML_LOG_ERROR("ggml-hex: Failed to query service config. Error: %lu\n", GetLastError());
+        CloseServiceHandle(schService);
+        CloseServiceHandle(schSCManager);
+        return result;
+    }
+    // Get the configuration of the service.
+    LPQUERY_SERVICE_CONFIGW serviceConfig =
+        static_cast<LPQUERY_SERVICE_CONFIGW>(LocalAlloc(LMEM_FIXED, bufferSize));
+    if (!QueryServiceConfigW(schService, serviceConfig, bufferSize, &bufferSize)) {
+        fprintf(stderr, "ggml-hex: Failed to query service config. Error: %lu\n", GetLastError());
+        LocalFree(serviceConfig);
+        CloseServiceHandle(schService);
+        CloseServiceHandle(schSCManager);
+        return result;
+    }
+
+    // Read the driver file path get its parent directory
+    std::wstring driverPath = std::wstring(serviceConfig->lpBinaryPathName);
+    driverPath = driverPath.substr(0, driverPath.find_last_of(L"\\"));
+
+    // Clean up resources
+    LocalFree(serviceConfig);
+    CloseServiceHandle(schService);
+    CloseServiceHandle(schSCManager);
+
+    // Driver path would contain invalid path string, like:
+    // \SystemRoot\System32\DriverStore\FileRepository\qcadsprpc8280.inf_arm64_c2b9460c9a072f37
+    // "\SystemRoot" should be replace with a correct one (e.g. C:\Windows)
+    const std::wstring systemRootPlaceholder = L"\\SystemRoot";
+    if (0 != driverPath.compare(0, systemRootPlaceholder.length(), systemRootPlaceholder)) {
+        GGML_LOG_ERROR("ggml-hex: String pattern not found in driver path.\n");
+        return result;
+    }
+
+    // Replace \SystemRoot with an absolute path from system ENV windir
+    const std::wstring systemRootEnv = L"windir";
+
+    // Query the number of wide charactors this variable requires
+    DWORD numWords = GetEnvironmentVariableW(systemRootEnv.c_str(), NULL, 0);
+    if (numWords == 0) {
+        GGML_LOG_ERROR("ggml-hex: Failed get systemRoot environment variable\n");
+        return result;
+    }
+
+    // Query the actual system root name from environment variable
+    std::vector<wchar_t> systemRoot(numWords + 1);
+    numWords = GetEnvironmentVariableW(systemRootEnv.c_str(), systemRoot.data(), numWords + 1);
+    if (numWords == 0) {
+        GGML_LOG_ERROR("ggml-hex: Failed to read windir environment variable\n");
+        return result;
+    }
+    driverPath.replace(0, systemRootPlaceholder.length(), std::wstring(systemRoot.data()));
+
+    return wstr_to_str(driverPath);
+}
+
+#endif
+
+using dl_handle_ptr = std::unique_ptr<dl_handle, dl_handle_deleter>;
+
+int htpdrv_init() {
+    static dl_handle_ptr lib_cdsp_rpc_handle = nullptr;
+    static bool initialized = false;
+#ifdef _WIN32
+    std::string drv_path = get_driver_path() + "\\" + "libcdsprpc.dll";
+#else
+    std::string drv_path = "libcdsprpc.so";
+#endif
+    if (initialized) {
+        GGML_LOG_INFO("ggml-hex: Driver already loaded\n");
+        return AEE_SUCCESS;
+    }
+    GGML_LOG_INFO("ggml-hex: Loading driver %s\n", drv_path.c_str());
+
+    fs::path path{ drv_path.c_str() };
+    dl_handle_ptr handle { dl_load_library(path) };
+    if (!handle) {
+        GGML_LOG_ERROR("ggml-hex: failed to load %s: %s\n", path.u8string().c_str(), dl_error());
+        return AEE_EUNABLETOLOAD;
+    }
+
+#define dlsym(drv, type, pfn, symbol, ignore)                               \
+    do {                                                                    \
+        pfn = (type) dl_get_sym(drv, #symbol);                              \
+        if (!ignore && nullptr == pfn) {                                    \
+            GGML_LOG_ERROR("ggml-hex: failed to dlsym %s\n", #symbol);      \
+            return AEE_EUNABLETOLOAD;                                       \
+        }                                                                   \
+    } while (0)
+
+    dlsym(handle.get(), rpcmem_alloc_pfn_t, rpcmem_alloc_pfn, rpcmem_alloc, false);
+    dlsym(handle.get(), rpcmem_alloc2_pfn_t, rpcmem_alloc2_pfn, rpcmem_alloc2, true);
+    dlsym(handle.get(), rpcmem_free_pfn_t, rpcmem_free_pfn, rpcmem_free, false);
+    dlsym(handle.get(), rpcmem_to_fd_pfn_t, rpcmem_to_fd_pfn, rpcmem_to_fd, false);
+    dlsym(handle.get(), fastrpc_mmap_pfn_t, fastrpc_mmap_pfn, fastrpc_mmap, false);
+    dlsym(handle.get(), fastrpc_munmap_pfn_t, fastrpc_munmap_pfn, fastrpc_munmap, false);
+    dlsym(handle.get(), dspqueue_create_pfn_t, dspqueue_create_pfn, dspqueue_create, false);
+    dlsym(handle.get(), dspqueue_close_pfn_t, dspqueue_close_pfn, dspqueue_close, false);
+    dlsym(handle.get(), dspqueue_export_pfn_t, dspqueue_export_pfn, dspqueue_export, false);
+    dlsym(handle.get(), dspqueue_write_pfn_t, dspqueue_write_pfn, dspqueue_write, false);
+    dlsym(handle.get(), dspqueue_read_pfn_t, dspqueue_read_pfn, dspqueue_read, false);
+    dlsym(handle.get(), remote_handle64_open_pfn_t, remote_handle64_open_pfn, remote_handle64_open, false);
+    dlsym(handle.get(), remote_handle64_invoke_pfn_t, remote_handle64_invoke_pfn, remote_handle64_invoke, false);
+    dlsym(handle.get(), remote_handle_control_pfn_t, remote_handle_control_pfn, remote_handle_control, false);
+    dlsym(handle.get(), remote_handle64_control_pfn_t, remote_handle64_control_pfn, remote_handle64_control, false);
+    dlsym(handle.get(), remote_session_control_pfn_t, remote_session_control_pfn, remote_session_control, false);
+    dlsym(handle.get(), remote_handle64_close_pfn_t, remote_handle64_close_pfn, remote_handle64_close, false);
+
+    lib_cdsp_rpc_handle = std::move(handle);
+    initialized         = true;
+
+    return AEE_SUCCESS;
+}
+
+domain * get_domain(int domain_id) {
+    int i    = 0;
+    int size = sizeof(supported_domains) / sizeof(domain);
+
+    for (i = 0; i < size; i++) {
+        if (supported_domains[i].id == domain_id) {
+            return &supported_domains[i];
+        }
+    }
+
+    return NULL;
+}
+
+int get_hex_arch_ver(int domain, int * arch) {
+    if (!remote_handle_control_pfn) {
+        GGML_LOG_ERROR("ggml-hex: remote_handle_control is not supported on this device\n");
+        return AEE_EUNSUPPORTEDAPI;
+    }
+
+    struct remote_dsp_capability arch_ver;
+    arch_ver.domain       = (uint32_t) domain;
+    arch_ver.attribute_ID = ARCH_VER;
+    arch_ver.capability   = (uint32_t) 0;
+
+    int err = remote_handle_control(DSPRPC_GET_DSP_INFO, &arch_ver, sizeof(arch_ver));
+    if ((err & 0xff) == (AEE_EUNSUPPORTEDAPI & 0xff)) {
+        GGML_LOG_ERROR("ggml-hex: FastRPC capability API is not supported on this device\n");
+        return AEE_EUNSUPPORTEDAPI;
+    }
+
+    if (err != AEE_SUCCESS) {
+        GGML_LOG_ERROR("ggml-hex: FastRPC capability query failed (err %d)\n", err);
+        return err;
+    }
+
+    switch (arch_ver.capability & 0xff) {
+        case 0x68:
+            *arch = 68;
+            return 0;
+        case 0x69:
+            *arch = 69;
+            return 0;
+        case 0x73:
+            *arch = 73;
+            return 0;
+        case 0x75:
+            *arch = 75;
+            return 0;
+        case 0x79:
+            *arch = 79;
+            return 0;
+        case 0x81:
+            *arch = 81;
+            return 0;
+    }
+    return -1;
+}
diff --git a/ggml/src/ggml-hexagon/htp-drv.h b/ggml/src/ggml-hexagon/htp-drv.h
new file mode 100644
index 0000000000..6eba7ba17d
--- /dev/null
+++ b/ggml/src/ggml-hexagon/htp-drv.h
@@ -0,0 +1,121 @@
+#pragma once
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifdef _WIN32
+#    pragma clang diagnostic ignored "-Wignored-attributes"
+#endif
+
+#include <AEEStdErr.h>
+#include <rpcmem.h>
+#include <remote.h>
+#include <dspqueue.h>
+
+#if defined(_WIN32) && !defined(__MINGW32__)
+#    ifdef GGML_BACKEND_BUILD
+#        define HTPDRV_API __declspec(dllexport) extern
+#    else
+#        define HTPDRV_API __declspec(dllimport) extern
+#    endif
+#else
+#    define HTPDRV_API __attribute__ ((visibility ("default"))) extern
+#endif
+
+/* Offset to differentiate HLOS and Hexagon error codes.
+   Stores the value of AEE_EOFFSET for Hexagon. */
+#ifndef DSP_OFFSET
+#    define DSP_OFFSET 0x80000400
+#endif
+
+/* Errno for connection reset by peer. */
+#ifndef ECONNRESET
+#    ifdef __hexagon__
+#        define ECONNRESET 104
+#    endif
+#endif
+
+/* Abstraction of different OS specific sleep APIs.
+   SLEEP accepts input in seconds. */
+#ifndef SLEEP
+#    ifdef __hexagon__
+#        define SLEEP(x)                      \
+            { /* Do nothing for simulator. */ \
+            }
+#    else
+#        ifdef _WIN32
+#            define SLEEP(x) Sleep(1000 * x) /* Sleep accepts input in milliseconds. */
+#        else
+#            define SLEEP(x) sleep(x)        /* sleep accepts input in seconds. */
+#        endif
+#    endif
+#endif
+
+/* Include windows specific header files. */
+#ifdef _WIN32
+#    include <windows.h>
+#    include <sysinfoapi.h>
+#    define _CRT_SECURE_NO_WARNINGS         1
+#    define _WINSOCK_DEPRECATED_NO_WARNINGS 1
+#endif
+
+/* Includes and defines for all HLOS except windows */
+#if !defined(__hexagon__) && !defined(_WIN32)
+#    include "unistd.h"
+
+#    include <sys/time.h>
+#endif
+
+/* Includes and defines for Hexagon and all HLOS except Windows. */
+#if !defined(_WIN32)
+/* Weak reference to remote symbol for compilation. */
+#    pragma weak remote_session_control
+#    pragma weak remote_handle_control
+#    pragma weak remote_handle64_control
+#    pragma weak fastrpc_mmap
+#    pragma weak fastrpc_munmap
+#    pragma weak rpcmem_alloc2
+#endif
+
+#if !defined(_WIN32)
+#    pragma weak remote_system_request
+#endif
+
+#ifdef _WIN32
+#     define DSPQUEUE_TIMEOUT DSPQUEUE_TIMEOUT_NONE
+#else
+#     define DSPQUEUE_TIMEOUT 1000000
+#endif
+
+/**
+ * htpdrv_init API: driver interface entry point
+ *
+ * @return      Return AEE error codes as defined in Hexagon SDK.
+ */
+HTPDRV_API int htpdrv_init(void);
+
+/**
+ * get_domain API: get domain struct from domain value.
+ *
+ * @param[in]  domain value of a domain
+ * @return     Returns domain struct of the domain if it is supported or else
+ *             returns NULL.
+ *
+ */
+HTPDRV_API domain * get_domain(int domain_id);
+
+/**
+ * get_hex_arch_ver API: query the Hexagon processor architecture version information
+ *
+ * @param[in]   domain_id value of a domain
+ * @param[out]  Arch version (73, 75, ...)
+ * @return      0 if query is successful.
+ *              non-zero if error, return value points to the error.
+ *
+ */
+HTPDRV_API int get_hex_arch_ver(int domain, int * arch);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/ggml/src/ggml-hexagon/htp-utils.c b/ggml/src/ggml-hexagon/htp-utils.c
deleted file mode 100644
index 3f335bf71c..0000000000
--- a/ggml/src/ggml-hexagon/htp-utils.c
+++ /dev/null
@@ -1,454 +0,0 @@
-
-#pragma clang diagnostic ignored "-Wgnu-anonymous-struct"
-#pragma clang diagnostic ignored "-Wmissing-prototypes"
-#pragma clang diagnostic ignored "-Wsign-compare"
-
-#define GGML_COMMON_IMPL_C
-#include "ggml-backend-impl.h"
-#include "ggml-common.h"
-#include "ggml-hexagon.h"
-#include "ggml-impl.h"
-
-#include "htp-utils.h"
-
-#include <domain.h>
-#include <remote.h>
-#include <stdbool.h>
-#include <stdint.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-
-domain * get_domain(int domain_id) {
-    int i    = 0;
-    int size = sizeof(supported_domains) / sizeof(domain);
-
-    for (i = 0; i < size; i++) {
-        if (supported_domains[i].id == domain_id) {
-            return &supported_domains[i];
-        }
-    }
-
-    return NULL;
-}
-
-bool is_valid_domain_id(int domain_id, int compute_only) {
-    int i    = 0;
-    int size = sizeof(supported_domains) / sizeof(domain);
-
-    if (compute_only) {
-        return is_CDSP(domain_id);
-    }
-
-    for (i = 0; i < size; i++) {
-        if (supported_domains[i].id == domain_id) {
-            return true;
-        }
-    }
-
-    return false;
-}
-
-int get_domains_info(char * domain_type, int * num_domains, fastrpc_domain ** domains_info) {
-    int nErr    = AEE_SUCCESS;
-    int ss_info = 0;
-    if (domain_type != NULL) {
-        if (strcmp(domain_type, "LPASS") == 0) {
-            ss_info = FASTRPC_LPASS;
-        } else if (strcmp(domain_type, "HPASS") == 0) {
-            ss_info = FASTRPC_HPASS;
-        } else {
-            ss_info = FASTRPC_NSP;
-        }
-    }
-    system_req_payload req  = { 0 };
-    req.id                  = FASTRPC_GET_DOMAINS;
-    req.sys.domains         = NULL;
-    fastrpc_domain * domain = NULL;
-    if (ss_info != 0) {
-        req.sys.flags = DOMAINS_LIST_FLAGS_SET_TYPE(req.sys.flags, ss_info);
-    } else {
-        req.sys.flags = 0;
-    }
-#ifdef _WIN32
-    nErr = AEE_EUNSUPPORTED;
-    goto bail;
-#endif
-    if (remote_system_request) {
-        nErr = remote_system_request(&req);
-        if (nErr != AEE_SUCCESS) {
-            GGML_LOG_ERROR("Failure in remote_system_request call: %d.\n", nErr);
-            goto bail;
-        }
-        // Allocate memory for domain-info array
-        req.sys.max_domains = req.sys.num_domains;
-        if ((req.sys.domains = calloc(req.sys.num_domains, sizeof(fastrpc_domain))) == NULL) {
-            nErr = AEE_ENOMEMORY;
-            GGML_LOG_ERROR("Unable to allocate memory for req.sys.domains");
-            goto bail;
-        }
-
-        nErr = remote_system_request(&req);
-        if (nErr != AEE_SUCCESS) {
-            GGML_LOG_ERROR("Failure in remote_system_request call: %d.\n", nErr);
-            goto bail;
-        }
-
-        for (int i = 0; i < req.sys.num_domains; i++) {
-            // Verify that only requested type domains were returned
-            domain = &req.sys.domains[i];
-            if (domain->type != ss_info && domain_type != NULL) {
-                nErr = -1;
-                GGML_LOG_ERROR("Incorrect data received from remote_system_request.\n");
-                goto bail;
-            }
-        }
-        *domains_info = req.sys.domains;
-        *num_domains  = req.sys.num_domains;
-    } else {
-        nErr = AEE_EUNSUPPORTED;
-        goto bail;
-    }
-bail:
-    if (nErr && !req.sys.domains) {
-        free(req.sys.domains);
-    }
-    return nErr;
-}
-
-int get_effective_domain_id(char * domain_name, int session_id, int * effec_domain_id) {
-    int                              err  = 0;
-    remote_rpc_effective_domain_id_t sess = { 0 };
-
-    sess.domain_name     = domain_name;
-    sess.domain_name_len = strlen(domain_name);
-    sess.session_id      = session_id;
-
-    err = remote_session_control(FASTRPC_GET_EFFECTIVE_DOMAIN_ID, &sess, sizeof(sess));
-    if (err) {
-        GGML_LOG_ERROR("Error 0x%x: failed to get effective domain id for %s, session id %d\n", err, sess.domain_name,
-               session_id);
-        return err;
-    }
-
-    *effec_domain_id = sess.effective_domain_id;
-    return err;
-}
-
-int get_dsp_support(int * domain) {
-    int nErr = AEE_SUCCESS;
-    *domain  = CDSP_DOMAIN_ID;  // DSP domain default value is CDSP_DOMAIN_ID
-
-    if (remote_handle_control) {
-        struct remote_dsp_capability dsp_capability_domain = { CDSP_DOMAIN_ID, DOMAIN_SUPPORT, 0 };
-        nErr = remote_handle_control(DSPRPC_GET_DSP_INFO, &dsp_capability_domain, sizeof(struct remote_dsp_capability));
-        if ((nErr & 0xFF) == (AEE_EUNSUPPORTEDAPI & 0xFF)) {
-            GGML_LOG_ERROR("\nFastRPC Capability API is not supported on this device\n");
-            goto bail;
-        }
-
-        if (dsp_capability_domain.capability == 0) {
-            dsp_capability_domain.domain       = ADSP_DOMAIN_ID;  // Check for ADSP support.
-            dsp_capability_domain.attribute_ID = DOMAIN_SUPPORT;
-            dsp_capability_domain.capability   = 0;
-            nErr                               = remote_handle_control(DSPRPC_GET_DSP_INFO, &dsp_capability_domain,
-                                                                       sizeof(struct remote_dsp_capability));
-            if (dsp_capability_domain.capability) {
-                *domain = ADSP_DOMAIN_ID;  // For targets like Agatti (not having cDSP), domain is ADSP_DOMAIN_ID
-            }
-        }
-
-        if (nErr != AEE_SUCCESS) {
-            GGML_LOG_ERROR("\nget_dsp_support failed with Error 0x%x\n", nErr);
-            goto bail;
-        }
-    } else {
-        nErr = AEE_EUNSUPPORTEDAPI;
-        GGML_LOG_ERROR("remote_dsp_capability interface is not supported on this device\n");
-    }
-
-bail:
-    return nErr;
-}
-
-int get_vtcm_info(int domain, uint32_t * capability, uint32_t attr) {
-    int nErr    = AEE_SUCCESS;
-    *capability = 0;
-
-    if (attr == VTCM_PAGE || attr == VTCM_COUNT) {
-    } else {
-        nErr = AEE_EBADPARM;
-        GGML_LOG_ERROR("Unsupported attr. Only VTCM_PAGE and VTCM_COUNT supported\n");
-        goto bail;
-    }
-    if (remote_handle_control) {
-        if (domain == ADSP_DOMAIN_ID || domain == CDSP_DOMAIN_ID) {
-            /*
-            * Query the DSP for VTCM information
-            * Since the ADSP does not have a dedicated VTCM, we expect the output to be 0
-            */
-            struct remote_dsp_capability dsp_capability_vtcm_dsp;
-            dsp_capability_vtcm_dsp.domain       = (uint32_t) domain;
-            dsp_capability_vtcm_dsp.attribute_ID = attr;
-            dsp_capability_vtcm_dsp.capability   = (uint32_t) 0;
-            nErr                                 = remote_handle_control(DSPRPC_GET_DSP_INFO, &dsp_capability_vtcm_dsp,
-                                                                         sizeof(struct remote_dsp_capability));
-            if ((nErr & 0xFF) == (AEE_EUNSUPPORTEDAPI & 0xFF)) {
-                GGML_LOG_ERROR("\nFastRPC Capability API is not supported on this device\n");
-                GGML_LOG_ERROR("Running the usecase without checking the capability\n");
-                nErr = AEE_SUCCESS;
-                goto bail;
-            } else if (nErr == AEE_SUCCESS) {
-                *capability = dsp_capability_vtcm_dsp.capability;
-            } else {
-                GGML_LOG_ERROR("\nget_vtcm_info failed with Error 0x%x\n", nErr);
-                goto bail;
-            }
-        } else {
-            nErr = AEE_EUNSUPPORTED;
-            GGML_LOG_ERROR("Unsupported domain %d\n", domain);
-            goto bail;
-        }
-    } else {
-        nErr = AEE_EUNSUPPORTEDAPI;
-        GGML_LOG_ERROR("remote_dsp_capability interface is not supported on this device\n");
-    }
-
-bail:
-    return nErr;
-}
-
-bool is_unsignedpd_supported(int domain_id) {
-    int nErr = AEE_SUCCESS;
-    if (remote_handle_control) {
-        struct remote_dsp_capability dsp_capability_domain = { domain_id, UNSIGNED_PD_SUPPORT, 0 };
-        nErr = remote_handle_control(DSPRPC_GET_DSP_INFO, &dsp_capability_domain, sizeof(struct remote_dsp_capability));
-        if ((nErr & 0xFF) == (AEE_EUNSUPPORTEDAPI & 0xFF)) {
-            GGML_LOG_ERROR("\nFastRPC Capability API is not supported on this device. Falling back to signed pd.\n");
-            return false;
-        }
-        if (nErr) {
-            GGML_LOG_ERROR("\nERROR 0x%x: FastRPC Capability API failed. Falling back to signed pd.", nErr);
-            return false;
-        }
-        if (dsp_capability_domain.capability == 1) {
-            return true;
-        }
-    } else {
-        nErr = AEE_EUNSUPPORTEDAPI;
-        GGML_LOG_ERROR("remote_dsp_capability interface is not supported on this device. Falling back to signed pd.\n");
-        return false;
-    }
-    return false;
-}
-
-bool get_unsignedpd_support(void) {
-    return is_unsignedpd_supported(CDSP_DOMAIN_ID);
-}
-
-bool is_async_fastrpc_supported(int domain) {
-    int nErr = AEE_SUCCESS;
-    if (remote_handle_control) {
-        if (domain == CDSP_DOMAIN_ID) {
-            /*
-            * Query the DSP for ASYNC_FASTRPC_SUPPORT information
-            * Async fastrpc is supported only on CDSP
-            */
-            struct remote_dsp_capability dsp_capability_async_support;
-            dsp_capability_async_support.domain       = (uint32_t) domain;
-            dsp_capability_async_support.attribute_ID = ASYNC_FASTRPC_SUPPORT;
-            dsp_capability_async_support.capability   = (uint32_t) 0;
-            nErr = remote_handle_control(DSPRPC_GET_DSP_INFO, &dsp_capability_async_support,
-                                         sizeof(struct remote_dsp_capability));
-            if ((nErr & 0xFF) == (AEE_EUNSUPPORTEDAPI & 0xFF)) {
-                GGML_LOG_ERROR("\nFastRPC Capability API is not supported on this device\n");
-                GGML_LOG_ERROR("Running the usecase without checking the capability\n");
-                nErr = AEE_SUCCESS;
-                goto bail;
-            } else if (dsp_capability_async_support.capability == 1) {
-                return true;
-            }
-            if (nErr != AEE_SUCCESS) {
-                GGML_LOG_ERROR("\nis_async_fastrpc_supported failed with Error 0x%x\n", nErr);
-                goto bail;
-            }
-        } else {
-            nErr = AEE_EUNSUPPORTED;
-            GGML_LOG_ERROR("Async fastrpc is not supported on domain %d\n", domain);
-            goto bail;
-        }
-    } else {
-        nErr = AEE_EUNSUPPORTEDAPI;
-        GGML_LOG_ERROR("remote_dsp_capability interface is not supported on this device\n");
-    }
-
-bail:
-    return false;
-}
-
-bool is_status_notification_supported(int domain) {
-    int nErr = AEE_SUCCESS;
-
-    if (remote_handle_control) {
-        /*
-        * Query the DSP for STATUS_NOTIFICATION_SUPPORT information
-        * DSP User PD status notification Support
-        */
-        struct remote_dsp_capability dsp_capability_status_notification_support;
-        dsp_capability_status_notification_support.domain       = (uint32_t) domain;
-        dsp_capability_status_notification_support.attribute_ID = STATUS_NOTIFICATION_SUPPORT;
-        dsp_capability_status_notification_support.capability   = (uint32_t) 0;
-        nErr = remote_handle_control(DSPRPC_GET_DSP_INFO, &dsp_capability_status_notification_support,
-                                     sizeof(struct remote_dsp_capability));
-        if ((nErr & 0xFF) == (AEE_EUNSUPPORTEDAPI & 0xFF)) {
-            GGML_LOG_ERROR("\nFastRPC Capability API is not supported on this device\n");
-            GGML_LOG_ERROR("Running the usecase without checking the capability\n");
-            nErr = AEE_SUCCESS;
-            goto bail;
-        } else if (dsp_capability_status_notification_support.capability == 1) {
-            return true;
-        }
-        if (nErr != AEE_SUCCESS) {
-            GGML_LOG_ERROR("\nis_status_notification_supported failed with Error 0x%x\n", nErr);
-            goto bail;
-        }
-    } else {
-        nErr = AEE_EUNSUPPORTEDAPI;
-        GGML_LOG_ERROR("remote_dsp_capability interface is not supported on this device\n");
-    }
-
-bail:
-    return false;
-}
-
-int get_hmx_support_info(int domain, uint32_t * capability, uint32_t attr) {
-    int nErr    = AEE_SUCCESS;
-    *capability = 0;
-
-    if (attr != HMX_SUPPORT_SPATIAL && attr != HMX_SUPPORT_DEPTH) {
-        nErr = AEE_EBADPARM;
-        GGML_LOG_ERROR("Unsupported attr. Only HMX_SUPPORT_SPATIAL and HMX_SUPPORT_DEPTH supported\n");
-        goto bail;
-    }
-    if (remote_handle_control) {
-        if (domain == CDSP_DOMAIN_ID) {
-            /*
-            * Query the DSP for HMX SUPPORT information
-            * HMX is supported on CDSP only
-            */
-            struct remote_dsp_capability dsp_capability_hmx_dsp;
-            dsp_capability_hmx_dsp.domain       = (uint32_t) domain;
-            dsp_capability_hmx_dsp.attribute_ID = attr;
-            dsp_capability_hmx_dsp.capability   = (uint32_t) 0;
-            nErr                                = remote_handle_control(DSPRPC_GET_DSP_INFO, &dsp_capability_hmx_dsp,
-                                                                        sizeof(struct remote_dsp_capability));
-            if ((nErr & 0xFF) == (AEE_EUNSUPPORTEDAPI & 0xFF)) {
-                GGML_LOG_ERROR("\nFastRPC Capability API is not supported on this device\n");
-                GGML_LOG_ERROR("Running the usecase without checking the capability\n");
-                nErr = AEE_SUCCESS;
-                goto bail;
-            } else if (nErr == AEE_SUCCESS) {
-                *capability = dsp_capability_hmx_dsp.capability;
-            } else {
-                GGML_LOG_ERROR("\nget_hmx_support_info failed with Error 0x%x\n", nErr);
-                goto bail;
-            }
-        } else {
-            nErr = AEE_EUNSUPPORTED;
-            GGML_LOG_ERROR("HMX support is not there for domain %d\n", domain);
-            goto bail;
-        }
-    } else {
-        nErr = AEE_EUNSUPPORTEDAPI;
-        GGML_LOG_ERROR("remote_dsp_capability interface is not supported on this device\n");
-    }
-
-bail:
-    return nErr;
-}
-
-int get_hex_arch_ver(int domain, int * arch) {
-    if (!remote_handle_control) {
-        GGML_LOG_ERROR("ggml-hex: remote_handle_control is not supported on this device\n");
-        return AEE_EUNSUPPORTEDAPI;
-    }
-
-    struct remote_dsp_capability arch_ver;
-    arch_ver.domain       = (uint32_t) domain;
-    arch_ver.attribute_ID = ARCH_VER;
-    arch_ver.capability   = (uint32_t) 0;
-
-    int err = remote_handle_control(DSPRPC_GET_DSP_INFO, &arch_ver, sizeof(arch_ver));
-    if ((err & 0xff) == (AEE_EUNSUPPORTEDAPI & 0xff)) {
-        GGML_LOG_ERROR("ggml-hex: FastRPC capability API is not supported on this device\n");
-        return AEE_EUNSUPPORTEDAPI;
-    }
-
-    if (err != AEE_SUCCESS) {
-        GGML_LOG_ERROR("ggml-hex: FastRPC capability query failed (err %d)\n", err);
-        return err;
-    }
-
-    switch (arch_ver.capability & 0xff) {
-        case 0x68:
-            *arch = 68;
-            return 0;
-        case 0x69:
-            *arch = 69;
-            return 0;
-        case 0x73:
-            *arch = 73;
-            return 0;
-        case 0x75:
-            *arch = 75;
-            return 0;
-        case 0x79:
-            *arch = 79;
-            return 0;
-        case 0x81:
-            *arch = 81;
-            return 0;
-    }
-    return -1;
-}
-
-int get_hvx_support_info(int domain, uint32_t * capability, uint32_t attr) {
-    int nErr    = AEE_SUCCESS;
-    *capability = 0;
-
-    if (remote_handle_control) {
-        if (domain == CDSP_DOMAIN_ID) {
-            /*
-            * Query the DSP for HVX SUPPORT information
-            * HVX is supported on CDSP only
-            */
-            struct remote_dsp_capability dsp_capability_hvx_dsp;
-            dsp_capability_hvx_dsp.domain       = (uint32_t) domain;
-            dsp_capability_hvx_dsp.attribute_ID = attr;
-            dsp_capability_hvx_dsp.capability   = (uint32_t) 0;
-            nErr                                = remote_handle_control(DSPRPC_GET_DSP_INFO, &dsp_capability_hvx_dsp,
-                                                                        sizeof(struct remote_dsp_capability));
-            if ((nErr & 0xFF) == (AEE_EUNSUPPORTEDAPI & 0xFF)) {
-                GGML_LOG_ERROR("\nFastRPC Capability API is not supported on this device\n");
-                GGML_LOG_ERROR("Running the usecase without checking the capability\n");
-                nErr = AEE_SUCCESS;
-                goto bail;
-            } else if (nErr == AEE_SUCCESS) {
-                *capability = dsp_capability_hvx_dsp.capability;
-            } else {
-                GGML_LOG_ERROR("\nget_hvx_support_info failed with Error 0x%x\n", nErr);
-                goto bail;
-            }
-        } else {
-            nErr = AEE_EUNSUPPORTED;
-            GGML_LOG_ERROR("HVX support is not available on domain %d\n", domain);
-            goto bail;
-        }
-    } else {
-        nErr = AEE_EUNSUPPORTEDAPI;
-        GGML_LOG_ERROR("remote_dsp_capability interface is not supported on this device\n");
-    }
-
-bail:
-    return nErr;
-}
diff --git a/ggml/src/ggml-hexagon/htp-utils.h b/ggml/src/ggml-hexagon/htp-utils.h
deleted file mode 100644
index 7bbae3a0b7..0000000000
--- a/ggml/src/ggml-hexagon/htp-utils.h
+++ /dev/null
@@ -1,221 +0,0 @@
-#ifndef HTP_UTILS_H
-#define HTP_UTILS_H
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#include <AEEStdErr.h>
-#include <inttypes.h>
-#include <remote.h>
-#include <rpcmem.h>
-#include <stdbool.h>
-
-/* Offset to differentiate HLOS and Hexagon error codes.
-   Stores the value of AEE_EOFFSET for Hexagon. */
-#ifndef DSP_OFFSET
-#    define DSP_OFFSET 0x80000400
-#endif
-
-/* Errno for connection reset by peer. */
-#ifndef ECONNRESET
-#    ifdef __hexagon__
-#        define ECONNRESET 104
-#    endif
-#endif
-
-/* Abstraction of different OS specific sleep APIs.
-   SLEEP accepts input in seconds. */
-#ifndef SLEEP
-#    ifdef __hexagon__
-#        define SLEEP(x)                      \
-            { /* Do nothing for simulator. */ \
-            }
-#    else
-#        ifdef _WINDOWS
-#            define SLEEP(x) Sleep(1000 * x) /* Sleep accepts input in milliseconds. */
-#        else
-#            define SLEEP(x) sleep(x)        /* sleep accepts input in seconds. */
-#        endif
-#    endif
-#endif
-
-/* Include windows specific header files. */
-#ifdef _WINDOWS
-#    include <sysinfoapi.h>
-#    include <windows.h>
-#    define _CRT_SECURE_NO_WARNINGS         1
-#    define _WINSOCK_DEPRECATED_NO_WARNINGS 1
-/* Including this file for custom implementation of getopt function. */
-#    include "getopt_custom.h"
-#endif
-
-/* Includes and defines for all HLOS except windows */
-#if !defined(__hexagon__) && !defined(_WINDOWS)
-#    include "unistd.h"
-
-#    include <sys/time.h>
-#endif
-
-/* Includes and defines for Hexagon and all HLOS except Windows. */
-#if !defined(_WINDOWS)
-/* Weak reference to remote symbol for compilation. */
-#    pragma weak remote_session_control
-#    pragma weak remote_handle_control
-#    pragma weak remote_handle64_control
-#    pragma weak fastrpc_mmap
-#    pragma weak fastrpc_munmap
-#    pragma weak rpcmem_alloc2
-#endif
-
-#if !defined(_WINDOWS)
-#    pragma weak remote_system_request
-#endif
-/**
- * Wrapper for FastRPC Capability API: query DSP support.
- *
- * @param[out]  domain pointer to supported domain.
- * @return      0          if query is successful.
- *              non-zero   if error, return value points to the error.
- */
-int get_dsp_support(int * domain);
-
-/**
- * Wrapper for FastRPC Capability API: query VTCM information.
- *
- * @param[in]   domain value of domain in the queried.
- * @param[out]  capability capability value of the attribute queried.
- * @param[in]   attr value of the attribute to the queried.
- * @return      0          if query is successful.
- *              non-zero   if error, return value points to the error.
- */
-int get_vtcm_info(int domain, uint32_t * capability, uint32_t attr);
-
-/**
- * Wrapper for FastRPC Capability API: query unsigned pd support on CDSP domain.
- *
- * @return      true          if unsigned pd is supported.
- *              false         if unsigned pd is not supported, capability query failed.
- */
-
-bool get_unsignedpd_support(void);
-
-/**
- * Wrapper for FastRPC Capability API: query unsigned pd support.
- *
- * @param[in]   domain value of domain in the queried.
- * @return      true          if unsigned pd is supported.
- *              false         if unsigned pd is not supported, capability query failed.
- */
-
-bool is_unsignedpd_supported(int domain_id);
-
-/**
- * is_valid_domain_id API: query a domain id is valid.
- *
- * @param[in]   domain value of domain in the queried.
- * @param[in]   compute_only value of domain is only compared with CDSP domains supported by the target when enabled.
- * @return      true          if value of domain is valid.
- *              false         if value of domain is not valid.
- */
-
-bool is_valid_domain_id(int domain_id, int compute_only);
-
-/**
- * get_domain API: get domain struct from domain value.
- *
- * @param[in]  domain value of a domain
- * @return     Returns domain struct of the domain if it is supported or else
- *             returns NULL.
- *
- */
-
-domain * get_domain(int domain_id);
-
-/**
- * get_domains_info API: get information for all the domains available on the device
- *
- * @param[in]  domain_type pointer to domain type
- * @param[in]  num_domains pointer to number of domains
- * @param[in]  domains_info pointer to save discovered domains information.
- * @return     0 if query is successful.
- *              non-zero if error, return value points to the error.
- *
- * It is user's responsibility to free the memory used to store the domains info whose address is present in domains_info before closing the application.
- *
- */
-
-int get_domains_info(char * domain_type, int * num_domains, fastrpc_domain ** domains_info);
-
-/**
- * get_effective_domain_id API: get effective domain id for given session id
- *
- * @param[in]  domain_name pointer to domain name
- * @param[in]  session_id
- * @param[in]  effec_domain_id pointer to save obtained effective domain id.
- * @return     0 if query is successful.
- *              non-zero if error, return value points to the error.
- *
- */
-
-int get_effective_domain_id(char * domain_name, int session_id, int * effec_domain_id);
-
-/**
- * is_async_fastrpc_supported API: query a domain id has async fastrpc supported or not
- *
- * @param[in]  domain_id value of a domain
- * @return     Returns true or false stating support of Async FastRPC
- *
- */
-
-bool is_async_fastrpc_supported(int domain_id);
-
-/**
- * is_status_notification_supported API: query the DSP for STATUS_NOTIFICATION_SUPPORT information
- *
- * @param[in]  domain_id value of a domain
- * @return     Returns true or false stating status notification support information
- *
- */
-bool is_status_notification_supported(int domain_id);
-
-/**
- * get_hmx_support_info API: query the DSP for HMX SUPPORT information
- *
- * @param[in]   domain_id value of a domain
- * @param[out]  capability capability value of the attribute queried.
- * @param[in]   attr value of the attribute to the queried.
- * @return      0 if query is successful.
- *              non-zero if error, return value points to the error.
- *
- */
-int get_hmx_support_info(int domain, uint32_t * capability, uint32_t attr);
-
-/**
- * get_hex_arch_ver API: query the Hexagon processor architecture version information
- *
- * @param[in]   domain_id value of a domain
- * @param[out]  Arch version (73, 75, ...)
- * @return      0 if query is successful.
- *              non-zero if error, return value points to the error.
- *
- */
-int get_hex_arch_ver(int domain, int * arch);
-
-/**
- * get_hvx_support_info API: query the DSP for HVX SUPPORT information
- *
- * @param[in]   domain_id value of a domain
- * @param[out]  capability capability value of the attribute queried.
- * @param[in]   attr value of the attribute to the queried.
- * @return      0 if query is successful.
- *              non-zero if error, return value points to the error.
- *
- */
-int get_hvx_support_info(int domain, uint32_t * capability, uint32_t attr);
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif  //DSP_CAPABILITIES_UTILS_H
diff --git a/ggml/src/ggml-hexagon/htp/flash-attn-ops.c b/ggml/src/ggml-hexagon/htp/flash-attn-ops.c
index 1de47d0f3d..c184637443 100644
--- a/ggml/src/ggml-hexagon/htp/flash-attn-ops.c
+++ b/ggml/src/ggml-hexagon/htp/flash-attn-ops.c
@@ -2,9 +2,9 @@
 #pragma clang diagnostic ignored "-Wunused-function"
 #pragma clang diagnostic ignored "-Wunused-but-set-variable"
 
+#include <assert.h>
 #include <HAP_farf.h>
 #include <HAP_perf.h>
-
 #include <math.h>
 #include <string.h>
 
@@ -17,6 +17,12 @@
 #include "htp-msg.h"
 #include "htp-ops.h"
 
+static inline HVX_Vector hvx_load_f32_to_f16(const HVX_Vector * restrict src, const HVX_Vector zero) {
+    HVX_Vector y0_qf = Q6_Vqf32_vsub_VsfVsf(src[0], zero);  // 32 elements
+    HVX_Vector y1_qf = Q6_Vqf32_vsub_VsfVsf(src[1], zero);  // 32 elements
+    return Q6_Vh_vdeal_Vh(Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(y1_qf, y0_qf)));
+}
+
 // Dot product of FP32 and FP16 vectors, accumulating to float
 static inline void hvx_dot_f32_f16_aa(float * restrict r, const void * restrict y, const void * restrict x, unsigned int n, float s) {
     const HVX_Vector * restrict vy = (const HVX_Vector * restrict) y; // fp32
@@ -33,23 +39,19 @@ static inline void hvx_dot_f32_f16_aa(float * restrict r, const void * restrict
     #pragma unroll(4)
     for (i = 0; i < nvec; i++) {
         // Load y (fp32) and convert into fp16
-        HVX_Vector y0_qf = Q6_Vqf32_vsub_VsfVsf(vy[i*2+0], zero);  // 32 elements
-        HVX_Vector y1_qf = Q6_Vqf32_vsub_VsfVsf(vy[i*2+1], zero);  // 32 elements
-        HVX_Vector y_hf  = Q6_Vh_vdeal_Vh(Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(y1_qf, y0_qf)));
+        HVX_Vector y_hf  = hvx_load_f32_to_f16(&vy[i*2], zero);
 
         // Load x (fp16)
         HVX_Vector x_hf  = vx[i];
 
         HVX_VectorPair xy_qf = Q6_Wqf32_vmpy_VhfVhf(x_hf, y_hf);
 
-        rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf), Q6_V_hi_W(xy_qf)));
+        rsum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf), Q6_V_hi_W(xy_qf)), rsum));
     }
 
     if (nloe) {
         // Load y (fp32) and convert into fp16
-        HVX_Vector y0_qf = Q6_Vqf32_vsub_VsfVsf(vy[i*2+0], zero);  // 32 elements
-        HVX_Vector y1_qf = Q6_Vqf32_vsub_VsfVsf(vy[i*2+1], zero);  // 32 elements
-        HVX_Vector y_hf  = Q6_Vh_vdeal_Vh(Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(y1_qf, y0_qf)));
+        HVX_Vector y_hf  = hvx_load_f32_to_f16(&vy[i*2], zero);
 
         // Load x (fp16)
         HVX_Vector x_hf  = vx[i];
@@ -62,13 +64,72 @@ static inline void hvx_dot_f32_f16_aa(float * restrict r, const void * restrict
 
         HVX_VectorPair xy_qf = Q6_Wqf32_vmpy_VhfVhf(x_hf, y_hf);
 
-        rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf), Q6_V_hi_W(xy_qf)));
+        rsum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf), Q6_V_hi_W(xy_qf)), rsum));
     }
 
-    rsum = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(rsum), hvx_vec_splat_f32(s));
-    rsum = Q6_Vsf_equals_Vqf32(hvx_vec_reduce_sum_qf32(rsum));
+    rsum = Q6_Vqf32_vmpy_VsfVsf(hvx_vec_splat_f32(s), hvx_vec_reduce_sum_f32(rsum));
+    hvx_vec_store_u(r, 4, Q6_Vsf_equals_Vqf32(rsum));
+}
 
-    hvx_vec_store_u(r, 4, rsum);
+// Dot product of FP32 and FP16 vectors, accumulating to float
+static inline void hvx_dot_f32_f16_aa_rx2(float * restrict r,
+                                          const void * restrict y,
+                                          const void * restrict x0,
+                                          const void * restrict x1,
+                                          unsigned int n,
+                                          float        s) {
+    const HVX_Vector * restrict vy  = (const HVX_Vector * restrict) y;   // fp32
+    const HVX_Vector * restrict vx0 = (const HVX_Vector * restrict) x0;  // fp16
+    const HVX_Vector * restrict vx1 = (const HVX_Vector * restrict) x1;  // fp16
+
+    uint32_t nvec = n / VLEN_FP16;                                       // num full fp16 hvx vectors
+    uint32_t nloe = n % VLEN_FP16;                                       // leftover elements
+
+    const HVX_Vector zero  = Q6_V_vsplat_R(0);
+    HVX_Vector       rsum0 = Q6_V_vsplat_R(0);
+    HVX_Vector       rsum1 = Q6_V_vsplat_R(0);
+
+    uint32_t i = 0;
+
+    #pragma unroll(2)
+    for (i = 0; i < nvec; i++) {
+        // Load y (fp32) and convert into fp16
+        HVX_Vector y_hf  = hvx_load_f32_to_f16(&vy[i*2], zero);
+        // Load x (fp16)
+        HVX_Vector x0_hf = vx0[i];
+        HVX_Vector x1_hf = vx1[i];
+
+        HVX_VectorPair xy0_qf = Q6_Wqf32_vmpy_VhfVhf(x0_hf, y_hf);
+        HVX_VectorPair xy1_qf = Q6_Wqf32_vmpy_VhfVhf(x1_hf, y_hf);
+
+        rsum0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy0_qf), Q6_V_hi_W(xy0_qf)), rsum0));
+        rsum1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy1_qf), Q6_V_hi_W(xy1_qf)), rsum1));
+    }
+
+    if (nloe) {
+        // Load y (fp32) and convert into fp16
+        HVX_Vector y_hf  = hvx_load_f32_to_f16(&vy[i*2], zero);
+
+        // Load x (fp16)
+        HVX_Vector x0_hf = vx0[i];
+        HVX_Vector x1_hf = vx1[i];
+
+        // Zero-out unused elements
+        // Note that we need to clear both x and y because they may contain NANs
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 2);
+        x0_hf                = Q6_V_vand_QV(bmask, x0_hf);
+        x1_hf                = Q6_V_vand_QV(bmask, x1_hf);
+        y_hf                 = Q6_V_vand_QV(bmask, y_hf);
+
+        HVX_VectorPair xy0_qf = Q6_Wqf32_vmpy_VhfVhf(x0_hf, y_hf);
+        HVX_VectorPair xy1_qf = Q6_Wqf32_vmpy_VhfVhf(x1_hf, y_hf);
+
+        rsum0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy0_qf), Q6_V_hi_W(xy0_qf)), rsum0));
+        rsum1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy1_qf), Q6_V_hi_W(xy1_qf)), rsum1));
+    }
+
+    HVX_Vector rsum = Q6_Vqf32_vmpy_VsfVsf(hvx_vec_splat_f32(s), hvx_vec_reduce_sum_f32x2(rsum0, rsum1));
+    hvx_vec_store_u(r, 8, Q6_Vsf_equals_Vqf32(rsum));
 }
 
 // Dot product of two F16 vectors, accumulating to float
@@ -91,7 +152,7 @@ static inline void hvx_dot_f16_f16_aa(float * restrict r, const void * restrict
 
         HVX_VectorPair xy_qf = Q6_Wqf32_vmpy_VhfVhf(x_hf, y_hf);
 
-        rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf),  Q6_V_hi_W(xy_qf)));
+        rsum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf), Q6_V_hi_W(xy_qf)), rsum));
     }
 
     if (nloe) {
@@ -103,15 +164,65 @@ static inline void hvx_dot_f16_f16_aa(float * restrict r, const void * restrict
 
         HVX_VectorPair xy_qf = Q6_Wqf32_vmpy_VhfVhf(x_hf, y_hf);
 
-        rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf),  Q6_V_hi_W(xy_qf)));
+        rsum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf), Q6_V_hi_W(xy_qf)), rsum));
     }
 
-    rsum = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(rsum), hvx_vec_splat_f32(s));
-    rsum = Q6_Vsf_equals_Vqf32(hvx_vec_reduce_sum_qf32(rsum));
-    hvx_vec_store_u(r, 4, rsum);
+    rsum = Q6_Vqf32_vmpy_VsfVsf(hvx_vec_splat_f32(s), hvx_vec_reduce_sum_f32(rsum));
+    hvx_vec_store_u(r, 4, Q6_Vsf_equals_Vqf32(rsum));
 }
 
-// MAD: y (F32) += x (F16) * v (float)
+static inline void hvx_dot_f16_f16_aa_rx2(float * restrict r,
+                                          const void * restrict y,
+                                          const void * restrict x0,
+                                          const void * restrict x1,
+                                          unsigned int n,
+                                          float        s) {
+    const HVX_Vector * restrict vx0 = (const HVX_Vector * restrict) x0;  // fp16
+    const HVX_Vector * restrict vx1 = (const HVX_Vector * restrict) x1;  // fp16
+    const HVX_Vector * restrict vy  = (const HVX_Vector * restrict) y;   // fp16
+
+    uint32_t nvec = n / VLEN_FP16;                                       // num full fp16 hvx vectors
+    uint32_t nloe = n % VLEN_FP16;                                       // leftover elements
+
+    const HVX_Vector zero  = Q6_V_vsplat_R(0);
+    HVX_Vector       rsum0 = Q6_V_vsplat_R(0);
+    HVX_Vector       rsum1 = Q6_V_vsplat_R(0);
+
+    uint32_t i = 0;
+
+    #pragma unroll(4)
+    for (i = 0; i < nvec; i++) {
+        HVX_Vector y_hf  = vy[i];
+        HVX_Vector x0_hf = vx0[i];
+        HVX_Vector x1_hf = vx1[i];
+
+        HVX_VectorPair xy0_qf = Q6_Wqf32_vmpy_VhfVhf(x0_hf, y_hf);
+        HVX_VectorPair xy1_qf = Q6_Wqf32_vmpy_VhfVhf(x1_hf, y_hf);
+
+        rsum0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy0_qf), Q6_V_hi_W(xy0_qf)), rsum0));
+        rsum1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy1_qf), Q6_V_hi_W(xy1_qf)), rsum1));
+    }
+
+    if (nloe) {
+        HVX_Vector y_hf = vy[i];
+
+        // Load x (fp16) and zero-out unused elements
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 2);
+        HVX_Vector     x0_hf = Q6_V_vand_QV(bmask, vx0[i]);
+        HVX_Vector     x1_hf = Q6_V_vand_QV(bmask, vx1[i]);
+
+        HVX_VectorPair xy0_qf = Q6_Wqf32_vmpy_VhfVhf(x0_hf, y_hf);
+        HVX_VectorPair xy1_qf = Q6_Wqf32_vmpy_VhfVhf(x1_hf, y_hf);
+
+        rsum0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy0_qf), Q6_V_hi_W(xy0_qf)), rsum0));
+        rsum1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy1_qf), Q6_V_hi_W(xy1_qf)), rsum1));
+    }
+
+    HVX_Vector rsum = Q6_Vqf32_vmpy_VsfVsf(hvx_vec_splat_f32(s), hvx_vec_reduce_sum_f32x2(rsum0, rsum1));
+    hvx_vec_store_u(r, 8, Q6_Vsf_equals_Vqf32(rsum));
+}
+
+// MAD: y (F32) += x (F16) * s (float)
 static inline void hvx_mad_f32_f16_aa(float * restrict y, const void * restrict x, int n, float s) {
     const HVX_Vector * restrict ptr_x = (const HVX_Vector *) x;
     HVX_Vector * restrict ptr_y = (HVX_Vector *) y;
@@ -317,17 +428,22 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
             // Inner loop processing the block from VTCM
             uint32_t ic = 0;
 
+            const bool is_q_fp32 = (q->type == HTP_TYPE_F32);
+
             // Process in blocks of 32 (VLEN_FP32)
-            for (; ic + VLEN_FP32 <= current_block_size; ic += VLEN_FP32) {
+            static_assert(FLASH_ATTN_BLOCK_SIZE / VLEN_FP32 <= 4, "FLASH_ATTN_BLOCK_SIZE changed, fix HVX_Vector_x4 usage");
+            HVX_Vector_x4 scores_x4;
+            HVX_Vector v_max = hvx_vec_splat_f32(-INFINITY);
+            for (uint32_t iv = 0; ic + VLEN_FP32 <= current_block_size; ic += VLEN_FP32, ++iv) {
                 // 1. Compute scores
                 float __attribute__((aligned(VLEN))) scores_arr[VLEN_FP32];
-                for (int j = 0; j < VLEN_FP32; ++j) {
+                for (int j = 0; j < VLEN_FP32; j += 2) {
                     const uint32_t cur_ic = ic + j;
                     const uint8_t * k_ptr = k_base + cur_ic * size_k_row_padded;
-                    if (q->type == HTP_TYPE_F32) {
-                        hvx_dot_f32_f16_aa(&scores_arr[j], q_ptr_vtcm, k_ptr, DK, scale);
+                    if (is_q_fp32) {
+                        hvx_dot_f32_f16_aa_rx2(&scores_arr[j], q_ptr_vtcm, k_ptr, k_ptr + size_k_row_padded, DK, scale);
                     } else {
-                        hvx_dot_f16_f16_aa(&scores_arr[j], q_ptr_vtcm, k_ptr, DK, scale);
+                        hvx_dot_f16_f16_aa_rx2(&scores_arr[j], q_ptr_vtcm, k_ptr, k_ptr + size_k_row_padded, DK, scale);
                     }
                 }
 
@@ -356,36 +472,43 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
                     scores = Q6_Vsf_equals_Vqf32(scores);
                 }
 
-                // 4. Online Softmax Update
-                HVX_Vector v_max = hvx_vec_reduce_max_f32(scores);
-                float m_block = hvx_vec_get_f32(v_max);
+                scores_x4.v[iv] = scores;
+                v_max = Q6_Vsf_vmax_VsfVsf(scores, v_max);
+            }
 
+            {
+                // 4. Online Softmax Update
+                v_max = hvx_vec_reduce_max_f32(v_max);
+                float m_block = hvx_vec_get_f32(v_max);
                 float M_old = M;
                 float M_new = (m_block > M) ? m_block : M;
                 M = M_new;
 
-                float ms = expf(M_old - M_new);
-
+                const float ms = expf(M_old - M_new);
                 hvx_scale_f32_aa((uint8_t *) VKQ32, (const uint8_t *) VKQ32, DV, ms);
-                S = S * ms;
 
                 HVX_Vector M_new_vec = hvx_vec_splat_f32(M_new);
-                HVX_Vector scores_shifted = Q6_Vqf32_vsub_VsfVsf(scores, M_new_vec);
-                HVX_Vector P = hvx_vec_exp_f32(Q6_Vsf_equals_Vqf32(scores_shifted));
+                HVX_Vector p_sum_vec = hvx_vec_splat_f32(0.0f);
+                for (uint32_t ic2 = 0, iv = 0; ic2 + VLEN_FP32 <= current_block_size; ic2 += VLEN_FP32, ++iv) {
+                    HVX_Vector scores = scores_x4.v[iv];
+                    HVX_Vector scores_shifted = Q6_Vqf32_vsub_VsfVsf(scores, M_new_vec);
+                    HVX_Vector P = hvx_vec_exp_f32(Q6_Vsf_equals_Vqf32(scores_shifted));
 
-                HVX_Vector p_sum_vec = hvx_vec_reduce_sum_f32(P);
-                float p_sum = hvx_vec_get_f32(p_sum_vec);
-                S += p_sum;
+                    p_sum_vec = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(p_sum_vec, P));
 
-                // 5. Accumulate V
-                float __attribute__((aligned(VLEN))) p_arr[VLEN_FP32];
-                *(HVX_Vector*)p_arr = P;
+                    // 5. Accumulate V
+                    float __attribute__((aligned(VLEN))) p_arr[VLEN_FP32];
+                    *(HVX_Vector*)p_arr = P;
 
-                for (int j = 0; j < VLEN_FP32; ++j) {
-                    const uint32_t cur_ic = ic + j;
-                    const uint8_t * v_ptr = v_base + cur_ic * size_v_row_padded;
-                    hvx_mad_f32_f16_aa(VKQ32, v_ptr, DV, p_arr[j]);
+                    for (int j = 0; j < VLEN_FP32; ++j) {
+                        const uint32_t cur_ic = ic2 + j;
+                        const uint8_t * v_ptr = v_base + cur_ic * size_v_row_padded;
+                        hvx_mad_f32_f16_aa(VKQ32, v_ptr, DV, p_arr[j]);
+                    }
                 }
+
+                p_sum_vec = hvx_vec_reduce_sum_f32(p_sum_vec);
+                S = S * ms + hvx_vec_get_f32(p_sum_vec);
             }
 
             // Leftover
@@ -393,7 +516,7 @@ static void flash_attn_ext_f16_thread(struct htp_ops_context * octx, int ith, in
                 float s_val;
                 const uint8_t * k_ptr = k_base + ic * size_k_row_padded;
 
-                if (q->type == HTP_TYPE_F32) {
+                if (is_q_fp32) {
                     hvx_dot_f32_f16_aa(&s_val, q_ptr_vtcm, k_ptr, DK, scale);
                 } else {
                     hvx_dot_f16_f16_aa(&s_val, q_ptr_vtcm, k_ptr, DK, scale);
diff --git a/ggml/src/ggml-hexagon/htp/hvx-dump.h b/ggml/src/ggml-hexagon/htp/hvx-dump.h
index e882227893..85201fc345 100644
--- a/ggml/src/ggml-hexagon/htp/hvx-dump.h
+++ b/ggml/src/ggml-hexagon/htp/hvx-dump.h
@@ -28,19 +28,16 @@ static void hvx_vec_dump_f16(char * pref, HVX_Vector v) {
 }
 
 static void hvx_vec_dump_f32_n(char * pref, HVX_Vector v, uint32_t n) {
-    union {
-        HVX_Vector v;
-        float      d[32];
-    } u = { .v = v };
+    HVX_VectorAlias u = { .v = v };
 
     const uint32_t n0 = n / 16;
     const uint32_t n1 = n % 16;
     int            i  = 0;
     for (; i < n0; i++) {
-        hex_dump_f32_line(pref, u.d + (16 * i), 16);
+        hex_dump_f32_line(pref, u.fp32 + (16 * i), 16);
     }
     if (n1) {
-        hex_dump_f32_line(pref, u.d + (16 * i), n1);
+        hex_dump_f32_line(pref, u.fp32 + (16 * i), n1);
     }
 }
 
diff --git a/ggml/src/ggml-hexagon/htp/hvx-reduce.h b/ggml/src/ggml-hexagon/htp/hvx-reduce.h
index 8845fe73ea..1ca7c05d98 100644
--- a/ggml/src/ggml-hexagon/htp/hvx-reduce.h
+++ b/ggml/src/ggml-hexagon/htp/hvx-reduce.h
@@ -44,6 +44,45 @@ static inline HVX_Vector hvx_vec_reduce_sum_qf32(HVX_Vector in) {
     return hvx_vec_reduce_sum_n_qf32(in, 32);
 }
 
+#if __HVX_ARCH__ > 75
+
+static inline HVX_Vector hvx_vec_reduce_sum_f32x2(HVX_Vector in0, HVX_Vector in1) {
+    HVX_VectorPair sump = Q6_W_vshuff_VVR(in1, in0, 4);
+    HVX_Vector  sum_sf  = Q6_Vsf_vadd_VsfVsf(Q6_V_lo_W(sump), Q6_V_hi_W(sump));
+
+    sum_sf = Q6_Vsf_vadd_VsfVsf(sum_sf, Q6_V_vror_VR(sum_sf, VLEN / 2));
+    sum_sf = Q6_Vsf_vadd_VsfVsf(sum_sf, Q6_V_vror_VR(sum_sf, VLEN / 4));
+    sum_sf = Q6_Vsf_vadd_VsfVsf(sum_sf, Q6_V_vror_VR(sum_sf, VLEN / 8));
+    sum_sf = Q6_Vsf_vadd_VsfVsf(sum_sf, Q6_V_vror_VR(sum_sf, VLEN / 16));
+    return sum_sf;
+}
+
+static inline HVX_Vector hvx_vec_reduce_sum_n_f32(HVX_Vector in, unsigned int n) {
+    unsigned int total = n * 4;  // total vec nbytes
+    unsigned int width = 4;      // fp32 nbytes
+
+    HVX_Vector sum = in, sum_t;
+    while (width < total) {
+        sum_t = Q6_V_vror_VR(sum, width);       // rotate right
+        sum   = Q6_Vsf_vadd_VsfVsf(sum, sum_t); // elementwise sum
+        width = width << 1;
+    }
+    return sum;
+}
+
+#else
+
+static inline HVX_Vector hvx_vec_reduce_sum_f32x2(HVX_Vector in0, HVX_Vector in1) {
+    HVX_VectorPair sump = Q6_W_vshuff_VVR(in1, in0, 4);
+    HVX_Vector  sum_qf  = Q6_Vqf32_vadd_VsfVsf(Q6_V_lo_W(sump), Q6_V_hi_W(sump));
+
+    sum_qf = Q6_Vqf32_vadd_Vqf32Vsf(sum_qf, Q6_V_vror_VR(Q6_Vsf_equals_Vqf32(sum_qf), VLEN / 2));
+    sum_qf = Q6_Vqf32_vadd_Vqf32Vsf(sum_qf, Q6_V_vror_VR(Q6_Vsf_equals_Vqf32(sum_qf), VLEN / 4));
+    sum_qf = Q6_Vqf32_vadd_Vqf32Vsf(sum_qf, Q6_V_vror_VR(Q6_Vsf_equals_Vqf32(sum_qf), VLEN / 8));
+    sum_qf = Q6_Vqf32_vadd_Vqf32Vsf(sum_qf, Q6_V_vror_VR(Q6_Vsf_equals_Vqf32(sum_qf), VLEN / 16));
+    return Q6_Vsf_equals_Vqf32(sum_qf);
+}
+
 static inline HVX_Vector hvx_vec_reduce_sum_n_f32(HVX_Vector in, unsigned int n) {
     unsigned int total = n * 4;  // total vec nbytes
     unsigned int width = 4;      // fp32 nbytes
@@ -57,6 +96,8 @@ static inline HVX_Vector hvx_vec_reduce_sum_n_f32(HVX_Vector in, unsigned int n)
     return sum;
 }
 
+#endif
+
 static inline HVX_Vector hvx_vec_reduce_sum_f32(HVX_Vector in) {
     return hvx_vec_reduce_sum_n_f32(in, 32);
 }
diff --git a/ggml/src/ggml-hexagon/htp/matmul-ops.c b/ggml/src/ggml-hexagon/htp/matmul-ops.c
index 1603ff2b3b..d251eeed33 100644
--- a/ggml/src/ggml-hexagon/htp/matmul-ops.c
+++ b/ggml/src/ggml-hexagon/htp/matmul-ops.c
@@ -11,6 +11,7 @@
 
 #include "hex-dma.h"
 #include "hvx-utils.h"
+#include "hvx-dump.h"
 
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
@@ -320,7 +321,7 @@ static void vec_dot_q4x4x2_q8x4x2(const int n, float * restrict s, const void *
     const uint8_t * restrict y_q = ((const uint8_t *) vy + 0);               // quants first
     const uint8_t * restrict y_d = ((const uint8_t *) vy + y_qrow_size);     // then scales
 
-    // Row sum (qf32)
+    // Row sum (sf)
     HVX_Vector r0_sum = Q6_V_vsplat_R(0);
 
     // Multiply and accumulate into int32.
@@ -344,7 +345,7 @@ static void vec_dot_q4x4x2_q8x4x2(const int n, float * restrict s, const void *
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
     }
 
     // Process leftovers, we still load full 4x4x2 block but zero out unused scales/blocks
@@ -362,14 +363,14 @@ static void vec_dot_q4x4x2_q8x4x2(const int n, float * restrict s, const void *
         // Zero out unused scales
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
         r0_dd                = Q6_V_vand_QV(bmask, r0_dd);
+        r0_ia                = Q6_V_vand_QV(bmask, r0_ia);
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
     }
 
-    // Reduce and convert into fp32
-    r0_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r0_sum));
+    r0_sum = hvx_vec_reduce_sum_f32(r0_sum);
 
     hvx_vec_store_u(&s[0], 4, r0_sum);
 }
@@ -402,7 +403,7 @@ static void vec_dot_q4x4x2_q8x4x2_rx2(const int n,
     const uint8_t * restrict y_q = ((const uint8_t *) vy + 0);                                     // quants first
     const uint8_t * restrict y_d = ((const uint8_t *) vy + y_qrow_size);                           // then scales
 
-    // Row sum (qf32)
+    // Row sum (sf)
     HVX_Vector r0_sum = Q6_V_vsplat_R(0);
     HVX_Vector r1_sum = Q6_V_vsplat_R(0);
 
@@ -432,8 +433,8 @@ static void vec_dot_q4x4x2_q8x4x2_rx2(const int n,
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
         HVX_Vector r1_fa = Q6_Vqf32_vmpy_VsfVsf(r1_ia, r1_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
-        r1_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r1_sum, r1_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
+        r1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_fa, r1_sum));
     }
 
     // Process leftovers, we still load full 4x4x2 block but zero out unused scales/blocks
@@ -456,20 +457,18 @@ static void vec_dot_q4x4x2_q8x4x2_rx2(const int n,
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
         r0_dd                = Q6_V_vand_QV(bmask, r0_dd);
         r1_dd                = Q6_V_vand_QV(bmask, r1_dd);
+        r0_ia                = Q6_V_vand_QV(bmask, r0_ia);
+        r1_ia                = Q6_V_vand_QV(bmask, r1_ia);
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
         HVX_Vector r1_fa = Q6_Vqf32_vmpy_VsfVsf(r1_ia, r1_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
-        r1_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r1_sum, r1_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
+        r1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_fa, r1_sum));
     }
 
-    // Convert into fp32 and reduce
-    r0_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r0_sum));
-    r1_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r1_sum));
-    HVX_VectorPair p0 = Q6_W_vshuff_VVR(r1_sum, r0_sum, 4);
-
-    hvx_vec_store_u(&s[0], 8, Q6_V_lo_W(p0));
+    HVX_Vector rsum = hvx_vec_reduce_sum_f32x2(r0_sum, r1_sum);
+    hvx_vec_store_u(&s[0], 8, rsum);
 }
 
 static void vec_dot_q8x4x2_q8x4x2(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
@@ -493,7 +492,7 @@ static void vec_dot_q8x4x2_q8x4x2(const int n, float * restrict s, const void *
     const uint8_t * restrict y_q = ((const uint8_t *) vy + 0);               // quants first
     const uint8_t * restrict y_d = ((const uint8_t *) vy + y_qrow_size);     // then scales
 
-    // Row sum (qf32)
+    // Row sum (sf)
     HVX_Vector r0_sum = Q6_V_vsplat_R(0);
 
     // Multiply and accumulate into int32.
@@ -517,7 +516,7 @@ static void vec_dot_q8x4x2_q8x4x2(const int n, float * restrict s, const void *
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
     }
 
     // Process leftovers, we still load full 4x4x2 block but zero out unused scales/blocks
@@ -535,14 +534,14 @@ static void vec_dot_q8x4x2_q8x4x2(const int n, float * restrict s, const void *
         // Zero out unused scales
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
         r0_dd                = Q6_V_vand_QV(bmask, r0_dd);
+        r0_ia                = Q6_V_vand_QV(bmask, r0_ia);
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
     }
 
-    // Reduce and convert into fp32
-    r0_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r0_sum));
+    r0_sum = hvx_vec_reduce_sum_f32(r0_sum);
 
     hvx_vec_store_u(&s[0], 4, r0_sum);
 }
@@ -605,8 +604,8 @@ static void vec_dot_q8x4x2_q8x4x2_rx2(const int n,
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
         HVX_Vector r1_fa = Q6_Vqf32_vmpy_VsfVsf(r1_ia, r1_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
-        r1_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r1_sum, r1_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
+        r1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_fa, r1_sum));
     }
 
     // Process leftovers, we still load full 4x4x2 block but zero out unused scales/blocks
@@ -629,20 +628,18 @@ static void vec_dot_q8x4x2_q8x4x2_rx2(const int n,
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
         r0_dd                = Q6_V_vand_QV(bmask, r0_dd);
         r1_dd                = Q6_V_vand_QV(bmask, r1_dd);
+        r0_ia                = Q6_V_vand_QV(bmask, r0_ia);
+        r1_ia                = Q6_V_vand_QV(bmask, r1_ia);
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
         HVX_Vector r1_fa = Q6_Vqf32_vmpy_VsfVsf(r1_ia, r1_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
-        r1_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r1_sum, r1_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
+        r1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_fa, r1_sum));
     }
 
-    // Convert into fp32 and reduce
-    r0_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r0_sum));
-    r1_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r1_sum));
-    HVX_VectorPair p0 = Q6_W_vshuff_VVR(r1_sum, r0_sum, 4);
-
-    hvx_vec_store_u(&s[0], 8, Q6_V_lo_W(p0));
+    HVX_Vector rsum = hvx_vec_reduce_sum_f32x2(r0_sum, r1_sum);
+    hvx_vec_store_u(&s[0], 8, rsum);
 }
 
 static void vec_dot_mxfp4x4x2_q8x4x2(const int n,
@@ -669,7 +666,7 @@ static void vec_dot_mxfp4x4x2_q8x4x2(const int n,
     const uint8_t * restrict y_q = ((const uint8_t *) vy + 0);               // quants first
     const uint8_t * restrict y_d = ((const uint8_t *) vy + y_qrow_size);     // then scales
 
-    // Row sum (qf32)
+    // Row sum (sf)
     HVX_Vector r0_sum = Q6_V_vsplat_R(0);
 
     // Multiply and accumulate into int32.
@@ -708,7 +705,7 @@ static void vec_dot_mxfp4x4x2_q8x4x2(const int n,
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
     }
 
     // Process leftovers
@@ -741,14 +738,14 @@ static void vec_dot_mxfp4x4x2_q8x4x2(const int n,
         // Zero-out unused scales
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
         r0_dd                = Q6_V_vand_QV(bmask, r0_dd);
+        r0_ia                = Q6_V_vand_QV(bmask, r0_ia);
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
     }
 
-    // Reduce and convert into fp32
-    r0_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r0_sum));
+    r0_sum = hvx_vec_reduce_sum_f32(r0_sum);
 
     hvx_vec_store_u(&s[0], 4, r0_sum);
 }
@@ -781,13 +778,13 @@ static void vec_dot_mxfp4x4x2_q8x4x2_rx2(const int n,
     const uint8_t * restrict y_q = ((const uint8_t *) vy + 0);                                     // quants first
     const uint8_t * restrict y_d = ((const uint8_t *) vy + y_qrow_size);                           // then scales
 
-    // Row sum (qf32)
+    // Row sum (sf)
     HVX_Vector r0_sum = Q6_V_vsplat_R(0);
     HVX_Vector r1_sum = Q6_V_vsplat_R(0);
 
     // Multiply and accumulate into int32.
     // Compute combined scale (fp32).
-    // Apply scale to acc and accumulate into the row sum (qf32).
+    // Apply scale to acc and accumulate into the row sum (f32).
 
     const uint32_t nb   = n / qk;  // num full blocks
     int32_t        nloe = n % qk;  // num leftover elemements (must be signed)
@@ -829,8 +826,8 @@ static void vec_dot_mxfp4x4x2_q8x4x2_rx2(const int n,
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
         HVX_Vector r1_fa = Q6_Vqf32_vmpy_VsfVsf(r1_ia, r1_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
-        r1_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r1_sum, r1_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
+        r1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_fa, r1_sum));
     }
 
     // Process leftovers
@@ -867,24 +864,22 @@ static void vec_dot_mxfp4x4x2_q8x4x2_rx2(const int n,
         HVX_Vector r0_dd = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(r0_d, vy_d));
         HVX_Vector r1_dd = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(r1_d, vy_d));
 
-        // Zero-out unused scales
+        // Zero-out unused values
         HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
         r0_dd                = Q6_V_vand_QV(bmask, r0_dd);
         r1_dd                = Q6_V_vand_QV(bmask, r1_dd);
+        r0_ia                = Q6_V_vand_QV(bmask, r0_ia);
+        r1_ia                = Q6_V_vand_QV(bmask, r1_ia);
 
         HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
         HVX_Vector r1_fa = Q6_Vqf32_vmpy_VsfVsf(r1_ia, r1_dd);
 
-        r0_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r0_sum, r0_fa);
-        r1_sum = Q6_Vqf32_vadd_Vqf32Vqf32(r1_sum, r1_fa);
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
+        r1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_fa, r1_sum));
     }
 
-    // Convert into fp32 and reduce
-    r0_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r0_sum));
-    r1_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(r1_sum));
-    HVX_VectorPair p0 = Q6_W_vshuff_VVR(r1_sum, r0_sum, 4);
-
-    hvx_vec_store_u(&s[0], 8, Q6_V_lo_W(p0));
+    HVX_Vector rsum = hvx_vec_reduce_sum_f32x2(r0_sum, r1_sum);
+    hvx_vec_store_u(&s[0], 8, rsum);
 }
 
 static void vec_dot_f16_f16_aa(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
@@ -913,7 +908,7 @@ static void vec_dot_f16_f16_aa(const int n, float * restrict s, const void * res
         rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf),  Q6_V_hi_W(xy_qf)));
     }
 
-    rsum = Q6_Vsf_equals_Vqf32(hvx_vec_reduce_sum_qf32(rsum));
+    rsum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(rsum));
     hvx_vec_store_u(&s[0], 4, rsum);
 }
 
@@ -957,11 +952,8 @@ static void vec_dot_f16_f16_aa_rx2(const int n,
         rsum1 = Q6_Vqf32_vadd_Vqf32Vqf32(rsum1, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy1_qf), Q6_V_hi_W(xy1_qf)));
     }
 
-    rsum0 = Q6_Vsf_equals_Vqf32(hvx_vec_reduce_sum_qf32(rsum0));
-    rsum1 = Q6_Vsf_equals_Vqf32(hvx_vec_reduce_sum_qf32(rsum1));
-    HVX_VectorPair p0 = Q6_W_vshuff_VVR(rsum1, rsum0, 4);
-
-    hvx_vec_store_u(&s[0], 8, Q6_V_lo_W(p0));
+    HVX_Vector rsum = hvx_vec_reduce_sum_f32x2(Q6_Vsf_equals_Vqf32(rsum0), Q6_Vsf_equals_Vqf32(rsum1));
+    hvx_vec_store_u(&s[0], 8, rsum);
 }
 
 static void vec_dot_f16_f16_uu(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
@@ -990,7 +982,7 @@ static void vec_dot_f16_f16_uu(const int n, float * restrict s, const void * res
         rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf),  Q6_V_hi_W(xy_qf)));
     }
 
-    rsum = Q6_Vsf_equals_Vqf32(hvx_vec_reduce_sum_qf32(rsum));
+    rsum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(rsum));
     hvx_vec_store_u(&s[0], 4, rsum);
 }
 
@@ -1042,7 +1034,8 @@ static void vec_dot_f16_f32_uu(const int n, float * restrict s, const void * res
         rsum = Q6_Vqf32_vadd_Vqf32Vqf32(rsum, Q6_Vqf32_vadd_Vqf32Vqf32(Q6_V_lo_W(xy_qf),  Q6_V_hi_W(xy_qf)));
     }
 
-    rsum = Q6_Vsf_equals_Vqf32(hvx_vec_reduce_sum_qf32(rsum));
+    // Convert into fp32 and reduce
+    rsum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(rsum));
     hvx_vec_store_u(&s[0], 4, rsum);
 }
 
diff --git a/ggml/src/ggml-hexagon/htp/softmax-ops.c b/ggml/src/ggml-hexagon/htp/softmax-ops.c
index 1b6b2eba4a..e91a16d947 100644
--- a/ggml/src/ggml-hexagon/htp/softmax-ops.c
+++ b/ggml/src/ggml-hexagon/htp/softmax-ops.c
@@ -154,8 +154,8 @@ static void hvx_fast_softmax_f32(const uint8_t * restrict src,
         v_pad[i] = v3;
     }
 
-    v       = hvx_vec_reduce_sum_qf32(sum_vec);
-    sum_vec = hvx_vec_repl4(Q6_Vsf_equals_Vqf32(v));
+    v       = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_vec));
+    sum_vec = hvx_vec_repl4(v);
 
     HVX_VectorPred pos_sum   = Q6_Q_vcmp_gt_VwVw(sum_vec, zero_v);
     HVX_Vector     v4        = hvx_vec_inverse_f32(sum_vec);
diff --git a/ggml/src/ggml-hexagon/htp/unary-ops.c b/ggml/src/ggml-hexagon/htp/unary-ops.c
index be8be8c4e6..1a27cb6e63 100644
--- a/ggml/src/ggml-hexagon/htp/unary-ops.c
+++ b/ggml/src/ggml-hexagon/htp/unary-ops.c
@@ -57,8 +57,8 @@ static void hvx_fast_rms_norm_f32(const uint8_t * restrict src,
         sum_v         = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, v2);
     }
 
-    HVX_Vector reduced_sum = hvx_vec_reduce_sum_qf32(sum_v);
-    sum_v                  = hvx_vec_repl4(Q6_Vsf_equals_Vqf32(reduced_sum));
+    HVX_Vector reduced_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_v));
+    sum_v                  = hvx_vec_repl4(reduced_sum);
 
     HVX_Vector t_v            = hvx_vec_splat_f32((float) num_elems);
     HVX_Vector denom_v        = hvx_vec_inverse_f32(t_v);
diff --git a/ggml/src/ggml-hexagon/libdl.h b/ggml/src/ggml-hexagon/libdl.h
new file mode 100644
index 0000000000..8ca5016f03
--- /dev/null
+++ b/ggml/src/ggml-hexagon/libdl.h
@@ -0,0 +1,79 @@
+#pragma once
+
+#ifdef _WIN32
+#   define WIN32_LEAN_AND_MEAN
+#   ifndef NOMINMAX
+#       define NOMINMAX
+#   endif
+#   include <windows.h>
+#   include <winevt.h>
+#else
+#    include <dlfcn.h>
+#    include <unistd.h>
+#endif
+#include <filesystem>
+
+namespace fs = std::filesystem;
+
+#ifdef _WIN32
+
+using dl_handle = std::remove_pointer_t<HMODULE>;
+
+struct dl_handle_deleter {
+    void operator()(HMODULE handle) {
+        FreeLibrary(handle);
+    }
+};
+
+static inline dl_handle * dl_load_library(const fs::path & path) {
+    // suppress error dialogs for missing DLLs
+    DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
+    SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
+
+    HMODULE handle = LoadLibraryW(path.wstring().c_str());
+
+    SetErrorMode(old_mode);
+
+    return handle;
+}
+
+static inline void * dl_get_sym(dl_handle * handle, const char * name) {
+    DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
+    SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
+
+    void * p = (void *) GetProcAddress(handle, name);
+
+    SetErrorMode(old_mode);
+
+    return p;
+}
+
+static inline const char * dl_error() {
+    return "";
+}
+
+#else
+
+using dl_handle = void;
+
+struct dl_handle_deleter {
+    void operator()(void * handle) {
+        dlclose(handle);
+    }
+};
+
+static inline dl_handle * dl_load_library(const fs::path & path) {
+    dl_handle * handle = dlopen(path.string().c_str(), RTLD_NOW | RTLD_LOCAL);
+    return handle;
+}
+
+static inline void * dl_get_sym(dl_handle * handle, const char * name) {
+    return dlsym(handle, name);
+}
+
+static inline const char * dl_error() {
+    const char *rslt = dlerror();
+    return rslt != nullptr ? rslt : "";
+}
+
+#endif
diff --git a/ggml/src/ggml-hexagon/libggml-htp.inf b/ggml/src/ggml-hexagon/libggml-htp.inf
new file mode 100644
index 0000000000..656d2d9ab2
--- /dev/null
+++ b/ggml/src/ggml-hexagon/libggml-htp.inf
@@ -0,0 +1,38 @@
+[Version]
+Signature   = "$WINDOWS NT$"
+Class       = ComputeAccelerator
+ClassGuid   = {F01A9D53-3FF6-48D2-9F97-C8A7004BE10C}
+Provider    = %GGML%
+DriverVer   = 01/01/2026,1.0.0.0
+CatalogFile = libggml-htp.cat
+PnpLockDown = 1
+
+[DestinationDirs]
+Drivers_Dir = 6
+
+[SourceDisksNames]
+1 = %DiskId%
+
+[SourceDisksFiles]
+libggml-htp-v68.so = 1
+libggml-htp-v69.so = 1
+libggml-htp-v73.so = 1
+libggml-htp-v75.so = 1
+libggml-htp-v81.so = 1
+
+[ControlFlags]
+ExcludeFromSelect = *
+
+[DefaultInstall.NTarm64]
+CopyFiles=Drivers_Dir
+
+[Drivers_Dir]
+libggml-htp-v68.so,,,0x10 ;COPYFLG_NO_OVERWRITE
+libggml-htp-v69.so,,,0x10 ;COPYFLG_NO_OVERWRITE
+libggml-htp-v73.so,,,0x10 ;COPYFLG_NO_OVERWRITE
+libggml-htp-v75.so,,,0x10 ;COPYFLG_NO_OVERWRITE
+libggml-htp-v81.so,,,0x10 ;COPYFLG_NO_OVERWRITE
+
+[Strings]
+GGML   = 'GGML'
+DiskId = 'GGML HTP library'
diff --git a/ggml/src/ggml-hip/CMakeLists.txt b/ggml/src/ggml-hip/CMakeLists.txt
index 23b6889919..80037d2436 100644
--- a/ggml/src/ggml-hip/CMakeLists.txt
+++ b/ggml/src/ggml-hip/CMakeLists.txt
@@ -62,6 +62,8 @@ file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-mma*.cu")
 list(APPEND GGML_SOURCES_ROCM ${SRCS})
 file(GLOB   SRCS "../ggml-cuda/template-instances/mmq*.cu")
 list(APPEND GGML_SOURCES_ROCM ${SRCS})
+file(GLOB   SRCS "../ggml-cuda/template-instances/mmf*.cu")
+list(APPEND GGML_SOURCES_ROCM ${SRCS})
 
 if (GGML_CUDA_FA_ALL_QUANTS)
     file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*.cu")
diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h
index 80e0fd2ff8..baadfe9a7b 100644
--- a/ggml/src/ggml-impl.h
+++ b/ggml/src/ggml-impl.h
@@ -611,6 +611,9 @@ static inline bool ggml_can_fuse_ext(const struct ggml_cgraph * cgraph, const in
         if (node->op != ops[i]) {
             return false;
         }
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            return false;
+        }
         if (i < num_ops - 1 && !ggml_node_has_n_uses(cgraph, node_idxs[i], 1)) {
             return false;
         }
diff --git a/ggml/src/ggml-metal/ggml-metal-device.m b/ggml/src/ggml-metal/ggml-metal-device.m
index c418afe9c3..7f9c384c34 100644
--- a/ggml/src/ggml-metal/ggml-metal-device.m
+++ b/ggml/src/ggml-metal/ggml-metal-device.m
@@ -785,8 +785,12 @@ ggml_metal_device_t ggml_metal_device_init(void) {
             dev->props.op_offload_min_batch_size  = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32;
 
             dev->props.max_buffer_size            = dev->mtl_device.maxBufferLength;
-            dev->props.max_working_set_size       = dev->mtl_device.recommendedMaxWorkingSetSize;
             dev->props.max_theadgroup_memory_size = dev->mtl_device.maxThreadgroupMemoryLength;
+            if (@available(macOS 10.12, iOS 16.0, *)) {
+                dev->props.max_working_set_size   = dev->mtl_device.recommendedMaxWorkingSetSize;
+            } else {
+                dev->props.max_working_set_size   = dev->mtl_device.maxBufferLength;
+            }
 
             strncpy(dev->props.name, [[dev->mtl_device name] UTF8String], sizeof(dev->props.name) - 1);
 
@@ -1078,12 +1082,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 op->src[0]->ne[0] != 112 &&
                 op->src[0]->ne[0] != 128 &&
                 op->src[0]->ne[0] != 192 &&
-                op->src[0]->ne[0] != 256) {
-                return false;
-            }
-            if (op->src[0]->ne[0] == 576) {
-                // DeepSeek sizes
-                // TODO: disabled for now, until optmized
+                op->src[0]->ne[0] != 256 &&
+                op->src[0]->ne[0] != 576) {
                 return false;
             }
             if (op->src[1]->type != op->src[2]->type) {
diff --git a/ggml/src/ggml-metal/ggml-metal-ops.cpp b/ggml/src/ggml-metal/ggml-metal-ops.cpp
index 680ad794de..7f4cfbba22 100644
--- a/ggml/src/ggml-metal/ggml-metal-ops.cpp
+++ b/ggml/src/ggml-metal/ggml-metal-ops.cpp
@@ -203,6 +203,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
         GGML_ABORT("unsupported op");
     }
 
+    if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+        return 1;
+    }
+
     int n_fuse = 1;
 
     // check if the current node can run concurrently with other nodes before it
@@ -2516,7 +2520,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 
         // simdgroups per threadgroup (a.k.a. warps)
         //nsg = ne01 <= nqptg ? MAX(4, MIN(nsgmax, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32))) : 4;
-        int32_t nsg = 4;
+        int32_t nsg = ne00 >= 512 ? 8 : 4;
 
         const size_t smem = FATTN_SMEM(nsg);
 
diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
index a4e1cafe55..17e358d1a8 100644
--- a/ggml/src/ggml-metal/ggml-metal.metal
+++ b/ggml/src/ggml-metal/ggml-metal.metal
@@ -5552,9 +5552,7 @@ void kernel_flash_attn_ext_impl(
 
                 constexpr short NC = (C/8)/NSG;
 
-                // note: do not unroll for large heads
-                #pragma unroll (DK <= 64 ? NC : 1)
-                for (short cc = 0; cc < NC; ++cc) {
+                FOR_UNROLL (short cc = 0; cc < NC; ++cc) {
                     qk8x8_t mqk = make_filled_simdgroup_matrix<qk_t, 8>((qk_t) 0.0f);
 
                     if (DK % 16 != 0) {
@@ -5575,7 +5573,9 @@ void kernel_flash_attn_ext_impl(
                         k8x8_t mk[2];
                         q8x8_t mq[2];
 
-                        FOR_UNROLL (short i = 0; i < DK8/2; ++i) {
+                        // note: too much unroll can tank the performance for large heads
+                        #pragma unroll (MIN(DK8/2, 4*NSG))
+                        for (short i = 0; i < DK8/2; ++i) {
                             simdgroup_barrier(mem_flags::mem_none);
 
                             simdgroup_load(mq[0], pq + 0*8 + 16*i, DK);
@@ -5749,7 +5749,9 @@ void kernel_flash_attn_ext_impl(
                                 pv  += 8*NS20;
                             }
                         } else {
-                            FOR_UNROLL (short cc = 0; cc < (C/8)/2; ++cc) {
+                            constexpr short NC = (C/8)/2;
+
+                            FOR_UNROLL (short cc = 0; cc < NC; ++cc) {
                                 s8x8_t vs[2];
 
                                 simdgroup_load(vs[0], ss + 16*cc + 0, SH, 0, false);
@@ -5952,6 +5954,7 @@ kernel void kernel_flash_attn_ext(
       //case 1: kernel_flash_attn_ext_impl<FWD_TMPL, 1>(FWD_ARGS); break;
       //case 2: kernel_flash_attn_ext_impl<FWD_TMPL, 2>(FWD_ARGS); break;
         case 4: kernel_flash_attn_ext_impl<FWD_TMPL, 4>(FWD_ARGS); break;
+        case 8: kernel_flash_attn_ext_impl<FWD_TMPL, 8>(FWD_ARGS); break;
     }
 #undef FWD_TMPL
 #undef FWD_ARGS
diff --git a/ggml/src/ggml-opencl/CMakeLists.txt b/ggml/src/ggml-opencl/CMakeLists.txt
index e000c576be..f9c92ef05a 100644
--- a/ggml/src/ggml-opencl/CMakeLists.txt
+++ b/ggml/src/ggml-opencl/CMakeLists.txt
@@ -59,6 +59,7 @@ set(GGML_OPENCL_KERNELS
     add
     add_id
     argsort
+    tri
     fill
     clamp
     cpy
@@ -86,7 +87,8 @@ set(GGML_OPENCL_KERNELS
     mul_mv_q4_0_f32_8x_flat
     mul_mv_q4_0_f32_1d_8x_flat
     mul_mv_q4_0_f32_1d_16x_flat
-    mul_mv_q6_k
+    mul_mv_q6_k_f32
+    mul_mv_q6_k_f32_flat
     mul_mv_q8_0_f32
     mul_mv_q8_0_f32_flat
     mul_mv_mxfp4_f32
@@ -101,6 +103,8 @@ set(GGML_OPENCL_KERNELS
     mul_mm_f32_f32_l4_lm
     mul_mm_f16_f32_l4_lm
     mul_mm_q8_0_f32_l4_lm
+    mul_mm_q8_0_f32_8x4
+    gemv_noshuffle_general_q8_0_f32
     mul
     norm
     relu
diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index 1b7bc382b8..28682357e4 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -230,7 +230,8 @@ static ADRENO_GPU_GEN get_adreno_gpu_gen(const char *device_name) {
         return ADRENO_GPU_GEN::A7X;
     }
 
-    if (strstr(device_name, "830")) {
+    if (strstr(device_name, "830") ||
+        strstr(device_name, "840")) {
         return ADRENO_GPU_GEN::A8X;
     }
 
@@ -402,6 +403,7 @@ struct ggml_backend_opencl_context {
     int adreno_wave_size;
 
     cl_bool non_uniform_workgroups;
+    size_t  image_max_buffer_size;
 
     cl_context context;
     cl_command_queue queue;
@@ -411,6 +413,10 @@ struct ggml_backend_opencl_context {
     ggml_cl_buffer prealloc_scales_trans;
     ggml_cl_buffer prealloc_act_trans;
 
+    // prealloc buffers for src0 and src1
+    ggml_cl_buffer prealloc_src0;
+    ggml_cl_buffer prealloc_src1;
+
     cl_program program_add;
     cl_program program_add_id;
     cl_program program_clamp;
@@ -493,6 +499,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_gelu_quick, kernel_gelu_quick_4;
     cl_kernel kernel_relu;
     cl_kernel kernel_sigmoid_f32, kernel_sigmoid_f16;
+    cl_kernel kernel_tri;
     cl_kernel kernel_fill;
     cl_kernel kernel_clamp;
     cl_kernel kernel_geglu, kernel_reglu, kernel_swiglu, kernel_swiglu_oai, kernel_geglu_erf, kernel_geglu_quick,
@@ -527,12 +534,14 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_mul_mat_q4_0_f32, kernel_mul_mat_q4_0_f32_v;
     cl_kernel kernel_convert_block_q4_0, kernel_restore_block_q4_0;
     cl_kernel kernel_convert_block_mxfp4, kernel_convert_block_mxfp4_trans, kernel_restore_block_mxfp4, kernel_restore_block_mxfp4_trans;
-    cl_kernel kernel_convert_block_q8_0, kernel_restore_block_q8_0;
+    cl_kernel kernel_convert_block_q8_0, kernel_restore_block_q8_0, kernel_restore_block_q8_0_trans;
     cl_kernel kernel_mul_mat_q4_0_f32_8x_flat;
     cl_kernel kernel_convert_block_q4_0_noshuffle;
     cl_kernel kernel_restore_block_q4_0_noshuffle;
+    cl_kernel kernel_convert_block_q6_K, kernel_restore_block_q6_K;
     cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
     cl_kernel kernel_mul_mv_q6_K_f32;
+    cl_kernel kernel_mul_mv_q6_K_f32_flat;
     cl_kernel kernel_mul_mv_mxfp4_f32, kernel_mul_mv_mxfp4_f32_flat;
     cl_kernel kernel_mul_mv_q8_0_f32, kernel_mul_mv_q8_0_f32_flat;
     cl_kernel kernel_solve_tri_f32;
@@ -692,6 +701,8 @@ struct ggml_backend_opencl_context {
     cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_4096_1_4096;
     cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_11008_1_4096;
     cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_32000_1_4096;
+    cl_kernel kernel_mul_mm_q8_0_f32_8x4;
+    cl_kernel CL_mul_mat_vec_q8_0_f32;
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
 
     void free() {
@@ -797,6 +808,24 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // tri
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "tri.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("tri.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_tri = clCreateKernel(prog, "kernel_tri_f32", &err), err));
+        GGML_LOG_CONT(".");
+
+        CL_CHECK(clReleaseProgram(prog));
+    }
+
     // fill
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -872,6 +901,9 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         CL_CHECK((backend_ctx->kernel_restore_block_mxfp4 = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_mxfp4", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_q8_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q8_0", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_q8_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q8_0", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q8_0_trans  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q8_0_trans", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_q6_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q6_K", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q6_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q6_K", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -1094,14 +1126,14 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
-    // mul_mv_q6_k
+    // mul_mv_q6_k_f32
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
         const std::string kernel_src {
-            #include "mul_mv_q6_k.cl.h"
+            #include "mul_mv_q6_k_f32.cl.h"
         };
 #else
-        const std::string kernel_src = read_file("mul_mv_q6_k.cl");
+        const std::string kernel_src = read_file("mul_mv_q6_k_f32.cl");
 #endif
         backend_ctx->program_mul_mv_q6_K =
             build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
@@ -1110,6 +1142,23 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mv_q6_k_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q6_k_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q6_k_f32_flat.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_q6_K_f32_flat = clCreateKernel(prog, "kernel_mul_mv_q6_K_f32_flat", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
     // mul_mv_q8_0_f32
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -2249,6 +2298,46 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mm_q8_0_f32_8x4
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src_q8_8x4_gemm {
+            #include "mul_mm_q8_0_f32_8x4.cl.h"
+       };
+#else
+        const std::string kernel_src_q8_8x4_gemm = read_file("mul_mm_q8_0_f32_8x4.cl");
+#endif
+        backend_ctx->program_CL_gemm = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_q8_8x4_gemm.c_str(), compile_opts);
+        CL_CHECK((backend_ctx->kernel_mul_mm_q8_0_f32_8x4 = clCreateKernel(backend_ctx->program_CL_gemm, "kernel_mul_mm_q8_0_f32_8x4", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // gemv_noshuffle_general_q8_0_f32
+    {
+        std::string CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std +
+                                       " -cl-mad-enable "
+                                       " -DSIMDGROUP_WIDTH=" +
+                                       std::to_string(backend_ctx->adreno_wave_size);
+        if (backend_ctx->has_vector_subgroup_broadcast) {
+            CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT ";
+        }
+
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src_CL_gemv_general {
+            #include "gemv_noshuffle_general_q8_0_f32.cl.h"
+        };
+#else
+        const std::string kernel_src_CL_gemv_general = read_file("gemv_noshuffle_general_q8_0_f32.cl");
+#endif
+
+        cl_program prog = build_program_from_source(
+            backend_ctx->context, backend_ctx->device, kernel_src_CL_gemv_general.c_str(), CL_gemv_compile_opts);
+
+        CL_CHECK((backend_ctx->CL_mul_mat_vec_q8_0_f32 = clCreateKernel(prog, "kernel_gemv_noshuffle", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
     std::string CL_moe_compile_opts = std::string("-cl-std=") + opencl_c_std +
             " -cl-mad-enable "
             " -cl-fast-relaxed-math";
@@ -2643,6 +2732,9 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
     clGetDeviceInfo(device, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(size_t), &backend_ctx->max_alloc_size, NULL);
     GGML_LOG_INFO("ggml_opencl: max mem alloc size: %zu MB\n", backend_ctx->max_alloc_size/1024/1024);
 
+    clGetDeviceInfo(device, CL_DEVICE_IMAGE_MAX_BUFFER_SIZE, sizeof(size_t), &backend_ctx->image_max_buffer_size, NULL);
+    GGML_LOG_INFO("ggml_opencl: device max image buffer size (pixels): %lu\n", backend_ctx->image_max_buffer_size);
+
     clGetDeviceInfo(device, CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(size_t), &backend_ctx->max_workgroup_size, NULL);
     GGML_LOG_INFO("ggml_opencl: device max workgroup size: %lu\n", backend_ctx->max_workgroup_size);
 
@@ -3022,6 +3114,50 @@ struct ggml_tensor_extra_cl_q8_0 {
     }
 };
 
+struct ggml_tensor_extra_cl_q6_K {
+    // Lower 4 bits of quantized weights.
+    cl_mem ql = nullptr;
+    // Upper 2 bits of quantized weights.
+    cl_mem qh = nullptr;
+    // Scales for each block.
+    cl_mem s  = nullptr;
+    // Scales for each super block.
+    cl_mem d  = nullptr;
+
+    size_t size_ql = 0;
+    size_t size_qh = 0;
+    size_t size_s  = 0;
+    size_t size_d  = 0;
+
+    ~ggml_tensor_extra_cl_q6_K() {
+        reset();
+    }
+
+    void reset() {
+        if (ql != nullptr) {
+            CL_CHECK(clReleaseMemObject(ql));
+            ql = nullptr;
+        }
+        if (qh != nullptr) {
+            CL_CHECK(clReleaseMemObject(qh));
+            qh = nullptr;
+        }
+        if (s != nullptr) {
+            CL_CHECK(clReleaseMemObject(s));
+            s = nullptr;
+        }
+        if (d != nullptr) {
+            CL_CHECK(clReleaseMemObject(d));
+            d = nullptr;
+        }
+
+        size_ql = 0;
+        size_qh = 0;
+        size_s  = 0;
+        size_d  = 0;
+    }
+};
+
 //------------------------------------------------------------------------------
 // Backend API
 //------------------------------------------------------------------------------
@@ -3188,6 +3324,10 @@ static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggm
             continue;
         }
 
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
+
         if (!backend_ctx->disable_fusion && ggml_opencl_can_fuse(cgraph, i, { GGML_OP_NORM, GGML_OP_MUL, GGML_OP_ADD })) {
             ggml_opencl_op_norm_fused(backend, node, cgraph->nodes[i+1], cgraph->nodes[i+2]);
             i += 2;
@@ -3331,6 +3471,8 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 default:
                     return false;
             }
+        case GGML_OP_TRI:
+            return op->type == GGML_TYPE_F32 && ggml_is_contiguous(op);
         case GGML_OP_FILL:
             return op->type == GGML_TYPE_F32 && ggml_is_contiguous(op);
         case GGML_OP_CLAMP:
@@ -3562,6 +3704,12 @@ struct ggml_backend_opencl_buffer_context {
         for (ggml_tensor_extra_cl_q8_0 * e : temp_tensor_extras_q8_0_in_use) {
             delete e;
         }
+        for (ggml_tensor_extra_cl_q6_K * e : temp_tensor_extras_q6_K) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_q6_K * e : temp_tensor_extras_q6_K_in_use) {
+            delete e;
+        }
     }
 
     ggml_tensor_extra_cl * ggml_opencl_alloc_temp_tensor_extra() {
@@ -3624,6 +3772,21 @@ struct ggml_backend_opencl_buffer_context {
         return extra;
     }
 
+    ggml_tensor_extra_cl_q6_K * ggml_opencl_alloc_temp_tensor_extra_q6_K() {
+        ggml_tensor_extra_cl_q6_K * extra;
+        if (temp_tensor_extras_q6_K.empty()) {
+            extra = new ggml_tensor_extra_cl_q6_K();
+        } else {
+            extra = temp_tensor_extras_q6_K.back();
+            temp_tensor_extras_q6_K.pop_back();
+        }
+
+        temp_tensor_extras_q6_K_in_use.push_back(extra);
+
+        extra->reset();
+        return extra;
+    }
+
     void reset() {
         for (ggml_tensor_extra_cl * e : temp_tensor_extras_in_use) {
             temp_tensor_extras.push_back(e);
@@ -3644,6 +3807,11 @@ struct ggml_backend_opencl_buffer_context {
             temp_tensor_extras_q8_0.push_back(e);
         }
         temp_tensor_extras_q8_0_in_use.clear();
+
+        for (ggml_tensor_extra_cl_q6_K * e : temp_tensor_extras_q6_K_in_use) {
+            temp_tensor_extras_q6_K.push_back(e);
+        }
+        temp_tensor_extras_q6_K_in_use.clear();
     }
 
     // Pools for extras. Available extras are in `temp_tensor_extras`. Extras
@@ -3659,6 +3827,8 @@ struct ggml_backend_opencl_buffer_context {
     std::vector<ggml_tensor_extra_cl_mxfp4 *> temp_tensor_extras_mxfp4_in_use;
     std::vector<ggml_tensor_extra_cl_q8_0 *> temp_tensor_extras_q8_0;
     std::vector<ggml_tensor_extra_cl_q8_0 *> temp_tensor_extras_q8_0_in_use;
+    std::vector<ggml_tensor_extra_cl_q6_K *> temp_tensor_extras_q6_K;
+    std::vector<ggml_tensor_extra_cl_q6_K *> temp_tensor_extras_q6_K_in_use;
 
     // The buffer_context is initially created by ggml_backend_buft_alloc_buffer
     // before any tensor is initialized (at the beginning of alloc_tensor_range).
@@ -3749,6 +3919,15 @@ inline bool use_adreno_moe_kernels(const ggml_backend_opencl_context *backend_ct
     return ((strstr(tensor->name, "ffn") != NULL) || (strstr(tensor->name, "as") != NULL)) && (ne01 % 64 == 0);
 }
 
+inline bool enable_adreno_trans_weight(const ggml_backend_opencl_context *backend_ctx, const ggml_tensor *tensor) {
+
+    bool adreno_kernel = use_adreno_kernels(backend_ctx, tensor);
+
+    size_t elem_num = tensor->ne[0] * tensor->ne[1] * tensor->ne[2] * tensor->ne[3];
+
+    return ((elem_num < 128 * 1024 * 1024) && adreno_kernel);  // max element num: 2**27
+}
+
 static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     ggml_backend_opencl_context *backend_ctx = ggml_cl2_init(buffer->buft->device);
 
@@ -4163,6 +4342,216 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
 
         tensor->extra = extra;
 
+        // Transpose the weights and scales
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (enable_adreno_trans_weight(backend_ctx, tensor)) {
+
+            int M = tensor->ne[1];   // ne01
+            int K = tensor->ne[0];   // ne00
+
+            GGML_ASSERT(K % 32 == 0);
+            GGML_ASSERT(M % 4 == 0);
+            GGML_ASSERT(tensor->ne[2] == 1);
+            GGML_ASSERT(tensor->ne[3] == 1);
+
+            // Transpose weights
+            size_t q_size_bytes = K * M / 4 * sizeof(float);
+            cl_buffer_region region;
+            region.origin = 0;
+            region.size = q_size_bytes;
+            cl_mem qT_d = clCreateSubBuffer(
+                backend_ctx->prealloc_quant_trans.buffer,
+                0,
+                CL_BUFFER_CREATE_TYPE_REGION,
+                &region,
+                &err);
+            CL_CHECK(err);
+
+            cl_mem q_d_image1D;
+            cl_mem qT_d_image1D;
+
+            cl_image_format img_fmt_1d;
+            cl_image_desc img_desc_1d;
+
+            img_fmt_1d = { CL_RGBA, CL_FLOAT };
+            memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+            img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+            img_desc_1d.image_width = M * K / 4 / 4;
+            img_desc_1d.buffer = extra->q;
+            q_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
+            CL_CHECK(err);
+
+            img_fmt_1d = { CL_RGBA, CL_FLOAT };
+            memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+            img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+            img_desc_1d.image_width = M * K / 4 / 4;
+            img_desc_1d.buffer = qT_d;
+            qT_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
+            CL_CHECK(err);
+
+            int height_q = M / 4;
+            int width_q = K / 4 / 4;
+            kernel = backend_ctx->kernel_transpose_32;
+
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &q_d_image1D));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &qT_d_image1D));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),    &height_q));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int),    &width_q));
+
+            size_t local_size_q[3] = {4, 16, 1};
+            size_t global_size_q[3] = {static_cast<size_t>(width_q), static_cast<size_t>(height_q), 1};
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_size_q, local_size_q, 0, NULL, &evt));
+            CL_CHECK(clWaitForEvents(1, &evt));
+
+            // Transpose scales
+            size_t d_size_bytes = M * (K / 32) * 2;
+            region.origin = 0;
+            region.size = d_size_bytes;
+            cl_mem dT_d = clCreateSubBuffer(
+                backend_ctx->prealloc_scales_trans.buffer,
+                0,
+                CL_BUFFER_CREATE_TYPE_REGION,
+                &region,
+                &err);
+            CL_CHECK(err);
+
+            cl_mem d_d_image1D;
+            cl_mem dT_d_image1D;
+
+            memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+            img_fmt_1d = { CL_R, CL_HALF_FLOAT };
+            img_desc_1d.image_width = M * K / 32;
+            img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+            img_desc_1d.buffer = extra->d;
+            d_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
+            CL_CHECK(err);
+
+            img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
+            memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+            img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+            img_desc_1d.image_width = M * K / 32 / 4;
+            img_desc_1d.buffer = dT_d;
+            dT_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
+            CL_CHECK(err);
+
+            int height_s = M / 4;
+            int width_s = K / 32;
+
+            kernel = backend_ctx->kernel_transpose_16_4x1;
+
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &d_d_image1D));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &dT_d_image1D));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int), &height_s));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int), &width_s));
+
+            size_t local_size_s[3] = {4, 16, 1};
+            size_t global_size_s[3] = {static_cast<size_t>(width_s), static_cast<size_t>(height_s), 1};
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_size_s, local_size_s, 0, NULL, &evt));
+            CL_CHECK(clWaitForEvents(1, &evt));
+
+            // copy transposed buffer contents to original buffers
+            CL_CHECK(clEnqueueCopyBuffer(queue, qT_d, extra->q, 0, 0, q_size_bytes, 0, NULL, &evt));
+            CL_CHECK(clWaitForEvents(1, &evt));
+
+            CL_CHECK(clEnqueueCopyBuffer(queue, dT_d, extra->d, 0, 0, d_size_bytes, 0, NULL, &evt));
+            CL_CHECK(clWaitForEvents(1, &evt));
+
+            CL_CHECK(clReleaseMemObject(qT_d));
+            CL_CHECK(clReleaseMemObject(dT_d));
+
+            CL_CHECK(clReleaseMemObject(q_d_image1D));
+            CL_CHECK(clReleaseMemObject(d_d_image1D));
+            CL_CHECK(clReleaseMemObject(qT_d_image1D));
+            CL_CHECK(clReleaseMemObject(dT_d_image1D));
+        } // end transpose
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
+
+        return;
+    }
+    if (tensor->type == GGML_TYPE_Q6_K) {
+        ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
+        GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized");
+
+        // Allocate the new extra and create aliases from the original.
+        ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+        ggml_tensor_extra_cl_q6_K * extra = ctx->ggml_opencl_alloc_temp_tensor_extra_q6_K();
+
+        size_t size_ql = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/2;
+        size_t size_qh = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/4;
+        size_t size_s  = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/16;
+        size_t size_d  = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
+        GGML_ASSERT(size_ql + size_qh + size_s + size_d == ggml_nbytes(tensor) &&
+            "Incorrect tensor size");
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+        CL_CHECK(clEnqueueWriteBuffer(
+            queue, data_device, CL_TRUE, 0,
+            ggml_nbytes(tensor), data, 0, NULL, NULL));
+
+        cl_buffer_region region;
+
+        // Subbuffer for ql
+        region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment);
+        region.size = size_ql;
+        extra->ql = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        auto previous_origin = region.origin;
+
+        // Subbuffer for qh
+        region.origin = align_to(previous_origin + size_ql, backend_ctx->alignment);
+        region.size = size_qh;
+        extra->qh = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        previous_origin = region.origin;
+
+        // Subbuffer for scales
+        region.origin = align_to(previous_origin + size_qh, backend_ctx->alignment);
+        region.size = size_s;
+        extra->s = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        previous_origin = region.origin;
+
+        // Create subbuffer for d.
+        region.origin = align_to(previous_origin + size_s, backend_ctx->alignment);
+        region.size = size_d;
+        extra->d = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        previous_origin = region.origin;
+
+        // Flatten the weights
+        cl_kernel kernel = backend_ctx->kernel_convert_block_q6_K;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->ql));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->qh));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->s));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra->d));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clReleaseMemObject(data_device));
+
+        extra->size_ql = size_ql;
+        extra->size_qh = size_qh;
+        extra->size_s  = size_s;
+        extra->size_d  = size_d;
+
+        tensor->extra  = extra;
         return;
     }
 #endif // GGML_OPENCL_SOA_Q
@@ -4366,6 +4755,36 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
             ggml_nbytes(tensor), NULL, &err);
         CL_CHECK(err);
 
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (enable_adreno_trans_weight(backend_ctx, tensor)) {
+            cl_kernel kernel = backend_ctx->kernel_restore_block_q8_0_trans;
+
+            int ne00 = tensor->ne[0];
+            int ne01 = tensor->ne[1];
+            GGML_ASSERT(tensor->ne[2] == 1);  // ???
+            GGML_ASSERT(tensor->ne[3] == 1);  // ???
+
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->d));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &data_device));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_int), &ne00));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_int), &ne01));
+
+            size_t global_work_size[3] = {static_cast<size_t>(((ne01 + 63) / 64) * 64), 1, 1};
+            size_t local_work_size[3] = {64, 1, 1};
+
+            cl_event evt;
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+                global_work_size, local_work_size, 0, NULL, &evt));
+            CL_CHECK(clWaitForEvents(1, &evt));
+
+            CL_CHECK(clEnqueueReadBuffer(
+                queue, data_device, CL_TRUE, offset,
+                size, data, 0, NULL, NULL));
+            CL_CHECK(clReleaseMemObject(data_device));
+            return;
+        }
+#endif
         cl_kernel kernel = backend_ctx->kernel_restore_block_q8_0;
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
         CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->d));
@@ -4374,6 +4793,34 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
         size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
         size_t local_work_size[] = {1, 1, 1};
 
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+            global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clEnqueueReadBuffer(
+            queue, data_device, CL_TRUE, offset,
+            size, data, 0, NULL, NULL));
+        CL_CHECK(clReleaseMemObject(data_device));
+        return;
+    }
+    if (tensor->type == GGML_TYPE_Q6_K) {
+        ggml_tensor_extra_cl_q6_K * extra = (ggml_tensor_extra_cl_q6_K *)tensor->extra;
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+
+        cl_kernel kernel = backend_ctx->kernel_restore_block_q6_K;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->ql));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->qh));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->s));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->d));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &data_device));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {1, 1, 1};
+
         cl_event evt;
         CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
             global_work_size, local_work_size, 0, NULL, &evt));
@@ -4816,6 +5263,81 @@ static bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct gg
             (ne0 >= 32 && ne1 >= 32 && ne10 >= 32);
 }
 
+// Copy a noncontiguous tensor to contiguous tensor. ne[] remains the same but
+// nb[] is recalculated such that tensor is contiguous.
+static void ggml_cl_copy_to_contiguous(ggml_backend_t backend, const ggml_tensor * src, cl_mem dst,
+                                       cl_ulong &nb0, cl_ulong &nb1, cl_ulong &nb2, cl_ulong &nb3) {
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    const int tensor_type_size = ggml_type_size(src->type);
+
+    const int ne00 = src->ne[0];
+    const int ne01 = src->ne[1];
+    const int ne02 = src->ne[2];
+    const int ne03 = src->ne[3];
+
+    const cl_ulong nb00 = src->nb[0];
+    const cl_ulong nb01 = src->nb[1];
+    const cl_ulong nb02 = src->nb[2];
+    const cl_ulong nb03 = src->nb[3];
+
+    const int ne0 = src->ne[0];
+    const int ne1 = src->ne[1];
+    const int ne2 = src->ne[2];
+    const int ne3 = src->ne[3];
+
+    nb0 = tensor_type_size;
+    nb1 = tensor_type_size*ne00;
+    nb2 = tensor_type_size*ne00*ne01;
+    nb3 = tensor_type_size*ne00*ne01*ne02;
+
+    ggml_tensor_extra_cl * extra = (ggml_tensor_extra_cl *)src->extra;
+
+    cl_ulong offset0 = extra->offset + src->view_offs;
+    cl_ulong offsetd = 0;
+
+    cl_kernel kernel;
+
+    switch (src->type) {
+        case GGML_TYPE_F32:
+            kernel = backend_ctx->kernel_cpy_f32_f32;
+            break;
+        case GGML_TYPE_F16:
+            kernel = backend_ctx->kernel_cpy_f16_f16;
+            break;
+        default:
+            GGML_ASSERT(false && "not implemented");
+    }
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &dst));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne1));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne2));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne3));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb0));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb3));
+
+    const int nth = MIN(64, ne00);
+
+    size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+    size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, src);
+}
+
 static void ggml_cl_nop(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     UNUSED(backend);
     UNUSED(src0);
@@ -5903,6 +6425,44 @@ static void ggml_cl_sigmoid(ggml_backend_t backend, const ggml_tensor * src0, co
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
 }
 
+static void ggml_cl_tri(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int tri_type = ggml_get_op_params_i32(dst, 0);
+    const int64_t n = ggml_nelements(dst);
+    const int     ne0  = dst->ne[0];
+    const int     ne1  = dst->ne[1];
+
+    cl_kernel kernel = backend_ctx->kernel_tri;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),      &n));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne1));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),      &tri_type));
+
+    size_t local_work_size[1] = { 256 };
+    size_t global_work_size[1] = { ((size_t)n + local_work_size[0] - 1) / local_work_size[0] * local_work_size[0] };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 1, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_fill(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(dst);
     GGML_ASSERT(dst->extra);
@@ -7524,6 +8084,252 @@ static void ggml_cl_mul_mat_kq_kqv_adreno(ggml_backend_t backend, const ggml_ten
     CL_CHECK(clReleaseMemObject(D_sub_buffer));
 }
 
+static void ggml_cl_mul_mat_q8_0_f32_adreno(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    const enum ggml_type src0t = src0->type;
+    const enum ggml_type src1t = src1->type;
+
+    GGML_ASSERT(src0t == GGML_TYPE_Q8_0);
+    GGML_ASSERT(src1t == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    ggml_tensor_extra_cl_q8_0 * extra0_q8_0 = (ggml_tensor_extra_cl_q8_0 *)src0->extra;
+
+    GGML_ASSERT(src1->view_offs == 0);
+    GGML_ASSERT(dst->view_offs == 0);
+
+    const int  ne00 = src0->ne[0];
+    const int  ne01 = src0->ne[1];
+    const int  ne02 = src0->ne[2];
+
+    const int  ne10 = src1->ne[0];
+    const int  ne12 = src1->ne[2];
+
+    const int  ne0 = dst->ne[0];
+    const int  ne1 = dst->ne[1];
+
+    GGML_ASSERT(ne00 == ne10);
+    GGML_ASSERT((ne00 % 32) == 0);
+    GGML_ASSERT(ne0 == ne01);
+
+    cl_context context = backend_ctx->context;
+    cl_kernel kernel;
+
+    // init CL objects
+    cl_int              status;
+    cl_image_format     img_fmt_1d;
+    cl_image_desc       img_desc_1d;
+    cl_buffer_region    region;
+    cl_mem              A_image1d;
+    cl_mem              B_image1d;
+    cl_mem              B_sub_buffer;
+    cl_mem              S_image1d;
+
+    cl_mem              D_image1d;
+    cl_mem              D_sub_buffer;
+
+    int M = ne01;
+    int N = ne1;
+    int K = ne00;
+
+    // create an image for A
+    img_fmt_1d = { CL_R, CL_FLOAT};
+    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+    img_desc_1d.image_width = M * K / 4;    // Divide by 4 for char -> float
+    img_desc_1d.buffer = extra0_q8_0->q;
+    A_image1d = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &status);
+    CL_CHECK(status);
+
+    // create an image for Scale
+    img_fmt_1d = { CL_R, CL_HALF_FLOAT};
+    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+    img_desc_1d.image_width = M * K / 32;    // Block size is 32
+    img_desc_1d.buffer = extra0_q8_0->d;
+    S_image1d = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &status);
+    CL_CHECK(status);
+
+    // create a sub_buffer for B
+    region.origin = (extra1->offset); // + src1->view_offs);
+    region.size = K * N * sizeof(float);
+    B_sub_buffer = clCreateSubBuffer((extra1->data_device), 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
+    CL_CHECK(status);
+
+    // create an image for B from sub_buffer: RGBA (OCL)
+    img_fmt_1d = {CL_RGBA, CL_FLOAT};
+    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+    img_desc_1d.image_width = K * N / 4;
+    img_desc_1d.buffer = B_sub_buffer;
+    B_image1d = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &status);
+    CL_CHECK(status);
+
+    // Create subbuffer and image1d_buffer for dst
+    region.origin = (extrad->offset); // + dst->view_offs;
+    region.size = M * N * sizeof(float);
+    D_sub_buffer = clCreateSubBuffer((extrad->data_device), 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
+    CL_CHECK(status);
+
+    img_fmt_1d = {CL_R, CL_FLOAT};
+    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+    img_desc_1d.image_width = M * N;
+    img_desc_1d.buffer = D_sub_buffer;
+    D_image1d = clCreateImage(context, CL_MEM_WRITE_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &status);
+    CL_CHECK(status);
+
+    size_t local_work_size[3] = {1, 1, 1};
+    size_t global_work_size[3] = {1, 1, 1};
+
+    if (N == 1) {
+        kernel = backend_ctx->CL_mul_mat_vec_q8_0_f32;
+
+        int r2 = 1;
+        int r3 = 1;
+        cl_uint k_arg = 0;
+
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem),   &A_image1d));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem),   &extra0_q8_0->d));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem),   &B_image1d));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_ulong), &extra1->offset));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_ulong), &extrad->offset));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne00));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne01));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne02));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne10));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne12));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne0));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne1));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &r2));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &r3));
+
+        size_t wavesize = backend_ctx->adreno_wave_size;
+        local_work_size[0] = wavesize;
+        local_work_size[1] = 4; // reduce factor
+        local_work_size[2] = 1;
+
+        global_work_size[0] = ((M + wavesize - 1) / wavesize) * wavesize;
+        global_work_size[1] = 4; // reduce factor
+        global_work_size[2] = 1;
+    } else {
+        cl_ulong offsetd = extrad->offset + dst->view_offs;
+        cl_mem              B_image1d_trans = nullptr;
+        // for B transpose
+        cl_mem B_d = nullptr;
+        int padding;
+
+        //how many extra elements beyond multiple of 8
+        int extra_elements = N % 8;
+
+        //how much padding to add
+        padding = 0;
+        if (extra_elements > 0){
+            padding = 8 - extra_elements;
+        }
+
+        // Specify the starting offset (in bytes)
+        region.origin = 0;
+        // Specify the size of the sub-buffer (divide by 2 for FP16)
+        region.size = K * (N + padding) * sizeof(float)/2;
+        backend_ctx->prealloc_act_trans.allocate(context, region.size);
+        B_d = clCreateSubBuffer(
+            backend_ctx->prealloc_act_trans.buffer,
+            0,
+            CL_BUFFER_CREATE_TYPE_REGION,
+            &region,
+            &status);
+        CL_CHECK(status);
+
+        cl_image_format image_format_B_d_output = { CL_RGBA, CL_HALF_FLOAT }; //(CL_HALF_FLOAT for FP16)
+        cl_image_desc image_desc_B_d_output = {
+            CL_MEM_OBJECT_IMAGE1D_BUFFER,
+            static_cast<size_t>(K * (N + padding)/4),
+            0, 0, 0, 0, 0, 0, 0, { B_d }
+        };
+        B_image1d_trans = clCreateImage(
+            context,
+            0,
+            &image_format_B_d_output,
+            &image_desc_B_d_output,
+            NULL,
+            &status);
+        CL_CHECK(status);
+
+        int height_B = N/4;
+        if (height_B == 0) {
+            height_B = 1;
+        }
+        int width_B = K/4;
+        int padded_height_B = (N + padding)/4;
+
+        kernel = backend_ctx->kernel_transpose_32_16;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &B_image1d));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &B_image1d_trans));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),    &height_B));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int),    &width_B));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),    &padded_height_B));
+
+        size_t local_size_t[2] = { 1, 16 };
+        size_t global_size_t[2] = {
+            static_cast<size_t>(width_B),
+            static_cast<size_t>(padded_height_B)
+        };
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_size_t, local_size_t, dst);
+
+        kernel = backend_ctx->kernel_mul_mm_q8_0_f32_8x4;
+
+        int N_with_padding = N + padding;
+
+        CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q8_0->q));
+        CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q8_0->d));
+        CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &B_image1d_trans));
+        CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),      &K));
+        CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),      &M));
+        CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &N_with_padding));
+        CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &N));
+        CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &offsetd));
+
+        global_work_size[0] = (size_t)(N + 7) / 8;
+        global_work_size[1] = (size_t)(M + 3) / 4;
+        global_work_size[2] = 1;
+
+        local_work_size[0] = 2;
+        local_work_size[1] = 128;
+        local_work_size[2] = 1;
+    }
+
+    // enqueue kernel with profiling
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+    // deallocate sub buffers and images
+    CL_CHECK(clReleaseMemObject(A_image1d));
+    CL_CHECK(clReleaseMemObject(B_sub_buffer));
+    CL_CHECK(clReleaseMemObject(B_image1d));
+    CL_CHECK(clReleaseMemObject(S_image1d));
+    CL_CHECK(clReleaseMemObject(D_sub_buffer));
+    CL_CHECK(clReleaseMemObject(D_image1d));
+#else
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+#endif
+}
+
 static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -7549,6 +8355,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
     ggml_tensor_extra_cl_q4_0 * extra0_q4_0 = (ggml_tensor_extra_cl_q4_0 *)src0->extra;
     ggml_tensor_extra_cl_mxfp4 * extra0_mxfp4 = (ggml_tensor_extra_cl_mxfp4 *)src0->extra;
     ggml_tensor_extra_cl_q8_0 * extra0_q8_0 = (ggml_tensor_extra_cl_q8_0 *)src0->extra;
+    ggml_tensor_extra_cl_q6_K * extra0_q6_K = (ggml_tensor_extra_cl_q6_K *)src0->extra;
 #endif
 
     const int  ne00 = src0 ? src0->ne[0] : 0;
@@ -7591,9 +8398,12 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
     cl_context context = backend_ctx->context;
 
     if(src0t == GGML_TYPE_F16 && src1t == GGML_TYPE_F32){
-        if (ne01 >= 64 && ne1 >= 32 && ne00 >= 16 && (ne12 % ne02) == 0) {
+        if (ne01 >= 64 && ne1 >= 32 && ne00 >= 16 && (ne12 % ne02) == 0  &&
+            // dst is wrapped with image1d_buffer, the size limit applies, also src0
+            (ne0 * ne1 * dst->ne[2] * dst->nb[0] / 4 <= backend_ctx->image_max_buffer_size)) {
             // For KQ
             if (ggml_is_permuted(src0) && ggml_is_permuted(src1) &&
+                ((nb01 * ne01 / 4)/4 <= backend_ctx->image_max_buffer_size) &&
                 nb00 <= nb02 &&
                 nb02 <= nb01 &&
                 nb01 <= nb03 &&
@@ -7604,7 +8414,8 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
                 return;
             }
             // For KQV
-            if (!ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+            if (!ggml_is_contiguous(src0) && ggml_is_contiguous(src1) &&
+                ((nb02 * ne02 / 4)/4 <= backend_ctx->image_max_buffer_size)) {
                 ggml_cl_mul_mat_kq_kqv_adreno(backend, src0, src1, dst);
                 return;
             }
@@ -7636,6 +8447,13 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
     int padding;
     // <--------------------------------------------> //
 
+    // q8_0 x fp32
+    if (src0t == GGML_TYPE_Q8_0 && src1t == GGML_TYPE_F32 &&
+        enable_adreno_trans_weight(backend_ctx, src0)) {
+            ggml_cl_mul_mat_q8_0_f32_adreno(backend, src0, src1, dst);
+            return;
+    }
+
     // q4_0 x fp32
     if(src0t == GGML_TYPE_Q4_0 && src1t == GGML_TYPE_F32) {
         // TODO: remove duplicate definitions of image description + format -- move to top
@@ -7910,9 +8728,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
 
     // GEMM using local memory
     // Current BK = 16, so ne00 % 16 == 0
-    if (ggml_is_contiguous(src0) &&
-        ggml_is_contiguous(src1) &&
-        src1t == GGML_TYPE_F32 &&
+    if (src1t == GGML_TYPE_F32 &&
         ne00 % 16 == 0 &&
         ne11 > 1) {
         switch(src0t) {
@@ -7924,10 +8740,42 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
                 int batch_stride_b = ne10*ne11;
                 int batch_stride_d = ne0*ne1;
 
-                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
-                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
-                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
-                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+                cl_mem mem_src0 = extra0->data_device;
+                cl_mem mem_src1 = extra1->data_device;
+
+                cl_ulong nb00_cont = nb00;
+                cl_ulong nb01_cont = nb01;
+                cl_ulong nb02_cont = nb02;
+                cl_ulong nb03_cont = nb03;
+
+                cl_ulong nb10_cont = nb10;
+                cl_ulong nb11_cont = nb11;
+                cl_ulong nb12_cont = nb12;
+                cl_ulong nb13_cont = nb13;
+
+                cl_ulong offset0_cont = offset0;
+                cl_ulong offset1_cont = offset1;
+
+                if (!ggml_is_contiguous(src0)) {
+                    backend_ctx->prealloc_src0.allocate(backend_ctx->context, ggml_nbytes(src0));
+                    ggml_cl_copy_to_contiguous(backend, src0, backend_ctx->prealloc_src0.buffer,
+                        nb00_cont, nb01_cont, nb02_cont, nb03_cont);
+                    mem_src0 = backend_ctx->prealloc_src0.buffer;
+                    offset0_cont = 0;
+                }
+
+                if (!ggml_is_contiguous(src1)) {
+                    backend_ctx->prealloc_src1.allocate(backend_ctx->context, ggml_nbytes(src1));
+                    ggml_cl_copy_to_contiguous(backend, src1, backend_ctx->prealloc_src1.buffer,
+                        nb10_cont, nb11_cont, nb12_cont, nb13_cont);
+                    mem_src1 = backend_ctx->prealloc_src1.buffer;
+                    offset1_cont = 0;
+                }
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &mem_src0));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0_cont));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &mem_src1));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1_cont));
                 CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
                 CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
                 CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
@@ -7959,10 +8807,42 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
                 int batch_stride_b = ne10*ne11;
                 int batch_stride_d = ne0*ne1;
 
-                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
-                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
-                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
-                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+                cl_mem mem_src0 = extra0->data_device;
+                cl_mem mem_src1 = extra1->data_device;
+
+                cl_ulong nb00_cont = nb00;
+                cl_ulong nb01_cont = nb01;
+                cl_ulong nb02_cont = nb02;
+                cl_ulong nb03_cont = nb03;
+
+                cl_ulong nb10_cont = nb10;
+                cl_ulong nb11_cont = nb11;
+                cl_ulong nb12_cont = nb12;
+                cl_ulong nb13_cont = nb13;
+
+                cl_ulong offset0_cont = offset0;
+                cl_ulong offset1_cont = offset1;
+
+                if (!ggml_is_contiguous(src0)) {
+                    backend_ctx->prealloc_src0.allocate(backend_ctx->context, ggml_nbytes(src0));
+                    ggml_cl_copy_to_contiguous(backend, src0, backend_ctx->prealloc_src0.buffer,
+                        nb00_cont, nb01_cont, nb02_cont, nb03_cont);
+                    mem_src0 = backend_ctx->prealloc_src0.buffer;
+                    offset0_cont = 0;
+                }
+
+                if (!ggml_is_contiguous(src1)) {
+                    backend_ctx->prealloc_src1.allocate(backend_ctx->context, ggml_nbytes(src1));
+                    ggml_cl_copy_to_contiguous(backend, src1, backend_ctx->prealloc_src1.buffer,
+                            nb10_cont, nb11_cont, nb12_cont, nb13_cont);
+                    mem_src1 = backend_ctx->prealloc_src1.buffer;
+                    offset1_cont = 0;
+                }
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &mem_src0));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0_cont));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &mem_src1));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1_cont));
                 CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
                 CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
                 CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
@@ -7990,6 +8870,10 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
                 if (ne11 < 32) {
                     break;
                 }
+                if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1)) {
+                    break;
+                }
+
                 kernel = backend_ctx->kernel_mul_mm_q8_0_f32_l4_lm;
                 nth0 = 128; // calculated as (BM*BN)/(TM*TN)
 
@@ -8362,14 +9246,49 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+#ifdef GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_q6_K_f32_flat;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = 4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = 4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q6_K->ql));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q6_K->qh));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra0_q6_K->s));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_mem),   &extra0_q6_K->d));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &r3));
+#else
             kernel = backend_ctx->kernel_mul_mv_q6_K_f32;
 
             if (backend_ctx->gpu_family == INTEL) {
-                nth0 = 2;
-                nth1 = 16;
+                nth0 = 16;
+                nth1 = 2;
+                ndst = 1;
             } else if (backend_ctx->gpu_family == ADRENO) {
-                nth0 = 2;
-                nth1 = 64;
+                nth0 = 64;
+                nth1 = 2;
+                ndst = 1;
             } else {
                 GGML_ASSERT(false && "TODO: Unknown GPU");
             }
@@ -8389,6 +9308,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
             CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne1));
             CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &r2));
             CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &r3));
+#endif // GGML_OPENCL_SOA_Q
             break;
         case GGML_TYPE_MXFP4: {
 #ifdef GGML_OPENCL_SOA_Q
@@ -8491,7 +9411,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
     } else if (src0t == GGML_TYPE_Q5_K) {
         GGML_ASSERT(false && "not implemented");
     } else if (src0t == GGML_TYPE_Q6_K) {
-        size_t global_work_size[] = {(size_t)(ne01+1)/2*nth0, (size_t)ne11*nth1, (size_t)ne12*ne13};
+        size_t global_work_size[] = {(size_t)(ne01+ndst*nth1-1)/(ndst*nth1)*nth0, (size_t)ne11*nth1, (size_t)ne12*ne13};
         size_t local_work_size[] = {(size_t)nth0, (size_t)nth1, 1};
 
         backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
@@ -9909,6 +10829,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_glu;
             break;
+        case GGML_OP_TRI:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_tri;
+            break;
         case GGML_OP_FILL:
             if (!any_on_device) {
                 return false;
diff --git a/ggml/src/ggml-opencl/kernels/cvt.cl b/ggml/src/ggml-opencl/kernels/cvt.cl
index 513a4d3e28..9fb434713d 100644
--- a/ggml/src/ggml-opencl/kernels/cvt.cl
+++ b/ggml/src/ggml-opencl/kernels/cvt.cl
@@ -46,6 +46,16 @@ struct block_q4_0
     uint8_t qs[QK4_0 / 2];
 };
 
+//------------------------------------------------------------------------------
+// block_q6_K
+//------------------------------------------------------------------------------
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    half d;                  // super-block scale
+};
+
 //------------------------------------------------------------------------------
 // kernel_convert_block_q4_0
 // Convert the block_q4_0 format to 2 separate arrays (AOS -> SOA).
@@ -263,3 +273,94 @@ kernel void kernel_restore_block_q8_0(
         b->qs[i] = q[i];
     }
 }
+
+kernel void kernel_restore_block_q8_0_trans(
+    global uchar * src_q,
+    global half  * src_d,
+    global block_q8_0 * dst,
+    uint ne00,
+    uint ne01
+){
+    uint num_blk_per_row = ne00 / QK8_0;
+
+    global block_q8_0 * b = (global block_q8_0 *) dst + get_global_id(0) * num_blk_per_row;
+    global uchar      * q = (global uchar *) src_q + get_global_id(0) * 4; // 4 8-bit packed
+    global half       * d = (global half *) src_d + get_global_id(0);
+
+    for (uint blk = 0; blk < num_blk_per_row; blk++) {
+        b->d = *d;
+
+        for (uint i = 0; i < QK8_0; i+=4) {
+            b->qs[i]   = q[0];
+            b->qs[i+1] = q[1];
+            b->qs[i+2] = q[2];
+            b->qs[i+3] = q[3];
+
+            q += 4 * ne01; // M stride
+        }
+
+        d += ne01;
+
+        b++;
+    }
+}
+
+//------------------------------------------------------------------------------
+// kernel_convert_block_q6_K
+// Convert the block_q6_K format to 3 separate arrays (AOS -> SOA).
+// This kernel does not deshuffle the bits.
+// Each thread processes a super block.
+//------------------------------------------------------------------------------
+kernel void kernel_convert_block_q6_K(
+    global struct block_q6_K * src0,
+    global uchar * dst_ql,
+    global uchar * dst_qh,
+    global char  * dst_s,
+    global half  * dst_d
+) {
+    global struct block_q6_K * b = (global struct block_q6_K *) src0 + get_global_id(0);
+    global uchar * ql = (global uchar *) dst_ql + QK_K/2*get_global_id(0);
+    global uchar * qh = (global uchar *) dst_qh + QK_K/4*get_global_id(0);
+    global char  * s  = (global char  *) dst_s  + QK_K/16*get_global_id(0);
+    global half  * d  = (global half  *) dst_d  + get_global_id(0);
+
+    *d = b->d;
+
+    for (int i = 0; i < QK_K/2; ++i) {
+        ql[i] = b->ql[i];
+    }
+    for (int i = 0; i < QK_K/4; ++i) {
+        qh[i] = b->qh[i];
+    }
+    for (int i = 0; i < QK_K/16; ++i) {
+        s[i] = b->scales[i];
+    }
+}
+
+// Restore block_q6_K from flattened arrays.
+// Each thread processes a super block.
+kernel void kernel_restore_block_q6_K(
+    global uchar * dst_ql,
+    global uchar * dst_qh,
+    global char  * dst_s,
+    global half  * dst_d,
+    global struct block_q6_K * dst
+) {
+    global struct block_q6_K * b = (global struct block_q6_K *) dst + get_global_id(0);
+    global uchar * ql = (global uchar *) dst_ql + QK_K/2*get_global_id(0);
+    global uchar * qh = (global uchar *) dst_qh + QK_K/4*get_global_id(0);
+    global char  * s  = (global char  *) dst_s  + QK_K/16*get_global_id(0);
+    global half  * d  = (global half  *) dst_d  + get_global_id(0);
+
+    b->d = *d;
+
+    for (int i = 0; i < QK_K/2; ++i) {
+        b->ql[i] = ql[i];
+    }
+    for (int i = 0; i < QK_K/4; ++i) {
+        b->qh[i] = qh[i];
+    }
+    for (int i = 0; i < QK_K/16; ++i) {
+        b->scales[i] = s[i];
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general_q8_0_f32.cl b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general_q8_0_f32.cl
new file mode 100644
index 0000000000..f944ef3a99
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general_q8_0_f32.cl
@@ -0,0 +1,195 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#endif
+
+#define QK8_0 32
+#define N_SIMDGROUP 4
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1(total_sums, bits8, scale, y) \
+    float shared_y; \
+    char elem; \
+                                             \
+    shared_y = sub_group_broadcast(y.s0, 0); \
+    elem = (char)(bits8.s0 & 0x000000FF); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 0); \
+    elem = (char)((bits8.s0 & 0x0000FF00) >> 8); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 0); \
+    elem = (char)((bits8.s0 & 0x00FF0000) >> 16); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 0); \
+    elem = (char)((bits8.s0 & 0xFF000000) >> 24); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+                                             \
+    shared_y = sub_group_broadcast(y.s4, 0); \
+    elem = (char)(bits8.s1 & 0x000000FF); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 0); \
+    elem = (char)((bits8.s1 & 0x0000FF00) >> 8); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 0); \
+    elem = (char)((bits8.s1 & 0x00FF0000) >> 16); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 0); \
+    elem = (char)((bits8.s1 & 0xFF000000) >> 24); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+                                             \
+    shared_y = sub_group_broadcast(y.s0, 1); \
+    elem = (char)(bits8.s2 & 0x000000FF); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 1); \
+    elem = (char)((bits8.s2 & 0x0000FF00) >> 8); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 1); \
+    elem = (char)((bits8.s2 & 0x00FF0000) >> 16); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 1); \
+    elem = (char)((bits8.s2 & 0xFF000000) >> 24); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+                                             \
+    shared_y = sub_group_broadcast(y.s4, 1); \
+    elem = (char)(bits8.s3 & 0x000000FF); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 1); \
+    elem = (char)((bits8.s3 & 0x0000FF00) >> 8); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 1); \
+    elem = (char)((bits8.s3 & 0x00FF0000) >> 16); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 1); \
+    elem = (char)((bits8.s3 & 0xFF000000) >> 24); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+                                             \
+    shared_y = sub_group_broadcast(y.s0, 2); \
+    elem = (char)(bits8.s4 & 0x000000FF); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 2); \
+    elem = (char)((bits8.s4 & 0x0000FF00) >> 8); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 2); \
+    elem = (char)((bits8.s4 & 0x00FF0000) >> 16); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 2); \
+    elem = (char)((bits8.s4 & 0xFF000000) >> 24); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+                                             \
+    shared_y = sub_group_broadcast(y.s4, 2); \
+    elem = (char)(bits8.s5 & 0x000000FF); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 2); \
+    elem = (char)((bits8.s5 & 0x0000FF00) >> 8); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 2); \
+    elem = (char)((bits8.s5 & 0x00FF0000) >> 16); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 2); \
+    elem = (char)((bits8.s5 & 0xFF000000) >> 24); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+                                             \
+    shared_y = sub_group_broadcast(y.s0, 3); \
+    elem = (char)(bits8.s6 & 0x000000FF); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 3); \
+    elem = (char)((bits8.s6 & 0x0000FF00) >> 8); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 3); \
+    elem = (char)((bits8.s6 & 0x00FF0000) >> 16); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 3); \
+    elem = (char)((bits8.s6 & 0xFF000000) >> 24); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+                                             \
+    shared_y = sub_group_broadcast(y.s4, 3); \
+    elem = (char)(bits8.s7 & 0x000000FF); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 3); \
+    elem = (char)((bits8.s7 & 0x0000FF00) >> 8); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 3); \
+    elem = (char)((bits8.s7 & 0x00FF0000) >> 16); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 3); \
+    elem = (char)((bits8.s7 & 0xFF000000) >> 24); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+__kernel void kernel_gemv_noshuffle(
+        __read_only  image1d_buffer_t src0_q,  // quantized A
+        global half  * src0_d,  // A scales
+        __read_only  image1d_buffer_t src1,    // B
+        ulong offset1,            // offset to B (0)
+        global float * dst,     // C
+        ulong offsetd,            // offset to C
+        int ne00,               // K
+        int ne01,               // M
+        int ne02,               // 1
+        int ne10,               // K
+        int ne12,               // 1
+        int ne0,                // M
+        int ne1,                // N
+        int r2,                 // 1
+        int r3)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid    = get_sub_group_local_id();
+
+    uint K = ne00;
+    uint M = ne01;
+
+    uint LINE_STRIDE_A = M;
+    uint BLOCK_STRIDE_A = 8 * M;   // 32 / 4 = 8
+
+    __private uint8     regA;
+    __private half      regS;
+    __private float8    regB;
+
+    __private float totalSum = (float)(0.0f);
+
+    // loop along K in block granularity, skip 4 blocks every iter
+    #pragma unroll 1 /* tell compiler not to unroll */
+    for (uint k = groupId; k < (K / QK8_0); k += N_SIMDGROUP) {
+        regS = src0_d[gid + k * LINE_STRIDE_A]; // each fiber loads scale of one rows
+        // first 4 fibers in each wave load 8 B values to its private scope
+        if (slid < 4) {
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
+        }
+
+        // load weights for one block in consecutive rows
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+        regA.s4 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA.s5 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA.s6 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA.s7 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+
+        dequantizeBlockAccum_ns_sgbroadcast_1(totalSum, regA, regS, regB);
+    }
+
+    // reduction in local memory, assumes #wave=4
+    __local float reduceLM[SIMDGROUP_WIDTH * 3];
+    if (groupId == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = totalSum;
+    if (groupId == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = totalSum;
+    if (groupId == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = totalSum;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 0 + slid];
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 1 + slid];
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 2 + slid];
+
+    // 1 outputs per fiber in wave 0
+    if (groupId == 0) {
+        dst = (global float*)((global char*)dst + offsetd);
+        dst[gid] = totalSum;
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_8x4.cl b/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_8x4.cl
new file mode 100644
index 0000000000..51ce2121ce
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_8x4.cl
@@ -0,0 +1,129 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_128
+#endif
+
+kernel void kernel_mul_mm_q8_0_f32_8x4(
+        global const uint * src0_q,
+        global const half  * src0_d,
+        __read_only image1d_buffer_t src1,
+        global float * dst,
+        int k,
+        int m,
+        int n,
+        int n_no_padding,
+        ulong offsetd
+) {
+
+    int m_4 = m >> 2;
+    int n_4 = n >> 2;
+
+    int gy   = get_global_id(0);
+    int gx   = get_global_id(1);
+    int gx_2 = gx << 2;
+    dst  = (global float *)((global char*)dst  + offsetd);
+
+
+    half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0;
+    half8 B;
+    half4 deq;
+
+    __global const uint* wptr = src0_q + gx_2;
+    __global const half* sptr = src0_d + gx_2;
+
+      for (int i = 0; i < k; i += 4) {
+        uint4 pack4 = vload4(0, wptr + (i / 4) * m);
+        half4 scale = vload4(0, sptr + (i / 32) * m);
+
+        char4 p0 = as_char4(pack4.s0);
+        char4 p1 = as_char4(pack4.s1);
+        char4 p2 = as_char4(pack4.s2);
+        char4 p3 = as_char4(pack4.s3);
+
+        // ------------------- j = 0 (k = i+0) -------------------
+        B.s0123 = read_imageh(src1, gy * 2 + (i + 0) * n_4);
+        B.s4567 = read_imageh(src1, gy * 2 + (i + 0) * n_4 + 1);
+
+        half4 wj0 = convert_half4((char4)(p0.s0, p1.s0, p2.s0, p3.s0)) * scale;
+
+        c0 += B * wj0.s0;
+        c1 += B * wj0.s1;
+        c2 += B * wj0.s2;
+        c3 += B * wj0.s3;
+
+        // ------------------- j = 1 (k = i+1) -------------------
+        B.s0123 = read_imageh(src1, gy * 2 + (i + 1) * n_4);
+        B.s4567 = read_imageh(src1, gy * 2 + (i + 1) * n_4 + 1);
+
+        half4 wj1 = convert_half4((char4)(p0.s1, p1.s1, p2.s1, p3.s1)) * scale;
+
+        c0 += B * wj1.s0;
+        c1 += B * wj1.s1;
+        c2 += B * wj1.s2;
+        c3 += B * wj1.s3;
+
+        // ------------------- j = 2 (k = i+2) -------------------
+        B.s0123 = read_imageh(src1, gy * 2 + (i + 2) * n_4);
+        B.s4567 = read_imageh(src1, gy * 2 + (i + 2) * n_4 + 1);
+
+        half4 wj2 = convert_half4((char4)(p0.s2, p1.s2, p2.s2, p3.s2)) * scale;
+
+        c0 += B * wj2.s0;
+        c1 += B * wj2.s1;
+        c2 += B * wj2.s2;
+        c3 += B * wj2.s3;
+
+        // ------------------- j = 3 (k = i+3) -------------------
+        B.s0123 = read_imageh(src1, gy * 2 + (i + 3) * n_4);
+        B.s4567 = read_imageh(src1, gy * 2 + (i + 3) * n_4 + 1);
+
+        half4 wj3 = convert_half4((char4)(p0.s3, p1.s3, p2.s3, p3.s3)) * scale;
+
+        c0 += B * wj3.s0;
+        c1 += B * wj3.s1;
+        c2 += B * wj3.s2;
+        c3 += B * wj3.s3;
+    }
+
+    int idx = (gy << 3) * m + (gx << 2);
+
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/mul_mv_q6_k.cl b/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl
similarity index 99%
rename from ggml/src/ggml-opencl/kernels/mul_mv_q6_k.cl
rename to ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl
index 8a17b9aae6..819e5192e3 100644
--- a/ggml/src/ggml-opencl/kernels/mul_mv_q6_k.cl
+++ b/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl
@@ -111,6 +111,10 @@ kernel void kernel_mul_mv_q6_K_f32(
 
     int row = N_SIMDGROUP * r0 + get_sub_group_id();
 
+    if (row >= ne01) {
+        return;
+    }
+
     int i12 = im%ne12;
     int i13 = im/ne12;
 
diff --git a/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl b/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl
new file mode 100644
index 0000000000..86fe09c6dd
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl
@@ -0,0 +1,194 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+//------------------------------------------------------------------------------
+// kernel_mul_mv_q6_K_f32_flat
+//------------------------------------------------------------------------------
+#define Q6_K_MASK1 0x03
+#define Q6_K_MASK2 0x0C
+#define Q6_K_MASK3 0x30
+#define Q6_K_MASK4 0xC0
+
+#define QK_K       256
+
+inline float block_q_6_K_dot_y_flat(
+    global uchar * blk_ql,
+    global uchar * blk_qh,
+    global char  * blk_scales,
+    global half  * blk_d,
+    global float * yy,
+    int ib,
+    int ip,
+    int is,
+    int l0
+) {
+    int y_offset   = 128*ip + l0;
+    int q_offset_l =  64*ip + l0;
+    int q_offset_h =  32*ip + l0;
+
+    global uchar * q1 = blk_ql     + ib*128 + q_offset_l;
+    global uchar * q2 = q1         + QK_K/8;
+    global uchar * qh = blk_qh     + ib*64 + q_offset_h;
+    global char  * sc = blk_scales + ib*16 + is;
+
+    global float * y = yy + ib * QK_K + y_offset;
+
+    float dall = blk_d[ib];
+
+    float  sumf = 0;
+    float4 sums = {0.f, 0.f, 0.f, 0.f};
+
+    sums.s0 += y[0+ 0] * ((float)((q1[0] & 0xF) | ((qh[0] & Q6_K_MASK1) << 4)) - 32.f);
+    sums.s1 += y[0+32] * ((float)((q2[0] & 0xF) | ((qh[0] & Q6_K_MASK2) << 2)) - 32.f);
+    sums.s2 += y[0+64] * ((float)((q1[0]  >> 4) | ((qh[0] & Q6_K_MASK3) << 0)) - 32.f);
+    sums.s3 += y[0+96] * ((float)((q2[0]  >> 4) | ((qh[0] & Q6_K_MASK4) >> 2)) - 32.f);
+
+    sums.s0 += y[1+ 0] * ((float)((q1[1] & 0xF) | ((qh[1] & Q6_K_MASK1) << 4)) - 32.f);
+    sums.s1 += y[1+32] * ((float)((q2[1] & 0xF) | ((qh[1] & Q6_K_MASK2) << 2)) - 32.f);
+    sums.s2 += y[1+64] * ((float)((q1[1]  >> 4) | ((qh[1] & Q6_K_MASK3) << 0)) - 32.f);
+    sums.s3 += y[1+96] * ((float)((q2[1]  >> 4) | ((qh[1] & Q6_K_MASK4) >> 2)) - 32.f);
+
+    sums.s0 += y[2+ 0] * ((float)((q1[2] & 0xF) | ((qh[2] & Q6_K_MASK1) << 4)) - 32.f);
+    sums.s1 += y[2+32] * ((float)((q2[2] & 0xF) | ((qh[2] & Q6_K_MASK2) << 2)) - 32.f);
+    sums.s2 += y[2+64] * ((float)((q1[2]  >> 4) | ((qh[2] & Q6_K_MASK3) << 0)) - 32.f);
+    sums.s3 += y[2+96] * ((float)((q2[2]  >> 4) | ((qh[2] & Q6_K_MASK4) >> 2)) - 32.f);
+
+    sums.s0 += y[3+ 0] * ((float)((q1[3] & 0xF) | ((qh[3] & Q6_K_MASK1) << 4)) - 32.f);
+    sums.s1 += y[3+32] * ((float)((q2[3] & 0xF) | ((qh[3] & Q6_K_MASK2) << 2)) - 32.f);
+    sums.s2 += y[3+64] * ((float)((q1[3]  >> 4) | ((qh[3] & Q6_K_MASK3) << 0)) - 32.f);
+    sums.s3 += y[3+96] * ((float)((q2[3]  >> 4) | ((qh[3] & Q6_K_MASK4) >> 2)) - 32.f);
+
+    sumf += dall * (sums.s0 * sc[0] + sums.s1 * sc[2] + sums.s2 * sc[4] + sums.s3 * sc[6]);
+
+    return sumf;
+}
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4
+#define N_SIMDGROUP 2
+#define N_SIMDWIDTH 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 2
+#define N_SIMDWIDTH 64
+#endif
+
+#define BLOCK_STRIDE (N_SIMDWIDTH/16) // number of blocks each subgroup processes
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q6_K_f32_flat(
+        global uchar * src0_ql,
+        global uchar * src0_qh,
+        global char  * src0_s,
+        global half  * src0_d,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int nb = ne00/QK_K;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    int first_row = (N_SIMDGROUP * r0 + get_sub_group_id()) * N_DST;
+
+    ulong offset_src0    = first_row*nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    ulong offset_src0_ql = offset_src0 * 128;
+    ulong offset_src0_qh = offset_src0 * 64;
+    ulong offset_src0_s  = offset_src0 * 16;
+    ulong offset_src0_d  = offset_src0;
+
+    global uchar * blk_ql     = (global uchar *) src0_ql + offset_src0_ql;
+    global uchar * blk_qh     = (global uchar *) src0_qh + offset_src0_qh;
+    global char  * blk_scales = (global char  *) src0_s  + offset_src0_s;
+    global half  * blk_d      = (global half  *) src0_d  + offset_src0_d;
+    global float * yy         = (global float *) src1    + r1*ne10 + im*ne00*ne1;
+
+    int tid = get_sub_group_local_id()/BLOCK_STRIDE; // first block_stride groups have tid=0
+    int ix  = get_sub_group_local_id()%BLOCK_STRIDE; // first block is 0..block_stride-1
+    int ip  = tid/8;   // first or second half of (super) block (0 or 1)
+    int il  = tid%8;   // each half has 8 parts, one per scale
+    int n   = 4;       // 4 scales at a time (and 4 sums)
+    int l0  = n*il;    // offset into half-block, 0..28
+    int is  = 8*ip + l0/16; // 0, 1, 8, 9
+
+    float4 sumf = 0;
+
+    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
+        if (first_row + 0 < ne01) {
+            sumf.s0 += block_q_6_K_dot_y_flat(blk_ql + 0*nb*128, blk_qh + 0*nb*64, blk_scales + 0*nb*16, blk_d + 0*nb, yy, ib, ip, is, l0);
+        }
+        if (first_row + 1 < ne01) {
+            sumf.s1 += block_q_6_K_dot_y_flat(blk_ql + 1*nb*128, blk_qh + 1*nb*64, blk_scales + 1*nb*16, blk_d + 1*nb, yy, ib, ip, is, l0);
+        }
+        if (first_row + 2 < ne01) {
+            sumf.s2 += block_q_6_K_dot_y_flat(blk_ql + 2*nb*128, blk_qh + 2*nb*64, blk_scales + 2*nb*16, blk_d + 2*nb, yy, ib, ip, is, l0);
+        }
+        if (first_row + 3 < ne01) {
+            sumf.s3 += block_q_6_K_dot_y_flat(blk_ql + 3*nb*128, blk_qh + 3*nb*64, blk_scales + 3*nb*16, blk_d + 3*nb, yy, ib, ip, is, l0);
+        }
+    }
+
+    float4 tot = (float4)(
+        sub_group_reduce_add(sumf.s0),
+        sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2),
+        sub_group_reduce_add(sumf.s3)
+    );
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/tri.cl b/ggml/src/ggml-opencl/kernels/tri.cl
new file mode 100644
index 0000000000..35cdd543bc
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/tri.cl
@@ -0,0 +1,32 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// tri
+//------------------------------------------------------------------------------
+__kernel void kernel_tri_f32(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int n,
+        int ne0,
+        int ne1,
+        int tri_type
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int idx = get_global_id(0);
+    if (idx >= n) return;
+
+    int i0 = idx % ne0;
+    int i1 = (idx / ne0) % ne1;
+
+    int keep = 0;
+    if (tri_type == 0) keep = (i0 >= i1);
+    else if (tri_type == 1) keep = (i0 >  i1);
+    else if (tri_type == 2) keep = (i0 <= i1);
+    else                    keep = (i0 <  i1);
+
+    dst[idx] = keep ? src0[idx] : 0.0f;
+}
diff --git a/ggml/src/ggml-sycl/dpct/helper.hpp b/ggml/src/ggml-sycl/dpct/helper.hpp
index 30ec1e8daf..8ae8098717 100644
--- a/ggml/src/ggml-sycl/dpct/helper.hpp
+++ b/ggml/src/ggml-sycl/dpct/helper.hpp
@@ -15,7 +15,6 @@
 
 #include <sycl/sycl.hpp>
 #include <sycl/half_type.hpp>
-#include <syclcompat/math.hpp>
 #include <map>
 
 #ifdef GGML_SYCL_USE_INTEL_ONEMKL
diff --git a/ggml/src/ggml-sycl/element_wise.cpp b/ggml/src/ggml-sycl/element_wise.cpp
index 8d83b2446b..651b875b63 100644
--- a/ggml/src/ggml-sycl/element_wise.cpp
+++ b/ggml/src/ggml-sycl/element_wise.cpp
@@ -123,6 +123,15 @@ static __dpct_inline__ T op_log(T x) {
     return sycl::log(x);
 }
 
+template<typename T>
+static __dpct_inline__ T op_softplus(T x) {
+    const float xf = (float) x;
+    const float ax = sycl::fabs(xf);
+    const float m  = sycl::fmax(xf, 0.0f);
+    const float y  = m + sycl::log1p(sycl::exp(-ax));
+    return (T) y;
+}
+
 template<typename T>
 static __dpct_inline__ T op_neg(T x) {
     return -x;
@@ -695,6 +704,12 @@ static inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, ggml_tensor
         });
 }
 
+static inline void ggml_sycl_op_softplus(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
+        return op_softplus(x);
+    });
+}
+
 static inline void ggml_sycl_op_neg(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {
         return op_neg(x);
@@ -1101,6 +1116,11 @@ void ggml_sycl_log(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     ggml_sycl_op_log(ctx, dst);
 }
 
+void ggml_sycl_softplus(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_softplus(ctx, dst);
+}
+
 void ggml_sycl_neg(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
     ggml_sycl_op_neg(ctx, dst);
diff --git a/ggml/src/ggml-sycl/element_wise.hpp b/ggml/src/ggml-sycl/element_wise.hpp
index 0913a2e529..7c71974687 100644
--- a/ggml/src/ggml-sycl/element_wise.hpp
+++ b/ggml/src/ggml-sycl/element_wise.hpp
@@ -61,6 +61,8 @@ void ggml_sycl_exp(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
 void ggml_sycl_log(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
+void ggml_sycl_softplus(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
 void ggml_sycl_neg(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
 void ggml_sycl_step(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
index 8f8176b678..74b4ed91cc 100644
--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -1157,13 +1157,28 @@ static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_
     GGML_UNUSED(buft);
 }
 
+inline void * aligned_malloc_host(size_t alignment, size_t size) {
+#ifdef _WIN32
+    return _aligned_malloc(size, alignment);
+#else
+    return aligned_alloc(alignment, size);
+#endif
+}
+
+inline void free_aligned_mem_host(void * memblock) {
+#ifdef _WIN32
+    _aligned_free(memblock);
+#else
+    free(memblock);
+#endif
+}
+
 static void ggml_backend_sycl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    ggml_sycl_host_free(buffer->context);
+    free_aligned_mem_host((void *)buffer->context);
 }
 
 static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    void * ptr = ggml_sycl_host_malloc(size);
-
+    void * ptr = aligned_malloc_host(TENSOR_ALIGNMENT, size);
     if (ptr == nullptr) {
         // fallback to cpu buffer
         return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
@@ -2248,6 +2263,65 @@ inline void ggml_sycl_op_diag_mask_inf(ggml_backend_sycl_context & ctx, ggml_ten
     diag_mask_inf_f32_sycl(src0_dd, dst_dd, ne00, nrows0, ne01, n_past, main_stream);
 }
 
+static void tri_f32_sycl(
+    const float * src,
+    float * dst,
+    const int64_t ne0,
+    const int64_t ne1,
+    const int64_t ne2,
+    const int64_t ne3,
+    const ggml_tri_type ttype,
+    dpct::queue_ptr main_stream
+) {
+    const size_t total = (size_t) ne0 * (size_t) ne1 * (size_t) ne2 * (size_t) ne3;
+
+    main_stream->parallel_for(sycl::range<1>(total), [=](sycl::id<1> tid) {
+        const int64_t idx = (int64_t) tid[0];
+
+        const int64_t i0 = idx % ne0;
+        const int64_t t1 = idx / ne0;
+        const int64_t i1 = t1 % ne1;
+
+        bool keep = false;
+        switch (ttype) {
+            case GGML_TRI_TYPE_LOWER:      keep = (i0 <  i1); break;
+            case GGML_TRI_TYPE_LOWER_DIAG: keep = (i0 <= i1); break;
+            case GGML_TRI_TYPE_UPPER:      keep = (i0 >  i1); break;
+            case GGML_TRI_TYPE_UPPER_DIAG: keep = (i0 >= i1); break;
+            default: keep = false; break;
+        }
+
+        dst[idx] = keep ? src[idx] : 0.0f;
+    });
+}
+
+static void ggml_sycl_op_tri(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    GGML_ASSERT(src0);
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+
+    dpct::queue_ptr main_stream = ctx.stream();
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+
+    const float * src0_dd = static_cast<const float *>(src0->data);
+    float *       dst_dd  = static_cast<float *>(dst->data);
+
+    const ggml_tri_type ttype = (ggml_tri_type) ggml_get_op_params_i32(dst, 0);
+
+    const int64_t ne0 = src0->ne[0];
+    const int64_t ne1 = src0->ne[1];
+    const int64_t ne2 = src0->ne[2];
+    const int64_t ne3 = src0->ne[3];
+
+    tri_f32_sycl(src0_dd, dst_dd, ne0, ne1, ne2, ne3, ttype, main_stream);
+}
+
+
 inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
@@ -3771,6 +3845,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
                 case GGML_UNARY_OP_EXP:
                     ggml_sycl_exp(ctx, dst);
                     break;
+                case GGML_UNARY_OP_SOFTPLUS:
+                    ggml_sycl_softplus(ctx, dst);
+                    break;
                 case GGML_UNARY_OP_SGN:
                     ggml_sycl_sgn(ctx, dst);
                     break;
@@ -3897,6 +3974,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
         case GGML_OP_TRANSPOSE:
             GGML_SYCL_DEBUG("%s: Tensor NO-OP\n", __func__);
             break;
+        case GGML_OP_TRI:
+            ggml_sycl_op_tri(ctx, dst);
+            break;
         case GGML_OP_DIAG_MASK_INF:
             ggml_sycl_diag_mask_inf(ctx, dst);
             break;
@@ -4109,6 +4189,9 @@ static void ggml_backend_sycl_graph_compute_impl(ggml_backend_sycl_context * syc
         if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
             continue;
         }
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
 #ifndef NDEBUG
         assert(node->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device));
         for (int j = 0; j < GGML_MAX_SRC; j++) {
@@ -4386,6 +4469,7 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_GELU_ERF:
                 case GGML_UNARY_OP_EXP:
+                case GGML_UNARY_OP_SOFTPLUS:
                 case GGML_UNARY_OP_ELU:
                     return true;
                 case GGML_UNARY_OP_FLOOR:
@@ -4588,18 +4672,23 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
             return (op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32) && (op->type == op->src[0]->type);
 #endif
         case GGML_OP_NORM:
-            return true;
         case GGML_OP_L2_NORM:
         case GGML_OP_GROUP_NORM:
-            return ggml_is_contiguous(op->src[0]);
         case GGML_OP_RMS_NORM:
-            return ((op->src[0]->ne[0] % WARP_SIZE) == 0);
+            return true;
         case GGML_OP_RMS_NORM_BACK:
-            return ((op->src[0]->ne[0] % WARP_SIZE) == 0);
+            return ggml_is_contiguous(op->src[0]);
         case GGML_OP_SCALE:
             return true;
         case GGML_OP_CONT:
             return op->src[0]->type != GGML_TYPE_BF16;
+        case GGML_OP_TRI:
+            {
+                const ggml_tensor * src0 = op->src[0];
+                return src0 &&
+                       op->type == GGML_TYPE_F32 &&
+                       ggml_is_contiguous(src0);
+            }
         case GGML_OP_DIAG_MASK_INF:
             return true;
         case GGML_OP_SOFT_MAX:
diff --git a/ggml/src/ggml-sycl/norm.cpp b/ggml/src/ggml-sycl/norm.cpp
index 823d3a4828..00702b5d09 100644
--- a/ggml/src/ggml-sycl/norm.cpp
+++ b/ggml/src/ggml-sycl/norm.cpp
@@ -251,7 +251,6 @@ static void norm_f32_sycl(const float * x, float * dst, const int ncols, const i
         const float eps, queue_ptr stream, int device) {
 
     const sycl::range<3> global_dims(nsamples, nchannels, nrows);
-    GGML_ASSERT(ncols % WARP_SIZE == 0);
     if (ncols < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
         stream->submit([&](sycl::handler& cgh) {
@@ -334,7 +333,6 @@ static void group_norm_f32_sycl(const float* x, float* dst,
 
 static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const int nrows, const int nchannels, const int nsamples,
         const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample, const float eps, queue_ptr stream, int device) {
-    GGML_ASSERT(ncols % WARP_SIZE == 0);
     // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE);
 
     const sycl::range<3> global_dims(nsamples, nchannels, nrows);
@@ -374,7 +372,6 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const
 static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols,
     const int nrows, const float eps,
     queue_ptr stream, int device) {
-    GGML_ASSERT(ncols % WARP_SIZE == 0);
     // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE);
     if (ncols < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
diff --git a/ggml/src/ggml-virtgpu/CMakeLists.txt b/ggml/src/ggml-virtgpu/CMakeLists.txt
new file mode 100644
index 0000000000..e6b020beb5
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/CMakeLists.txt
@@ -0,0 +1,70 @@
+cmake_minimum_required(VERSION 3.19)
+cmake_policy(SET CMP0114 NEW)
+
+include(ExternalProject)
+
+message(STATUS "Including the VirtGPU/Virglrenderer API Remoting")
+
+# Download venus_hw.h from virglrenderer repository
+ExternalProject_Add(
+    venus_hw_header
+    URL https://gitlab.freedesktop.org/virgl/virglrenderer/-/raw/virglrenderer-1.2.0/src/venus_hw.h
+    DOWNLOAD_NO_EXTRACT YES
+    DOWNLOAD_DIR ${CMAKE_CURRENT_SOURCE_DIR}/include
+    DOWNLOAD_NAME venus_hw.h
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND ""
+    INSTALL_COMMAND ""
+    LOG_DOWNLOAD ON
+)
+
+if (NOT GGML_VIRTGPU_BACKEND STREQUAL "ONLY")
+    message(STATUS "Enable the VirtGPU/Virglrenderer API Remoting frontend library")
+
+    find_package(PkgConfig REQUIRED)
+    pkg_check_modules(DRM REQUIRED libdrm)
+    if (NOT GGML_BACKEND_DL)
+      # cannot simply use USE_VIRTGPU, as in the 'else()' case the
+      # frontend isn't compiled
+      target_compile_definitions(ggml PUBLIC "GGML_USE_VIRTGPU_FRONTEND")
+    endif()
+
+    ggml_add_backend_library(ggml-virtgpu
+                             ggml-backend-buffer.cpp
+                             ggml-backend.cpp
+                             ggml-backend-device.cpp
+                             ggml-backend-reg.cpp
+                             ggml-backend-buffer-type.cpp
+                             virtgpu-apir.h
+                             virtgpu-forward.gen.h
+                             virtgpu.cpp
+                             virtgpu-shm.cpp
+                             virtgpu-utils.cpp
+                             virtgpu-forward-device.cpp
+                             virtgpu-forward-buffer-type.cpp
+                             virtgpu-forward-buffer.cpp
+                             virtgpu-forward-backend.cpp
+                             virtgpu-forward-impl.h
+                             apir_cs_ggml-rpc-front.cpp
+                             ../../include/ggml-virtgpu.h)
+
+    target_include_directories(ggml-virtgpu PUBLIC /usr/include/libdrm/)
+
+    target_link_libraries(ggml-virtgpu PUBLIC ${DRM_LIBRARIES})
+    target_include_directories(ggml-virtgpu PUBLIC ${DRM_INCLUDE_DIRS})
+    target_compile_options(ggml-virtgpu PUBLIC ${DRM_CFLAGS_OTHER})
+
+    target_include_directories(ggml-virtgpu PUBLIC ./include)
+    target_include_directories(ggml-virtgpu PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
+
+    # Ensure venus_hw.h is downloaded before building ggml-virtgpu
+    add_dependencies(ggml-virtgpu venus_hw_header)
+
+    target_compile_options(ggml-virtgpu PRIVATE -std=c++20)
+else()
+    message(STATUS "Not building the VirtGPU/Virglrenderer API Remoting frontend library")
+endif()
+
+if (NOT GGML_VIRTGPU_BACKEND STREQUAL "OFF")
+    add_subdirectory("backend")
+endif()
diff --git a/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp b/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp
new file mode 100644
index 0000000000..f60ae3556c
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/apir_cs_ggml-rpc-front.cpp
@@ -0,0 +1,87 @@
+#include "backend/shared/apir_cs_rpc.h"
+#include "ggml-backend-impl.h"
+#include "ggml-impl.h"
+#include "ggml-remoting.h"
+
+#include <cinttypes>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+apir_rpc_tensor apir_serialize_tensor(const ggml_tensor * tensor) {
+    apir_rpc_tensor result;
+    result.id   = reinterpret_cast<uint64_t>(tensor);
+    result.type = tensor->type;
+    if (tensor->buffer) {
+        ggml_backend_buffer_t buffer = tensor->buffer;
+
+        result.buffer = BUFFER_TO_HOST_HANDLE(buffer);
+    } else {
+        result.buffer = 0;
+    }
+    for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) {
+        result.ne[i] = tensor->ne[i];
+        result.nb[i] = tensor->nb[i];
+    }
+    result.op = tensor->op;
+    for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
+        result.op_params[i] = tensor->op_params[i];
+    }
+    result.flags = tensor->flags;
+    for (uint32_t i = 0; i < GGML_MAX_SRC; i++) {
+        result.src[i] = reinterpret_cast<uint64_t>(tensor->src[i]);
+    }
+    result.view_src  = reinterpret_cast<uint64_t>(tensor->view_src);
+    result.view_offs = tensor->view_offs;
+    result.data      = reinterpret_cast<uint64_t>(tensor->data);
+    if (tensor->data) {
+        if (!tensor->buffer) {
+            GGML_ABORT("tensor has data but not buffer");
+        }
+        // tensor->data is serialized as an offset to the buffer base address
+        result.data -= reinterpret_cast<uint64_t>(BUFFER_TO_GGML_CONTEXT(tensor->buffer)->base);
+    }
+    snprintf(result.name, GGML_MAX_NAME, "%s", tensor->name);
+    return result;
+}
+
+void apir_add_tensor(ggml_tensor *                       tensor,
+                     std::vector<apir_rpc_tensor> &      tensors,
+                     std::unordered_set<ggml_tensor *> & visited) {
+    if (tensor == nullptr) {
+        return;
+    }
+    if (visited.find(tensor) != visited.end()) {
+        return;
+    }
+    visited.insert(tensor);
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        apir_add_tensor(tensor->src[i], tensors, visited);
+    }
+    apir_add_tensor(tensor->view_src, tensors, visited);
+    tensors.push_back(apir_serialize_tensor(tensor));
+}
+
+void apir_serialize_graph(const ggml_cgraph * cgraph, std::vector<uint8_t> & output) {
+    uint32_t                          n_nodes = cgraph->n_nodes;
+    std::vector<apir_rpc_tensor>      tensors;
+    std::unordered_set<ggml_tensor *> visited;
+    for (uint32_t i = 0; i < n_nodes; i++) {
+        apir_add_tensor(cgraph->nodes[i], tensors, visited);
+    }
+    // serialization format:
+    // | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(apir_rpc_tensor)) |
+    uint32_t n_tensors = tensors.size();
+    int      output_size =
+        sizeof(uint32_t) + n_nodes * sizeof(uint64_t) + sizeof(uint32_t) + n_tensors * sizeof(apir_rpc_tensor);
+    output.resize(output_size, 0);
+    memcpy(output.data(), &n_nodes, sizeof(n_nodes));
+    for (uint32_t i = 0; i < n_nodes; i++) {
+        memcpy(output.data() + sizeof(n_nodes) + i * sizeof(uint64_t), &cgraph->nodes[i], sizeof(uint64_t));
+    }
+    uint32_t * out_ntensors = (uint32_t *) (output.data() + sizeof(n_nodes) + n_nodes * sizeof(uint64_t));
+    *out_ntensors           = n_tensors;
+    apir_rpc_tensor * out_tensors =
+        (apir_rpc_tensor *) (output.data() + sizeof(n_nodes) + n_nodes * sizeof(uint64_t) + sizeof(uint32_t));
+    memcpy(out_tensors, tensors.data(), n_tensors * sizeof(apir_rpc_tensor));
+}
diff --git a/ggml/src/ggml-virtgpu/backend/CMakeLists.txt b/ggml/src/ggml-virtgpu/backend/CMakeLists.txt
new file mode 100644
index 0000000000..0b49c403b9
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/CMakeLists.txt
@@ -0,0 +1,21 @@
+cmake_minimum_required(VERSION 3.19)
+cmake_policy(SET CMP0114 NEW)
+
+message(STATUS "Enable the VirtGPU/Virglrenderer backend library")
+
+ggml_add_backend_library(ggml-virtgpu-backend
+                         backend.cpp
+                         backend-dispatched.cpp
+                         backend-dispatched-backend.cpp
+                         backend-dispatched-device.cpp
+                         backend-dispatched-buffer.cpp
+                         backend-dispatched-buffer-type.cpp
+                         shared/api_remoting.h
+                         shared/apir_backend.h
+                         shared/apir_cs.h
+                         apir_cs_ggml-rpc-back.cpp)
+
+target_compile_options(ggml-virtgpu-backend PRIVATE -std=c++20)
+
+# Add include directory for ggml-backend-impl.h and other core headers
+target_include_directories(ggml-virtgpu-backend PRIVATE ../..)
diff --git a/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp b/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp
new file mode 100644
index 0000000000..60a8a93bfb
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/apir_cs_ggml-rpc-back.cpp
@@ -0,0 +1,115 @@
+#include "ggml-backend-impl.h"
+#include "ggml-impl.h"
+#include "shared/apir_cs_rpc.h"
+
+#include <cinttypes>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+std::unordered_set<ggml_backend_buffer_t> backend_buffers;
+
+void apir_track_backend_buffer(ggml_backend_buffer_t buffer) {
+    backend_buffers.insert(buffer);
+}
+
+bool apir_untrack_backend_buffer(ggml_backend_buffer_t buffer) {
+    auto it = backend_buffers.find(buffer);
+    if (it == backend_buffers.end()) {
+        return false;
+    }
+
+    backend_buffers.erase(it);
+    return true;
+}
+
+std::unordered_set<ggml_backend_buffer_t> apir_get_track_backend_buffers() {
+    return backend_buffers;
+}
+
+ggml_tensor * apir_deserialize_tensor(ggml_context * ctx, const apir_rpc_tensor * tensor) {
+    ggml_tensor * result =
+        ggml_new_tensor_4d(ctx, (ggml_type) tensor->type, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
+    for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) {
+        result->nb[i] = tensor->nb[i];
+    }
+    result->buffer = reinterpret_cast<ggml_backend_buffer_t>(tensor->buffer);
+    if (result->buffer && backend_buffers.find(result->buffer) == backend_buffers.end()) {
+        printf("WARNING: HOST BUFFER NOT FOUND | %p\n", (void *) result->buffer);
+        result->buffer = nullptr;
+    }
+
+    uint64_t tensor_data = tensor->data;
+    if (result->buffer) {
+        // require that the tensor data does not go beyond the buffer end
+        uint64_t tensor_size  = (uint64_t) ggml_nbytes(result);
+        uint64_t buffer_start = (uint64_t) ggml_backend_buffer_get_base(result->buffer);
+        uint64_t buffer_size  = (uint64_t) ggml_backend_buffer_get_size(result->buffer);
+
+        // tensor->data is serialized as an offset to the buffer base address
+        tensor_data += buffer_start;
+
+        GGML_ASSERT(tensor_data + tensor_size >= tensor_data);  // check for overflow
+        GGML_ASSERT(tensor_data >= buffer_start && tensor_data + tensor_size <= buffer_start + buffer_size);
+    }
+
+    result->op = (ggml_op) tensor->op;
+    for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
+        result->op_params[i] = tensor->op_params[i];
+    }
+    result->flags = tensor->flags;
+    result->data  = reinterpret_cast<void *>(tensor_data);
+    ggml_set_name(result, tensor->name);
+    return result;
+}
+
+ggml_tensor * apir_create_node(uint64_t                                                      id,
+                               ggml_context *                                                ctx,
+                               const std::unordered_map<uint64_t, const apir_rpc_tensor *> & tensor_ptrs,
+                               std::unordered_map<uint64_t, ggml_tensor *> &                 tensor_map) {
+    if (id == 0) {
+        return nullptr;
+    }
+    if (tensor_map.find(id) != tensor_map.end()) {
+        return tensor_map[id];
+    }
+    const apir_rpc_tensor * tensor = tensor_ptrs.at(id);
+    ggml_tensor *           result = apir_deserialize_tensor(ctx, tensor);
+    if (result == nullptr) {
+        return nullptr;
+    }
+    tensor_map[id] = result;
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        result->src[i] = apir_create_node(tensor->src[i], ctx, tensor_ptrs, tensor_map);
+    }
+    result->view_src  = apir_create_node(tensor->view_src, ctx, tensor_ptrs, tensor_map);
+    result->view_offs = tensor->view_offs;
+    return result;
+}
+
+ggml_cgraph * apir_deserialize_graph(uint32_t                n_nodes,
+                                     uint32_t                n_tensors,
+                                     const apir_rpc_tensor * tensors,
+                                     const uint64_t *        nodes) {
+    size_t buf_size = ggml_tensor_overhead() * (n_nodes + n_tensors) + ggml_graph_overhead_custom(n_nodes, false);
+    ggml_init_params params = {
+        /*.mem_size   =*/buf_size,
+        /*.mem_buffer =*/NULL,
+        /*.no_alloc   =*/true,
+    };
+    ggml_context * ctx   = ggml_init(params);
+    ggml_cgraph *  graph = ggml_new_graph_custom(ctx, n_nodes, false);
+    graph->n_nodes       = n_nodes;
+    std::unordered_map<uint64_t, const apir_rpc_tensor *> tensor_ptrs;
+    for (uint32_t i = 0; i < n_tensors; i++) {
+        tensor_ptrs[tensors[i].id] = &tensors[i];
+    }
+    std::unordered_map<uint64_t, ggml_tensor *> tensor_map;
+    for (uint32_t i = 0; i < n_nodes; i++) {
+        int64_t id;
+        memcpy(&id, &nodes[i], sizeof(id));
+        graph->nodes[i] = apir_create_node(id, ctx, tensor_ptrs, tensor_map);
+    }
+
+    return graph;
+}
diff --git a/ggml/src/ggml-virtgpu/backend/backend-convert.h b/ggml/src/ggml-virtgpu/backend/backend-convert.h
new file mode 100644
index 0000000000..1978d21f7e
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/backend-convert.h
@@ -0,0 +1,13 @@
+#include "shared/apir_backend.h"
+
+#define BUFFER_TO_HOST_HANDLE(name) ggml_buffer_to_apir_handle(name)
+
+static inline apir_buffer_host_handle_t ggml_buffer_to_apir_handle(ggml_backend_buffer_t buffer) {
+    // in the backend, the buffer handle is the buffer pointer
+    return (apir_buffer_host_handle_t) buffer;
+}
+
+static inline apir_buffer_type_host_handle_t ggml_buffer_type_to_apir_handle(ggml_backend_buffer_type_t buft) {
+    // in the backend, the buffer handle is the buffer pointer
+    return (apir_buffer_type_host_handle_t) buft;
+}
diff --git a/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp b/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp
new file mode 100644
index 0000000000..77b4ee71e1
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/backend-dispatched-backend.cpp
@@ -0,0 +1,65 @@
+#include "backend-dispatched.h"
+#include "backend-virgl-apir.h"
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+#include "shared/apir_backend.h"
+
+#include <cstdint>
+
+uint32_t backend_backend_graph_compute(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(enc);
+
+    static bool async_backend_initialized = false;
+    static bool async_backend;
+
+    if (!async_backend_initialized) {
+        ggml_backend_dev_props props;
+
+        dev->iface.get_props(dev, &props);
+        async_backend             = props.caps.async;
+        async_backend_initialized = true;
+    }
+
+    uint32_t shmem_res_id;
+    apir_decode_virtgpu_shmem_res_id(dec, &shmem_res_id);
+
+    const void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
+    if (!shmem_data) {
+        GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
+        apir_decoder_set_fatal(dec);
+        return 1;
+    }
+    size_t cgraph_size;
+    apir_decode_size_t(dec, &cgraph_size);
+
+    apir_decoder secondary_dec = apir_new_decoder((const char *) shmem_data, cgraph_size);
+
+    ggml_cgraph * cgraph = apir_decode_ggml_cgraph(&secondary_dec, cgraph_size);
+
+    ggml_status status;
+#if APIR_BACKEND_CHECK_SUPPORTS_OP == 1
+    for (int idx = 0; idx < cgraph->n_nodes; idx++) {
+        ggml_tensor * op = ggml_graph_node(cgraph, idx);
+        if (dev->iface.supports_op(dev, op)) {
+            continue;
+        }
+        GGML_LOG_ERROR("Graph node %d (%s) not supported by the backend\n", idx, ggml_op_desc(op));
+
+        status = GGML_STATUS_ABORTED;
+        apir_encode_ggml_status(enc, &status);
+
+        return 0;
+    }
+#endif
+    status = bck->iface.graph_compute(bck, cgraph);
+
+    if (async_backend) {
+        bck->iface.synchronize(bck);
+    }
+
+    apir_encode_ggml_status(enc, &status);
+
+    return 0;
+}
diff --git a/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp b/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp
new file mode 100644
index 0000000000..8ea1bb4fb4
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp
@@ -0,0 +1,89 @@
+#include "backend-dispatched.h"
+#include "backend-virgl-apir.h"
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+
+#include <cstdint>
+
+uint32_t backend_buffer_type_get_name(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    ggml_backend_buffer_type_t buft;
+    buft = apir_decode_ggml_buffer_type(dec);
+
+    const char * string = buft->iface.get_name(buft);
+
+    const size_t string_size = strlen(string) + 1;
+    apir_encode_array_size(enc, string_size);
+    apir_encode_char_array(enc, string, string_size);
+
+    return 0;
+}
+
+uint32_t backend_buffer_type_get_alignment(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    ggml_backend_buffer_type_t buft;
+    buft = apir_decode_ggml_buffer_type(dec);
+
+    size_t value = buft->iface.get_alignment(buft);
+    apir_encode_size_t(enc, &value);
+
+    return 0;
+}
+
+uint32_t backend_buffer_type_get_max_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    ggml_backend_buffer_type_t buft;
+    buft = apir_decode_ggml_buffer_type(dec);
+
+    size_t value = buft->iface.get_max_size(buft);
+    apir_encode_size_t(enc, &value);
+
+    return 0;
+}
+
+uint32_t backend_buffer_type_is_host(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    ggml_backend_buffer_type_t buft;
+    buft = apir_decode_ggml_buffer_type(dec);
+
+    bool is_host = buft->iface.is_host(buft);
+    apir_encode_bool_t(enc, &is_host);
+
+    return 0;
+}
+
+uint32_t backend_buffer_type_alloc_buffer(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    ggml_backend_buffer_type_t buft;
+    buft = apir_decode_ggml_buffer_type(dec);
+
+    size_t size;
+    apir_decode_size_t(dec, &size);
+
+    ggml_backend_buffer_t buffer;
+
+    buffer = buft->iface.alloc_buffer(buft, size);
+
+    apir_encode_ggml_buffer(enc, buffer);
+
+    if (buffer) {
+        apir_track_backend_buffer(buffer);
+    }
+
+    return 0;
+}
+
+uint32_t backend_buffer_type_get_alloc_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    ggml_backend_buffer_type_t buft;
+    buft = apir_decode_ggml_buffer_type(dec);
+
+    const ggml_tensor * op = apir_decode_ggml_tensor_inplace(dec);
+
+    size_t value = buft->iface.get_alloc_size(buft, op);
+
+    apir_encode_size_t(enc, &value);
+
+    return 0;
+}
diff --git a/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp b/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp
new file mode 100644
index 0000000000..cf81888e98
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer.cpp
@@ -0,0 +1,131 @@
+#include "backend-dispatched.h"
+#include "backend-virgl-apir.h"
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+
+#include <cstdint>
+
+uint32_t backend_buffer_get_base(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    ggml_backend_buffer_t buffer;
+    buffer = apir_decode_ggml_buffer(dec);
+
+    uintptr_t base = (uintptr_t) buffer->iface.get_base(buffer);
+    apir_encode_uintptr_t(enc, &base);
+
+    return 0;
+}
+
+uint32_t backend_buffer_set_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(enc);
+
+    ggml_backend_buffer_t buffer;
+    buffer = apir_decode_ggml_buffer(dec);
+
+    ggml_tensor * tensor;
+    // safe to remove the const qualifier here
+    tensor = (ggml_tensor *) (uintptr_t) apir_decode_ggml_tensor(dec);
+
+    uint32_t shmem_res_id;
+    apir_decode_virtgpu_shmem_res_id(dec, &shmem_res_id);
+
+    size_t offset;
+    apir_decode_size_t(dec, &offset);
+
+    size_t size;
+    apir_decode_size_t(dec, &size);
+
+    void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
+
+    if (!shmem_data) {
+        GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
+        return 1;
+    }
+
+    buffer->iface.set_tensor(buffer, tensor, shmem_data, offset, size);
+
+    return 0;
+}
+
+uint32_t backend_buffer_get_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(enc);
+
+    ggml_backend_buffer_t buffer;
+    buffer = apir_decode_ggml_buffer(dec);
+
+    const ggml_tensor * tensor;
+    // safe to remove the const qualifier here
+    tensor = apir_decode_ggml_tensor(dec);
+
+    uint32_t shmem_res_id;
+    apir_decode_virtgpu_shmem_res_id(dec, &shmem_res_id);
+
+    size_t offset;
+    apir_decode_size_t(dec, &offset);
+
+    size_t size;
+    apir_decode_size_t(dec, &size);
+
+    void * shmem_data = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
+    if (!shmem_data) {
+        GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
+        return 1;
+    }
+
+    buffer->iface.get_tensor(buffer, tensor, shmem_data, offset, size);
+
+    return 0;
+}
+
+uint32_t backend_buffer_cpy_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+
+    ggml_backend_buffer_t buffer;
+    buffer = apir_decode_ggml_buffer(dec);
+
+    const ggml_tensor * src;
+    // safe to remove the const qualifier here
+    src               = apir_decode_ggml_tensor(dec);
+    ggml_tensor * dst = (ggml_tensor *) (uintptr_t) apir_decode_ggml_tensor(dec);
+
+    bool ret = buffer->iface.cpy_tensor(buffer, src, (ggml_tensor *) dst);
+
+    apir_encode_bool_t(enc, &ret);
+
+    return 0;
+}
+
+uint32_t backend_buffer_clear(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(enc);
+
+    ggml_backend_buffer_t buffer;
+    buffer = apir_decode_ggml_buffer(dec);
+
+    uint8_t value;
+    apir_decode_uint8_t(dec, &value);
+
+    buffer->iface.clear(buffer, value);
+
+    return 0;
+}
+
+uint32_t backend_buffer_free_buffer(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(enc);
+
+    ggml_backend_buffer_t buffer;
+    buffer = apir_decode_ggml_buffer(dec);
+
+    if (!apir_untrack_backend_buffer(buffer)) {
+        GGML_LOG_WARN("%s: unknown buffer %p\n", __func__, (void *) buffer);
+        return 1;
+    }
+
+    buffer->iface.free_buffer(buffer);
+
+    return 0;
+}
diff --git a/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp b/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp
new file mode 100644
index 0000000000..497f737a88
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/backend-dispatched-device.cpp
@@ -0,0 +1,148 @@
+#include "backend-dispatched.h"
+#include "backend-virgl-apir.h"
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+
+#include <cstdint>
+
+uint32_t backend_device_get_device_count(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    int32_t dev_count = reg->iface.get_device_count(reg);
+    apir_encode_int32_t(enc, &dev_count);
+
+    return 0;
+}
+
+uint32_t backend_device_get_count(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    int32_t dev_count = reg->iface.get_device_count(reg);
+    apir_encode_int32_t(enc, &dev_count);
+
+    return 0;
+}
+
+uint32_t backend_device_get_name(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    const char * string = dev->iface.get_name(dev);
+
+    const size_t string_size = strlen(string) + 1;
+    apir_encode_array_size(enc, string_size);
+    apir_encode_char_array(enc, string, string_size);
+
+    return 0;
+}
+
+uint32_t backend_device_get_description(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    const char * string = dev->iface.get_description(dev);
+
+    const size_t string_size = strlen(string) + 1;
+    apir_encode_array_size(enc, string_size);
+    apir_encode_char_array(enc, string, string_size);
+
+    return 0;
+}
+
+uint32_t backend_device_get_type(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    uint32_t type = dev->iface.get_type(dev);
+    apir_encode_uint32_t(enc, &type);
+
+    return 0;
+}
+
+uint32_t backend_device_get_memory(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    size_t free, total;
+    dev->iface.get_memory(dev, &free, &total);
+
+    apir_encode_size_t(enc, &free);
+    apir_encode_size_t(enc, &total);
+
+    return 0;
+}
+
+uint32_t backend_device_supports_op(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+
+    const ggml_tensor * op = apir_decode_ggml_tensor_inplace(dec);
+
+    bool supports_op = dev->iface.supports_op(dev, op);
+
+    apir_encode_bool_t(enc, &supports_op);
+
+    return 0;
+}
+
+uint32_t backend_device_get_buffer_type(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    ggml_backend_buffer_type_t bufft = dev->iface.get_buffer_type(dev);
+
+    apir_encode_ggml_buffer_type(enc, bufft);
+
+    return 0;
+}
+
+uint32_t backend_device_get_props(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    ggml_backend_dev_props props;
+    dev->iface.get_props(dev, &props);
+
+    apir_encode_bool_t(enc, &props.caps.async);
+    apir_encode_bool_t(enc, &props.caps.host_buffer);
+    apir_encode_bool_t(enc, &props.caps.buffer_from_host_ptr);
+    apir_encode_bool_t(enc, &props.caps.events);
+
+    return 0;
+}
+
+uint32_t backend_device_buffer_from_ptr(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(dec);
+
+    uint32_t shmem_res_id;
+    apir_decode_virtgpu_shmem_res_id(dec, &shmem_res_id);
+
+    void * shmem_ptr = ctx->iface->get_shmem_ptr(ctx->ctx_id, shmem_res_id);
+    if (!shmem_ptr) {
+        GGML_LOG_ERROR("Couldn't get the shmem addr from virgl\n");
+        apir_decoder_set_fatal(dec);
+        return 1;
+    }
+
+    size_t size;
+    apir_decode_size_t(dec, &size);
+    size_t max_tensor_size;
+    apir_decode_size_t(dec, &max_tensor_size);
+
+    ggml_backend_buffer_t buffer;
+    buffer = dev->iface.buffer_from_host_ptr(dev, shmem_ptr, size, max_tensor_size);
+
+    apir_encode_ggml_buffer(enc, buffer);
+    apir_encode_ggml_buffer_type(enc, buffer->buft);
+
+    if (buffer) {
+        apir_track_backend_buffer(buffer);
+    }
+
+    return 0;
+}
diff --git a/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp b/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp
new file mode 100644
index 0000000000..51d445725f
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/backend-dispatched.cpp
@@ -0,0 +1,46 @@
+#include "backend-dispatched.h"
+#include "backend-virgl-apir.h"
+
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+
+#include <cstdint>
+
+ggml_backend_reg_t reg = NULL;
+ggml_backend_dev_t dev = NULL;
+ggml_backend_t     bck = NULL;
+
+uint64_t timer_start = 0;
+uint64_t timer_total = 0;
+uint64_t timer_count = 0;
+
+uint32_t backend_dispatch_initialize(void * ggml_backend_reg_fct_p) {
+    if (reg != NULL) {
+        GGML_LOG_WARN("%s: already initialized\n", __func__);
+        return APIR_BACKEND_INITIALIZE_ALREADY_INITED;
+    }
+    ggml_backend_reg_t (*ggml_backend_reg_fct)(void) = (ggml_backend_reg_t (*)()) ggml_backend_reg_fct_p;
+
+    reg = ggml_backend_reg_fct();
+    if (reg == NULL) {
+        GGML_LOG_ERROR("%s: backend registration failed\n", __func__);
+        return APIR_BACKEND_INITIALIZE_BACKEND_REG_FAILED;
+    }
+
+    if (!reg->iface.get_device_count(reg)) {
+        GGML_LOG_ERROR("%s: backend initialization failed: no device found\n", __func__);
+        return APIR_BACKEND_INITIALIZE_NO_DEVICE;
+    }
+
+    dev = reg->iface.get_device(reg, 0);
+
+    if (!dev) {
+        GGML_LOG_ERROR("%s: backend initialization failed: no device received\n", __func__);
+        return APIR_BACKEND_INITIALIZE_NO_DEVICE;
+    }
+
+    bck = dev->iface.init_backend(dev, NULL);
+
+    return APIR_BACKEND_INITIALIZE_SUCCESS;
+}
diff --git a/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h b/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h
new file mode 100644
index 0000000000..b81fd5039b
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h
@@ -0,0 +1,130 @@
+#pragma once
+
+/* device */
+uint32_t backend_device_get_device_count(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_get_count(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_get_name(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_get_description(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_get_type(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_get_memory(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_supports_op(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_get_buffer_type(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_get_props(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_device_buffer_from_ptr(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+
+/* buffer-type */
+uint32_t backend_buffer_type_get_name(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_type_get_alignment(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_type_get_max_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_type_is_host(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_type_alloc_buffer(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_type_get_alloc_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+
+/* buffer */
+uint32_t backend_buffer_get_base(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_set_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_get_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_cpy_tensor(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_clear(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+uint32_t backend_buffer_free_buffer(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+
+/* backend */
+uint32_t backend_backend_graph_compute(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+
+static inline const char * backend_dispatch_command_name(ApirBackendCommandType type) {
+    switch (type) {
+        /* device */
+        case APIR_COMMAND_TYPE_DEVICE_GET_DEVICE_COUNT:
+            return "backend_device_get_device_count";
+        case APIR_COMMAND_TYPE_DEVICE_GET_COUNT:
+            return "backend_device_get_count";
+        case APIR_COMMAND_TYPE_DEVICE_GET_NAME:
+            return "backend_device_get_name";
+        case APIR_COMMAND_TYPE_DEVICE_GET_DESCRIPTION:
+            return "backend_device_get_description";
+        case APIR_COMMAND_TYPE_DEVICE_GET_TYPE:
+            return "backend_device_get_type";
+        case APIR_COMMAND_TYPE_DEVICE_GET_MEMORY:
+            return "backend_device_get_memory";
+        case APIR_COMMAND_TYPE_DEVICE_SUPPORTS_OP:
+            return "backend_device_supports_op";
+        case APIR_COMMAND_TYPE_DEVICE_GET_BUFFER_TYPE:
+            return "backend_device_get_buffer_type";
+        case APIR_COMMAND_TYPE_DEVICE_GET_PROPS:
+            return "backend_device_get_props";
+        case APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR:
+            return "backend_device_buffer_from_ptr";
+        /* buffer-type */
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME:
+            return "backend_buffer_type_get_name";
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT:
+            return "backend_buffer_type_get_alignment";
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE:
+            return "backend_buffer_type_get_max_size";
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST:
+            return "backend_buffer_type_is_host";
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER:
+            return "backend_buffer_type_alloc_buffer";
+        case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE:
+            return "backend_buffer_type_get_alloc_size";
+        /* buffer */
+        case APIR_COMMAND_TYPE_BUFFER_GET_BASE:
+            return "backend_buffer_get_base";
+        case APIR_COMMAND_TYPE_BUFFER_SET_TENSOR:
+            return "backend_buffer_set_tensor";
+        case APIR_COMMAND_TYPE_BUFFER_GET_TENSOR:
+            return "backend_buffer_get_tensor";
+        case APIR_COMMAND_TYPE_BUFFER_CPY_TENSOR:
+            return "backend_buffer_cpy_tensor";
+        case APIR_COMMAND_TYPE_BUFFER_CLEAR:
+            return "backend_buffer_clear";
+        case APIR_COMMAND_TYPE_BUFFER_FREE_BUFFER:
+            return "backend_buffer_free_buffer";
+        /* backend */
+        case APIR_COMMAND_TYPE_BACKEND_GRAPH_COMPUTE:
+            return "backend_backend_graph_compute";
+
+        default:
+            return "unknown";
+    }
+}
+
+extern "C" {
+static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATCH_TABLE_COUNT] = {
+
+    /* device */
+
+    /* APIR_COMMAND_TYPE_DEVICE_GET_DEVICE_COUNT  = */ backend_device_get_device_count,
+    /* APIR_COMMAND_TYPE_DEVICE_GET_COUNT  = */ backend_device_get_count,
+    /* APIR_COMMAND_TYPE_DEVICE_GET_NAME  = */ backend_device_get_name,
+    /* APIR_COMMAND_TYPE_DEVICE_GET_DESCRIPTION  = */ backend_device_get_description,
+    /* APIR_COMMAND_TYPE_DEVICE_GET_TYPE  = */ backend_device_get_type,
+    /* APIR_COMMAND_TYPE_DEVICE_GET_MEMORY  = */ backend_device_get_memory,
+    /* APIR_COMMAND_TYPE_DEVICE_SUPPORTS_OP  = */ backend_device_supports_op,
+    /* APIR_COMMAND_TYPE_DEVICE_GET_BUFFER_TYPE  = */ backend_device_get_buffer_type,
+    /* APIR_COMMAND_TYPE_DEVICE_GET_PROPS  = */ backend_device_get_props,
+    /* APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR  = */ backend_device_buffer_from_ptr,
+
+    /* buffer-type */
+
+    /* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME  = */ backend_buffer_type_get_name,
+    /* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT  = */ backend_buffer_type_get_alignment,
+    /* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE  = */ backend_buffer_type_get_max_size,
+    /* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST  = */ backend_buffer_type_is_host,
+    /* APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER  = */ backend_buffer_type_alloc_buffer,
+    /* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE  = */ backend_buffer_type_get_alloc_size,
+
+    /* buffer */
+
+    /* APIR_COMMAND_TYPE_BUFFER_GET_BASE  = */ backend_buffer_get_base,
+    /* APIR_COMMAND_TYPE_BUFFER_SET_TENSOR  = */ backend_buffer_set_tensor,
+    /* APIR_COMMAND_TYPE_BUFFER_GET_TENSOR  = */ backend_buffer_get_tensor,
+    /* APIR_COMMAND_TYPE_BUFFER_CPY_TENSOR  = */ backend_buffer_cpy_tensor,
+    /* APIR_COMMAND_TYPE_BUFFER_CLEAR  = */ backend_buffer_clear,
+    /* APIR_COMMAND_TYPE_BUFFER_FREE_BUFFER  = */ backend_buffer_free_buffer,
+
+    /* backend */
+
+    /* APIR_COMMAND_TYPE_BACKEND_GRAPH_COMPUTE  = */ backend_backend_graph_compute,
+};
+}
diff --git a/ggml/src/ggml-virtgpu/backend/backend-dispatched.h b/ggml/src/ggml-virtgpu/backend/backend-dispatched.h
new file mode 100644
index 0000000000..6ccbecf078
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/backend-dispatched.h
@@ -0,0 +1,23 @@
+#pragma once
+
+#include <cstdint>
+#include <cstddef>
+
+#include <ggml-backend.h>
+
+#include "backend-convert.h"
+#include "backend-virgl-apir.h"
+#include "shared/apir_backend.h"
+#include "shared/apir_cs.h"
+#include "shared/apir_cs_ggml.h"
+
+struct virgl_apir_context {
+    uint32_t               ctx_id;
+    virgl_apir_callbacks * iface;
+};
+
+typedef uint32_t (*backend_dispatch_t)(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
+
+#include "backend-dispatched.gen.h"
+
+uint32_t backend_dispatch_initialize(void * ggml_backend_reg_fct_p);
diff --git a/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h b/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h
new file mode 100644
index 0000000000..44b347f853
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/backend-virgl-apir.h
@@ -0,0 +1,32 @@
+#pragma once
+
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+#include "shared/api_remoting.h"
+
+#include <cstdarg>
+#include <cstdio>
+#include <cstdlib>
+
+extern ggml_backend_reg_t reg;
+extern ggml_backend_dev_t dev;
+extern ggml_backend_t     bck;
+
+struct virgl_apir_callbacks {
+    const char * (*get_config)(uint32_t virgl_ctx_id, const char * key);
+    void * (*get_shmem_ptr)(uint32_t virgl_ctx_id, uint32_t res_id);
+};
+
+extern "C" {
+ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct virgl_apir_callbacks *virgl_cbs);
+void                      apir_backend_deinit(uint32_t virgl_ctx_id);
+uint32_t                  apir_backend_dispatcher(uint32_t               virgl_ctx_id,
+                                                  virgl_apir_callbacks * virgl_cbs,
+                                                  uint32_t               cmd_type,
+                                                  char *                 dec_cur,
+                                                  const char *           dec_end,
+                                                  char *                 enc_cur,
+                                                  const char *           enc_end,
+                                                  char **                enc_cur_after);
+}
diff --git a/ggml/src/ggml-virtgpu/backend/backend.cpp b/ggml/src/ggml-virtgpu/backend/backend.cpp
new file mode 100644
index 0000000000..95d602ed60
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/backend.cpp
@@ -0,0 +1,148 @@
+#include "backend-dispatched.h"
+#include "backend-virgl-apir.h"
+
+#include "shared/api_remoting.h"
+#include "shared/apir_backend.h"
+#include "shared/apir_cs.h"
+
+#include <dlfcn.h>
+#include <ggml-backend.h>
+
+#include <iostream>
+
+#define APIR_LLAMA_CPP_GGML_LIBRARY_PATH_ENV "APIR_LLAMA_CPP_GGML_LIBRARY_PATH"
+#define APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV  "APIR_LLAMA_CPP_GGML_LIBRARY_REG"
+#define APIR_LLAMA_CPP_LOG_TO_FILE_ENV       "APIR_LLAMA_CPP_LOG_TO_FILE"
+
+#define GGML_DEFAULT_BACKEND_REG "ggml_backend_init"
+
+static void * backend_library_handle = NULL;
+static FILE * apir_logfile = NULL;
+
+static void log_to_file_callback(enum ggml_log_level level, const char * text, void * user_data) {
+    FILE * logfile = (FILE *)user_data;
+    fprintf(logfile, "[%d] %s", level, text);
+    fflush(logfile);
+}
+
+extern "C" {
+void apir_backend_deinit(uint32_t virgl_ctx_id) {
+    GGML_UNUSED(virgl_ctx_id);
+
+    auto buffers = apir_get_track_backend_buffers();
+    for (const auto & buffer : buffers) {
+        apir_untrack_backend_buffer(buffer);
+        buffer->iface.free_buffer(buffer);
+    }
+
+    if (dev) {
+        size_t free, total;
+        dev->iface.get_memory(dev, &free, &total);
+        GGML_LOG_INFO("%s: free memory: %ld MB\n", __func__, (size_t) free / 1024 / 1024);
+    }
+
+    if (backend_library_handle) {
+        GGML_LOG_INFO("%s: The GGML backend library was loaded. Unloading it.\n", __func__);
+        dlclose(backend_library_handle);
+        backend_library_handle = NULL;
+    }
+
+    if (apir_logfile) {
+        fclose(apir_logfile);
+        apir_logfile = NULL;
+    }
+}
+
+#define APIR_GGML_LIBRARY_PATH_KEY "ggml.library.path"
+#define APIR_GGML_LIBRARY_REG_KEY "ggml.library.reg"
+
+ApirLoadLibraryReturnCode apir_backend_initialize(uint32_t virgl_ctx_id, struct virgl_apir_callbacks *virgl_cbs) {
+    const char * dlsym_error;
+
+    const char * apir_log_to_file = getenv(APIR_LLAMA_CPP_LOG_TO_FILE_ENV);
+    if (apir_log_to_file) {
+        apir_logfile = fopen(apir_log_to_file, "w");
+        if (apir_logfile) {
+            ggml_log_set(log_to_file_callback, apir_logfile);
+        } else {
+            GGML_LOG_INFO("Could not open the log file at '%s'\n", apir_log_to_file);
+        }
+    }
+
+    const char * library_name = virgl_cbs->get_config(virgl_ctx_id, APIR_GGML_LIBRARY_PATH_KEY);
+    const char * virgl_library_reg = virgl_cbs->get_config(virgl_ctx_id, APIR_GGML_LIBRARY_REG_KEY);
+    const char * library_reg = virgl_library_reg ? virgl_library_reg : GGML_DEFAULT_BACKEND_REG;
+
+    if (!library_name) {
+        GGML_LOG_ERROR("cannot open the GGML library: env var '%s' not defined\n", APIR_LLAMA_CPP_GGML_LIBRARY_PATH_ENV);
+
+        return APIR_LOAD_LIBRARY_ENV_VAR_MISSING;
+    }
+
+    backend_library_handle = dlopen(library_name, RTLD_LAZY);
+
+    if (!backend_library_handle) {
+        GGML_LOG_ERROR("cannot open the GGML library: %s\n", dlerror());
+
+        return APIR_LOAD_LIBRARY_CANNOT_OPEN;
+    }
+
+    if (!library_reg) {
+        GGML_LOG_ERROR("cannot register the GGML library: env var '%s' not defined\n", APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV);
+
+        return APIR_LOAD_LIBRARY_ENV_VAR_MISSING;
+    }
+
+    void * ggml_backend_reg_fct = dlsym(backend_library_handle, library_reg);
+    dlsym_error                 = dlerror();
+    if (dlsym_error) {
+        GGML_LOG_ERROR("cannot find the GGML backend registration symbol '%s' (from %s): %s\n", library_reg,
+              APIR_LLAMA_CPP_GGML_LIBRARY_REG_ENV, dlsym_error);
+
+        return APIR_LOAD_LIBRARY_SYMBOL_MISSING;
+    }
+
+    uint32_t ret = backend_dispatch_initialize(ggml_backend_reg_fct);
+
+    return (ApirLoadLibraryReturnCode) (APIR_LOAD_LIBRARY_INIT_BASE_INDEX + ret);
+}
+
+uint32_t apir_backend_dispatcher(uint32_t               virgl_ctx_id,
+                                 virgl_apir_callbacks * virgl_cbs,
+                                 uint32_t               cmd_type,
+                                 char *                 dec_cur,
+                                 const char *           dec_end,
+                                 char *                 enc_cur,
+                                 const char *           enc_end,
+                                 char **                enc_cur_after) {
+    apir_encoder enc = {
+        .cur   = enc_cur,
+        .start = enc_cur,
+        .end   = enc_end,
+        .fatal = false,
+    };
+
+    apir_decoder dec = {
+        .cur   = dec_cur,
+        .end   = dec_end,
+        .fatal = false,
+    };
+
+    virgl_apir_context ctx = {
+        .ctx_id = virgl_ctx_id,
+        .iface = virgl_cbs,
+    };
+
+    if (cmd_type >= APIR_BACKEND_DISPATCH_TABLE_COUNT) {
+        GGML_LOG_ERROR("Received an invalid dispatch index (%d >= %d)\n", cmd_type, APIR_BACKEND_DISPATCH_TABLE_COUNT);
+        return APIR_BACKEND_FORWARD_INDEX_INVALID;
+    }
+
+    backend_dispatch_t forward_fct = apir_backend_dispatch_table[cmd_type];
+    uint32_t           ret         = forward_fct(&enc, &dec, &ctx);
+
+    *enc_cur_after = enc.cur;
+
+    return ret;
+}
+}
diff --git a/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h b/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h
new file mode 100644
index 0000000000..f19a5d12d1
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/shared/api_remoting.h
@@ -0,0 +1,90 @@
+#pragma once
+
+/* the rest of this file must match virglrenderer/src/apir-protocol.h */
+
+#include <unistd.h>
+
+#include <cstdint>
+
+#define APIR_PROTOCOL_MAJOR 0
+#define APIR_PROTOCOL_MINOR 1
+
+#define APIR_HANDSHAKE_MAGIC 0xab1e
+
+enum ApirCommandType {
+    APIR_COMMAND_TYPE_HANDSHAKE   = 0,
+    APIR_COMMAND_TYPE_LOADLIBRARY = 1,
+    APIR_COMMAND_TYPE_FORWARD     = 2,
+
+    APIR_COMMAND_TYPE_LENGTH      = 3,
+};
+
+typedef uint64_t ApirCommandFlags;
+
+enum ApirLoadLibraryReturnCode {
+    APIR_LOAD_LIBRARY_SUCCESS                        = 0,
+    APIR_LOAD_LIBRARY_HYPERCALL_INITIALIZATION_ERROR = 1,
+    APIR_LOAD_LIBRARY_ALREADY_LOADED                 = 2,
+    APIR_LOAD_LIBRARY_ENV_VAR_MISSING                = 3,
+    APIR_LOAD_LIBRARY_CANNOT_OPEN                    = 4,
+    APIR_LOAD_LIBRARY_SYMBOL_MISSING                 = 5,
+    APIR_LOAD_LIBRARY_INIT_BASE_INDEX                = 6,  // anything above this is a APIR backend library initialization return code
+};
+
+enum ApirForwardReturnCode {
+    APIR_FORWARD_SUCCESS         = 0,
+    APIR_FORWARD_NO_DISPATCH_FCT = 1,
+    APIR_FORWARD_TIMEOUT         = 2,
+
+    APIR_FORWARD_BASE_INDEX      = 3,  // anything above this is a APIR backend library forward return code
+} ;
+
+__attribute__((unused)) static inline const char * apir_command_name(ApirCommandType type) {
+    switch (type) {
+        case APIR_COMMAND_TYPE_HANDSHAKE:
+            return "HandShake";
+        case APIR_COMMAND_TYPE_LOADLIBRARY:
+            return "LoadLibrary";
+        case APIR_COMMAND_TYPE_FORWARD:
+            return "Forward";
+        default:
+            return "unknown";
+    }
+}
+
+__attribute__((unused)) static const char * apir_load_library_error(ApirLoadLibraryReturnCode code) {
+#define APIR_LOAD_LIBRARY_ERROR(code_name) \
+    do {                                   \
+        if (code == code_name)             \
+            return #code_name;             \
+    } while (0)
+
+    APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_SUCCESS);
+    APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_HYPERCALL_INITIALIZATION_ERROR);
+    APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_ALREADY_LOADED);
+    APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_ENV_VAR_MISSING);
+    APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_CANNOT_OPEN);
+    APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_SYMBOL_MISSING);
+    APIR_LOAD_LIBRARY_ERROR(APIR_LOAD_LIBRARY_INIT_BASE_INDEX);
+
+    return "Unknown APIR_COMMAND_TYPE_LoadLibrary error";
+
+#undef APIR_LOAD_LIBRARY_ERROR
+}
+
+__attribute__((unused)) static const char * apir_forward_error(ApirForwardReturnCode code) {
+#define APIR_FORWARD_ERROR(code_name) \
+    do {                              \
+        if (code == code_name)        \
+            return #code_name;        \
+    } while (0)
+
+    APIR_FORWARD_ERROR(APIR_FORWARD_SUCCESS);
+    APIR_FORWARD_ERROR(APIR_FORWARD_NO_DISPATCH_FCT);
+    APIR_FORWARD_ERROR(APIR_FORWARD_TIMEOUT);
+    APIR_FORWARD_ERROR(APIR_FORWARD_BASE_INDEX);
+
+    return "Unknown APIR_COMMAND_TYPE_FORWARD error";
+
+#undef APIR_FORWARD_ERROR
+}
diff --git a/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h b/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h
new file mode 100644
index 0000000000..d214b6f2a9
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/shared/apir_backend.gen.h
@@ -0,0 +1,36 @@
+typedef enum ApirBackendCommandType {
+
+    /* device */
+    APIR_COMMAND_TYPE_DEVICE_GET_DEVICE_COUNT = 0,
+    APIR_COMMAND_TYPE_DEVICE_GET_COUNT        = 1,
+    APIR_COMMAND_TYPE_DEVICE_GET_NAME         = 2,
+    APIR_COMMAND_TYPE_DEVICE_GET_DESCRIPTION  = 3,
+    APIR_COMMAND_TYPE_DEVICE_GET_TYPE         = 4,
+    APIR_COMMAND_TYPE_DEVICE_GET_MEMORY       = 5,
+    APIR_COMMAND_TYPE_DEVICE_SUPPORTS_OP      = 6,
+    APIR_COMMAND_TYPE_DEVICE_GET_BUFFER_TYPE  = 7,
+    APIR_COMMAND_TYPE_DEVICE_GET_PROPS        = 8,
+    APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR  = 9,
+
+    /* buffer-type */
+    APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME       = 10,
+    APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT  = 11,
+    APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE   = 12,
+    APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST        = 13,
+    APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER   = 14,
+    APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE = 15,
+
+    /* buffer */
+    APIR_COMMAND_TYPE_BUFFER_GET_BASE    = 16,
+    APIR_COMMAND_TYPE_BUFFER_SET_TENSOR  = 17,
+    APIR_COMMAND_TYPE_BUFFER_GET_TENSOR  = 18,
+    APIR_COMMAND_TYPE_BUFFER_CPY_TENSOR  = 19,
+    APIR_COMMAND_TYPE_BUFFER_CLEAR       = 20,
+    APIR_COMMAND_TYPE_BUFFER_FREE_BUFFER = 21,
+
+    /* backend */
+    APIR_COMMAND_TYPE_BACKEND_GRAPH_COMPUTE = 22,
+
+    // last command_type index + 1
+    APIR_BACKEND_DISPATCH_TABLE_COUNT = 23,
+} ApirBackendCommandType;
diff --git a/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h b/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h
new file mode 100644
index 0000000000..f3efa52c72
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/shared/apir_backend.h
@@ -0,0 +1,46 @@
+#pragma once
+
+#include "apir_backend.gen.h"
+
+#include <stdint.h>  // for uintptr_t
+#include <time.h>    // for timespec, clock_gettime
+
+#define APIR_BACKEND_INITIALIZE_SUCCESS                     0
+#define APIR_BACKEND_INITIALIZE_CANNOT_OPEN_BACKEND_LIBRARY 1
+#define APIR_BACKEND_INITIALIZE_CANNOT_OPEN_GGML_LIBRARY    2
+#define APIR_BACKEND_INITIALIZE_MISSING_BACKEND_SYMBOLS     3
+#define APIR_BACKEND_INITIALIZE_MISSING_GGML_SYMBOLS        4
+#define APIR_BACKEND_INITIALIZE_BACKEND_FAILED              5
+#define APIR_BACKEND_INITIALIZE_BACKEND_REG_FAILED          6
+#define APIR_BACKEND_INITIALIZE_ALREADY_INITED              7
+#define APIR_BACKEND_INITIALIZE_NO_DEVICE                   8
+
+
+// new entries here need to be added to the apir_backend_initialize_error function below
+
+#define APIR_BACKEND_FORWARD_INDEX_INVALID 6
+
+// 0 is fast, 1 avoids the backend to crash if an unsupported tensor is received
+#define APIR_BACKEND_CHECK_SUPPORTS_OP 0
+
+typedef uintptr_t apir_buffer_type_host_handle_t;
+typedef uintptr_t apir_buffer_host_handle_t;
+
+static const char * apir_backend_initialize_error(int code) {
+#define APIR_BACKEND_INITIALIZE_ERROR(code_name) \
+    do {                                         \
+        if (code == code_name)                   \
+            return #code_name;                   \
+    } while (0)
+
+    APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_SUCCESS);
+    APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_CANNOT_OPEN_BACKEND_LIBRARY);
+    APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_CANNOT_OPEN_GGML_LIBRARY);
+    APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_MISSING_BACKEND_SYMBOLS);
+    APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_MISSING_GGML_SYMBOLS);
+    APIR_BACKEND_INITIALIZE_ERROR(APIR_BACKEND_INITIALIZE_BACKEND_FAILED);
+
+    return "Unknown APIR_BACKEND_INITIALIZE error:/";
+
+#undef APIR_BACKEND_INITIALIZE_ERROR
+}
diff --git a/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h b/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h
new file mode 100644
index 0000000000..27a61091ff
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/shared/apir_cs.h
@@ -0,0 +1,383 @@
+#pragma once
+
+#include "ggml-impl.h"
+
+#include <cassert>
+#include <cstring>
+
+#define likely(x)   __builtin_expect(!!(x), 1)
+#define unlikely(x) __builtin_expect(!!(x), 0)
+
+struct apir_encoder {
+    char *       cur;
+    const char * start;
+    const char * end;
+    bool         fatal;
+
+};
+
+struct apir_decoder {
+    const char * cur;
+    const char * end;
+    bool         fatal;
+};
+
+/*
+ * new encoder and decoder
+ */
+
+static apir_decoder apir_new_decoder(const char * ptr, size_t size) {
+    apir_decoder dec = {
+        .cur = ptr,
+        .end = ptr + size,
+        .fatal = false,
+    };
+
+    return dec;
+}
+
+static apir_encoder apir_new_encoder(char * ptr, size_t size) {
+    apir_encoder enc = {
+        .cur   = ptr,
+        .start = ptr,
+        .end   = ptr + size,
+        .fatal = false,
+    };
+
+    return enc;
+}
+
+/*
+ * fatal flag handling
+ */
+
+static inline void apir_encoder_reset_fatal(apir_encoder * enc) {
+    enc->fatal = false;
+}
+
+static inline void apir_encoder_set_fatal(apir_encoder * enc) {
+    enc->fatal = true;
+}
+
+static inline bool apir_encoder_get_fatal(const apir_encoder * enc) {
+    return enc->fatal;
+}
+
+static inline void apir_decoder_reset_fatal(apir_decoder * dec) {
+    dec->fatal = false;
+}
+
+static inline void apir_decoder_set_fatal(apir_decoder * dec) {
+    dec->fatal = true;
+}
+
+static inline bool apir_decoder_get_fatal(const apir_decoder * dec) {
+    return dec->fatal;
+}
+
+/*
+ * encode peek
+ */
+
+static inline bool apir_decoder_peek_internal(apir_decoder * dec,
+                                              size_t                size,
+                                              void *                val,
+                                              size_t                val_size) {
+    assert(val_size <= size);
+
+    if (unlikely(size > (size_t) (dec->end - dec->cur))) {
+        GGML_LOG_ERROR("reading too much from the decoder ...\n");
+        apir_decoder_set_fatal(dec);
+        memset(val, 0, val_size);
+        return false;
+    }
+
+    /* we should not rely on the compiler to optimize away memcpy... */
+    memcpy(val, dec->cur, val_size);
+    return true;
+}
+
+static inline void apir_decoder_peek(apir_decoder * dec, size_t size, void * val, size_t val_size) {
+    apir_decoder_peek_internal(dec, size, val, val_size);
+}
+
+static inline const void * apir_decoder_use_inplace(apir_decoder * dec, size_t size) {
+    if (unlikely(size > (size_t) (dec->end - dec->cur))) {
+        GGML_LOG_ERROR("reading too much from the decoder ...\n");
+        apir_decoder_set_fatal(dec);
+        return NULL;
+    }
+    const void * addr = dec->cur;
+    dec->cur += size;
+
+    return addr;
+}
+
+/*
+ * read/write
+ */
+
+static inline void apir_decoder_read(apir_decoder * dec, size_t size, void * val, size_t val_size) {
+    if (apir_decoder_peek_internal(dec, size, val, val_size)) {
+        dec->cur += size;
+    }
+}
+
+static inline char * apir_encoder_write(apir_encoder * enc, size_t size, const void * val, size_t val_size) {
+    assert(val_size <= size);
+    assert(size <= ((size_t) (enc->end - enc->cur)));
+
+    char * write_addr = enc->cur;
+    /* we should not rely on the compiler to optimize away memcpy... */
+    memcpy(write_addr, val, val_size);
+    enc->cur += size;
+
+    return write_addr;
+}
+
+/*
+ * encode/decode
+ */
+
+static inline void apir_decode(apir_decoder * dec, size_t size, void * data, size_t data_size) {
+    assert(size % 4 == 0);
+    apir_decoder_read(dec, size, data, data_size);
+}
+
+static inline void apir_encode(apir_encoder * enc, size_t size, const void * data, size_t data_size) {
+    assert(size % 4 == 0);
+    apir_encoder_write(enc, size, data, data_size);
+}
+
+/*
+ * typed encode/decode
+ */
+
+/* uint8_t */
+
+static inline void apir_encode_uint8_t(apir_encoder * enc, const uint8_t * val) {
+    apir_encode(enc, sizeof(int), val, sizeof(*val));
+}
+
+static inline void apir_decode_uint8_t(apir_decoder * dec, uint8_t * val) {
+    apir_decode(dec, sizeof(int), val, sizeof(*val));
+}
+
+/* uint64_t */
+
+static inline void apir_encode_uint64_t(apir_encoder * enc, const uint64_t * val) {
+    apir_encode(enc, 8, val, sizeof(*val));
+}
+
+static inline void apir_decode_uint64_t(apir_decoder * dec, uint64_t * val) {
+    apir_decode(dec, 8, val, sizeof(*val));
+}
+
+static inline void apir_encode_uint64_t_array(apir_encoder * enc, const uint64_t * val, uint32_t count) {
+    const size_t size = sizeof(*val) * count;
+    assert(size >= count);
+    apir_encode(enc, size, val, size);
+}
+
+static inline void apir_decode_uint64_t_array(apir_decoder * dec, uint64_t * val, uint32_t count) {
+    const size_t size = sizeof(*val) * count;
+    assert(size >= count);
+    apir_decode(dec, size, val, size);
+}
+
+static inline const uint64_t * apir_decode_uint64_t_array_inplace(apir_decoder * dec, uint32_t count) {
+    return (uint64_t *) (uintptr_t) apir_decoder_use_inplace(dec, count * sizeof(uint64_t));
+}
+
+/* int32_t */
+
+static inline void apir_encode_int32_t(apir_encoder * enc, const int32_t * val) {
+    apir_encode(enc, 4, val, sizeof(*val));
+}
+
+static inline void apir_decode_int32_t(apir_decoder * dec, int32_t * val) {
+    apir_decode(dec, 4, val, sizeof(*val));
+}
+
+static inline void apir_encode_int32_t_array(apir_encoder * enc, const int32_t * val, uint32_t count) {
+    const size_t size = sizeof(*val) * count;
+    assert(size >= count);
+    apir_encode(enc, size, val, size);
+}
+
+static inline void apir_decode_int32_t_array(apir_decoder * dec, int32_t * val, uint32_t count) {
+    const size_t size = sizeof(*val) * count;
+    assert(size >= count);
+    apir_decode(dec, size, val, size);
+}
+
+/* array size (uint64_t) */
+
+static inline void apir_encode_array_size(apir_encoder * enc, uint64_t size) {
+    apir_encode_uint64_t(enc, &size);
+}
+
+static inline uint64_t apir_decode_array_size(apir_decoder * dec, uint64_t expected_size) {
+    uint64_t size;
+    apir_decode_uint64_t(dec, &size);
+    if (size != expected_size) {
+        GGML_LOG_ERROR("Couldn't decode array from the decoder\n");
+        apir_decoder_set_fatal(dec);
+        size = 0;
+    }
+    return size;
+}
+
+static inline uint64_t apir_decode_array_size_unchecked(apir_decoder * dec) {
+    uint64_t size;
+    apir_decode_uint64_t(dec, &size);
+    return size;
+}
+
+/* non-array pointer */
+
+static inline bool apir_encode_simple_pointer(apir_encoder * enc, const void * val) {
+    apir_encode_array_size(enc, val ? 1 : 0);
+    return val;
+}
+
+static inline bool apir_decode_simple_pointer(apir_decoder * dec) {
+    return apir_decode_array_size_unchecked(dec);
+}
+
+/* uint32_t */
+
+static inline void apir_encode_uint32_t(apir_encoder * enc, const uint32_t * val) {
+    apir_encode(enc, 4, val, sizeof(*val));
+}
+
+static inline void apir_decode_uint32_t(apir_decoder * dec, uint32_t * val) {
+    apir_decode(dec, 4, val, sizeof(*val));
+}
+
+static inline void apir_encode_uint32_t_array(apir_encoder * enc, const uint32_t * val, uint32_t count) {
+    const size_t size = sizeof(*val) * count;
+    assert(size >= count);
+    apir_encode(enc, size, val, size);
+}
+
+static inline void apir_decode_uint32_t_array(apir_decoder * dec, uint32_t * val, uint32_t count) {
+    const size_t size = sizeof(*val) * count;
+    assert(size >= count);
+    apir_decode(dec, size, val, size);
+}
+
+/* size_t */
+
+static inline void apir_encode_size_t(apir_encoder * enc, const size_t * val) {
+    const uint64_t tmp = *val;
+    apir_encode_uint64_t(enc, &tmp);
+}
+
+static inline void apir_decode_size_t(apir_decoder * dec, size_t * val) {
+    uint64_t tmp;
+    apir_decode_uint64_t(dec, &tmp);
+    *val = tmp;
+}
+
+static inline void apir_encode_size_t_array(apir_encoder * enc, const size_t * val, uint32_t count) {
+    if (sizeof(size_t) == sizeof(uint64_t)) {
+        apir_encode_uint64_t_array(enc, (const uint64_t *) val, count);
+    } else {
+        for (uint32_t i = 0; i < count; i++) {
+            apir_encode_size_t(enc, &val[i]);
+        }
+    }
+}
+
+static inline void apir_decode_size_t_array(apir_decoder * dec, size_t * val, uint32_t count) {
+    if (sizeof(size_t) == sizeof(uint64_t)) {
+        apir_decode_uint64_t_array(dec, (uint64_t *) val, count);
+    } else {
+        for (uint32_t i = 0; i < count; i++) {
+            apir_decode_size_t(dec, &val[i]);
+        }
+    }
+}
+
+/* opaque blob */
+
+static inline void apir_encode_blob_array(apir_encoder * enc, const void * val, size_t size) {
+    apir_encode(enc, (size + 3) & ~3, val, size);
+}
+
+static inline void apir_decode_blob_array(apir_decoder * dec, void * val, size_t size) {
+    apir_decode(dec, (size + 3) & ~3, val, size);
+}
+
+/* string */
+
+static inline void apir_encode_char_array(apir_encoder * enc, const char * val, size_t size) {
+    assert(size && strlen(val) < size);
+    apir_encode_blob_array(enc, val, size);
+}
+
+static inline void apir_decode_char_array(apir_decoder * dec, char * val, size_t size) {
+    apir_decode_blob_array(dec, val, size);
+    if (size) {
+        val[size - 1] = '\0';
+    } else {
+        GGML_LOG_ERROR("Couldn't decode the blog array\n");
+        apir_decoder_set_fatal(dec);
+    }
+}
+
+/* (temp) buffer allocation */
+
+static inline void * apir_decoder_alloc_array(size_t size, size_t count) {
+    size_t alloc_size;
+    if (unlikely(__builtin_mul_overflow(size, count, &alloc_size))) {
+        GGML_LOG_ERROR("overflow in array allocation of %zu * %zu bytes\n", size, count);
+        return NULL;
+    }
+
+    return malloc(alloc_size);
+}
+
+/* bool */
+
+static inline void apir_encode_bool_t(apir_encoder * enc, const bool * val) {
+    apir_encode(enc, sizeof(int), val, sizeof(bool));
+}
+
+static inline void apir_decode_bool_t(apir_decoder * dec, bool * val) {
+    apir_decode(dec, sizeof(int), val, sizeof(bool));
+}
+
+/* apir_buffer_type_host_handle_t */
+
+static inline void apir_encode_apir_buffer_type_host_handle_t(apir_encoder *                  enc,
+                                                              const apir_buffer_type_host_handle_t * val) {
+    apir_encode(enc, sizeof(apir_buffer_type_host_handle_t), val, sizeof(apir_buffer_type_host_handle_t));
+}
+
+static inline void apir_decode_apir_buffer_type_host_handle_t(apir_decoder *            dec,
+                                                              apir_buffer_type_host_handle_t * val) {
+    apir_decode(dec, sizeof(apir_buffer_type_host_handle_t), val, sizeof(apir_buffer_type_host_handle_t));
+}
+
+/* apir_buffer_host_handle_t */
+
+static inline void apir_encode_apir_buffer_host_handle_t(apir_encoder *             enc,
+                                                         const apir_buffer_host_handle_t * val) {
+    apir_encode(enc, sizeof(apir_buffer_host_handle_t), val, sizeof(apir_buffer_host_handle_t));
+}
+
+static inline void apir_decode_apir_buffer_host_handle_t(apir_decoder * dec, apir_buffer_host_handle_t * val) {
+    apir_decode(dec, sizeof(apir_buffer_host_handle_t), val, sizeof(apir_buffer_host_handle_t));
+}
+
+/* uintptr_t */
+
+static inline void apir_encode_uintptr_t(apir_encoder * enc, const uintptr_t * val) {
+    apir_encode(enc, sizeof(*val), val, sizeof(*val));
+}
+
+static inline void apir_decode_uintptr_t(apir_decoder * dec, uintptr_t * val) {
+    apir_decode(dec, sizeof(*val), val, sizeof(*val));
+}
diff --git a/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h b/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h
new file mode 100644
index 0000000000..070c3b25fb
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h
@@ -0,0 +1,211 @@
+#include "ggml-impl.h"
+#include "apir_cs.h"
+#include "apir_cs_rpc.h"
+
+// ggml_buffer_to_apir_host_handle(ggml_backend_buffer_t buffer);
+
+static inline void apir_encode_ggml_buffer_host_handle(apir_encoder *                    enc,
+                                                       const apir_buffer_host_handle_t * handle);
+
+static inline ggml_backend_buffer_t apir_decode_ggml_buffer(apir_decoder * dec);
+
+/* apir_rpc_tensor */
+
+static inline void apir_encode_rcp_tensor(apir_encoder * enc, const apir_rpc_tensor * apir_rpc_tensor) {
+    size_t apir_rpc_tensor_size = sizeof(*apir_rpc_tensor);
+    apir_encode(enc, apir_rpc_tensor_size, apir_rpc_tensor, apir_rpc_tensor_size);
+}
+
+static inline apir_rpc_tensor * apir_decode_apir_rpc_tensor_inplace(apir_decoder * dec) {
+    size_t apir_rpc_tensor_size = sizeof(apir_rpc_tensor);
+
+    return (apir_rpc_tensor *) (uintptr_t) apir_decoder_use_inplace(dec, apir_rpc_tensor_size);
+}
+
+static inline apir_rpc_tensor * apir_decode_apir_rpc_tensor_array_inplace(apir_decoder * dec,
+                                                                          uint32_t       n_tensors) {
+    size_t apir_rpc_tensor_size = sizeof(apir_rpc_tensor) * n_tensors;
+
+    return (apir_rpc_tensor *) (uintptr_t) apir_decoder_use_inplace(dec, apir_rpc_tensor_size);
+}
+
+/* ggml_tensor */
+
+static inline void apir_encode_ggml_tensor(apir_encoder * enc, const ggml_tensor * tensor) {
+    apir_rpc_tensor serialized = apir_serialize_tensor(tensor);
+
+    apir_encode_rcp_tensor(enc, &serialized);
+}
+
+static inline const ggml_tensor * apir_decode_ggml_tensor(apir_decoder * dec) {
+    const apir_rpc_tensor * apir_rpc_tensor = apir_decode_apir_rpc_tensor_inplace(dec);
+    ggml_init_params params{
+        /*.mem_size   =*/ ggml_tensor_overhead(),
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ true,
+    };
+    ggml_context * ctx = ggml_init(params);
+
+    const ggml_tensor * tensor = apir_deserialize_tensor(ctx, apir_rpc_tensor);
+
+    return tensor;
+}
+
+/* *** ggml_backend_buffer_type_t *** */
+
+// ggml_backend_buffer_type_t is a POINTER (to a struct).
+// Only the host pointer is shared between the host and guest.
+// The guest stores it in `buft->context`.
+// The host simply writes the pointer address in the buffer variable.
+
+static inline void apir_encode_ggml_buffer_type(apir_encoder * enc, ggml_backend_buffer_type_t buft) {
+    apir_buffer_type_host_handle_t handle = ggml_buffer_type_to_apir_handle(buft);
+    apir_encoder_write(enc, sizeof(handle), &handle, sizeof(handle));
+}
+
+static inline ggml_backend_buffer_type_t apir_decode_ggml_buffer_type(apir_decoder * dec) {
+    apir_buffer_type_host_handle_t handle;
+
+    apir_decoder_read(dec, sizeof(handle), &handle, sizeof(handle));
+
+    return (ggml_backend_buffer_type_t) handle;
+}
+
+static inline apir_buffer_type_host_handle_t apir_decode_apir_buffer_type_host_handle(apir_decoder * dec) {
+    apir_buffer_type_host_handle_t handle;
+
+    apir_decoder_read(dec, sizeof(handle), &handle, sizeof(handle));
+
+    return handle;
+}
+
+/* *** ggml_backend_type_t *** */
+
+// ggml_backend_buffer_t is a POINTER.
+// same logic as for ggml_backend_buffer_type_t
+
+static inline void apir_encode_ggml_buffer(apir_encoder * enc, const ggml_backend_buffer_t buffer) {
+    apir_buffer_host_handle_t handle = BUFFER_TO_HOST_HANDLE(buffer);
+    apir_encoder_write(enc, sizeof(handle), &handle, sizeof(handle));
+}
+
+static inline ggml_backend_buffer_t apir_decode_ggml_buffer(apir_decoder * dec) {
+    ggml_backend_buffer_t buffer;
+    size_t                buffer_ptr_size = sizeof(buffer);
+
+    apir_decoder_read(dec, buffer_ptr_size, &buffer, buffer_ptr_size);
+
+    return buffer;
+}
+
+/* enum ggml_status */
+
+static inline void apir_encode_ggml_status(apir_encoder * enc, const ggml_status * status) {
+    apir_encoder_write(enc, sizeof(*status), status, sizeof(*status));
+}
+
+static inline void apir_decode_ggml_status(apir_decoder * dec, ggml_status * status) {
+    apir_decoder_read(dec, sizeof(*status), status, sizeof(*status));
+}
+
+/* virtgpu_shmem */
+
+static inline void apir_encode_virtgpu_shmem_res_id(apir_encoder * enc, uint32_t shmem_res_id) {
+    apir_encode_uint32_t(enc, &shmem_res_id);
+}
+
+static inline void apir_decode_virtgpu_shmem_res_id(apir_decoder * dec, uint32_t * shmem_res_id) {
+    apir_decode_uint32_t(dec, shmem_res_id);
+}
+
+/* ggml_cgraph */
+
+static inline size_t apir_serialize_ggml_cgraph(ggml_cgraph * cgraph, std::vector<uint8_t> & cgraph_data) {
+    apir_serialize_graph(cgraph, cgraph_data);
+
+    return cgraph_data.size();
+}
+
+static inline void apir_encode_cgraph_data(apir_encoder * enc, std::vector<uint8_t> & cgraph_data) {
+    size_t cgraph_size = cgraph_data.size();
+
+    apir_encode(enc, cgraph_size, cgraph_data.data(), cgraph_size);
+}
+
+static inline ggml_cgraph * apir_decode_ggml_cgraph(apir_decoder * dec, size_t cgraph_size) {
+    GGML_UNUSED(cgraph_size);
+
+    uint32_t n_nodes;
+    apir_decode_uint32_t(dec, &n_nodes);
+    const uint64_t * nodes = apir_decode_uint64_t_array_inplace(dec, n_nodes);
+
+    uint32_t n_tensors;
+    apir_decode_uint32_t(dec, &n_tensors);
+    const apir_rpc_tensor * tensors = apir_decode_apir_rpc_tensor_array_inplace(dec, n_tensors);
+
+    return apir_deserialize_graph(n_nodes, n_tensors, tensors, nodes);
+}
+
+static inline void apir_encode_ggml_buffer_handle(apir_encoder * enc, const apir_buffer_host_handle_t * handle) {
+    apir_encoder_write(enc, sizeof(*handle), &handle, sizeof(*handle));
+}
+
+static inline void apir_encode_ggml_tensor_inline(apir_encoder * enc, const ggml_tensor * tensor) {
+    size_t tensor_size = sizeof(*tensor);
+
+    if (tensor->extra) {
+        GGML_ABORT("Cannot pass tensors with extra");
+    }
+
+    if (tensor->src[0] && tensor->buffer) {
+        static int first = 1;
+        if (first) {
+            GGML_LOG_WARN("Cannot pass tensors with src and buffer\n");
+            first = 0;
+        }
+    }
+
+    apir_encoder_write(enc, tensor_size, tensor, tensor_size);
+
+    // tensor->data is a pointer inside the device buffer. No need to touch it
+    // tensor->buffer is a pointer to a buffer. Encoding the buffer handle in sequence.
+    // (could also make a copy of the tensor, and update locally.)
+
+    if (tensor->buffer) {
+        apir_buffer_host_handle_t buffer_handle = ggml_buffer_to_apir_handle(tensor->buffer);
+        apir_encode_ggml_buffer_handle(enc, &buffer_handle);
+    }
+
+    if (tensor->view_src) {
+        apir_encoder_write(enc, tensor_size, tensor->view_src, tensor_size);
+    }
+
+    for (int i = 0; tensor->src[i]; i++) {
+        const ggml_tensor * tensor_src = tensor->src[i];
+        apir_encoder_write(enc, tensor_size, tensor_src, tensor_size);
+    }
+}
+
+static inline const ggml_tensor * apir_decode_ggml_tensor_inplace(apir_decoder * dec) {
+    // it safe to remove the `const` qualifier here, we *do* want to
+    // modify the shared memory data to fix the `src` pointers.
+    ggml_tensor * tensor = (ggml_tensor *) (uintptr_t) apir_decoder_use_inplace(dec, sizeof(ggml_tensor));
+
+    // tensor->data is a pointer inside the device buffer. No need to touch it
+    // tensor->buffer is a pointer to a buffer. Decode the buffer handle encoded in sequence.
+    if (tensor->buffer) {
+        tensor->buffer = apir_decode_ggml_buffer(dec);
+    }
+
+    if (tensor->view_src) {
+        ggml_tensor * tensor_view_src = (ggml_tensor *) (uintptr_t) apir_decoder_use_inplace(dec, sizeof(ggml_tensor));
+        tensor->view_src              = tensor_view_src;
+    }
+
+    for (int i = 0; tensor->src[i]; i++) {
+        ggml_tensor * tensor_src = (ggml_tensor *) (uintptr_t) apir_decoder_use_inplace(dec, sizeof(ggml_tensor));
+        tensor->src[i] = tensor_src;  // overwrite op->src[i] pointer with the actual location of the src tensor
+    }
+
+    return tensor;
+}
diff --git a/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h b/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h
new file mode 100644
index 0000000000..f681798952
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/backend/shared/apir_cs_rpc.h
@@ -0,0 +1,54 @@
+#include "ggml.h"
+#include "ggml-backend-impl.h"
+
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+#include <cstdint>
+
+// ggml_tensor is serialized into apir_rpc_tensor
+struct apir_rpc_tensor {
+    uint64_t id;
+    uint32_t type;
+    uint64_t buffer;
+    uint32_t ne[GGML_MAX_DIMS];
+    uint32_t nb[GGML_MAX_DIMS];
+    uint32_t op;
+    int32_t  op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
+    int32_t  flags;
+    uint64_t src[GGML_MAX_SRC];
+    uint64_t view_src;
+    uint64_t view_offs;
+    uint64_t data;
+    char     name[GGML_MAX_NAME];
+
+    char padding[4];
+};
+
+/* frontend */
+
+apir_rpc_tensor apir_serialize_tensor(const ggml_tensor * tensor);
+
+void apir_serialize_graph(const ggml_cgraph * cgraph, std::vector<uint8_t> & output);
+
+/* backend */
+
+void                                      apir_track_backend_buffer(ggml_backend_buffer_t buffer);
+bool                                      apir_untrack_backend_buffer(ggml_backend_buffer_t buffer);
+std::unordered_set<ggml_backend_buffer_t> apir_get_track_backend_buffers();
+
+void apir_add_tensor(ggml_tensor *                       tensor,
+                     std::vector<apir_rpc_tensor> &      tensors,
+                     std::unordered_set<ggml_tensor *> & visited);
+
+ggml_tensor * apir_deserialize_tensor(ggml_context * ctx, const apir_rpc_tensor * tensor);
+
+ggml_tensor * apir_create_node(uint64_t                                                      id,
+                               ggml_context *                                                ctx,
+                               const std::unordered_map<uint64_t, const apir_rpc_tensor *> & tensor_ptrs,
+                               std::unordered_map<uint64_t, ggml_tensor *> &                 tensor_map);
+
+ggml_cgraph * apir_deserialize_graph(uint32_t                n_nodes,
+                                     uint32_t                n_tensors,
+                                     const apir_rpc_tensor * tensors,
+                                     const uint64_t *        nodes);
diff --git a/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp b/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp
new file mode 100644
index 0000000000..7f650659b8
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp
@@ -0,0 +1,98 @@
+#include "ggml-remoting.h"
+
+static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
+                                                                            size_t                     size) {
+    virtgpu * gpu = BUFT_TO_GPU(buft);
+
+    ggml_backend_remoting_buffer_context * context = (ggml_backend_remoting_buffer_context *) malloc(sizeof(*context));
+    if (!context) {
+        GGML_ABORT("Couldn't allocate the buffer context ...");
+    }
+
+    context->gpu = gpu;
+
+    bool async__unused, host_buffer__unused, events__unused;
+    bool buffer_from_host_ptr;
+    apir_device_get_props(gpu, &async__unused, &host_buffer__unused, &buffer_from_host_ptr, &events__unused);
+
+    if (buffer_from_host_ptr) {
+        context->apir_context = apir_device_buffer_from_ptr(gpu, size, size);
+        context->base         = context->apir_context.shmem.mmap_ptr;
+        context->is_from_ptr  = true;
+    } else {
+        context->apir_context = apir_buffer_type_alloc_buffer(gpu, buft, size);
+        context->is_from_ptr  = false;
+        context->base         = NULL;
+    }
+
+    ggml_backend_buffer_t buffer =
+        ggml_backend_buffer_init(buft, ggml_backend_remoting_buffer_interface, (void *) context, size);
+
+    return buffer;
+}
+
+static const char * ggml_backend_remoting_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    virtgpu * gpu = BUFT_TO_GPU(buft);
+
+    return apir_buffer_type_get_name(gpu, buft);
+}
+
+static size_t ggml_backend_remoting_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    virtgpu * gpu = BUFT_TO_GPU(buft);
+
+    static size_t align = 0;
+
+    if (align == 0) {
+        align = apir_buffer_type_get_alignment(gpu, buft);
+    }
+
+    return align;
+}
+
+static size_t ggml_backend_remoting_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
+    virtgpu * gpu = BUFT_TO_GPU(buft);
+
+    static size_t max_size = 0;
+    if (max_size == 0) {
+        max_size = apir_buffer_type_get_max_size(gpu, buft);
+    }
+
+    return max_size;
+}
+
+static bool ggml_backend_remoting_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    virtgpu * gpu = BUFT_TO_GPU(buft);
+
+    return apir_buffer_type_is_host(gpu, buft);
+}
+
+static size_t ggml_backend_remoting_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft,
+                                                               const ggml_tensor *        tensor) {
+    virtgpu * gpu = BUFT_TO_GPU(buft);
+
+    if (tensor->buffer == NULL
+        || !tensor->buffer->context
+        || !buft->device->iface.supports_buft(buft->device, tensor->buffer->buft)) {
+        return ggml_nbytes(tensor);
+    }
+
+    return apir_buffer_type_get_alloc_size(gpu, buft, tensor);
+}
+
+const ggml_backend_buffer_type_i ggml_backend_remoting_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_remoting_buffer_type_get_name,
+    /* .alloc_buffer     = */ ggml_backend_remoting_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_remoting_buffer_type_get_alignment,
+    /* .get_max_size     = */ ggml_backend_remoting_buffer_type_get_max_size,
+    /* .get_alloc_size   = */ ggml_backend_remoting_buffer_type_get_alloc_size,
+    /* .is_host          = */ NULL,
+};
+
+const ggml_backend_buffer_type_i ggml_backend_remoting_buffer_from_ptr_type_interface = {
+    /* .get_name         = */ ggml_backend_remoting_buffer_type_get_name,
+    /* .alloc_buffer     = */ NULL,
+    /* .get_alignment    = */ ggml_backend_remoting_buffer_type_get_alignment,
+    /* .get_max_size     = */ ggml_backend_remoting_buffer_type_get_max_size,
+    /* .get_alloc_size   = */ ggml_backend_remoting_buffer_type_get_alloc_size,
+    /* .is_host          = */ NULL,
+};
diff --git a/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp b/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp
new file mode 100644
index 0000000000..6b95362dd8
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/ggml-backend-buffer.cpp
@@ -0,0 +1,119 @@
+#include "ggml-remoting.h"
+
+#define BUFFER_TO_GPU(name) ((ggml_backend_remoting_buffer_context *) (name)->context)->gpu
+
+static void * ggml_backend_remoting_buffer_get_base(ggml_backend_buffer_t buffer) {
+    ggml_backend_remoting_buffer_context * context = (ggml_backend_remoting_buffer_context *) buffer->context;
+    if (context->base) {
+        return context->base;
+    }
+
+    context->base = apir_buffer_get_base(BUFFER_TO_GPU(buffer), BUFFER_TO_APIR_CONTEXT(buffer));
+
+    return context->base;
+}
+
+static void ggml_backend_remoting_buffer_set_tensor(ggml_backend_buffer_t buffer,
+                                                    ggml_tensor *         tensor,
+                                                    const void *          data,
+                                                    size_t                offset,
+                                                    size_t                size) {
+    virtgpu * gpu = BUFFER_TO_GPU(buffer);
+
+    ggml_backend_remoting_buffer_context * context = BUFFER_TO_GGML_CONTEXT(buffer);
+    if (context->is_from_ptr) {
+        memcpy((char *) tensor->data + offset, data, size);
+    } else {
+        apir_buffer_set_tensor(gpu, BUFFER_TO_APIR_CONTEXT(buffer), tensor, data, offset, size);
+    }
+
+    return;
+}
+
+static void ggml_backend_remoting_buffer_get_tensor(ggml_backend_buffer_t buffer,
+                                                    const ggml_tensor *   tensor,
+                                                    void *                data,
+                                                    size_t                offset,
+                                                    size_t                size) {
+    virtgpu *                              gpu     = BUFFER_TO_GPU(buffer);
+    ggml_backend_remoting_buffer_context * context = BUFFER_TO_GGML_CONTEXT(buffer);
+    if (context->is_from_ptr) {
+        memcpy(data, (const char *) tensor->data + offset, size);
+    } else {
+        apir_buffer_get_tensor(gpu, BUFFER_TO_APIR_CONTEXT(buffer), tensor, data, offset, size);
+    }
+}
+
+static void ggml_backend_remoting_buffer_set_tensor_from_ptr(ggml_backend_buffer_t buffer,
+                                                             ggml_tensor *         tensor,
+                                                             const void *          data,
+                                                             size_t                offset,
+                                                             size_t                size) {
+    UNUSED(buffer);
+
+    memcpy((char *) tensor->data + offset, data, size);
+
+    return;
+}
+
+static void ggml_backend_remoting_buffer_get_tensor_from_ptr(ggml_backend_buffer_t buffer,
+                                                             const ggml_tensor *   tensor,
+                                                             void *                data,
+                                                             size_t                offset,
+                                                             size_t                size) {
+    UNUSED(buffer);
+
+    memcpy(data, (const char *) tensor->data + offset, size);
+}
+
+static bool ggml_backend_remoting_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
+                                                    const ggml_tensor *   src,
+                                                    ggml_tensor *         dst) {
+    virtgpu * gpu = BUFFER_TO_GPU(buffer);
+
+    bool ret = apir_buffer_cpy_tensor(gpu, BUFFER_TO_APIR_CONTEXT(buffer), src, dst);
+
+    return ret;
+}
+
+static void ggml_backend_remoting_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    virtgpu * gpu = BUFFER_TO_GPU(buffer);
+
+    apir_buffer_clear(gpu, BUFFER_TO_APIR_CONTEXT(buffer), value);
+
+    return;
+}
+
+static void ggml_backend_remoting_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    virtgpu * gpu = BUFFER_TO_GPU(buffer);
+
+    apir_buffer_free_buffer(gpu, BUFFER_TO_APIR_CONTEXT(buffer));
+
+    ggml_backend_remoting_buffer_context * context = BUFFER_TO_GGML_CONTEXT(buffer);
+    free(context);
+    buffer->context = NULL;
+}
+
+const ggml_backend_buffer_i ggml_backend_remoting_buffer_interface = {
+    /* .free_buffer     = */ ggml_backend_remoting_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_remoting_buffer_get_base,
+    /* .init_tensor     = */ NULL,
+    /* .memset_tensor   = */ NULL,
+    /* .set_tensor      = */ ggml_backend_remoting_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_remoting_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_remoting_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_remoting_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+const ggml_backend_buffer_i ggml_backend_remoting_buffer_from_ptr_interface = {
+    /* .free_buffer     = */ ggml_backend_remoting_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_remoting_buffer_get_base,
+    /* .init_tensor     = */ NULL,
+    /* .memset_tensor   = */ NULL,
+    /* .set_tensor      = */ ggml_backend_remoting_buffer_set_tensor_from_ptr,
+    /* .get_tensor      = */ ggml_backend_remoting_buffer_get_tensor_from_ptr,
+    /* .cpy_tensor      = */ ggml_backend_remoting_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_remoting_buffer_clear,
+    /* .reset           = */ NULL,
+};
diff --git a/ggml/src/ggml-virtgpu/ggml-backend-device.cpp b/ggml/src/ggml-virtgpu/ggml-backend-device.cpp
new file mode 100644
index 0000000000..579eb99078
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/ggml-backend-device.cpp
@@ -0,0 +1,144 @@
+#include "ggml-remoting.h"
+
+static const char * ggml_backend_remoting_device_get_name(ggml_backend_dev_t dev) {
+    virtgpu * gpu = DEV_TO_GPU(dev);
+
+    return apir_device_get_name(gpu);
+}
+
+static const char * ggml_backend_remoting_device_get_description(ggml_backend_dev_t dev) {
+    virtgpu * gpu = DEV_TO_GPU(dev);
+
+    return apir_device_get_description(gpu);
+}
+
+static enum ggml_backend_dev_type ggml_backend_remoting_device_get_type(ggml_backend_dev_t dev) {
+    virtgpu * gpu = DEV_TO_GPU(dev);
+
+    static enum ggml_backend_dev_type type;
+    static bool                       has_type = false;
+    if (!has_type) {
+        has_type = true;
+        type     = (enum ggml_backend_dev_type) apir_device_get_type(gpu);
+    }
+
+    return type;
+}
+
+static void ggml_backend_remoting_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    virtgpu * gpu = DEV_TO_GPU(dev);
+
+    return apir_device_get_memory(gpu, free, total);
+}
+
+static bool ggml_backend_remoting_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
+#if USE_ALWAYS_TRUE_SUPPORTS_OP == 1
+    /* ggml-rpc cheats it like this */
+    /* with the current implementation of serialize_tensor, the src/view aren't properly passed */
+    UNUSED(dev);
+    UNUSED(op);
+
+    return true;
+#else
+    virtgpu * gpu = DEV_TO_GPU(dev);
+
+    return apir_device_supports_op(gpu, op);
+#endif
+}
+
+static bool ggml_backend_remoting_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    bool supported = buft->device == dev;
+
+    return supported;
+}
+
+static bool ggml_backend_remoting_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
+    UNUSED(dev);
+    UNUSED(op);
+
+    return false;
+}
+
+static void ggml_backend_remoting_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_remoting_device_get_name(dev);
+    props->description = ggml_backend_remoting_device_get_description(dev);
+    props->type        = ggml_backend_remoting_device_get_type(dev);
+    ggml_backend_remoting_device_get_memory(dev, &props->memory_free, &props->memory_total);
+
+    virtgpu * gpu = DEV_TO_GPU(dev);
+    apir_device_get_props(gpu, &props->caps.async, &props->caps.host_buffer, &props->caps.buffer_from_host_ptr,
+                          &props->caps.events);
+
+    props->caps.buffer_from_host_ptr = false;
+    props->caps.async                = false;
+    props->caps.events               = false;
+}
+
+ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_type(ggml_backend_dev_t dev) {
+    virtgpu * gpu = DEV_TO_GPU(dev);
+
+    apir_buffer_type_host_handle_t ctx = apir_device_get_buffer_type(gpu);
+
+    static ggml_backend_buffer_type buft{
+        /* .iface    = */ ggml_backend_remoting_buffer_type_interface,
+        /* .device   = */ dev,
+        /* .context  = */ (void *) ctx,
+    };
+
+    return &buft;
+}
+
+static ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_from_ptr_type(ggml_backend_dev_t dev) {
+    virtgpu * gpu = DEV_TO_GPU(dev);
+
+    apir_buffer_type_host_handle_t ctx = apir_device_get_buffer_type(gpu);
+
+    static ggml_backend_buffer_type buft{
+        /* .iface    = */ ggml_backend_remoting_buffer_from_ptr_type_interface,
+        /* .device   = */ dev,
+        /* .context  = */ (void *) ctx,
+    };
+
+    return &buft;
+}
+
+static ggml_backend_buffer_t ggml_backend_remoting_device_buffer_from_ptr(ggml_backend_dev_t dev,
+                                                                          void *             ptr,
+                                                                          size_t             size,
+                                                                          size_t             max_tensor_size) {
+    virtgpu * gpu = DEV_TO_GPU(dev);
+
+    ggml_backend_remoting_buffer_context * context = (ggml_backend_remoting_buffer_context *) malloc(sizeof(*context));
+    if (!context) {
+        GGML_ABORT("Couldn't allocate the buffer context ...");
+    }
+
+    context->gpu          = gpu;
+    context->apir_context = apir_device_buffer_from_ptr(gpu, size, max_tensor_size);
+    context->base         = ptr;
+    context->is_from_ptr  = true;
+
+    ggml_backend_buffer_t buffer =
+        ggml_backend_buffer_init(ggml_backend_remoting_device_get_buffer_from_ptr_type(dev),
+                                 ggml_backend_remoting_buffer_from_ptr_interface, (void *) context, size);
+
+    return buffer;
+}
+
+const ggml_backend_device_i ggml_backend_remoting_device_interface = {
+    /* .get_name             = */ ggml_backend_remoting_device_get_name,
+    /* .get_description      = */ ggml_backend_remoting_device_get_description,
+    /* .get_memory           = */ ggml_backend_remoting_device_get_memory,
+    /* .get_type             = */ ggml_backend_remoting_device_get_type,
+    /* .get_props            = */ ggml_backend_remoting_device_get_props,
+    /* .init_backend         = */ ggml_backend_remoting_device_init,
+    /* .get_buffer_type      = */ ggml_backend_remoting_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ ggml_backend_remoting_device_buffer_from_ptr,
+    /* .supports_op          = */ ggml_backend_remoting_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_remoting_device_supports_buft,
+    /* .offload_op           = */ ggml_backend_remoting_device_offload_op,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
diff --git a/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp b/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp
new file mode 100644
index 0000000000..c46cf51c02
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/ggml-backend-reg.cpp
@@ -0,0 +1,137 @@
+#include "ggml-remoting.h"
+#include "ggml-virtgpu.h"
+
+#include <iostream>
+#include <mutex>
+
+static virtgpu * apir_initialize() {
+    static virtgpu * apir_gpu_instance = NULL;
+    static bool      apir_initialized  = false;
+
+    {
+        static std::mutex           mutex;
+        std::lock_guard<std::mutex> lock(mutex);
+
+        if (apir_initialized) {
+            return apir_gpu_instance;
+        }
+
+        apir_gpu_instance = create_virtgpu();
+        if (!apir_gpu_instance) {
+            GGML_ABORT("failed to initialize the virtgpu");
+        }
+
+        apir_initialized = true;
+    }
+
+    return apir_gpu_instance;
+}
+
+static int ggml_backend_remoting_get_device_count() {
+    virtgpu * gpu = apir_initialize();
+    if (!gpu) {
+        GGML_LOG_WARN("apir_initialize failed\n");
+        return 0;
+    }
+
+    return apir_device_get_count(gpu);
+}
+
+static size_t ggml_backend_remoting_reg_get_device_count(ggml_backend_reg_t reg) {
+    UNUSED(reg);
+
+    return ggml_backend_remoting_get_device_count();
+}
+
+static std::vector<ggml_backend_dev_t> devices;
+
+ggml_backend_dev_t ggml_backend_remoting_get_device(size_t device) {
+    GGML_ASSERT(device < devices.size());
+    return devices[device];
+}
+
+static void ggml_backend_remoting_reg_init_devices(ggml_backend_reg_t reg) {
+    if (devices.size() > 0) {
+        GGML_LOG_INFO("%s: already initialized\n", __func__);
+        return;
+    }
+
+    virtgpu * gpu = apir_initialize();
+    if (!gpu) {
+        GGML_LOG_ERROR("apir_initialize failed\n");
+        return;
+    }
+
+    static bool initialized = false;
+
+    {
+        static std::mutex           mutex;
+        std::lock_guard<std::mutex> lock(mutex);
+        if (!initialized) {
+            for (int i = 0; i < ggml_backend_remoting_get_device_count(); i++) {
+                ggml_backend_remoting_device_context * ctx       = new ggml_backend_remoting_device_context;
+                char                                   desc[256] = "API Remoting device";
+
+                ctx->device      = i;
+                ctx->name        = GGML_REMOTING_FRONTEND_NAME + std::to_string(i);
+                ctx->description = desc;
+                ctx->gpu         = gpu;
+
+                ggml_backend_dev_t dev = new ggml_backend_device{
+                    /* .iface   = */ ggml_backend_remoting_device_interface,
+                    /* .reg     = */ reg,
+                    /* .context = */ ctx,
+                };
+                devices.push_back(dev);
+            }
+            initialized = true;
+        }
+    }
+}
+
+static ggml_backend_dev_t ggml_backend_remoting_reg_get_device(ggml_backend_reg_t reg, size_t device) {
+    UNUSED(reg);
+
+    return ggml_backend_remoting_get_device(device);
+}
+
+static const char * ggml_backend_remoting_reg_get_name(ggml_backend_reg_t reg) {
+    UNUSED(reg);
+
+    return GGML_REMOTING_FRONTEND_NAME;
+}
+
+static const ggml_backend_reg_i ggml_backend_remoting_reg_i = {
+    /* .get_name         = */ ggml_backend_remoting_reg_get_name,
+    /* .get_device_count = */ ggml_backend_remoting_reg_get_device_count,
+    /* .get_device       = */ ggml_backend_remoting_reg_get_device,
+    /* .get_proc_address = */ NULL,
+};
+
+ggml_backend_reg_t ggml_backend_virtgpu_reg() {
+    virtgpu * gpu = apir_initialize();
+    if (!gpu) {
+        GGML_LOG_ERROR("virtgpu_apir_initialize failed\n");
+        return NULL;
+    }
+
+    static ggml_backend_reg reg = {
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_remoting_reg_i,
+        /* .context     = */ gpu,
+    };
+
+    static bool initialized = false;
+    if (initialized) {
+        return &reg;
+    }
+    initialized = true;
+
+    ggml_backend_remoting_reg_init_devices(&reg);
+
+    GGML_LOG_INFO("%s: initialized\n", __func__);
+
+    return &reg;
+}
+
+GGML_BACKEND_DL_IMPL(ggml_backend_virtgpu_reg)
diff --git a/ggml/src/ggml-virtgpu/ggml-backend.cpp b/ggml/src/ggml-virtgpu/ggml-backend.cpp
new file mode 100644
index 0000000000..5cd6c0c060
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/ggml-backend.cpp
@@ -0,0 +1,69 @@
+#include "ggml-remoting.h"
+#include "../../include/ggml-virtgpu.h"
+
+static const char * ggml_backend_remoting_get_name(ggml_backend_t backend) {
+    UNUSED(backend);
+
+    return "API Remoting backend";
+}
+
+static void ggml_backend_remoting_free(ggml_backend_t backend) {
+    delete backend;
+}
+
+static ggml_status ggml_backend_remoting_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+    virtgpu * gpu = DEV_TO_GPU(backend->device);
+
+    return apir_backend_graph_compute(gpu, cgraph);
+}
+
+static void ggml_backend_remoting_graph_optimize(ggml_backend_t backend, ggml_cgraph * cgraph) {
+    virtgpu * gpu = DEV_TO_GPU(backend->device);
+#if true
+    UNUSED(gpu);
+    UNUSED(cgraph);
+#else
+    // not working yet
+
+    apir_backend_graph_optimize(gpu, cgraph);
+#endif
+}
+
+static ggml_backend_i ggml_backend_remoting_interface = {
+    /* .get_name                = */ ggml_backend_remoting_get_name,
+    /* .free                    = */ ggml_backend_remoting_free,
+    /* .set_tensor_async        = */ NULL,  // ggml_backend_remoting_set_tensor_async,
+    /* .get_tensor_async        = */ NULL,  // ggml_backend_remoting_get_tensor_async,
+    /* .cpy_tensor_async        = */ NULL,  // ggml_backend_remoting_cpy_tensor_async,
+    /* .synchronize             = */ NULL,  // ggml_backend_remoting_synchronize,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_remoting_graph_compute,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+    /* .graph_optimize          = */ ggml_backend_remoting_graph_optimize,
+};
+
+static ggml_guid_t ggml_backend_remoting_guid() {
+    static ggml_guid guid = { 0xb8, 0xf7, 0x4f, 0x86, 0x14, 0x03, 0x86, 0x02,
+                              0x91, 0xc8, 0xdd, 0xe9, 0x02, 0x3f, 0xc0, 0x2b };
+
+    return &guid;
+}
+
+ggml_backend_t ggml_backend_remoting_device_init(ggml_backend_dev_t dev, const char * params) {
+    UNUSED(params);
+
+    ggml_backend_remoting_device_context * ctx = (ggml_backend_remoting_device_context *) dev->context;
+
+    ggml_backend_t remoting_backend = new ggml_backend{
+        /* .guid      = */ ggml_backend_remoting_guid(),
+        /* .interface = */ ggml_backend_remoting_interface,
+        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_virtgpu_reg(), ctx->device),
+        /* .context   = */ ctx,
+    };
+
+    return remoting_backend;
+}
diff --git a/ggml/src/ggml-virtgpu/ggml-remoting.h b/ggml/src/ggml-virtgpu/ggml-remoting.h
new file mode 100644
index 0000000000..36fc6b2a7b
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/ggml-remoting.h
@@ -0,0 +1,68 @@
+#pragma once
+
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+#include "virtgpu.h"
+
+#include <memory>
+#include <string>
+
+// USE_ALWAYS_TRUE_SUPPORTS_OP: 1 is fast, 0 avoid micro-benchmark crashes
+
+#define USE_ALWAYS_TRUE_SUPPORTS_OP 1
+#define USE_METAL_GUEST_SUPPORTS_OP 0
+
+#define DEV_TO_GPU(name) ((ggml_backend_remoting_device_context *) (name)->context)->gpu
+
+#define BUFFER_TO_GGML_CONTEXT(name) ((ggml_backend_remoting_buffer_context *) (name)->context)
+
+#define BUFFER_TO_APIR_CONTEXT(name) &((ggml_backend_remoting_buffer_context *) (name)->context)->apir_context
+
+#define BUFFER_TO_HOST_HANDLE(name) ((ggml_backend_remoting_buffer_context *) (name)->context)->apir_context.host_handle
+
+#define GET_DEVICE_CONTEXT() (ggml_backend_remoting_device_context *) ggml_backend_remoting_get_device(0)->context
+
+#define BUFT_TO_GPU(name) ((ggml_backend_remoting_device_context *) (name)->device->context)->gpu
+
+struct ggml_backend_remoting_device_context {
+    size_t      device;
+    std::string name;
+    std::string description;
+
+    std::vector<std::tuple<void *, size_t, virtgpu_shmem *>> shared_memory;
+
+    virtgpu * gpu;
+};
+
+struct ggml_backend_remoting_buffer_context {
+    apir_buffer_context_t apir_context;
+
+    virtgpu * gpu;
+
+    void * base;
+
+    bool is_from_ptr;
+};
+
+extern const ggml_backend_buffer_type_i ggml_backend_remoting_buffer_type_interface;
+extern const ggml_backend_device_i      ggml_backend_remoting_device_interface;
+extern const ggml_backend_buffer_i      ggml_backend_remoting_buffer_interface;
+extern const ggml_backend_buffer_type_i ggml_backend_remoting_buffer_from_ptr_type_interface;
+extern const ggml_backend_buffer_i      ggml_backend_remoting_buffer_from_ptr_interface;
+
+ggml_backend_dev_t         ggml_backend_remoting_get_device(size_t device);
+ggml_backend_t             ggml_backend_remoting_device_init(ggml_backend_dev_t dev, const char * params);
+ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_type(ggml_backend_dev_t dev);
+
+static inline apir_buffer_type_host_handle_t ggml_buffer_type_to_apir_handle(ggml_backend_buffer_type_t buft) {
+    // in the backend, the buffer handle is the buffer pointer
+    return (apir_buffer_type_host_handle_t) buft->context;
+}
+
+static inline apir_buffer_host_handle_t ggml_buffer_to_apir_handle(ggml_backend_buffer_t buffer) {
+    if (!buffer->context) {
+        GGML_ABORT("%s: no context available :/", __func__);
+    }
+    return BUFFER_TO_HOST_HANDLE(buffer);
+}
diff --git a/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml b/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml
new file mode 100644
index 0000000000..0b7cccfe9c
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml
@@ -0,0 +1,168 @@
+# YAML schema for GGML remoting API functions
+# This defines the structure for generating the remoting layer code
+
+# Configuration for the generated files
+config:
+  # Base path for the generated files
+  base_path: "ggml/src"
+
+  # Header files to update
+  files:
+    apir_backend_header: "ggml-virtgpu-apir/backend/shared/apir_backend.gen.h"
+    backend_dispatched_header: "ggml-virtgpu-apir/backend/backend-dispatched.gen.h"
+    virtgpu_forward_header: "ggml-virtgpu-apir/virtgpu-forward.gen.h"
+
+# Simplified function definitions with grouping and metadata combined
+functions:
+  device:
+    group_description: "device"
+    functions:
+      get_device_count:
+        # No specific metadata - uses default void return and base params
+
+      get_count:
+        frontend_return: "int"
+
+      get_name:
+        frontend_return: "const char *"
+
+      get_description:
+        frontend_return: "const char *"
+
+      get_type:
+        frontend_return: "uint32_t"
+
+      get_memory:
+        frontend_return: "void"
+        frontend_extra_params:
+        - "size_t *free"
+        - "size_t *total"
+
+      supports_op:
+        frontend_return: "bool"
+        frontend_extra_params:
+        - "const ggml_tensor *op"
+
+      get_buffer_type:
+        frontend_return: "apir_buffer_type_host_handle_t"
+
+      get_props:
+        frontend_return: "void"
+        frontend_extra_params:
+        - "bool *async"
+        - "bool *host_buffer"
+        - "bool *buffer_from_host_ptr"
+        - "bool *events"
+
+      buffer_from_ptr:
+        frontend_return: "apir_buffer_context_t"
+        frontend_extra_params:
+        - "size_t size"
+        - "size_t max_tensor_size"
+
+  buffer_type:
+    group_description: "buffer-type"
+    functions:
+      get_name:
+        frontend_return: "const char *"
+        frontend_extra_params:
+        - "ggml_backend_buffer_type_t buft"
+
+      get_alignment:
+        frontend_return: "size_t"
+        frontend_extra_params:
+        - "ggml_backend_buffer_type_t buft"
+
+      get_max_size:
+        frontend_return: "size_t"
+        frontend_extra_params:
+        - "ggml_backend_buffer_type_t buft"
+
+      is_host:
+        frontend_return: "bool"
+        frontend_extra_params:
+        - "ggml_backend_buffer_type_t buft"
+
+      alloc_buffer:
+        frontend_return: "apir_buffer_context_t"
+        frontend_extra_params:
+        - "ggml_backend_buffer_type_t buffer_buft"
+        - "size_t size"
+
+      get_alloc_size:
+        frontend_return: "size_t"
+        frontend_extra_params:
+        - "ggml_backend_buffer_type_t buft"
+        - "const ggml_tensor *op"
+
+  buffer:
+    group_description: "buffer"
+    functions:
+      get_base:
+        frontend_return: "void *"
+        frontend_extra_params:
+        - "apir_buffer_context_t *buffer_context"
+
+      set_tensor:
+        frontend_return: "void"
+        frontend_extra_params:
+        - "apir_buffer_context_t *buffer_context"
+        - "ggml_tensor *tensor"
+        - "const void *data"
+        - "size_t offset"
+        - "size_t size"
+
+      get_tensor:
+        frontend_return: "void"
+        frontend_extra_params:
+        - "apir_buffer_context_t *buffer_context"
+        - "const ggml_tensor *tensor"
+        - "void *data"
+        - "size_t offset"
+        - "size_t size"
+
+      cpy_tensor:
+        frontend_return: "bool"
+        frontend_extra_params:
+        - "apir_buffer_context_t *buffer_context"
+        - "const ggml_tensor *src"
+        - "const ggml_tensor *dst"
+
+      clear:
+        frontend_return: "void"
+        frontend_extra_params:
+        - "apir_buffer_context_t *buffer_context"
+        - "uint8_t value"
+
+      free_buffer:
+        frontend_return: "void"
+        frontend_extra_params:
+        - "apir_buffer_context_t *buffer_context"
+
+  backend:
+    group_description: "backend"
+    functions:
+      graph_compute:
+        frontend_return: "ggml_status"
+        frontend_extra_params:
+        - "ggml_cgraph *cgraph"
+
+      graph_optimize:
+        frontend_return: "ggml_cgraph *"
+        frontend_extra_params:
+        - "ggml_cgraph *cgraph"
+        enabled: false
+
+# Naming patterns used for code generation
+naming_patterns:
+  # How to generate enum names
+  enum_prefix: "APIR_COMMAND_TYPE_"
+
+  # How to generate backend function names
+  backend_function_prefix: "backend_"
+
+  # How to generate frontend function names
+  frontend_function_prefix: "apir_"
+
+  # Standard frontend first parameter
+  frontend_base_param: "struct virtgpu *gpu"
diff --git a/ggml/src/ggml-virtgpu/include/apir_hw.h b/ggml/src/ggml-virtgpu/include/apir_hw.h
new file mode 100644
index 0000000000..33af045ca2
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/include/apir_hw.h
@@ -0,0 +1,9 @@
+#pragma once
+
+#include <stdint.h>
+
+struct virgl_renderer_capset_apir {
+   uint32_t apir_version;
+   uint32_t supports_blob_resources;
+   uint32_t reserved[4];           // For future expansion
+};
diff --git a/ggml/src/ggml-virtgpu/regenerate_remoting.py b/ggml/src/ggml-virtgpu/regenerate_remoting.py
new file mode 100755
index 0000000000..4174a24327
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/regenerate_remoting.py
@@ -0,0 +1,322 @@
+#!/usr/bin/env python3
+"""
+# Generated by Claude AI
+
+Script to completely regenerate the GGML remoting codebase from YAML configuration.
+
+This script reads api_functions.yaml and regenerates all the header files and
+implementation templates for the GGML remoting layer.
+
+Usage:
+  python regenerate_remoting.py
+
+The script will:
+1. Read ggmlremoting_functions.yaml configuration
+2. Generate updated header files
+3. Generate implementation templates in dedicated files
+4. Show a summary of what was generated
+"""
+
+import yaml
+from typing import Dict, List, Any
+from pathlib import Path
+import os
+import subprocess
+import shutil
+import logging
+
+NL = '\n' # can't have f"{'\n'}" in f-strings
+
+
+class RemotingCodebaseGenerator:
+    def __init__(self, yaml_path: str = "ggmlremoting_functions.yaml"):
+        """Initialize the generator with the YAML configuration."""
+        self.yaml_path = yaml_path
+
+        if not Path(yaml_path).exists():
+            raise FileNotFoundError(f"Configuration file {yaml_path} not found")
+
+        with open(yaml_path, 'r') as f:
+            self.config = yaml.safe_load(f)
+
+        self.functions = self.config['functions']
+        self.naming_patterns = self.config['naming_patterns']
+        self.config_data = self.config['config']
+
+        # Check if clang-format is available
+        self.clang_format_available = self._check_clang_format_available()
+
+    def _check_clang_format_available(self) -> bool:
+        """Check if clang-format is available in the system PATH."""
+        return shutil.which("clang-format") is not None
+
+    def _format_file_with_clang_format(self, file_path: Path) -> bool:
+        """Format a file with clang-format -i. Returns True if successful, False otherwise."""
+        if not self.clang_format_available:
+            return False
+
+        try:
+            subprocess.run(
+                ["clang-format", "-i", str(file_path)],
+                check=True,
+                capture_output=True,
+                text=True
+            )
+            return True
+        except subprocess.CalledProcessError:
+            logging.exception(f"   ⚠️  clang-format failed for {file_path}")
+            return False
+        except Exception as e:
+            logging.exception(f"   ⚠️  Unexpected error formatting {file_path}: {e}")
+            return False
+
+    def generate_enum_name(self, group_name: str, function_name: str) -> str:
+        """Generate the APIR_COMMAND_TYPE enum name for a function."""
+        prefix = self.naming_patterns['enum_prefix']
+        return f"{prefix}{group_name.upper()}_{function_name.upper()}"
+
+    def generate_backend_function_name(self, group_name: str, function_name: str) -> str:
+        """Generate the backend function name."""
+        function_key = f"{group_name}_{function_name}"
+        overrides = self.naming_patterns.get('backend_function_overrides', {})
+
+        if function_key in overrides:
+            return overrides[function_key]
+
+        prefix = self.naming_patterns['backend_function_prefix']
+        return f"{prefix}{group_name}_{function_name}"
+
+    def generate_frontend_function_name(self, group_name: str, function_name: str) -> str:
+        """Generate the frontend function name."""
+        prefix = self.naming_patterns['frontend_function_prefix']
+        return f"{prefix}{group_name}_{function_name}"
+
+    def get_enabled_functions(self) -> List[Dict[str, Any]]:
+        """Get all enabled functions with their metadata."""
+        functions = []
+        enum_value = 0
+
+        for group_name, group_data in self.functions.items():
+            group_description = group_data['group_description']
+
+            for function_name, func_metadata in group_data['functions'].items():
+                # Handle case where func_metadata is None or empty (functions with only comments)
+                if func_metadata is None:
+                    func_metadata = {}
+
+                # Functions are enabled by default unless explicitly disabled
+                if func_metadata.get('enabled', True):
+                    functions.append({
+                        'group_name': group_name,
+                        'function_name': function_name,
+                        'enum_name': self.generate_enum_name(group_name, function_name),
+                        'enum_value': enum_value,
+                        'backend_function': self.generate_backend_function_name(group_name, function_name),
+                        'frontend_function': self.generate_frontend_function_name(group_name, function_name),
+                        'frontend_return': func_metadata.get('frontend_return', 'void'),
+                        'frontend_extra_params': func_metadata.get('frontend_extra_params', []),
+                        'group_description': group_description,
+                        'newly_added': func_metadata.get('newly_added', False)
+                    })
+                    enum_value += 1
+
+        return functions
+
+    def generate_apir_backend_header(self) -> str:
+        """Generate the complete apir_backend.h file."""
+        functions = self.get_enabled_functions()
+
+        # Generate the enum section
+        enum_lines = ["typedef enum ApirBackendCommandType {"]
+        current_group = None
+
+        for func in functions:
+            # Add comment for new group
+            if func['group_name'] != current_group:
+                enum_lines.append("")
+                enum_lines.append(f"  /* {func['group_description']} */")
+                current_group = func['group_name']
+
+            enum_lines.append(f"  {func['enum_name']} = {func['enum_value']},")
+
+        # Add the count
+        total_count = len(functions)
+        enum_lines.append("\n  // last command_type index + 1")
+        enum_lines.append(f"  APIR_BACKEND_DISPATCH_TABLE_COUNT = {total_count},")
+        enum_lines.append("} ApirBackendCommandType;")
+
+        # Full header template
+        header_content = NL.join(enum_lines) + "\n"
+
+        return header_content
+
+    def generate_backend_dispatched_header(self) -> str:
+        """Generate the complete backend-dispatched.h file."""
+        functions = self.get_enabled_functions()
+
+        # Function declarations
+        decl_lines = []
+        current_group = None
+
+        for func in functions:
+            if func['group_name'] != current_group:
+                decl_lines.append(f"\n/* {func['group_description']} */")
+                current_group = func['group_name']
+
+            signature = "uint32_t"
+            params = "apir_encoder *enc, apir_decoder *dec, virgl_apir_context *ctx"
+            decl_lines.append(f"{signature} {func['backend_function']}({params});")
+
+        # Switch cases
+        switch_lines = []
+        current_group = None
+
+        for func in functions:
+            if func['group_name'] != current_group:
+                switch_lines.append(f"  /* {func['group_description']} */")
+                current_group = func['group_name']
+
+            switch_lines.append(f"  case {func['enum_name']}: return \"{func['backend_function']}\";")
+
+        # Dispatch table
+        table_lines = []
+        current_group = None
+
+        for func in functions:
+            if func['group_name'] != current_group:
+                table_lines.append(f"\n  /* {func['group_description']} */")
+                table_lines.append("")
+                current_group = func['group_name']
+
+            table_lines.append(f"  /* {func['enum_name']}  = */ {func['backend_function']},")
+
+        header_content = f'''\
+#pragma once
+
+{NL.join(decl_lines)}
+
+static inline const char *backend_dispatch_command_name(ApirBackendCommandType type)
+{{
+  switch (type) {{
+{NL.join(switch_lines)}
+
+  default: return "unknown";
+  }}
+}}
+
+extern "C" {{
+static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATCH_TABLE_COUNT] = {{
+  {NL.join(table_lines)}
+}};
+}}
+'''
+        return header_content
+
+    def generate_virtgpu_forward_header(self) -> str:
+        """Generate the complete virtgpu-forward.gen.h file."""
+        functions = self.get_enabled_functions()
+
+        decl_lines = []
+        current_group = None
+
+        for func in functions:
+            if func['group_name'] != current_group:
+                decl_lines.append("")
+                decl_lines.append(f"/* {func['group_description']} */")
+                current_group = func['group_name']
+
+            # Build parameter list
+            params = [self.naming_patterns['frontend_base_param']]
+            params.extend(func['frontend_extra_params'])
+            param_str = ', '.join(params)
+
+            decl_lines.append(f"{func['frontend_return']} {func['frontend_function']}({param_str});")
+
+        header_content = f'''\
+#pragma once
+{NL.join(decl_lines)}
+'''
+        return header_content
+
+    def regenerate_codebase(self) -> None:
+        """Regenerate the entire remoting codebase."""
+        logging.info("🔄 Regenerating GGML Remoting Codebase...")
+        logging.info("=" * 50)
+
+        # Detect if we're running from frontend directory
+        current_dir = os.getcwd()
+        is_frontend_dir = current_dir.endswith('ggml-virtgpu')
+
+        if is_frontend_dir:
+            # Running from ggml/src/ggml-virtgpu-apir
+            logging.info("📍 Detected frontend directory execution")
+            frontend_base = Path(".")
+        else:
+            # Running from project root (fallback to original behavior)
+            logging.info("📍 Detected project root execution")
+            base_path = self.config_data.get('base_path', 'ggml/src')
+            frontend_base = Path(base_path) / "ggml-virtgpu"
+
+        # Compute final file paths
+        backend_base = frontend_base / "backend"
+        apir_backend_path = backend_base / "shared" / "apir_backend.gen.h"
+        backend_dispatched_path = backend_base / "backend-dispatched.gen.h"
+        virtgpu_forward_path = frontend_base / "virtgpu-forward.gen.h"
+
+        # Create output directories for each file
+        apir_backend_path.parent.mkdir(parents=True, exist_ok=True)
+        backend_dispatched_path.parent.mkdir(parents=True, exist_ok=True)
+        virtgpu_forward_path.parent.mkdir(parents=True, exist_ok=True)
+
+        # Generate header files
+        logging.info("📁 Generating header files...")
+
+        apir_backend_content = self.generate_apir_backend_header()
+        apir_backend_path.write_text(apir_backend_content)
+        logging.info(f"   ✅ {apir_backend_path.resolve()}")
+
+        backend_dispatched_content = self.generate_backend_dispatched_header()
+        backend_dispatched_path.write_text(backend_dispatched_content)
+        logging.info(f"   ✅ {backend_dispatched_path.resolve()}")
+
+        virtgpu_forward_content = self.generate_virtgpu_forward_header()
+        virtgpu_forward_path.write_text(virtgpu_forward_content)
+        logging.info(f"   ✅ {virtgpu_forward_path.resolve()}")
+
+        # Format generated files with clang-format
+        generated_files = [apir_backend_path, backend_dispatched_path, virtgpu_forward_path]
+
+        if not self.clang_format_available:
+            logging.warning("\n⚠️clang-format not found in PATH. Generated files will not be formatted."
+                            "   Install clang-format to enable automatic code formatting.")
+        else:
+            logging.info("\n🎨 Formatting files with clang-format...")
+            for file_path in generated_files:
+                if self._format_file_with_clang_format(file_path):
+                    logging.info(f"   ✅ Formatted {file_path.name}")
+                else:
+                    logging.warning(f"   ❌ Failed to format {file_path.name}")
+
+        # Generate summary
+        functions = self.get_enabled_functions()
+        total_functions = len(functions)
+
+        logging.info("\n📊 Generation Summary:")
+        logging.info("=" * 50)
+        logging.info(f"   Total functions: {total_functions}")
+        logging.info(f"   Function groups: {len(self.functions)}")
+        logging.info("   Header files: 3")
+        logging.info(f"   Working directory: {current_dir}")
+
+
+def main():
+    try:
+        generator = RemotingCodebaseGenerator()
+        generator.regenerate_codebase()
+    except Exception as e:
+        logging.exception(f"❌ Error: {e}")
+        exit(1)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/ggml/src/ggml-virtgpu/virtgpu-apir.h b/ggml/src/ggml-virtgpu/virtgpu-apir.h
new file mode 100644
index 0000000000..238f960acd
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/virtgpu-apir.h
@@ -0,0 +1,15 @@
+#include "backend/shared/apir_backend.h"
+#include "ggml-alloc.h"
+#include "ggml-impl.h"
+#include "ggml.h"
+#include "virtgpu-shm.h"
+#include "virtgpu-utils.h"
+
+struct apir_buffer_context_t {
+    apir_buffer_host_handle_t host_handle;
+
+    struct virtgpu_shmem           shmem;
+    apir_buffer_type_host_handle_t buft_host_handle;
+};
+
+#include "virtgpu-forward.gen.h"
diff --git a/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp b/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp
new file mode 100644
index 0000000000..bf3c41011a
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp
@@ -0,0 +1,50 @@
+#include "virtgpu-forward-impl.h"
+
+static long long current_time_ms() {
+    timespec ts;
+    clock_gettime(CLOCK_REALTIME, &ts);  // Use CLOCK_MONOTONIC for elapsed time
+    return (long long) ts.tv_sec * 1000000000LL + ts.tv_nsec;
+}
+
+ggml_status apir_backend_graph_compute(virtgpu * gpu, ggml_cgraph * cgraph) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BACKEND_GRAPH_COMPUTE);
+
+    std::vector<uint8_t> cgraph_data;
+    size_t               cgraph_size = apir_serialize_ggml_cgraph(cgraph, cgraph_data);
+
+    virtgpu_shmem   temp_shmem;  // Local storage for large buffers
+    virtgpu_shmem * shmem = &temp_shmem;
+
+    if (cgraph_size <= gpu->data_shmem.mmap_size) {
+        // prefer the init-time allocated page, if large enough
+        shmem = &gpu->data_shmem;
+    } else if (virtgpu_shmem_create(gpu, cgraph_size, shmem)) {
+        GGML_ABORT("Couldn't allocate the guest-host shared buffer");
+    }
+
+    apir_encode_virtgpu_shmem_res_id(encoder, shmem->res_id);
+
+    apir_encode_size_t(encoder, &cgraph_size);
+
+    char *       shmem_data    = (char *) shmem->mmap_ptr;
+    apir_encoder secondary_enc = apir_new_encoder(shmem_data, cgraph_size);
+
+    apir_encode_cgraph_data(&secondary_enc, cgraph_data);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    ggml_status status = GGML_STATUS_ABORTED;
+    apir_decode_ggml_status(decoder, &status);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    if (shmem != &gpu->data_shmem) {
+        virtgpu_shmem_destroy(gpu, shmem);
+    }
+
+    return status;
+}
diff --git a/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp b/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp
new file mode 100644
index 0000000000..03cb09e064
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp
@@ -0,0 +1,125 @@
+#include "virtgpu-forward-impl.h"
+
+const char * apir_buffer_type_get_name(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME);
+
+    apir_encode_ggml_buffer_type(encoder, buft);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    const size_t string_size = apir_decode_array_size_unchecked(decoder);
+    char *       string      = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
+    if (!string) {
+        GGML_LOG_ERROR("%s: Could not allocate the device name buffer\n", __func__);
+        apir_decoder_set_fatal(decoder);
+    }
+    apir_decode_char_array(decoder, string, string_size);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    return string;
+}
+
+size_t apir_buffer_type_get_alignment(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT);
+
+    apir_encode_ggml_buffer_type(encoder, buft);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    size_t alignment;
+    apir_decode_size_t(decoder, &alignment);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    return alignment;
+}
+
+size_t apir_buffer_type_get_max_size(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE);
+
+    apir_encode_ggml_buffer_type(encoder, buft);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    size_t max_size;
+    apir_decode_size_t(decoder, &max_size);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    return max_size;
+}
+
+bool apir_buffer_type_is_host(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST);
+
+    apir_encode_ggml_buffer_type(encoder, buft);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    bool is_host;
+    apir_decode_bool_t(decoder, &is_host);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    return is_host;
+}
+
+apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, ggml_backend_buffer_type_t buft, size_t size) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    apir_buffer_context_t buffer_context;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER);
+
+    apir_encode_ggml_buffer_type(encoder, buft);
+
+    apir_encode_size_t(encoder, &size);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    apir_decode_apir_buffer_host_handle_t(decoder, &buffer_context.host_handle);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    return buffer_context;
+}
+
+size_t apir_buffer_type_get_alloc_size(virtgpu * gpu, ggml_backend_buffer_type_t buft, const ggml_tensor * op) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE);
+
+    apir_encode_ggml_buffer_type(encoder, buft);
+
+    apir_encode_ggml_tensor_inline(encoder, op);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    size_t alloc_size;
+    apir_decode_size_t(decoder, &alloc_size);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    return alloc_size;
+}
diff --git a/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp b/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp
new file mode 100644
index 0000000000..3181e39440
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp
@@ -0,0 +1,157 @@
+#include "virtgpu-forward-impl.h"
+
+void * apir_buffer_get_base(virtgpu * gpu, apir_buffer_context_t * buffer_context) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_GET_BASE);
+
+    apir_encode_apir_buffer_host_handle_t(encoder, &buffer_context->host_handle);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    uintptr_t base;
+    apir_decode_uintptr_t(decoder, &base);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    return (void *) base;
+}
+
+void apir_buffer_set_tensor(virtgpu *               gpu,
+                            apir_buffer_context_t * buffer_context,
+                            ggml_tensor *           tensor,
+                            const void *            data,
+                            size_t                  offset,
+                            size_t                  size) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_SET_TENSOR);
+
+    apir_encode_apir_buffer_host_handle_t(encoder, &buffer_context->host_handle);
+    apir_encode_ggml_tensor(encoder, tensor);
+
+    virtgpu_shmem   temp_shmem;  // Local storage for large buffers
+    virtgpu_shmem * shmem = &temp_shmem;
+
+    if (size <= gpu->data_shmem.mmap_size) {
+        // prefer the init-time allocated page, if large enough
+        shmem = &gpu->data_shmem;
+
+    } else if (virtgpu_shmem_create(gpu, size, shmem)) {
+        GGML_ABORT("Couldn't allocate the guest-host shared buffer");
+    }
+
+    memcpy(shmem->mmap_ptr, data, size);
+    apir_encode_virtgpu_shmem_res_id(encoder, shmem->res_id);
+
+    apir_encode_size_t(encoder, &offset);
+    apir_encode_size_t(encoder, &size);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    if (shmem != &gpu->data_shmem) {
+        virtgpu_shmem_destroy(gpu, shmem);
+    }
+
+    return;
+}
+
+void apir_buffer_get_tensor(virtgpu *               gpu,
+                            apir_buffer_context_t * buffer_context,
+                            const ggml_tensor *     tensor,
+                            void *                  data,
+                            size_t                  offset,
+                            size_t                  size) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_GET_TENSOR);
+
+    apir_encode_apir_buffer_host_handle_t(encoder, &buffer_context->host_handle);
+    apir_encode_ggml_tensor(encoder, tensor);
+
+    virtgpu_shmem   temp_shmem;  // Local storage for large buffers
+    virtgpu_shmem * shmem = &temp_shmem;
+
+    if (size <= gpu->data_shmem.mmap_size) {
+        // prefer the init-time allocated page, if large enough
+        shmem = &gpu->data_shmem;
+
+    } else if (virtgpu_shmem_create(gpu, size, shmem)) {
+        GGML_ABORT("Couldn't allocate the guest-host shared buffer");
+    }
+
+    apir_encode_virtgpu_shmem_res_id(encoder, shmem->res_id);
+    apir_encode_size_t(encoder, &offset);
+    apir_encode_size_t(encoder, &size);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    memcpy(data, shmem->mmap_ptr, size);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    if (shmem != &gpu->data_shmem) {
+        virtgpu_shmem_destroy(gpu, shmem);
+    }
+}
+
+bool apir_buffer_cpy_tensor(virtgpu *               gpu,
+                            apir_buffer_context_t * buffer_context,
+                            const ggml_tensor *     src,
+                            const ggml_tensor *     dst) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_CPY_TENSOR);
+
+    apir_encode_apir_buffer_host_handle_t(encoder, &buffer_context->host_handle);
+    apir_encode_ggml_tensor(encoder, src);
+    apir_encode_ggml_tensor(encoder, dst);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    bool ret_val;
+    apir_decode_bool_t(decoder, &ret_val);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    return ret_val;
+}
+
+void apir_buffer_clear(virtgpu * gpu, apir_buffer_context_t * buffer_context, uint8_t value) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_CLEAR);
+
+    apir_encode_apir_buffer_host_handle_t(encoder, &buffer_context->host_handle);
+    apir_encode_uint8_t(encoder, &value);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    remote_call_finish(gpu, encoder, decoder);
+}
+
+void apir_buffer_free_buffer(virtgpu * gpu, apir_buffer_context_t * buffer_context) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_FREE_BUFFER);
+
+    apir_encode_apir_buffer_host_handle_t(encoder, &buffer_context->host_handle);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    remote_call_finish(gpu, encoder, decoder);
+}
diff --git a/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp b/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp
new file mode 100644
index 0000000000..3e45e55bdc
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp
@@ -0,0 +1,200 @@
+#include "virtgpu-forward-impl.h"
+#include "virtgpu-shm.h"
+
+int apir_device_get_count(virtgpu * gpu) {
+    static int32_t dev_count = -1;
+    if (dev_count != -1) {
+        return dev_count;
+    }
+
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_COUNT);
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    apir_decode_int32_t(decoder, &dev_count);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    return dev_count;
+}
+
+const char * apir_device_get_name(virtgpu * gpu) {
+    static char * string = nullptr;
+    if (string) {
+        return string;
+    }
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_NAME);
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    const size_t string_size = apir_decode_array_size_unchecked(decoder);
+    string                   = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
+    if (!string) {
+        GGML_LOG_ERROR("%s: Could not allocate the device name buffer\n", __func__);
+        return NULL;
+    }
+    apir_decode_char_array(decoder, string, string_size);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    return string;
+}
+
+const char * apir_device_get_description(virtgpu * gpu) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_DESCRIPTION);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    const size_t string_size = apir_decode_array_size_unchecked(decoder);
+    char *       string      = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
+    if (!string) {
+        GGML_LOG_ERROR("%s: Could not allocate the device description buffer\n", __func__);
+
+        return NULL;
+    }
+    apir_decode_char_array(decoder, string, string_size);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    return string;
+}
+
+uint32_t apir_device_get_type(virtgpu * gpu) {
+    static uint32_t dev_type = 255;
+    if (dev_type != 255) {
+        return dev_type;
+    }
+
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_TYPE);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    apir_decode_uint32_t(decoder, &dev_type);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    return dev_type;
+}
+
+void apir_device_get_memory(virtgpu * gpu, size_t * free, size_t * total) {
+    static size_t         dev_free  = 0;
+    static size_t         dev_total = 0;
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_MEMORY);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    apir_decode_size_t(decoder, &dev_free);
+    apir_decode_size_t(decoder, &dev_total);
+
+    *free  = dev_free;
+    *total = dev_total;
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    return;
+}
+
+bool apir_device_supports_op(virtgpu * gpu, const ggml_tensor * op) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_SUPPORTS_OP);
+
+    apir_encode_ggml_tensor_inline(encoder, op);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    bool supports_op;
+    apir_decode_bool_t(decoder, &supports_op);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    return supports_op;
+}
+
+apir_buffer_type_host_handle_t apir_device_get_buffer_type(virtgpu * gpu) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_BUFFER_TYPE);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    apir_buffer_type_host_handle_t buft_handle;
+    apir_decode_apir_buffer_type_host_handle_t(decoder, &buft_handle);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    return buft_handle;
+}
+
+void apir_device_get_props(virtgpu * gpu,
+                           bool *    async,
+                           bool *    host_buffer,
+                           bool *    buffer_from_host_ptr,
+                           bool *    events) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_PROPS);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    apir_decode_bool_t(decoder, async);
+    apir_decode_bool_t(decoder, host_buffer);
+    apir_decode_bool_t(decoder, buffer_from_host_ptr);
+    apir_decode_bool_t(decoder, events);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    return;
+}
+
+apir_buffer_context_t apir_device_buffer_from_ptr(virtgpu * gpu, size_t size, size_t max_tensor_size) {
+    apir_encoder *        encoder;
+    apir_decoder *        decoder;
+    ApirForwardReturnCode ret;
+
+    apir_buffer_context_t buffer_context;
+
+    REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_BUFFER_FROM_PTR);
+
+    if (virtgpu_shmem_create(gpu, size, &buffer_context.shmem)) {
+        GGML_ABORT("Couldn't allocate the guest-host shared buffer");
+    }
+
+    apir_encode_virtgpu_shmem_res_id(encoder, buffer_context.shmem.res_id);
+
+    apir_encode_size_t(encoder, &size);
+    apir_encode_size_t(encoder, &max_tensor_size);
+
+    REMOTE_CALL(gpu, encoder, decoder, ret);
+
+    apir_decode_apir_buffer_host_handle_t(decoder, &buffer_context.host_handle);
+    buffer_context.buft_host_handle = apir_decode_apir_buffer_type_host_handle(decoder);
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    return buffer_context;
+}
diff --git a/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h b/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h
new file mode 100644
index 0000000000..eea3e7e5a9
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/virtgpu-forward-impl.h
@@ -0,0 +1,29 @@
+#include "virtgpu.h"
+
+#include "ggml-remoting.h"
+#include "backend/shared/apir_backend.h"
+#include "backend/shared/apir_cs_ggml.h"
+
+#include "ggml-backend-impl.h"
+
+#define REMOTE_CALL_PREPARE(gpu_dev_name, encoder_name, apir_command_type__)                               \
+    do {                                                                                                   \
+        int32_t forward_flag = (int32_t) apir_command_type__;                                              \
+        encoder_name         = remote_call_prepare(gpu_dev_name, APIR_COMMAND_TYPE_FORWARD, forward_flag); \
+        if (!encoder_name) {                                                                               \
+            GGML_ABORT("%s: failed to prepare the remote call encoder", __func__);                       \
+        }                                                                                                  \
+    } while (0)
+
+#define REMOTE_CALL(gpu_dev_name, encoder_name, decoder_name, ret_name)                                           \
+    do {                                                                                                          \
+        ret_name = (ApirForwardReturnCode) remote_call(gpu_dev_name, encoder_name, &decoder_name, 0, NULL);       \
+        if (!decoder_name) {                                                                                      \
+            GGML_ABORT("%s: failed to kick the remote call", __func__);                                         \
+        }                                                                                                         \
+        if (ret_name < APIR_FORWARD_BASE_INDEX) {                                                                 \
+            GGML_ABORT("%s: failed to forward the API call: %s: code %d", __func__,                             \
+                       apir_forward_error(ret_name), ret_name);                                                   \
+        }                                                                                                         \
+        ret_name = (ApirForwardReturnCode) (ret_name - APIR_FORWARD_BASE_INDEX);                                  \
+    } while (0)
diff --git a/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h b/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h
new file mode 100644
index 0000000000..c27c07f086
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/virtgpu-forward.gen.h
@@ -0,0 +1,51 @@
+#pragma once
+
+/* device */
+void                           apir_device_get_device_count(struct virtgpu * gpu);
+int                            apir_device_get_count(struct virtgpu * gpu);
+const char *                   apir_device_get_name(struct virtgpu * gpu);
+const char *                   apir_device_get_description(struct virtgpu * gpu);
+uint32_t                       apir_device_get_type(struct virtgpu * gpu);
+void                           apir_device_get_memory(struct virtgpu * gpu, size_t * free, size_t * total);
+bool                           apir_device_supports_op(struct virtgpu * gpu, const ggml_tensor * op);
+apir_buffer_type_host_handle_t apir_device_get_buffer_type(struct virtgpu * gpu);
+void                           apir_device_get_props(struct virtgpu * gpu,
+                                                     bool *           async,
+                                                     bool *           host_buffer,
+                                                     bool *           buffer_from_host_ptr,
+                                                     bool *           events);
+apir_buffer_context_t          apir_device_buffer_from_ptr(struct virtgpu * gpu, size_t size, size_t max_tensor_size);
+
+/* buffer-type */
+const char *          apir_buffer_type_get_name(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
+size_t                apir_buffer_type_get_alignment(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
+size_t                apir_buffer_type_get_max_size(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
+bool                  apir_buffer_type_is_host(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
+apir_buffer_context_t apir_buffer_type_alloc_buffer(struct virtgpu *           gpu,
+                                                    ggml_backend_buffer_type_t buffer_buft,
+                                                    size_t                     size);
+size_t apir_buffer_type_get_alloc_size(struct virtgpu * gpu, ggml_backend_buffer_type_t buft, const ggml_tensor * op);
+
+/* buffer */
+void * apir_buffer_get_base(struct virtgpu * gpu, apir_buffer_context_t * buffer_context);
+void   apir_buffer_set_tensor(struct virtgpu *        gpu,
+                              apir_buffer_context_t * buffer_context,
+                              ggml_tensor *           tensor,
+                              const void *            data,
+                              size_t                  offset,
+                              size_t                  size);
+void   apir_buffer_get_tensor(struct virtgpu *        gpu,
+                              apir_buffer_context_t * buffer_context,
+                              const ggml_tensor *     tensor,
+                              void *                  data,
+                              size_t                  offset,
+                              size_t                  size);
+bool   apir_buffer_cpy_tensor(struct virtgpu *        gpu,
+                              apir_buffer_context_t * buffer_context,
+                              const ggml_tensor *     src,
+                              const ggml_tensor *     dst);
+void   apir_buffer_clear(struct virtgpu * gpu, apir_buffer_context_t * buffer_context, uint8_t value);
+void   apir_buffer_free_buffer(struct virtgpu * gpu, apir_buffer_context_t * buffer_context);
+
+/* backend */
+ggml_status apir_backend_graph_compute(struct virtgpu * gpu, ggml_cgraph * cgraph);
diff --git a/ggml/src/ggml-virtgpu/virtgpu-shm.cpp b/ggml/src/ggml-virtgpu/virtgpu-shm.cpp
new file mode 100644
index 0000000000..4def405a62
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/virtgpu-shm.cpp
@@ -0,0 +1,99 @@
+#include "virtgpu-shm.h"
+
+#include "virtgpu.h"
+
+#include <assert.h>
+
+static uint32_t virtgpu_ioctl_resource_create_blob(virtgpu *  gpu,
+                                                   uint32_t   blob_mem,
+                                                   uint32_t   blob_flags,
+                                                   size_t     blob_size,
+                                                   uint64_t   blob_id,
+                                                   uint32_t * res_id) {
+#ifdef SIMULATE_BO_SIZE_FIX
+    blob_size = align64(blob_size, 4096);
+#endif
+
+    drm_virtgpu_resource_create_blob args = {
+        .blob_mem   = blob_mem,
+        .blob_flags = blob_flags,
+        .bo_handle  = 0,
+        .res_handle = 0,
+        .size       = blob_size,
+        .pad        = 0,
+        .cmd_size   = 0,
+        .cmd        = 0,
+        .blob_id    = blob_id,
+    };
+
+    if (virtgpu_ioctl(gpu, DRM_IOCTL_VIRTGPU_RESOURCE_CREATE_BLOB, &args)) {
+        return 0;
+    }
+
+    *res_id = args.res_handle;
+    return args.bo_handle;
+}
+
+static void virtgpu_ioctl_gem_close(virtgpu * gpu, uint32_t gem_handle) {
+    drm_gem_close args = {
+        .handle = gem_handle,
+        .pad    = 0,
+    };
+
+    const int ret = virtgpu_ioctl(gpu, DRM_IOCTL_GEM_CLOSE, &args);
+    assert(!ret);
+#ifdef NDEBUG
+    UNUSED(ret);
+#endif
+}
+
+static void * virtgpu_ioctl_map(virtgpu * gpu, uint32_t gem_handle, size_t size) {
+    drm_virtgpu_map args = {
+        .offset = 0,
+        .handle = gem_handle,
+        .pad    = 0,
+    };
+
+    if (virtgpu_ioctl(gpu, DRM_IOCTL_VIRTGPU_MAP, &args)) {
+        return NULL;
+    }
+
+    void * ptr = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED, gpu->fd, args.offset);
+    if (ptr == MAP_FAILED) {
+        return NULL;
+    }
+
+    return ptr;
+}
+
+void virtgpu_shmem_destroy(virtgpu * gpu, virtgpu_shmem * shmem) {
+    munmap(shmem->mmap_ptr, shmem->mmap_size);
+    virtgpu_ioctl_gem_close(gpu, shmem->gem_handle);
+}
+
+int virtgpu_shmem_create(virtgpu * gpu, size_t size, virtgpu_shmem * shmem) {
+    size = align64(size, 16384);
+
+    uint32_t res_id;
+    uint32_t gem_handle = virtgpu_ioctl_resource_create_blob(gpu, VIRTGPU_BLOB_MEM_HOST3D,
+                                                             VIRTGPU_BLOB_FLAG_USE_MAPPABLE, size, 0, &res_id);
+
+    if (!gem_handle) {
+        return 1;
+    }
+
+    void * ptr = virtgpu_ioctl_map(gpu, gem_handle, size);
+    if (!ptr) {
+        virtgpu_ioctl_gem_close(gpu, gem_handle);
+        GGML_LOG_ERROR("virtgpu_ioctl_map FAILED\n");
+        exit(1);
+        return 1;
+    }
+
+    shmem->res_id     = res_id;
+    shmem->mmap_size  = size;
+    shmem->mmap_ptr   = ptr;
+    shmem->gem_handle = gem_handle;
+
+    return 0;
+}
diff --git a/ggml/src/ggml-virtgpu/virtgpu-shm.h b/ggml/src/ggml-virtgpu/virtgpu-shm.h
new file mode 100644
index 0000000000..606860a094
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/virtgpu-shm.h
@@ -0,0 +1,23 @@
+#pragma once
+
+#include "virtgpu-utils.h"
+
+#include <sys/mman.h>
+
+#include <atomic>
+#include <cassert>
+#include <cstddef>
+#include <cstdint>
+
+struct virtgpu;
+
+struct virtgpu_shmem {
+    uint32_t res_id;
+    size_t   mmap_size;
+    void *   mmap_ptr;
+
+    uint32_t gem_handle;
+};
+
+int  virtgpu_shmem_create(virtgpu * gpu, size_t size, virtgpu_shmem * shmem);
+void virtgpu_shmem_destroy(virtgpu * gpu, virtgpu_shmem * shmem);
diff --git a/ggml/src/ggml-virtgpu/virtgpu-utils.cpp b/ggml/src/ggml-virtgpu/virtgpu-utils.cpp
new file mode 100644
index 0000000000..8a2805e990
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/virtgpu-utils.cpp
@@ -0,0 +1,179 @@
+#include "virtgpu-utils.h"
+
+#include <malloc.h>
+#include <stdlib.h>
+
+#include <cstring>
+
+#define NODE_ALLOC_ALIGN 64
+#define NODE_PTR_MASK    (~((uintptr_t) NODE_ALLOC_ALIGN - 1))
+#define NODE_LEVEL_MASK  ((uintptr_t) NODE_ALLOC_ALIGN - 1)
+#define NULL_NODE        0
+
+#define os_malloc_aligned(_size, _align) _aligned_malloc(_size, _align)
+#define os_free_aligned(_ptr)            free(_ptr)
+#define p_atomic_cmpxchg(v, old, _new)   __sync_val_compare_and_swap((v), (old), (_new))
+
+static inline uint64_t util_logbase2_64(uint64_t n) {
+#if defined(HAVE___BUILTIN_CLZLL)
+    return ((sizeof(uint64_t) * 8 - 1) - __builtin_clzll(n | 1));
+#else
+    uint64_t pos = 0ull;
+    if (n >= 1ull << 32) {
+        n >>= 32;
+        pos += 32;
+    }
+    if (n >= 1ull << 16) {
+        n >>= 16;
+        pos += 16;
+    }
+    if (n >= 1ull << 8) {
+        n >>= 8;
+        pos += 8;
+    }
+    if (n >= 1ull << 4) {
+        n >>= 4;
+        pos += 4;
+    }
+    if (n >= 1ull << 2) {
+        n >>= 2;
+        pos += 2;
+    }
+    if (n >= 1ull << 1) {
+        pos += 1;
+    }
+    return pos;
+#endif
+}
+
+void util_sparse_array_init(util_sparse_array * arr, size_t elem_size, size_t node_size) {
+    memset(arr, 0, sizeof(*arr));
+    arr->elem_size      = elem_size;
+    arr->node_size_log2 = util_logbase2_64(node_size);
+    assert(node_size >= 2 && node_size == (1ull << arr->node_size_log2));
+}
+
+static inline void * os_malloc_aligned(size_t size, size_t alignment) {
+    void * ptr;
+    alignment = (alignment + sizeof(void *) - 1) & ~(sizeof(void *) - 1);
+    if (posix_memalign(&ptr, alignment, size) != 0) {
+        return NULL;
+    }
+    return ptr;
+}
+
+static inline void * _util_sparse_array_node_data(uintptr_t handle) {
+    return (void *) (handle & NODE_PTR_MASK);
+}
+
+static inline unsigned _util_sparse_array_node_level(uintptr_t handle) {
+    return handle & NODE_LEVEL_MASK;
+}
+
+static inline void _util_sparse_array_node_finish(util_sparse_array * arr, uintptr_t node) {
+    if (_util_sparse_array_node_level(node) > 0) {
+        uintptr_t * children  = (uintptr_t *) _util_sparse_array_node_data(node);
+        size_t      node_size = 1ull << arr->node_size_log2;
+        for (size_t i = 0; i < node_size; i++) {
+            if (children[i]) {
+                _util_sparse_array_node_finish(arr, children[i]);
+            }
+        }
+    }
+
+    os_free_aligned(_util_sparse_array_node_data(node));
+}
+
+static inline uintptr_t _util_sparse_array_node(void * data, unsigned level) {
+    assert(data != NULL);
+    assert(((uintptr_t) data & NODE_LEVEL_MASK) == 0);
+    assert((level & NODE_PTR_MASK) == 0);
+    return (uintptr_t) data | level;
+}
+
+inline uintptr_t _util_sparse_array_node_alloc(util_sparse_array * arr, unsigned level) {
+    size_t size;
+    if (level == 0) {
+        size = arr->elem_size << arr->node_size_log2;
+    } else {
+        size = sizeof(uintptr_t) << arr->node_size_log2;
+    }
+
+    void * data = os_malloc_aligned(size, NODE_ALLOC_ALIGN);
+    memset(data, 0, size);
+
+    return _util_sparse_array_node(data, level);
+}
+
+static inline uintptr_t _util_sparse_array_set_or_free_node(uintptr_t * node_ptr, uintptr_t cmp_node, uintptr_t node) {
+    uintptr_t prev_node = p_atomic_cmpxchg(node_ptr, cmp_node, node);
+
+    if (prev_node != cmp_node) {
+        /* We lost the race.  Free this one and return the one that was already
+       * allocated.
+       */
+        os_free_aligned(_util_sparse_array_node_data(node));
+        return prev_node;
+    } else {
+        return node;
+    }
+}
+
+void * util_sparse_array_get(util_sparse_array * arr, uint64_t idx) {
+    const unsigned node_size_log2 = arr->node_size_log2;
+    uintptr_t      root           = p_atomic_read(&arr->root);
+    if (unlikely(!root)) {
+        unsigned root_level = 0;
+        uint64_t idx_iter   = idx >> node_size_log2;
+        while (idx_iter) {
+            idx_iter >>= node_size_log2;
+            root_level++;
+        }
+        uintptr_t new_root = _util_sparse_array_node_alloc(arr, root_level);
+        root               = _util_sparse_array_set_or_free_node(&arr->root, NULL_NODE, new_root);
+    }
+
+    while (1) {
+        unsigned root_level = _util_sparse_array_node_level(root);
+        uint64_t root_idx   = idx >> (root_level * node_size_log2);
+        if (likely(root_idx < (1ull << node_size_log2))) {
+            break;
+        }
+
+        /* In this case, we have a root but its level is low enough that the
+       * requested index is out-of-bounds.
+       */
+        uintptr_t new_root = _util_sparse_array_node_alloc(arr, root_level + 1);
+
+        uintptr_t * new_root_children = (uintptr_t *) _util_sparse_array_node_data(new_root);
+        new_root_children[0]          = root;
+
+        /* We only add one at a time instead of the whole tree because it's
+       * easier to ensure correctness of both the tree building and the
+       * clean-up path.  Because we're only adding one node we never have to
+       * worry about trying to free multiple things without freeing the old
+       * things.
+       */
+        root = _util_sparse_array_set_or_free_node(&arr->root, root, new_root);
+    }
+
+    void *   node_data  = _util_sparse_array_node_data(root);
+    unsigned node_level = _util_sparse_array_node_level(root);
+    while (node_level > 0) {
+        uint64_t child_idx = (idx >> (node_level * node_size_log2)) & ((1ull << node_size_log2) - 1);
+
+        uintptr_t * children = (uintptr_t *) node_data;
+        uintptr_t   child    = p_atomic_read(&children[child_idx]);
+
+        if (unlikely(!child)) {
+            child = _util_sparse_array_node_alloc(arr, node_level - 1);
+            child = _util_sparse_array_set_or_free_node(&children[child_idx], NULL_NODE, child);
+        }
+
+        node_data  = _util_sparse_array_node_data(child);
+        node_level = _util_sparse_array_node_level(child);
+    }
+
+    uint64_t elem_idx = idx & ((1ull << node_size_log2) - 1);
+    return (void *) ((char *) node_data + (elem_idx * arr->elem_size));
+}
diff --git a/ggml/src/ggml-virtgpu/virtgpu-utils.h b/ggml/src/ggml-virtgpu/virtgpu-utils.h
new file mode 100644
index 0000000000..a0036b4e2b
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/virtgpu-utils.h
@@ -0,0 +1,86 @@
+#pragma once
+
+#include <atomic>
+#include <cassert>
+#include <cerrno>
+#include <cstdarg>
+#include <cstddef>
+#include <cstdint>
+#include <cstdio>
+#include <cstdlib>
+#include <ctime>
+
+#define unlikely(x) __builtin_expect(!!(x), 0)
+#define likely(x)   __builtin_expect(!!(x), 1)
+
+#ifndef UNUSED
+#    define UNUSED(x) (void) (x)
+#endif
+
+/** Checks is a value is a power of two. Does not handle zero. */
+#define IS_POT(v) (((v) & ((v) - 1)) == 0)
+
+/** Checks is a value is a power of two. Zero handled. */
+#define IS_POT_NONZERO(v) ((v) != 0 && IS_POT(v))
+
+/** Align a value to a power of two */
+#define ALIGN_POT(x, pot_align) (((x) + (pot_align) - 1) & ~((pot_align) - 1))
+
+#define p_atomic_read(_v) __atomic_load_n((_v), __ATOMIC_ACQUIRE)
+
+static inline bool util_is_power_of_two_nonzero64(uint64_t v) {
+    return IS_POT_NONZERO(v);
+}
+
+static inline uint64_t align64(uint64_t value, uint64_t alignment) {
+    assert(util_is_power_of_two_nonzero64(alignment));
+    return ALIGN_POT(value, alignment);
+}
+
+struct list_head {
+    list_head * prev;
+    list_head * next;
+};
+
+struct util_sparse_array {
+    size_t   elem_size;
+    unsigned node_size_log2;
+
+    uintptr_t root;
+};
+
+void * util_sparse_array_get(util_sparse_array * arr, uint64_t idx);
+void   util_sparse_array_init(util_sparse_array * arr, size_t elem_size, size_t node_size);
+
+inline void os_time_sleep(int64_t usecs) {
+    timespec time;
+    time.tv_sec  = usecs / 1000000;
+    time.tv_nsec = (usecs % 1000000) * 1000;
+    while (clock_nanosleep(CLOCK_MONOTONIC, 0, &time, &time) == EINTR)
+        ;
+}
+
+struct timer_data {
+    long long start;
+    long long total;
+    long long count;
+};
+
+static inline void start_timer(timer_data * timer) {
+    timespec ts;
+    clock_gettime(CLOCK_MONOTONIC, &ts);
+    timer->start = (long long) ts.tv_sec * 1000000000LL + ts.tv_nsec;
+}
+
+// returns the duration in ns
+static inline long long stop_timer(timer_data * timer) {
+    timespec ts;
+    clock_gettime(CLOCK_MONOTONIC, &ts);
+    long long timer_end = (long long) ts.tv_sec * 1000000000LL + ts.tv_nsec;
+
+    long long duration = (timer_end - timer->start);
+    timer->total += duration;
+    timer->count += 1;
+
+    return duration;
+}
diff --git a/ggml/src/ggml-virtgpu/virtgpu.cpp b/ggml/src/ggml-virtgpu/virtgpu.cpp
new file mode 100644
index 0000000000..005c8e21db
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/virtgpu.cpp
@@ -0,0 +1,498 @@
+#include "virtgpu.h"
+
+#include <stdio.h>
+#include <unistd.h>
+
+#include <cassert>
+#include <cerrno>
+#include <cstdlib>
+
+static virt_gpu_result_t virtgpu_open_device(virtgpu * gpu, const drmDevicePtr dev);
+static virt_gpu_result_t virtgpu_open(virtgpu * gpu);
+
+static virt_gpu_result_t virtgpu_init_capset(virtgpu * gpu);
+static virt_gpu_result_t virtgpu_init_context(virtgpu * gpu);
+
+static int      virtgpu_ioctl_context_init(virtgpu * gpu, virgl_renderer_capset capset_id);
+static int      virtgpu_ioctl_get_caps(virtgpu *             gpu,
+                                       virgl_renderer_capset id,
+                                       uint32_t              version,
+                                       void *                capset,
+                                       size_t                capset_size);
+static uint64_t virtgpu_ioctl_getparam(virtgpu * gpu, uint64_t param);
+static void     virtgpu_init_renderer_info(virtgpu * gpu);
+
+static void log_call_duration(long long call_duration_ns, const char * name);
+
+const uint64_t APIR_HANDSHAKE_MAX_WAIT_MS   = 2 * 1000;   // 2s
+const uint64_t APIR_LOADLIBRARY_MAX_WAIT_MS = 60 * 1000;  // 60s
+
+static int virtgpu_handshake(virtgpu * gpu) {
+    apir_encoder * encoder;
+    apir_decoder * decoder;
+
+    encoder = remote_call_prepare(gpu, APIR_COMMAND_TYPE_HANDSHAKE, 0);
+    if (!encoder) {
+        GGML_ABORT("%s: failed to prepare the remote call encoder", __func__);
+        return 1;
+    }
+
+    /* write handshake props */
+
+    uint32_t guest_major = APIR_PROTOCOL_MAJOR;
+    uint32_t guest_minor = APIR_PROTOCOL_MINOR;
+    apir_encode_uint32_t(encoder, &guest_major);
+    apir_encode_uint32_t(encoder, &guest_minor);
+
+    /* *** */
+
+    uint32_t  ret_magic;
+    long long call_duration_ns;
+    ret_magic = remote_call(gpu, encoder, &decoder, APIR_HANDSHAKE_MAX_WAIT_MS, &call_duration_ns);
+    log_call_duration(call_duration_ns, "API Remoting handshake");
+
+    if (!decoder) {
+        GGML_ABORT(
+            "%s: failed to initiate the communication with the virglrenderer library. "
+            "Most likely, the wrong virglrenderer library was loaded in the hypervisor.",
+            __func__);
+        return 1;
+    }
+
+    /* read handshake return values */
+
+    uint32_t host_major;
+    uint32_t host_minor;
+
+    if (ret_magic != APIR_HANDSHAKE_MAGIC) {
+        GGML_ABORT("%s: handshake with the virglrenderer failed (code=%d | %s)", __func__, ret_magic,
+                   apir_backend_initialize_error(ret_magic));
+    } else {
+        apir_decode_uint32_t(decoder, &host_major);
+        apir_decode_uint32_t(decoder, &host_minor);
+    }
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    if (ret_magic != APIR_HANDSHAKE_MAGIC) {
+        return 1;
+    }
+
+    GGML_LOG_INFO("%s: Guest is running with %u.%u\n", __func__, guest_major, guest_minor);
+    GGML_LOG_INFO("%s: Host is running with %u.%u\n", __func__, host_major, host_minor);
+
+    if (guest_major != host_major) {
+        GGML_LOG_ERROR("Host major (%d) and guest major (%d) version differ\n", host_major, guest_major);
+    } else if (guest_minor != host_minor) {
+        GGML_LOG_WARN("Host minor (%d) and guest minor (%d) version differ\n", host_minor, guest_minor);
+    }
+
+    return 0;
+}
+
+static ApirLoadLibraryReturnCode virtgpu_load_library(virtgpu * gpu) {
+    apir_encoder *            encoder;
+    apir_decoder *            decoder;
+    ApirLoadLibraryReturnCode ret;
+
+    encoder = remote_call_prepare(gpu, APIR_COMMAND_TYPE_LOADLIBRARY, 0);
+    if (!encoder) {
+        GGML_ABORT("%s: hypercall error: failed to prepare the remote call encoder", __func__);
+        return APIR_LOAD_LIBRARY_HYPERCALL_INITIALIZATION_ERROR;
+    }
+
+    long long call_duration_ns;
+
+    ret = (ApirLoadLibraryReturnCode) remote_call(gpu, encoder, &decoder, APIR_LOADLIBRARY_MAX_WAIT_MS,
+                                                  &call_duration_ns);
+    log_call_duration(call_duration_ns, "API Remoting LoadLibrary");
+
+    if (!decoder) {
+        GGML_ABORT("%s: hypercall error: failed to kick the API remoting hypercall.\n", __func__);
+        return APIR_LOAD_LIBRARY_HYPERCALL_INITIALIZATION_ERROR;
+    }
+
+    remote_call_finish(gpu, encoder, decoder);
+
+    if (ret == APIR_LOAD_LIBRARY_SUCCESS) {
+        GGML_LOG_INFO("%s: The API Remoting backend was successfully loaded and initialized\n", __func__);
+
+        return ret;
+    }
+
+    // something wrong happened, find out what.
+
+    if (ret < APIR_LOAD_LIBRARY_INIT_BASE_INDEX) {
+        GGML_ABORT("%s: virglrenderer could not load the API Remoting backend library: %s (code %d)", __func__,
+                   apir_load_library_error(ret), ret);
+        return ret;
+    }
+
+    GGML_LOG_INFO("%s: virglrenderer successfully loaded the API Remoting backend library", __func__);
+
+    ApirLoadLibraryReturnCode apir_ret = (ApirLoadLibraryReturnCode) (ret - APIR_LOAD_LIBRARY_INIT_BASE_INDEX);
+
+    if (apir_ret < APIR_LOAD_LIBRARY_INIT_BASE_INDEX) {
+        GGML_ABORT("%s: the API Remoting backend library couldn't load the backend library: apir code=%d | %s)",
+                   __func__, apir_ret, apir_load_library_error(apir_ret));
+    } else {
+        uint32_t lib_ret = apir_ret - APIR_LOAD_LIBRARY_INIT_BASE_INDEX;
+        GGML_ABORT("%s: the API Remoting backend library initialize its backend library: apir code=%d)", __func__,
+                   lib_ret);
+    }
+    return ret;
+}
+
+virtgpu * create_virtgpu() {
+    virtgpu * gpu = new virtgpu();
+
+    gpu->use_apir_capset = getenv("GGML_REMOTING_USE_APIR_CAPSET") != nullptr;
+    util_sparse_array_init(&gpu->shmem_array, sizeof(virtgpu_shmem), 1024);
+
+    if (virtgpu_open(gpu) != APIR_SUCCESS) {
+        GGML_ABORT("%s: failed to open the virtgpu device", __func__);
+        return NULL;
+    }
+
+    if (virtgpu_init_capset(gpu) != APIR_SUCCESS) {
+        GGML_ABORT("%s: failed to initialize the GPU capset", __func__);
+        return NULL;
+    }
+
+    if (virtgpu_init_context(gpu) != APIR_SUCCESS) {
+        GGML_ABORT("%s: failed to initialize the GPU context", __func__);
+        return NULL;
+    }
+
+    if (virtgpu_shmem_create(gpu, SHMEM_REPLY_SIZE, &gpu->reply_shmem)) {
+        GGML_ABORT("%s: failed to create the shared reply memory pages", __func__);
+        return NULL;
+    }
+
+    if (virtgpu_shmem_create(gpu, SHMEM_DATA_SIZE, &gpu->data_shmem)) {
+        GGML_ABORT("%s: failed to create the shared data memory pages", __func__);
+        return NULL;
+    }
+
+    if (virtgpu_handshake(gpu)) {
+        GGML_ABORT("%s: failed to handshake with the virglrenderer library", __func__);
+        return NULL;
+    }
+
+    if (virtgpu_load_library(gpu) != APIR_LOAD_LIBRARY_SUCCESS) {
+        GGML_ABORT("%s: failed to load the backend library", __func__);
+        return NULL;
+    }
+
+    return gpu;
+}
+
+static virt_gpu_result_t virtgpu_open(virtgpu * gpu) {
+    drmDevicePtr devs[8];
+    int          count = drmGetDevices2(0, devs, ARRAY_SIZE(devs));
+    if (count < 0) {
+        GGML_LOG_ERROR("%s: failed to enumerate DRM devices\n", __func__);
+        return APIR_ERROR_INITIALIZATION_FAILED;
+    }
+
+    virt_gpu_result_t result = APIR_ERROR_INITIALIZATION_FAILED;
+    for (int i = 0; i < count; i++) {
+        result = virtgpu_open_device(gpu, devs[i]);
+        if (result == APIR_SUCCESS) {
+            break;
+        }
+    }
+
+    drmFreeDevices(devs, count);
+
+    return result;
+}
+
+static virt_gpu_result_t virtgpu_open_device(virtgpu * gpu, const drmDevicePtr dev) {
+    const char * node_path = dev->nodes[DRM_NODE_RENDER];
+
+    int fd = open(node_path, O_RDWR | O_CLOEXEC);
+    if (fd < 0) {
+        GGML_ABORT("failed to open %s", node_path);
+        return APIR_ERROR_INITIALIZATION_FAILED;
+    }
+
+    drmVersionPtr version = drmGetVersion(fd);
+    if (!version || strcmp(version->name, "virtio_gpu") || version->version_major != 0) {
+        if (version) {
+            GGML_ABORT("unknown DRM driver %s version %d", version->name, version->version_major);
+        } else {
+            GGML_ABORT("failed to get DRM driver version");
+        }
+
+        if (version) {
+            drmFreeVersion(version);
+        }
+        close(fd);
+        return APIR_ERROR_INITIALIZATION_FAILED;
+    }
+
+    gpu->fd = fd;
+
+    drmFreeVersion(version);
+
+    GGML_LOG_INFO("using DRM device %s\n", node_path);
+
+    return APIR_SUCCESS;
+}
+
+static virt_gpu_result_t virtgpu_init_context(virtgpu * gpu) {
+    assert(!gpu->capset.version);
+    const int ret = virtgpu_ioctl_context_init(gpu, gpu->capset.id);
+    if (ret) {
+        GGML_LOG_INFO("failed to initialize context: %s\n", strerror(errno));
+        return APIR_ERROR_INITIALIZATION_FAILED;
+    }
+
+    return APIR_SUCCESS;
+}
+
+static virt_gpu_result_t virtgpu_init_capset(virtgpu * gpu) {
+    if (gpu->use_apir_capset) {
+        GGML_LOG_INFO("Using the APIR capset\n");
+        gpu->capset.id = VIRTGPU_DRM_CAPSET_APIR;
+    } else {
+        GGML_LOG_INFO("Using the Venus capset\n");
+        gpu->capset.id = VIRTGPU_DRM_CAPSET_VENUS;
+    }
+    gpu->capset.version = 0;
+
+    int ret =
+        virtgpu_ioctl_get_caps(gpu, gpu->capset.id, gpu->capset.version, &gpu->capset.data, sizeof(gpu->capset.data));
+
+    if (ret) {
+        GGML_LOG_INFO("failed to get APIR v%d capset: %s\n", gpu->capset.version, strerror(errno));
+        return APIR_ERROR_INITIALIZATION_FAILED;
+    }
+
+    assert(gpu->capset.data.supports_blob_resources);
+
+    return APIR_SUCCESS;
+}
+
+static int virtgpu_ioctl_context_init(virtgpu * gpu, virgl_renderer_capset capset_id) {
+    drm_virtgpu_context_set_param ctx_set_params[3] = {
+        {
+         .param = VIRTGPU_CONTEXT_PARAM_CAPSET_ID,
+         .value = capset_id,
+         },
+        {
+         .param = VIRTGPU_CONTEXT_PARAM_NUM_RINGS,
+         .value = 1,
+         },
+        {
+         .param = VIRTGPU_CONTEXT_PARAM_POLL_RINGS_MASK,
+         .value = 0, /* don't generate drm_events on fence signaling */
+        },
+    };
+
+    drm_virtgpu_context_init args = {
+        .num_params     = ARRAY_SIZE(ctx_set_params),
+        .pad            = 0,
+        .ctx_set_params = (uintptr_t) &ctx_set_params,
+    };
+
+    return virtgpu_ioctl(gpu, DRM_IOCTL_VIRTGPU_CONTEXT_INIT, &args);
+}
+
+static int virtgpu_ioctl_get_caps(virtgpu *             gpu,
+                                  virgl_renderer_capset id,
+                                  uint32_t              version,
+                                  void *                capset,
+                                  size_t                capset_size) {
+    drm_virtgpu_get_caps args = {
+        .cap_set_id  = id,
+        .cap_set_ver = version,
+        .addr        = (uintptr_t) capset,
+        .size        = (__u32) capset_size,
+        .pad         = 0,
+    };
+
+    return virtgpu_ioctl(gpu, DRM_IOCTL_VIRTGPU_GET_CAPS, &args);
+}
+
+static uint64_t virtgpu_ioctl_getparam(virtgpu * gpu, uint64_t param) {
+    /* val must be zeroed because kernel only writes the lower 32 bits */
+    uint64_t             val  = 0;
+    drm_virtgpu_getparam args = {
+        .param = param,
+        .value = (uintptr_t) &val,
+    };
+
+    const int ret = virtgpu_ioctl(gpu, DRM_IOCTL_VIRTGPU_GETPARAM, &args);
+    return ret ? 0 : val;
+}
+
+apir_encoder * remote_call_prepare(virtgpu * gpu, ApirCommandType apir_cmd_type, int32_t cmd_flags) {
+    /*
+     * Prepare the command encoder and its buffer
+     */
+
+    static char encoder_buffer[4096];
+
+    static apir_encoder enc;
+    enc = {
+        .cur   = encoder_buffer,
+        .start = encoder_buffer,
+        .end   = encoder_buffer + sizeof(encoder_buffer),
+        .fatal = false,
+    };
+
+    /*
+     * Fill the command encoder with the common args:
+     * - cmd_type (int32_t)
+     * - cmd_flags (int32_t)
+     * - reply res id (uint32_t)
+   */
+
+    int32_t cmd_type = apir_cmd_type;
+
+    // for testing during the hypervisor transition
+    if (!gpu->use_apir_capset) {
+        cmd_type += VENUS_COMMAND_TYPE_LENGTH;
+    }
+    apir_encode_int32_t(&enc, &cmd_type);
+    apir_encode_int32_t(&enc, &cmd_flags);
+
+    uint32_t reply_res_id = gpu->reply_shmem.res_id;
+    apir_encode_uint32_t(&enc, &reply_res_id);
+
+    return &enc;
+}
+
+void remote_call_finish(virtgpu * gpu, apir_encoder * enc, apir_decoder * dec) {
+    UNUSED(gpu);
+
+    if (!enc) {
+        GGML_LOG_ERROR("Invalid (null) encoder\n");
+    }
+
+    if (!dec) {
+        GGML_LOG_ERROR("Invalid (null) decoder\n");
+    }
+
+    if (apir_encoder_get_fatal(enc)) {
+        GGML_LOG_ERROR("Failed to encode the output parameters.\n");
+    }
+
+    if (apir_decoder_get_fatal(dec)) {
+        GGML_LOG_ERROR("Failed to decode the input parameters.\n");
+    }
+}
+
+uint32_t remote_call(virtgpu *       gpu,
+                     apir_encoder *  encoder,
+                     apir_decoder ** decoder,
+                     float           max_wait_ms,
+                     long long *     call_duration_ns) {
+    /*
+     * Prepare the reply notification pointer
+     */
+
+    volatile std::atomic_uint * atomic_reply_notif = (volatile std::atomic_uint *) gpu->reply_shmem.mmap_ptr;
+    *atomic_reply_notif                            = 0;
+
+    /*
+     * Trigger the execbuf ioctl
+     */
+
+    drm_virtgpu_execbuffer args = {
+        .flags   = VIRTGPU_EXECBUF_RING_IDX,
+        .size    = (uint32_t) (encoder->cur - encoder->start),
+        .command = (uintptr_t) encoder->start,
+
+        .bo_handles     = 0,
+        .num_bo_handles = 0,
+
+        .fence_fd         = 0,
+        .ring_idx         = 0,
+        .syncobj_stride   = 0,
+        .num_in_syncobjs  = 0,
+        .num_out_syncobjs = 0,
+        .in_syncobjs      = 0,
+        .out_syncobjs     = 0,
+    };
+
+    *decoder = NULL;
+
+    int ret = drmIoctl(gpu->fd, DRM_IOCTL_VIRTGPU_EXECBUFFER, &args);
+
+    if (ret != 0) {
+        GGML_ABORT("%s: the virtgpu EXECBUFFER ioctl failed (%d)", __func__, ret);
+    }
+
+    /*
+     * Wait for the response notification
+     */
+    timer_data wait_host_reply_timer = { 0, 0, 0 };
+
+    start_timer(&wait_host_reply_timer);
+
+    timespec ts_start, ts_end;
+    clock_gettime(CLOCK_MONOTONIC, &ts_start);
+    long long start_time = (long long) ts_start.tv_sec * 1000000000LL + ts_start.tv_nsec;
+
+    bool     timedout    = false;
+    uint32_t notif_value = 0;
+    while (true) {
+        notif_value = std::atomic_load_explicit(atomic_reply_notif, std::memory_order_acquire);
+
+        if (notif_value != 0) {
+            break;
+        }
+
+        int64_t base_sleep_us = 15;
+
+        os_time_sleep(base_sleep_us);
+
+        if (max_wait_ms) {
+            clock_gettime(CLOCK_MONOTONIC, &ts_end);
+            long long end_time    = (long long) ts_end.tv_sec * 1000000000LL + ts_end.tv_nsec;
+            float     duration_ms = (end_time - start_time) / 1000000;
+
+            if (duration_ms > max_wait_ms) {
+                timedout = true;
+                break;
+            }
+        }
+    }
+
+    if (call_duration_ns) {
+        *call_duration_ns = stop_timer(&wait_host_reply_timer);
+    }
+
+    if (max_wait_ms && timedout) {
+        GGML_LOG_ERROR("timed out waiting for the host answer...\n");
+        return APIR_FORWARD_TIMEOUT;
+    }
+
+    /*
+     * Prepare the decoder
+     */
+    static apir_decoder response_dec;
+    response_dec.cur = (char *) gpu->reply_shmem.mmap_ptr + sizeof(*atomic_reply_notif);
+    response_dec.end = (char *) gpu->reply_shmem.mmap_ptr + gpu->reply_shmem.mmap_size;
+    *decoder         = &response_dec;
+
+    // extract the actual return value from the notif flag
+    uint32_t returned_value = notif_value - 1;
+    return returned_value;
+}
+
+static void log_call_duration(long long call_duration_ns, const char * name) {
+    double call_duration_ms = (double) call_duration_ns / 1e6;  // 1 millisecond = 1e6 nanoseconds
+    double call_duration_s  = (double) call_duration_ns / 1e9;  // 1 second = 1e9 nanoseconds
+
+    if (call_duration_s > 1) {
+        GGML_LOG_INFO("%s: waited %.2fs for the %s host reply...\n", __func__, call_duration_s, name);
+    } else if (call_duration_ms > 1) {
+        GGML_LOG_INFO("%s: waited %.2fms for the %s host reply...\n", __func__, call_duration_ms, name);
+    } else {
+        GGML_LOG_INFO("%s: waited %lldns for the %s host reply...\n", __func__, call_duration_ns, name);
+    }
+}
diff --git a/ggml/src/ggml-virtgpu/virtgpu.h b/ggml/src/ggml-virtgpu/virtgpu.h
new file mode 100644
index 0000000000..d4bb42e20b
--- /dev/null
+++ b/ggml/src/ggml-virtgpu/virtgpu.h
@@ -0,0 +1,92 @@
+#pragma once
+
+#include "virtgpu-utils.h"
+#include "virtgpu-shm.h"
+#include "virtgpu-apir.h"
+
+#include "backend/shared/api_remoting.h"
+#include "backend/shared/apir_cs.h"
+
+#include <fcntl.h>
+#include <stdbool.h>
+#include <stdio.h>
+#include <sys/stat.h>
+#include <sys/sysmacros.h>
+#include <threads.h>
+#include <xf86drm.h>
+
+#include <cstring>
+
+#define VIRGL_RENDERER_UNSTABLE_APIS 1
+#include "apir_hw.h"
+#include <drm/virtgpu_drm.h>
+#include "venus_hw.h"
+
+#ifndef VIRTGPU_DRM_CAPSET_APIR
+// Will be defined include/drm/virtgpu_drm.h when
+// https://gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/1590/diffs
+// is merged
+#define VIRTGPU_DRM_CAPSET_APIR 10
+#endif
+
+// Mesa/Virlgrenderer Venus internal. Only necessary during the
+// Venus->APIR transition in Virglrenderer
+#define VENUS_COMMAND_TYPE_LENGTH 331
+
+#ifndef VIRTGPU_DRM_CAPSET_VENUS // only available with Linux >= v6.16
+#define VIRTGPU_DRM_CAPSET_VENUS 4
+#endif
+
+typedef uint32_t virgl_renderer_capset;
+
+/* from src/virtio/vulkan/vn_renderer_virtgpu.c */
+#define VIRTGPU_PCI_VENDOR_ID       0x1af4
+#define VIRTGPU_PCI_DEVICE_ID       0x1050
+#define VIRTGPU_BLOB_MEM_GUEST_VRAM 0x0004
+#define VIRTGPU_PARAM_GUEST_VRAM    9
+
+#define SHMEM_DATA_SIZE  0x1830000  // 24MiB
+#define SHMEM_REPLY_SIZE 0x4000
+
+#define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
+
+enum virt_gpu_result_t {
+    APIR_SUCCESS                     = 0,
+    APIR_ERROR_INITIALIZATION_FAILED = -1,
+};
+
+#define PRINTFLIKE(f, a) __attribute__((format(__printf__, f, a)))
+
+struct virtgpu {
+    bool use_apir_capset;
+
+    int fd;
+
+    struct {
+        virgl_renderer_capset      id;
+        uint32_t                   version;
+        virgl_renderer_capset_apir data;
+    } capset;
+
+    util_sparse_array shmem_array;
+
+    /* APIR communication pages */
+    virtgpu_shmem reply_shmem;
+    virtgpu_shmem data_shmem;
+};
+
+static inline int virtgpu_ioctl(virtgpu * gpu, unsigned long request, void * args) {
+    return drmIoctl(gpu->fd, request, args);
+}
+
+virtgpu * create_virtgpu();
+
+apir_encoder * remote_call_prepare(virtgpu * gpu, ApirCommandType apir_cmd_type, int32_t cmd_flags);
+
+uint32_t remote_call(virtgpu *       gpu,
+                     apir_encoder *  enc,
+                     apir_decoder ** dec,
+                     float           max_wait_ms,
+                     long long *     call_duration_ns);
+
+void remote_call_finish(virtgpu * gpu, apir_encoder * enc, apir_decoder * dec);
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 0fabbcec31..a99375c088 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -991,6 +991,8 @@ struct vk_mat_vec_id_push_constants {
     uint32_t fusion_flags;
     uint32_t nei0;
     uint32_t ne11;
+    uint32_t expert_i1;
+    uint32_t nbi1;
 };
 
 struct vk_flash_attn_push_constants {
@@ -1516,6 +1518,15 @@ struct vk_quantize_q8_1_push_constants {
     uint32_t num_blocks;
 };
 
+struct vk_op_flash_attn_split_k_reduce_push_constants {
+    uint32_t D;
+    uint32_t ne1;
+    uint32_t ne2;
+    uint32_t ne3;
+    uint32_t k_num;
+    uint32_t sinks;
+};
+
 // Allow pre-recording command buffers
 struct vk_staging_memcpy {
     vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {}
@@ -1802,7 +1813,6 @@ struct ggml_backend_vk_context {
     bool prealloc_x_need_sync, prealloc_y_need_sync, prealloc_split_k_need_sync;
 
     vk_context_ref compute_ctx;
-    vk_context_ref transfer_ctx;
 
     std::vector<vk_context_ref> tensor_ctxs;
 
@@ -1812,7 +1822,6 @@ struct ggml_backend_vk_context {
     uint32_t pipeline_descriptor_set_requirements {};
 
     vk_command_pool compute_cmd_pool;
-    vk_command_pool transfer_cmd_pool;
 
     // number of additional consecutive nodes that are being fused with the
     // node currently being processed
@@ -3153,17 +3162,31 @@ static void ggml_vk_load_shaders(vk_device& device) {
         // For scalar, use 128 (arbitrary)
         // The same D_split value is used for both HSK and HSV, so just base it on the union of the LSBs.
         const uint32_t D = (hsk|hsv);
-        uint32_t wg_size = (path == FA_SCALAR || path == FA_COOPMAT1)
-                            ? scalar_flash_attention_workgroup_size
-                            : ((small_rows && (D % 32) == 0) ? 256 : 128);
         auto rows_cols = fa_rows_cols(path, hsk, hsv, clamp, type, small_rows, small_cache);
 
+        uint32_t wg_size;
+        switch (path) {
+        case FA_COOPMAT2:
+            wg_size = ((small_rows && (D % 32) == 0) ? 256 : 128);
+            break;
+        case FA_COOPMAT1:
+            wg_size = (rows_cols[1] / 16) * device->subgroup_size; // enough subgroups for Bc/MatBc
+            break;
+        default:
+            wg_size = scalar_flash_attention_workgroup_size;
+            break;
+        }
+
         // D_split can't be larger than a subgroup because we use subgroupShuffle to reduce it.
         // D_split can't be larger than the LSB of D divided by 4 due to vectorization in the shader.
         const uint32_t D_lsb = D ^ (D & (D-1));
         uint32_t D_split = std::min(std::min(device->subgroup_size, 8u), D_lsb / 4);
 
-        return {wg_size, rows_cols[0], rows_cols[1], hsk, hsv, clamp, D_split};
+        // Nvidia prefers shared memory use to load large tiles of K
+        // AMD prefers loading K directly from global memory
+        const uint32_t k_load_shmem = device->vendor_id == VK_VENDOR_ID_NVIDIA ? 1 : 0;
+
+        return {wg_size, rows_cols[0], rows_cols[1], hsk, hsv, clamp, D_split, device->subgroup_size, k_load_shmem};
     };
 
 #define CREATE_FA(TYPE, NAMELC, FAPATH, SUFFIX) \
@@ -3178,15 +3201,15 @@ static void ggml_vk_load_shaders(vk_device& device) {
             if (path == FAPATH) { \
                 if (aligned) { \
                     if (f32acc) { \
-                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0));     \
                     } else { \
-                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0));     \
                     } \
                 } else { \
                     if (f32acc) { \
-                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f32acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), 1,                                        true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f32acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), 1,                                        true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0));     \
                     } else { \
-                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f16acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), 1,                                        true, true, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f16acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 6, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), 1,                                        true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0));     \
                     } \
                 } \
             } \
@@ -3980,7 +4003,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_MXFP4],   "get_rows_mxfp4_f32",   get_rows_mxfp4_f32_len,   get_rows_mxfp4_f32_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 3, 5 * sizeof(uint32_t), {1, device->subgroup_size, 1}, {device->subgroup_size}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 3, sizeof(vk_op_flash_attn_split_k_reduce_push_constants), {1, device->subgroup_size, 1}, {device->subgroup_size}, 1, true);
 
     if (device->subgroup_clustered && device->subgroup_require_full_support) {
         ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1_x4, "quantize_q8_1_x4", quantize_q8_1_x4_subgroup_len, quantize_q8_1_x4_subgroup_data, "main", 2, sizeof(vk_quantize_q8_1_push_constants), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1, true, true);
@@ -5513,22 +5536,32 @@ static void ggml_vk_instance_init() {
 
             if ((new_props.properties.deviceType == vk::PhysicalDeviceType::eDiscreteGpu || new_props.properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu) && ggml_vk_device_is_supported(devices[i])) {
                 // Check if there are two physical devices corresponding to the same GPU
+                // This handles the case where the same GPU appears with different drivers (e.g., RADV + AMDVLK on Linux),
+                // see https://github.com/ggml-org/llama.cpp/pull/7582 for original deduplication.
+                // However, for MoltenVK on macOS, multiple GPUs on the same card may report the same UUID,
+                // see https://github.com/KhronosGroup/MoltenVK/issues/2683. Until this is fixed, we'll only deduplicate
+                // when drivers differ (same driver + same UUID = likely different GPUs)
                 auto old_device = std::find_if(
                     vk_instance.device_indices.begin(),
                     vk_instance.device_indices.end(),
-                    [&devices, &new_id](const size_t k){
+                    [&devices, &new_id, &new_driver](const size_t k){
                         vk::PhysicalDeviceProperties2 old_props;
+                        vk::PhysicalDeviceDriverProperties old_driver;
                         vk::PhysicalDeviceIDProperties old_id;
-                        old_props.pNext = &old_id;
+                        old_props.pNext = &old_driver;
+                        old_driver.pNext = &old_id;
                         devices[k].getProperties2(&old_props);
 
-                        bool equals = std::equal(std::begin(old_id.deviceUUID), std::end(old_id.deviceUUID), std::begin(new_id.deviceUUID));
-                        equals = equals || (
+                        bool same_uuid = std::equal(std::begin(old_id.deviceUUID), std::end(old_id.deviceUUID), std::begin(new_id.deviceUUID));
+                        same_uuid = same_uuid || (
                             old_id.deviceLUIDValid && new_id.deviceLUIDValid &&
                             std::equal(std::begin(old_id.deviceLUID), std::end(old_id.deviceLUID), std::begin(new_id.deviceLUID))
                         );
 
-                        return equals;
+                        // Only deduplicate if same UUID AND different drivers
+                        // (same driver + same UUID on MoltenVK = likely different GPUs on multi-GPU card)
+                        bool different_driver = (old_driver.driverID != new_driver.driverID);
+                        return same_uuid && different_driver;
                     }
                 );
                 if (old_device == vk_instance.device_indices.end()) {
@@ -5647,7 +5680,6 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
     ctx->almost_ready_fence = ctx->device->device.createFence({});
 
     ctx->compute_cmd_pool.init(ctx->device, &ctx->device->compute_queue);
-    ctx->transfer_cmd_pool.init(ctx->device, &ctx->device->transfer_queue);
 
     if (vk_perf_logger_enabled) {
         ctx->perf_logger = std::unique_ptr<vk_perf_logger>(new vk_perf_logger());
@@ -8083,8 +8115,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
 
     const uint64_t nei0 = ids->ne[0];
     const uint64_t nei1 = ids->ne[1];
-
-    GGML_ASSERT(nei1 == 1);
+    const uint32_t nbi1 = (uint32_t)(ids->nb[1] / sizeof(int));
 
     const uint64_t ne20 = dst->ne[0];
     const uint64_t ne21 = dst->ne[1];
@@ -8168,7 +8199,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
         if (quantize_y) {
             ggml_pipeline_request_descriptor_sets(ctx, to_q8_1, 1);
         }
-        ggml_pipeline_request_descriptor_sets(ctx, dmmv, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, dmmv, nei1);
     }
 
     vk_subbuffer d_D = ggml_vk_tensor_subbuffer(ctx, cgraph->nodes[node_idx + ctx->num_additional_fused_ops]);
@@ -8226,7 +8257,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
     uint32_t stride_batch_y = ne10*ne11;
 
     if (!ggml_vk_dim01_contiguous(src1) && !qy_needs_dequant) {
-        stride_batch_y = src1->nb[0] / ggml_type_size(src1->type);
+        stride_batch_y = src1->nb[2] / ggml_type_size(src1->type);
     }
 
     const uint32_t max_groups_x = ctx->device->properties.limits.maxComputeWorkGroupCount[0];
@@ -8262,23 +8293,25 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
         fusion_flags |= MAT_VEC_FUSION_FLAGS_SCALE1;
     }
 
-    // compute
-    const vk_mat_vec_id_push_constants pc = {
-        (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
-        (uint32_t)(ne00 * ne01), stride_batch_y, (uint32_t)(ne20 * ne21),
-        fusion_flags,
-        (uint32_t)nei0, (uint32_t)ne11,
-    };
-    ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
-        {
-            d_X,
-            d_Y,
-            d_D,
-            d_F0,
-            d_F1,
-            d_ids,
-        },
-        pc, { groups_x, (uint32_t)nei0, groups_z });
+    // Loop over the batch dimension
+    for (uint32_t expert_i1 = 0; expert_i1 < nei1; ++expert_i1) {
+        const vk_mat_vec_id_push_constants pc = {
+            (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
+            (uint32_t)(ne00 * ne01), stride_batch_y, (uint32_t)(ne20 * ne21),
+            fusion_flags,
+            (uint32_t)nei0, (uint32_t)ne11, expert_i1, nbi1
+        };
+        ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
+            {
+                d_X,
+                d_Y,
+                d_D,
+                d_F0,
+                d_F1,
+                d_ids,
+            },
+            pc, { groups_x, (uint32_t)nei0, groups_z });
+    }
 
     if (x_non_contig) {
         ctx->prealloc_x_need_sync = true;
@@ -8292,7 +8325,7 @@ static bool ggml_vk_use_mul_mat_vec_id(const struct ggml_cgraph * cgraph, int no
     ggml_tensor * dst = cgraph->nodes[node_idx];
     ggml_tensor * src0 = dst->src[0];
     ggml_tensor * src2 = dst->src[2];
-    return src2->ne[1] == 1 && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type));
+    return (src2->ne[1] <= 8) && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type));
 }
 
 static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx, const struct ggml_cgraph * cgraph, int node_idx) {
@@ -8325,41 +8358,49 @@ static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, con
     const uint32_t total_size = tmpsh + tmpshv4 + masksh + Qf;
     const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize;
 
-    VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", total_size=" << total_size << ", supported=" << supported);
+    VK_LOG_DEBUG("ggml_vk_flash_attn_scalar_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", total_size=" << total_size << ", supported=" << supported);
 
     return supported;
 }
 
-static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const uint32_t hsk, uint32_t hsv, bool f32acc) {
+static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const uint32_t hsk, uint32_t hsv, bool f32acc, ggml_type kv_type) {
     // Needs to be kept up to date on shader changes
     GGML_UNUSED(hsv);
-    const uint32_t wg_size = scalar_flash_attention_workgroup_size;
-    const uint32_t Br = coopmat1_flash_attention_num_large_rows;
-    const uint32_t Bc = scalar_flash_attention_Bc;
+    const auto rows_cols = fa_rows_cols(FA_COOPMAT1, hsk, hsv, 0, kv_type, false, false);
+    const uint32_t Br = rows_cols[0];
+    const uint32_t Bc = rows_cols[1];
+
+    const uint32_t MatBr = 16, MatBc = 16;
+
+    const uint32_t row_split = Bc / MatBc;
 
     const uint32_t hsk_pad = ROUNDUP_POW2(hsk, 16);
 
     const uint32_t acctype = f32acc ? 4 : 2;
     const uint32_t f16vec4 = 8;
 
-    const uint32_t tmpsh = wg_size * sizeof(float);
-    const uint32_t tmpshv4 = wg_size * 4 * acctype;
+    const uint32_t tmpsh = (Bc / MatBc) * sizeof(float);
 
     const uint32_t qstride = hsk_pad / 4 + 2;
     const uint32_t Qf = Br * qstride * f16vec4;
 
+    const uint32_t psh_stride = Br / 4 + 2;
+    const uint32_t Psh = Bc * psh_stride * f16vec4;
+
     const uint32_t sfshstride = (hsk <= 128) ? (Br + 8) : Br;
     const uint32_t sfsh = Bc * sfshstride * acctype;
 
-    const uint32_t kshstride = hsk_pad / 4 + 2;
-    const uint32_t ksh = Bc * kshstride * f16vec4;
+    const bool k_load_shmem = device->vendor_id == VK_VENDOR_ID_NVIDIA;
+    const uint32_t kshstride = (k_load_shmem ? hsk_pad : MatBr) / 4 + 2;
+    const uint32_t vsh_stride = MatBc / 4 * row_split;
+    const uint32_t ksh = ((kshstride >= vsh_stride) ? (Bc * kshstride) : (Bc * vsh_stride)) * f16vec4;
 
-    const uint32_t slope = Br * sizeof(float);
+    const uint32_t slope = Br * acctype;
 
-    const uint32_t total_size = tmpsh + tmpshv4 + Qf + sfsh + ksh + slope;
+    const uint32_t total_size = tmpsh + Qf + Psh + sfsh + ksh + slope;
     const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize;
 
-    VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", f32acc=" << f32acc << ", total_size=" << total_size << ", supported=" << supported);
+    VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", f32acc=" << f32acc << ", kv_type=" << kv_type << ", total_size=" << total_size << ", supported=" << supported);
 
     return supported;
 }
@@ -8423,7 +8464,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         const bool coopmat_shape_supported = (dst->op_params[3] == GGML_PREC_F32 && ctx->device->coopmat_support_16x16x16_f32acc) ||
                                              (dst->op_params[3] != GGML_PREC_F32 && ctx->device->coopmat_support_16x16x16_f16acc);
 
-        const bool coopmat_shmem_supported = ggml_vk_flash_attn_coopmat_shmem_support(ctx->device, HSK, HSV, dst->op_params[3] == GGML_PREC_F32);
+        const bool coopmat_shmem_supported = ggml_vk_flash_attn_coopmat_shmem_support(ctx->device, HSK, HSV, dst->op_params[3] == GGML_PREC_F32, k->type);
 
         if (!coopmat_shape_supported || !coopmat_shmem_supported) {
             path = FA_SCALAR;
@@ -8454,14 +8495,14 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         GGML_ASSERT(0);
     }
 
-    if (N == 1 && qk_ratio > 1 && qk_ratio <= max_gqa &&
+    if (N <= 8 && qk_ratio > 1 && qk_ratio <= max_gqa &&
         qk_ratio * nek2 == neq2 && nek2 == nev2 && nem2 <= 1) {
         // grouped query attention - make the N dimension equal to gqa_ratio, reduce
         // workgroups proportionally in y dimension. The shader will detect gqa_ratio > 1
         // and change addressing calculations to index Q's dimension 2.
         gqa_ratio = qk_ratio;
         N = gqa_ratio;
-        workgroups_y /= N;
+        workgroups_y /= gqa_ratio;
     }
 
     bool small_rows = N <= get_fa_num_small_rows(path);
@@ -8523,6 +8564,8 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     }
 
     assert(pipeline);
+    // Compile early to initialize wg_denoms.
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
 
     uint32_t split_kv = KV;
     uint32_t split_k = 1;
@@ -8530,22 +8573,24 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     // Use a placeholder core count if one isn't available. split_k is a big help for perf.
     const uint32_t shader_core_count = ctx->device->shader_core_count ? ctx->device->shader_core_count : 16;
 
-    // Try to use split_k when KV is large enough to be worth the overhead
-    if (workgroups_x == 1 && shader_core_count > 0) {
+    // Try to use split_k when KV is large enough to be worth the overhead.
+    // Must either be a single batch or be using gqa, we can't mix the two.
+    if (workgroups_x <= pipeline->wg_denoms[0] && (workgroups_x == 1 || gqa_ratio > 1)) {
         // Try to run two workgroups per SM.
-        split_k = shader_core_count * 2 / (workgroups_y * workgroups_z);
+        split_k = shader_core_count * 2 / (workgroups_x * workgroups_y * workgroups_z);
         if (split_k > 1) {
             // Try to evenly split KV into split_k chunks, but it needs to be a multiple
             // of "align", so recompute split_k based on that.
             split_kv = ROUNDUP_POW2(std::max(1u, KV / split_k), alignment);
             split_k = CEIL_DIV(KV, split_kv);
-            workgroups_x = split_k;
         }
     }
 
     // Reserve space for split_k temporaries. For each split x batch, we need to store the O matrix (D x ne1)
     // and the per-row m and L values (ne1 rows). We store all the matrices first, followed by the rows.
-    const uint64_t split_k_size = split_k > 1 ? (HSV * ne1 * sizeof(float) + ne1 * sizeof(float) * 2) * split_k * ne3 : 0;
+    // For matrices, the order is (inner to outer) [HSV, ne1, k, ne2, ne3].
+    // For L/M, the order is (inner to outer) [ne1, k, ne2, ne3].
+    const uint64_t split_k_size = split_k > 1 ? (HSV * ne1 * sizeof(float) + ne1 * sizeof(float) * 2) * split_k * ne2 * ne3 : 0;
     if (split_k_size > ctx->device->properties.limits.maxStorageBufferRange) {
         GGML_ABORT("Requested preallocation size is too large");
     }
@@ -8556,7 +8601,6 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
 
     {
         // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
         if (split_k > 1) {
             ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_flash_attn_split_k_reduce, 1);
         }
@@ -8605,7 +8649,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         if (ctx->prealloc_split_k_need_sync) {
             ggml_vk_sync_buffers(ctx, subctx);
         }
-
+        workgroups_x *= pipeline->wg_denoms[0];
         vk_subbuffer split_k_buf = ggml_vk_subbuffer(ctx, ctx->prealloc_split_k, 0);
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
                                     {q_buf, k_buf, v_buf, mask_buf, sinks_buf, split_k_buf},
@@ -8613,15 +8657,19 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                     // there's no more than one tile of rows (i.e. workgroups_x would have been
                                     // one). We reuse workgroups_x to mean the number of splits, so we need to
                                     // cancel out the divide by wg_denoms[0].
-                                    pc, { workgroups_x * pipeline->wg_denoms[0], workgroups_y, workgroups_z });
+                                    pc, { split_k * workgroups_x, workgroups_y, workgroups_z });
 
         ggml_vk_sync_buffers(ctx, subctx);
-        const std::array<uint32_t, 5> pc2 = { HSV, (uint32_t)ne1, (uint32_t)ne3, split_k, (sinks != nullptr) };
+        const vk_op_flash_attn_split_k_reduce_push_constants pc2 = { HSV, (uint32_t)ne1, (uint32_t)ne2, (uint32_t)ne3, split_k, (sinks != nullptr) };
         ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_flash_attn_split_k_reduce,
                                     {split_k_buf, sinks_buf, dst_buf},
-                                    pc2, { (uint32_t)ne1, HSV, (uint32_t)ne3 });
+                                    pc2, { (uint32_t)ne1, HSV, (uint32_t)(ne2 * ne3) });
         ctx->prealloc_split_k_need_sync = true;
     } else {
+        if (gqa_ratio > 1) {
+            // When using gqa, we want one actual workgroup per batch, so cancel out wg_denoms
+            workgroups_x *= pipeline->wg_denoms[0];
+        }
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
                                     {q_buf, k_buf, v_buf, mask_buf, sinks_buf, dst_buf},
                                     pc, { workgroups_x, workgroups_y, workgroups_z });
@@ -11560,7 +11608,6 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
     free(d_chk);
 
     ggml_vk_command_pool_cleanup(ctx->device, ctx->compute_cmd_pool);
-    ggml_vk_command_pool_cleanup(ctx->device, ctx->transfer_cmd_pool);
 
     ggml_vk_destroy_buffer(d_X);
     ggml_vk_destroy_buffer(d_Y);
@@ -11909,7 +11956,8 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
         }
     }
     if (mmq) {
-        ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_quantize_q8_1, num_it);
+        vk_pipeline pipeline_quantize_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline_quantize_q8_1, num_it);
     }
 
     ggml_pipeline_allocate_descriptor_sets(ctx);
@@ -12145,7 +12193,9 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx, vk_contex
         ggml_vk_submit(subctx, {});
         ctx->submit_pending = true;
         ggml_vk_synchronize(ctx);
+        GGML_ASSERT(ctx->compute_ctx.expired());
         ggml_vk_ctx_begin(ctx->device, subctx);
+        ctx->compute_ctx = subctx;
     }
 
     if (ctx->prealloc_x == nullptr || (ctx->prealloc_size_x > 0 && ctx->prealloc_x->size < ctx->prealloc_size_x)) {
@@ -12163,6 +12213,7 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx, vk_contex
             ggml_vk_destroy_buffer(ctx->prealloc_y);
         }
         ctx->prealloc_y = ggml_vk_create_buffer_device(ctx->device, ctx->prealloc_size_y);
+        ctx->prealloc_y_last_tensor_used = nullptr;
     }
     if (ctx->prealloc_split_k == nullptr || (ctx->prealloc_size_split_k > 0 && ctx->prealloc_split_k->size < ctx->prealloc_size_split_k)) {
         VK_LOG_MEMORY("ggml_vk_preallocate_buffers(split_k_size: " << ctx->prealloc_size_split_k << ")");
@@ -12191,6 +12242,9 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     if (ggml_is_empty(node) || ggml_op_is_empty(node->op) || !node->buffer) {
         return false;
     }
+    if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+        return false;
+    }
 
     VK_LOG_DEBUG("ggml_vk_build_graph(" << node << ", " << ggml_op_name(node->op) << ")");
     ctx->semaphore_idx = 0;
@@ -12740,7 +12794,6 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) {
     ctx->prealloc_x_need_sync = ctx->prealloc_y_need_sync = ctx->prealloc_split_k_need_sync = false;
 
     ggml_vk_command_pool_cleanup(ctx->device, ctx->compute_cmd_pool);
-    ggml_vk_command_pool_cleanup(ctx->device, ctx->transfer_cmd_pool);
 
     for (size_t i = 0; i < ctx->gc.semaphores.size(); i++) {
         ctx->device->device.destroySemaphore({ ctx->gc.semaphores[i].s });
@@ -12769,7 +12822,7 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) {
 static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) {
     VK_LOG_DEBUG("ggml_vk_cleanup(" << ctx->name << ")");
     // discard any unsubmitted command buffers
-    ctx->transfer_ctx.reset();
+    ctx->compute_ctx.reset();
     // wait for any pending command buffers to finish
     ggml_vk_synchronize(ctx);
 
@@ -12802,7 +12855,6 @@ static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) {
     ctx->descriptor_sets.clear();
 
     ctx->compute_cmd_pool.destroy(ctx->device->device);
-    ctx->transfer_cmd_pool.destroy(ctx->device->device);
     if (vk_perf_logger_enabled) {
         ctx->perf_logger->print_timings(true);
     }
@@ -13074,34 +13126,34 @@ static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, ggml_tensor
 
     ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context;
 
-    vk_context transfer_ctx;
+    vk_context compute_ctx;
 
-    if (ctx->transfer_ctx.expired()) {
+    if (ctx->compute_ctx.expired()) {
         // Initialize new transfer context
-        transfer_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
-        ctx->transfer_ctx = transfer_ctx;
-        ggml_vk_ctx_begin(ctx->device, transfer_ctx);
+        compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
+        ctx->compute_ctx = compute_ctx;
+        ggml_vk_ctx_begin(ctx->device, compute_ctx);
     } else {
-        transfer_ctx = ctx->transfer_ctx.lock();
+        compute_ctx = ctx->compute_ctx.lock();
     }
 
     vk_buffer buf = buf_ctx->dev_buffer;
 
     auto dst_offset = vk_tensor_offset(tensor) + tensor->view_offs + offset;
 
-    bool ret = ggml_vk_buffer_write_async(transfer_ctx, buf, dst_offset, data, size);
+    bool ret = ggml_vk_buffer_write_async(compute_ctx, buf, dst_offset, data, size);
 
     if (!ret) {
         ggml_vk_ensure_sync_staging_buffer(ctx, size);
-        ggml_vk_sync_buffers(nullptr, transfer_ctx);
+        ggml_vk_sync_buffers(nullptr, compute_ctx);
 
         vk::BufferCopy buffer_cpy;
         buffer_cpy.srcOffset = 0;
         buffer_cpy.dstOffset = dst_offset;
         buffer_cpy.size = size;
 
-        transfer_ctx->s->buffer.copyBuffer(ctx->sync_staging->buffer, buf->buffer, { buffer_cpy });
-        deferred_memcpy(ctx->sync_staging->ptr, data, size, &transfer_ctx->in_memcpys);
+        compute_ctx->s->buffer.copyBuffer(ctx->sync_staging->buffer, buf->buffer, { buffer_cpy });
+        deferred_memcpy(ctx->sync_staging->ptr, data, size, &compute_ctx->in_memcpys);
         ggml_vk_synchronize(ctx);
     }
 }
@@ -13113,34 +13165,34 @@ static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, const ggml_
 
     ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context;
 
-    vk_context transfer_ctx;
+    vk_context compute_ctx;
 
-    if (ctx->transfer_ctx.expired()) {
+    if (ctx->compute_ctx.expired()) {
         // Initialize new transfer context
-        transfer_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
-        ctx->transfer_ctx = transfer_ctx;
-        ggml_vk_ctx_begin(ctx->device, transfer_ctx);
+        compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
+        ctx->compute_ctx = compute_ctx;
+        ggml_vk_ctx_begin(ctx->device, compute_ctx);
     } else {
-        transfer_ctx = ctx->transfer_ctx.lock();
+        compute_ctx = ctx->compute_ctx.lock();
     }
 
     vk_buffer buf = buf_ctx->dev_buffer;
 
     auto src_offset = vk_tensor_offset(tensor) + tensor->view_offs + offset;
-    bool ret = ggml_vk_buffer_read_async(transfer_ctx, buf, src_offset, data, size);
+    bool ret = ggml_vk_buffer_read_async(compute_ctx, buf, src_offset, data, size);
 
     // If that failed, copy synchronously through a staging buffer
     if (!ret) {
         ggml_vk_ensure_sync_staging_buffer(ctx, size);
-        ggml_vk_sync_buffers(nullptr, transfer_ctx);
+        ggml_vk_sync_buffers(nullptr, compute_ctx);
 
         vk::BufferCopy buffer_cpy;
         buffer_cpy.srcOffset = src_offset;
         buffer_cpy.dstOffset = 0;
         buffer_cpy.size = size;
 
-        transfer_ctx->s->buffer.copyBuffer(buf->buffer, ctx->sync_staging->buffer, { buffer_cpy });
-        deferred_memcpy(data, ctx->sync_staging->ptr, size, &transfer_ctx->out_memcpys);
+        compute_ctx->s->buffer.copyBuffer(buf->buffer, ctx->sync_staging->buffer, { buffer_cpy });
+        deferred_memcpy(data, ctx->sync_staging->ptr, size, &compute_ctx->out_memcpys);
         ggml_vk_synchronize(ctx);
     }
 }
@@ -13152,21 +13204,21 @@ static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_
         ggml_backend_vk_buffer_context * src_buf_ctx = (ggml_backend_vk_buffer_context *)src->buffer->context;
         ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
 
-        vk_context transfer_ctx;
+        vk_context compute_ctx;
 
-        if (ctx->transfer_ctx.expired()) {
+        if (ctx->compute_ctx.expired()) {
             // Initialize new transfer context
-            transfer_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
-            ctx->transfer_ctx = transfer_ctx;
-            ggml_vk_ctx_begin(ctx->device, transfer_ctx);
+            compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
+            ctx->compute_ctx = compute_ctx;
+            ggml_vk_ctx_begin(ctx->device, compute_ctx);
         } else {
-            transfer_ctx = ctx->transfer_ctx.lock();
+            compute_ctx = ctx->compute_ctx.lock();
         }
 
         vk_buffer src_buf = src_buf_ctx->dev_buffer;
         vk_buffer dst_buf = dst_buf_ctx->dev_buffer;
 
-        ggml_vk_buffer_copy_async(transfer_ctx, dst_buf, vk_tensor_offset(dst) + dst->view_offs, src_buf, vk_tensor_offset(src) + src->view_offs, ggml_nbytes(src));
+        ggml_vk_buffer_copy_async(compute_ctx, dst_buf, vk_tensor_offset(dst) + dst->view_offs, src_buf, vk_tensor_offset(src) + src->view_offs, ggml_nbytes(src));
         return true;
     }
 
@@ -13176,19 +13228,19 @@ static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_
 static void ggml_vk_synchronize(ggml_backend_vk_context * ctx) {
     VK_LOG_DEBUG("ggml_vk_synchronize()");
 
-    bool do_transfer = !ctx->transfer_ctx.expired();
+    bool do_transfer = !ctx->compute_ctx.expired();
 
-    vk_context transfer_ctx;
+    vk_context compute_ctx;
     if (do_transfer) {
-        transfer_ctx = ctx->transfer_ctx.lock();
+        compute_ctx = ctx->compute_ctx.lock();
 
-        ggml_vk_ctx_end(transfer_ctx);
+        ggml_vk_ctx_end(compute_ctx);
 
-        for (auto& cpy : transfer_ctx->in_memcpys) {
+        for (auto& cpy : compute_ctx->in_memcpys) {
             memcpy(cpy.dst, cpy.src, cpy.n);
         }
 
-        ggml_vk_submit(transfer_ctx, {});
+        ggml_vk_submit(compute_ctx, {});
         ctx->submit_pending = true;
     }
 
@@ -13202,10 +13254,10 @@ static void ggml_vk_synchronize(ggml_backend_vk_context * ctx) {
     }
 
     if (do_transfer) {
-        for (auto& cpy : transfer_ctx->out_memcpys) {
+        for (auto& cpy : compute_ctx->out_memcpys) {
             memcpy(cpy.dst, cpy.src, cpy.n);
         }
-        ctx->transfer_ctx.reset();
+        ctx->compute_ctx.reset();
     }
 }
 
@@ -13645,7 +13697,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
     int last_node = cgraph->n_nodes - 1;
 
     // If the last op in the cgraph isn't backend GPU, the command buffer doesn't get closed properly
-    while (last_node > 0 && ggml_vk_is_empty(cgraph->nodes[last_node])) {
+    while (last_node > 0 && (ggml_vk_is_empty(cgraph->nodes[last_node]) || ((cgraph->nodes[last_node]->flags & GGML_TENSOR_FLAG_COMPUTE) == 0))) {
         last_node -= 1;
     }
 
@@ -13874,6 +13926,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         ggml_vk_submit(compute_ctx, ctx->device->fence);
         VK_CHECK(ctx->device->device.waitForFences({ ctx->device->fence }, true, UINT64_MAX), "GGML_VULKAN_PERF waitForFences");
         ctx->device->device.resetFences({ ctx->device->fence });
+        ctx->compute_ctx.reset();
 
         // Get the results and pass them to the logger
         std::vector<uint64_t> timestamps(cgraph->n_nodes + 1);
@@ -14160,15 +14213,15 @@ static void ggml_backend_vk_event_record(ggml_backend_t backend, ggml_backend_ev
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
     vk_event *vkev = (vk_event *)event->context;
 
-    vk_context transfer_ctx;
+    vk_context compute_ctx;
 
-    if (ctx->transfer_ctx.expired()) {
+    if (ctx->compute_ctx.expired()) {
         // Initialize new transfer context
-        transfer_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
-        ctx->transfer_ctx = transfer_ctx;
-        ggml_vk_ctx_begin(ctx->device, transfer_ctx);
+        compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
+        ctx->compute_ctx = compute_ctx;
+        ggml_vk_ctx_begin(ctx->device, compute_ctx);
     } else {
-        transfer_ctx = ctx->transfer_ctx.lock();
+        compute_ctx = ctx->compute_ctx.lock();
     }
 
     // the backend interface doesn't have an explicit reset, so reset it here
@@ -14176,13 +14229,13 @@ static void ggml_backend_vk_event_record(ggml_backend_t backend, ggml_backend_ev
     ctx->device->device.resetEvent(vkev->event);
     ctx->device->device.resetFences({ vkev->fence });
 
-    ggml_vk_set_event(transfer_ctx, vkev->event);
+    ggml_vk_set_event(compute_ctx, vkev->event);
 
-    ggml_vk_ctx_end(transfer_ctx);
+    ggml_vk_ctx_end(compute_ctx);
 
-    ggml_vk_submit(transfer_ctx, {vkev->fence});
+    ggml_vk_submit(compute_ctx, {vkev->fence});
     ctx->submit_pending = true;
-    ctx->transfer_ctx.reset();
+    ctx->compute_ctx.reset();
 }
 
 static void ggml_backend_vk_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
@@ -14190,20 +14243,20 @@ static void ggml_backend_vk_event_wait(ggml_backend_t backend, ggml_backend_even
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
     vk_event *vkev = (vk_event *)event->context;
 
-    vk_context transfer_ctx;
+    vk_context compute_ctx;
 
-    if (ctx->transfer_ctx.expired()) {
+    if (ctx->compute_ctx.expired()) {
         // Initialize new transfer context
-        transfer_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
-        ctx->transfer_ctx = transfer_ctx;
-        ggml_vk_ctx_begin(ctx->device, transfer_ctx);
+        compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
+        ctx->compute_ctx = compute_ctx;
+        ggml_vk_ctx_begin(ctx->device, compute_ctx);
     } else {
-        transfer_ctx = ctx->transfer_ctx.lock();
+        compute_ctx = ctx->compute_ctx.lock();
     }
 
-    ggml_vk_wait_events(transfer_ctx, {vkev->event});
-    ggml_vk_ctx_end(transfer_ctx);
-    ctx->transfer_ctx.reset();
+    ggml_vk_wait_events(compute_ctx, {vkev->event});
+    ggml_vk_ctx_end(compute_ctx);
+    ctx->compute_ctx.reset();
 }
 
 // TODO: enable async and synchronize
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
index 0379e5d502..3ce8d07be8 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
@@ -53,7 +53,7 @@ void main() {
     const uint32_t d_tid = gl_LocalInvocationIndex % D_split;
     const uint32_t col_tid = gl_LocalInvocationIndex / D_split;
 
-    uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
+    uint32_t q_offset = gqa_iq1*p.nb01 + (iq2*p.nb02 + iq3*p.nb03) / 4;
 
     [[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
         uint32_t d = (idx + tid) % (HSK / 4);
@@ -101,9 +101,9 @@ void main() {
     uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2;
     uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
 #endif
-    uint32_t m_offset = 0;
+    uint32_t m_offset = gqa_iq1*KV;
     if (p.nem2 != 1 || p.nem3 != 1) {
-        m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
+        m_offset += ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
     }
 
     [[dont_unroll]]
@@ -320,7 +320,8 @@ void main() {
     // If there is split_k, then the split_k resolve shader does the final
     // division by L. Store the intermediate O value and per-row m and L values.
     if (p.k_num > 1) {
-        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
+        // note: O and Q have swapped coord 1,2.
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
 
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
@@ -332,7 +333,7 @@ void main() {
             }
         }
 
-        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
+        o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
                 perElemOpStoreCol0(r, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
@@ -378,7 +379,7 @@ void main() {
         }
     }
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
+    uint32_t o_offset = gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV;
 
     if (p.gqa_ratio > 1) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl
index eb93903c46..23a4d2c005 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl
@@ -8,6 +8,8 @@ layout (constant_id = 3) const uint32_t HSK = 32;
 layout (constant_id = 4) const uint32_t HSV = 32;
 layout (constant_id = 5) const uint32_t Clamp = 0;
 layout (constant_id = 6) const uint32_t D_split = 16;
+layout (constant_id = 7) const uint32_t SubGroupSize = 32;
+layout (constant_id = 8) const uint32_t K_LOAD_SHMEM = 0;
 
 // Round up head sizes to a multiple of 16, for coopmat1/coopmat2 paths
 const uint32_t HSK_pad = (HSK + 15) & ~15;
@@ -74,6 +76,10 @@ layout (binding = 1) readonly buffer K_PACKED16 {A_TYPE_PACKED16 k_data_packed16
 layout (binding = 2) readonly buffer V_PACKED16 {A_TYPE_PACKED16 v_data_packed16[];} v_packed;
 #endif
 
+#ifndef BLOCK_SIZE
+#define BLOCK_SIZE 1
+#endif
+
 #if defined(DATA_A_F32)
 #undef BLOCK_SIZE
 #define BLOCK_SIZE 4
@@ -165,7 +171,7 @@ ACC_TYPE perElemOpGetSink(const in uint32_t r, const in uint32_t c, const in ACC
 }
 
 uint32_t i, N, KV, split_k_index, Tr, start_j, end_j,
-         iq2, iq3, rk2, rk3, rv2, rv3, ik2, ik3, iv2, iv3,
+         gqa_iq1, iq2, iq3, rk2, rk3, rv2, rv3, ik2, ik3, iv2, iv3,
          q_stride, k_stride, v_stride, m_stride;
 
 void init_indices()
@@ -173,12 +179,19 @@ void init_indices()
     N = p.N;
     KV = p.KV;
 
-    i = gl_WorkGroupID.x;
-    split_k_index = 0;
-
     if (p.k_num > 1) {
         i = 0;
-        split_k_index = gl_WorkGroupID.x;
+        // batch and split_k share gl_WorkGroupID.x
+        gqa_iq1 = gl_WorkGroupID.x / p.k_num;
+        split_k_index = gl_WorkGroupID.x % p.k_num;
+    } else if (p.gqa_ratio > 1) {
+        i = 0;
+        gqa_iq1 = gl_WorkGroupID.x;
+        split_k_index = 0;
+    } else {
+        i = gl_WorkGroupID.x;
+        gqa_iq1 = 0;
+        split_k_index = 0;
     }
 
     Tr = CEIL_DIV(N, Br);
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
index c995ab140e..83d52d19d6 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
@@ -7,6 +7,7 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
 
 #extension GL_KHR_shader_subgroup_basic : enable
+#extension GL_KHR_shader_subgroup_arithmetic : enable
 #extension GL_KHR_shader_subgroup_vote : enable
 #extension GL_KHR_memory_scope_semantics : enable
 #extension GL_KHR_cooperative_matrix : enable
@@ -14,12 +15,13 @@
 #include "types.glsl"
 #include "flash_attn_base.glsl"
 
-const uint32_t HSK_per_thread = HSK / D_split;
-const uint32_t HSV_per_thread = HSV / D_split;
+// These need to be supported N,M values for a MatBc x MatBr x 16 coopmatmuladd
+const uint32_t MatBr = 16;
+const uint32_t MatBc = 16;
 
-const uint32_t row_split = 4;
+const uint32_t row_split = Bc / MatBc;
 const uint32_t rows_per_thread = Br / row_split;
-const uint32_t cols_per_iter = gl_WorkGroupSize.x / D_split / row_split;
+const uint32_t cols_per_iter = gl_WorkGroupSize.x / row_split;
 const uint32_t cols_per_thread = Bc / cols_per_iter;
 
 
@@ -40,24 +42,24 @@ D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TY
     return elem;
 }
 
-// These need to be supported N,M values for a MatBc x MatBr x 16 coopmatmuladd
-const uint32_t MatBr = 16;
-const uint32_t MatBc = 16;
-
-shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x];
-shared ACC_TYPEV4 tmpshv4[gl_WorkGroupSize.x];
+shared float tmpsh[row_split];
 
 const uint32_t qstride = HSK_pad / 4 + 2; // in units of f16vec4
 shared f16vec4 Qf[Br * qstride];
 
+const uint psh_stride = Br / 4 + 2;
+shared f16vec4 Psh[Bc * psh_stride];
+
 // Avoid padding for hsk==256 to make it fit in 48KB shmem.
-const uint32_t sfshstride = (HSK <= 128) ? (Br + 8) : Br;
-shared ACC_TYPE sfsh[Bc * sfshstride];
+const uint32_t sfshstride = (HSK <= 128) ? (Br / 4 + 2) : Br / 4;
+shared ACC_TYPEV4 sfsh[Bc * sfshstride];
 
-const uint32_t kshstride = HSK_pad / 4 + 2; // in units of f16vec4
-shared f16vec4 ksh[Bc * kshstride];
+const uint32_t kshstride = (K_LOAD_SHMEM != 0 ? HSK_pad : MatBr) / 4 + 2; // in units of f16vec4
+const uint v_cols = MatBc / 4 * row_split; // total cols, 4 vec4s per MatBc * number of subgroups
+const uint vsh_stride = v_cols;
+shared f16vec4 ksh[(kshstride >= vsh_stride) ? (Bc * kshstride) : (Bc * vsh_stride)];
 
-shared float slope[Br];
+shared ACC_TYPE slope[Br];
 
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
@@ -69,9 +71,9 @@ void main() {
     const uint32_t tid = gl_LocalInvocationIndex;
 
     const uint32_t threads_per_rowgroup = gl_WorkGroupSize.x / row_split;
+    const uint32_t d_per_thread = (HSV/4 + threads_per_rowgroup - 1) / threads_per_rowgroup;
     const uint32_t row_tid = gl_LocalInvocationIndex / threads_per_rowgroup;
-    const uint32_t d_tid = gl_LocalInvocationIndex % D_split;
-    const uint32_t col_tid = (gl_LocalInvocationIndex % threads_per_rowgroup) / D_split;
+    const uint32_t col_tid = gl_LocalInvocationIndex % threads_per_rowgroup;
 
 #define tile_row(r) (row_tid * rows_per_thread + (r))
 
@@ -90,7 +92,7 @@ void main() {
         barrier();
     }
 
-    uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
+    uint32_t q_offset = gqa_iq1*p.nb01 + (iq2*p.nb02+iq3*p.nb03) / 4;
 
     [[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
         uint32_t d = (idx + tid) % (HSK / 4);
@@ -102,9 +104,9 @@ void main() {
     }
     barrier();
 
-    ACC_TYPEV4 Of[rows_per_thread][HSV_per_thread / 4];
-    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+    ACC_TYPEV4 Of[rows_per_thread][d_per_thread];
+    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+        [[unroll]] for (uint32_t d = 0; d < d_per_thread; ++d) {
             Of[r][d] = ACC_TYPEV4(0.0);
         }
     }
@@ -125,13 +127,11 @@ void main() {
             uint r = tid;
             slope[r] = perElemOpComputeSlope(r, col_tid, ACC_TYPE(0), iq2);
         }
-        barrier();
     } else {
         if (tid < Br) {
             uint r = tid;
-            slope[r] = 1.0;
+            slope[r] = ACC_TYPE(1.0);
         }
-        barrier();
     }
 
 #if BLOCK_SIZE > 1
@@ -141,27 +141,53 @@ void main() {
     uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2;
     uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
 #endif
-    uint32_t m_offset = 0;
+    uint32_t m_offset = gqa_iq1*KV;
     if (p.nem2 != 1 || p.nem3 != 1) {
-        m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
+        m_offset += ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
     }
 
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
 
-        float mask_cache[Bc * Br / WorkGroupSize];
+        f16vec4 mask_cache[Bc * Br / 4 / WorkGroupSize];
         if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
             bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
 
             float max_mask = NEG_FLT_MAX_OVER_2;
-            [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
-                uint32_t c = (idx + tid) % Bc;
-                uint32_t r = (idx + tid) / Bc;
-                if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) {
-                    if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
-                        float m = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * Br / 4; idx += gl_WorkGroupSize.x) {
+                uint32_t c = (idx + tid) / (Br / 4);
+                uint32_t r = (idx + tid) % (Br / 4);
+                if (idx + tid < Bc * Br / 4 || idx + gl_WorkGroupSize.x <= Bc * Br / 4) {
+                    if ((!KV_bounds_check || j * Bc + c < KV)) {
+                        f16vec4 m;
+                        if (!nem1_bounds_check || i * Br + r * 4 + 3 < p.nem1) {
+                            m = f16vec4(data_m[m_offset + (i * Br + r * 4    ) * m_stride + (j * Bc + c)],
+                                        data_m[m_offset + (i * Br + r * 4 + 1) * m_stride + (j * Bc + c)],
+                                        data_m[m_offset + (i * Br + r * 4 + 2) * m_stride + (j * Bc + c)],
+                                        data_m[m_offset + (i * Br + r * 4 + 3) * m_stride + (j * Bc + c)]);
+                            max_mask = max(max(max(max(max_mask, float(m[0])), float(m[1])), float(m[2])), float(m[3]));
+                        } else if (i * Br + r * 4 + 2 < p.nem1) {
+                            m = f16vec4(data_m[m_offset + (i * Br + r * 4    ) * m_stride + (j * Bc + c)],
+                                        data_m[m_offset + (i * Br + r * 4 + 1) * m_stride + (j * Bc + c)],
+                                        data_m[m_offset + (i * Br + r * 4 + 2) * m_stride + (j * Bc + c)],
+                                        0.0);
+                            max_mask = max(max(max(max_mask, float(m[0])), float(m[1])), float(m[2]));
+                        } else if (i * Br + r * 4 + 1 < p.nem1) {
+                            m = f16vec4(data_m[m_offset + (i * Br + r * 4    ) * m_stride + (j * Bc + c)],
+                                        data_m[m_offset + (i * Br + r * 4 + 1) * m_stride + (j * Bc + c)],
+                                        0.0,
+                                        0.0);
+                            max_mask = max(max(max_mask, float(m[0])), float(m[1]));
+                        } else if (i * Br + r * 4 < p.nem1) {
+                            m = f16vec4(data_m[m_offset + (i * Br + r * 4    ) * m_stride + (j * Bc + c)],
+                                        0.0,
+                                        0.0,
+                                        0.0);
+                            max_mask = max(max_mask, float(m[0]));
+                        } else {
+                            m = f16vec4(0.0);
+                        }
                         mask_cache[idx / WorkGroupSize] = m;
-                        max_mask = max(max_mask, m);
                     }
                 }
             }
@@ -180,26 +206,28 @@ void main() {
             }
         }
 
-        [[unroll]] for (uint32_t idx = 0; idx < Bc * HSK / 4; idx += gl_WorkGroupSize.x) {
-            uint32_t d = (idx + tid) % (HSK / 4);
-            uint32_t c = (idx + tid) / (HSK / 4);
-            if (c < Bc && d < HSK / 4) {
-                f16vec4 K_Tf = f16vec4(0);
-                if (!KV_bounds_check || j * Bc + c < KV) {
+        if (K_LOAD_SHMEM != 0) {
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * HSK / 4; idx += gl_WorkGroupSize.x) {
+                uint32_t d = (idx + tid) % (HSK / 4);
+                uint32_t c = (idx + tid) / (HSK / 4);
+                if (c < Bc && d < HSK / 4) {
+                    f16vec4 K_Tf = f16vec4(0);
+                    if (!KV_bounds_check || j * Bc + c < KV) {
 #if BLOCK_SIZE > 1
-                    uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d;
-                    uint ib = coord / BLOCK_SIZE;
-                    uint iqs = (coord % BLOCK_SIZE);
-                    K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K));
+                        uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d;
+                        uint ib = coord / BLOCK_SIZE;
+                        uint iqs = (coord % BLOCK_SIZE);
+                        K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K));
 #else
-                    K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + c) * k_stride / 4 + d]);
+                        K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + c) * k_stride / 4 + d]);
 #endif
-                }
+                    }
 
-                ksh[c * kshstride + d] = K_Tf;
+                    ksh[c * kshstride + d] = K_Tf;
+                }
             }
+            barrier();
         }
-        barrier();
 
         // K * Q^T -> S^T: Bc x HSK_pad * HSK_pad x Br -> Bc x Br
         // Bc split across workgroup (four subgroups), loop over HSK in chunks of 16: 16 x 16 * 16 x 16 -> 16 x 16
@@ -208,11 +236,55 @@ void main() {
         coopmat<float16_t, gl_ScopeSubgroup, MatBc, 16, gl_MatrixUseA> KMat;
         coopmat<float16_t, gl_ScopeSubgroup, 16, MatBr, gl_MatrixUseB> QMat;
 
-        for (uint32_t d = 0; d < HSK_pad / 16; ++d) {
-            coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor);
+        [[unroll]] for (uint32_t d = 0; d < HSK_pad / 16; ++d) {
+            if (K_LOAD_SHMEM == 0) {
+#if BLOCK_SIZE == 1
+            if (KV_bounds_check || d * 16 + 16 > HSK) {
+#endif
+            barrier();
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * MatBr / 4; idx += gl_WorkGroupSize.x) {
+                uint32_t col_vec = (idx + tid) % (MatBr / 4);
+                uint32_t row = (idx + tid) / (MatBr / 4);
+                if (idx + tid < Bc * MatBr / 4) {
+                    f16vec4 K_Tf = f16vec4(0);
+                    if ((!KV_bounds_check || j * Bc + row < KV) && (HSK == HSK_pad || d * 16 + col_vec * 4 < HSK)) {
+#if BLOCK_SIZE > 1
+                        uint coord = (j * Bc + row) * k_stride * BLOCK_SIZE + d * 16 + col_vec * 4;
+                        uint ib = coord / BLOCK_SIZE;
+                        uint iqs = (coord % BLOCK_SIZE);
+                        K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K));
+#else
+                        K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + row) * k_stride / 4 + d * 16 / 4 + col_vec]);
+#endif
+                    }
 
-            uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4;
-            coopMatLoad(KMat, ksh, coord, kshstride, gl_CooperativeMatrixLayoutRowMajor);
+                    ksh[row * kshstride + col_vec] = K_Tf;
+                }
+            }
+            barrier();
+#if BLOCK_SIZE == 1
+            }
+#endif
+
+#if BLOCK_SIZE == 1
+            if (KV_bounds_check || d * 16 + 16 > HSK)
+#endif
+            {
+                uint coord = (gl_SubgroupID * MatBc) * kshstride;
+                coopMatLoad(KMat, ksh, coord, kshstride, gl_CooperativeMatrixLayoutRowMajor);
+            }
+#if BLOCK_SIZE == 1
+            else {
+                const uint coord = k_offset / 4 + (j * Bc + gl_SubgroupID * MatBc) * k_stride / 4 + d * 16 / 4;
+                coopMatLoad(KMat, data_kv4, coord, k_stride / 4, gl_CooperativeMatrixLayoutRowMajor);
+            }
+#endif
+            } else {
+                uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4;
+                coopMatLoad(KMat, ksh, coord, kshstride, gl_CooperativeMatrixLayoutRowMajor);
+            }
+
+            coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor);
 
             SfMat = coopMatMulAdd(KMat, QMat, SfMat);
         }
@@ -222,26 +294,26 @@ void main() {
         barrier();
 
         if (p.logit_softcap != 0.0f) {
-            [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
-                uint32_t c = (idx + tid) / Br;
-                uint32_t r = (idx + tid) % Br;
-                if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) {
-                    sfsh[c * sfshstride + r] = ACC_TYPE(p.logit_softcap * tanh(sfsh[c * sfshstride + r]));
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * Br / 4; idx += gl_WorkGroupSize.x) {
+                uint32_t c = (idx + tid) / (Br / 4);
+                uint32_t r = (idx + tid) % (Br / 4);
+                if (idx + tid < Bc * Br / 4 || idx + gl_WorkGroupSize.x <= Bc * Br / 4) {
+                    sfsh[c * sfshstride + r] = ACC_TYPEV4(p.logit_softcap * tanh(sfsh[c * sfshstride + r]));
                 }
             }
             barrier();
         }
 
         if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
-            bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
-
-            [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
-                uint32_t c = (idx + tid) % Bc;
-                uint32_t r = (idx + tid) / Bc;
-                if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) {
-                    if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
-                        float f = mask_cache[idx / WorkGroupSize];
-                        sfsh[c * sfshstride + r] += ACC_TYPE(slope[r] * f);
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * Br / 4; idx += gl_WorkGroupSize.x) {
+                uint32_t c = (idx + tid) / (Br / 4);
+                uint32_t r = (idx + tid) % (Br / 4);
+                if (idx + tid < Bc * Br / 4 || idx + gl_WorkGroupSize.x <= Bc * Br / 4) {
+                    if (!KV_bounds_check || j * Bc + c < KV) {
+                        // Mask nem1 bounds check is handled when loading masks
+                        ACC_TYPEV4 masks = ACC_TYPEV4(mask_cache[idx / WorkGroupSize]);
+                        ACC_TYPEV4 slopes = ACC_TYPEV4(slope[r * 4], slope[r * 4 + 1], slope[r * 4 + 2], slope[r * 4 + 3]);
+                        sfsh[c * sfshstride + r] += slopes * masks;
                     }
                 }
             }
@@ -250,139 +322,176 @@ void main() {
 
         float eMf[rows_per_thread];
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            const uint r_vec  = tile_row(r) / 4;
+            const uint r_comp = tile_row(r) % 4;
+
             float rowmaxf = NEG_FLT_MAX_OVER_2;
             [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
                 if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
                     continue;
                 }
-                rowmaxf = max(rowmaxf, float(sfsh[tile_row(r) + (c * cols_per_iter + col_tid) * sfshstride]));
+                rowmaxf = max(rowmaxf, float(sfsh[r_vec + (c * cols_per_iter + col_tid) * sfshstride][r_comp]));
             }
             float Moldf = Mf[r];
 
+            // Compute max across the row
+            rowmaxf = subgroupMax(rowmaxf);
+
             // M = max(rowmax, Mold)
             // P = e^(S - M)
             // eM = e^(Mold - M)
             Mf[r] = max(rowmaxf, Moldf);
             eMf[r] = exp(Moldf - Mf[r]);
-        }
 
-        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                Of[r][d] = ACC_TYPE(eMf[r]) * Of[r][d];
-            }
-        }
-        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             Lf[r] = eMf[r]*Lf[r];
         }
 
-        [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-            if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
-                continue;
-            }
-            float Pf[rows_per_thread];
+        [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
+            const uint d_local = d0 / threads_per_rowgroup;
             [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                Pf[r] = exp(sfsh[tile_row(r) + (c * cols_per_iter + col_tid) * sfshstride] - Mf[r]);
-                Lf[r] += Pf[r];
+                Of[r][d_local] = ACC_TYPE(eMf[r]) * Of[r][d_local];
             }
-            [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-#if BLOCK_SIZE > 1
-                uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
-                uint ib = coord / BLOCK_SIZE;
-                uint iqs = (coord % BLOCK_SIZE);
-                vec4 Vf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
-#else
-                vec4 Vf = vec4(data_vv4[v_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * v_stride / 4 + d * D_split + d_tid]);
-#endif
-                [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                    Of[r][d] += ACC_TYPE(Pf[r]) * ACC_TYPEV4(Vf);
+        }
+
+        // Calculate and store Pf in Psh
+        [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+            const uint col = c * cols_per_iter + col_tid;
+
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; r += 4) {
+                const uint row = tile_row(r);
+                if (KV_bounds_check && j * Bc + col >= KV) {
+                    Psh[col * psh_stride + row / 4] = f16vec4(0.0f);
+                } else {
+                    const vec4 mfvec = vec4(Mf[r], Mf[r + 1], Mf[r + 2], Mf[r + 3]);
+                    const f16vec4 Pf = f16vec4(exp(vec4(sfsh[row / 4 + col * sfshstride]) - mfvec));
+                    [[unroll]] for (uint32_t vec_idx = 0; vec_idx < 4; ++vec_idx) {
+                        Lf[r + vec_idx] += Pf[vec_idx];
+                    }
+                    Psh[col * psh_stride + row / 4] = Pf;
                 }
             }
         }
 
-        barrier();
-    }
+        const uint num_hsv_tiles = (HSV + MatBc * row_split - 1) / (MatBc * row_split); // round up
 
-    // prevent race on tmpsh
-    barrier();
+        // Each subgroup handles HSV/4 columns
+        [[unroll]] for (uint32_t hsv_tile = 0; hsv_tile < num_hsv_tiles; ++hsv_tile) {
+            const uint hsv_offset = (hsv_tile * row_split + gl_SubgroupID) * 16;
 
-    // reduce across threads
+            SfMat = coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator>(0);
 
-    float rowmaxf[rows_per_thread], eMf[rows_per_thread], Moldf[rows_per_thread];
-    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-        FLOAT_TYPE M = Mf[r];
-        tmpsh[tid] = M;
-        // Compute max across the row
-        barrier();
-        [[unroll]] for (int s = int(gl_WorkGroupSize.x / row_split) / 2; s >= D_split; s >>= 1) {
-            M = max(M, tmpsh[tid ^ s]);
-            barrier();
-            tmpsh[tid] = M;
-            barrier();
-        }
-        rowmaxf[r] = tmpsh[d_tid + row_tid * threads_per_rowgroup];
-        barrier();
-    }
+            // Preload V tiles for [Bc, 16 * num subgroups]
+            const uint v_rows = Bc;
+            const uint v_total = v_rows * v_cols;
+            const uint v_loads_per_thread = v_total / gl_WorkGroupSize.x;
 
-    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-        Moldf[r] = Mf[r];
+#if BLOCK_SIZE == 1
+            // For f16, only preload if not aligned
+            if (KV_bounds_check) {
+#endif
+            [[unroll]] for (uint32_t i = 0; i < v_loads_per_thread; ++i) {
+                const uint idx = i * gl_WorkGroupSize.x + tid;
+                const uint row = idx / v_cols;
+                const uint col = idx % v_cols;
 
-        // M = max(rowmax, Mold)
-        // eM = e^(Mold - M)
-        Mf[r] = max(rowmaxf[r], Moldf[r]);
-        eMf[r] = exp(Moldf[r] - Mf[r]);
+                const uint v_row = j * Bc + row;
+                const uint v_col = hsv_tile * MatBc * row_split + col * 4;
 
-        Lf[r] = eMf[r]*Lf[r];
-    }
+                const uint coord = v_row * v_stride * BLOCK_SIZE + v_col;
+                const uint ib = coord / BLOCK_SIZE;
+                const uint iqs = coord % BLOCK_SIZE;
 
-    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-        FLOAT_TYPE L = Lf[r];
-        tmpsh[tid] = L;
-        // Compute sum across the row
-        barrier();
-        [[unroll]] for (int s = int(gl_WorkGroupSize.x / row_split) / 2; s >= D_split; s >>= 1) {
-            L += tmpsh[tid ^ s];
-            barrier();
-            tmpsh[tid] = L;
-            barrier();
-        }
-        Lf[r] = tmpsh[d_tid + row_tid * threads_per_rowgroup];
-        barrier();
-    }
-
-    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-
-            Of[r][d] = ACC_TYPE(eMf[r]) * Of[r][d];
-            tmpshv4[tid] = Of[r][d];
-
-            barrier();
-            [[unroll]] for (int s = int(gl_WorkGroupSize.x / row_split) / 2; s >= D_split; s >>= 1) {
-                Of[r][d] += tmpshv4[tid ^ s];
-                barrier();
-                tmpshv4[tid] = Of[r][d];
-                barrier();
+                if (!KV_bounds_check || (v_row < KV && v_col < HSV)) {
+#if BLOCK_SIZE > 1
+                    ksh[row * vsh_stride + col] = f16vec4(dequantize4(ib, iqs, v_offset, BINDING_IDX_V));
+#else
+                    ksh[row * vsh_stride + col] = data_vv4[(v_offset + v_row * v_stride + v_col) / 4];
+#endif
+                } else {
+                    ksh[row * vsh_stride + col] = f16vec4(0.0f);
+                }
             }
-            Of[r][d] = tmpshv4[d_tid + row_tid * threads_per_rowgroup];
+#if BLOCK_SIZE == 1
+            }
+#endif
+
             barrier();
-        }
-    }
 
-    // If there is split_k, then the split_k resolve shader does the final
-    // division by L. Store the intermediate O value and per-row m and L values.
-    if (p.k_num > 1) {
-        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
+            [[unroll]] for (uint32_t bc_chunk = 0; bc_chunk < Bc / MatBc; ++bc_chunk) {
+                coopMatLoad(KMat, Psh, bc_chunk * MatBc * psh_stride, psh_stride, gl_CooperativeMatrixLayoutColumnMajor);
 
-        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-            if (tile_row(r) < N) {
-                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N);
+#if BLOCK_SIZE == 1
+                if (!KV_bounds_check) {
+                    // F16 values can be loaded directly from global memory
+                    const uint v_tile_row = j * Bc + bc_chunk * MatBc;
+                    const uint v_tile_offset = v_offset / 4 + v_tile_row * v_stride / 4 + hsv_offset / 4;
+                    coopMatLoad(QMat, data_vv4, v_tile_offset, v_stride / 4, gl_CooperativeMatrixLayoutRowMajor);
+                } else
+#endif
+                {
+                    const uint v_tile_offset = bc_chunk * MatBr * v_cols + gl_SubgroupID * (MatBc / 4);
+                    coopMatLoad(QMat, ksh, v_tile_offset, vsh_stride, gl_CooperativeMatrixLayoutRowMajor);
+                }
+
+                SfMat = coopMatMulAdd(KMat, QMat, SfMat);
+            }
+
+            // Store SfMat to sfsh and load into Of
+            const uint osh_stride = row_split * MatBc / 4;
+            const uint o_offset = gl_SubgroupID * MatBc / 4;
+            coopMatStore(SfMat, sfsh, o_offset, osh_stride, gl_CooperativeMatrixLayoutRowMajor);
+
+            barrier();
+
+            const uint hsv_per_tile = row_split * MatBc;
+            const uint hsv_base = hsv_tile * hsv_per_tile;
+            const uint d_values_per_tile = hsv_per_tile / 4;
+
+            const uint d_start = hsv_tile * d_values_per_tile;
+            const uint d_end = min(d_start + d_values_per_tile, HSV / 4);
+
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                const uint row = tile_row(r);
+
+                [[unroll]] for (uint32_t d_local = 0; d_local < d_per_thread; ++d_local) {
+                    const uint d = d_local * threads_per_rowgroup + col_tid;
+                    const uint hsv_col = 4 * d;
+
+                    if (hsv_col >= hsv_base && hsv_col < hsv_base + hsv_per_tile && hsv_col < HSV) {
+                        const uint local_hsv = (hsv_col - hsv_base) / 4;
+                        Of[r][d_local] += ACC_TYPEV4(sfsh[row * osh_stride + local_hsv]);
                     }
                 }
             }
         }
 
-        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
+        barrier();
+    }
+
+    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+        Lf[r] = subgroupAdd(Lf[r]);
+    }
+
+    // If there is split_k, then the split_k resolve shader does the final
+    // division by L. Store the intermediate O value and per-row m and L values.
+    if (p.k_num > 1) {
+        // note: O and Q have swapped coord 1,2.
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
+
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            if (tile_row(r) < N) {
+                [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
+                    const uint d = d0 + col_tid;
+                    if (d >= HSV/4) break;
+                    const uint d_local = d0 / threads_per_rowgroup;
+                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
+                        perElemOpGqaStore(tile_row(r), 4 * d + comp, float(Of[r][d_local][comp]), o_offset, iq2, N);
+                    }
+                }
+            }
+        }
+
+        o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (tile_row(r) < N) {
                 perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
@@ -403,8 +512,9 @@ void main() {
             if (sink > Mf[r]) {
                 ms = exp(Mf[r] - sink);
 
-                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-                    Of[r][d] *= ACC_TYPE(ms);
+                [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
+                    const uint d_local = d0 / threads_per_rowgroup;
+                    Of[r][d_local] *= ACC_TYPE(ms);
                 }
             } else {
                 vs = exp(sink - Mf[r]);
@@ -419,23 +529,27 @@ void main() {
         Lfrcp[r] = (Lf[r] == 0.0) ? 0.0 : (1.0 / Lf[r]);
     }
 
-    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+    [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
+        const uint d_local = d0 / threads_per_rowgroup;
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-            Of[r][d] *= ACC_TYPE(Lfrcp[r]);
+            Of[r][d_local] *= ACC_TYPE(Lfrcp[r]);
 #if defined(ACC_TYPE_MAX)
-            Of[r][d] = clamp(Of[r][d], -ACC_TYPE_MAX, ACC_TYPE_MAX);
+            Of[r][d_local] = clamp(Of[r][d_local], -ACC_TYPE_MAX, ACC_TYPE_MAX);
 #endif
         }
     }
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
+    uint32_t o_offset = gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV;
 
     if (p.gqa_ratio > 1) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (tile_row(r) < N) {
-                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
+                    const uint d = d0 + col_tid;
+                    if (d >= HSV / 4) break;
+                    const uint d_local = d0 / threads_per_rowgroup;
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N);
+                        perElemOpGqaStore(tile_row(r), 4 * d + comp, float(Of[r][d_local][comp]), o_offset, iq2, N);
                     }
                 }
             }
@@ -443,9 +557,12 @@ void main() {
     } else {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (i * Br + tile_row(r) < N) {
-                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
+                    const uint d = d0 + col_tid;
+                    if (d >= HSV / 4) break;
+                    const uint d_local = d0 / threads_per_rowgroup;
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        data_o[o_offset + iq2 * HSV + (i * Br + tile_row(r)) * p.ne1 * HSV + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
+                        data_o[o_offset + iq2 * HSV + (i * Br + tile_row(r)) * p.ne1 * HSV + 4 * d + comp] = D_TYPE(Of[r][d_local][comp]);
                     }
                 }
             }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
index 9a71996383..54f1b0b622 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
@@ -55,7 +55,7 @@ ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE ele
     return max(elem0, elem1);
 }
 
-#if defined(BLOCK_SIZE)
+#if BLOCK_SIZE > 1
 #define DECODEFUNC , DEQUANTFUNC
 #else
 #define DECODEFUNC
@@ -85,7 +85,7 @@ void main() {
 
     tensorViewNV<2, false, 1, 0> tensorViewTranspose = createTensorViewNV(2, false, 1, 0);
 
-#if defined(BLOCK_SIZE)
+#if BLOCK_SIZE > 1
     tensorLayoutK = setTensorLayoutBlockSizeNV(tensorLayoutK, 1, BLOCK_SIZE);
     tensorLayoutV = setTensorLayoutBlockSizeNV(tensorLayoutV, 1, BLOCK_SIZE);
 #endif
@@ -98,7 +98,7 @@ void main() {
     if (Clamp != gl_CooperativeMatrixClampModeConstantNV)
     {
         q_stride &= ~7;
-#if !defined(BLOCK_SIZE)
+#if BLOCK_SIZE == 1
         k_stride &= ~7;
         v_stride &= ~7;
 #endif
@@ -111,7 +111,7 @@ void main() {
     coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseAccumulator> Q;
     coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseA> Qf16;
 
-    uint32_t q_offset = iq2*p.nb02+iq3*p.nb03;
+    uint32_t q_offset = gqa_iq1*p.nb01*4/*sizeof(float)*/ + iq2*p.nb02+iq3*p.nb03;
     coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, HSK_pad));
 
     Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK_pad, gl_MatrixUseA>(Q);
@@ -138,9 +138,9 @@ void main() {
         coopMatPerElementNV(slopeMat, slopeMat, perElemOpComputeSlope, iq2);
     }
 
-    uint32_t m_offset = 0;
+    uint32_t m_offset = gqa_iq1*KV * 2 /*sizeof(float16_t)*/;
     if (p.nem2 != 1 || p.nem3 != 1) {
-        m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV * 2 /*sizeof(float16_t)*/;
+        m_offset += ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV * 2 /*sizeof(float16_t)*/;
     }
 
     [[dont_unroll]]
@@ -272,10 +272,11 @@ void main() {
     if (p.k_num > 1) {
         coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
 
-        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
+        // note: O and Q have swapped coord 1,2.
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
         coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
 
-        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
+        o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
         coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N);
         coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N);
         return;
@@ -325,7 +326,7 @@ void main() {
     [[unroll]] for (uint i = 0; i < O.length(); ++i) { O[i] = clamp(O[i], -ACC_TYPE_MAX, ACC_TYPE_MAX); }
 #endif
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
+    uint32_t o_offset = gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV;
 
     coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
     if (p.gqa_ratio > 1) {
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp
index 4eaddd31a8..68917fc0bb 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp
@@ -12,7 +12,8 @@ layout (binding = 2) writeonly buffer D {float data_d[];};
 
 layout (push_constant) uniform parameter {
     uint D;
-    uint N;
+    uint ne1;
+    uint ne2;
     uint ne3;
     uint k_num;
     uint sinks;
@@ -24,15 +25,15 @@ void main() {
     // Each workgroup handles a row
     const uint n = gl_WorkGroupID.x;
     const uint tid = gl_LocalInvocationID.x;
-    const uint iq3 = gl_WorkGroupID.z;
+    const uint i2 = gl_WorkGroupID.z % p.ne2;
+    const uint i3 = gl_WorkGroupID.z / p.ne2;
 
     uint D = p.D;
-    uint N = p.N;
     uint k_num = p.k_num;
 
-    uint l_offset = D * N * p.ne3 * k_num + N * iq3 * k_num * 2 + n;
-    uint m_offset = D * N * p.ne3 * k_num + N * iq3 * k_num * 2 + N + n;
-    uint lm_stride = N * 2;
+    uint l_offset = D * p.ne1 * p.ne2 * p.ne3 * k_num + p.ne1 * 2 * (0/*split_k_index*/ + p.k_num * (i2 + p.ne2 * i3)) + n;
+    uint m_offset = D * p.ne1 * p.ne2 * p.ne3 * k_num + p.ne1 * 2 * (0/*split_k_index*/ + p.k_num * (i2 + p.ne2 * i3)) + p.ne1 + n;
+    uint lm_stride = p.ne1 * 2;
 
     // Compute the max m value for the row
     float m_max = -1.0/0.0;
@@ -99,7 +100,7 @@ void main() {
     if (d < D) {
         float O = 0.0;
         [[unroll]] for (uint k = 0; k < k_num; ++k) {
-            uint o_offset = D * N * (k + iq3 * k_num) + D * n + d;
+            uint o_offset = D * p.ne1 * (k + p.k_num * (i2 + p.ne2 * i3)) + D * n + d;
             float m = data_a[m_offset + k * lm_stride];
             O += exp(m - m_max) * data_a[o_offset];
         }
@@ -115,6 +116,6 @@ void main() {
         const float FLT_MAX = uintBitsToFloat(0x7F7FFFFF);
         O = clamp(O, -FLT_MAX, FLT_MAX);
 
-        data_d[iq3 * D * N + D * n + d] = O;
+        data_d[(i3 * p.ne2 + i2) * p.ne1 * D + D * n + d] = O;
     }
 }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl
index dfb7865936..4f2c700306 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl
@@ -29,6 +29,8 @@ layout (push_constant) uniform parameter
 #ifdef MUL_MAT_ID
     uint nei0;
     uint ne11;
+    uint expert_i1;
+    uint nbi1;
 #else
     uint ne02;
     uint ne12;
@@ -43,7 +45,7 @@ uint expert_id;
 
 void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
 #ifdef MUL_MAT_ID
-    const uint expert_idx = gl_GlobalInvocationID.y;
+    const uint expert_i0 = gl_GlobalInvocationID.y;
 #else
     const uint batch_idx = gl_GlobalInvocationID.y;
 #endif
@@ -60,7 +62,7 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
         batch_idx_a = i03 * p.ne02 + i02;
     }
 #else
-    expert_id = data_ids[expert_idx];
+    expert_id = data_ids[expert_i0 + p.expert_i1 * p.nbi1];
 #endif
 
     a_offset =
@@ -71,13 +73,13 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
 #endif
     b_offset =
 #ifdef MUL_MAT_ID
-            (expert_idx % p.ne11) * p.stride_b;
+            (expert_i0 % p.ne11) * p.stride_b + p.expert_i1 * p.batch_stride_b;
 #else
             batch_idx * p.batch_stride_b;
 #endif
     d_offset =
 #ifdef MUL_MAT_ID
-            expert_idx * p.stride_d;
+            expert_i0 * p.stride_d + p.expert_i1 * p.batch_stride_d;
 #else
             batch_idx * p.batch_stride_d;
 #endif
@@ -103,12 +105,12 @@ void reduce_result(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t
                     temp[j][n] += FLOAT_TYPE(data_fuse0[expert_id*p.stride_d + first_row + n]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE0) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    temp[j][n] *= FLOAT_TYPE(data_fuse0[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse0[expert_i0]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE1) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    temp[j][n] *= FLOAT_TYPE(data_fuse1[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse1[expert_i0]);
                 }
 #else
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
@@ -158,12 +160,12 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
                     temp[j][n] += FLOAT_TYPE(data_fuse0[expert_id*p.stride_d + first_row + n]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE0) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    temp[j][n] *= FLOAT_TYPE(data_fuse0[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse0[expert_i0]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE1) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    temp[j][n] *= FLOAT_TYPE(data_fuse1[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    temp[j][n] *= FLOAT_TYPE(data_fuse1[expert_i0]);
                 }
 #else
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
@@ -203,12 +205,12 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
                     tmpsh[j][n][0] += FLOAT_TYPE(data_fuse0[expert_id*p.stride_d + first_row + n]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE0) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    tmpsh[j][n][0] *= FLOAT_TYPE(data_fuse0[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    tmpsh[j][n][0] *= FLOAT_TYPE(data_fuse0[expert_i0]);
                 }
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_SCALE1) != 0) {
-                    const uint expert_idx = gl_GlobalInvocationID.y;
-                    tmpsh[j][n][0] *= FLOAT_TYPE(data_fuse1[expert_idx]);
+                    const uint expert_i0 = gl_GlobalInvocationID.y;
+                    tmpsh[j][n][0] *= FLOAT_TYPE(data_fuse1[expert_i0]);
                 }
 #else
                 if ((p.fusion_flags & MAT_VEC_FUSION_FLAGS_BIAS0) != 0) {
diff --git a/ggml/src/ggml-webgpu/ggml-webgpu.cpp b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
index 1470378af0..22e2bfeb4c 100644
--- a/ggml/src/ggml-webgpu/ggml-webgpu.cpp
+++ b/ggml/src/ggml-webgpu/ggml-webgpu.cpp
@@ -47,7 +47,6 @@
         double cpu_total_time_##id =                                                                      \
             std::chrono::duration<double, std::milli>(cpu_total_end_##id - cpu_total_start_##id).count(); \
         (ctx)->cpu_time_ms[#id] += cpu_total_time_##id;
-
 // fine-grained timing (not included in totals)
 #    define WEBGPU_CPU_PROFILE_DETAIL_START(id) auto cpu_detail_start_##id = std::chrono::high_resolution_clock::now();
 
@@ -74,13 +73,13 @@
 #define WEBGPU_MAX_WG_SIZE 288
 
 #define WEBGPU_MUL_MAT_WG_SIZE               256
-#define WEBGPU_NUM_PARAM_BUFS                32u
+#define WEBGPU_NUM_PARAM_BUFS                16u
 #define WEBGPU_COMMAND_SUBMIT_BATCH_SIZE     8u
 #define WEBGPU_WAIT_ANY_TIMEOUT_MS           0
 // Maximum number of in-flight submissions per-thread, to avoid exhausting the parameter buffer pool
 #define WEBGPU_MAX_INFLIGHT_SUBS_PER_THREAD  WEBGPU_NUM_PARAM_BUFS / WEBGPU_COMMAND_SUBMIT_BATCH_SIZE
 #define WEBGPU_PARAMS_BUF_SIZE_BYTES         128  // enough for 32 parameters
-#define WEBGPU_NUM_SET_ROWS_ERROR_BUFS       32
+#define WEBGPU_NUM_SET_ROWS_ERROR_BUFS       16
 #define WEBGPU_SET_ROWS_ERROR_BUF_SIZE_BYTES 4
 #define WEBGPU_STORAGE_BUF_BINDING_MULT      4  // a storage buffer binding size must be a multiple of 4
 
@@ -267,30 +266,67 @@ struct webgpu_command {
 #endif
 };
 
-// All the base objects needed to run operations on a WebGPU device
-struct webgpu_context_struct {
+struct webgpu_capabilities_base {
+    wgpu::Limits limits;
+    bool         supports_subgroup_matrix = false;
+
+    uint32_t sg_mat_m = 0;
+    uint32_t sg_mat_n = 0;
+    uint32_t sg_mat_k = 0;
+
+    uint32_t subgroup_size     = 0;
+    uint32_t max_subgroup_size = 0;
+    size_t   memset_bytes_per_thread;
+};
+
+// Stores global webgpu members
+struct webgpu_global_context_struct {
     wgpu::Instance instance;
     wgpu::Adapter  adapter;
     wgpu::Device   device;
     wgpu::Queue    queue;
-    wgpu::Limits   limits;
 
-    uint32_t max_subgroup_size;
+    webgpu_capabilities_base capabilities;
+    // Shared buffer to move data from device to host
+    wgpu::Buffer             get_tensor_staging_buf;
+    // Global mutex for pipeline and staging buffer, will be refactored to exclude pipeline caches.
+    std::recursive_mutex     mutex;
 
-    bool     supports_subgroup_matrix = false;
-    uint32_t sg_mat_m;
-    uint32_t sg_mat_n;
-    uint32_t sg_mat_k;
+    webgpu_buf_pool                memset_buf_pool;
+    std::map<int, webgpu_pipeline> memset_pipelines;  // variant or type index
+    std::atomic_uint               inflight_threads = 0;
 
-    std::recursive_mutex mutex;
-    std::atomic_uint     inflight_threads = 0;
+#ifdef GGML_WEBGPU_CPU_PROFILE
+    // Profiling: labeled CPU time in ms (total)
+    std::unordered_map<std::string, double> cpu_time_ms;
+    // Profiling: detailed CPU time in ms
+    std::unordered_map<std::string, double> cpu_detail_ms;
+#endif
 
-    webgpu_buf_pool param_buf_pool;
-    webgpu_buf_pool set_rows_error_buf_pool;
+#ifdef GGML_WEBGPU_GPU_PROFILE
+    // Profiling: per-shader GPU time in ms
+    std::unordered_map<std::string, double> shader_gpu_time_ms;
+    // Profiling: pool of timestamp query buffers (one per operation)
+    webgpu_gpu_profile_buf_pool             timestamp_query_buf_pool;
+#endif
+
+#ifdef GGML_WEBGPU_DEBUG
+    wgpu::Buffer debug_host_buf;
+    wgpu::Buffer debug_dev_buf;
+#endif
+};
+
+typedef std::shared_ptr<webgpu_global_context_struct> webgpu_global_context;
+
+// All the base objects needed to run operations on a WebGPU device
+struct webgpu_context_struct {
+    // Points to global instances owned by ggml_backend_webgpu_reg_context
+    webgpu_global_context global_ctx;
 
     pre_wgsl::Preprocessor p;
 
-    std::map<int, webgpu_pipeline> memset_pipelines;                                 // variant or type index
+    webgpu_buf_pool param_buf_pool;
+    webgpu_buf_pool set_rows_error_buf_pool;
 
     std::map<int, std::map<int, std::map<int, webgpu_pipeline>>> mul_mat_pipelines;  // src0_type, src1_type, vectorized
     std::map<int, std::map<int, std::map<int, webgpu_pipeline>>>
@@ -326,57 +362,42 @@ struct webgpu_context_struct {
 
     size_t memset_bytes_per_thread;
 
-    // Staging buffer for reading data from the GPU
-    wgpu::Buffer get_tensor_staging_buf;
-
-#ifdef GGML_WEBGPU_DEBUG
-    wgpu::Buffer debug_host_buf;
-    wgpu::Buffer debug_dev_buf;
-#endif
-
-#ifdef GGML_WEBGPU_CPU_PROFILE
-    // Profiling: labeled CPU time in ms (total)
-    std::unordered_map<std::string, double> cpu_time_ms;
-    // Profiling: detailed CPU time in ms
-    std::unordered_map<std::string, double> cpu_detail_ms;
-#endif
-
-#ifdef GGML_WEBGPU_GPU_PROFILE
-    // Profiling: per-shader GPU time in ms
-    std::unordered_map<std::string, double> shader_gpu_time_ms;
-    // Profiling: pool of timestamp query buffers (one per operation)
-    webgpu_gpu_profile_buf_pool             timestamp_query_buf_pool;
-#endif
 };
 
 typedef std::shared_ptr<webgpu_context_struct> webgpu_context;
 
+// Metadata required for the ggml backend registration/discovery interface
 struct ggml_backend_webgpu_reg_context {
-    webgpu_context webgpu_ctx;
-    size_t         device_count;
-    const char *   name;
+    // Since the Instance is a global entrypoint into the WebGPU API, it lives here
+    webgpu_global_context webgpu_global_ctx;
+    size_t                device_count;
+    const char *          name;
 };
 
+// Per-device struct for the global logical device interface
 struct ggml_backend_webgpu_device_context {
-    webgpu_context webgpu_ctx;
-    std::string    device_name;
-    std::string    device_desc;
+    webgpu_global_context webgpu_global_ctx;
+    std::string           device_name;
+    std::string           device_desc;
 };
 
+// Per-thread data required to actually run WebGPU operations in a backend instance
 struct ggml_backend_webgpu_context {
-    webgpu_context webgpu_ctx;
-    std::string    name;
+    webgpu_context        webgpu_ctx;
+    std::once_flag        init_once;
+    std::string           name;
 };
 
+// Per-thread data related to buffers
 struct ggml_backend_webgpu_buffer_context {
-    webgpu_context webgpu_ctx;
-    wgpu::Buffer   buffer;
-    std::string    label;
+    wgpu::Buffer          buffer;
+    std::string           label;
+    webgpu_global_context global_ctx;
 
-    ggml_backend_webgpu_buffer_context(webgpu_context ctx, wgpu::Buffer buf, std::string lbl) :
-        webgpu_ctx(std::move(ctx)),
+    ggml_backend_webgpu_buffer_context(wgpu::Buffer buf, std::string lbl, webgpu_global_context global_ctx_) :
         buffer(std::move(buf)),
-        label(std::move(lbl)) {}
+        label(std::move(lbl)),
+        global_ctx(std::move(global_ctx_)) {}
 };
 
 /* WebGPU object initializations */
@@ -444,7 +465,7 @@ static void ggml_webgpu_create_buffer(wgpu::Device &    device,
 /** WebGPU Actions */
 
 // Wait for the queue to finish processing all submitted work
-static void ggml_backend_webgpu_wait(webgpu_context &                         ctx,
+static void ggml_backend_webgpu_wait(webgpu_global_context &                  ctx,
                                      std::vector<webgpu_submission_futures> & futures,
                                      bool                                     block = true) {
     // If we have too many in-flight submissions, wait on the oldest one first. If there are many threads,
@@ -476,11 +497,11 @@ static void ggml_backend_webgpu_wait(webgpu_context &                         ct
     }
 }
 
-static void ggml_backend_webgpu_map_buffer(webgpu_context & ctx,
-                                           wgpu::Buffer &   buffer,
-                                           wgpu::MapMode    mode,
-                                           size_t           offset,
-                                           size_t           size) {
+static void ggml_backend_webgpu_map_buffer(webgpu_global_context & ctx,
+                                           wgpu::Buffer &          buffer,
+                                           wgpu::MapMode           mode,
+                                           size_t                  offset,
+                                           size_t                  size) {
     ctx->instance.WaitAny(buffer.MapAsync(mode, offset, size, wgpu::CallbackMode::AllowSpontaneous,
                                           [](wgpu::MapAsyncStatus status, wgpu::StringView message) {
                                               if (status != wgpu::MapAsyncStatus::Success) {
@@ -495,7 +516,7 @@ static void ggml_backend_webgpu_map_buffer(webgpu_context & ctx,
 // This function adds debugging information to shaders, as WebGPU does not support printing directly.
 // To use, add a bind group entry to the setup for the shader you are debugging, add the buffer and
 // debug statements in the shader, and then call this function after encoding the commands and submitting them.
-static void ggml_backend_webgpu_debug(webgpu_context & ctx) {
+static void ggml_backend_webgpu_debug(webgpu_global_context & ctx) {
     wgpu::CommandEncoder encoder = ctx->device.CreateCommandEncoder();
     encoder.CopyBufferToBuffer(ctx->debug_dev_buf, 0, ctx->debug_host_buf, 0, ctx->debug_host_buf.GetSize());
     wgpu::CommandBuffer commands = encoder.Finish();
@@ -507,7 +528,10 @@ static void ggml_backend_webgpu_debug(webgpu_context & ctx) {
 }
 #endif
 
-static webgpu_submission_futures ggml_backend_webgpu_submit(webgpu_context ctx, std::vector<webgpu_command> commands) {
+static webgpu_submission_futures ggml_backend_webgpu_submit(webgpu_global_context       ctx,
+                                                            std::vector<webgpu_command> commands,
+                                                            webgpu_buf_pool &           param_buf_pool,
+                                                            webgpu_buf_pool * set_rows_error_buf_pool = nullptr) {
     std::vector<wgpu::CommandBuffer> command_buffers;
     std::vector<webgpu_pool_bufs>    params_bufs;
     std::vector<webgpu_pool_bufs>    set_rows_error_bufs;
@@ -528,19 +552,19 @@ static webgpu_submission_futures ggml_backend_webgpu_submit(webgpu_context ctx,
 
     wgpu::Future p_f = ctx->queue.OnSubmittedWorkDone(
         wgpu::CallbackMode::AllowSpontaneous,
-        [ctx, params_bufs](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) {
+        [&param_buf_pool, params_bufs](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) {
             if (status != wgpu::QueueWorkDoneStatus::Success) {
                 GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n", std::string(message).c_str());
             }
             // Free the staged buffers
-            ctx->param_buf_pool.free_bufs(params_bufs);
+            param_buf_pool.free_bufs(params_bufs);
         });
     futures.push_back({ p_f });
 
     for (const auto & bufs : set_rows_error_bufs) {
         wgpu::Future f = bufs.host_buf.MapAsync(
             wgpu::MapMode::Read, 0, bufs.host_buf.GetSize(), wgpu::CallbackMode::AllowSpontaneous,
-            [ctx, bufs](wgpu::MapAsyncStatus status, wgpu::StringView message) {
+            [set_rows_error_buf_pool, bufs](wgpu::MapAsyncStatus status, wgpu::StringView message) {
                 if (status != wgpu::MapAsyncStatus::Success) {
                     GGML_LOG_ERROR("ggml_webgpu: Failed to map error buffer: %s\n", std::string(message).c_str());
                 } else {
@@ -549,7 +573,9 @@ static webgpu_submission_futures ggml_backend_webgpu_submit(webgpu_context ctx,
                         GGML_ABORT("ggml_webgpu: SET_ROWS index > 2^32, unsupported.");
                     }
                     // We can't unmap in here due to WebGPU reentrancy limitations.
-                    ctx->set_rows_error_buf_pool.free_bufs({ bufs });
+                    if (set_rows_error_buf_pool) {
+                        set_rows_error_buf_pool->free_bufs({ bufs });
+                    }
                 }
             });
         futures.push_back({ f });
@@ -581,7 +607,8 @@ static webgpu_submission_futures ggml_backend_webgpu_submit(webgpu_context ctx,
 }
 
 static webgpu_command ggml_backend_webgpu_build_multi(
-    webgpu_context &                                       ctx,
+    webgpu_global_context &                                ctx,
+    webgpu_buf_pool &                                      param_buf_pool,
     const std::vector<webgpu_pipeline> &                   pipelines,
     const std::vector<std::vector<uint32_t>> &             params_list,
     const std::vector<std::vector<wgpu::BindGroupEntry>> & bind_group_entries_list,
@@ -595,7 +622,7 @@ static webgpu_command ggml_backend_webgpu_build_multi(
     std::vector<wgpu::BindGroup>  bind_groups;
 
     for (size_t i = 0; i < pipelines.size(); i++) {
-        webgpu_pool_bufs params_bufs = ctx->param_buf_pool.alloc_bufs();
+        webgpu_pool_bufs params_bufs = param_buf_pool.alloc_bufs();
 
         ggml_backend_webgpu_map_buffer(ctx, params_bufs.host_buf, wgpu::MapMode::Write, 0,
                                        params_bufs.host_buf.GetSize());
@@ -672,34 +699,37 @@ static webgpu_command ggml_backend_webgpu_build_multi(
     return result;
 }
 
-static webgpu_command ggml_backend_webgpu_build(webgpu_context &                  ctx,
+static webgpu_command ggml_backend_webgpu_build(webgpu_global_context &           ctx,
+                                                webgpu_buf_pool &                 param_buf_pool,
                                                 webgpu_pipeline &                 pipeline,
                                                 std::vector<uint32_t>             params,
                                                 std::vector<wgpu::BindGroupEntry> bind_group_entries,
                                                 uint32_t                          wg_x,
                                                 uint32_t                          wg_y                = 1,
                                                 std::optional<webgpu_pool_bufs>   set_rows_error_bufs = std::nullopt) {
-    return ggml_backend_webgpu_build_multi(ctx,
+    return ggml_backend_webgpu_build_multi(ctx, param_buf_pool,
                                            {
                                                pipeline
     },
                                            { params }, { bind_group_entries }, { { wg_x, wg_y } }, set_rows_error_bufs);
 }
 
-static void ggml_backend_webgpu_buffer_memset(webgpu_context & ctx,
-                                              wgpu::Buffer &   buf,
-                                              uint32_t         value,
-                                              size_t           offset,
-                                              size_t           size) {
+static void ggml_backend_webgpu_buffer_memset(webgpu_global_context & ctx,
+                                              wgpu::Buffer &          buf,
+                                              uint32_t                value,
+                                              size_t                  offset,
+                                              size_t                  size) {
     std::vector<uint32_t>             params  = { (uint32_t) offset, (uint32_t) size, value };
     std::vector<wgpu::BindGroupEntry> entries = {
         { .binding = 0, .buffer = buf, .offset = 0, .size = buf.GetSize() }
     };
-    size_t   bytes_per_wg = WEBGPU_MAX_WG_SIZE * ctx->memset_bytes_per_thread;
+    size_t   bytes_per_wg = WEBGPU_MAX_WG_SIZE * ctx->capabilities.memset_bytes_per_thread;
     uint32_t wg_x         = CEIL_DIV(size + 3, bytes_per_wg);
 
-    webgpu_command command = ggml_backend_webgpu_build(ctx, ctx->memset_pipelines[0], params, entries, wg_x);
-    std::vector<webgpu_submission_futures> futures = { ggml_backend_webgpu_submit(ctx, { command }) };
+    webgpu_command command =
+        ggml_backend_webgpu_build(ctx, ctx->memset_buf_pool, ctx->memset_pipelines[0], params, entries, wg_x);
+    std::vector<webgpu_submission_futures> futures = { ggml_backend_webgpu_submit(ctx, { command },
+                                                                                  ctx->memset_buf_pool) };
     ggml_backend_webgpu_wait(ctx, futures);
 }
 
@@ -720,19 +750,19 @@ static void ggml_backend_webgpu_free(ggml_backend_t backend) {
 #ifdef GGML_WEBGPU_CPU_PROFILE
     std::cout << "\n[ggml_webgpu cpu profiling summary]\n";
     double total_cpu = 0.0;
-    for (const auto & kv : ctx->webgpu_ctx->cpu_time_ms) {
+    for (const auto & kv : ctx->webgpu_ctx->global_ctx->cpu_time_ms) {
         total_cpu += kv.second;
     }
     std::cout << "ggml_webgpu: total cpu time: " << total_cpu << " ms\n";
     std::cout << "ggml_webgpu: cpu breakdown:\n";
-    for (const auto & kv : ctx->webgpu_ctx->cpu_time_ms) {
+    for (const auto & kv : ctx->webgpu_ctx->global_ctx->cpu_time_ms) {
         double pct = (total_cpu > 0.0) ? (kv.second / total_cpu * 100.0) : 0.0;
         std::cout << "ggml_webgpu:  " << kv.first << ": " << kv.second << " ms (" << pct << "%)\n";
     }
-    if (ctx->webgpu_ctx->cpu_detail_ms.size() > 0) {
+    if (ctx->webgpu_ctx->global_ctx->cpu_detail_ms.size() > 0) {
         std::cout << "ggml_webgpu: cpu detailed breakdown:\n";
     }
-    for (const auto & kv : ctx->webgpu_ctx->cpu_detail_ms) {
+    for (const auto & kv : ctx->webgpu_ctx->global_ctx->cpu_detail_ms) {
         double pct = (total_cpu > 0.0) ? (kv.second / total_cpu * 100.0) : 0.0;
         std::cout << "ggml_webgpu:  " << kv.first << ": " << kv.second << " ms (" << pct << "%)\n";
     }
@@ -741,12 +771,12 @@ static void ggml_backend_webgpu_free(ggml_backend_t backend) {
 #ifdef GGML_WEBGPU_GPU_PROFILE
     std::cout << "\n[ggml_webgpu gpu profiling summary]\n";
     double total_gpu = 0.0;
-    for (const auto & kv : ctx->webgpu_ctx->shader_gpu_time_ms) {
+    for (const auto & kv : ctx->webgpu_ctx->global_ctx->shader_gpu_time_ms) {
         total_gpu += kv.second;
     }
     std::cout << "ggml_webgpu: total gpu time (all shaders): " << total_gpu << " ms\n";
     std::cout << "\nggml_webgpu: gpu breakdown:\n";
-    for (const auto & kv : ctx->webgpu_ctx->shader_gpu_time_ms) {
+    for (const auto & kv : ctx->webgpu_ctx->global_ctx->shader_gpu_time_ms) {
         double pct = (total_gpu > 0.0) ? (kv.second / total_gpu * 100.0) : 0.0;
         std::cout << "ggml_webgpu:  " << kv.first << ": " << kv.second << " ms (" << pct << "%)\n";
     }
@@ -772,12 +802,12 @@ static wgpu::Buffer ggml_webgpu_tensor_buf(const ggml_tensor * tensor) {
 
 static size_t ggml_webgpu_tensor_misalignment(webgpu_context & ctx, const ggml_tensor * t) {
     size_t offset = ggml_webgpu_tensor_offset(t);
-    return offset & (ctx->limits.minStorageBufferOffsetAlignment - 1);
+    return offset & (ctx->global_ctx->capabilities.limits.minStorageBufferOffsetAlignment - 1);
 }
 
 static size_t ggml_webgpu_tensor_align_offset(webgpu_context & ctx, const ggml_tensor * t) {
     size_t offset = ggml_webgpu_tensor_offset(t);
-    return offset & ~(ctx->limits.minStorageBufferOffsetAlignment - 1);
+    return offset & ~(ctx->global_ctx->capabilities.limits.minStorageBufferOffsetAlignment - 1);
 }
 
 static size_t ggml_webgpu_tensor_binding_size(webgpu_context & ctx, ggml_tensor * t) {
@@ -818,28 +848,30 @@ static webgpu_command ggml_webgpu_cpy(webgpu_context & ctx, ggml_tensor * src, g
     };
 
     uint32_t wg_x = CEIL_DIV(ne, WEBGPU_MAX_WG_SIZE);
-    return ggml_backend_webgpu_build(ctx, ctx->cpy_pipelines[src->type][dst->type], params, entries, wg_x);
+    return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, ctx->cpy_pipelines[src->type][dst->type],
+                                     params, entries, wg_x);
 }
 
 static webgpu_command ggml_webgpu_pad(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) {
     const bool circular = ggml_get_op_params_i32(dst, 8) != 0;
 
     ggml_webgpu_pad_pipeline_key       pipeline_key   = { .circular = circular };
-    ggml_webgpu_pad_shader_lib_context shader_lib_ctx = { .key = pipeline_key,
-                                                          .max_wg_size =
-                                                              ctx->limits.maxComputeInvocationsPerWorkgroup };
+    ggml_webgpu_pad_shader_lib_context shader_lib_ctx = {
+        .key = pipeline_key, .max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup
+    };
 
     webgpu_pipeline pipeline;
     {
         // TODO: remove guard once pipeline caches are per-thread
-        std::lock_guard<std::recursive_mutex> lock(ctx->mutex);
+        std::lock_guard<std::recursive_mutex> lock(ctx->global_ctx->mutex);
         auto                                  it = ctx->pad_pipelines.find(pipeline_key);
         if (it != ctx->pad_pipelines.end()) {
             pipeline = it->second;
         } else {
             ggml_webgpu_processed_shader processed =
                 ggml_webgpu_preprocess_pad_shader(ctx->p, wgsl_pad, shader_lib_ctx);
-            pipeline = ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
+            pipeline =
+                ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
             pipeline.context = processed.decisions;
             ctx->pad_pipelines.emplace(pipeline_key, pipeline);
         }
@@ -891,7 +923,7 @@ static webgpu_command ggml_webgpu_pad(webgpu_context & ctx, ggml_tensor * src, g
     };
 
     uint32_t wg_x = CEIL_DIV(ne, decisions.wg_size);
-    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
+    return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
 }
 
 static std::optional<webgpu_command> ggml_webgpu_set_rows(webgpu_context & ctx,
@@ -907,21 +939,22 @@ static std::optional<webgpu_command> ggml_webgpu_set_rows(webgpu_context & ctx,
                                               .vec4     = src->ne[0] % 4 == 0,
                                               .i64_idx  = idx->type == GGML_TYPE_I64 };
 
-    ggml_webgpu_set_rows_shader_lib_context shader_lib_ctx = { .key = key,
-                                                               .max_wg_size =
-                                                                   ctx->limits.maxComputeInvocationsPerWorkgroup };
+    ggml_webgpu_set_rows_shader_lib_context shader_lib_ctx = {
+        .key = key, .max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup
+    };
 
     webgpu_pipeline pipeline;
     // TODO: remove guard once pipeline caches are per-thread
     {
-        std::lock_guard<std::recursive_mutex> lock(ctx->mutex);
+        std::lock_guard<std::recursive_mutex> lock(ctx->global_ctx->mutex);
         auto                                  it = ctx->set_rows_pipelines.find(key);
         if (it != ctx->set_rows_pipelines.end()) {
             pipeline = it->second;
         } else {
             ggml_webgpu_processed_shader processed =
                 ggml_webgpu_preprocess_set_rows_shader(ctx->p, wgsl_set_rows, shader_lib_ctx);
-            pipeline = ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
+            pipeline =
+                ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
             pipeline.context = processed.decisions;
             ctx->set_rows_pipelines.emplace(key, pipeline);
         }
@@ -981,7 +1014,8 @@ static std::optional<webgpu_command> ggml_webgpu_set_rows(webgpu_context & ctx,
         threads = src->ne[0] * src->ne[1] * src->ne[2] * src->ne[3];
     }
     uint32_t wg_x = CEIL_DIV(threads, decisions.wg_size);
-    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x, 1, error_bufs);
+    return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x, 1,
+                                     error_bufs);
 }
 
 static webgpu_command ggml_webgpu_get_rows(webgpu_context & ctx,
@@ -1023,7 +1057,7 @@ static webgpu_command ggml_webgpu_get_rows(webgpu_context & ctx,
 
     uint32_t        vectorized = src->type == GGML_TYPE_F32 && dst->ne[0] % 4 == 0;
     webgpu_pipeline pipeline   = ctx->get_rows_pipelines[src->type][vectorized];
-    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
+    return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
 }
 
 static webgpu_command ggml_webgpu_mul_mat(webgpu_context & ctx,
@@ -1098,19 +1132,21 @@ static webgpu_command ggml_webgpu_mul_mat(webgpu_context & ctx,
             uint32_t batches       = dst->ne[2] * dst->ne[3];
             uint32_t output_groups = CEIL_DIV(dst->ne[0], WEBGPU_MUL_MAT_VEC_OUTPUTS_PER_WG);
             uint32_t total_wg      = output_groups * batches;
-            wg_x                   = total_wg % ctx->limits.maxComputeWorkgroupsPerDimension;
-            wg_y                   = CEIL_DIV(total_wg, ctx->limits.maxComputeWorkgroupsPerDimension);
+            wg_x                   = total_wg % ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension;
+            wg_y = CEIL_DIV(total_wg, ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension);
         } else {
             pipeline = ctx->mul_mat_pipelines[src0->type][src1->type][vectorized];
             uint32_t wg_m;
             uint32_t wg_n;
 #ifndef __EMSCRIPTEN__
-            if (ctx->supports_subgroup_matrix) {
+            if (ctx->global_ctx->capabilities.supports_subgroup_matrix) {
                 // The total number of subgroups/workgroups needed per matrix.
-                uint32_t wg_m_sg_tile = WEBGPU_MUL_MAT_SUBGROUP_M * WEBGPU_MUL_MAT_SUBGROUP_MATRIX_M * ctx->sg_mat_m;
+                uint32_t wg_m_sg_tile = WEBGPU_MUL_MAT_SUBGROUP_M * WEBGPU_MUL_MAT_SUBGROUP_MATRIX_M *
+                                        ctx->global_ctx->capabilities.sg_mat_m;
                 wg_m                  = CEIL_DIV(dst->ne[0], wg_m_sg_tile);
-                uint32_t wg_n_sg_tile = WEBGPU_MUL_MAT_SUBGROUP_N * WEBGPU_MUL_MAT_SUBGROUP_MATRIX_N * ctx->sg_mat_n;
-                wg_n                  = CEIL_DIV(dst->ne[1], wg_n_sg_tile);
+                uint32_t wg_n_sg_tile = WEBGPU_MUL_MAT_SUBGROUP_N * WEBGPU_MUL_MAT_SUBGROUP_MATRIX_N *
+                                        ctx->global_ctx->capabilities.sg_mat_n;
+                wg_n = CEIL_DIV(dst->ne[1], wg_n_sg_tile);
             } else {
 #endif
                 uint32_t tile_m_s = WEBGPU_MUL_MAT_TILE_M * WEBGPU_MUL_MAT_WG_SIZE_M;
@@ -1124,9 +1160,10 @@ static webgpu_command ggml_webgpu_mul_mat(webgpu_context & ctx,
             wg_x = wg_m * wg_n * dst->ne[2] * dst->ne[3];
         }
     }
-    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x, wg_y);
+    return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x, wg_y);
 }
 
+#ifndef __EMSCRIPTEN__
 static webgpu_command ggml_webgpu_flash_attn(webgpu_context & ctx,
                                              ggml_tensor *    Q,
                                              ggml_tensor *    K,
@@ -1210,8 +1247,8 @@ static webgpu_command ggml_webgpu_flash_attn(webgpu_context & ctx,
                         .offset  = ggml_webgpu_tensor_align_offset(ctx, dst),
                         .size    = ggml_webgpu_tensor_binding_size(ctx, dst) });
 
-    bool kv_direct =
-        (K->type == GGML_TYPE_F16) && (Q->ne[0] % ctx->sg_mat_k == 0) && (K->ne[1] % GGML_WEBGPU_KV_SEQ_PAD == 0);
+    bool kv_direct = (K->type == GGML_TYPE_F16) && (Q->ne[0] % ctx->global_ctx->capabilities.sg_mat_k == 0) &&
+                     (K->ne[1] % GGML_WEBGPU_KV_SEQ_PAD == 0);
 
     ggml_webgpu_flash_attn_pipeline_key key = {
         .kv_type            = K->type,
@@ -1223,25 +1260,27 @@ static webgpu_command ggml_webgpu_flash_attn(webgpu_context & ctx,
         .uses_logit_softcap = logit_softcap != 0.0f,
     };
 
-    webgpu_pipeline                         pipeline;
+    webgpu_pipeline pipeline;
     // TODO: remove guard once pipeline caches are per-thread
     {
-        std::lock_guard<std::recursive_mutex> lock(ctx->mutex);
+        std::lock_guard<std::recursive_mutex> lock(ctx->global_ctx->mutex);
         auto                                  it = ctx->flash_attn_pipelines.find(key);
         if (it != ctx->flash_attn_pipelines.end()) {
-            pipeline  = it->second;
+            pipeline = it->second;
         } else {
-            ggml_webgpu_flash_attn_shader_lib_context shader_lib_ctx = { .key      = key,
-                                                                         .sg_mat_m = ctx->sg_mat_m,
-                                                                         .sg_mat_n = ctx->sg_mat_n,
-                                                                         .sg_mat_k = ctx->sg_mat_k,
-                                                                         .wg_mem_limit_bytes =
-                                                                             ctx->limits.maxComputeWorkgroupStorageSize,
-                                                                         .max_subgroup_size = ctx->max_subgroup_size };
+            ggml_webgpu_flash_attn_shader_lib_context shader_lib_ctx = {
+                .key                = key,
+                .sg_mat_m           = ctx->global_ctx->capabilities.sg_mat_m,
+                .sg_mat_n           = ctx->global_ctx->capabilities.sg_mat_n,
+                .sg_mat_k           = ctx->global_ctx->capabilities.sg_mat_k,
+                .wg_mem_limit_bytes = ctx->global_ctx->capabilities.limits.maxComputeWorkgroupStorageSize,
+                .max_subgroup_size  = ctx->global_ctx->capabilities.max_subgroup_size
+            };
 
             ggml_webgpu_processed_shader processed =
                 ggml_webgpu_preprocess_flash_attn_shader(ctx->p, wgsl_flash_attn, shader_lib_ctx);
-            pipeline = ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
+            pipeline =
+                ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
             pipeline.context = processed.decisions;
             ctx->flash_attn_pipelines.emplace(key, pipeline);
         }
@@ -1250,11 +1289,11 @@ static webgpu_command ggml_webgpu_flash_attn(webgpu_context & ctx,
     ggml_webgpu_flash_attn_shader_decisions decisions =
         *static_cast<ggml_webgpu_flash_attn_shader_decisions *>(pipeline.context);
 
-
     uint32_t wg_per_head = CEIL_DIV(Q->ne[1], decisions.q_tile);
     uint32_t wg_x        = wg_per_head * Q->ne[2] * Q->ne[3];  // wg per head * number of heads * number of batches
-    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
+    return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
 }
+#endif
 
 static webgpu_command ggml_webgpu_unary_op(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) {
     bool is_unary = dst->op == GGML_OP_UNARY;
@@ -1264,21 +1303,22 @@ static webgpu_command ggml_webgpu_unary_op(webgpu_context & ctx, ggml_tensor * s
     ggml_webgpu_unary_pipeline_key pipeline_key = {
         .type = dst->type, .op = op, .is_unary = is_unary, .inplace = inplace
     };
-    ggml_webgpu_unary_shader_lib_context shader_lib_ctx = { .key = pipeline_key,
-                                                            .max_wg_size =
-                                                                ctx->limits.maxComputeInvocationsPerWorkgroup };
+    ggml_webgpu_unary_shader_lib_context shader_lib_ctx = {
+        .key = pipeline_key, .max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup
+    };
 
     webgpu_pipeline pipeline;
     {
         // TODO: remove guard once pipeline caches are per-thread
-        std::lock_guard<std::recursive_mutex> lock(ctx->mutex);
+        std::lock_guard<std::recursive_mutex> lock(ctx->global_ctx->mutex);
         auto                                  it = ctx->unary_pipelines.find(pipeline_key);
         if (it != ctx->unary_pipelines.end()) {
             pipeline = it->second;
         } else {
             ggml_webgpu_processed_shader processed =
                 ggml_webgpu_preprocess_unary_shader(ctx->p, wgsl_unary, shader_lib_ctx);
-            pipeline = ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
+            pipeline =
+                ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
             pipeline.context = processed.decisions;
             ctx->unary_pipelines.emplace(pipeline_key, pipeline);
         }
@@ -1346,7 +1386,7 @@ static webgpu_command ggml_webgpu_unary_op(webgpu_context & ctx, ggml_tensor * s
     }
 
     uint32_t wg_x = CEIL_DIV(ne, decisions.wg_size);
-    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
+    return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
 }
 
 static webgpu_command ggml_webgpu_binary_op(webgpu_context &  ctx,
@@ -1391,7 +1431,7 @@ static webgpu_command ggml_webgpu_binary_op(webgpu_context &  ctx,
     }
 
     uint32_t wg_x = CEIL_DIV(ggml_nelements(dst), WEBGPU_MAX_WG_SIZE);
-    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
+    return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
 }
 
 static webgpu_command ggml_webgpu_rms_norm(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) {
@@ -1426,7 +1466,8 @@ static webgpu_command ggml_webgpu_rms_norm(webgpu_context & ctx, ggml_tensor * s
                             .size    = ggml_webgpu_tensor_binding_size(ctx, dst) });
     }
 
-    return ggml_backend_webgpu_build(ctx, ctx->rms_norm_pipelines[inplace], params, entries, ggml_nrows(src));
+    return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, ctx->rms_norm_pipelines[inplace], params,
+                                     entries, ggml_nrows(src));
 }
 
 static webgpu_command ggml_webgpu_rope(webgpu_context & ctx,
@@ -1513,7 +1554,7 @@ static webgpu_command ggml_webgpu_rope(webgpu_context & ctx,
 
     webgpu_pipeline pipeline = ctx->rope_pipelines[dst->type][has_freq_factor][inplace];
     uint32_t        wg_x     = CEIL_DIV(ggml_nelements(dst), WEBGPU_MAX_WG_SIZE);
-    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
+    return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
 }
 
 static webgpu_command ggml_webgpu_glu(webgpu_context & ctx, ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst) {
@@ -1565,7 +1606,7 @@ static webgpu_command ggml_webgpu_glu(webgpu_context & ctx, ggml_tensor * src0,
 
     webgpu_pipeline pipeline = ctx->glu_pipelines[ggml_get_glu_op(dst)][dst->type][split];
     uint32_t        wg_x     = CEIL_DIV(ggml_nelements(dst), WEBGPU_MAX_WG_SIZE);
-    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
+    return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
 }
 
 static webgpu_command ggml_webgpu_scale(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) {
@@ -1602,7 +1643,8 @@ static webgpu_command ggml_webgpu_scale(webgpu_context & ctx, ggml_tensor * src,
     }
 
     uint32_t wg_x = CEIL_DIV(ggml_nelements(dst), WEBGPU_MAX_WG_SIZE);
-    return ggml_backend_webgpu_build(ctx, ctx->scale_pipelines[inplace], params, entries, wg_x);
+    return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, ctx->scale_pipelines[inplace], params,
+                                     entries, wg_x);
 }
 
 static webgpu_command ggml_webgpu_soft_max(webgpu_context & ctx,
@@ -1674,7 +1716,8 @@ static webgpu_command ggml_webgpu_soft_max(webgpu_context & ctx,
                             .size    = ggml_webgpu_tensor_binding_size(ctx, dst) });
     }
 
-    return ggml_backend_webgpu_build(ctx, ctx->soft_max_pipelines[mask_type][has_sink][inplace], params, entries,
+    return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool,
+                                     ctx->soft_max_pipelines[mask_type][has_sink][inplace], params, entries,
                                      ggml_nrows(dst));
 }
 
@@ -1696,25 +1739,26 @@ static webgpu_command ggml_webgpu_argmax(webgpu_context & ctx, ggml_tensor * src
 
     ggml_webgpu_generic_shader_lib_context shader_lib_ctx = {
         .vec4        = src->ne[0] % 4 == 0,
-        .max_wg_size = ctx->limits.maxComputeInvocationsPerWorkgroup,
+        .max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup,
     };
 
     webgpu_pipeline pipeline;
     {
         // TODO: remove guard once pipeline caches are per-thread
-        std::lock_guard<std::recursive_mutex> lock(ctx->mutex);
+        std::lock_guard<std::recursive_mutex> lock(ctx->global_ctx->mutex);
         auto                                  it = ctx->argmax_pipelines.find(shader_lib_ctx.vec4);
         if (it != ctx->argmax_pipelines.end()) {
             pipeline = it->second;
         } else {
             ggml_webgpu_processed_shader processed =
                 ggml_webgpu_preprocess_generic_shader(ctx->p, wgsl_argmax, shader_lib_ctx, "argmax");
-            pipeline = ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
+            pipeline =
+                ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
             ctx->argmax_pipelines.emplace(shader_lib_ctx.vec4, pipeline);
         }
     }
     uint32_t wg_x = ggml_nelements(dst);
-    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
+    return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
 }
 
 static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) {
@@ -1722,13 +1766,13 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr
     // ascending order is 0, descending order is 1
     const int32_t order    = is_top_k ? (int32_t) GGML_SORT_ORDER_DESC : (int32_t) ggml_get_op_params_i32(dst, 0);
 
-    ggml_webgpu_argsort_shader_lib_context shader_lib_ctx = { .max_wg_size =
-                                                                  ctx->limits.maxComputeInvocationsPerWorkgroup,
-                                                              .wg_mem_limit_bytes =
-                                                                  ctx->limits.maxComputeWorkgroupStorageSize,
-                                                              .order = order };
+    ggml_webgpu_argsort_shader_lib_context shader_lib_ctx = {
+        .max_wg_size        = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup,
+        .wg_mem_limit_bytes = ctx->global_ctx->capabilities.limits.maxComputeWorkgroupStorageSize,
+        .order              = order
+    };
 
-    std::lock_guard<std::recursive_mutex> lock(ctx->mutex);
+    std::lock_guard<std::recursive_mutex> lock(ctx->global_ctx->mutex);
     webgpu_pipeline                       argsort_pipeline;
     auto                                  it = ctx->argsort_pipelines.find(order);
     if (it != ctx->argsort_pipelines.end()) {
@@ -1736,7 +1780,8 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr
     } else {
         ggml_webgpu_processed_shader processed =
             ggml_webgpu_preprocess_argsort_shader(ctx->p, wgsl_argsort, shader_lib_ctx);
-        argsort_pipeline = ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
+        argsort_pipeline =
+            ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
         argsort_pipeline.context = processed.decisions;
         ctx->argsort_pipelines.emplace(order, argsort_pipeline);
     }
@@ -1751,7 +1796,7 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr
         ggml_webgpu_processed_shader processed =
             ggml_webgpu_preprocess_argsort_merge_shader(ctx->p, wgsl_argsort_merge, shader_lib_ctx);
         argsort_merge_pipeline =
-            ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
+            ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
         argsort_merge_pipeline.context = processed.decisions;
         ctx->argsort_merge_pipelines.emplace(order, argsort_merge_pipeline);
     }
@@ -1780,9 +1825,10 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr
 
     const bool start_in_tmp = (merge_passes % 2) == 1;
 
-    const size_t dst_offset       = ggml_webgpu_tensor_offset(dst);
-    const size_t idx_nbytes       = out_ne0 * ggml_nrows(dst) * sizeof(int32_t);
-    const size_t tmp_offset       = ROUNDUP_POW2(dst_offset + idx_nbytes, ctx->limits.minStorageBufferOffsetAlignment);
+    const size_t dst_offset = ggml_webgpu_tensor_offset(dst);
+    const size_t idx_nbytes = out_ne0 * ggml_nrows(dst) * sizeof(int32_t);
+    const size_t tmp_offset =
+        ROUNDUP_POW2(dst_offset + idx_nbytes, ctx->global_ctx->capabilities.limits.minStorageBufferOffsetAlignment);
     const size_t tmp_binding_size = ROUNDUP_POW2(idx_nbytes, WEBGPU_STORAGE_BUF_BINDING_MULT);
     const size_t dst_binding_size =
         ROUNDUP_POW2(idx_nbytes + ggml_webgpu_tensor_misalignment(ctx, dst), WEBGPU_STORAGE_BUF_BINDING_MULT);
@@ -1813,10 +1859,10 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr
     };
 
     const uint32_t                    total_wg_init = npr * nrows;
-    const uint32_t                    max_wg        = ctx->limits.maxComputeWorkgroupsPerDimension;
-    const uint32_t                    wg_x_init     = std::min(total_wg_init, max_wg);
-    const uint32_t                    wg_y_init     = CEIL_DIV(total_wg_init, wg_x_init);
-    std::vector<wgpu::BindGroupEntry> init_entries  = {
+    const uint32_t                    max_wg    = ctx->global_ctx->capabilities.limits.maxComputeWorkgroupsPerDimension;
+    const uint32_t                    wg_x_init = std::min(total_wg_init, max_wg);
+    const uint32_t                    wg_y_init = CEIL_DIV(total_wg_init, wg_x_init);
+    std::vector<wgpu::BindGroupEntry> init_entries = {
         { .binding = 0,
          .buffer  = ggml_webgpu_tensor_buf(src),
          .offset  = ggml_webgpu_tensor_align_offset(ctx, src),
@@ -1830,7 +1876,8 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr
     workgroups_list.push_back({ wg_x_init, wg_y_init });
 
     if (merge_passes == 0) {
-        return ggml_backend_webgpu_build_multi(ctx, pipelines, params_list, entries_list, workgroups_list);
+        return ggml_backend_webgpu_build_multi(ctx->global_ctx, ctx->param_buf_pool, pipelines, params_list,
+                                               entries_list, workgroups_list);
     }
 
     bool     in_is_tmp = start_in_tmp;
@@ -1891,7 +1938,8 @@ static webgpu_command ggml_webgpu_argsort(webgpu_context & ctx, ggml_tensor * sr
         in_is_tmp = !in_is_tmp;
     }
 
-    return ggml_backend_webgpu_build_multi(ctx, pipelines, params_list, entries_list, workgroups_list);
+    return ggml_backend_webgpu_build_multi(ctx->global_ctx, ctx->param_buf_pool, pipelines, params_list, entries_list,
+                                           workgroups_list);
 }
 
 static webgpu_command ggml_webgpu_cumsum(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) {
@@ -1912,24 +1960,25 @@ static webgpu_command ggml_webgpu_cumsum(webgpu_context & ctx, ggml_tensor * src
 
     ggml_webgpu_generic_shader_lib_context shader_lib_ctx = {
         .vec4        = false,
-        .max_wg_size = ctx->limits.maxComputeInvocationsPerWorkgroup,
+        .max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup,
     };
     webgpu_pipeline pipeline;
     // TODO: remove guard once pipeline caches are per-thread
     {
-        std::lock_guard<std::recursive_mutex> lock(ctx->mutex);
+        std::lock_guard<std::recursive_mutex> lock(ctx->global_ctx->mutex);
         auto                                  it = ctx->cumsum_pipelines.find(1);
         if (it != ctx->cumsum_pipelines.end()) {
             pipeline = it->second;
         } else {
             ggml_webgpu_processed_shader processed =
                 ggml_webgpu_preprocess_generic_shader(ctx->p, wgsl_cumsum, shader_lib_ctx, "cumsum");
-            pipeline = ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
+            pipeline =
+                ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
             ctx->cumsum_pipelines.emplace(1, pipeline);
         }
     }
     uint32_t wg_x = ggml_nrows(dst);
-    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
+    return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
 }
 
 static webgpu_command ggml_webgpu_sum_rows(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) {
@@ -1956,25 +2005,26 @@ static webgpu_command ggml_webgpu_sum_rows(webgpu_context & ctx, ggml_tensor * s
 
     ggml_webgpu_generic_shader_lib_context shader_lib_ctx = {
         .vec4        = false,
-        .max_wg_size = ctx->limits.maxComputeInvocationsPerWorkgroup,
+        .max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup,
     };
 
     webgpu_pipeline pipeline;
     {
         // TODO: remove guard once pipeline caches are per-thread
-        std::lock_guard<std::recursive_mutex> lock(ctx->mutex);
+        std::lock_guard<std::recursive_mutex> lock(ctx->global_ctx->mutex);
         auto                                  it = ctx->sum_rows_pipelines.find(1);
         if (it != ctx->sum_rows_pipelines.end()) {
             pipeline = it->second;
         } else {
             ggml_webgpu_processed_shader processed =
                 ggml_webgpu_preprocess_generic_shader(ctx->p, wgsl_sum_rows, shader_lib_ctx, "sum_rows");
-            pipeline = ggml_webgpu_create_pipeline(ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
+            pipeline =
+                ggml_webgpu_create_pipeline(ctx->global_ctx->device, processed.wgsl.c_str(), processed.variant.c_str());
             ctx->sum_rows_pipelines.emplace(1, pipeline);
         }
     }
     uint32_t wg_x = total_sum ? 1 : ggml_nrows(dst);
-    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x);
+    return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
 }
 
 // Returns the encoded command, or std::nullopt if the operation is a no-op
@@ -1982,6 +2032,9 @@ static std::optional<webgpu_command> ggml_webgpu_encode_node(webgpu_context ctx,
     if (ggml_is_empty(node)) {
         return std::nullopt;
     }
+    if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+        return std::nullopt;
+    }
     WEBGPU_LOG_DEBUG("ggml_webgpu_encode_node(" << node << ", " << ggml_op_name(node->op) << ")");
 
     ggml_tensor * src0 = node->src[0];
@@ -2006,7 +2059,11 @@ static std::optional<webgpu_command> ggml_webgpu_encode_node(webgpu_context ctx,
         case GGML_OP_MUL_MAT:
             return ggml_webgpu_mul_mat(ctx, src0, src1, node);
         case GGML_OP_FLASH_ATTN_EXT:
+#ifndef __EMSCRIPTEN__
             return ggml_webgpu_flash_attn(ctx, src0, src1, src2, node->src[3], node->src[4], node);
+#else
+            return std::nullopt;
+#endif
         case GGML_OP_ADD:
             {
                 int inplace = ggml_webgpu_tensor_equal(src0, node);
@@ -2067,12 +2124,12 @@ static std::optional<webgpu_command> ggml_webgpu_encode_node(webgpu_context ctx,
 static ggml_status ggml_backend_webgpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     WEBGPU_LOG_DEBUG("ggml_backend_webgpu_graph_compute(" << cgraph->n_nodes << " nodes)");
 
-    ggml_backend_webgpu_context * backend_ctx = static_cast<ggml_backend_webgpu_context *>(backend->context);
+    ggml_backend_webgpu_context * backend_ctx = (ggml_backend_webgpu_context *) backend->context;
     webgpu_context                ctx         = backend_ctx->webgpu_ctx;
 
     WEBGPU_CPU_PROFILE_TOTAL_START(graph_compute);
 
-    ctx->inflight_threads++;
+    ctx->global_ctx->inflight_threads++;
 
     std::vector<webgpu_command>            commands;
     std::vector<webgpu_submission_futures> futures;
@@ -2081,25 +2138,27 @@ static ggml_status ggml_backend_webgpu_graph_compute(ggml_backend_t backend, str
             commands.push_back(*cmd);
         }
         // compute the batch size based on the number of inflight threads
-        uint32_t inflight_threads = ctx->inflight_threads;
+        uint32_t inflight_threads = ctx->global_ctx->inflight_threads;
         uint32_t batch_size       = std::min(std::max(1u, WEBGPU_NUM_PARAM_BUFS / std::max(inflight_threads, 1u)),
                                              WEBGPU_COMMAND_SUBMIT_BATCH_SIZE);
         if (commands.size() >= batch_size) {
-            futures.push_back(ggml_backend_webgpu_submit(ctx, commands));
+            futures.push_back(ggml_backend_webgpu_submit(ctx->global_ctx, commands, ctx->param_buf_pool,
+                                                         &ctx->set_rows_error_buf_pool));
             // Process events and check for completed submissions
-            ctx->instance.ProcessEvents();
-            ggml_backend_webgpu_wait(ctx, futures, false);
+            ctx->global_ctx->instance.ProcessEvents();
+            ggml_backend_webgpu_wait(ctx->global_ctx, futures, false);
             commands.clear();
         }
     }
     if (!commands.empty()) {
-        webgpu_submission_futures new_futures = ggml_backend_webgpu_submit(ctx, commands);
+        webgpu_submission_futures new_futures =
+            ggml_backend_webgpu_submit(ctx->global_ctx, commands, ctx->param_buf_pool, &ctx->set_rows_error_buf_pool);
         futures.push_back(new_futures);
     }
 
-    ggml_backend_webgpu_wait(ctx, futures);
-    ctx->inflight_threads--;
-    WEBGPU_CPU_PROFILE_TOTAL_END(graph_compute, ctx);
+    ggml_backend_webgpu_wait(ctx->global_ctx, futures);
+    ctx->global_ctx->inflight_threads--;
+    WEBGPU_CPU_PROFILE_TOTAL_END(graph_compute, ctx->global_ctx);
     return GGML_STATUS_SUCCESS;
 }
 
@@ -2156,8 +2215,8 @@ static void ggml_backend_webgpu_buffer_memset_tensor(ggml_backend_buffer_t buffe
 
     // This is a trick to set all bytes of a u32 to the same 1 byte value.
     uint32_t val32 = (uint32_t) value * 0x01010101;
-    ggml_backend_webgpu_buffer_memset(buf_ctx->webgpu_ctx, buf_ctx->buffer, val32, total_offset, size);
-    WEBGPU_CPU_PROFILE_TOTAL_END(memset_tensor, buf_ctx->webgpu_ctx);
+    ggml_backend_webgpu_buffer_memset(buf_ctx->global_ctx, buf_ctx->buffer, val32, total_offset, size);
+    WEBGPU_CPU_PROFILE_TOTAL_END(memset_tensor, buf_ctx->global_ctx);
 }
 
 static void ggml_backend_webgpu_buffer_set_tensor(ggml_backend_buffer_t buffer,
@@ -2166,15 +2225,14 @@ static void ggml_backend_webgpu_buffer_set_tensor(ggml_backend_buffer_t buffer,
                                                   size_t                offset,
                                                   size_t                size) {
     WEBGPU_CPU_PROFILE_TOTAL_START(set_tensor);
-    ggml_backend_webgpu_buffer_context * buf_ctx    = (ggml_backend_webgpu_buffer_context *) buffer->context;
-    webgpu_context                       webgpu_ctx = buf_ctx->webgpu_ctx;
+    ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
 
     WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_set_tensor(" << buf_ctx->label << ", " << tensor << ", " << data
                                                               << ", " << offset << ", " << size << ")");
 
     size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset;
 
-    webgpu_ctx->queue.WriteBuffer(buf_ctx->buffer, total_offset, data, (size / 4) * 4);
+    buf_ctx->global_ctx->queue.WriteBuffer(buf_ctx->buffer, total_offset, data, (size / 4) * 4);
 
     if (size % 4 != 0) {
         // If size is not a multiple of 4, we need to memset the remaining bytes
@@ -2187,21 +2245,21 @@ static void ggml_backend_webgpu_buffer_set_tensor(ggml_backend_buffer_t buffer,
             ((uint8_t *) &val32)[i] = ((const uint8_t *) data)[size - remaining_size + i];
         }
         // memset the remaining bytes
-        ggml_backend_webgpu_buffer_memset(webgpu_ctx, buf_ctx->buffer, val32, total_offset + (size - remaining_size),
-                                          remaining_size);
+        ggml_backend_webgpu_buffer_memset(buf_ctx->global_ctx, buf_ctx->buffer, val32,
+                                          total_offset + (size - remaining_size), remaining_size);
     } else {
         // wait for WriteBuffer to complete
-        webgpu_ctx->instance.WaitAny(
-            webgpu_ctx->queue.OnSubmittedWorkDone(wgpu::CallbackMode::AllowSpontaneous,
+        buf_ctx->global_ctx->instance.WaitAny(buf_ctx->global_ctx->queue.OnSubmittedWorkDone(
+                                                  wgpu::CallbackMode::AllowSpontaneous,
                                                   [](wgpu::QueueWorkDoneStatus status, wgpu::StringView message) {
                                                       if (status != wgpu::QueueWorkDoneStatus::Success) {
                                                           GGML_LOG_ERROR("ggml_webgpu: Failed to submit commands: %s\n",
                                                                          std::string(message).c_str());
                                                       }
                                                   }),
-            UINT64_MAX);
+                                              UINT64_MAX);
     }
-    WEBGPU_CPU_PROFILE_TOTAL_END(set_tensor, webgpu_ctx);
+    WEBGPU_CPU_PROFILE_TOTAL_END(set_tensor, buf_ctx->global_ctx);
 }
 
 static void ggml_backend_webgpu_buffer_get_tensor(ggml_backend_buffer_t buffer,
@@ -2213,8 +2271,7 @@ static void ggml_backend_webgpu_buffer_get_tensor(ggml_backend_buffer_t buffer,
     ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
     WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_get_tensor(" << buf_ctx->label << ", " << tensor << ", " << data
                                                               << ", " << offset << ", " << size << ")");
-    webgpu_context webgpu_ctx = buf_ctx->webgpu_ctx;
-    wgpu::Device   device     = webgpu_ctx->device;
+    wgpu::Device device = buf_ctx->global_ctx->device;
 
     size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset;
 
@@ -2224,42 +2281,45 @@ static void ggml_backend_webgpu_buffer_get_tensor(ggml_backend_buffer_t buffer,
         final_size = size + (4 - (size % 4));
     }
 
-    std::lock_guard<std::recursive_mutex> lock(webgpu_ctx->mutex);
+    std::lock_guard<std::recursive_mutex> lock(buf_ctx->global_ctx->mutex);
 
-    if (webgpu_ctx->get_tensor_staging_buf == nullptr || webgpu_ctx->get_tensor_staging_buf.GetSize() < final_size) {
+    if (buf_ctx->global_ctx->get_tensor_staging_buf == nullptr ||
+        buf_ctx->global_ctx->get_tensor_staging_buf.GetSize() < final_size) {
         // Create a new staging buffer if it doesn't exist or is too small
-        if (webgpu_ctx->get_tensor_staging_buf) {
-            webgpu_ctx->get_tensor_staging_buf.Destroy();
+        if (buf_ctx->global_ctx->get_tensor_staging_buf) {
+            buf_ctx->global_ctx->get_tensor_staging_buf.Destroy();
         }
-        ggml_webgpu_create_buffer(device, webgpu_ctx->get_tensor_staging_buf, final_size,
+        ggml_webgpu_create_buffer(device, buf_ctx->global_ctx->get_tensor_staging_buf, final_size,
                                   wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead, "get_tensor_staging_buf");
     }
 
     // Copy the data from the buffer to the staging buffer
     wgpu::CommandEncoder encoder = device.CreateCommandEncoder();
-    encoder.CopyBufferToBuffer(buf_ctx->buffer, total_offset, webgpu_ctx->get_tensor_staging_buf, 0, final_size);
+    encoder.CopyBufferToBuffer(buf_ctx->buffer, total_offset, buf_ctx->global_ctx->get_tensor_staging_buf, 0,
+                               final_size);
     wgpu::CommandBuffer commands = encoder.Finish();
 
     // Submit the command buffer to the queue
-    webgpu_ctx->queue.Submit(1, &commands);
+    buf_ctx->global_ctx->queue.Submit(1, &commands);
 
     // Map the staging buffer to read the data
-    ggml_backend_webgpu_map_buffer(webgpu_ctx, webgpu_ctx->get_tensor_staging_buf, wgpu::MapMode::Read, 0, final_size);
+    ggml_backend_webgpu_map_buffer(buf_ctx->global_ctx, buf_ctx->global_ctx->get_tensor_staging_buf,
+                                   wgpu::MapMode::Read, 0, final_size);
     // Must specify size here since the staging buffer might be larger than the tensor size
-    const void * mapped_range = webgpu_ctx->get_tensor_staging_buf.GetConstMappedRange(0, final_size);
+    const void * mapped_range = buf_ctx->global_ctx->get_tensor_staging_buf.GetConstMappedRange(0, final_size);
 
     // Copy the data from the mapped range to the output buffer
     std::memcpy(data, mapped_range, size);
-    webgpu_ctx->get_tensor_staging_buf.Unmap();
-    WEBGPU_CPU_PROFILE_TOTAL_END(get_tensor, webgpu_ctx);
+    buf_ctx->global_ctx->get_tensor_staging_buf.Unmap();
+    WEBGPU_CPU_PROFILE_TOTAL_END(get_tensor, buf_ctx->global_ctx);
 }
 
 static void ggml_backend_webgpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
     WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_clear(" << buffer << ", " << (uint32_t) value << ")");
     WEBGPU_CPU_PROFILE_TOTAL_START(clear);
     ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
-    ggml_backend_webgpu_buffer_memset(buf_ctx->webgpu_ctx, buf_ctx->buffer, value, 0, buffer->size);
-    WEBGPU_CPU_PROFILE_TOTAL_END(clear, buf_ctx->webgpu_ctx);
+    ggml_backend_webgpu_buffer_memset(buf_ctx->global_ctx, buf_ctx->buffer, value, 0, buffer->size);
+    WEBGPU_CPU_PROFILE_TOTAL_END(clear, buf_ctx->global_ctx);
 }
 
 static ggml_backend_buffer_i ggml_backend_webgpu_buffer_interface = {
@@ -2289,28 +2349,30 @@ static ggml_backend_buffer_t ggml_backend_webgpu_buffer_type_alloc_buffer(ggml_b
     int                     buffer_id = buffer_count++;
     std::string             buf_name  = "tensor_buf" + std::to_string(buffer_id);
     WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_type_alloc_buffer_" << buffer_id << ": " << size << " bytes");
-    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(buft->device->context);
 
-    wgpu::Buffer buf;
-    ggml_webgpu_create_buffer(ctx->webgpu_ctx->device, buf, ROUNDUP_POW2(size, WEBGPU_STORAGE_BUF_BINDING_MULT),
+    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(buft->device->context);
+    wgpu::Buffer                         buf;
+    ggml_webgpu_create_buffer(ctx->webgpu_global_ctx->device, buf, ROUNDUP_POW2(size, WEBGPU_STORAGE_BUF_BINDING_MULT),
                               wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::CopyDst,
                               buf_name.c_str());
 
     ggml_backend_webgpu_buffer_context * buf_ctx =
-        new ggml_backend_webgpu_buffer_context(ctx->webgpu_ctx, buf, buf_name);
+        new ggml_backend_webgpu_buffer_context(buf, buf_name, ctx->webgpu_global_ctx);
 
     return ggml_backend_buffer_init(buft, ggml_backend_webgpu_buffer_interface, buf_ctx, size);
 }
 
 static size_t ggml_backend_webgpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
-    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(buft->device->context);
-    return ctx->webgpu_ctx->limits.minStorageBufferOffsetAlignment;
+    ggml_backend_webgpu_device_context * dev_ctx =
+        static_cast<ggml_backend_webgpu_device_context *>(buft->device->context);
+    return dev_ctx->webgpu_global_ctx->capabilities.limits.minStorageBufferOffsetAlignment;
 }
 
 // maxBufferSize might be larger, but you can't bind more than maxStorageBufferBindingSize to a single binding.
 static size_t ggml_backend_webgpu_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
-    ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(buft->device->context);
-    return ctx->webgpu_ctx->limits.maxStorageBufferBindingSize;
+    ggml_backend_webgpu_device_context * dev_ctx =
+        static_cast<ggml_backend_webgpu_device_context *>(buft->device->context);
+    return dev_ctx->webgpu_global_ctx->capabilities.limits.maxStorageBufferBindingSize;
 }
 
 static size_t ggml_backend_webgpu_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft,
@@ -2319,7 +2381,7 @@ static size_t ggml_backend_webgpu_buffer_type_get_alloc_size(ggml_backend_buffer
     size_t                               res = ggml_nbytes(tensor);
     switch (tensor->op) {
         case GGML_OP_ARGSORT:
-            res = ROUNDUP_POW2(res * 2 + ctx->webgpu_ctx->limits.minStorageBufferOffsetAlignment,
+            res = ROUNDUP_POW2(res * 2 + ctx->webgpu_global_ctx->capabilities.limits.minStorageBufferOffsetAlignment,
                                WEBGPU_STORAGE_BUF_BINDING_MULT);
             break;
         case GGML_OP_TOP_K:
@@ -2327,8 +2389,9 @@ static size_t ggml_backend_webgpu_buffer_type_get_alloc_size(ggml_backend_buffer
                 const ggml_tensor * src0 = tensor->src[0];
                 if (src0) {
                     const size_t full = sizeof(int32_t) * ggml_nelements(src0);
-                    res               = ROUNDUP_POW2(full * 2 + ctx->webgpu_ctx->limits.minStorageBufferOffsetAlignment,
-                                                     WEBGPU_STORAGE_BUF_BINDING_MULT);
+                    res               = ROUNDUP_POW2(
+                        full * 2 + ctx->webgpu_global_ctx->capabilities.limits.minStorageBufferOffsetAlignment,
+                        WEBGPU_STORAGE_BUF_BINDING_MULT);
                 }
             }
             break;
@@ -2356,7 +2419,7 @@ static void ggml_backend_webgpu_device_get_memory(ggml_backend_dev_t dev, size_t
     ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(dev->context);
     // TODO: for now, return maxBufferSize as both free and total memory
     // Track https://github.com/gpuweb/gpuweb/issues/5505 for updates.
-    uint64_t                             max_buffer_size = ctx->webgpu_ctx->limits.maxBufferSize;
+    uint64_t                             max_buffer_size = ctx->webgpu_global_ctx->capabilities.limits.maxBufferSize;
     // If we're on a 32-bit system, clamp to UINTPTR_MAX
 #if UINTPTR_MAX < UINT64_MAX
     uint64_t max_ptr_size = static_cast<uint64_t>(UINTPTR_MAX);
@@ -2399,66 +2462,67 @@ static std::vector<wgpu::ConstantEntry> ggml_webgpu_wg_size_entry(uint32_t wg_si
     return constants;
 }
 
-static void ggml_webgpu_init_memset_pipeline(webgpu_context & webgpu_ctx) {
+static void ggml_webgpu_init_memset_pipeline(webgpu_global_context & ctx) {
     // we use the maximum workgroup size for the memset pipeline
-    size_t max_threads                  = WEBGPU_MAX_WG_SIZE * webgpu_ctx->limits.maxComputeWorkgroupsPerDimension;
+    size_t max_threads = WEBGPU_MAX_WG_SIZE * ctx->capabilities.limits.maxComputeWorkgroupsPerDimension;
     // Size the bytes_per_thread so that the largest buffer size can be handled
-    webgpu_ctx->memset_bytes_per_thread = CEIL_DIV(webgpu_ctx->limits.maxStorageBufferBindingSize, max_threads);
+    ctx->capabilities.memset_bytes_per_thread =
+        CEIL_DIV(ctx->capabilities.limits.maxStorageBufferBindingSize, max_threads);
     std::vector<wgpu::ConstantEntry> constants(2);
-    constants[0].key                = "wg_size";
-    constants[0].value              = WEBGPU_MAX_WG_SIZE;
-    constants[1].key                = "bytes_per_thread";
-    constants[1].value              = webgpu_ctx->memset_bytes_per_thread;
-    webgpu_ctx->memset_pipelines[0] = ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_memset, "memset", constants);
+    constants[0].key         = "wg_size";
+    constants[0].value       = WEBGPU_MAX_WG_SIZE;
+    constants[1].key         = "bytes_per_thread";
+    constants[1].value       = ctx->capabilities.memset_bytes_per_thread;
+    ctx->memset_pipelines[0] = ggml_webgpu_create_pipeline(ctx->device, wgsl_memset, "memset", constants);
 }
 
 static void ggml_webgpu_init_mul_mat_pipeline(webgpu_context & webgpu_ctx) {
     // Q4/Q5/Q8 classic quantizations
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_0][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q4_0_f32, "mul_mat_q4_0_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q4_0_f32, "mul_mat_q4_0_f32");
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_1][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q4_1_f32, "mul_mat_q4_1_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q4_1_f32, "mul_mat_q4_1_f32");
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q5_0][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q5_0_f32, "mul_mat_q5_0_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q5_0_f32, "mul_mat_q5_0_f32");
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q5_1][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q5_1_f32, "mul_mat_q5_1_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q5_1_f32, "mul_mat_q5_1_f32");
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q8_0][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q8_0_f32, "mul_mat_q8_0_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q8_0_f32, "mul_mat_q8_0_f32");
 
     // K-quantizations
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q2_K][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q2_k_f32, "mul_mat_q2_k_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q2_k_f32, "mul_mat_q2_k_f32");
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q3_K][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q3_k_f32, "mul_mat_q3_k_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q3_k_f32, "mul_mat_q3_k_f32");
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_K][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q4_k_f32, "mul_mat_q4_k_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q4_k_f32, "mul_mat_q4_k_f32");
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q5_K][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q5_k_f32, "mul_mat_q5_k_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q5_k_f32, "mul_mat_q5_k_f32");
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q6_K][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_q6_k_f32, "mul_mat_q6_k_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_q6_k_f32, "mul_mat_q6_k_f32");
 
     // IQ quantizations (2-, 3-, 4-bit variants)
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ2_XXS][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq2_xxs_f32, "mul_mat_iq2_xxs_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq2_xxs_f32, "mul_mat_iq2_xxs_f32");
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ2_XS][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq2_xs_f32, "mul_mat_iq2_xs_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq2_xs_f32, "mul_mat_iq2_xs_f32");
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ2_S][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq2_s_f32, "mul_mat_iq2_s_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq2_s_f32, "mul_mat_iq2_s_f32");
 
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ3_XXS][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq3_xxs_f32, "mul_mat_iq3_xxs_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq3_xxs_f32, "mul_mat_iq3_xxs_f32");
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ3_S][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq3_s_f32, "mul_mat_iq3_s_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq3_s_f32, "mul_mat_iq3_s_f32");
 
     // 1-bit and 4-bit IQ variants
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ1_S][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq1_s_f32, "mul_mat_iq1_s_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq1_s_f32, "mul_mat_iq1_s_f32");
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ1_M][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq1_m_f32, "mul_mat_iq1_m_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq1_m_f32, "mul_mat_iq1_m_f32");
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ4_NL][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq4_nl_f32, "mul_mat_iq4_nl_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq4_nl_f32, "mul_mat_iq4_nl_f32");
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_IQ4_XS][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_mat_iq4_xs_f32, "mul_mat_iq4_xs_f32");
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_mat_iq4_xs_f32, "mul_mat_iq4_xs_f32");
 
     std::string proc_mul_mat_f32_f32;
     std::string proc_mul_mat_f32_f32_vec;
@@ -2471,18 +2535,18 @@ static void ggml_webgpu_init_mul_mat_pipeline(webgpu_context & webgpu_ctx) {
 
     std::vector<wgpu::ConstantEntry> mul_mat_constants;
 #ifndef __EMSCRIPTEN__
-    if (webgpu_ctx->supports_subgroup_matrix) {
+    if (webgpu_ctx->global_ctx->capabilities.supports_subgroup_matrix) {
         std::map<std::string, std::string> sg_matrix_repls;
-        sg_matrix_repls["WEBGPU_MAX_SUBGROUP_SIZE"] = std::to_string(webgpu_ctx->max_subgroup_size);
+        sg_matrix_repls["WEBGPU_MAX_SUBGROUP_SIZE"] =
+            std::to_string(webgpu_ctx->global_ctx->capabilities.max_subgroup_size);
         sg_matrix_repls["WEBGPU_TILE_K"]            = std::to_string(WEBGPU_MUL_MAT_TILE_K);
         sg_matrix_repls["WEBGPU_SUBGROUP_M"]        = std::to_string(WEBGPU_MUL_MAT_SUBGROUP_M);
         sg_matrix_repls["WEBGPU_SUBGROUP_N"]        = std::to_string(WEBGPU_MUL_MAT_SUBGROUP_N);
         sg_matrix_repls["WEBGPU_SUBGROUP_MATRIX_M"] = std::to_string(WEBGPU_MUL_MAT_SUBGROUP_MATRIX_M);
         sg_matrix_repls["WEBGPU_SUBGROUP_MATRIX_N"] = std::to_string(WEBGPU_MUL_MAT_SUBGROUP_MATRIX_N);
-        sg_matrix_repls["WEBGPU_SG_MAT_M_SIZE"]     = std::to_string(webgpu_ctx->sg_mat_m);
-        sg_matrix_repls["WEBGPU_SG_MAT_N_SIZE"]     = std::to_string(webgpu_ctx->sg_mat_n);
-        sg_matrix_repls["WEBGPU_SG_MAT_K_SIZE"]     = std::to_string(webgpu_ctx->sg_mat_k);
-
+        sg_matrix_repls["WEBGPU_SG_MAT_M_SIZE"]     = std::to_string(webgpu_ctx->global_ctx->capabilities.sg_mat_m);
+        sg_matrix_repls["WEBGPU_SG_MAT_N_SIZE"]     = std::to_string(webgpu_ctx->global_ctx->capabilities.sg_mat_n);
+        sg_matrix_repls["WEBGPU_SG_MAT_K_SIZE"]     = std::to_string(webgpu_ctx->global_ctx->capabilities.sg_mat_k);
         proc_mul_mat_f32_f32 = ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_f32_f32, sg_matrix_repls);
         proc_mul_mat_f32_f32_vec =
             ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_f32_f32_vec, sg_matrix_repls);
@@ -2519,21 +2583,21 @@ static void ggml_webgpu_init_mul_mat_pipeline(webgpu_context & webgpu_ctx) {
 #endif
 
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, proc_mul_mat_f32_f32.c_str(), "mul_mat_f32_f32", mul_mat_constants);
+        webgpu_ctx->global_ctx->device, proc_mul_mat_f32_f32.c_str(), "mul_mat_f32_f32", mul_mat_constants);
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, proc_mul_mat_f32_f32_vec.c_str(), "mul_mat_f32_f32_vec", mul_mat_constants);
+        webgpu_ctx->global_ctx->device, proc_mul_mat_f32_f32_vec.c_str(), "mul_mat_f32_f32_vec", mul_mat_constants);
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, proc_mul_mat_f16_f32.c_str(), "mul_mat_f16_f32", mul_mat_constants);
+        webgpu_ctx->global_ctx->device, proc_mul_mat_f16_f32.c_str(), "mul_mat_f16_f32", mul_mat_constants);
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, proc_mul_mat_f16_f32_vec.c_str(), "mul_mat_f16_f32_vec", mul_mat_constants);
+        webgpu_ctx->global_ctx->device, proc_mul_mat_f16_f32_vec.c_str(), "mul_mat_f16_f32_vec", mul_mat_constants);
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][0] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, proc_mul_mat_f16_f16.c_str(), "mul_mat_f16_f16", mul_mat_constants);
+        webgpu_ctx->global_ctx->device, proc_mul_mat_f16_f16.c_str(), "mul_mat_f16_f16", mul_mat_constants);
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][1] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, proc_mul_mat_f16_f16_vec.c_str(), "mul_mat_f16_f16_vec", mul_mat_constants);
+        webgpu_ctx->global_ctx->device, proc_mul_mat_f16_f16_vec.c_str(), "mul_mat_f16_f16_vec", mul_mat_constants);
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_0][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, proc_mul_mat_q4_0_f32.c_str(), "mul_mat_q4_0_f32", mul_mat_constants);
+        webgpu_ctx->global_ctx->device, proc_mul_mat_q4_0_f32.c_str(), "mul_mat_q4_0_f32", mul_mat_constants);
     webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_0][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, proc_mul_mat_q4_0_f32_vec.c_str(), "mul_mat_q4_0_f32_vec", mul_mat_constants);
+        webgpu_ctx->global_ctx->device, proc_mul_mat_q4_0_f32_vec.c_str(), "mul_mat_q4_0_f32_vec", mul_mat_constants);
 
     std::vector<wgpu::ConstantEntry> mul_mat_vec_constants(3);
     mul_mat_vec_constants[0].key   = "WORKGROUP_SIZE";
@@ -2544,171 +2608,171 @@ static void ggml_webgpu_init_mul_mat_pipeline(webgpu_context & webgpu_ctx) {
     mul_mat_vec_constants[2].value = WEBGPU_MUL_MAT_VEC_OUTPUTS_PER_WG;
 
     webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, wgsl_mul_mat_vec_f32_f32, "mul_mat_vec_f32_f32", mul_mat_vec_constants);
+        webgpu_ctx->global_ctx->device, wgsl_mul_mat_vec_f32_f32, "mul_mat_vec_f32_f32", mul_mat_vec_constants);
     webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, wgsl_mul_mat_vec_f32_f32_vec, "mul_mat_vec_f32_f32_vec", mul_mat_vec_constants);
+        webgpu_ctx->global_ctx->device, wgsl_mul_mat_vec_f32_f32_vec, "mul_mat_vec_f32_f32_vec", mul_mat_vec_constants);
     webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, wgsl_mul_mat_vec_f16_f32, "mul_mat_vec_f16_f32", mul_mat_vec_constants);
+        webgpu_ctx->global_ctx->device, wgsl_mul_mat_vec_f16_f32, "mul_mat_vec_f16_f32", mul_mat_vec_constants);
     webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, wgsl_mul_mat_vec_f16_f32_vec, "mul_mat_vec_f16_f32_vec", mul_mat_vec_constants);
+        webgpu_ctx->global_ctx->device, wgsl_mul_mat_vec_f16_f32_vec, "mul_mat_vec_f16_f32_vec", mul_mat_vec_constants);
     webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][0] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, wgsl_mul_mat_vec_f16_f16, "mul_mat_vec_f16_f16", mul_mat_vec_constants);
+        webgpu_ctx->global_ctx->device, wgsl_mul_mat_vec_f16_f16, "mul_mat_vec_f16_f16", mul_mat_vec_constants);
     webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][1] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, wgsl_mul_mat_vec_f16_f16_vec, "mul_mat_vec_f16_f16_vec", mul_mat_vec_constants);
+        webgpu_ctx->global_ctx->device, wgsl_mul_mat_vec_f16_f16_vec, "mul_mat_vec_f16_f16_vec", mul_mat_vec_constants);
     webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_Q4_0][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, wgsl_mul_mat_vec_q4_0_f32, "mul_mat_vec_q4_0_f32", mul_mat_vec_constants);
+        webgpu_ctx->global_ctx->device, wgsl_mul_mat_vec_q4_0_f32, "mul_mat_vec_q4_0_f32", mul_mat_vec_constants);
 }
 
 static void ggml_webgpu_init_get_rows_pipeline(webgpu_context & webgpu_ctx) {
     std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
 
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_f32, "get_rows_f32", constants);
-    webgpu_ctx->get_rows_pipelines[GGML_TYPE_F32][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_f32_vec, "get_rows_f32_vec", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_f32, "get_rows_f32", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_get_rows_f32_vec, "get_rows_f32_vec", constants);
 
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_F16][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_f16, "get_rows_f16", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_f16, "get_rows_f16", constants);
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_I32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_i32, "get_rows_i32", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_i32, "get_rows_i32", constants);
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q4_0][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q4_0, "get_rows_q4_0", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q4_0, "get_rows_q4_0", constants);
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q4_1][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q4_1, "get_rows_q4_1", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q4_1, "get_rows_q4_1", constants);
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q5_0][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q5_0, "get_rows_q5_0", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q5_0, "get_rows_q5_0", constants);
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q5_1][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q5_1, "get_rows_q5_1", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q5_1, "get_rows_q5_1", constants);
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q8_0][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q8_0, "get_rows_q8_0", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q8_0, "get_rows_q8_0", constants);
 
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q2_K][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q2_k, "get_rows_q2_k", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q2_k, "get_rows_q2_k", constants);
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q3_K][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q3_k, "get_rows_q3_k", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q3_k, "get_rows_q3_k", constants);
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q4_K][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q4_k, "get_rows_q4_k", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q4_k, "get_rows_q4_k", constants);
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q5_K][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q5_k, "get_rows_q5_k", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q5_k, "get_rows_q5_k", constants);
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_Q6_K][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_q6_k, "get_rows_q6_k", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_q6_k, "get_rows_q6_k", constants);
 
-    webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ2_XXS][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq2_xxs, "get_rows_iq2_xxs", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ2_XXS][0] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_get_rows_iq2_xxs, "get_rows_iq2_xxs", constants);
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ2_XS][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq2_xs, "get_rows_iq2_xs", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_iq2_xs, "get_rows_iq2_xs", constants);
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ2_S][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq2_s, "get_rows_iq2_s", constants);
-    webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ3_XXS][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq3_xxs, "get_rows_iq3_xxs", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_iq2_s, "get_rows_iq2_s", constants);
+    webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ3_XXS][0] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_get_rows_iq3_xxs, "get_rows_iq3_xxs", constants);
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ3_S][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq3_s, "get_rows_iq3_s", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_iq3_s, "get_rows_iq3_s", constants);
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ1_S][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq1_s, "get_rows_iq1_s", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_iq1_s, "get_rows_iq1_s", constants);
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ1_M][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq1_m, "get_rows_iq1_m", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_iq1_m, "get_rows_iq1_m", constants);
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ4_NL][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq4_nl, "get_rows_iq4_nl", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_iq4_nl, "get_rows_iq4_nl", constants);
     webgpu_ctx->get_rows_pipelines[GGML_TYPE_IQ4_XS][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_get_rows_iq4_xs, "get_rows_iq4_xs", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_get_rows_iq4_xs, "get_rows_iq4_xs", constants);
 }
 
 static void ggml_webgpu_init_cpy_pipeline(webgpu_context & webgpu_ctx) {
     std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
 
     webgpu_ctx->cpy_pipelines[GGML_TYPE_F32][GGML_TYPE_F32] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_cpy_f32_f32, "cpy_f32_f32", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_cpy_f32_f32, "cpy_f32_f32", constants);
     webgpu_ctx->cpy_pipelines[GGML_TYPE_F32][GGML_TYPE_I32] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_cpy_f32_i32, "cpy_f32_i32", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_cpy_f32_i32, "cpy_f32_i32", constants);
     webgpu_ctx->cpy_pipelines[GGML_TYPE_F32][GGML_TYPE_F16] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_cpy_f32_f16, "cpy_f32_f16", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_cpy_f32_f16, "cpy_f32_f16", constants);
     webgpu_ctx->cpy_pipelines[GGML_TYPE_F16][GGML_TYPE_F32] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_cpy_f16_f32, "cpy_f16_f32", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_cpy_f16_f32, "cpy_f16_f32", constants);
     webgpu_ctx->cpy_pipelines[GGML_TYPE_F16][GGML_TYPE_F16] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_cpy_f16_f16, "cpy_f16_f16", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_cpy_f16_f16, "cpy_f16_f16", constants);
 }
 
 static void ggml_webgpu_init_add_pipeline(webgpu_context & webgpu_ctx) {
     std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
 
     webgpu_ctx->add_pipelines[GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_add_f32, "add_f32", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_add_f32, "add_f32", constants);
     webgpu_ctx->add_pipelines[GGML_TYPE_F16][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_add_f16, "add_f16", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_add_f16, "add_f16", constants);
     webgpu_ctx->add_pipelines[GGML_TYPE_F32][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_add_f32_inplace, "add_f32_inplace", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_add_f32_inplace, "add_f32_inplace", constants);
     webgpu_ctx->add_pipelines[GGML_TYPE_F16][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_add_f16_inplace, "add_f16_inplace", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_add_f16_inplace, "add_f16_inplace", constants);
 }
 
 static void ggml_webgpu_init_sub_pipeline(webgpu_context & webgpu_ctx) {
     std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
 
     webgpu_ctx->sub_pipelines[GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sub_f32, "sub_f32", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_sub_f32, "sub_f32", constants);
     webgpu_ctx->sub_pipelines[GGML_TYPE_F16][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sub_f16, "sub_f16", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_sub_f16, "sub_f16", constants);
     webgpu_ctx->sub_pipelines[GGML_TYPE_F32][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sub_f32_inplace, "sub_f32_inplace", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_sub_f32_inplace, "sub_f32_inplace", constants);
     webgpu_ctx->sub_pipelines[GGML_TYPE_F16][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_sub_f16_inplace, "sub_f16_inplace", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_sub_f16_inplace, "sub_f16_inplace", constants);
 }
 
 static void ggml_webgpu_init_mul_pipeline(webgpu_context & webgpu_ctx) {
     std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
 
     webgpu_ctx->mul_pipelines[GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_f32, "mul_f32", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_f32, "mul_f32", constants);
     webgpu_ctx->mul_pipelines[GGML_TYPE_F16][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_f16, "mul_f16", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_f16, "mul_f16", constants);
     webgpu_ctx->mul_pipelines[GGML_TYPE_F32][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_f32_inplace, "mul_f32_inplace", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_f32_inplace, "mul_f32_inplace", constants);
     webgpu_ctx->mul_pipelines[GGML_TYPE_F16][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_mul_f16_inplace, "mul_f16_inplace", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_mul_f16_inplace, "mul_f16_inplace", constants);
 }
 
 static void ggml_webgpu_init_div_pipeline(webgpu_context & webgpu_ctx) {
     std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
 
     webgpu_ctx->div_pipelines[GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_div_f32, "div_f32", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_div_f32, "div_f32", constants);
     webgpu_ctx->div_pipelines[GGML_TYPE_F16][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_div_f16, "div_f16", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_div_f16, "div_f16", constants);
     webgpu_ctx->div_pipelines[GGML_TYPE_F32][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_div_f32_inplace, "div_f32_inplace", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_div_f32_inplace, "div_f32_inplace", constants);
     webgpu_ctx->div_pipelines[GGML_TYPE_F16][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_div_f16_inplace, "div_f16_inplace", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_div_f16_inplace, "div_f16_inplace", constants);
 }
 
 static void ggml_webgpu_init_rms_norm_pipeline(webgpu_context & webgpu_ctx) {
     std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_ROW_SPLIT_WG_SIZE);
 
     webgpu_ctx->rms_norm_pipelines[0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rms_norm, "rms_norm", constants);
-    webgpu_ctx->rms_norm_pipelines[1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rms_norm_inplace, "rms_norm_inplace", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_rms_norm, "rms_norm", constants);
+    webgpu_ctx->rms_norm_pipelines[1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_rms_norm_inplace, "rms_norm_inplace", constants);
 }
 
 static void ggml_webgpu_init_rope_pipeline(webgpu_context & webgpu_ctx) {
     std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
 
     webgpu_ctx->rope_pipelines[GGML_TYPE_F32][0][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f32, "rope_f32", constants);
-    webgpu_ctx->rope_pipelines[GGML_TYPE_F32][0][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f32_inplace, "rope_f32_inplace", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_rope_f32, "rope_f32", constants);
+    webgpu_ctx->rope_pipelines[GGML_TYPE_F32][0][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_rope_f32_inplace, "rope_f32_inplace", constants);
     webgpu_ctx->rope_pipelines[GGML_TYPE_F32][1][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f32_ff, "rope_f32_ff", constants);
-    webgpu_ctx->rope_pipelines[GGML_TYPE_F32][1][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f32_ff_inplace, "rope_f32_ff_inplace", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_rope_f32_ff, "rope_f32_ff", constants);
+    webgpu_ctx->rope_pipelines[GGML_TYPE_F32][1][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_rope_f32_ff_inplace, "rope_f32_ff_inplace", constants);
 
     webgpu_ctx->rope_pipelines[GGML_TYPE_F16][0][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f16, "rope_f16", constants);
-    webgpu_ctx->rope_pipelines[GGML_TYPE_F16][0][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f16_inplace, "rope_f16_inplace", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_rope_f16, "rope_f16", constants);
+    webgpu_ctx->rope_pipelines[GGML_TYPE_F16][0][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_rope_f16_inplace, "rope_f16_inplace", constants);
     webgpu_ctx->rope_pipelines[GGML_TYPE_F16][1][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f16_ff, "rope_f16_ff", constants);
-    webgpu_ctx->rope_pipelines[GGML_TYPE_F16][1][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_rope_f16_ff_inplace, "rope_f16_ff_inplace", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_rope_f16_ff, "rope_f16_ff", constants);
+    webgpu_ctx->rope_pipelines[GGML_TYPE_F16][1][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_rope_f16_ff_inplace, "rope_f16_ff_inplace", constants);
 }
 
 static void ggml_webgpu_init_glu_pipeline(webgpu_context & webgpu_ctx) {
@@ -2716,68 +2780,68 @@ static void ggml_webgpu_init_glu_pipeline(webgpu_context & webgpu_ctx) {
 
     // REGLU
     webgpu_ctx->glu_pipelines[GGML_GLU_OP_REGLU][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_reglu_f32, "reglu_f32", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_reglu_f32, "reglu_f32", constants);
     webgpu_ctx->glu_pipelines[GGML_GLU_OP_REGLU][GGML_TYPE_F16][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_reglu_f16, "reglu_f16", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_reglu_f16, "reglu_f16", constants);
     webgpu_ctx->glu_pipelines[GGML_GLU_OP_REGLU][GGML_TYPE_F32][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_reglu_f32_split, "reglu_f32_split", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_reglu_f32_split, "reglu_f32_split", constants);
     webgpu_ctx->glu_pipelines[GGML_GLU_OP_REGLU][GGML_TYPE_F16][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_reglu_f16_split, "reglu_f16_split", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_reglu_f16_split, "reglu_f16_split", constants);
 
     // GEGLU
     webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_f32, "geglu_f32", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_geglu_f32, "geglu_f32", constants);
     webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU][GGML_TYPE_F16][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_f16, "geglu_f16", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_geglu_f16, "geglu_f16", constants);
     webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU][GGML_TYPE_F32][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_f32_split, "geglu_f32_split", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_geglu_f32_split, "geglu_f32_split", constants);
     webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU][GGML_TYPE_F16][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_f16_split, "geglu_f16_split", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_geglu_f16_split, "geglu_f16_split", constants);
 
     // SWIGLU
     webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_f32, "swiglu_f32", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_swiglu_f32, "swiglu_f32", constants);
     webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU][GGML_TYPE_F16][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_f16, "swiglu_f16", constants);
-    webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU][GGML_TYPE_F32][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_f32_split, "swiglu_f32_split", constants);
-    webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU][GGML_TYPE_F16][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_f16_split, "swiglu_f16_split", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_swiglu_f16, "swiglu_f16", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_swiglu_f32_split, "swiglu_f32_split", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU][GGML_TYPE_F16][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_swiglu_f16_split, "swiglu_f16_split", constants);
 
     // SWIGLU_OAI
     webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU_OAI][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_oai_f32, "swiglu_oai_f32", constants);
-    webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU_OAI][GGML_TYPE_F32][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_swiglu_oai_f32_split, "swiglu_oai_f32_split", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_swiglu_oai_f32, "swiglu_oai_f32", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_SWIGLU_OAI][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_swiglu_oai_f32_split, "swiglu_oai_f32_split", constants);
 
     // GEGLU_ERF
     webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_ERF][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_erf_f32, "geglu_erf_f32", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_geglu_erf_f32, "geglu_erf_f32", constants);
     webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_ERF][GGML_TYPE_F16][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_erf_f16, "geglu_erf_f16", constants);
-    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_ERF][GGML_TYPE_F32][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_erf_f32_split, "geglu_erf_f32_split", constants);
-    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_ERF][GGML_TYPE_F16][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_erf_f16_split, "geglu_erf_f16_split", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_geglu_erf_f16, "geglu_erf_f16", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_ERF][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_geglu_erf_f32_split, "geglu_erf_f32_split", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_ERF][GGML_TYPE_F16][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_geglu_erf_f16_split, "geglu_erf_f16_split", constants);
 
     // GEGLU_QUICK
     webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_QUICK][GGML_TYPE_F32][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_quick_f32, "geglu_quick_f32", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_geglu_quick_f32, "geglu_quick_f32", constants);
     webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_QUICK][GGML_TYPE_F16][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_quick_f16, "geglu_quick_f16", constants);
-    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_QUICK][GGML_TYPE_F32][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_quick_f32_split, "geglu_quick_f32_split", constants);
-    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_QUICK][GGML_TYPE_F16][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_geglu_quick_f16_split, "geglu_quick_f16_split", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_geglu_quick_f16, "geglu_quick_f16", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_QUICK][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_geglu_quick_f32_split, "geglu_quick_f32_split", constants);
+    webgpu_ctx->glu_pipelines[GGML_GLU_OP_GEGLU_QUICK][GGML_TYPE_F16][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_geglu_quick_f16_split, "geglu_quick_f16_split", constants);
 }
 
 static void ggml_webgpu_init_scale_pipeline(webgpu_context & webgpu_ctx) {
     std::vector<wgpu::ConstantEntry> constants = ggml_webgpu_wg_size_entry(WEBGPU_MAX_WG_SIZE);
 
     webgpu_ctx->scale_pipelines[0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_scale_f32, "scale_f32", constants);
-    webgpu_ctx->scale_pipelines[1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_scale_f32_inplace, "scale_f32_inplace", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_scale_f32, "scale_f32", constants);
+    webgpu_ctx->scale_pipelines[1] = ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_scale_f32_inplace,
+                                                                 "scale_f32_inplace", constants);
 }
 
 static void ggml_webgpu_init_soft_max_pipeline(webgpu_context & webgpu_ctx) {
@@ -2785,56 +2849,243 @@ static void ggml_webgpu_init_soft_max_pipeline(webgpu_context & webgpu_ctx) {
 
     // f32 (no mask)
     webgpu_ctx->soft_max_pipelines[2][0][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_soft_max_f32, "soft_max_f32", constants);
-    webgpu_ctx->soft_max_pipelines[2][0][1] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_soft_max_f32_inplace, "soft_max_f32_inplace", constants);
-    webgpu_ctx->soft_max_pipelines[2][1][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_soft_max_f32_sink, "soft_max_f32_sink", constants);
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_soft_max_f32, "soft_max_f32", constants);
+    webgpu_ctx->soft_max_pipelines[2][0][1] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_inplace, "soft_max_f32_inplace", constants);
+    webgpu_ctx->soft_max_pipelines[2][1][0] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_sink, "soft_max_f32_sink", constants);
     webgpu_ctx->soft_max_pipelines[2][1][1] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, wgsl_soft_max_f32_sink_inplace, "soft_max_f32_sink_inplace", constants);
+        webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_sink_inplace, "soft_max_f32_sink_inplace", constants);
 
     // f32 mask (mask_type = 0)
-    webgpu_ctx->soft_max_pipelines[0][0][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_soft_max_f32_mask_f32, "soft_max_f32_mask_f32", constants);
+    webgpu_ctx->soft_max_pipelines[0][0][0] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_mask_f32, "soft_max_f32_mask_f32", constants);
     webgpu_ctx->soft_max_pipelines[0][0][1] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, wgsl_soft_max_f32_mask_f32_inplace, "soft_max_f32_mask_f32_inplace", constants);
+        webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_mask_f32_inplace, "soft_max_f32_mask_f32_inplace", constants);
     webgpu_ctx->soft_max_pipelines[0][1][0] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, wgsl_soft_max_f32_mask_f32_sink, "soft_max_f32_mask_f32_sink", constants);
-    webgpu_ctx->soft_max_pipelines[0][1][1] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, wgsl_soft_max_f32_mask_f32_sink_inplace, "soft_max_f32_mask_f32_sink_inplace", constants);
+        webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_mask_f32_sink, "soft_max_f32_mask_f32_sink", constants);
+    webgpu_ctx->soft_max_pipelines[0][1][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_mask_f32_sink_inplace,
+                                    "soft_max_f32_mask_f32_sink_inplace", constants);
 
     // f16 mask (mask_type = 1)
-    webgpu_ctx->soft_max_pipelines[1][0][0] =
-        ggml_webgpu_create_pipeline(webgpu_ctx->device, wgsl_soft_max_f32_mask_f16, "soft_max_f32_mask_f16", constants);
+    webgpu_ctx->soft_max_pipelines[1][0][0] = ggml_webgpu_create_pipeline(
+        webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_mask_f16, "soft_max_f32_mask_f16", constants);
     webgpu_ctx->soft_max_pipelines[1][0][1] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, wgsl_soft_max_f32_mask_f16_inplace, "soft_max_f32_mask_f16_inplace", constants);
+        webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_mask_f16_inplace, "soft_max_f32_mask_f16_inplace", constants);
     webgpu_ctx->soft_max_pipelines[1][1][0] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, wgsl_soft_max_f32_mask_f16_sink, "soft_max_f32_mask_f16_sink", constants);
-    webgpu_ctx->soft_max_pipelines[1][1][1] = ggml_webgpu_create_pipeline(
-        webgpu_ctx->device, wgsl_soft_max_f32_mask_f16_sink_inplace, "soft_max_f32_mask_f16_sink_inplace", constants);
+        webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_mask_f16_sink, "soft_max_f32_mask_f16_sink", constants);
+    webgpu_ctx->soft_max_pipelines[1][1][1] =
+        ggml_webgpu_create_pipeline(webgpu_ctx->global_ctx->device, wgsl_soft_max_f32_mask_f16_sink_inplace,
+                                    "soft_max_f32_mask_f16_sink_inplace", constants);
 }
 
-// TODO: move most initialization logic here
-static ggml_backend_t ggml_backend_webgpu_device_init(ggml_backend_dev_t dev, const char * params) {
+static bool create_webgpu_device(ggml_backend_webgpu_reg_context * ctx) {
+    wgpu::RequestAdapterOptions options = {};
+
+#ifndef __EMSCRIPTEN__
+    // TODO: track need for these toggles: https://issues.chromium.org/issues/42251215
+    const char * const          adapterEnabledToggles[] = { "vulkan_enable_f16_on_nvidia", "use_vulkan_memory_model" };
+    wgpu::DawnTogglesDescriptor adapterTogglesDesc;
+    adapterTogglesDesc.enabledToggles     = adapterEnabledToggles;
+    adapterTogglesDesc.enabledToggleCount = 2;
+    options.nextInChain                   = &adapterTogglesDesc;
+#endif
+
+    ctx->webgpu_global_ctx->instance.WaitAny(
+        ctx->webgpu_global_ctx->instance.RequestAdapter(
+            &options, wgpu::CallbackMode::AllowSpontaneous,
+            [&ctx](wgpu::RequestAdapterStatus status, wgpu::Adapter adapter, const char * message) {
+                if (status != wgpu::RequestAdapterStatus::Success) {
+                    GGML_LOG_ERROR("ggml_webgpu: Failed to get an adapter: %s\n", message);
+                    return;
+                }
+                ctx->webgpu_global_ctx->adapter = std::move(adapter);
+            }),
+        UINT64_MAX);
+    GGML_ASSERT(ctx->webgpu_global_ctx->adapter != nullptr);
+
+    ctx->webgpu_global_ctx->adapter.GetLimits(&ctx->webgpu_global_ctx->capabilities.limits);
+
+    wgpu::AdapterInfo info{};
+#ifndef __EMSCRIPTEN__
+    wgpu::AdapterPropertiesSubgroupMatrixConfigs subgroup_matrix_configs{};
+    if (ctx->webgpu_global_ctx->adapter.HasFeature(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix)) {
+        info.nextInChain = &subgroup_matrix_configs;
+    }
+#endif
+    ctx->webgpu_global_ctx->adapter.GetInfo(&info);
+    wgpu::SupportedFeatures features;
+    ctx->webgpu_global_ctx->adapter.GetFeatures(&features);
+    // we require f16 support
+    GGML_ASSERT(ctx->webgpu_global_ctx->adapter.HasFeature(wgpu::FeatureName::ShaderF16));
+
+#ifndef __EMSCRIPTEN__
+    // Only support square f16 matrices of size 8 or 16 for now
+    bool valid_subgroup_matrix_config = false;
+    if (ctx->webgpu_global_ctx->adapter.HasFeature(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix)) {
+        for (size_t i = 0; i < subgroup_matrix_configs.configCount; i++) {
+            const wgpu::SubgroupMatrixConfig config = subgroup_matrix_configs.configs[i];
+            if (config.M == config.N && config.N == config.K && (config.K == 8 || config.K == 16) &&
+                config.componentType == wgpu::SubgroupMatrixComponentType::F16 &&
+                config.resultComponentType == wgpu::SubgroupMatrixComponentType::F16) {
+                ctx->webgpu_global_ctx->capabilities.sg_mat_m = config.M;
+                ctx->webgpu_global_ctx->capabilities.sg_mat_n = config.N;
+                ctx->webgpu_global_ctx->capabilities.sg_mat_k = config.K;
+                valid_subgroup_matrix_config                  = true;
+                break;
+            }
+        }
+    }
+    ctx->webgpu_global_ctx->capabilities.supports_subgroup_matrix = valid_subgroup_matrix_config;
+#endif
+
+    // For subgroup matrix code to be the most efficient, we would like the subgroup size to be consistent and accurate.
+    // Unfortunately, that is not possible, so we use the maximum subgroup size reported by the adapter.
+    ctx->webgpu_global_ctx->capabilities.max_subgroup_size = info.subgroupMaxSize;
+    // Initialize device
+    std::vector<wgpu::FeatureName> required_features       = { wgpu::FeatureName::ShaderF16 };
+
+#ifndef __EMSCRIPTEN__
+    required_features.push_back(wgpu::FeatureName::ImplicitDeviceSynchronization);
+    if (ctx->webgpu_global_ctx->capabilities.supports_subgroup_matrix) {
+        required_features.push_back(wgpu::FeatureName::Subgroups);
+        required_features.push_back(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix);
+    }
+#endif
+
+#ifdef GGML_WEBGPU_GPU_PROFILE
+    required_features.push_back(wgpu::FeatureName::TimestampQuery);
+#endif
+
+    wgpu::DeviceDescriptor dev_desc;
+    dev_desc.requiredLimits       = &ctx->webgpu_global_ctx->capabilities.limits;
+    dev_desc.requiredFeatures     = required_features.data();
+    dev_desc.requiredFeatureCount = required_features.size();
+    dev_desc.SetDeviceLostCallback(
+        wgpu::CallbackMode::AllowSpontaneous,
+        [](const wgpu::Device & device, wgpu::DeviceLostReason reason, wgpu::StringView message) {
+            GGML_UNUSED(device);
+            GGML_UNUSED(reason);
+            GGML_UNUSED(message);
+            //TODO: uncomment once proper free logic is in place
+            //GGML_LOG_ERROR("ggml_webgpu: Device lost! Reason: %d, Message: %s\n", static_cast<int>(reason),
+            //std::string(message).c_str());
+        });
+    dev_desc.SetUncapturedErrorCallback(
+        [](const wgpu::Device & device, wgpu::ErrorType reason, wgpu::StringView message) {
+            GGML_UNUSED(device);
+            GGML_ABORT("ggml_webgpu: Device error! Reason: %d, Message: %s\n", static_cast<int>(reason),
+                       std::string(message).c_str());
+        });
+
+#ifndef __EMSCRIPTEN__
+    // Enable Dawn-specific toggles to increase native performance
+    // TODO: Maybe WebGPU needs a "fast" mode where you can request compilers skip adding checks like these,
+    //       only for native performance?
+    const char * const deviceEnabledToggles[]  = { "skip_validation", "disable_robustness", "disable_workgroup_init",
+                                                   "disable_polyfills_on_integer_div_and_mod" };
+    const char * const deviceDisabledToggles[] = { "timestamp_quantization" };
+    wgpu::DawnTogglesDescriptor deviceTogglesDesc;
+    deviceTogglesDesc.enabledToggles      = deviceEnabledToggles;
+    deviceTogglesDesc.enabledToggleCount  = 4;
+    deviceTogglesDesc.disabledToggles     = deviceDisabledToggles;
+    deviceTogglesDesc.disabledToggleCount = 1;
+
+    dev_desc.nextInChain = &deviceTogglesDesc;
+#endif
+
+    ctx->webgpu_global_ctx->instance.WaitAny(
+        ctx->webgpu_global_ctx->adapter.RequestDevice(
+            &dev_desc, wgpu::CallbackMode::AllowSpontaneous,
+            [ctx](wgpu::RequestDeviceStatus status, wgpu::Device device, wgpu::StringView message) {
+                if (status != wgpu::RequestDeviceStatus::Success) {
+                    GGML_LOG_ERROR("ggml_webgpu: Failed to get a device: %s\n", std::string(message).c_str());
+                    return;
+                }
+                ctx->webgpu_global_ctx->device = std::move(device);
+            }),
+        UINT64_MAX);
+    GGML_ASSERT(ctx->webgpu_global_ctx->device != nullptr);
+
+    ggml_webgpu_init_memset_pipeline(ctx->webgpu_global_ctx);
+    ctx->webgpu_global_ctx->memset_buf_pool.init(ctx->webgpu_global_ctx->device, 1, WEBGPU_PARAMS_BUF_SIZE_BYTES,
+                                                 wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::Uniform,
+                                                 wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::MapWrite);
+    ctx->webgpu_global_ctx->queue = ctx->webgpu_global_ctx->device.GetQueue();
+
+#ifdef GGML_WEBGPU_GPU_PROFILE
+    // Initialize buffer pool for timestamp queries, used for profiling
+    ctx->webgpu_global_ctx->timestamp_query_buf_pool.init(ctx->webgpu_global_ctx->device, WEBGPU_NUM_TIMESTAMP_QUERY_BUFS,
+                                              WEBGPU_TIMESTAMP_QUERY_BUF_SIZE_BYTES,
+                                              wgpu::BufferUsage::QueryResolve | wgpu::BufferUsage::CopySrc,
+                                              wgpu::BufferUsage::MapRead | wgpu::BufferUsage::CopyDst);
+#endif
+
+    GGML_LOG_INFO(
+        "ggml_webgpu: adapter_info: vendor_id: %u | vendor: %s | architecture: %s | device_id: %u | name: %s | "
+        "device_desc: %s\n",
+        info.vendorID, std::string(info.vendor).c_str(), std::string(info.architecture).c_str(), info.deviceID,
+        std::string(info.device).c_str(), std::string(info.description).c_str());
+    return true;
+}
+
+static webgpu_context initialize_webgpu_context(ggml_backend_dev_t dev) {
+    ggml_backend_webgpu_device_context * dev_ctx    = (ggml_backend_webgpu_device_context *) dev->context;
+    webgpu_context                       webgpu_ctx = std::make_shared<webgpu_context_struct>();
+    webgpu_ctx->global_ctx                          = dev_ctx->webgpu_global_ctx;
+    webgpu_ctx->param_buf_pool.init(webgpu_ctx->global_ctx->device, WEBGPU_NUM_PARAM_BUFS, WEBGPU_PARAMS_BUF_SIZE_BYTES,
+                                    wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::Uniform,
+                                    wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::MapWrite);
+    webgpu_ctx->set_rows_error_buf_pool.init(webgpu_ctx->global_ctx->device, WEBGPU_NUM_SET_ROWS_ERROR_BUFS,
+                                             WEBGPU_SET_ROWS_ERROR_BUF_SIZE_BYTES,
+                                             wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::Storage,
+                                             wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead);
+
+    ggml_webgpu_init_mul_mat_pipeline(webgpu_ctx);
+    ggml_webgpu_init_get_rows_pipeline(webgpu_ctx);
+    ggml_webgpu_init_cpy_pipeline(webgpu_ctx);
+    ggml_webgpu_init_add_pipeline(webgpu_ctx);
+    ggml_webgpu_init_sub_pipeline(webgpu_ctx);
+    ggml_webgpu_init_mul_pipeline(webgpu_ctx);
+    ggml_webgpu_init_div_pipeline(webgpu_ctx);
+    ggml_webgpu_init_rms_norm_pipeline(webgpu_ctx);
+    ggml_webgpu_init_rope_pipeline(webgpu_ctx);
+    ggml_webgpu_init_glu_pipeline(webgpu_ctx);
+    ggml_webgpu_init_scale_pipeline(webgpu_ctx);
+    ggml_webgpu_init_soft_max_pipeline(webgpu_ctx);
+#ifdef GGML_WEBGPU_DEBUG
+    // Initialize debug buffers
+    ggml_webgpu_create_buffer(webgpu_ctx->global_ctx->device, webgpu_ctx->global_ctx->debug_host_buf,
+                              WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t),
+                              wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead, "debug_host_buf");
+    ggml_webgpu_create_buffer(webgpu_ctx->global_ctx->device, webgpu_ctx->global_ctx->debug_dev_buf,
+                              WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t),
+                              wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc, "debug_dev_buf");
+#endif
+    return webgpu_ctx;
+}
+
+static ggml_backend_t ggml_backend_webgpu_backend_init(ggml_backend_dev_t dev, const char * params) {
     GGML_UNUSED(params);
 
-    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_device_init()");
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_backend_init()");
 
-    ggml_backend_webgpu_device_context * dev_ctx    = static_cast<ggml_backend_webgpu_device_context *>(dev->context);
-    webgpu_context                       webgpu_ctx = dev_ctx->webgpu_ctx;
+    ggml_backend_webgpu_device_context * dev_ctx = static_cast<ggml_backend_webgpu_device_context *>(dev->context);
 
-    static ggml_backend_webgpu_context backend_ctx;
-    backend_ctx.name       = GGML_WEBGPU_NAME + std::string(": ") + dev_ctx->device_name;
-    backend_ctx.webgpu_ctx = webgpu_ctx;
+    auto * backend_ctx      = new ggml_backend_webgpu_context();
+    backend_ctx->name       = GGML_WEBGPU_NAME + std::string(": ") + dev_ctx->device_name;
+    backend_ctx->webgpu_ctx = initialize_webgpu_context(dev);
 
     // See GGML Backend Interface section
-    static ggml_backend backend = {
+    auto * backend = new ggml_backend();
+    *backend       = {
         /* .guid      = */ ggml_backend_webgpu_guid(),
         /* .interface = */ ggml_backend_webgpu_i,
         /* .device    = */ dev,
-        /* .context   = */ &backend_ctx,
+        /* .context   = */ backend_ctx,
     };
-    return &backend;
+    return backend;
 }
 
 static ggml_backend_buffer_type_t ggml_backend_webgpu_device_get_buffer_type(ggml_backend_dev_t dev) {
@@ -2851,7 +3102,8 @@ static ggml_backend_buffer_type_t ggml_backend_webgpu_device_get_buffer_type(ggm
         },
         /* .device  = */
         dev,
-        /* .context = */ NULL,
+        /* .context = */
+        NULL
     };
 
     return &ggml_backend_webgpu_buffer_type;
@@ -2892,16 +3144,16 @@ static bool ggml_webgpu_supported_qtype(ggml_type type) {
 static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
     ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(dev->context);
 
-    webgpu_context webgpu_ctx = ctx->webgpu_ctx;
-
     ggml_tensor * src0 = op->src[0];
     ggml_tensor * src1 = op->src[1];
     ggml_tensor * src2 = op->src[2];
 
     // on smaller devices (or CI), tensors may be larger than the max storage buffer size
-    if (ggml_nbytes(op) > webgpu_ctx->limits.maxStorageBufferBindingSize ||
-        (src0 != nullptr && ggml_nbytes(src0) > webgpu_ctx->limits.maxStorageBufferBindingSize) ||
-        (src1 != nullptr && ggml_nbytes(src1) > webgpu_ctx->limits.maxStorageBufferBindingSize)) {
+    if (ggml_nbytes(op) > ctx->webgpu_global_ctx->capabilities.limits.maxStorageBufferBindingSize ||
+        (src0 != nullptr &&
+         ggml_nbytes(src0) > ctx->webgpu_global_ctx->capabilities.limits.maxStorageBufferBindingSize) ||
+        (src1 != nullptr &&
+         ggml_nbytes(src1) > ctx->webgpu_global_ctx->capabilities.limits.maxStorageBufferBindingSize)) {
         return false;
     }
 
@@ -2981,17 +3233,19 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
             }
         case GGML_OP_FLASH_ATTN_EXT:
             {
-                if (!webgpu_ctx->supports_subgroup_matrix) {
+#ifndef __EMSCRIPTEN__
+                if (!ctx->webgpu_global_ctx->capabilities.supports_subgroup_matrix) {
                     break;
                 }
                 // Head dimensions must fit in workgroup memory with minimum tile sizes
-                size_t     limit_bytes = webgpu_ctx->limits.maxComputeWorkgroupStorageSize;
+                size_t     limit_bytes = ctx->webgpu_global_ctx->capabilities.limits.maxComputeWorkgroupStorageSize;
                 const bool has_mask    = op->src[3] != nullptr;
-                const bool kv_direct   = src1->type == GGML_TYPE_F16 && (src0->ne[0] % webgpu_ctx->sg_mat_k) == 0 &&
+                const bool kv_direct   = src1->type == GGML_TYPE_F16 &&
+                                       (src0->ne[0] % ctx->webgpu_global_ctx->capabilities.sg_mat_k) == 0 &&
                                        (src1->ne[1] % GGML_WEBGPU_KV_SEQ_PAD) == 0;
                 const size_t min_bytes = ggml_webgpu_flash_attn_wg_mem_bytes(
-                    webgpu_ctx->sg_mat_m, webgpu_ctx->sg_mat_n, (uint32_t) src0->ne[0], (uint32_t) src2->ne[0],
-                    has_mask, kv_direct);
+                    ctx->webgpu_global_ctx->capabilities.sg_mat_m, ctx->webgpu_global_ctx->capabilities.sg_mat_n,
+                    (uint32_t) src0->ne[0], (uint32_t) src2->ne[0], has_mask, kv_direct);
                 if (min_bytes > limit_bytes) {
                     break;
                 }
@@ -3000,6 +3254,7 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
                               (src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16 ||
                                src1->type == GGML_TYPE_Q4_0 || src1->type == GGML_TYPE_Q8_0) &&
                               src2->type == src1->type && op->type == GGML_TYPE_F32;
+#endif
                 break;
             }
         case GGML_OP_RMS_NORM:
@@ -3096,10 +3351,13 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
         default:
             break;
     }
-    if (ggml_nbytes(op) > webgpu_ctx->limits.maxStorageBufferBindingSize ||
-        (src0 != nullptr && ggml_nbytes(src0) > webgpu_ctx->limits.maxStorageBufferBindingSize) ||
-        (src1 != nullptr && ggml_nbytes(src1) > webgpu_ctx->limits.maxStorageBufferBindingSize) ||
-        (src2 != nullptr && ggml_nbytes(src2) > webgpu_ctx->limits.maxStorageBufferBindingSize)) {
+    if (ggml_nbytes(op) > ctx->webgpu_global_ctx->capabilities.limits.maxStorageBufferBindingSize ||
+        (src0 != nullptr &&
+         ggml_nbytes(src0) > ctx->webgpu_global_ctx->capabilities.limits.maxStorageBufferBindingSize) ||
+        (src1 != nullptr &&
+         ggml_nbytes(src1) > ctx->webgpu_global_ctx->capabilities.limits.maxStorageBufferBindingSize) ||
+        (src2 != nullptr &&
+         ggml_nbytes(src2) > ctx->webgpu_global_ctx->capabilities.limits.maxStorageBufferBindingSize)) {
         supports_op = false;
         WEBGPU_LOG_DEBUG("ggml_webgpu op not supported due to size: ");
     }
@@ -3124,7 +3382,7 @@ static struct ggml_backend_device_i ggml_backend_webgpu_device_i = {
     /* .get_memory           = */ ggml_backend_webgpu_device_get_memory,
     /* .get_type             = */ ggml_backend_webgpu_device_get_type,
     /* .get_props            = */ ggml_backend_webgpu_device_get_props,
-    /* .init_backend         = */ ggml_backend_webgpu_device_init,
+    /* .init_backend         = */ ggml_backend_webgpu_backend_init,
     /* .get_buffer_type      = */ ggml_backend_webgpu_device_get_buffer_type,
     /* .get_host_buffer_type = */ NULL,
     /* .buffer_from_host_ptr = */ NULL,
@@ -3153,6 +3411,7 @@ static size_t ggml_backend_webgpu_reg_get_device_count(ggml_backend_reg_t reg) {
 // TODO: Does this need to be thread safe? Is it only called once?
 // TODO: move most logic to device_init function so backend can be freed/initialized properly
 // Only one device is supported for now
+
 static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
     GGML_ASSERT(index == 0);
     WEBGPU_LOG_DEBUG("ggml_backend_reg_get_device()");
@@ -3161,189 +3420,12 @@ static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t
 
     ggml_backend_webgpu_reg_context * reg_ctx = static_cast<ggml_backend_webgpu_reg_context *>(reg->context);
 
-    webgpu_context ctx = reg_ctx->webgpu_ctx;
-
-    wgpu::RequestAdapterOptions options = {};
-
-#ifndef __EMSCRIPTEN__
-    // TODO: track need for these toggles: https://issues.chromium.org/issues/42251215
-    const char * const          adapterEnabledToggles[] = { "vulkan_enable_f16_on_nvidia", "use_vulkan_memory_model" };
-    wgpu::DawnTogglesDescriptor adapterTogglesDesc;
-    adapterTogglesDesc.enabledToggles     = adapterEnabledToggles;
-    adapterTogglesDesc.enabledToggleCount = 2;
-    options.nextInChain                   = &adapterTogglesDesc;
-#endif
-
-    ctx->instance.WaitAny(ctx->instance.RequestAdapter(
-                              &options, wgpu::CallbackMode::AllowSpontaneous,
-                              [&ctx](wgpu::RequestAdapterStatus status, wgpu::Adapter adapter, const char * message) {
-                                  if (status != wgpu::RequestAdapterStatus::Success) {
-                                      GGML_LOG_ERROR("ggml_webgpu: Failed to get an adapter: %s\n", message);
-                                      return;
-                                  }
-                                  ctx->adapter = std::move(adapter);
-                              }),
-                          UINT64_MAX);
-    GGML_ASSERT(ctx->adapter != nullptr);
-
-    ctx->adapter.GetLimits(&ctx->limits);
-
-    wgpu::AdapterInfo info{};
-#ifndef __EMSCRIPTEN__
-    wgpu::AdapterPropertiesSubgroupMatrixConfigs subgroup_matrix_configs{};
-    if (ctx->adapter.HasFeature(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix)) {
-        info.nextInChain = &subgroup_matrix_configs;
-    }
-#endif
-    ctx->adapter.GetInfo(&info);
-
-    wgpu::SupportedFeatures features;
-    ctx->adapter.GetFeatures(&features);
-    // we require f16 support
-    GGML_ASSERT(ctx->adapter.HasFeature(wgpu::FeatureName::ShaderF16));
-
-#ifndef __EMSCRIPTEN__
-    // Only support square f16 matrices of size 8 or 16 for now
-    bool valid_subgroup_matrix_config = false;
-    if (ctx->adapter.HasFeature(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix)) {
-        for (size_t i = 0; i < subgroup_matrix_configs.configCount; i++) {
-            const wgpu::SubgroupMatrixConfig config = subgroup_matrix_configs.configs[i];
-            if (config.M == config.N && config.N == config.K && (config.K == 8 || config.K == 16) &&
-                config.componentType == wgpu::SubgroupMatrixComponentType::F16 &&
-                config.resultComponentType == wgpu::SubgroupMatrixComponentType::F16) {
-                ctx->sg_mat_m                = config.M;
-                ctx->sg_mat_n                = config.N;
-                ctx->sg_mat_k                = config.K;
-                valid_subgroup_matrix_config = true;
-                break;
-            }
-        }
-    }
-
-    ctx->supports_subgroup_matrix = valid_subgroup_matrix_config;
-#endif
-    // For subgroup matrix code to be the most efficient, we would like the subgroup size to be consistent and accurate.
-    // Unfortunately, that is not possible, so we use the maximum subgroup size reported by the adapter.
-    ctx->max_subgroup_size = info.subgroupMaxSize;
-
-    // Initialize device
-    std::vector<wgpu::FeatureName> required_features = { wgpu::FeatureName::ShaderF16 };
-
-#ifndef __EMSCRIPTEN__
-    required_features.push_back(wgpu::FeatureName::ImplicitDeviceSynchronization);
-    if (ctx->supports_subgroup_matrix) {
-        required_features.push_back(wgpu::FeatureName::Subgroups);
-        required_features.push_back(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix);
-    }
-#endif
-
-#ifdef GGML_WEBGPU_GPU_PROFILE
-    required_features.push_back(wgpu::FeatureName::TimestampQuery);
-#endif
-
-    wgpu::DeviceDescriptor dev_desc;
-    dev_desc.requiredLimits       = &ctx->limits;
-    dev_desc.requiredFeatures     = required_features.data();
-    dev_desc.requiredFeatureCount = required_features.size();
-    dev_desc.SetDeviceLostCallback(
-        wgpu::CallbackMode::AllowSpontaneous,
-        [](const wgpu::Device & device, wgpu::DeviceLostReason reason, wgpu::StringView message) {
-            GGML_UNUSED(device);
-            GGML_UNUSED(reason);
-            GGML_UNUSED(message);
-            //TODO: uncomment once proper free logic is in place
-            //GGML_LOG_ERROR("ggml_webgpu: Device lost! Reason: %d, Message: %s\n", static_cast<int>(reason),
-            //std::string(message).c_str());
-        });
-    dev_desc.SetUncapturedErrorCallback(
-        [](const wgpu::Device & device, wgpu::ErrorType reason, wgpu::StringView message) {
-            GGML_UNUSED(device);
-            GGML_ABORT("ggml_webgpu: Device error! Reason: %d, Message: %s\n", static_cast<int>(reason),
-                       std::string(message).c_str());
-        });
-
-#ifndef __EMSCRIPTEN__
-    // Enable Dawn-specific toggles to increase native performance
-    // TODO: Maybe WebGPU needs a "fast" mode where you can request compilers skip adding checks like these,
-    //       only for native performance?
-    const char * const deviceEnabledToggles[]  = { "skip_validation", "disable_robustness", "disable_workgroup_init",
-                                                   "disable_polyfills_on_integer_div_and_mod" };
-    const char * const deviceDisabledToggles[] = { "timestamp_quantization" };
-    wgpu::DawnTogglesDescriptor deviceTogglesDesc;
-    deviceTogglesDesc.enabledToggles      = deviceEnabledToggles;
-    deviceTogglesDesc.enabledToggleCount  = 4;
-    deviceTogglesDesc.disabledToggles     = deviceDisabledToggles;
-    deviceTogglesDesc.disabledToggleCount = 1;
-
-    dev_desc.nextInChain = &deviceTogglesDesc;
-#endif
-
-    ctx->instance.WaitAny(ctx->adapter.RequestDevice(
-                              &dev_desc, wgpu::CallbackMode::AllowSpontaneous,
-                              [ctx](wgpu::RequestDeviceStatus status, wgpu::Device device, wgpu::StringView message) {
-                                  if (status != wgpu::RequestDeviceStatus::Success) {
-                                      GGML_LOG_ERROR("ggml_webgpu: Failed to get a device: %s\n",
-                                                     std::string(message).c_str());
-                                      return;
-                                  }
-                                  ctx->device = std::move(device);
-                              }),
-                          UINT64_MAX);
-    GGML_ASSERT(ctx->device != nullptr);
-
-    // Initialize (compute) queue
-    ctx->queue = ctx->device.GetQueue();
-
-    // Create buffer pool for shader parameters
-    ctx->param_buf_pool.init(ctx->device, WEBGPU_NUM_PARAM_BUFS, WEBGPU_PARAMS_BUF_SIZE_BYTES,
-                             wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::Uniform,
-                             wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::MapWrite);
-
-#ifdef GGML_WEBGPU_GPU_PROFILE
-    // Initialize buffer pool for timestamp queries (profiling)
-    ctx->timestamp_query_buf_pool.init(ctx->device, WEBGPU_NUM_TIMESTAMP_QUERY_BUFS,
-                                       WEBGPU_TIMESTAMP_QUERY_BUF_SIZE_BYTES,
-                                       wgpu::BufferUsage::QueryResolve | wgpu::BufferUsage::CopySrc,
-                                       wgpu::BufferUsage::MapRead | wgpu::BufferUsage::CopyDst);
-#endif
-
-    ctx->set_rows_error_buf_pool.init(ctx->device, WEBGPU_NUM_SET_ROWS_ERROR_BUFS, WEBGPU_SET_ROWS_ERROR_BUF_SIZE_BYTES,
-                                      wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::Storage,
-                                      wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead);
-
-    ggml_webgpu_init_memset_pipeline(ctx);
-    ggml_webgpu_init_mul_mat_pipeline(ctx);
-    ggml_webgpu_init_get_rows_pipeline(ctx);
-    ggml_webgpu_init_cpy_pipeline(ctx);
-    ggml_webgpu_init_add_pipeline(ctx);
-    ggml_webgpu_init_sub_pipeline(ctx);
-    ggml_webgpu_init_mul_pipeline(ctx);
-    ggml_webgpu_init_div_pipeline(ctx);
-    ggml_webgpu_init_rms_norm_pipeline(ctx);
-    ggml_webgpu_init_rope_pipeline(ctx);
-    ggml_webgpu_init_glu_pipeline(ctx);
-    ggml_webgpu_init_scale_pipeline(ctx);
-    ggml_webgpu_init_soft_max_pipeline(ctx);
-
-#ifdef GGML_WEBGPU_DEBUG
-    // Initialize debug buffers
-    ggml_webgpu_create_buffer(ctx->device, ctx->debug_host_buf, WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t),
-                              wgpu::BufferUsage::CopyDst | wgpu::BufferUsage::MapRead, "debug_host_buf");
-    ggml_webgpu_create_buffer(ctx->device, ctx->debug_dev_buf, WEBGPU_DEBUG_BUF_ELEMS * sizeof(uint32_t),
-                              wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc, "debug_dev_buf");
-#endif
+    create_webgpu_device(reg_ctx);
 
     static ggml_backend_webgpu_device_context device_ctx;
-    device_ctx.webgpu_ctx  = ctx;
-    device_ctx.device_name = GGML_WEBGPU_NAME;
-    device_ctx.device_desc = info.description;
-
-    GGML_LOG_INFO(
-        "ggml_webgpu: adapter_info: vendor_id: %u | vendor: %s | architecture: %s | device_id: %u | name: %s | "
-        "device_desc: %s\n",
-        info.vendorID, std::string(info.vendor).c_str(), std::string(info.architecture).c_str(), info.deviceID,
-        std::string(info.device).c_str(), std::string(info.description).c_str());
-
+    device_ctx.device_name            = GGML_WEBGPU_NAME;
+    device_ctx.device_desc            = GGML_WEBGPU_NAME;
+    device_ctx.webgpu_global_ctx      = reg_ctx->webgpu_global_ctx;
     // See GGML Backend Device Interface section
     static ggml_backend_device device = {
         /* .iface   = */ ggml_backend_webgpu_device_i,
@@ -3351,7 +3433,7 @@ static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t
         /* .context = */ &device_ctx,
     };
 
-    WEBGPU_CPU_PROFILE_TOTAL_END(reg_get_device, ctx);
+    WEBGPU_CPU_PROFILE_TOTAL_END(reg_get_device, reg_ctx->webgpu_global_ctx);
     return &device;
 }
 
@@ -3367,10 +3449,7 @@ static const struct ggml_backend_reg_i ggml_backend_webgpu_reg_i = {
 ggml_backend_reg_t ggml_backend_webgpu_reg() {
     WEBGPU_LOG_DEBUG("ggml_backend_webgpu_reg()");
 
-    webgpu_context webgpu_ctx = std::make_shared<webgpu_context_struct>();
-
     static ggml_backend_webgpu_reg_context ctx;
-    ctx.webgpu_ctx   = webgpu_ctx;
     ctx.name         = GGML_WEBGPU_NAME;
     ctx.device_count = 1;
 
@@ -3387,15 +3466,17 @@ ggml_backend_reg_t ggml_backend_webgpu_reg() {
     instance_descriptor.nextInChain        = &instanceTogglesDesc;
 #endif
 
-    webgpu_ctx->instance = wgpu::CreateInstance(&instance_descriptor);
+    wgpu::Instance inst             = wgpu::CreateInstance(&instance_descriptor);
+    ctx.webgpu_global_ctx           = webgpu_global_context(new webgpu_global_context_struct());
+    ctx.webgpu_global_ctx->instance = std::move(inst);
 
 #ifdef __EMSCRIPTEN__
-    if (webgpu_ctx->instance == nullptr) {
+    if (ctx.webgpu_global_ctx->instance == nullptr) {
         GGML_LOG_ERROR("ggml_webgpu: Failed to create WebGPU instance. Make sure either -sASYNCIFY or -sJSPI is set\n");
         return nullptr;
     }
 #endif
-    GGML_ASSERT(webgpu_ctx->instance != nullptr);
+    GGML_ASSERT(ctx.webgpu_global_ctx->instance != nullptr);
 
     static ggml_backend_reg reg = {
         /* .api_version = */ GGML_BACKEND_API_VERSION,
@@ -3408,7 +3489,7 @@ ggml_backend_reg_t ggml_backend_webgpu_reg() {
 ggml_backend_t ggml_backend_webgpu_init(void) {
     ggml_backend_dev_t dev = ggml_backend_reg_dev_get(ggml_backend_webgpu_reg(), 0);
 
-    return ggml_backend_webgpu_device_init(dev, nullptr);
+    return ggml_backend_webgpu_backend_init(dev, nullptr);
 }
 
 GGML_BACKEND_DL_IMPL(ggml_backend_webgpu_reg)
diff --git a/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl b/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl
index de7c132a62..b682216146 100644
--- a/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl
+++ b/ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl
@@ -114,7 +114,7 @@ struct Params {
 #define PARAMS_BINDING 4
 #endif
 
-@group(0) @binding(DST_BINDING) var<storage, read_write> dst: array<f32>;
+@group(0) @binding(DST_BINDING) var<storage, read_write> dst: array<vec4<f32>>;
 @group(0) @binding(PARAMS_BINDING) var<uniform> params: Params;
 
 // Just a very small float value.
@@ -160,14 +160,21 @@ fn calc_softmax_term(kv_idx: u32, q_tile_row: u32, slope: f32) -> f32 {
     return v;
 }
 
+fn load_f32x4(buf: ptr<storage, array<vec4<f32>>, read_write>, scalar_index: u32) -> vec4<f32> {
+    return (*buf)[scalar_index >> 2u];
+}
+
+fn load_kvx4(buf: ptr<storage, array<vec4<KV_TYPE>>, read_write>, scalar_index: u32) -> vec4<KV_TYPE> {
+    return (*buf)[scalar_index >> 2u];
+}
 
 @compute @workgroup_size(WG_SIZE)
 fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
-        @builtin(local_invocation_id) local_id: vec3<u32>,
-        @builtin(subgroup_id) subgroup_id: u32,
-        @builtin(subgroup_size) subgroup_size: u32,
-        @builtin(num_subgroups) num_subgroups: u32,
-        @builtin(subgroup_invocation_id) sg_inv_id: u32) {
+    @builtin(local_invocation_id) local_id: vec3<u32>,
+    @builtin(subgroup_id) subgroup_id: u32,
+    @builtin(subgroup_size) subgroup_size: u32,
+    @builtin(num_subgroups) num_subgroups: u32,
+    @builtin(subgroup_invocation_id) sg_inv_id: u32) {
 
     // initialize row max for online softmax
     for (var i = local_id.x; i < Q_TILE; i += WG_SIZE) {
@@ -231,9 +238,9 @@ fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
 
     for (var kv_tile = 0u; kv_tile < params.seq_len_kv; kv_tile += KV_TILE) {
       // clear inter_shmem to ensure zero-initialized accumulators
-      for (var elem_idx = local_id.x; elem_idx < Q_TILE * KV_TILE; elem_idx += WG_SIZE) {
-          inter_shmem[elem_idx] = 0.0;
-      }
+        for (var elem_idx = local_id.x; elem_idx < Q_TILE * KV_TILE; elem_idx += WG_SIZE) {
+            inter_shmem[elem_idx] = 0.0;
+        }
 
       // load k tile into shared memory
 #if defined(KV_Q4_0)
@@ -309,48 +316,77 @@ fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
 
       // accumulate q block * k block into registers across the entire KV tile
       // TODO: this loop seems to be the current largest bottleneck
-      for (var kv_block = subgroup_id; kv_block < KV_BLOCKS; kv_block += num_subgroups) {
-          let inter_offset = kv_block * SG_MAT_N;
-          var acc: subgroup_matrix_result<f16, SG_MAT_M, SG_MAT_N> = subgroupMatrixLoad<
-              subgroup_matrix_result<f16, SG_MAT_M, SG_MAT_N>>(&inter_shmem, inter_offset, false, KV_TILE);
+      // this bracket exists to scope the lifetime of variables, reducing register pressure
+      {
 #ifdef KV_DIRECT
-          let k_block_row = kv_tile + kv_block * SG_MAT_N;
-          let k_global_offset = k_head_offset + k_block_row * params.stride_k1;
+          let k_block_row = kv_tile + subgroup_id * SG_MAT_N;
+          var k_global_offset = k_head_offset + k_block_row * params.stride_k1;
 #else
-          let k_block_offset = kv_block * SG_MAT_N * HEAD_DIM_QK;
+          var k_block_offset = subgroup_id * SG_MAT_N * HEAD_DIM_QK;
 #endif
-          for (var head_dim_block = 0u; head_dim_block < HEAD_DIM_QK; head_dim_block += SG_MAT_K) {
-              // load q submatrix from shared memory
-              var q_sg_mat: subgroup_matrix_left<f16, SG_MAT_M, SG_MAT_K> = subgroupMatrixLoad<subgroup_matrix_left<f16, SG_MAT_M, SG_MAT_K>>(
-                  &q_shmem,
-                  head_dim_block,
-                  false,
-                  HEAD_DIM_QK
-              );
+          for (var kv_block = subgroup_id; kv_block < KV_BLOCKS; kv_block += num_subgroups) {
+              let inter_offset = kv_block * SG_MAT_N;
+              var acc: subgroup_matrix_result<f16, SG_MAT_M, SG_MAT_N> = subgroupMatrixLoad<subgroup_matrix_result<f16, SG_MAT_M, SG_MAT_N>>(&inter_shmem, inter_offset, false, KV_TILE);
+
+              var q_cur = subgroupMatrixLoad<subgroup_matrix_left<f16, SG_MAT_M, SG_MAT_K>>(&q_shmem, 0u, false, HEAD_DIM_QK);
 
-              // load k submatrix from device or shared memory
 #ifdef KV_DIRECT
-              var k_sg_mat: subgroup_matrix_right<f16, SG_MAT_K, SG_MAT_N> = subgroupMatrixLoad<subgroup_matrix_right<f16, SG_MAT_K, SG_MAT_N>>(
-                  &K,
-                  k_global_offset + head_dim_block,
-                  true,
-                  params.stride_k1
-              );
+              var k_cur = subgroupMatrixLoad<subgroup_matrix_right<f16, SG_MAT_K, SG_MAT_N>>(&K, k_global_offset + 0u, true, params.stride_k1);
 #else
-              var k_sg_mat: subgroup_matrix_right<f16, SG_MAT_K, SG_MAT_N> = subgroupMatrixLoad<subgroup_matrix_right<f16, SG_MAT_K, SG_MAT_N>>(
-                  &kv_shmem,
-                  k_block_offset + head_dim_block,
-                  true,
-                  HEAD_DIM_QK
-              );
+              var k_cur = subgroupMatrixLoad<subgroup_matrix_right<f16, SG_MAT_K, SG_MAT_N>>(&kv_shmem, k_block_offset + 0u, true, HEAD_DIM_QK);
 #endif
-              acc = subgroupMatrixMultiplyAccumulate(q_sg_mat, k_sg_mat, acc);
+
+              var t: u32 = 1u;
+              for (; t + 1u < HEAD_DIM_QK / SG_MAT_K; t += 2u) {
+                  let h0 = t * SG_MAT_K;
+                  var q0 = subgroupMatrixLoad<subgroup_matrix_left<f16, SG_MAT_M, SG_MAT_K>>(&q_shmem, h0, false, HEAD_DIM_QK);
+#ifdef KV_DIRECT
+                  var k0 = subgroupMatrixLoad<subgroup_matrix_right<f16, SG_MAT_K, SG_MAT_N>>(&K, k_global_offset + h0, true, params.stride_k1);
+#else
+                  var k0 = subgroupMatrixLoad<subgroup_matrix_right<f16, SG_MAT_K, SG_MAT_N>>(&kv_shmem, k_block_offset + h0, true, HEAD_DIM_QK);
+#endif
+                  acc = subgroupMatrixMultiplyAccumulate(q_cur, k_cur, acc);
+                  q_cur = q0;
+                  k_cur = k0;
+
+                  let h1 = (t + 1u) * SG_MAT_K;
+                  var q1g = subgroupMatrixLoad<subgroup_matrix_left<f16, SG_MAT_M, SG_MAT_K>>(&q_shmem, h1, false, HEAD_DIM_QK);
+#ifdef KV_DIRECT
+                  var k1g = subgroupMatrixLoad<subgroup_matrix_right<f16, SG_MAT_K, SG_MAT_N>>(&K, k_global_offset + h1, true, params.stride_k1);
+#else
+                  var k1g = subgroupMatrixLoad<subgroup_matrix_right<f16, SG_MAT_K, SG_MAT_N>>(&kv_shmem, k_block_offset + h1, true, HEAD_DIM_QK);
+#endif
+                  acc = subgroupMatrixMultiplyAccumulate(q_cur, k_cur, acc);
+                  q_cur = q1g;
+                  k_cur = k1g;
+              }
+
+              // handle odd tail
+              if (t < HEAD_DIM_QK / SG_MAT_K) {
+                  let h = t * SG_MAT_K;
+                  var qn = subgroupMatrixLoad<subgroup_matrix_left<f16, SG_MAT_M, SG_MAT_K>>(&q_shmem, h, false, HEAD_DIM_QK);
+#ifdef KV_DIRECT
+                  var kn = subgroupMatrixLoad<subgroup_matrix_right<f16, SG_MAT_K, SG_MAT_N>>(&K, k_global_offset + h, true, params.stride_k1);
+#else
+                  var kn = subgroupMatrixLoad<subgroup_matrix_right<f16, SG_MAT_K, SG_MAT_N>>(&kv_shmem, k_block_offset + h, true, HEAD_DIM_QK);
+#endif
+                  acc = subgroupMatrixMultiplyAccumulate(q_cur, k_cur, acc);
+                  q_cur = qn;
+                  k_cur = kn;
+              }
+
+              acc = subgroupMatrixMultiplyAccumulate(q_cur, k_cur, acc);
+
+#ifdef KV_DIRECT
+              k_global_offset += num_subgroups * SG_MAT_N * params.stride_k1;
+#else
+              k_block_offset += num_subgroups * SG_MAT_N * HEAD_DIM_QK;
+#endif
+              subgroupMatrixStore(&inter_shmem, inter_offset, acc, false, KV_TILE);
           }
-
-          // store acc to shared memory for softmax (S matrix from paper)
-          subgroupMatrixStore(&inter_shmem, inter_offset, acc, false, KV_TILE);
       }
 
+
 #ifdef MASK
       // load mask tile into shared memory for this KV block
       // TODO: optimize and skip if mask is -INF for the entire tile
@@ -495,7 +531,6 @@ fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
                   false,
                   HEAD_DIM_V
               );
-
               for (var kv_block = 0u; kv_block < KV_BLOCKS; kv_block++) {
                   let p_offset = kv_block * SG_MAT_N;
                   var p_sg_mat: subgroup_matrix_left<f16, SG_MAT_M, SG_MAT_K> = subgroupMatrixLoad<subgroup_matrix_left<f16, SG_MAT_M, SG_MAT_K>>(
@@ -527,11 +562,9 @@ fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
                   // O += P * V
                   o_sg_mat = subgroupMatrixMultiplyAccumulate(p_sg_mat, v_sg_mat, o_sg_mat);
               }
-
               // store O back to shared memory
               subgroupMatrixStore(&o_shmem, head_dim_block, o_sg_mat, false, HEAD_DIM_V);
       }
-
       workgroupBarrier();
     }
 
@@ -566,26 +599,38 @@ fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
                 o_shmem[idx] = f16(val);
             }
     }
-
     workgroupBarrier();
 #endif
-
-    // write output back to global memory
     for (var q_tile_row = subgroup_id;
-         q_tile_row < Q_TILE;
-         q_tile_row += num_subgroups) {
-            let global_q_row = q_row_start + q_tile_row;
-            if (global_q_row >= params.seq_len_q) {
-                break;
-            }
+        q_tile_row < Q_TILE;
+        q_tile_row += num_subgroups) {
 
-            let exp_sum = exp_sum_shmem[q_tile_row];
-            let scale = select(0.0, 1.0 / exp_sum, exp_sum != 0);
+        let global_q_row = q_row_start + q_tile_row;
+        if (global_q_row >= params.seq_len_q) { break; }
 
-            for (var elem_idx = sg_inv_id; elem_idx < HEAD_DIM_V; elem_idx += subgroup_size) {
-                let o_val = o_shmem[q_tile_row * HEAD_DIM_V + elem_idx];
-                let scaled = f32(o_val) * scale;
-                dst[dst_global_offset + q_tile_row * dst2_stride + elem_idx] = scaled;
-            }
+        let exp_sum = exp_sum_shmem[q_tile_row];
+        let scale = select(0.0, 1.0 / exp_sum, exp_sum != 0.0);
+
+        let row_base: u32 = dst_global_offset + q_tile_row * dst2_stride;
+
+        for (var elem_base = sg_inv_id * 4u;
+            elem_base < HEAD_DIM_V;
+            elem_base += subgroup_size * 4u) {
+
+            let i0 = q_tile_row * HEAD_DIM_V + (elem_base + 0u);
+            let i1 = q_tile_row * HEAD_DIM_V + (elem_base + 1u);
+            let i2 = q_tile_row * HEAD_DIM_V + (elem_base + 2u);
+            let i3 = q_tile_row * HEAD_DIM_V + (elem_base + 3u);
+
+            let v = vec4<f32>(
+                f32(o_shmem[i0]) * scale,
+                f32(o_shmem[i1]) * scale,
+                f32(o_shmem[i2]) * scale,
+                f32(o_shmem[i3]) * scale
+            );
+
+            let dst_vec_index: u32 = (row_base + elem_base) >> 2u;
+            dst[dst_vec_index] = v;
+        }
     }
 }
diff --git a/ggml/src/ggml-zdnn/ggml-zdnn.cpp b/ggml/src/ggml-zdnn/ggml-zdnn.cpp
index edbeb8eef2..9b6938abf7 100644
--- a/ggml/src/ggml-zdnn/ggml-zdnn.cpp
+++ b/ggml/src/ggml-zdnn/ggml-zdnn.cpp
@@ -58,6 +58,10 @@ static enum ggml_status ggml_zdnn_graph_compute(ggml_backend_t backend, ggml_cgr
             continue;
         }
 
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
+
         bool ok = ggml_zdnn_compute_forward(ctx, node);
         if (!ok) {
             GGML_LOG_ERROR("%s: unsupported op %s (%s)\n",
@@ -368,7 +372,8 @@ static size_t ggml_backend_zdnn_buffer_type_get_alignment(ggml_backend_buffer_ty
 }
 
 static bool ggml_backend_zdnn_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
-    return true;
+    /* while it resides in host memory, additional transformation is needed */
+    return false;
 
     GGML_UNUSED(buft);
 }
diff --git a/ggml/src/ggml-zendnn/CMakeLists.txt b/ggml/src/ggml-zendnn/CMakeLists.txt
index bdbfc74369..f5cf6eedd3 100644
--- a/ggml/src/ggml-zendnn/CMakeLists.txt
+++ b/ggml/src/ggml-zendnn/CMakeLists.txt
@@ -21,7 +21,7 @@ if (NOT ZENDNN_ROOT OR ZENDNN_ROOT STREQUAL "" OR ZENDNN_ROOT STREQUAL "OFF")
     ExternalProject_Add(
         zendnn
         GIT_REPOSITORY https://github.com/amd/ZenDNN.git
-        GIT_TAG zendnnl
+        GIT_TAG 21ce8f7879c86bf3637f707fae6f29e0951db5fe
         PREFIX      ${ZENDNN_PREFIX}
         SOURCE_DIR  ${ZENDNN_SOURCE_DIR}
         BINARY_DIR  ${ZENDNN_BUILD_DIR}
diff --git a/ggml/src/ggml-zendnn/ggml-zendnn.cpp b/ggml/src/ggml-zendnn/ggml-zendnn.cpp
index fd07f983da..551c15bb4a 100644
--- a/ggml/src/ggml-zendnn/ggml-zendnn.cpp
+++ b/ggml/src/ggml-zendnn/ggml-zendnn.cpp
@@ -2,7 +2,6 @@
 
 #include "ggml-backend-impl.h"
 #include "ggml-impl.h"
-#include "ggml-cpu.h"
 #include "zendnnl.hpp"
 
 #include <cstring>
@@ -122,8 +121,8 @@ static void ggml_zendnn_compute_forward_mul_mat(
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
-    ggml_type         const vec_dot_type = ggml_get_type_traits_cpu(src0->type)->vec_dot_type;
-    ggml_from_float_t const from_float = ggml_get_type_traits_cpu(vec_dot_type)->from_float;
+    ggml_type         const vec_dot_type = src0->type;
+    ggml_from_float_t const from_float = ggml_get_type_traits(vec_dot_type)->from_float_ref;
 
     GGML_ASSERT(ne0 == ne01);
     GGML_ASSERT(ne1 == ne11);
@@ -211,6 +210,10 @@ static ggml_status ggml_backend_zendnn_graph_compute(ggml_backend_t backend, ggm
     for (int i = 0; i < cgraph->n_nodes; i++) {
         struct ggml_tensor * node = cgraph->nodes[i];
 
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
+
         switch (node->op) {
             case GGML_OP_MUL_MAT:
                 ggml_zendnn_compute_forward_mul_mat(ctx, node);
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index c75fe7d271..1725ad1654 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -3441,7 +3441,8 @@ struct ggml_tensor * ggml_cast(
 
     result->op     = GGML_OP_CPY;
     result->src[0] = a;
-    result->src[1] = result;
+    result->src[1] = result; // note: this self-reference might seem redundant, but it's actually needed by some
+                             //       backends for consistency with ggml_cpy_impl() above
 
     return result;
 }
@@ -6725,20 +6726,35 @@ static void ggml_compute_backward(
     GGML_ASSERT(!src2_needs_grads || ggml_are_same_shape(src2, cgraph->grads[isrc2]));
 }
 
-static size_t ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * node) {
-    // check if already visited
-    size_t node_hash_pos = ggml_hash_find(&cgraph->visited_hash_set, node);
+static size_t ggml_visit_parents_graph(struct ggml_cgraph * cgraph, struct ggml_tensor * node, bool compute) {
+    if (node->op != GGML_OP_NONE && compute) {
+        node->flags |= GGML_TENSOR_FLAG_COMPUTE;
+    }
+
+    const size_t node_hash_pos = ggml_hash_find(&cgraph->visited_hash_set, node);
     GGML_ASSERT(node_hash_pos != GGML_HASHSET_FULL);
-    if (!ggml_bitset_get(cgraph->visited_hash_set.used, node_hash_pos)) {
-        // This is the first time we see this node in the current graph.
-        cgraph->visited_hash_set.keys[node_hash_pos] = node;
-        ggml_bitset_set(cgraph->visited_hash_set.used, node_hash_pos);
-        cgraph->use_counts[node_hash_pos] = 0;
-    } else {
+
+    if (ggml_bitset_get(cgraph->visited_hash_set.used, node_hash_pos)) {
         // already visited
+
+        if (compute) {
+            // update the compute flag regardless
+            for (int i = 0; i < GGML_MAX_SRC; ++i) {
+                struct ggml_tensor * src = node->src[i];
+                if (src && ((src->flags & GGML_TENSOR_FLAG_COMPUTE) == 0)) {
+                    ggml_visit_parents_graph(cgraph, src, true);
+                }
+            }
+        }
+
         return node_hash_pos;
     }
 
+    // This is the first time we see this node in the current graph.
+    cgraph->visited_hash_set.keys[node_hash_pos] = node;
+    ggml_bitset_set(cgraph->visited_hash_set.used, node_hash_pos);
+    cgraph->use_counts[node_hash_pos] = 0;
+
     for (int i = 0; i < GGML_MAX_SRC; ++i) {
         const int k =
             (cgraph->order == GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT) ? i :
@@ -6747,7 +6763,7 @@ static size_t ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor
 
         struct ggml_tensor * src = node->src[k];
         if (src) {
-            size_t src_hash_pos = ggml_visit_parents(cgraph, src);
+            const size_t src_hash_pos = ggml_visit_parents_graph(cgraph, src, compute);
 
             // Update the use count for this operand.
             cgraph->use_counts[src_hash_pos]++;
@@ -6778,17 +6794,17 @@ static size_t ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor
     return node_hash_pos;
 }
 
-static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor, bool expand) {
+static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor, bool expand, bool compute) {
     if (!expand) {
         // TODO: this branch isn't accessible anymore, maybe move this to ggml_build_forward_expand
         ggml_graph_clear(cgraph);
     }
 
-    const int n0 = cgraph->n_nodes;
+    const int n_old = cgraph->n_nodes;
 
-    ggml_visit_parents(cgraph, tensor);
+    ggml_visit_parents_graph(cgraph, tensor, compute);
 
-    const int n_new = cgraph->n_nodes - n0;
+    const int n_new = cgraph->n_nodes - n_old;
     GGML_PRINT_DEBUG("%s: visited %d new nodes\n", __func__, n_new);
 
     if (n_new > 0) {
@@ -6797,8 +6813,22 @@ static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_ten
     }
 }
 
+struct ggml_tensor * ggml_build_forward_select(
+        struct ggml_cgraph  * cgraph,
+        struct ggml_tensor ** tensors,
+        int                   n_tensors,
+        int                   idx) {
+    GGML_ASSERT(idx >= 0 && idx < n_tensors);
+
+    for (int i = 0; i < n_tensors; i++) {
+        ggml_build_forward_impl(cgraph, tensors[i], true, i == idx ? true : false);
+    }
+
+    return tensors[idx];
+}
+
 void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor) {
-    ggml_build_forward_impl(cgraph, tensor, true);
+    ggml_build_forward_impl(cgraph, tensor, true, true);
 }
 
 void ggml_build_backward_expand(
@@ -7229,6 +7259,10 @@ bool ggml_can_fuse_subgraph_ext(const struct ggml_cgraph * cgraph,
             return false;
         }
 
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            return false;
+        }
+
         if (ggml_node_list_find_tensor(cgraph, outputs, num_outputs, node) != -1) {
             continue;
         }
@@ -7310,7 +7344,7 @@ static void ggml_graph_dump_dot_leaf_edge(FILE * fp, struct ggml_tensor * node,
             label);
 }
 
-void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename) {
+void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * cgraph, const char * filename) {
     char color[16];
 
     FILE * fp = ggml_fopen(filename, "w");
@@ -7331,7 +7365,7 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
         if (node->flags & GGML_TENSOR_FLAG_PARAM) {
             snprintf(color, sizeof(color), "yellow");
         } else if (grad) {
-            if (ggml_graph_find(gf, node)) {
+            if (ggml_graph_find(cgraph, node)) {
                 snprintf(color, sizeof(color), "green");
             } else {
                 snprintf(color, sizeof(color), "lightblue");
diff --git a/ggml/src/gguf.cpp b/ggml/src/gguf.cpp
index b165d8bdc6..ed0d7f2cae 100644
--- a/ggml/src/gguf.cpp
+++ b/ggml/src/gguf.cpp
@@ -585,6 +585,14 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
                 break;
             }
 
+            // check that the size of the tensor in bytes is representable
+            if (ok && uint64_t(ggml_nelements(&info.t)/ggml_blck_size(info.t.type)) > SIZE_MAX/ggml_type_size(info.t.type)) {
+                GGML_LOG_ERROR("%s: tensor '%s' with shape (%" PRIi64 ", %" PRIi64 ", %" PRIi64 ", %" PRIi64 ") has a size in bytes > %zu\n",
+                    __func__, info.t.name, info.t.ne[0], info.t.ne[1], info.t.ne[2], info.t.ne[3], SIZE_MAX);
+                ok = false;
+                break;
+            }
+
             // calculate byte offsets given the tensor shape and type
             info.t.nb[0] = type_size;
             info.t.nb[1] = info.t.nb[0]*(info.t.ne[0]/blck_size);
@@ -734,7 +742,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
     FILE * file = ggml_fopen(fname, "rb");
 
     if (!file) {
-        GGML_LOG_ERROR("%s: failed to open GGUF file '%s'\n", __func__, fname);
+        GGML_LOG_ERROR("%s: failed to open GGUF file '%s' (%s)\n", __func__, fname, strerror(errno));
         return nullptr;
     }
 
diff --git a/include/llama.h b/include/llama.h
index 280745713e..bf4e28a8be 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -309,7 +309,7 @@ extern "C" {
         // Keep the booleans together to avoid misalignment during copy-by-value.
         bool vocab_only;      // only load the vocabulary, no weights
         bool use_mmap;        // use mmap if possible
-        bool use_direct_io;   // use direct io, takes precedence over use_mmap
+        bool use_direct_io;   // use direct io, takes precedence over use_mmap when supported
         bool use_mlock;       // force system to keep model in RAM
         bool check_tensors;   // validate model tensor data
         bool use_extra_bufts; // use extra buffer types (used for weight repacking)
@@ -489,6 +489,7 @@ extern "C" {
     //   - returns true if the parameters could be successfully modified to fit device memory
     //   - this function is NOT thread safe because it modifies the global llama logger state
     //   - only parameters that have the same value as in llama_default_model_params are modified
+    //     with the exception of the context size which is modified if and only if equal to 0
     LLAMA_API enum llama_params_fit_status llama_params_fit(
                                    const char   * path_model,
                     struct llama_model_params   * mparams,
@@ -1475,12 +1476,12 @@ extern "C" {
     /// @details Build a split GGUF final path for this chunk.
     ///          llama_split_path(split_path, sizeof(split_path), "/models/ggml-model-q4_0", 2, 4) => split_path = "/models/ggml-model-q4_0-00002-of-00004.gguf"
     //  Returns the split_path length.
-    LLAMA_API int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count);
+    LLAMA_API int32_t llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int32_t split_no, int32_t split_count);
 
     /// @details Extract the path prefix from the split_path if and only if the split_no and split_count match.
     ///          llama_split_prefix(split_prefix, 64, "/models/ggml-model-q4_0-00002-of-00004.gguf", 2, 4) => split_prefix = "/models/ggml-model-q4_0"
     //  Returns the split_prefix length.
-    LLAMA_API int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count);
+    LLAMA_API int32_t llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int32_t split_no, int32_t split_count);
 
     // Print system information
     LLAMA_API const char * llama_print_system_info(void);
diff --git a/models/templates/upstage-Solar-Open-100B.jinja b/models/templates/upstage-Solar-Open-100B.jinja
new file mode 100644
index 0000000000..13268c1a84
--- /dev/null
+++ b/models/templates/upstage-Solar-Open-100B.jinja
@@ -0,0 +1,156 @@
+{#- ======== Template Parameters ========  #}
+{%- set add_generation_prompt = add_generation_prompt if add_generation_prompt is defined else true %}
+{%- set default_system_prompt = default_system_prompt if default_system_prompt is defined else true %}
+{%- set reasoning_effort = reasoning_effort if reasoning_effort is defined else "high" %}
+{%- set think_render_option = think_render_option if think_render_option is defined else "lastthink" %}
+
+{#- ======== System Block State ========  #}
+{%- set sys_ns = namespace(is_first_block=true) -%}
+
+{#- ======== Find last user message index ========  #}
+{%- set last_user_idx = namespace(value=-1) -%}
+{%- for message in messages -%}
+    {%- if message.role == 'user' -%}
+        {%- set last_user_idx.value = loop.index0 -%}
+    {%- endif -%}
+{%- endfor -%}
+
+{#- ======== System messages renderers ========  #}
+{%- macro render_system_message(user_system_messages) %}
+    {%- if default_system_prompt %}
+        {%- if not sys_ns.is_first_block %}{{- "\n\n" }}{%- endif %}
+        {%- set sys_ns.is_first_block = false %}
+        {{- "## Provider System Prompt\n\nYou are Solar Open 100B, a large language model trained by Upstage AI, a Korean startup. Your knowledge cutoff is 2025-07. The current date is " + strftime_now("%Y-%m-%d") + "." }}
+    {%- endif -%}
+    {%- if user_system_messages %}
+        {%- if not sys_ns.is_first_block %}{{- "\n\n" }}{%- endif %}
+        {%- set sys_ns.is_first_block = false %}
+        {{- "## System Prompt" }}
+        {%- for system_message in user_system_messages %}
+            {{- "\n\n" }}
+            {{- system_message }}
+        {%- endfor %}
+    {%- endif -%}
+{%- endmacro %}
+
+{%- macro render_tool_instruction(tools) %}
+    {%- if not sys_ns.is_first_block %}{{- "\n\n" }}{%- endif %}
+    {%- set sys_ns.is_first_block = false %}
+    {{- "## Tools\n\n### Tool Call Instruction" }}
+    {{- "\nYou may invoke one or more tools to assist with the user's query. Available tools are provided in JSON Schema format: <|tools:begin|><|tool:begin|><tools-json-object><|tool:end|>...<|tools:end|>\n" }}
+    {{- "\n### Available Tools\n" }}
+    {{- "<|tools:begin|>" }}
+    {%- for tool in tools %}
+        {{- "<|tool:begin|>" }}
+        {{- tool.function | tojson }}
+        {{- "<|tool:end|>" }}
+    {%- endfor %}
+    {{- "<|tools:end|>\n" }}
+    {{- "\n### Tool Call Format\n" }}
+    {{- "For each tool call, return a JSON object with the following structure, enclosed within <|tool_call:begin|> and <|tool_call:end|> tags: \n<|tool_call:begin|><tool-call-id><|tool_call:name|><tool-name><|tool_call:args|><args-json-object><|tool_call:end|>\n" }}
+    {{- "- The <tool-call-id> must be a randomly generated string consisting of 10 lowercase letters (a-z) and/or digits (0-9) (e.g., a1b2c3d4e5)\n" }}
+    {{- "\n### Tool Response Format\n" }}
+    {{- "Each tool is responded by `tool` with the following structure:\n<|tool_response:id|><tool-call-id><|tool_response:name|><tool-name><|tool_response:result|><results><|tool_response:end|>\n" }}
+    {{- "- Ensure the <tool-call-id> matches the corresponding tool call" -}}
+{%- endmacro %}
+
+{%- macro render_json_response_format_instruction(response_format) %}
+    {%- if not sys_ns.is_first_block %}{{- "\n\n" }}{%- endif %}
+    {%- set sys_ns.is_first_block = false %}
+    {{- "## Output Format Constraint" }}
+    {{- "\n\nYour final response should follow the JSON schema: \n[Start of schema]" }}
+    {{- response_format }}
+    {{- "\n[End of schema]\nPlease ensure your answers adhere to this format and do not contain any unnecessary text." }}
+{%- endmacro %}
+
+{%- macro get_tool_name(messages, tool_call_id) %}
+    {%- for msg in messages -%}
+        {%- if msg.role == 'assistant' and msg.tool_calls -%}
+            {%- for tool_call in msg.tool_calls -%}
+                {%- if tool_call.id == tool_call_id -%}
+                    {{- tool_call.function.name }}
+                {%- endif -%}
+            {%- endfor -%}
+        {%- endif -%}
+    {%- endfor -%}
+{%- endmacro %}
+
+{%- macro render_tool_arguments(tool_arguments) %}
+    {%- if tool_arguments is mapping -%}
+        {{- tool_arguments | tojson }}
+    {%- else -%}
+        {{- tool_arguments }}
+    {%- endif -%}
+{%- endmacro %}
+
+{#- ======== Render system message ========  #}
+{%- set ns = namespace(system_messages=[]) -%}
+{%- for message in messages -%}
+    {%- if message.role == 'system' -%}
+        {%- set ns.system_messages = ns.system_messages + [message.content] -%}
+    {%- endif -%}
+{%- endfor -%}
+
+{%- if ns.system_messages or default_system_prompt or tools or response_format -%}
+    {{- "<|begin|>system<|content|>" }}
+        {{- render_system_message(ns.system_messages) }}
+        {%- if tools -%}
+            {{- render_tool_instruction(tools) }}
+        {%- endif %}
+        {%- if response_format -%}
+            {{- render_json_response_format_instruction(response_format) }}
+        {%- endif %}
+    {{- "<|end|>" }}
+{%- endif -%}
+
+{#- ======== Render main messages ========  #}
+{%- for message in messages -%}
+    {%- if message.role == 'user' -%}
+         {{- "<|begin|>user<|content|>" + message.content + "<|end|>" }}
+    {%- elif message.role == 'tool' -%}
+        {%- set prev_is_tool = loop.index0 > 0 and messages[loop.index0 - 1].role == 'tool' -%}
+        {%- set next_is_tool = loop.index0 < (messages | length - 1) and messages[loop.index0 + 1].role == 'tool' -%}
+        {%- if not prev_is_tool -%}
+            {{- "<|begin|>tool<|tool_response|>" }}
+        {%- endif -%}
+        {{- "<|tool_response:begin|>" + message.tool_call_id + "<|tool_response:name|>" }}
+        {{- get_tool_name(messages, message.tool_call_id) }}
+        {{- "<|tool_response:result|>" }}
+        {{- message.content }}
+        {{- "<|tool_response:end|>" }}
+        {%- if not next_is_tool -%}
+            {{- "<|end|>" }}
+        {%- endif -%}
+    {%- elif message.role == 'assistant' -%}
+        {#- ======== Assistant Thinking ========  #}
+        {%- if think_render_option == "all" -%}
+            {%- if message.reasoning -%}
+                {{- "<|begin|>assistant<|think|>" + message.reasoning + "<|end|>" }}
+            {%- endif -%}
+        {%- elif think_render_option == "lastthink" -%}
+            {%- if message.reasoning and loop.index0 > last_user_idx.value -%}
+                {{- "<|begin|>assistant<|think|>" + message.reasoning + "<|end|>" }}
+            {%- endif -%}
+        {%- endif -%}
+
+        {#- ======== Assistant Messages ========  #}
+        {%- if message.tool_calls -%}
+            {{- "<|begin|>assistant<|tool_calls|>" }}
+            {%- for tool_call in message.tool_calls -%}
+                {{- "<|tool_call:begin|>" + tool_call.id +"<|tool_call:name|>" + tool_call.function.name + "<|tool_call:args|>" }}
+                {{- render_tool_arguments(tool_call.function.arguments) }}
+                {{- "<|tool_call:end|>" }}
+            {%- endfor -%}
+            {{- "<|calls|>" }}
+        {%- else -%}
+            {{- "<|begin|>assistant<|content|>" + message.content + "<|end|>" }}
+        {%- endif -%}
+    {%- endif -%}
+{%- endfor -%}
+
+{%- if add_generation_prompt -%}
+    {%- if reasoning_effort in ["low", "minimal"] -%}
+        {{- "<|begin|>assistant<|think|><|end|>" }}
+    {%- endif -%}
+    {{- "<|begin|>assistant" }}
+{%- endif -%}
diff --git a/requirements/requirements-tool_bench.txt b/requirements/requirements-tool_bench.txt
index f7912aff72..3bb74fb9d0 100644
--- a/requirements/requirements-tool_bench.txt
+++ b/requirements/requirements-tool_bench.txt
@@ -3,7 +3,7 @@ pytest~=8.3.3
 huggingface_hub>=0.34.0,<1.0
 matplotlib~=3.10.0
 numpy~=1.26.4
-openai~=1.55.3
+openai~=2.14.0
 pandas~=2.2.3
 prometheus-client~=0.20.0
 requests~=2.32.3
diff --git a/scripts/compare-llama-bench.py b/scripts/compare-llama-bench.py
index c45c83fdb5..9541b89eb9 100755
--- a/scripts/compare-llama-bench.py
+++ b/scripts/compare-llama-bench.py
@@ -29,7 +29,7 @@ LLAMA_BENCH_DB_FIELDS = [
     "cpu_mask",     "cpu_strict",   "poll",           "type_k",     "type_v",       "n_gpu_layers",
     "split_mode",   "main_gpu",     "no_kv_offload",  "flash_attn", "tensor_split", "tensor_buft_overrides",
     "use_mmap",     "embeddings",   "no_op_offload",  "n_prompt",   "n_gen",        "n_depth",
-    "test_time",    "avg_ns",       "stddev_ns",      "avg_ts",     "stddev_ts",
+    "test_time",    "avg_ns",       "stddev_ns",      "avg_ts",     "stddev_ts",    "n_cpu_moe"
 ]
 
 LLAMA_BENCH_DB_TYPES = [
@@ -38,7 +38,7 @@ LLAMA_BENCH_DB_TYPES = [
     "TEXT",    "INTEGER", "INTEGER", "TEXT",    "TEXT",    "INTEGER",
     "TEXT",    "INTEGER", "INTEGER", "INTEGER", "TEXT",    "TEXT",
     "INTEGER", "INTEGER", "INTEGER", "INTEGER", "INTEGER", "INTEGER",
-    "TEXT",    "INTEGER", "INTEGER", "REAL",    "REAL",
+    "TEXT",    "INTEGER", "INTEGER", "REAL",    "REAL",    "INTEGER",
 ]
 
 # All test-backend-ops SQL fields
@@ -59,7 +59,7 @@ assert len(TEST_BACKEND_OPS_DB_FIELDS) == len(TEST_BACKEND_OPS_DB_TYPES)
 
 # Properties by which to differentiate results per commit for llama-bench:
 LLAMA_BENCH_KEY_PROPERTIES = [
-    "cpu_info", "gpu_info", "backends", "n_gpu_layers", "tensor_buft_overrides", "model_filename", "model_type",
+    "cpu_info", "gpu_info", "backends", "n_gpu_layers", "n_cpu_moe", "tensor_buft_overrides", "model_filename", "model_type",
     "n_batch", "n_ubatch", "embeddings", "cpu_mask", "cpu_strict", "poll", "n_threads", "type_k", "type_v",
     "use_mmap", "no_kv_offload", "split_mode", "main_gpu", "tensor_split", "flash_attn", "n_prompt", "n_gen", "n_depth"
 ]
diff --git a/scripts/snapdragon/windows/run-bench.ps1 b/scripts/snapdragon/windows/run-bench.ps1
new file mode 100644
index 0000000000..21fd063ebe
--- /dev/null
+++ b/scripts/snapdragon/windows/run-bench.ps1
@@ -0,0 +1,40 @@
+
+#!/usr/bin/env pwsh
+
+# Basedir on device
+$basedir=".\pkg-snapdragon"
+
+$cli_opts=$args
+
+$model="Llama-3.2-3B-Instruct-Q4_0.gguf"
+if ($null -ne $env:M) {
+    $model=$env:M
+}
+
+$device="HTP0"
+if ($null -ne $env:D) {
+    $device=$env:D
+}
+
+if ($null -ne $env:V) {
+    $env:GGML_HEXAGON_VERBOSE=$env:V
+}
+
+if ($null -ne $env:OPMASK) {
+    $env:GGML_HEXAGON_OPMASK=$env:OPMASK
+}
+
+if ($null -ne $env:NHVX) {
+    $env:GGML_HEXAGON_NHVX=$env:NHVX
+}
+
+if ($null -ne $env:NDEV) {
+    $env:GGML_HEXAGON_NDEV=$env:NDEV
+}
+
+$env:ADSP_LIBRARY_PATH="$basedir\lib"
+
+& "$basedir\bin\llama-bench.exe" `
+    --mmap 0 -m $basedir\..\..\gguf\$model `
+    --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1 `
+    --batch-size 128 -ngl 99 --device $device $cli_opts
diff --git a/scripts/snapdragon/windows/run-cli.ps1 b/scripts/snapdragon/windows/run-cli.ps1
new file mode 100644
index 0000000000..b13161aa63
--- /dev/null
+++ b/scripts/snapdragon/windows/run-cli.ps1
@@ -0,0 +1,53 @@
+
+#!/usr/bin/env pwsh
+
+# Basedir on device
+$basedir=".\pkg-snapdragon"
+
+$cli_opts=$args
+
+$model="Llama-3.2-3B-Instruct-Q4_0.gguf"
+if ($null -ne $env:M) {
+    $model=$env:M
+}
+
+$device="HTP0"
+if ($null -ne $env:D) {
+    $device=$env:D
+}
+
+if ($null -ne $env:V) {
+    $env:GGML_HEXAGON_VERBOSE=$env:V
+}
+
+if ($null -ne $env:E) {
+    $env:GGML_HEXAGON_EXPERIMENTAL=$env:E
+}
+
+if ($null -ne $env:SCHED) {
+    $env:GGML_SCHED_DEBUG=$env:SCHED; $cli_opts="$cli_opts -v"
+}
+
+if ($null -ne $env:PROF) {
+    $env:GGML_HEXAGON_PROFILE=$env:PROF; $env:GGML_HEXAGON_OPSYNC=1
+}
+
+if ($null -ne $env:OPMASK) {
+    $env:GGML_HEXAGON_OPMASK=$env:OPMASK
+}
+
+if ($null -ne $env:NHVX) {
+    $env:GGML_HEXAGON_NHVX=$env:NHVX
+}
+
+if ($null -ne $env:NDEV) {
+    $env:GGML_HEXAGON_NDEV=$env:NDEV
+}
+
+$env:ADSP_LIBRARY_PATH="$basedir\lib"
+
+& "$basedir\bin\llama-completion.exe" `
+    --no-mmap -no-cnv -m $basedir\..\..\gguf\$model `
+    --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1 `
+    --ctx-size 8192 --batch-size 128 -ctk q8_0 -ctv q8_0 -fa on `
+    -ngl 99 --device $device $cli_opts
diff --git a/scripts/snapdragon/windows/run-tool.ps1 b/scripts/snapdragon/windows/run-tool.ps1
new file mode 100644
index 0000000000..70094af9bc
--- /dev/null
+++ b/scripts/snapdragon/windows/run-tool.ps1
@@ -0,0 +1,56 @@
+
+#!/usr/bin/env pwsh
+
+# Basedir on device
+$basedir=".\pkg-snapdragon"
+
+if ($args.Count -eq 0) {
+    Write-Host "No arguments provided.Expected the tool and argument to run."
+    exit -1
+}
+
+$tool=$args[0]
+$cli_opts=@()
+
+if ($args.Count -gt 1) {
+    $cli_opts=$args[1..($args.Count - 1)]
+    $remainingArgs = $args[1..($args.Count - 1)]
+}
+
+$device="HTP0"
+if ($null -ne $env:D) {
+    $device=$env:D
+}
+
+if ($null -ne $env:V) {
+    $env:GGML_HEXAGON_VERBOSE=$env:V
+}
+
+if ($null -ne $env:E) {
+    $env:GGML_HEXAGON_EXPERIMENTAL=$env:E
+}
+
+if ($null -ne $env:SCHED) {
+    $env:GGML_SCHED_DEBUG=$env:SCHED; $cli_opts="$cli_opts -v"
+}
+
+if ($null -ne $env:PROF) {
+    $env:GGML_HEXAGON_PROFILE=$env:PROF; $env:GGML_HEXAGON_OPSYNC=1
+}
+
+if ($null -ne $env:OPMASK) {
+    $env:GGML_HEXAGON_OPMASK=$env:OPMASK
+}
+
+if ($null -ne $env:NHVX) {
+    $env:GGML_HEXAGON_NHVX=$env:NHVX
+}
+
+if ($null -ne $env:NDEV) {
+    $env:GGML_HEXAGON_NDEV=$env:NDEV
+}
+
+$env:ADSP_LIBRARY_PATH="$basedir\lib"
+
+& "$basedir\bin\$tool" `
+    $cli_opts
diff --git a/scripts/snapdragon/windows/setup-build.ps1 b/scripts/snapdragon/windows/setup-build.ps1
new file mode 100644
index 0000000000..0f3244cc9d
--- /dev/null
+++ b/scripts/snapdragon/windows/setup-build.ps1
@@ -0,0 +1,105 @@
+# Requires Run as Administrator is NOT strictly necessary for User-scope env vars,
+# but recommended for creating directories in C:\ root if permissions are restricted.
+
+$ErrorActionPreference = "Stop"
+
+# --- Configuration ---
+$BaseDir = "C:\Qualcomm"
+
+# SDK 1: Hexagon
+$HexagonUrl     = "https://github.com/snapdragon-toolchain/hexagon-sdk/releases/download/v6.4.0.2/hexagon-sdk-v6.4.0.2-arm64-wos.tar.xz"
+$HexagonParent  = Join-Path $BaseDir "Hexagon_SDK"
+$HexagonSdkVersion   = "6.4.0.2"
+$HexagonToolsVersion = "19.0.04"
+$HexagonSdkTarget    = Join-Path $HexagonParent $HexagonSdkVersion
+$HexagonToolsTarget  = Join-Path $HexagonSdkTarget "\tools\HEXAGON_Tools\$HexagonToolsVersion"
+
+# SDK 2: OpenCL
+$OpenCLUrl      = "https://github.com/snapdragon-toolchain/opencl-sdk/releases/download/v2.3.2/adreno-opencl-sdk-v2.3.2-arm64-wos.tar.xz"
+$OpenCLParent   = Join-Path $BaseDir "OpenCL_SDK"
+$OpenCLVersion  = "2.3.2"
+$OpenCLTarget   = Join-Path $OpenCLParent $OpenCLVersion
+
+# --- Helper Function ---
+function Install-QualcommSDK {
+    param (
+        [string]$Url,
+        [string]$ParentDir,
+        [string]$TargetDir,
+        [string]$Name
+    )
+
+    # 1. Create Parent Directory
+    if (-not (Test-Path -Path $ParentDir)) {
+        Write-Host "Creating directory: $ParentDir" -ForegroundColor Cyan
+        New-Item -Path $ParentDir -ItemType Directory -Force | Out-Null
+    }
+
+    # 2. Check for Specific Version Directory
+    if (Test-Path -Path $TargetDir) {
+        Write-Host "$Name ($TargetDir) already exists. Skipping download." -ForegroundColor Green
+    }
+    else {
+        Write-Host "$Name not found. preparing to download..." -ForegroundColor Yellow
+
+        # Create the target directory to extract into
+        New-Item -Path $TargetDir -ItemType Directory -Force | Out-Null
+
+        # Define temporary archive path
+        $TempFile = Join-Path $ParentDir "temp_sdk.tar.xz"
+
+        try {
+            # Download
+            Write-Host "Downloading from: $Url"
+            Invoke-WebRequest -Uri $Url -OutFile $TempFile
+
+            # Untar
+            # Note: We assume Windows includes tar.exe (Win 10 build 17063+)
+            Write-Host "Extracting archive to $TargetDir..."
+
+            # We use -C to extract contents INTO the target directory created above
+            tar -xJvf $TempFile -C $TargetDir\..
+
+            Write-Host "Extraction complete." -ForegroundColor Green
+        }
+        catch {
+            Write-Error "Failed to download or extract $Name. Error: $_"
+            # Cleanup target dir if failed so script tries again next time
+            Remove-Item -Path $TargetDir -Recurse -Force -ErrorAction SilentlyContinue
+        }
+        finally {
+            # Cleanup Archive
+            if (Test-Path $TempFile) { Remove-Item $TempFile -Force }
+        }
+    }
+}
+
+# --- Execution ---
+
+# 1. Ensure Base C:\Qualcomm exists
+if (-not (Test-Path $BaseDir)) {
+    New-Item -Path $BaseDir -ItemType Directory -Force | Out-Null
+}
+
+# 2. Run Install Logic
+Install-QualcommSDK -Url $HexagonUrl -ParentDir $HexagonParent -TargetDir $HexagonSdkTarget -Name "Hexagon SDK"
+Install-QualcommSDK -Url $OpenCLUrl -ParentDir $OpenCLParent -TargetDir $OpenCLTarget -Name "OpenCL SDK"
+
+# --- Environment Variables ---
+
+Write-Host "`nSetting Environment Variables..." -ForegroundColor Cyan
+
+# Set OPENCL_SDK_ROOT
+[System.Environment]::SetEnvironmentVariable('OPENCL_SDK_ROOT', $OpenCLTarget, [System.EnvironmentVariableTarget]::User)
+$env:OPENCL_SDK_ROOT = $OpenCLTarget # Set for current session as well
+Write-Host "OPENCL_SDK_ROOT set to:  $OpenCLTarget"
+
+# Set HEXAGON_SDK_ROOT
+[System.Environment]::SetEnvironmentVariable('HEXAGON_SDK_ROOT', $HexagonSdkTarget, [System.EnvironmentVariableTarget]::User)
+$env:HEXAGON_SDK_ROOT = $HexagonSdkTarget # Set for current session as well
+Write-Host "HEXAGON_SDK_ROOT set to: $HexagonSdkTarget"
+
+# Set HEXAGON_SDK_ROOT
+[System.Environment]::SetEnvironmentVariable('HEXAGON_TOOLS_ROOT', $HexagonToolsTarget, [System.EnvironmentVariableTarget]::User)
+$env:HEXAGON_TOOLS_ROOT = $HexagonToolsTarget # Set for current session as well
+Write-Host "HEXAGON_TOOLS_ROOT set to: $HexagonToolsTarget"
diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last
index c838276158..81e79a9470 100644
--- a/scripts/sync-ggml.last
+++ b/scripts/sync-ggml.last
@@ -1 +1 @@
-ebc3a0f4a56be1c9424a89fbec09962ac34fde85
+a8db410a252c8c8f2d120c6f2e7133ebe032f35d
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index c15c281a5e..f337afd6b3 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -24,6 +24,7 @@ add_library(llama
             llama-kv-cache-iswa.cpp
             llama-memory.cpp
             llama-memory-hybrid.cpp
+            llama-memory-hybrid-iswa.cpp
             llama-memory-recurrent.cpp
             llama-mmap.cpp
             llama-model-loader.cpp
diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index a6d5ddfa33..10b306a853 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -253,11 +253,7 @@ llama_context::llama_context(
 
         // graph outputs buffer
         {
-            // resized during inference when a batch uses more outputs
-            // Create a dummy batch for initialization.
-            llama_batch dummy_batch = {};
-            dummy_batch.n_tokens = 0;
-            if (output_reserve(params.n_seq_max, dummy_batch) < params.n_seq_max) {
+            if (output_reserve(params.n_seq_max) < params.n_seq_max) {
                 throw std::runtime_error("failed to reserve initial output buffer");
             }
 
@@ -793,7 +789,7 @@ float * llama_context::get_embeddings_ith(int32_t i) {
             throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
         }
 
-        const uint32_t n_embd_out = model.hparams.get_n_embd_out();
+        const uint32_t n_embd_out = model.hparams.n_embd_out();
         return embd + j*n_embd_out;
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: invalid embeddings id %d, reason: %s\n", __func__, i, err.what());
@@ -1225,7 +1221,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
     n_queued_tokens += n_tokens;
 
     // reserve output buffer
-    if (output_reserve(n_tokens, batch_inp) < n_tokens) {
+    if (output_reserve(n_tokens) < n_tokens) {
         LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_tokens);
         return -2;
     };
@@ -1279,7 +1275,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
                 {
                     // extract token embeddings
                     GGML_ASSERT(embd != nullptr);
-                    const uint32_t n_embd_out = hparams.get_n_embd_out();
+                    const uint32_t n_embd_out = hparams.n_embd_out();
 
                     GGML_ASSERT(n_tokens*n_embd_out <= (int64_t) embd_size);
                     ggml_backend_tensor_get_async(backend_embd, t_embd, embd, 0, n_tokens*n_embd_out*sizeof(float));
@@ -1456,6 +1452,23 @@ static void copy_tensor_async_candidates(
     }
 }
 
+static bool needs_raw_logits(const llama_ubatch & ubatch, const std::map<llama_seq_id, llama_sampler *> & samplers) {
+    for (uint32_t i = 0; i < ubatch.n_tokens; i++) {
+        if (!ubatch.output[i]) {
+            continue;
+        }
+
+        // Check if the output token has at least one sequence without a backend sampler.
+        for (int32_t j = 0; j < ubatch.n_seq_id[i]; ++j) {
+            llama_seq_id seq_id = ubatch.seq_id[i][j];
+            if (samplers.find(seq_id) == samplers.end()) {
+                return true;
+            }
+        }
+    }
+    return false; // all sequences use backend sampling
+}
+
 int llama_context::decode(const llama_batch & batch_inp) {
     GGML_ASSERT((!batch_inp.token && batch_inp.embd) || (batch_inp.token && !batch_inp.embd)); // NOLINT
 
@@ -1588,7 +1601,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
     }
 
     // reserve output buffer
-    if (output_reserve(n_outputs_all, balloc->get_batch()) < n_outputs_all) {
+    if (output_reserve(n_outputs_all) < n_outputs_all) {
         LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all);
         return -2;
     };
@@ -1661,10 +1674,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
         }
 
         // extract logits
-        // For multi-sequence batches that mix backend samplers and CPU sampler
-        // this is currently inefficient as we copy all logits even for the
-        // backend sampled tokens.
-        if (logits && t_logits && n_outputs > 0) {
+        if (logits && t_logits && n_outputs > 0 && needs_raw_logits(ubatch, sampling.samplers)) {
             ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(sched.get(), t_logits);
             GGML_ASSERT(backend_res != nullptr);
             GGML_ASSERT(logits != nullptr);
@@ -1688,7 +1698,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
                     {
                         // extract token embeddings
                         GGML_ASSERT(embd != nullptr);
-                        const uint32_t n_embd_out = hparams.get_n_embd_out();
+                        const uint32_t n_embd_out = hparams.n_embd_out();
                         float * embd_out = embd + n_outputs_prev*n_embd_out;
 
                         if (n_outputs) {
@@ -1734,11 +1744,8 @@ int llama_context::decode(const llama_batch & batch_inp) {
             }
         }
 
-        // This flag indicates whether a backend sampler has actually sampled a specific
-        // token, or if it has produced probabilites. If true, we can skip the normal copying of logits and embeddings.
-        const bool has_sampled = !res->t_sampled.empty() || !res->t_sampled_probs.empty() || !res->t_sampled_logits.empty();
-
-        if (has_samplers && has_sampled) {
+        // Copy backend sampling output if this ubatch produced any sampling tensors.
+        if (has_samplers && (!res->t_sampled.empty() || !res->t_sampled_probs.empty() || !res->t_sampled_logits.empty())) {
             const auto seq_to_output_row = build_seq_to_output_row(ubatch, n_outputs_prev);
             const auto stride = n_vocab;
 
@@ -1813,7 +1820,8 @@ int llama_context::decode(const llama_batch & batch_inp) {
 // output
 //
 
-uint32_t llama_context::output_reserve(int32_t n_outputs, const llama_batch & batch) {
+uint32_t llama_context::output_reserve(int32_t n_outputs) {
+
     const auto & hparams = model.hparams;
     const auto & vocab   = model.vocab;
 
@@ -1821,7 +1829,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs, const llama_batch & ba
 
     const auto n_batch    = cparams.n_batch;
     const auto n_vocab    = vocab.n_tokens();
-    const auto n_embd_out = hparams.get_n_embd_out();
+    const auto n_embd_out = hparams.n_embd_out();
 
     bool has_logits = true;
     bool has_embd   = cparams.embeddings;
@@ -1832,45 +1840,16 @@ uint32_t llama_context::output_reserve(int32_t n_outputs, const llama_batch & ba
         has_embd   = true;
     }
 
-    // Check which sampling modes are needed for the current batch.
-    // TODO: avoid this branching by working with the worst-case
-    bool has_sampling = false;
-    bool cpu_logits   = false;
-
-    if (batch.logits) {
-        for (int32_t i = 0; i < batch.n_tokens; i++) {
-            if (!batch.logits[i]) {
-                continue;
-            }
-            for (int32_t j = 0; j < batch.n_seq_id[i]; j++) {
-                llama_seq_id seq_id = batch.seq_id[i][j];
-                if (sampling.samplers.find(seq_id) != sampling.samplers.end()) {
-                    has_sampling = true;
-                } else {
-                    cpu_logits = true;
-                }
-            }
-        }
-    } else {
-        // When batch.logits is nullptr (when loading state with a dummy batch),
-        // allocate CPU logits.
-        cpu_logits = true;
-    }
 
     size_t backend_float_count = 0;
     size_t backend_token_count = 0;
 
-    // Allocate CPU logits buffer only if needed by sequences in this batch
-    logits_size = (has_logits && cpu_logits) ? n_vocab*n_outputs_max : 0;
+    logits_size = has_logits ? n_vocab*n_outputs_max : 0;
     embd_size   = has_embd ? n_embd_out*n_outputs_max : 0;
 
-    // TODO: avoid this branching by working with the worst-case
-    if (!has_sampling) {
-        sampling.logits_size     = 0;
-        sampling.probs_size      = 0;
-        sampling.sampled_size    = 0;
-        sampling.candidates_size = 0;
-    } else {
+    // Allocate backend sampling output buffers if there are backend samplers configured.
+    const bool has_sampling = !sampling.samplers.empty();
+    if (has_sampling) {
         sampling.logits_size     = n_vocab*n_outputs_max;
         sampling.probs_size      = n_vocab*n_outputs_max;
         sampling.sampled_size    =         n_outputs_max;
@@ -1928,7 +1907,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs, const llama_batch & ba
     size_t offset = 0;
     uint8_t * base = (uint8_t *) output_base;
 
-    logits = (has_logits && cpu_logits) ? output_base : nullptr;
+    logits = has_logits ? output_base : nullptr;
     offset += logits_size * sizeof(float);
 
     embd = has_embd ? (float *) (base + offset) : nullptr;
@@ -2173,13 +2152,6 @@ llm_graph_cb llama_context::graph_get_cb() const {
             ggml_set_name(cur, name);
         }
 
-        if (!cparams.offload_kqv) {
-            if (strcmp(name, "kqv_merged_cont") == 0) {
-                // all nodes between the KV store and the attention output are run on the CPU
-                ggml_backend_sched_set_tensor_backend(sched.get(), cur, backend_cpu);
-            }
-        }
-
         // norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends
         // FIXME: fix in ggml_backend_sched
         const bool full_offload = model.n_gpu_layers() > model.hparams.n_layer;
@@ -2559,6 +2531,7 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
         }
     }
 
+    // [TAG_CONTEXT_STATE_LOGITS]
     // write logits
     {
         LLAMA_LOG_DEBUG("%s: - writing logits\n", __func__);
@@ -2620,10 +2593,7 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
         auto n_outputs = this->n_outputs;
         io.read_to(&n_outputs, sizeof(n_outputs));
 
-        // Create a dummy batch for state loading.
-        llama_batch dummy_batch = {};
-        dummy_batch.n_tokens = 0;
-        if (n_outputs > output_reserve(n_outputs, dummy_batch)) {
+        if (n_outputs > output_reserve(n_outputs)) {
             throw std::runtime_error("could not reserve outputs");
         }
 
@@ -2868,7 +2838,7 @@ void llama_context::opt_epoch_iter(
         }
 
         // reserve output buffer
-        if (output_reserve(n_outputs_all, balloc->get_batch()) < n_outputs_all) {
+        if (output_reserve(n_outputs_all) < n_outputs_all) {
             LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all);
             GGML_ABORT("TODO: handle this error");
         };
@@ -2903,7 +2873,7 @@ void llama_context::opt_epoch_iter(
                 };
                 ctx_compute_opt = ggml_init(params);
             }
-            ggml_opt_prepare_alloc(opt_ctx, ctx_compute_opt, gf, res->get_tokens(), res->get_logits());
+            ggml_opt_prepare_alloc(opt_ctx, ctx_compute_opt, gf, res->get_inp_tokens(), res->get_logits());
             ggml_opt_alloc(opt_ctx, train);
 
             res->set_inputs(&ubatch);
diff --git a/src/llama-context.h b/src/llama-context.h
index 86decc05fb..8e71cdd1dc 100644
--- a/src/llama-context.h
+++ b/src/llama-context.h
@@ -212,7 +212,7 @@ private:
 
     // Make sure enough space is available for outputs.
     // Returns max number of outputs for which space was reserved.
-    uint32_t output_reserve(int32_t n_outputs, const llama_batch & batch);
+    uint32_t output_reserve(int32_t n_outputs);
 
     void output_reorder();
 
diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 944c7e53bd..16d42c4ae3 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -7,6 +7,7 @@
 #include "llama-kv-cache.h"
 #include "llama-kv-cache-iswa.h"
 #include "llama-memory-hybrid.h"
+#include "llama-memory-hybrid-iswa.h"
 #include "llama-memory-recurrent.h"
 
 #include <cassert>
@@ -22,7 +23,8 @@ void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) {
     }
 
     if (ubatch->embd) {
-        const int64_t n_embd   = embd->ne[0];
+        GGML_ASSERT(n_embd == embd->ne[0]);
+
         const int64_t n_tokens = ubatch->n_tokens;
 
         ggml_backend_tensor_set(embd, ubatch->embd, 0, n_tokens*n_embd*ggml_element_size(embd));
@@ -32,8 +34,8 @@ void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) {
 bool llm_graph_input_embd::can_reuse(const llm_graph_params & params) {
     bool res = true;
 
-    res &= (!tokens && !params.ubatch.token) || (tokens && tokens->ne[0] == params.ubatch.n_tokens);
-    res &= (!embd   && !params.ubatch.embd)  || (embd   &&   embd->ne[1] == params.ubatch.n_tokens);
+    res &= (!params.ubatch.token) || (tokens && tokens->ne[0] == params.ubatch.n_tokens);
+    res &= (!params.ubatch.embd)  || (embd   &&   embd->ne[1] == params.ubatch.n_tokens);
 
     return res;
 }
@@ -405,6 +407,27 @@ bool llm_graph_input_attn_kv::can_reuse(const llm_graph_params & params) {
     return res;
 }
 
+void llm_graph_input_attn_k::set_input(const llama_ubatch * ubatch) {
+    mctx->set_input_k_idxs(self_k_idxs, ubatch);
+
+    mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
+}
+
+bool llm_graph_input_attn_k::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_kv_cache_context *>(params.mctx);
+
+    this->mctx = mctx;
+
+    bool res = true;
+
+    res &= self_k_idxs->ne[0] == params.ubatch.n_tokens;
+
+    res &= self_kq_mask->ne[0] == mctx->get_n_kv();
+    res &= self_kq_mask->ne[1] == params.ubatch.n_tokens;
+
+    return res;
+}
+
 void llm_graph_input_attn_kv_iswa::set_input(const llama_ubatch * ubatch) {
     mctx->get_base()->set_input_k_idxs(self_k_idxs, ubatch);
     mctx->get_base()->set_input_v_idxs(self_v_idxs, ubatch);
@@ -510,6 +533,76 @@ bool llm_graph_input_mem_hybrid::can_reuse(const llm_graph_params & params) {
     return res;
 }
 
+void llm_graph_input_mem_hybrid_iswa::set_input(const llama_ubatch * ubatch) {
+    const auto * attn_ctx = mctx->get_attn();
+
+    // base tensors may not be allocated if there are no non-SWA attention layers
+    if (inp_attn->self_k_idxs && inp_attn->self_k_idxs->buffer) {
+        attn_ctx->get_base()->set_input_k_idxs(inp_attn->self_k_idxs, ubatch);
+        attn_ctx->get_base()->set_input_v_idxs(inp_attn->self_v_idxs, ubatch);
+
+        attn_ctx->get_base()->set_input_kq_mask(inp_attn->self_kq_mask, ubatch, cparams.causal_attn);
+    }
+
+    // swa tensors may not be allocated if there are no SWA attention layers
+    if (inp_attn->self_k_idxs_swa && inp_attn->self_k_idxs_swa->buffer) {
+        attn_ctx->get_swa()->set_input_k_idxs(inp_attn->self_k_idxs_swa, ubatch);
+        attn_ctx->get_swa()->set_input_v_idxs(inp_attn->self_v_idxs_swa, ubatch);
+
+        attn_ctx->get_swa()->set_input_kq_mask(inp_attn->self_kq_mask_swa, ubatch, cparams.causal_attn);
+    }
+
+    const int64_t n_rs = mctx->get_recr()->get_n_rs();
+
+    if (inp_rs->s_copy) {
+        GGML_ASSERT(ggml_backend_buffer_is_host(inp_rs->s_copy->buffer));
+        int32_t * data = (int32_t *) inp_rs->s_copy->data;
+
+        // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n
+        for (uint32_t i = 0; i < n_rs; ++i) {
+            data[i] = mctx->get_recr()->s_copy(i);
+        }
+    }
+}
+
+bool llm_graph_input_mem_hybrid_iswa::can_reuse(const llm_graph_params & params) {
+    const auto * mctx = static_cast<const llama_memory_hybrid_iswa_context *>(params.mctx);
+
+    this->mctx = mctx;
+
+    bool res = true;
+
+    const auto * attn_ctx = mctx->get_attn();
+
+    // base tensors may not be allocated if there are no non-SWA attention layers
+    if (inp_attn->self_k_idxs && inp_attn->self_k_idxs->buffer) {
+        res &= inp_attn->self_k_idxs->ne[0] == params.ubatch.n_tokens;
+      //res &= inp_attn->self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
+
+        res &= inp_attn->self_kq_mask->ne[0] == attn_ctx->get_base()->get_n_kv();
+        res &= inp_attn->self_kq_mask->ne[1] == params.ubatch.n_tokens;
+    }
+
+    // swa tensors may not be allocated if there are no SWA attention layers
+    if (inp_attn->self_k_idxs_swa && inp_attn->self_k_idxs_swa->buffer) {
+        res &= inp_attn->self_k_idxs_swa->ne[0] == params.ubatch.n_tokens;
+      //res &= inp_attn->self_v_idxs_swa->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
+
+        res &= inp_attn->self_kq_mask_swa->ne[0] == attn_ctx->get_swa()->get_n_kv();
+        res &= inp_attn->self_kq_mask_swa->ne[1] == params.ubatch.n_tokens;
+    }
+
+    res &= inp_rs->s_copy->ne[0] == mctx->get_recr()->get_n_rs();
+
+    res &= inp_rs->s_copy_main->ne[0]  == params.ubatch.n_seqs;
+    res &= inp_rs->s_copy_extra->ne[0] == mctx->get_recr()->get_n_rs() - params.ubatch.n_seqs;
+
+    res &= inp_rs->head == mctx->get_recr()->get_head();
+    res &= inp_rs->rs_z == mctx->get_recr()->get_rs_z();
+
+    return res;
+}
+
 void llm_graph_input_sampling::set_input(const llama_ubatch * ubatch) {
     // set the inputs only for the active samplers in the current ubatch
     std::unordered_set<llama_seq_id> active_samplers;
@@ -563,7 +656,8 @@ int64_t llm_graph_result::get_max_nodes() const {
 }
 
 void llm_graph_result::reset() {
-    t_tokens      = nullptr;
+    t_inp_tokens  = nullptr;
+    t_inp_embd    = nullptr;
     t_logits      = nullptr;
     t_embd        = nullptr;
     t_embd_pooled = nullptr;
@@ -1267,17 +1361,29 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
 
 // input embeddings with optional lora
 ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
-    const int64_t n_embd = hparams.n_embd_inp();
+    const int64_t n_embd_inp = hparams.n_embd_inp();
+    const int64_t n_embd     = hparams.n_embd;
 
-    auto inp = std::make_unique<llm_graph_input_embd>();
+    assert(n_embd_inp >= n_embd);
 
-    ggml_tensor * cur = nullptr;
+    auto inp = std::make_unique<llm_graph_input_embd>(n_embd_inp);
 
-    if (ubatch.token) {
-        inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
-        //cb(inp->tokens, "inp_tokens", -1);
-        ggml_set_input(inp->tokens);
-        res->t_tokens = inp->tokens;
+    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
+    cb(inp->tokens, "inp_tokens", -1);
+    ggml_set_input(inp->tokens);
+    res->t_inp_tokens = inp->tokens;
+
+    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd_inp, ubatch.n_tokens);
+    cb(inp->embd, "inp_embd", -1);
+    ggml_set_input(inp->embd);
+
+    // select one of the 2 inputs, based on the batch contents
+    // ref: https://github.com/ggml-org/llama.cpp/pull/18550
+    std::array<ggml_tensor *, 2> inps;
+
+    // token embeddings path (ubatch.token != nullptr)
+    {
+        auto & cur = inps[0];
 
         cur = ggml_get_rows(ctx0, tok_embd, inp->tokens);
 
@@ -1298,19 +1404,36 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
 
             cur = ggml_add(ctx0, cur, inpL_delta);
         }
-    } else {
-        inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, ubatch.n_tokens);
-        ggml_set_input(inp->embd);
+
+        if (n_embd_inp != n_embd) {
+            cur = ggml_pad(ctx0, cur, hparams.n_embd_inp() - n_embd, 0, 0, 0);
+        }
+    }
+
+    // vector embeddings path (ubatch.embd != nullptr)
+    {
+        auto & cur = inps[1];
 
         cur = inp->embd;
     }
 
+    assert(ggml_are_same_shape (inps[0], inps[1]));
+    assert(ggml_are_same_stride(inps[0], inps[1]));
+
+    ggml_tensor * cur = ggml_build_forward_select(gf, inps.data(), inps.size(), ubatch.token ? 0 : 1);
+
+    if (n_embd_inp != n_embd) {
+        cur = ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0);
+    }
+
+    res->t_inp_embd = cur;
+
     // For Granite architecture
     if (hparams.f_embedding_scale != 0.0f) {
         cur = ggml_scale(ctx0, cur, hparams.f_embedding_scale);
     }
 
-    cb(cur, "inp_embd", -1);
+    cb(cur, "embd", -1);
 
     res->add_input(std::move(inp));
 
@@ -1409,7 +1532,7 @@ ggml_tensor * llm_graph_context::build_inp_cross_embd() const {
     //}
 
     const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd_inp();
-    const auto n_enc  = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;
+    const auto n_enc  = !cross->v_embd.empty() ? cross->n_enc  : hparams.n_ctx_train;
 
     cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_enc);
     ggml_set_input(cur);
@@ -1716,9 +1839,11 @@ ggml_tensor * llm_graph_context::build_attn(
         ggml_tensor * v_cur,
         ggml_tensor * kq_b,
         ggml_tensor * sinks,
-        ggml_tensor * v_mla,
+        ggml_tensor * v_mla, // TODO: remove
             float     kq_scale,
             int       il) const {
+    GGML_ASSERT(v_mla == nullptr);
+
     // these nodes are added to the graph together so that they are not reordered
     // by doing so, the number of splits in the graph is reduced
     // expand k later to enable rope fusion which directly writes into k-v cache
@@ -1761,6 +1886,93 @@ ggml_tensor * llm_graph_context::build_attn(
     return cur;
 }
 
+static std::unique_ptr<llm_graph_input_attn_k> build_attn_inp_k_impl(
+           ggml_context * ctx0,
+     const llama_ubatch & ubatch,
+    const llama_hparams & hparams,
+    const llama_cparams & cparams,
+    const llama_kv_cache_context * mctx_cur) {
+
+    auto inp = std::make_unique<llm_graph_input_attn_k>(hparams, cparams, mctx_cur);
+
+    {
+        GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_iswa for SWA");
+
+        const auto n_kv     = mctx_cur->get_n_kv();
+        const auto n_tokens = ubatch.n_tokens;
+        const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
+
+        inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);
+
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
+        ggml_set_input(inp->self_kq_mask);
+
+        inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
+    }
+
+    return inp;
+}
+
+llm_graph_input_attn_k * llm_graph_context::build_attn_inp_k() const {
+    const auto * mctx_cur = static_cast<const llama_kv_cache_context *>(mctx);
+
+    auto inp = build_attn_inp_k_impl(ctx0, ubatch, hparams, cparams, mctx_cur);
+
+    return (llm_graph_input_attn_k *) res->add_input(std::move(inp));
+}
+
+ggml_tensor * llm_graph_context::build_attn(
+        llm_graph_input_attn_k * inp,
+        ggml_tensor * wo,
+        ggml_tensor * wo_b,
+        ggml_tensor * q_cur,
+        ggml_tensor * k_cur,
+        ggml_tensor * v_cur,
+        ggml_tensor * kq_b,
+        ggml_tensor * sinks,
+        ggml_tensor * v_mla,
+            float     kq_scale,
+            int       il) const {
+    // these nodes are added to the graph together so that they are not reordered
+    // by doing so, the number of splits in the graph is reduced
+    // expand k later to enable rope fusion which directly writes into k-v cache
+    ggml_build_forward_expand(gf, q_cur);
+    ggml_build_forward_expand(gf, v_cur);
+    ggml_build_forward_expand(gf, k_cur);
+
+    const auto * mctx_cur = inp->mctx;
+
+    // store to KV cache
+    {
+        const auto & k_idxs = inp->get_k_idxs();
+
+        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
+    }
+
+    const auto & kq_mask = inp->get_kq_mask();
+
+    ggml_tensor * q = q_cur;
+    ggml_tensor * k = mctx_cur->get_k(ctx0, il);
+    ggml_tensor * v = ggml_view_4d(ctx0, k, v_cur->ne[0], k->ne[1], k->ne[2], k->ne[3], k->nb[1], k->nb[2], k->nb[3], 0);
+
+    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
+    cb(cur, "kqv_out", il);
+
+    if (wo) {
+        cur = build_lora_mm(wo, cur);
+        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
+            // GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
+            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
+        }
+    }
+
+    if (wo_b) {
+        cur = ggml_add(ctx0, cur, wo_b);
+    }
+
+    return cur;
+}
+
 ggml_tensor * llm_graph_context::build_attn(
         llm_graph_input_attn_kv_iswa * inp,
         ggml_tensor * wo,
@@ -2056,6 +2268,47 @@ llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
     return (llm_graph_input_mem_hybrid *) res->add_input(std::move(inp));
 }
 
+llm_graph_input_mem_hybrid_iswa * llm_graph_context::build_inp_mem_hybrid_iswa() const {
+    const auto * mctx_cur = static_cast<const llama_memory_hybrid_iswa_context *>(mctx);
+
+    auto inp_rs = build_rs_inp_impl(ctx0, ubatch, mctx_cur->get_recr());
+
+    // build iswa attention input
+    const auto * attn_ctx = mctx_cur->get_attn();
+
+    auto inp_attn = std::make_unique<llm_graph_input_attn_kv_iswa>(hparams, cparams, attn_ctx);
+
+    const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
+
+    {
+        const auto n_kv = attn_ctx->get_base()->get_n_kv();
+
+        inp_attn->self_k_idxs = attn_ctx->get_base()->build_input_k_idxs(ctx0, ubatch);
+        inp_attn->self_v_idxs = attn_ctx->get_base()->build_input_v_idxs(ctx0, ubatch);
+
+        inp_attn->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
+        ggml_set_input(inp_attn->self_kq_mask);
+
+        inp_attn->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp_attn->self_kq_mask, GGML_TYPE_F16) : inp_attn->self_kq_mask;
+    }
+
+    {
+        const auto n_kv = attn_ctx->get_swa()->get_n_kv();
+
+        inp_attn->self_k_idxs_swa = attn_ctx->get_swa()->build_input_k_idxs(ctx0, ubatch);
+        inp_attn->self_v_idxs_swa = attn_ctx->get_swa()->build_input_v_idxs(ctx0, ubatch);
+
+        inp_attn->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
+        ggml_set_input(inp_attn->self_kq_mask_swa);
+
+        inp_attn->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp_attn->self_kq_mask_swa, GGML_TYPE_F16) : inp_attn->self_kq_mask_swa;
+    }
+
+    auto inp = std::make_unique<llm_graph_input_mem_hybrid_iswa>(cparams, std::move(inp_attn), std::move(inp_rs), mctx_cur);
+
+    return (llm_graph_input_mem_hybrid_iswa *) res->add_input(std::move(inp));
+}
+
 void llm_graph_context::build_dense_out(
     ggml_tensor * dense_2,
     ggml_tensor * dense_3) const {
diff --git a/src/llama-graph.h b/src/llama-graph.h
index 503ffd695a..4090d8116c 100644
--- a/src/llama-graph.h
+++ b/src/llama-graph.h
@@ -24,6 +24,7 @@ class llama_kv_cache_context;
 class llama_kv_cache_iswa_context;
 class llama_memory_recurrent_context;
 class llama_memory_hybrid_context;
+class llama_memory_hybrid_iswa_context;
 
 // certain models (typically multi-modal) can produce different types of graphs
 enum llm_graph_type {
@@ -105,7 +106,7 @@ using llm_graph_input_ptr = std::unique_ptr<llm_graph_input_i>;
 
 class llm_graph_input_embd : public llm_graph_input_i {
 public:
-    llm_graph_input_embd()          = default;
+    llm_graph_input_embd(int64_t n_embd) : n_embd(n_embd) {}
     virtual ~llm_graph_input_embd() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
@@ -114,6 +115,8 @@ public:
 
     ggml_tensor * tokens = nullptr; // I32 [n_batch]
     ggml_tensor * embd   = nullptr; // F32 [n_embd, n_batch]
+
+    const int64_t n_embd = 0;
 };
 
 class llm_graph_input_pos : public llm_graph_input_i {
@@ -314,6 +317,39 @@ public:
     const llama_kv_cache_context * mctx;
 };
 
+// V-less input for the KV cache
+// ref: https://github.com/ggml-org/llama.cpp/pull/19067
+class llm_graph_input_attn_k : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_k(
+            const llama_hparams & hparams,
+            const llama_cparams & cparams,
+            const llama_kv_cache_context * mctx) :
+        hparams(hparams),
+        cparams(cparams),
+        mctx(mctx) {
+    }
+    ~llm_graph_input_attn_k() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    bool can_reuse(const llm_graph_params & params) override;
+
+    ggml_tensor * get_k_idxs() const { return self_k_idxs; }
+
+    ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
+
+    ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
+
+    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
+
+    const llama_hparams hparams;
+    const llama_cparams cparams;
+
+    const llama_kv_cache_context * mctx;
+};
+
 class llm_graph_input_attn_kv_iswa : public llm_graph_input_i {
 public:
     llm_graph_input_attn_kv_iswa(
@@ -397,6 +433,34 @@ public:
     const llama_memory_hybrid_context * mctx;
 };
 
+class llm_graph_input_mem_hybrid_iswa : public llm_graph_input_i {
+public:
+    llm_graph_input_mem_hybrid_iswa(
+            const llama_cparams & cparams,
+            std::unique_ptr<llm_graph_input_attn_kv_iswa> inp_attn,
+            std::unique_ptr<llm_graph_input_rs>          inp_rs,
+            const llama_memory_hybrid_iswa_context *     mctx) :
+        inp_attn(std::move(inp_attn)),
+        inp_rs(std::move(inp_rs)),
+        cparams(cparams),
+        mctx(mctx) { }
+    virtual ~llm_graph_input_mem_hybrid_iswa() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    bool can_reuse(const llm_graph_params & params) override;
+
+    std::unique_ptr<llm_graph_input_attn_kv_iswa> inp_attn;
+    std::unique_ptr<llm_graph_input_rs>          inp_rs;
+
+    llm_graph_input_attn_kv_iswa * get_attn() const { return inp_attn.get(); }
+    llm_graph_input_rs           * get_recr() const { return inp_rs.get(); }
+
+    const llama_cparams cparams;
+
+    const llama_memory_hybrid_iswa_context * mctx;
+};
+
 class llm_graph_input_sampling : public llm_graph_input_i {
 public:
     llm_graph_input_sampling(std::map<llama_seq_id, llama_sampler *> samplers) :
@@ -537,7 +601,7 @@ public:
 
     virtual ~llm_graph_result() = default;
 
-    ggml_tensor * get_tokens()      const { return t_tokens; }
+    ggml_tensor * get_inp_tokens()  const { return t_inp_tokens; }
     ggml_tensor * get_logits()      const { return t_logits; }
     ggml_tensor * get_embd()        const { return t_embd; }
     ggml_tensor * get_embd_pooled() const { return t_embd_pooled; }
@@ -564,7 +628,8 @@ public:
     void set_params(const llm_graph_params & params);
 
     // important graph nodes
-    ggml_tensor * t_tokens      = nullptr;
+    ggml_tensor * t_inp_tokens  = nullptr;
+    ggml_tensor * t_inp_embd    = nullptr; // [n_embd_inp, n_tokens]
     ggml_tensor * t_logits      = nullptr;
     ggml_tensor * t_embd        = nullptr;
     ggml_tensor * t_embd_pooled = nullptr;
@@ -801,6 +866,21 @@ struct llm_graph_context {
             ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
             ggml_tensor * kq_b,
             ggml_tensor * sinks, // [n_head_q]
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v] // TODO: remove
+                  float   kq_scale,
+                    int   il) const;
+
+    llm_graph_input_attn_k  * build_attn_inp_k() const;
+
+    ggml_tensor * build_attn(
+            llm_graph_input_attn_k * inp,
+            ggml_tensor * wo,
+            ggml_tensor * wo_b,
+            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
+            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
+            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
+            ggml_tensor * kq_b,
+            ggml_tensor * sinks, // [n_head_q]
             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
@@ -881,6 +961,8 @@ struct llm_graph_context {
 
     llm_graph_input_mem_hybrid * build_inp_mem_hybrid() const;
 
+    llm_graph_input_mem_hybrid_iswa * build_inp_mem_hybrid_iswa() const;
+
     //
     // pooling
     //
diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp
index c847ef91b7..392f9160ce 100644
--- a/src/llama-hparams.cpp
+++ b/src/llama-hparams.cpp
@@ -72,8 +72,8 @@ uint32_t llama_hparams::n_embd_inp() const {
     return n_embd_inp;
 }
 
-uint32_t llama_hparams::get_n_embd_out() const {
-    return n_embd_out > 0 ? n_embd_out : n_embd;
+uint32_t llama_hparams::n_embd_out() const {
+    return n_embd_out_impl > 0 ? n_embd_out_impl : n_embd;
 }
 
 uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
@@ -175,6 +175,21 @@ bool llama_hparams::is_swa(uint32_t il) const {
     GGML_ABORT("fatal error");
 }
 
+bool llama_hparams::is_mla() const {
+    assert((n_embd_head_k_mla_impl == 0 && n_embd_head_v_mla_impl == 0) ||
+           (n_embd_head_k_mla_impl != 0 && n_embd_head_v_mla_impl != 0));
+
+    return n_embd_head_k_mla_impl != 0 && n_embd_head_v_mla_impl != 0;
+}
+
+uint32_t llama_hparams::n_embd_head_k_mla() const {
+    return is_mla() ? n_embd_head_k_mla_impl : n_embd_head_k;
+}
+
+uint32_t llama_hparams::n_embd_head_v_mla() const {
+    return is_mla() ? n_embd_head_v_mla_impl : n_embd_head_v;
+}
+
 bool llama_hparams::has_kv(uint32_t il) const {
     if (n_layer_kv_from_start >= 0) {
         if (il < (uint32_t) n_layer_kv_from_start) {
@@ -200,42 +215,6 @@ uint32_t llama_hparams::n_layer_kv() const {
     return res;
 }
 
-bool llama_hparams::is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama_pos p0, llama_pos p1) {
-    assert(p0 >= 0 && p1 >= 0);
-
-    switch (swa_type) {
-        case LLAMA_SWA_TYPE_NONE:
-            {
-            } break;
-        case LLAMA_SWA_TYPE_STANDARD:
-            {
-                if (p1 - p0 >= (int32_t) n_swa) {
-                    return true;
-                }
-            } break;
-        case LLAMA_SWA_TYPE_CHUNKED:
-            {
-                const llama_pos pos_chunk_start = (p1 / n_swa) * n_swa;
-
-                if (p0 < pos_chunk_start) {
-                    return true;
-                }
-            } break;
-        case LLAMA_SWA_TYPE_SYMMETRIC:
-            {
-                const int32_t half_n_swa = (int32_t) n_swa / 2;
-                const int32_t pos_diff = p1 - p0;
-
-                // Mask if outside the symmetric window
-                if (pos_diff < -half_n_swa || pos_diff > half_n_swa) {
-                    return true;
-                }
-            } break;
-    }
-
-    return false;
-}
-
 bool llama_hparams::use_mrope() const {
     return rope_sections[0] > 0 && rope_sections[1] > 0;
 }
diff --git a/src/llama-hparams.h b/src/llama-hparams.h
index 7ae3ec292e..caed0ec1b7 100644
--- a/src/llama-hparams.h
+++ b/src/llama-hparams.h
@@ -3,6 +3,7 @@
 #include "llama.h"
 
 #include <array>
+#include <cassert>
 
 // bump if necessary
 #define LLAMA_MAX_LAYERS  512
@@ -52,8 +53,8 @@ struct llama_hparams {
     uint32_t n_rel_attn_bkts = 0;
 
     // note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
-    uint32_t n_embd_head_k_mla = 0;
-    uint32_t n_embd_head_v_mla = 0;
+    uint32_t n_embd_head_k_mla_impl = 0;
+    uint32_t n_embd_head_v_mla_impl = 0;
 
     // for WavTokenizer
     struct llama_hparams_posnet   posnet;
@@ -163,7 +164,7 @@ struct llama_hparams {
     uint32_t n_cls_out = 1;
 
     // output embedding dimension (0 = use n_embd)
-    uint32_t n_embd_out = 0;
+    uint32_t n_embd_out_impl = 0;
 
     // llama4 smallthinker
     uint32_t n_moe_layer_step        = 0;
@@ -238,7 +239,7 @@ struct llama_hparams {
     uint32_t n_embd_inp() const;
 
     // dimension of output embeddings
-    uint32_t get_n_embd_out() const;
+    uint32_t n_embd_out() const;
 
     // dimension of key embeddings across all k-v heads
     uint32_t n_embd_k_gqa(uint32_t il = 0) const;
@@ -268,15 +269,57 @@ struct llama_hparams {
 
     bool is_swa(uint32_t il) const;
 
+    // note: currently only support if either all or none of the layers are MLA
+    bool is_mla() const;
+
+    uint32_t n_embd_head_k_mla() const;
+    uint32_t n_embd_head_v_mla() const;
+
     bool has_kv(uint32_t il) const;
 
     // number of layers for which has_kv() returns true
     uint32_t n_layer_kv() const;
 
     // note that this function uses different SWA parameters from those in the hparams
+    // note: inlined on purpose for performance reasons
     // TODO: think of a better place for this function
     // TODO: pack the SWA params in a struct?
-    static bool is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama_pos p0, llama_pos p1);
+    static bool is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama_pos p0, llama_pos p1) {
+        assert(p0 >= 0 && p1 >= 0);
+
+        switch (swa_type) {
+            case LLAMA_SWA_TYPE_NONE:
+                {
+                } break;
+            case LLAMA_SWA_TYPE_STANDARD:
+                {
+                    if (p1 - p0 >= (int32_t) n_swa) {
+                        return true;
+                    }
+                } break;
+            case LLAMA_SWA_TYPE_CHUNKED:
+                {
+                    const llama_pos pos_chunk_start = (p1 / n_swa) * n_swa;
+
+                    if (p0 < pos_chunk_start) {
+                        return true;
+                    }
+                } break;
+            case LLAMA_SWA_TYPE_SYMMETRIC:
+                {
+                    const int32_t half_n_swa = (int32_t) n_swa / 2;
+                    const int32_t pos_diff = p1 - p0;
+
+                    // Mask if outside the symmetric window
+                    if (pos_diff < -half_n_swa || pos_diff > half_n_swa) {
+                        return true;
+                    }
+                } break;
+        }
+
+        return false;
+    }
+
 
     bool use_mrope() const;
 };
diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
index 3186242d60..c35cd6761b 100644
--- a/src/llama-kv-cache.cpp
+++ b/src/llama-kv-cache.cpp
@@ -97,6 +97,8 @@ llama_kv_cache::llama_kv_cache(
                 __func__, hparams.n_embd_v_gqa_max());
     }
 
+    const bool is_mla = hparams.is_mla();
+
     for (uint32_t il = 0; il < hparams.n_layer; il++) {
         if (!hparams.has_kv(il)) {
             LLAMA_LOG_DEBUG("%s: layer %3d: does not have KV cache\n", __func__, il);
@@ -130,18 +132,21 @@ llama_kv_cache::llama_kv_cache(
             throw std::runtime_error("failed to create ggml context for kv cache");
         }
 
-        ggml_tensor * k = ggml_new_tensor_3d(ctx, type_k, n_embd_k_gqa, kv_size, n_stream);
-        ggml_tensor * v = ggml_new_tensor_3d(ctx, type_v, n_embd_v_gqa, kv_size, n_stream);
+        const bool has_k = true;
+        const bool has_v = !is_mla;
 
-        ggml_format_name(k, "cache_k_l%d", il);
-        ggml_format_name(v, "cache_v_l%d", il);
+        ggml_tensor * k = has_k ? ggml_new_tensor_3d(ctx, type_k, n_embd_k_gqa, kv_size, n_stream) : nullptr;
+        ggml_tensor * v = has_v ? ggml_new_tensor_3d(ctx, type_v, n_embd_v_gqa, kv_size, n_stream) : nullptr;
+
+        has_k && ggml_format_name(k, "cache_k_l%d", il);
+        has_v && ggml_format_name(v, "cache_v_l%d", il);
 
         std::vector<ggml_tensor *> k_stream;
         std::vector<ggml_tensor *> v_stream;
 
         for (uint32_t s = 0; s < n_stream; ++s) {
-            k_stream.push_back(ggml_view_2d(ctx, k, n_embd_k_gqa, kv_size, k->nb[1], s*k->nb[2]));
-            v_stream.push_back(ggml_view_2d(ctx, v, n_embd_v_gqa, kv_size, v->nb[1], s*v->nb[2]));
+            k_stream.push_back(has_k ? ggml_view_2d(ctx, k, n_embd_k_gqa, kv_size, k->nb[1], s*k->nb[2]) : nullptr);
+            v_stream.push_back(has_v ? ggml_view_2d(ctx, v, n_embd_v_gqa, kv_size, v->nb[1], s*v->nb[2]) : nullptr);
         }
 
         map_layer_ids[il] = layers.size();
@@ -647,7 +652,10 @@ bool llama_kv_cache::update(llama_context * lctx, bool do_shift, const stream_co
                 const auto & layer = layers[il];
 
                 ggml_backend_tensor_copy(layer.k_stream[ssrc], layer.k_stream[sdst]);
-                ggml_backend_tensor_copy(layer.v_stream[ssrc], layer.v_stream[sdst]);
+
+                if (layer.v_stream[ssrc]) {
+                    ggml_backend_tensor_copy(layer.v_stream[ssrc], layer.v_stream[sdst]);
+                }
             }
         }
     }
@@ -852,7 +860,7 @@ llama_kv_cache::slot_info llama_kv_cache::find_slot(const llama_ubatch & ubatch,
                         const llama_seq_id seq_id_cell = cells.seq_get(idx);
 
                         // SWA mask
-                        if (is_masked_swa(pos_cell, cells.seq_pos_max(seq_id_cell) + 1)) {
+                        if (llama_hparams::is_masked_swa(n_swa, swa_type, pos_cell, cells.seq_pos_max(seq_id_cell) + 1)) {
                             can_use = true;
                         }
                     }
@@ -1237,6 +1245,197 @@ void llama_kv_cache::set_input_k_shift(ggml_tensor * dst) const {
     }
 }
 
+struct args_set_input_kq_mask {
+    const llama_hparams & hparams;
+    const llama_ubatch  * ubatch;
+
+    const std::vector<llama_kv_cells> & v_cells;
+    const std::vector<uint32_t>       & seq_to_stream;
+
+    uint32_t       n_swa;
+    llama_swa_type swa_type;
+
+    int64_t n_kv;
+    int64_t n_stream;
+    int64_t n_tps;
+};
+
+template<bool causal, bool swa, bool is_2d, bool alibi>
+static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, float * data) {
+  //const auto & hparams = args.hparams;
+    const auto & ubatch  = args.ubatch;
+
+    const auto & v_cells       = args.v_cells;
+    const auto & seq_to_stream = args.seq_to_stream;
+
+    const uint32_t       n_swa    = args.n_swa;
+    const llama_swa_type swa_type = args.swa_type;
+
+    const int64_t n_kv     = args.n_kv;
+    const int64_t n_stream = args.n_stream;
+    const int64_t n_tps    = args.n_tps;
+
+    // the min position in the batch for each sequence
+    llama_pos seq_pos_min[LLAMA_MAX_SEQ];
+    std::fill(seq_pos_min, seq_pos_min + LLAMA_MAX_SEQ, INT32_MAX);
+
+    for (uint32_t i = 0; i < ubatch->n_tokens; ++i) {
+        const llama_seq_id seq_id = ubatch->seq_id[i][0];
+
+        seq_pos_min[seq_id] = std::min(seq_pos_min[seq_id], ubatch->pos[i]);
+    }
+
+    for (uint32_t s = 0; s < n_stream; ++s) {
+        // bookeeping of the KQ mask cells that could change for other tokens of the same sequence
+        std::unordered_map<llama_seq_id, uint32_t>              seq_srct;
+        std::unordered_map<llama_seq_id, std::vector<uint32_t>> seq_idxs;
+
+        for (uint32_t ii = 0; ii < n_tps; ++ii) {
+            const uint32_t i = s*n_tps + ii;
+
+            const llama_seq_id seq_id = ubatch->seq_id[i][0];
+
+            const auto & cells = v_cells.at(seq_to_stream[seq_id]);
+
+                  llama_pos p0 = -1;
+            const llama_pos p1 = ubatch->pos[i];
+
+            // for M-RoPE
+            const llama_pos p1_x = is_2d ? ubatch->pos[i + ubatch->n_tokens*2] : 0;
+            const llama_pos p1_y = is_2d ? ubatch->pos[i + ubatch->n_tokens]   : 0;
+
+            const uint64_t idst = n_kv*i;
+
+            // for tokens of the same sequence, the mask is mostly the same, so we can reuse it
+            // the only cells that could change are the ones that are with similar positions as the
+            //   ones in the batch (i.e. due to causal masking, SWA, etc.)
+            // keep track of those cells and shortcut the loop to save time
+            // note: this optimization is not compatible with Alibi position encoding
+            // ref:  https://github.com/ggml-org/llama.cpp/pull/18842
+            bool prev = false;
+
+            auto & idxs = seq_idxs[seq_id];
+
+            if (!alibi) {
+                if (seq_srct.find(seq_id) != seq_srct.end()) {
+                    const uint32_t srct = seq_srct[seq_id];
+
+                    const uint64_t idst_prev = n_kv*srct;
+
+                    std::copy(data + idst_prev, data + idst_prev + n_kv, data + idst);
+
+                    prev = true;
+                } else {
+                    idxs.clear();
+                    idxs.reserve(ubatch->n_tokens + n_swa + 32);
+
+                    seq_srct[seq_id] = i;
+                }
+            }
+
+            for (uint32_t jj = 0; jj < n_kv; ++jj) {
+                uint32_t j = jj;
+
+                // we have an exiting mask for this sequence -> update just seq_idxs
+                if (!alibi) {
+                    if (prev) {
+                        if (jj >= idxs.size()) {
+                            break;
+                        }
+
+                        j = idxs[jj];
+                    }
+                }
+
+                if (cells.is_empty(j)) {
+                    goto skip;
+                }
+
+                // mask the token if not the same sequence
+                if (!cells.seq_has(j, seq_id)) {
+                    goto skip;
+                }
+
+                p0 = cells.pos_get(j);
+
+                if (!alibi) {
+                    if (!prev) {
+                        // record all cells for which: p0 >= seq_pos_min[seq_id] - n_swa - 32
+                        if (p0 + (int32_t) (n_swa + 32) >= seq_pos_min[seq_id]) {
+                            idxs.push_back(j);
+                        }
+                    }
+                }
+
+                if (causal) {
+                    // mask future tokens
+                    if (p0 > p1) {
+                        goto skip;
+                    }
+
+                    // M-RoPE causal mask
+                    if (is_2d) {
+                        if (p0 == p1) {
+                            const auto & p0_ext = cells.ext_get(j);
+
+                            if (p0_ext.is_2d_gt(p1_x, p1_y)) {
+                                goto skip;
+                            }
+                        }
+                    }
+                }
+
+                // apply SWA if any
+                if (swa) {
+                    if (llama_hparams::is_masked_swa(n_swa, swa_type, p0, p1)) {
+                        goto skip;
+                    }
+                }
+
+                if (alibi) {
+                    data[idst + j] = -std::abs(p0 - p1);
+                } else {
+                    data[idst + j] = 0.0f;
+                }
+
+                continue;
+skip:
+                data[idst + j] = -INFINITY;
+            }
+        }
+    }
+}
+
+template<bool causal, bool swa, bool is_2d>
+static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, float * data) {
+    const bool alibi = args.hparams.use_alibi;
+    if (alibi) {
+        set_input_kq_mask_impl<causal, swa, is_2d, true> (args, data);
+    } else {
+        set_input_kq_mask_impl<causal, swa, is_2d, false>(args, data);
+    }
+}
+
+template<bool causal, bool swa>
+static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, float * data) {
+    const bool is_2d = args.ubatch->is_pos_2d();
+    if (is_2d) {
+        set_input_kq_mask_impl<causal, swa, true> (args, data);
+    } else {
+        set_input_kq_mask_impl<causal, swa, false>(args, data);
+    }
+}
+
+template<bool causal>
+static void set_input_kq_mask_impl(const args_set_input_kq_mask & args, float * data) {
+    const bool swa = args.swa_type != LLAMA_SWA_TYPE_NONE;
+    if (swa) {
+        set_input_kq_mask_impl<causal, true> (args, data);
+    } else {
+        set_input_kq_mask_impl<causal, false>(args, data);
+    }
+}
+
 void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const {
     const uint32_t n_tokens = ubatch->n_tokens;
 
@@ -1251,74 +1450,29 @@ void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * u
     // n_tps == n_tokens_per_stream
     const int64_t n_tps = n_tokens/n_stream;
 
-    std::fill(data, data + ggml_nelements(dst), -INFINITY);
+    //const int64_t t_start = ggml_time_us();
 
-    // Use only the previous KV cells of the correct sequence for each token of the ubatch.
-    // It's assumed that if a token in the batch has multiple sequences, they are equivalent.
-    // Example with a cache of 10 tokens, 2 tokens populated in cache and 3 tokens in batch:
-    //   Causal mask:
-    //      xxx-------
-    //      xxxx------
-    //      xxxxx-----
-    //   Non-causal mask:
-    //      xxxxx-----
-    //      xxxxx-----
-    //      xxxxx-----
-    // To visualize the mask, see https://github.com/ggml-org/llama.cpp/pull/12615
-    // TODO: optimize this section
-    for (uint32_t h = 0; h < 1; ++h) {
-        for (uint32_t s = 0; s < n_stream; ++s) {
-            for (uint32_t ii = 0; ii < n_tps; ++ii) {
-                const uint32_t i = s*n_tps + ii;
+    const args_set_input_kq_mask args = {
+        /*.hparams          =*/ hparams,
+        /*.ubatch           =*/ ubatch,
+        /*.v_cells          =*/ v_cells,
+        /*.seq_to_stream    =*/ seq_to_stream,
+        /*.n_swa            =*/ n_swa,
+        /*.swa_type         =*/ swa_type,
+        /*.n_kv             =*/ n_kv,
+        /*.n_stream         =*/ n_stream,
+        /*.n_tps            =*/ n_tps,
+    };
 
-                const llama_seq_id seq_id = ubatch->seq_id[i][0];
-
-                const auto & cells = v_cells[seq_to_stream[seq_id]];
-
-                const llama_pos p1 = ubatch->pos[i];
-
-                // for M-RoPE
-                const bool is_2d = ubatch->is_pos_2d();
-                const llama_pos p1_x = is_2d ? ubatch->pos[i + ubatch->n_tokens*2] : 0;
-                const llama_pos p1_y = is_2d ? ubatch->pos[i + ubatch->n_tokens]   : 0;
-
-                const uint64_t idst = n_kv*(h*n_stream*n_tps + s*n_tps + ii);
-
-                for (uint32_t j = 0; j < n_kv; ++j) {
-                    if (cells.is_empty(j)) {
-                        continue;
-                    }
-
-                    // mask the token if not the same sequence
-                    if (!cells.seq_has(j, seq_id)) {
-                        continue;
-                    }
-
-                    const llama_pos p0 = cells.pos_get(j);
-
-                    // mask future tokens
-                    if (causal_attn && p0 > p1) {
-                        continue;
-                    }
-
-                    // M-RoPE causal mask
-                    if (causal_attn && is_2d && p0 == p1) {
-                        const auto & p0_ext = cells.ext_get(j);
-                        if (p0_ext.is_2d_gt(p1_x, p1_y)) {
-                            continue;
-                        }
-                    }
-
-                    // apply SWA if any
-                    if (is_masked_swa(p0, p1)) {
-                        continue;
-                    }
-
-                    data[idst + j] = hparams.use_alibi ? -std::abs(p0 - p1) : 0.0f;
-                }
-            }
-        }
+    if (causal_attn) {
+        set_input_kq_mask_impl<true> (args, data);
+    } else {
+        set_input_kq_mask_impl<false>(args, data);
     }
+
+    //const int64_t t_end = ggml_time_us();
+
+    //LLAMA_LOG_ERROR("%s: kq mask time: %0.3f ms\n", __func__, (t_end - t_start)/1000.0);
 }
 
 void llama_kv_cache::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const {
@@ -1370,7 +1524,7 @@ size_t llama_kv_cache::size_v_bytes() const {
     size_t size_v_bytes = 0;
 
     for (const auto & layer : layers) {
-        size_v_bytes += ggml_nbytes(layer.v);
+        size_v_bytes += layer.v ? ggml_nbytes(layer.v) : 0;
     }
 
     return size_v_bytes;
@@ -1448,6 +1602,10 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
     const auto & n_embd_head_k = hparams.n_embd_head_k;
   //const auto & n_embd_head_v = hparams.n_embd_head_v;
 
+    const auto & n_rot = hparams.n_rot;
+
+    const auto n_embd_nope = hparams.n_lora_kv > 0 ? n_embd_head_k - n_rot : 0;
+
     auto inp = std::make_unique<llm_graph_input_k_shift>(this);
 
     inp->k_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, (int64_t) get_size()*n_stream);
@@ -1468,10 +1626,10 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
 
         ggml_tensor * k =
             ggml_view_3d(ctx, layer.k,
-                n_embd_head_k, n_head_kv, get_size()*n_stream,
+                n_rot, n_head_kv, get_size()*n_stream,
                 ggml_row_size(layer.k->type, n_embd_head_k),
                 ggml_row_size(layer.k->type, n_embd_k_gqa),
-                0);
+                ggml_row_size(layer.k->type, n_embd_nope));
 
         ggml_tensor * cur = build_rope_shift(cparams, ctx, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l);
 
@@ -1483,10 +1641,6 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
     return gf;
 }
 
-bool llama_kv_cache::is_masked_swa(llama_pos p0, llama_pos p1) const {
-    return llama_hparams::is_masked_swa(n_swa, swa_type, p0, p1);
-}
-
 void llama_kv_cache::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
     GGML_UNUSED(flags);
 
@@ -1618,8 +1772,6 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
     io.write(&v_trans, sizeof(v_trans));
     io.write(&n_layer, sizeof(n_layer));
 
-    std::vector<uint8_t> tmp_buf;
-
     // Iterate and write all the keys first, each row is a cell
     // Get whole range at a time
     for (const auto & layer : layers) {
@@ -1637,7 +1789,7 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
         const uint64_t k_size_row = ggml_row_size(k->type, n_embd_k_gqa);
         io.write(&k_size_row, sizeof(k_size_row));
 
-        // Read each range of cells of k_size length each into tmp_buf and write out
+        // Read each range of cells of k_size length and write out
         for (const auto & range : cr.data) {
             const size_t range_size = range.second - range.first;
             const size_t buf_size = range_size * k_size_row;
@@ -1652,6 +1804,9 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
             const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
             auto * v = layer.v_stream[cr.strm];
+            if (!v) {
+                continue;
+            }
 
             // Write value type
             const int32_t v_type_i = (int32_t) v->type;
@@ -1661,7 +1816,7 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
             const uint64_t v_size_row = ggml_row_size(v->type, n_embd_v_gqa);
             io.write(&v_size_row, sizeof(v_size_row));
 
-            // Read each range of cells of v_size length each into tmp_buf and write out
+            // Read each range of cells of v_size length and write out
             for (const auto & range : cr.data) {
                 const size_t range_size = range.second - range.first;
                 const size_t buf_size = range_size * v_size_row;
@@ -1678,6 +1833,9 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
             const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
             auto * v = layer.v_stream[cr.strm];
+            if (!v) {
+                continue;
+            }
 
             // Write value type
             const int32_t v_type_i = (int32_t) v->type;
@@ -1692,7 +1850,7 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
 
             // For each row, we get the element values of each cell
             for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
-                // Read each range of cells of v_size_el length each into tmp_buf and write out
+                // Read each range of cells of v_size_el length and write out
                 for (const auto & range : cr.data) {
                     const size_t range_size = range.second - range.first;
                     const size_t src_offset = (range.first + j * kv_size) * v_size_el;
@@ -1881,6 +2039,9 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
             const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
             auto * v = layer.v_stream[strm];
+            if (!v) {
+                continue;
+            }
 
             // Read type of value
             int32_t v_type_i_ref;
@@ -1922,6 +2083,9 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
             const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
             auto * v = layer.v_stream[strm];
+            if (!v) {
+                continue;
+            }
 
             // Read type of value
             int32_t v_type_i_ref;
diff --git a/src/llama-kv-cache.h b/src/llama-kv-cache.h
index 0c4ed64845..e194bf3e26 100644
--- a/src/llama-kv-cache.h
+++ b/src/llama-kv-cache.h
@@ -257,8 +257,6 @@ private:
     size_t size_k_bytes() const;
     size_t size_v_bytes() const;
 
-    bool is_masked_swa(llama_pos p0, llama_pos p1) const;
-
     ggml_tensor * build_rope_shift(
             const llama_cparams & cparams,
                    ggml_context * ctx,
diff --git a/src/llama-memory-hybrid-iswa.cpp b/src/llama-memory-hybrid-iswa.cpp
new file mode 100644
index 0000000000..411769672a
--- /dev/null
+++ b/src/llama-memory-hybrid-iswa.cpp
@@ -0,0 +1,275 @@
+#include "llama-memory-hybrid-iswa.h"
+
+#include "llama-impl.h"
+#include "llama-model.h"
+#include "llama-context.h"
+
+//
+// llama_memory_hybrid_iswa
+//
+
+llama_memory_hybrid_iswa::llama_memory_hybrid_iswa(
+        const llama_model & model,
+                            /* attn */
+                ggml_type   type_k,
+                ggml_type   type_v,
+                     bool   v_trans,
+                     bool   swa_full,
+                 uint32_t   kv_size,
+                 uint32_t   n_ubatch,
+                 uint32_t   n_pad,
+                            /* recurrent */
+                ggml_type   type_r,
+                ggml_type   type_s,
+                 uint32_t   rs_size,
+                            /* common */
+                 uint32_t   n_seq_max,
+                     bool   offload,
+                     bool   unified,
+                            /* layer filters */
+    const layer_filter_cb & filter_attn,
+    const layer_filter_cb & filter_recr) :
+    hparams(model.hparams),
+    mem_attn(new llama_kv_cache_iswa(
+        model,
+        type_k,
+        type_v,
+        v_trans,
+        offload,
+        swa_full,
+        unified,
+        kv_size,
+        n_seq_max,
+        n_ubatch,
+        n_pad,
+        filter_attn == nullptr ?
+            [&](int32_t il) { return !hparams.is_recurrent(il); }
+            : filter_attn,
+        nullptr
+    )),
+    mem_recr(new llama_memory_recurrent(
+        model,
+        type_r,
+        type_s,
+        offload,
+        rs_size,
+        n_seq_max,
+        filter_recr == nullptr ?
+            [&](int32_t il) { return hparams.is_recurrent(il); }
+            : filter_recr
+    )) {}
+
+llama_memory_context_ptr llama_memory_hybrid_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
+    do {
+        balloc.split_reset();
+
+        // follow the recurrent pattern for creating the ubatch splits
+        std::vector<llama_ubatch> ubatches;
+
+        while (true) {
+            llama_ubatch ubatch;
+
+            if (embd_all) {
+                // if all tokens are output, split by sequence
+                ubatch = balloc.split_seq(n_ubatch);
+            } else {
+                // TODO: non-sequential equal split can be done if using unified KV cache
+                //       for simplicity, we always use sequential equal split for now
+                ubatch = balloc.split_equal(n_ubatch, true);
+            }
+
+            if (ubatch.n_tokens == 0) {
+                break;
+            }
+
+            ubatches.push_back(std::move(ubatch)); // NOLINT
+        }
+
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
+            break;
+        }
+
+        // prepare the recurrent batches first
+        if (!mem_recr->prepare(ubatches)) {
+            // TODO: will the recurrent cache be in an undefined context at this point?
+            LLAMA_LOG_ERROR("%s: failed to prepare recurrent ubatches\n", __func__);
+            return std::make_unique<llama_memory_hybrid_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+        }
+
+        // prepare the attention cache (iswa version returns both base and swa slot infos)
+        auto sinfos_base = mem_attn->get_base()->prepare(ubatches);
+        if (sinfos_base.empty()) {
+            LLAMA_LOG_ERROR("%s: failed to prepare attention base ubatches\n", __func__);
+            return std::make_unique<llama_memory_hybrid_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+        }
+
+        auto sinfos_swa = mem_attn->get_swa()->prepare(ubatches);
+        if (sinfos_swa.empty()) {
+            LLAMA_LOG_ERROR("%s: failed to prepare attention swa ubatches\n", __func__);
+            return std::make_unique<llama_memory_hybrid_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+        }
+
+        return std::make_unique<llama_memory_hybrid_iswa_context>(
+                this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
+    } while(false);
+
+    return std::make_unique<llama_memory_hybrid_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+}
+
+llama_memory_context_ptr llama_memory_hybrid_iswa::init_full() {
+    return std::make_unique<llama_memory_hybrid_iswa_context>(this);
+}
+
+llama_memory_context_ptr llama_memory_hybrid_iswa::init_update(llama_context * lctx, bool optimize) {
+    return std::make_unique<llama_memory_hybrid_iswa_context>(this, lctx, optimize);
+}
+
+bool llama_memory_hybrid_iswa::get_can_shift() const {
+    // Shifting is trivially supported for recurrent
+    return mem_attn->get_can_shift();
+}
+
+void llama_memory_hybrid_iswa::clear(bool data) {
+    mem_attn->clear(data);
+    mem_recr->clear(data);
+}
+
+bool llama_memory_hybrid_iswa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+    // Try removing from the recurrent cache first since it may fail. If it does
+    // fail, the cache will not have been mutated.
+    if (!mem_recr->seq_rm(seq_id, p0, p1)) {
+        return false;
+    }
+    return mem_attn->seq_rm(seq_id, p0, p1);
+}
+
+void llama_memory_hybrid_iswa::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+    mem_attn->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+    mem_recr->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+}
+
+void llama_memory_hybrid_iswa::seq_keep(llama_seq_id seq_id) {
+    mem_attn->seq_keep(seq_id);
+    mem_recr->seq_keep(seq_id);
+}
+
+void llama_memory_hybrid_iswa::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
+    mem_attn->seq_add(seq_id, p0, p1, shift);
+    mem_recr->seq_add(seq_id, p0, p1, shift);
+}
+
+void llama_memory_hybrid_iswa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+    mem_attn->seq_div(seq_id, p0, p1, d);
+    mem_recr->seq_div(seq_id, p0, p1, d);
+}
+
+llama_pos llama_memory_hybrid_iswa::seq_pos_min(llama_seq_id seq_id) const {
+    // the min of the total cache is the max of the two caches' min values
+    return std::max(mem_attn->seq_pos_min(seq_id), mem_recr->seq_pos_min(seq_id));
+}
+
+llama_pos llama_memory_hybrid_iswa::seq_pos_max(llama_seq_id seq_id) const {
+    // the max of the total cache is the min of the two caches' max values
+    return std::min(mem_attn->seq_pos_max(seq_id), mem_recr->seq_pos_max(seq_id));
+}
+
+std::map<ggml_backend_buffer_type_t, size_t> llama_memory_hybrid_iswa::memory_breakdown() const {
+    std::map<ggml_backend_buffer_type_t, size_t> mb = mem_attn->memory_breakdown();
+    for (const auto & buft_size : mem_recr->memory_breakdown()) {
+        mb[buft_size.first] += buft_size.second;
+    }
+    return mb;
+}
+
+void llama_memory_hybrid_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    mem_attn->state_write(io, seq_id, flags);
+    mem_recr->state_write(io, seq_id, flags);
+}
+
+void llama_memory_hybrid_iswa::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    mem_attn->state_read(io, seq_id, flags);
+    mem_recr->state_read(io, seq_id, flags);
+}
+
+llama_kv_cache_iswa * llama_memory_hybrid_iswa::get_mem_attn() const {
+    return mem_attn.get();
+}
+
+llama_memory_recurrent * llama_memory_hybrid_iswa::get_mem_recr() const {
+    return mem_recr.get();
+}
+
+//
+// llama_memory_hybrid_iswa_context
+//
+
+llama_memory_hybrid_iswa_context::llama_memory_hybrid_iswa_context(llama_memory_status status) : status(status) {}
+
+llama_memory_hybrid_iswa_context::llama_memory_hybrid_iswa_context(llama_memory_hybrid_iswa * mem) :
+    ctx_attn(mem->get_mem_attn()->init_full()),
+    ctx_recr(mem->get_mem_recr()->init_full()),
+    status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
+}
+
+llama_memory_hybrid_iswa_context::llama_memory_hybrid_iswa_context(
+        llama_memory_hybrid_iswa * mem,
+                   llama_context * lctx,
+                            bool   optimize) :
+    ctx_attn(mem->get_mem_attn()->init_update(lctx, optimize)),
+    ctx_recr(mem->get_mem_recr()->init_update(lctx, optimize)),
+    status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
+}
+
+llama_memory_hybrid_iswa_context::llama_memory_hybrid_iswa_context(
+           llama_memory_hybrid_iswa * mem,
+                    slot_info_vec_t   sinfos_base,
+                    slot_info_vec_t   sinfos_swa,
+          std::vector<llama_ubatch>   ubatches) :
+    ubatches(std::move(ubatches)),
+    // note: here we copy the ubatches. not sure if this is ideal
+    ctx_attn(new llama_kv_cache_iswa_context(mem->get_mem_attn(), std::move(sinfos_base), std::move(sinfos_swa), this->ubatches)),
+    ctx_recr(new llama_memory_recurrent_context(mem->get_mem_recr(), this->ubatches)),
+    status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
+}
+
+bool llama_memory_hybrid_iswa_context::next() {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    ctx_attn->next();
+    ctx_recr->next();
+
+    if (++i_next >= ubatches.size()) {
+        return false;
+    }
+
+    return true;
+}
+
+bool llama_memory_hybrid_iswa_context::apply() {
+    assert(!llama_memory_status_is_fail(status));
+
+    bool res = true;
+
+    res = res & ctx_attn->apply();
+    res = res & ctx_recr->apply();
+
+    return res;
+}
+
+llama_memory_status llama_memory_hybrid_iswa_context::get_status() const {
+    return status;
+}
+
+const llama_ubatch & llama_memory_hybrid_iswa_context::get_ubatch() const {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+    return ubatches[i_next];
+}
+
+const llama_kv_cache_iswa_context * llama_memory_hybrid_iswa_context::get_attn() const {
+    return static_cast<const llama_kv_cache_iswa_context *>(ctx_attn.get());
+}
+
+const llama_memory_recurrent_context * llama_memory_hybrid_iswa_context::get_recr() const {
+    return static_cast<const llama_memory_recurrent_context *>(ctx_recr.get());
+}
diff --git a/src/llama-memory-hybrid-iswa.h b/src/llama-memory-hybrid-iswa.h
new file mode 100644
index 0000000000..807c8aac96
--- /dev/null
+++ b/src/llama-memory-hybrid-iswa.h
@@ -0,0 +1,140 @@
+#pragma once
+
+#include "llama-batch.h"
+#include "llama-graph.h"
+#include "llama-kv-cache-iswa.h"
+#include "llama-memory.h"
+#include "llama-memory-recurrent.h"
+
+#include <memory>
+#include <vector>
+
+//
+// llama_memory_hybrid_iswa
+//
+
+// utilizes instances of llama_memory_recurrent and llama_kv_cache_iswa to
+//   support models where each layer may be either attention-based (with SWA support) or recurrent
+
+class llama_memory_hybrid_iswa : public llama_memory_i {
+public:
+    llama_memory_hybrid_iswa(
+        const llama_model & model,
+                            /* attn */
+                ggml_type   type_k,
+                ggml_type   type_v,
+                     bool   v_trans,
+                     bool   swa_full,
+                 uint32_t   kv_size,
+                 uint32_t   n_ubatch,
+                 uint32_t   n_pad,
+                            /* recurrent */
+                ggml_type   type_r,
+                ggml_type   type_s,
+                 uint32_t   rs_size,
+                            /* common */
+                 uint32_t   n_seq_max,
+                     bool   offload,
+                     bool   unified,
+                            /* layer filters */
+    const layer_filter_cb & filter_attn = nullptr,
+    const layer_filter_cb & filter_recr = nullptr);
+
+    ~llama_memory_hybrid_iswa() = default;
+
+    //
+    // llama_memory_i
+    //
+
+    llama_memory_context_ptr init_batch(
+            llama_batch_allocr & balloc,
+            uint32_t n_ubatch,
+            bool embd_all) override;
+
+    llama_memory_context_ptr init_full() override;
+
+    llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
+
+    bool get_can_shift() const override;
+
+    void clear(bool data) override;
+
+    bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
+    void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
+    void seq_keep(llama_seq_id seq_id)                                                          override;
+    void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos shift) override;
+    void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) override;
+
+    llama_pos seq_pos_min(llama_seq_id seq_id) const override;
+    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
+
+    std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const override;
+
+    // state write/load
+
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0)       override;
+
+    //
+    // llama_memory_hybrid_iswa specific API
+    //
+
+    llama_kv_cache_iswa * get_mem_attn() const;
+    llama_memory_recurrent * get_mem_recr() const;
+
+private:
+    const llama_hparams & hparams;
+
+    const std::unique_ptr<llama_kv_cache_iswa> mem_attn;
+    const std::unique_ptr<llama_memory_recurrent> mem_recr;
+};
+
+class llama_memory_hybrid_iswa_context : public llama_memory_context_i {
+public:
+    using slot_info_vec_t = llama_kv_cache::slot_info_vec_t;
+
+    // init failure
+    explicit llama_memory_hybrid_iswa_context(llama_memory_status status);
+
+    // init full
+    explicit llama_memory_hybrid_iswa_context(llama_memory_hybrid_iswa * mem);
+
+    // init update
+    explicit llama_memory_hybrid_iswa_context(
+        llama_memory_hybrid_iswa * mem,
+                   llama_context * lctx,
+                            bool   optimize);
+
+    // init success
+    llama_memory_hybrid_iswa_context(
+           llama_memory_hybrid_iswa * mem,
+                    slot_info_vec_t   sinfos_base,
+                    slot_info_vec_t   sinfos_swa,
+          std::vector<llama_ubatch>   ubatches);
+
+    ~llama_memory_hybrid_iswa_context() = default;
+
+    bool next()  override;
+    bool apply() override;
+
+    llama_memory_status  get_status() const override;
+    const llama_ubatch & get_ubatch() const override;
+
+    //
+    // llama_memory_hybrid_iswa_context
+    //
+
+    const llama_kv_cache_iswa_context * get_attn() const;
+    const llama_memory_recurrent_context * get_recr() const;
+
+private:
+    // the index of the next ubatch to process
+    size_t i_next = 0;
+
+    std::vector<llama_ubatch> ubatches;
+
+    const llama_memory_context_ptr ctx_attn;
+    const llama_memory_context_ptr ctx_recr;
+
+    const llama_memory_status status;
+};
diff --git a/src/llama-memory-recurrent.cpp b/src/llama-memory-recurrent.cpp
index 812bf25304..f0038036dc 100644
--- a/src/llama-memory-recurrent.cpp
+++ b/src/llama-memory-recurrent.cpp
@@ -785,23 +785,21 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
     io.write(&s_trans, sizeof(s_trans));
     io.write(&n_layer,   sizeof(n_layer));
 
-    std::vector<uint8_t> tmp_buf;
-
-    // Iterate and write all the keys first, each row is a cell
+    // Iterate and write all the R tensors first, each row is a cell
     // Get whole range at a time
     for (uint32_t il = 0; il < n_layer; ++il) {
         // skip null layers (read_data will handle this by checking "r_l" and "s_l" for null)
         if (r_l[il] == nullptr) continue;
 
-        // Write key type
+        // Write R tensor type
         const int32_t r_type_i = (int32_t)r_l[il]->type;
         io.write(&r_type_i, sizeof(r_type_i));
 
-        // Write row size of key
+        // Write row size of R tensor
         const uint64_t r_size_row = ggml_row_size(r_l[il]->type, hparams.n_embd_r());
         io.write(&r_size_row, sizeof(r_size_row));
 
-        // Read each range of cells of k_size length each into tmp_buf and write out
+        // Write each range of cells of r_size_row length
         for (const auto & range : cell_ranges) {
             const size_t range_size = range.second - range.first;
             const size_t buf_size = range_size * r_size_row;
@@ -814,15 +812,15 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
             // skip null layers (read_data will handle this by checking "r_l" and "s_l" for null)
             if (s_l[il] == nullptr) continue;
 
-            // Write value type
+            // Write S tensor type
             const int32_t s_type_i = (int32_t)s_l[il]->type;
             io.write(&s_type_i, sizeof(s_type_i));
 
-            // Write row size of value
+            // Write row size of S tensor
             const uint64_t s_size_row = ggml_row_size(s_l[il]->type, hparams.n_embd_s());
             io.write(&s_size_row, sizeof(s_size_row));
 
-            // Read each range of cells of s_size length each into tmp_buf and write out
+            // Write each range of S tensor rows
             for (const auto & range : cell_ranges) {
                 const size_t range_size = range.second - range.first;
                 const size_t buf_size = range_size * s_size_row;
@@ -830,7 +828,7 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
             }
         }
     } else {
-        // When v is transposed, we also need the element size and get the element ranges from each row
+        // When S tensor is transposed, we also need the element size and get the element ranges from each row
         const uint32_t mem_size = size;
         for (uint32_t il = 0; il < n_layer; ++il) {
             // skip null layers (read_data will handle this by checking "r_l" and "s_l" for null)
@@ -838,7 +836,7 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
 
             const uint32_t n_embd_s = hparams.n_embd_s();
 
-            // Write value type
+            // Write S tensor type
             const int32_t s_type_i = (int32_t)s_l[il]->type;
             io.write(&s_type_i, sizeof(s_type_i));
 
@@ -851,7 +849,7 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
 
             // For each row, we get the element values of each cell
             for (uint32_t j = 0; j < n_embd_s; ++j) {
-                // Read each range of cells of v_size_el length each into tmp_buf and write out
+                // Write each range of cells of s_size_el length
                 for (const auto & range : cell_ranges) {
                     const size_t range_size = range.second - range.first;
                     const size_t src_offset = (range.first + j * mem_size) * s_size_el;
diff --git a/src/llama-mmap.cpp b/src/llama-mmap.cpp
index fe0847fe1a..0261e4c72c 100644
--- a/src/llama-mmap.cpp
+++ b/src/llama-mmap.cpp
@@ -265,7 +265,8 @@ struct llama_file::impl {
                         continue;  // Interrupted by signal, retry
                     }
                     // Fallback to std::fread in case the DMA controller cannot access the buffer
-                    if (errno == EFAULT) {
+                    if (errno == EFAULT || errno == EINVAL) {
+                        LLAMA_LOG_WARN("%s: Falling back to buffered IO due to %s\n", __func__, strerror(errno));
                         auto curr_off = tell();
                         close(fd);
                         fd = -1;
@@ -384,6 +385,9 @@ int llama_file::file_id() const {
 #ifdef _WIN32
     return _fileno(pimpl->fp);
 #else
+    if (pimpl->fd != -1) {
+        return pimpl->fd;
+    }
 #if defined(fileno)
     return fileno(pimpl->fp);
 #else
diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
index 300a322c51..1501e392ca 100644
--- a/src/llama-model-loader.cpp
+++ b/src/llama-model-loader.cpp
@@ -539,12 +539,18 @@ llama_model_loader::llama_model_loader(
     files.emplace_back(new llama_file(fname.c_str(), "rb", use_direct_io));
     contexts.emplace_back(ctx);
 
-    use_direct_io = use_direct_io && files.back()->has_direct_io();
+    if (use_mmap && use_direct_io) {
+        if (files.back()->has_direct_io()) {
+            LLAMA_LOG_WARN("%s: direct I/O is enabled, disabling mmap\n", __func__);
+            use_mmap = false;
+        } else {
+            LLAMA_LOG_WARN("%s: direct I/O is not available, using mmap\n", __func__);
+            use_direct_io = false;
 
-    // Disable mmap in case Direct I/O is enabled and available
-    if (use_direct_io && use_mmap) {
-        use_mmap = false;
-        LLAMA_LOG_WARN("%s: direct I/O is enabled, disabling mmap\n", __func__);
+            // reopen file using std::fopen for mmap
+            files.pop_back();
+            files.emplace_back(new llama_file(fname.c_str(), "rb", false));
+        }
     }
 
     // Save tensors data offset of the main file.
diff --git a/src/llama-model-saver.cpp b/src/llama-model-saver.cpp
index ae27c71ce2..36e353074e 100644
--- a/src/llama-model-saver.cpp
+++ b/src/llama-model-saver.cpp
@@ -146,8 +146,8 @@ void llama_model_saver::add_kv_from_model() {
     add_kv(LLM_KV_VOCAB_SIZE,                        vocab.n_tokens());
     add_kv(LLM_KV_CONTEXT_LENGTH,                    hparams.n_ctx_train);
     add_kv(LLM_KV_EMBEDDING_LENGTH,                  hparams.n_embd);
-    if (hparams.n_embd_out > 0) {
-        add_kv(LLM_KV_EMBEDDING_LENGTH_OUT,          hparams.n_embd_out);
+    if (hparams.n_embd_out_impl > 0) {
+        add_kv(LLM_KV_EMBEDDING_LENGTH_OUT,          hparams.n_embd_out_impl);
     }
     add_kv(LLM_KV_BLOCK_COUNT,                       hparams.n_layer);
     add_kv(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead);
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index 94c47dc248..72490a89b5 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -8,6 +8,7 @@
 #include "llama-kv-cache.h"
 #include "llama-kv-cache-iswa.h"
 #include "llama-memory-hybrid.h"
+#include "llama-memory-hybrid-iswa.h"
 #include "llama-memory-recurrent.h"
 
 #include "ggml-cpp.h"
@@ -511,7 +512,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
     ml.get_key(LLM_KV_CONTEXT_LENGTH,          hparams.n_ctx_train);
     ml.get_key(LLM_KV_EMBEDDING_LENGTH,        hparams.n_embd);
-    ml.get_key(LLM_KV_EMBEDDING_LENGTH_OUT,    hparams.n_embd_out, false);
+    ml.get_key(LLM_KV_EMBEDDING_LENGTH_OUT,    hparams.n_embd_out_impl, false);
     ml.get_key(LLM_KV_BLOCK_COUNT,             hparams.n_layer);
     ml.get_key(LLM_KV_EXPERT_COUNT,            hparams.n_expert,        false);
     ml.get_key(LLM_KV_EXPERT_USED_COUNT,       hparams.n_expert_used,   false);
@@ -1696,15 +1697,16 @@ void llama_model::load_hparams(llama_model_loader & ml) {
         case LLM_ARCH_DEEPSEEK2:
             {
                 // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
-                bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
+                const bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
+
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead);
                 if (!is_lite) {
                     ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q);
                 }
                 ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,     hparams.n_lora_kv);
-                ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   hparams.n_embd_head_k_mla, false);
-                ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla, false);
+                ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   hparams.n_embd_head_k_mla_impl, false);
+                ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl, false);
                 ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
                 ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
                 ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,       hparams.expert_weights_scale, false);
@@ -1713,7 +1715,12 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
                     // for compatibility with existing DeepSeek V2 and V2.5 GGUFs
                     // that have no expert_gating_func model parameter set
-                    hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX;
+                    if ((hparams.n_layer == 47 || hparams.n_layer == 48) && n_vocab == 154880) {
+                        // GLM 4.7 Lite
+                        hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
+                    } else {
+                        hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX;
+                    }
                 }
 
                 if (ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul, 0.0f)) {
@@ -1730,6 +1737,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
 
                 switch (hparams.n_layer) {
                     case 27: type = LLM_TYPE_16B; break;
+                    case 47: type = LLM_TYPE_30B_A3B; break;
                     case 60: type = LLM_TYPE_236B; break;
                     case 61: type = LLM_TYPE_671B; break;
                     default: type = LLM_TYPE_UNKNOWN;
@@ -4903,14 +4911,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                 } break;
             case LLM_ARCH_DEEPSEEK2:
                 {
-                    // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
-                    const bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
-
-                    const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
+                    const bool is_mla = hparams.is_mla();
 
                     // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
-                    const int64_t n_embd_head_k_mla = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
-                    const int64_t n_embd_head_v_mla = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
+                    const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla();
+                    const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla();
 
                     const int64_t n_embd_head_qk_rope = hparams.n_rot;
                     const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
@@ -4935,13 +4940,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         auto & layer = layers[i];
 
                         layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
-                        if (!is_lite) {
+                        if (q_lora_rank > 0) {
                             layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, 0);
                         }
 
                         layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, 0);
 
-                        if (!is_lite) {
+                        if (q_lora_rank > 0) {
                             layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, 0);
                             layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, 0);
                         } else {
@@ -6591,7 +6596,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     }
 
                     // for LFM2-ColBert-350M
-                    dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.get_n_embd_out()}, TENSOR_NOT_REQUIRED);
+                    dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.n_embd_out()}, TENSOR_NOT_REQUIRED);
                 } break;
             case LLM_ARCH_SMALLTHINKER:
                 {
@@ -7310,8 +7315,8 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: n_layer_dense_lead    = %d\n",     __func__, hparams.n_layer_dense_lead);
         LLAMA_LOG_INFO("%s: n_lora_q              = %d\n",     __func__, hparams.n_lora_q);
         LLAMA_LOG_INFO("%s: n_lora_kv             = %d\n",     __func__, hparams.n_lora_kv);
-        LLAMA_LOG_INFO("%s: n_embd_head_k_mla     = %d\n",     __func__, hparams.n_embd_head_k_mla);
-        LLAMA_LOG_INFO("%s: n_embd_head_v_mla     = %d\n",     __func__, hparams.n_embd_head_v_mla);
+        LLAMA_LOG_INFO("%s: n_embd_head_k_mla     = %d\n",     __func__, hparams.n_embd_head_k_mla());
+        LLAMA_LOG_INFO("%s: n_embd_head_v_mla     = %d\n",     __func__, hparams.n_embd_head_v_mla());
         LLAMA_LOG_INFO("%s: n_ff_exp              = %d\n",     __func__, hparams.n_ff_exp);
         LLAMA_LOG_INFO("%s: n_expert_shared       = %d\n",     __func__, hparams.n_expert_shared);
         LLAMA_LOG_INFO("%s: expert_weights_scale  = %.1f\n",   __func__, hparams.expert_weights_scale);
@@ -7523,23 +7528,44 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                         };
                     }
 
-                    res = new llama_memory_hybrid(
-                        /* model             */ *this,
-                        /* attn_type_k       */ params.type_k,
-                        /* attn_type_v       */ params.type_v,
-                        /* attn_v_trans      */ !cparams.flash_attn,
-                        /* attn_kv_size      */ cparams.n_ctx,
-                        /* attn_n_pad        */ 1,
-                        /* attn_n_swa        */ hparams.n_swa,
-                        /* attn_swa_type     */ hparams.swa_type,
-                        /* recurrent_type_k  */ GGML_TYPE_F32,
-                        /* recurrent_type_v  */ GGML_TYPE_F32,
-                        /* recurrent_kv_size */ std::max((uint32_t) 1, cparams.n_seq_max),
-                        /* n_seq_max         */ cparams.n_seq_max,
-                        /* offload           */ cparams.offload_kqv,
-                        /* unified           */ cparams.kv_unified,
-                        /* filter_attn       */ std::move(filter_attn),
-                        /* filter_recr       */ std::move(filter_recr));
+                    if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
+                        // Use hybrid-iswa for hybrid models with SWA
+                        res = new llama_memory_hybrid_iswa(
+                            /* model             */ *this,
+                            /* attn_type_k       */ params.type_k,
+                            /* attn_type_v       */ params.type_v,
+                            /* attn_v_trans      */ !cparams.flash_attn,
+                            /* attn_swa_full     */ params.swa_full,
+                            /* attn_kv_size      */ cparams.n_ctx,
+                            /* attn_n_ubatch     */ cparams.n_ubatch,
+                            /* attn_n_pad        */ 1,
+                            /* recurrent_type_r  */ GGML_TYPE_F32,
+                            /* recurrent_type_s  */ GGML_TYPE_F32,
+                            /* recurrent_rs_size */ std::max((uint32_t) 1, cparams.n_seq_max),
+                            /* n_seq_max         */ cparams.n_seq_max,
+                            /* offload           */ cparams.offload_kqv,
+                            /* unified           */ cparams.kv_unified,
+                            /* filter_attn       */ std::move(filter_attn),
+                            /* filter_recr       */ std::move(filter_recr));
+                    } else {
+                        res = new llama_memory_hybrid(
+                            /* model             */ *this,
+                            /* attn_type_k       */ params.type_k,
+                            /* attn_type_v       */ params.type_v,
+                            /* attn_v_trans      */ !cparams.flash_attn,
+                            /* attn_kv_size      */ cparams.n_ctx,
+                            /* attn_n_pad        */ 1,
+                            /* attn_n_swa        */ hparams.n_swa,
+                            /* attn_swa_type     */ hparams.swa_type,
+                            /* recurrent_type_k  */ GGML_TYPE_F32,
+                            /* recurrent_type_v  */ GGML_TYPE_F32,
+                            /* recurrent_kv_size */ std::max((uint32_t) 1, cparams.n_seq_max),
+                            /* n_seq_max         */ cparams.n_seq_max,
+                            /* offload           */ cparams.offload_kqv,
+                            /* unified           */ cparams.kv_unified,
+                            /* filter_attn       */ std::move(filter_attn),
+                            /* filter_recr       */ std::move(filter_recr));
+                    }
                 } else {
                     llama_memory_i::layer_reuse_cb reuse = nullptr;
 
@@ -8099,7 +8125,7 @@ llama_model_params llama_model_default_params() {
         /*.kv_overrides                =*/ nullptr,
         /*.vocab_only                  =*/ false,
         /*.use_mmap                    =*/ true,
-        /*.use_direct_io               =*/ true,
+        /*.use_direct_io               =*/ false,
         /*.use_mlock                   =*/ false,
         /*.check_tensors               =*/ false,
         /*.use_extra_bufts             =*/ true,
@@ -8135,7 +8161,7 @@ int32_t llama_model_n_embd_inp(const llama_model * model) {
 }
 
 int32_t llama_model_n_embd_out(const llama_model * model) {
-    return model->hparams.get_n_embd_out();
+    return model->hparams.n_embd_out();
 }
 
 int32_t llama_model_n_layer(const llama_model * model) {
diff --git a/src/llama-quant.cpp b/src/llama-quant.cpp
index 048d65a75c..776222cb6f 100644
--- a/src/llama-quant.cpp
+++ b/src/llama-quant.cpp
@@ -422,57 +422,6 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
         ++qs.i_ffn_up;
     }
 
-    //    if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
-    //}
-    // IK: let's remove this, else Q2_K is almost the same as Q3_K_S
-    //else if (name.find("ffn_gate") != std::string::npos || name.find("ffn_up") != std::string::npos) {
-    //    if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
-    //}
-    // This can be used to reduce the size of the Q5_K_S model.
-    // The associated PPL increase is fully in line with the size reduction
-    //else {
-    //    if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q4_K;
-    //}
-    bool convert_incompatible_tensor = false;
-    {
-        const int64_t nx = tensor->ne[0];
-        const int64_t ny = tensor->ne[1];
-        const int64_t qk_k = ggml_blck_size(new_type);
-
-        if (nx % qk_k != 0) {
-            LLAMA_LOG_WARN("\n\n%s : tensor cols %" PRId64 " x %" PRId64 " are not divisible by %" PRId64 ", required for %s", __func__, nx, ny, qk_k, ggml_type_name(new_type));
-            convert_incompatible_tensor = true;
-        } else {
-            ++qs.n_k_quantized;
-        }
-    }
-
-    if (convert_incompatible_tensor) {
-        switch (new_type) {
-            case GGML_TYPE_TQ1_0:
-            case GGML_TYPE_TQ2_0:  new_type = GGML_TYPE_Q4_0; break;  // TODO: use a symmetric type instead
-            case GGML_TYPE_IQ2_XXS:
-            case GGML_TYPE_IQ2_XS:
-            case GGML_TYPE_IQ2_S:
-            case GGML_TYPE_IQ3_XXS:
-            case GGML_TYPE_IQ3_S:
-            case GGML_TYPE_IQ1_S:
-            case GGML_TYPE_IQ1_M:
-            case GGML_TYPE_Q2_K:
-            case GGML_TYPE_Q3_K:
-            case GGML_TYPE_IQ4_XS: new_type = GGML_TYPE_IQ4_NL; break;
-            case GGML_TYPE_Q4_K:   new_type = GGML_TYPE_Q5_0;   break;
-            case GGML_TYPE_Q5_K:   new_type = GGML_TYPE_Q5_1;   break;
-            case GGML_TYPE_Q6_K:   new_type = GGML_TYPE_Q8_0;   break;
-            default: throw std::runtime_error("\nUnsupported tensor size encountered\n");
-        }
-        if (tensor->ne[0] % ggml_blck_size(new_type) != 0) {
-            new_type = GGML_TYPE_F16;
-        }
-        LLAMA_LOG_WARN(" - using fallback quantization %s\n", ggml_type_name(new_type));
-        ++qs.n_fallback;
-    }
-
     return new_type;
 }
 
@@ -596,7 +545,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     }
 
     std::vector<std::string> splits = {};
-    llama_model_loader ml(fname_inp, splits, use_mmap, /*use_direct_io*/ true, /*check_tensors*/ true, /*no_alloc*/ false, kv_overrides, nullptr);
+    llama_model_loader ml(fname_inp, splits, use_mmap, /*use_direct_io*/ false, /*check_tensors*/ true, /*no_alloc*/ false, kv_overrides, nullptr);
     ml.init_mappings(false); // no prefetching
 
     llama_model model(llama_model_default_params());
@@ -875,21 +824,69 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 
             // get more optimal quantization type based on the tensor shape, layer, etc.
             if (!params->pure && ggml_is_quantized(default_type)) {
-                int fallback = qs.n_fallback;
-                new_type = llama_tensor_get_type(qs, new_type, tensor, ftype);
-                // unless the user specifies a type, and the tensor geometry will not require fallback quantisation
-                if (params->tensor_types && qs.n_fallback - fallback == 0) {
+                // if the user provided tensor types - use those
+                bool manual = false;
+                if (params->tensor_types) {
                     const std::vector<tensor_quantization> & tensor_types = *static_cast<const std::vector<tensor_quantization> *>(params->tensor_types);
                     const std::string tensor_name(tensor->name);
                     for (const auto & [tname, qtype] : tensor_types) {
                         if (std::regex pattern(tname); std::regex_search(tensor_name, pattern)) {
                             if  (qtype != new_type) {
-                                LLAMA_LOG_DEBUG("(overriding %s) ", ggml_type_name(new_type));
+                                LLAMA_LOG_WARN("(manual override: %s -> %s) ", ggml_type_name(new_type), ggml_type_name(qtype));
                                 new_type = qtype; // if two or more types are specified for the same tensor, the last match wins
+                                manual = true;
+                                break;
                             }
                         }
                     }
                 }
+
+                // if not manual - use the standard logic for choosing the quantization type based on the selected mixture
+                if (!manual) {
+                    new_type = llama_tensor_get_type(qs, new_type, tensor, ftype);
+                }
+
+                // incompatible tensor shapes are handled here - fallback to a compatible type
+                {
+                    bool convert_incompatible_tensor = false;
+
+                    const int64_t nx = tensor->ne[0];
+                    const int64_t ny = tensor->ne[1];
+                    const int64_t qk_k = ggml_blck_size(new_type);
+
+                    if (nx % qk_k != 0) {
+                        LLAMA_LOG_WARN("\n\n%s : tensor cols %" PRId64 " x %" PRId64 " are not divisible by %" PRId64 ", required for %s", __func__, nx, ny, qk_k, ggml_type_name(new_type));
+                        convert_incompatible_tensor = true;
+                    } else {
+                        ++qs.n_k_quantized;
+                    }
+
+                    if (convert_incompatible_tensor) {
+                        switch (new_type) {
+                            case GGML_TYPE_TQ1_0:
+                            case GGML_TYPE_TQ2_0:  new_type = GGML_TYPE_Q4_0; break;  // TODO: use a symmetric type instead
+                            case GGML_TYPE_IQ2_XXS:
+                            case GGML_TYPE_IQ2_XS:
+                            case GGML_TYPE_IQ2_S:
+                            case GGML_TYPE_IQ3_XXS:
+                            case GGML_TYPE_IQ3_S:
+                            case GGML_TYPE_IQ1_S:
+                            case GGML_TYPE_IQ1_M:
+                            case GGML_TYPE_Q2_K:
+                            case GGML_TYPE_Q3_K:
+                            case GGML_TYPE_IQ4_XS: new_type = GGML_TYPE_IQ4_NL; break;
+                            case GGML_TYPE_Q4_K:   new_type = GGML_TYPE_Q5_0;   break;
+                            case GGML_TYPE_Q5_K:   new_type = GGML_TYPE_Q5_1;   break;
+                            case GGML_TYPE_Q6_K:   new_type = GGML_TYPE_Q8_0;   break;
+                            default: throw std::runtime_error("\nUnsupported tensor size encountered\n");
+                        }
+                        if (tensor->ne[0] % ggml_blck_size(new_type) != 0) {
+                            new_type = GGML_TYPE_F16;
+                        }
+                        LLAMA_LOG_WARN(" - using fallback quantization %s\n", ggml_type_name(new_type));
+                        ++qs.n_fallback;
+                    }
+                }
             }
             if (params->token_embedding_type < GGML_TYPE_COUNT && strcmp(tensor->name, "token_embd.weight") == 0) {
                 new_type = params->token_embedding_type;
diff --git a/src/llama.cpp b/src/llama.cpp
index f1096d960e..6da90d6f1f 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -311,8 +311,12 @@ static void llama_params_fit_impl(
                             __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
                     }
                 } else {
-                    LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
-                        __func__, hp_nct, n_ctx_min);
+                    if (n_ctx_min == UINT32_MAX) {
+                        LLAMA_LOG_INFO("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, hp_nct);
+                    } else {
+                        LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
+                            __func__, hp_nct, n_ctx_min);
+                    }
                 }
             } else {
                 LLAMA_LOG_INFO("%s: context size set by user to %" PRIu32 " -> no change\n", __func__, cparams->n_ctx);
@@ -1091,25 +1095,55 @@ int32_t llama_chat_apply_template(
 // model split
 //
 
-int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count) {
+int32_t llama_split_path(
+    char * split_path,
+    size_t maxlen,
+    const char * path_prefix,
+    int32_t split_no,
+    int32_t split_count) {
+
     static const char * const SPLIT_PATH_FORMAT = "%s-%05d-of-%05d.gguf";
-    if (snprintf(split_path, maxlen, SPLIT_PATH_FORMAT, path_prefix, split_no + 1, split_count)) {
-        return strlen(split_path);
+
+    const int written = snprintf(
+        split_path,
+        maxlen,
+        SPLIT_PATH_FORMAT,
+        path_prefix,
+        split_no + 1,
+        split_count
+    );
+
+    if (written < 0 || (size_t) written >= maxlen) {
+        return 0;
     }
-    return 0;
+
+    return (int32_t) written;
 }
 
-int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count) {
-    std::string str_split_path(split_path);
-    char postfix[32];
-    snprintf(postfix, 32, "-%05d-of-%05d.gguf", split_no + 1, split_count);
-    std::string str_postfix(postfix);
+int32_t llama_split_prefix(
+    char * split_prefix,
+    size_t maxlen,
+    const char * split_path,
+    int32_t split_no,
+    int32_t split_count) {
 
-    // check if split_prefix ends with postfix
-    int size_prefix = str_split_path.size() - str_postfix.size();
-    if (size_prefix > 0 && str_split_path.find(str_postfix, size_prefix) != std::string::npos) {
-        snprintf(split_prefix, std::min((size_t) size_prefix + 1, maxlen), "%s", split_path);
-        return size_prefix;
+    const std::string str_split_path(split_path);
+
+    char postfix[32];
+    snprintf(postfix, sizeof(postfix), "-%05d-of-%05d.gguf", split_no + 1, split_count);
+
+    const std::string str_postfix(postfix);
+    if (str_split_path.size() <= str_postfix.size()) {
+        return 0;
+    }
+
+    const size_t size_prefix = str_split_path.size() - str_postfix.size();
+
+    if (str_split_path.compare(size_prefix, std::string::npos, str_postfix) == 0) {
+        const size_t copy_len = std::min(size_prefix + 1, maxlen);
+        snprintf(split_prefix, copy_len, "%s", split_path);
+
+        return (int32_t) size_prefix;
     }
 
     return 0;
diff --git a/src/models/deepseek2.cpp b/src/models/deepseek2.cpp
index ca63a62ad1..297dca5136 100644
--- a/src/models/deepseek2.cpp
+++ b/src/models/deepseek2.cpp
@@ -2,14 +2,11 @@
 
 llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
-    // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
-    bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
-
-    const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
+    const bool is_mla = hparams.is_mla();
 
     // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
-    const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
-    const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
+    const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();
+    const int64_t n_embd_head_v = hparams.n_embd_head_v_mla();
 
     const int64_t n_embd_head_qk_rope = hparams.n_rot;
     const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
@@ -43,7 +40,8 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
-    auto * inp_attn = build_attn_inp_kv();
+    auto * inp_attn_kv = !is_mla ? build_attn_inp_kv() : nullptr;
+    auto * inp_attn_k  =  is_mla ? build_attn_inp_k()  : nullptr;
 
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
@@ -57,6 +55,9 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
         // self_attention
         {
             ggml_tensor * q = NULL;
+
+            const bool is_lite = model.layers[il].wq;
+
             if (!is_lite) {
                 q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
                 cb(q, "q", il);
@@ -124,14 +125,14 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
 
                 // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}
                 // note: rope must go first for in-place context shifting in build_rope_shift()
-                ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope_absorbed, 0);
+                ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
                 cb(Qcur, "Qcur", il);
 
                 kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
                 cb(kv_cmpr, "kv_cmpr_reshape", il);
 
                 // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
-                ggml_tensor * Kcur = ggml_concat(ctx0, k_pe, kv_cmpr, 0);
+                ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);
                 cb(Kcur, "Kcur", il);
 
                 // {kv_lora_rank, 1, n_tokens}
@@ -145,7 +146,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                 }
 
                 // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
-                cur = build_attn(inp_attn,
+                cur = build_attn(inp_attn_k,
                         model.layers[il].wo, NULL,
                         Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);
             } else {
@@ -169,11 +170,10 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                 Vcur = ggml_cont(ctx0, Vcur);
                 cb(Vcur, "Vcur_cont", il);
 
-                // note: rope must go first for in-place context shifting in build_rope_shift()
-                ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope, 0);
+                ggml_tensor * Qcur = ggml_concat(ctx0, q_nope, q_pe, 0);
                 cb(Qcur, "Qcur", il);
 
-                ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0);
+                ggml_tensor * Kcur = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
                 cb(Kcur, "Kcur", il);
 
                 if (inp_attn_scale) {
@@ -183,7 +183,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                 }
 
                 // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
-                cur = build_attn(inp_attn,
+                cur = build_attn(inp_attn_kv,
                             model.layers[il].wo, NULL,
                             Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
             }
diff --git a/src/models/gemma3n-iswa.cpp b/src/models/gemma3n-iswa.cpp
index 51acab1490..7db6d3bf4e 100644
--- a/src/models/gemma3n-iswa.cpp
+++ b/src/models/gemma3n-iswa.cpp
@@ -245,12 +245,12 @@ ggml_tensor * llm_build_gemma3n_iswa::view_2d_slice(ggml_tensor * x, int idx) {
 // equivalent to get_per_layer_inputs() in python code
 // output shape: [n_embd_altup, n_layer, n_tokens]
 ggml_tensor * llm_build_gemma3n_iswa::get_per_layer_inputs() {
-    auto inp = std::make_unique<llm_graph_input_embd>();
+    auto inp = std::make_unique<llm_graph_input_embd>(n_embd);
     ggml_tensor * inp_per_layer;
     if (ubatch.token) {
         inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
         ggml_set_input(inp->tokens);
-        res->t_tokens = inp->tokens;
+        res->t_inp_tokens = inp->tokens;
         inp_per_layer = ggml_get_rows(ctx0, model.tok_embd_per_layer, inp->tokens);
         inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, n_tokens);
         inp_per_layer = ggml_scale(ctx0, inp_per_layer, sqrtf((float) n_embd_altup));
diff --git a/src/models/minicpm3.cpp b/src/models/minicpm3.cpp
index f374a9fd03..297cc34ba5 100644
--- a/src/models/minicpm3.cpp
+++ b/src/models/minicpm3.cpp
@@ -9,6 +9,7 @@ llm_build_minicpm3::llm_build_minicpm3(const llama_model & model, const llm_grap
 
     const uint32_t n_embd_head_qk_rope = hparams.n_rot;
     const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+
     const uint32_t kv_lora_rank = hparams.n_lora_kv;
 
     ggml_tensor * cur;
diff --git a/src/models/nemotron-h.cpp b/src/models/nemotron-h.cpp
index eb135e63f1..079c730ac2 100644
--- a/src/models/nemotron-h.cpp
+++ b/src/models/nemotron-h.cpp
@@ -67,7 +67,7 @@ ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor *
                                                           const llama_model &       model,
                                                           const int64_t             n_embd_head,
                                                           const int                 il) {
-    // compute Q and K and (optionally) RoPE them
+    // compute Q and K
     ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
     cb(Qcur, "Qcur", il);
     if (model.layers[il].bq) {
diff --git a/src/models/plm.cpp b/src/models/plm.cpp
index 481cbba690..612a487c56 100644
--- a/src/models/plm.cpp
+++ b/src/models/plm.cpp
@@ -5,6 +5,7 @@ llm_build_plm::llm_build_plm(const llama_model & model, const llm_graph_params &
 
     const uint32_t n_embd_head_qk_rope = hparams.n_rot;
     const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
+
     const uint32_t kv_lora_rank = hparams.n_lora_kv;
 
     ggml_tensor * cur;
diff --git a/src/models/qwen3vl-moe.cpp b/src/models/qwen3vl-moe.cpp
index f72f80a837..e5e1a2150c 100644
--- a/src/models/qwen3vl-moe.cpp
+++ b/src/models/qwen3vl-moe.cpp
@@ -2,7 +2,8 @@
 
 llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const size_t n_deepstack_layers = hparams.n_deepstack_layers;
-    const int64_t n_embd = hparams.n_embd;
+
+    const int64_t n_embd      = hparams.n_embd;
     const int64_t n_embd_head = hparams.n_embd_head_v;
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -16,17 +17,6 @@ llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_
     int sections[4];
     std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
 
-    std::vector<ggml_tensor *> deepstack_features(n_deepstack_layers, nullptr);
-
-    if (ubatch.embd) {
-        // Image input: split main embd and deepstack embds
-        ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], 0);
-        for (size_t i = 0; i < n_deepstack_layers; i++) {
-            deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float));
-        }
-        inpL = inpL_main;
-    }
-
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
@@ -120,8 +110,9 @@ llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_
         cur = build_cvec(cur, il);
         cb(cur, "l_out", il);
 
-        if (ubatch.embd && (size_t)il < n_deepstack_layers) {
-            cur = ggml_add(ctx0, cur, deepstack_features[il]);
+        if (il < (int) n_deepstack_layers) {
+            ggml_tensor * ds = ggml_view_2d(ctx0, res->t_inp_embd, n_embd, n_tokens, res->t_inp_embd->nb[1], (il + 1) * n_embd * sizeof(float));
+            cur = ggml_add(ctx0, cur, ds);
             cb(cur, "deepstack_out", il);
         }
 
diff --git a/src/models/qwen3vl.cpp b/src/models/qwen3vl.cpp
index 0bae52239c..0f8315b324 100644
--- a/src/models/qwen3vl.cpp
+++ b/src/models/qwen3vl.cpp
@@ -2,7 +2,8 @@
 
 llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const size_t n_deepstack_layers = hparams.n_deepstack_layers;
-    const int64_t n_embd = hparams.n_embd;
+
+    const int64_t n_embd      = hparams.n_embd;
     const int64_t n_embd_head = hparams.n_embd_head_v;
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
@@ -16,17 +17,6 @@ llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_
     int sections[4];
     std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
 
-    std::vector<ggml_tensor *> deepstack_features(n_deepstack_layers, nullptr);
-
-    if (ubatch.embd) {
-        // Image input: split main embd and deepstack embds
-        ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], 0);
-        for (size_t i = 0; i < n_deepstack_layers; i++) {
-            deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float));
-        }
-        inpL = inpL_main;
-    }
-
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
@@ -113,8 +103,9 @@ llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_
         cur = build_cvec(cur, il);
         cb(cur, "l_out", il);
 
-        if (ubatch.embd && (size_t)il < n_deepstack_layers) {
-            cur = ggml_add(ctx0, cur, deepstack_features[il]);
+        if (il < (int) n_deepstack_layers) {
+            ggml_tensor * ds = ggml_view_2d(ctx0, res->t_inp_embd, n_embd, n_tokens, res->t_inp_embd->nb[1], (il + 1) * n_embd * sizeof(float));
+            cur = ggml_add(ctx0, cur, ds);
             cb(cur, "deepstack_out", il);
         }
 
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index 3eae18eefd..c9436c5995 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -187,6 +187,7 @@ llama_build_and_test(test-chat-parser.cpp)
 llama_build_and_test(test-chat-peg-parser.cpp peg-parser/simple-tokenize.cpp)
 llama_build_and_test(test-chat-template.cpp)
 llama_build_and_test(test-jinja.cpp)
+llama_test(test-jinja NAME test-jinja-py ARGS -py LABEL python)
 llama_build_and_test(test-json-partial.cpp)
 llama_build_and_test(test-log.cpp)
 llama_build_and_test(
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 6bb781737e..411467e968 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -6122,7 +6122,19 @@ struct test_flash_attn_ext : public test_case {
         ggml_tensor * k = create_permuted(type_KV,       hsk_padded, kv, nh,         nr23[1], true); // the K tensor is usually a view of the K cache
         ggml_set_name(k, "k");
 
-        ggml_tensor * v = create_permuted(type_KV,       hsv_padded, kv, nh,         nr23[1], true); // the V tensor is usually a view of the V cache
+        ggml_tensor * v = nullptr;
+        if (hsk_padded == 576 && hsv_padded == 512) {
+            // TODO: this branch should become a separate test case parameter instead of hardcoding this for these head shapes
+
+            // in this branch, the V cache is sub-view of the K cache. this is used by some MLA-based models
+            // for more info:
+            //   - https://github.com/ggml-org/llama.cpp/pull/13435
+            //   - https://github.com/ggml-org/llama.cpp/pull/18953#issuecomment-3774948392
+            //   - https://github.com/ggml-org/llama.cpp/pull/18986
+            v = ggml_view_4d(ctx, k, hsv_padded, kv, nh, nr23[1], k->nb[1], k->nb[2], k->nb[3], 0);
+        } else {
+            v = create_permuted(type_KV,       hsv_padded, kv, nh,         nr23[1], true); // the V tensor is usually a view of the V cache
+        }
         ggml_set_name(v, "v");
 
         ggml_tensor * m = nullptr;
@@ -8201,11 +8213,13 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                         if (!mask && max_bias > 0.0f) continue;
                         for (float logit_softcap : {0.0f, 10.0f}) {
                             if (hsk != 128 && logit_softcap != 0.0f) continue;
-                            for (int nh : { 4, }) {
+                            for (int nh : { 1, 4 }) {
+                                if (nh == 1 && hsk != 576) continue; // GLM 4.7 Flash
                                 for (int nr3 : { 1, 3, }) {
                                     if (hsk > 64 && nr3 > 1) continue; // skip broadcast for large head sizes
-                                    for (int nr2 : { 1, 4, 16 }) {
-                                        if (nr2 == 16 && hsk != 128) continue;
+                                    for (int nr2 : { 1, 4, 12, 20 }) {
+                                        if (nr2 == 12 && hsk != 128) continue;
+                                        if (nr2 == 20 && (nh != 1 || hsk != 576)) continue;
                                         //for (int kv : { 1, 17, 31, 33, 61, 113, 65, 127, 129, 130, 255, 260, 371, 380, 407, 512, 1024, }) {
                                         for (int kv : { 113, 512, 1024, }) {
                                             if (nr2 != 1 && kv != 512) continue;
@@ -8460,6 +8474,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     // Qwen3-VL-8B https://github.com/ggml-org/llama.cpp/issues/17012
     test_cases.emplace_back(new test_flash_attn_ext(72, 72, 16, {1, 1}, 5776, 5776, false, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16));
 
+    test_cases.emplace_back(new test_flash_attn_ext(64, 64, 8, {8, 1}, 7680, 1, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16));
+    test_cases.emplace_back(new test_flash_attn_ext(64, 64, 8, {8, 1}, 7680, 4, true, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16));
+
     for (int kv : { 4096, 8192, 16384, }) {
         for (int hs : { 64, 128, }) {
             for (int nr : { 1, 4, }) {
diff --git a/tests/test-chat-parser.cpp b/tests/test-chat-parser.cpp
index 4766518fe6..6f44a2b421 100644
--- a/tests/test-chat-parser.cpp
+++ b/tests/test-chat-parser.cpp
@@ -54,113 +54,109 @@ static void assert_throws(const std::function<void()> & fn, const std::string &
 static void test_reasoning() {
   //common_log_set_verbosity_thold(LOG_DEFAULT_DEBUG);
   {
-    common_chat_msg_parser builder("<tnk>Cogito</tnk>Ergo sum", /* is_partial= */ false, {
-        /* .format = */ COMMON_CHAT_FORMAT_CONTENT_ONLY,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_NONE,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ false,
-    });
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    params.reasoning_format = COMMON_REASONING_FORMAT_NONE;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = false;
+    common_chat_msg_parser builder("<tnk>Cogito</tnk>Ergo sum", /* is_partial= */ false, params);
     assert_equals(false, builder.try_parse_reasoning("<tnk>", "</tnk>"));
     assert_equals("<tnk>Cogito</tnk>Ergo sum", builder.consume_rest());
   }
   {
-    common_chat_msg_parser builder("<tnk>Cogito</tnk>Ergo sum", /* is_partial= */ false, {
-        /* .format = */ COMMON_CHAT_FORMAT_CONTENT_ONLY,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ false,
-    });
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = false;
+    common_chat_msg_parser builder("<tnk>Cogito</tnk>Ergo sum", /* is_partial= */ false, params);
     assert_equals(true, builder.try_parse_reasoning("<tnk>", "</tnk>"));
     assert_equals(std::string("Cogito"), builder.result().reasoning_content);
     assert_equals("Ergo sum", builder.consume_rest());
   }
   {
-    common_chat_msg_parser builder("Cogito</tnk>Ergo sum", /* is_partial= */ false, {
-        /* .format = */ COMMON_CHAT_FORMAT_CONTENT_ONLY,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_NONE,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ false,
-    });
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    params.reasoning_format = COMMON_REASONING_FORMAT_NONE;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = false;
+    common_chat_msg_parser builder("Cogito</tnk>Ergo sum", /* is_partial= */ false, params);
     assert_equals(false, builder.try_parse_reasoning("<tnk>", "</tnk>"));
     assert_equals("Cogito</tnk>Ergo sum", builder.consume_rest());
   }
   {
-    common_chat_msg_parser builder("Cogito</tnk>Ergo sum", /* is_partial= */ false, {
-        /* .format = */ COMMON_CHAT_FORMAT_CONTENT_ONLY,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ true,
-    });
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = true;
+    common_chat_msg_parser builder("Cogito</tnk>Ergo sum", /* is_partial= */ false, params);
     assert_equals(true, builder.try_parse_reasoning("<tnk>", "</tnk>"));
     assert_equals(std::string("Cogito"), builder.result().reasoning_content);
     assert_equals("Ergo sum", builder.consume_rest());
   }
   {
-    common_chat_msg_parser builder("Cogito</tnk>Ergo sum", /* is_partial= */ false, {
-        /* .format = */ COMMON_CHAT_FORMAT_CONTENT_ONLY,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-        /* .reasoning_in_content = */ true,
-        /* .thinking_forced_open = */ true,
-    });
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    params.reasoning_in_content = true;
+    params.thinking_forced_open = true;
+    common_chat_msg_parser builder("Cogito</tnk>Ergo sum", /* is_partial= */ false, params);
     assert_equals(true, builder.try_parse_reasoning("<tnk>", "</tnk>"));
     assert_equals("<think>Cogito</think>", builder.result().content);
     assert_equals("Ergo sum", builder.consume_rest());
   }
   {
     const std::string variant("content_only_inline_think");
-    common_chat_syntax syntax = {
-        /* .format = */ COMMON_CHAT_FORMAT_CONTENT_ONLY,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ false,
-        /* .parse_tool_calls = */ false,
-    };
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = false;
+    params.parse_tool_calls = false;
     const std::string input = "<think>Pense</think>Bonjour";
-    auto msg = common_chat_parse(input, false, syntax);
+    auto msg = common_chat_parse(input, false, params);
     assert_equals(variant, std::string("Pense"), msg.reasoning_content);
     assert_equals(variant, std::string("Bonjour"), msg.content);
   }
   {
     const std::string variant("llama_3_inline_think");
-    common_chat_syntax syntax = {
-        /* .format = */ COMMON_CHAT_FORMAT_LLAMA_3_X,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ false,
-        /* .parse_tool_calls = */ false,
-    };
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_LLAMA_3_X;
+    params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = false;
+    params.parse_tool_calls = false;
     const std::string input = "<think>Plan</think>Réponse";
-    auto msg = common_chat_parse(input, false, syntax);
+    auto msg = common_chat_parse(input, false, params);
     assert_equals(variant, std::string("Plan"), msg.reasoning_content);
     assert_equals(variant, std::string("Réponse"), msg.content);
   }
   // Test DeepSeek V3.1 parsing - reasoning content followed by "</think>" and then regular content
   {
-    common_chat_syntax syntax = {
-        /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ true,
-        /* .parse_tool_calls = */ true,
-    };
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+    params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = true;
+    params.parse_tool_calls = true;
     const std::string variant("deepseek_v3_1_reasoning_format_deepseek");
-    common_chat_msg_parser builder("REASONING</think>ok", /* is_partial= */ false, syntax);
+    common_chat_msg_parser builder("REASONING</think>ok", /* is_partial= */ false, params);
     assert_equals(variant, true, builder.try_parse_reasoning("<think>", "</think>"));
     assert_equals(variant, std::string("REASONING"), builder.result().reasoning_content);
     assert_equals(variant, std::string("ok"), builder.consume_rest());
   }
   // Test DeepSeek V3.1 parsing - reasoning_format none - reasoning content followed by "</think>" and then regular content
   {
-    common_chat_syntax syntax = {
-        /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_NONE,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ true,
-        /* .parse_tool_calls = */ true,
-    };
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+    params.reasoning_format = COMMON_REASONING_FORMAT_NONE;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = true;
+    params.parse_tool_calls = true;
     const std::string variant("deepseek_v3_1_reasoning_format_none");
     const std::string input = "REASONING</think>ok";
-    auto msg = common_chat_parse(input, false, syntax);
+    auto msg = common_chat_parse(input, false, params);
     assert_equals(variant, std::string("REASONING</think>ok"), msg.content);
     assert_equals(variant, std::string(""), msg.reasoning_content);
   }
@@ -256,15 +252,14 @@ static void test_deepseek_v3_1_tool_calls() {
     //common_log_set_verbosity_thold(LOG_DEFAULT_DEBUG);
     // variant: happy path for when it works as the model card says it should
     const std::string variant("simple");
-    common_chat_syntax syntax = {
-        /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-        /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-        /* .reasoning_in_content = */ false,
-        /* .thinking_forced_open = */ false,
-        /* .parse_tool_calls = */ true,
-    };
+    common_chat_parser_params params;
+    params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+    params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    params.reasoning_in_content = false;
+    params.thinking_forced_open = false;
+    params.parse_tool_calls = true;
     const std::string input = "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>";
-    auto msg = common_chat_parse(input, false, syntax);
+    auto msg = common_chat_parse(input, false, params);
     assert_equals<std::size_t>(variant, 1, msg.tool_calls.size());
     assert_equals(variant, std::string("get_time"), msg.tool_calls[0].name);
     // JSON arguments are dumped without spaces
@@ -274,16 +269,15 @@ static void test_deepseek_v3_1_tool_calls() {
 
     // variant: simple + thinking open
     {
-        common_chat_syntax syntax = {
-            /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-            /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-            /* .reasoning_in_content = */ false,
-            /* .thinking_forced_open = */ true,
-            /* .parse_tool_calls = */ true,
-        };
+        common_chat_parser_params params;
+        params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+        params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+        params.reasoning_in_content = false;
+        params.thinking_forced_open = true;
+        params.parse_tool_calls = true;
         const std::string variant("simple_thinking");
         const std::string in = "REASONING</think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>";
-        auto m = common_chat_parse(in, false, syntax);
+        auto m = common_chat_parse(in, false, params);
         assert_equals<std::size_t>(variant, 1, m.tool_calls.size());
         assert_equals(variant, std::string("get_time"), m.tool_calls[0].name);
         assert_equals(variant, std::string("{\"city\":\"Tokyo\"}"), m.tool_calls[0].arguments);
@@ -292,16 +286,15 @@ static void test_deepseek_v3_1_tool_calls() {
     }
     // variant: simple + multiple tool calls
     {
-        common_chat_syntax syntax = {
-            /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-            /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-            /* .reasoning_in_content = */ false,
-            /* .thinking_forced_open = */ false,
-            /* .parse_tool_calls = */ true,
-        };
+        common_chat_parser_params params;
+        params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+        params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+        params.reasoning_in_content = false;
+        params.thinking_forced_open = false;
+        params.parse_tool_calls = true;
         const std::string variant("simple_multiple_tool_calls");
         const std::string in = "CONTENT<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Paris\"}<｜tool▁call▁end｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{\"city\": \"Paris\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>";
-        auto m = common_chat_parse(in, false, syntax);
+        auto m = common_chat_parse(in, false, params);
         assert_equals<std::size_t>(variant, 2, m.tool_calls.size());
         assert_equals(variant, std::string("get_time"), m.tool_calls[0].name);
         assert_equals(variant, std::string("{\"city\":\"Paris\"}"), m.tool_calls[0].arguments);
@@ -314,16 +307,15 @@ static void test_deepseek_v3_1_tool_calls() {
 
     // variant: thinking forced open + tool call in reasoning content
     {
-        common_chat_syntax syntax = {
-            /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-            /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-            /* .reasoning_in_content = */ false,
-            /* .thinking_forced_open = */ true,
-            /* .parse_tool_calls = */ true,
-        };
+        common_chat_parser_params params;
+        params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+        params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+        params.reasoning_in_content = false;
+        params.thinking_forced_open = true;
+        params.parse_tool_calls = true;
         const std::string variant("thinking_forced_open_tool_call_in_reasoning");
         const std::string in = "REASONING<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time2<｜tool▁sep｜>{\"city\": \"Tokyo2\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>REASONING</think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>";
-        auto m = common_chat_parse(in, false, syntax);
+        auto m = common_chat_parse(in, false, params);
         assert_equals<std::size_t>(variant, 1, m.tool_calls.size());
         assert_equals(variant, std::string("get_time"), m.tool_calls[0].name);
         assert_equals(variant, std::string("{\"city\":\"Tokyo\"}"), m.tool_calls[0].arguments);
@@ -336,16 +328,15 @@ static void test_deepseek_v3_1_tool_calls() {
     //          to make tool calls in reasoning content according to the model card, but it does sometimes, so
     //          add the reasoning content as regular content and parse the tool calls.
     {
-        common_chat_syntax syntax = {
-            /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-            /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-            /* .reasoning_in_content = */ false,
-            /* .thinking_forced_open = */ true,
-            /* .parse_tool_calls = */ true,
-        };
+        common_chat_parser_params params;
+        params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+        params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+        params.reasoning_in_content = false;
+        params.thinking_forced_open = true;
+        params.parse_tool_calls = true;
         const std::string variant("thinking_forced_open_tool_call_in_reasoning_no_closing_think_not_partial");
         const std::string in = "REASONING<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>";
-        auto m = common_chat_parse(in, false, syntax);
+        auto m = common_chat_parse(in, false, params);
         assert_equals(variant, std::string("REASONING"), m.content);
         assert_equals(variant, std::string(""), m.reasoning_content);
         assert_equals<std::size_t>(variant, 1, m.tool_calls.size());
@@ -355,16 +346,15 @@ static void test_deepseek_v3_1_tool_calls() {
 
     // variant: thinking forced open + tool call in reasoning content + no closing think + partial
     {
-        common_chat_syntax syntax = {
-            /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-            /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-            /* .reasoning_in_content = */ false,
-            /* .thinking_forced_open = */ true,
-            /* .parse_tool_calls = */ true,
-        };
+        common_chat_parser_params params;
+        params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+        params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+        params.reasoning_in_content = false;
+        params.thinking_forced_open = true;
+        params.parse_tool_calls = true;
         const std::string variant("thinking_forced_open_tool_call_in_reasoning_no_closing_think_partial");
         const std::string in = "REASONING<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>";
-        auto m = common_chat_parse(in, /* is_partial= */ true, syntax);
+        auto m = common_chat_parse(in, /* is_partial= */ true, params);
         assert_equals(variant, std::string("REASONING<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>"), m.reasoning_content);
         assert_equals(variant, std::string(""), m.content);
         assert_equals<std::size_t>(variant, 0, m.tool_calls.size());
@@ -372,32 +362,30 @@ static void test_deepseek_v3_1_tool_calls() {
 
     // variant: thinking not forced open + reasoning + regular content + no tool calls
     {
-        common_chat_syntax syntax = {
-            /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-            /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-            /* .reasoning_in_content = */ false,
-            /* .thinking_forced_open = */ true,
-            /* .parse_tool_calls = */ true,
-        };
+        common_chat_parser_params params;
+        params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+        params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+        params.reasoning_in_content = false;
+        params.thinking_forced_open = true;
+        params.parse_tool_calls = true;
         const std::string variant("thinking_forced_open_reasoning_regular_content_no_tool_calls");
         const std::string in = "REASONING</think>CONTENT";
-        auto m = common_chat_parse(in, false, syntax);
+        auto m = common_chat_parse(in, false, params);
         assert_equals<std::size_t>(variant, 0, m.tool_calls.size());
         assert_equals(variant, std::string("CONTENT"), m.content);
         assert_equals(variant, std::string("REASONING"), m.reasoning_content);
     }
     // variant: thinking not forced open + missing reasoning + no tool calls
     {
-        common_chat_syntax syntax = {
-            /* .format = */ COMMON_CHAT_FORMAT_DEEPSEEK_V3_1,
-            /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
-            /* .reasoning_in_content = */ false,
-            /* .thinking_forced_open = */ false,
-            /* .parse_tool_calls = */ true,
-        };
+        common_chat_parser_params params;
+        params.format = COMMON_CHAT_FORMAT_DEEPSEEK_V3_1;
+        params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+        params.reasoning_in_content = false;
+        params.thinking_forced_open = false;
+        params.parse_tool_calls = true;
         const std::string variant("thinking_not_forced_open_missing_reasoning_no_tool_calls");
         const std::string in = "CONTENT";
-        auto m = common_chat_parse(in, false, syntax);
+        auto m = common_chat_parse(in, false, params);
         assert_equals<std::size_t>(variant, 0, m.tool_calls.size());
         assert_equals(variant, std::string("CONTENT"), m.content);
         assert_equals(variant, std::string(""), m.reasoning_content);
diff --git a/tests/test-chat-peg-parser.cpp b/tests/test-chat-peg-parser.cpp
index d3a4cfd226..f767c73c27 100644
--- a/tests/test-chat-peg-parser.cpp
+++ b/tests/test-chat-peg-parser.cpp
@@ -616,15 +616,15 @@ void test_command7_parser_compare(testing & t) {
 
     auto test_legacy = [&](const std::string & input, bool need_more_input, bool print_results) {
         // Original common_chat_combinator_parser taken from chat.cpp
+        common_chat_parser_params params;
+        params.format = COMMON_CHAT_FORMAT_GENERIC;
+        params.reasoning_format = COMMON_REASONING_FORMAT_AUTO;
+        params.reasoning_in_content = false;
+        params.thinking_forced_open = false;
         common_chat_msg_parser builder(
             input,
             /* .is_partial = */ need_more_input,
-            {
-                /* .format = */ COMMON_CHAT_FORMAT_GENERIC,
-                /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
-                /* .reasoning_in_content = */ false,
-                /* .thinking_forced_open = */ false,
-            }
+            params
         );
 
         builder.try_parse_reasoning("<|START_THINKING|>", "<|END_THINKING|>");
diff --git a/tests/test-chat-template.cpp b/tests/test-chat-template.cpp
index e142900723..27b537a036 100644
--- a/tests/test-chat-template.cpp
+++ b/tests/test-chat-template.cpp
@@ -54,7 +54,6 @@ std::string DEFAULT_JSON = R"({
     ],
     "bos_token": "<s>",
     "eos_token": "</s>",
-    "tools": [],
     "add_generation_prompt": true
 })";
 
diff --git a/tests/test-chat.cpp b/tests/test-chat.cpp
index e1264b8e8d..4378a8db71 100644
--- a/tests/test-chat.cpp
+++ b/tests/test-chat.cpp
@@ -341,10 +341,11 @@ static void test_templates(const struct common_chat_templates * tmpls, const std
         }
 
         if (expect_grammar_triggered) {
-            common_chat_syntax syntax;
-            syntax.format = data.params.format;
-            syntax.reasoning_format = reasoning_format;
-            const auto msg = common_chat_parse(data.delta, /* is_partial= */ false, syntax);
+            // TODO @ngxson : refactor common_chat_parse to avoid passing format/reasoning_format every time
+            common_chat_parser_params params;
+            params.format = data.params.format;
+            params.reasoning_format = reasoning_format;
+            const auto msg = common_chat_parse(data.delta, /* is_partial= */ false, params);
             assert_msg_equals(test_message, msg, ignore_whitespace_differences);
         }
 
@@ -461,9 +462,9 @@ static void test_parser_with_streaming(const common_chat_msg & expected, const s
     for (size_t i = 1; i <= raw_message.size(); ++i) {
         auto curr_msg = parse_msg(std::string(utf8_truncate_safe_view(std::string_view(raw_message).substr(0, i))));
         if (curr_msg == simple_assist_msg("")) continue;
-        LOG_INF("Streaming msg: %s\n", common_chat_msgs_to_json_oaicompat<json>({curr_msg}).dump().c_str());
+        LOG_INF("Streaming msg: %s\n", common_chat_msgs_to_json_oaicompat({curr_msg}).dump().c_str());
         for (auto diff: common_chat_msg_diff::compute_diffs(last_msg, curr_msg)) {
-            LOG_INF("Streaming diff: %s\n", common_chat_msg_diff_to_json_oaicompat<json>(diff).dump().c_str());
+            LOG_INF("Streaming diff: %s\n", common_chat_msg_diff_to_json_oaicompat(diff).dump().c_str());
             if (!diff.reasoning_content_delta.empty()) {
                 merged.reasoning_content += diff.reasoning_content_delta;
             }
@@ -479,7 +480,7 @@ static void test_parser_with_streaming(const common_chat_msg & expected, const s
                     merged.tool_calls.back().arguments += diff.tool_call_delta.arguments;
                 }
             }
-            LOG_INF("Streaming merged: %s\n", common_chat_msgs_to_json_oaicompat<json>({merged}).dump().c_str());
+            LOG_INF("Streaming merged: %s\n", common_chat_msgs_to_json_oaicompat({merged}).dump().c_str());
         }
         assert_msg_equals(curr_msg, merged, true);
         last_msg = curr_msg;
@@ -556,7 +557,9 @@ struct make_peg_parser {
     }
 
     common_chat_msg parse(const std::string & msg, bool is_partial) {
-        return common_chat_peg_parse(arena_, msg, is_partial, /* syntax = */ {params_.format});
+        common_chat_parser_params parser_params;
+        parser_params.format = params_.format;
+        return common_chat_peg_parse(arena_, msg, is_partial, parser_params);
     }
 };
 
@@ -589,7 +592,7 @@ static void test_peg_parser(common_chat_templates * tmpls, const std::function<v
             }
             if (diff.tool_call_index != std::string::npos) {
                 if (!diff.tool_call_delta.name.empty()) {
-                    msg_accum.tool_calls.push_back({diff.tool_call_delta.name, "", ""});
+                    msg_accum.tool_calls.push_back({diff.tool_call_delta.name, "", diff.tool_call_delta.id});
                 }
                 if (!diff.tool_call_delta.arguments.empty()) {
                     msg_accum.tool_calls.back().arguments += diff.tool_call_delta.arguments;
@@ -619,7 +622,7 @@ static void test_msgs_oaicompat_json_conversion() {
         message_assist_call_code_interpreter,
     };
     for (const auto & msg : msgs) {
-        auto oai_json = common_chat_msgs_to_json_oaicompat<json>({msg});
+        auto oai_json = common_chat_msgs_to_json_oaicompat({msg});
         auto msgs2 = common_chat_msgs_parse_oaicompat(oai_json);
         assert_equals((size_t) 1, msgs2.size());
         auto msg2 = msgs2[0];
@@ -643,7 +646,7 @@ static void test_msgs_oaicompat_json_conversion() {
             "  }\n"
             "]"
         ),
-        common_chat_msgs_to_json_oaicompat<json>({message_user_parts}).dump(2));
+        common_chat_msgs_to_json_oaicompat({message_user_parts}).dump(2));
 
     assert_equals(
         std::string(
@@ -663,7 +666,7 @@ static void test_msgs_oaicompat_json_conversion() {
             "  }\n"
             "]"
         ),
-        common_chat_msgs_to_json_oaicompat<json>({message_assist_call_python}).dump(2));
+        common_chat_msgs_to_json_oaicompat({message_assist_call_python}).dump(2));
 
     auto res = common_chat_msgs_parse_oaicompat(json::parse("[{\"role\": \"assistant\", \"tool_calls\": []}]"));
     assert_equals<size_t>(1, res.size());
@@ -690,7 +693,7 @@ static void test_tools_oaicompat_json_conversion() {
     };
 
     for (const auto & tool : tools) {
-        auto oai_json = common_chat_tools_to_json_oaicompat<json>({tool});
+        auto oai_json = common_chat_tools_to_json_oaicompat({tool});
         auto tools2 = common_chat_tools_parse_oaicompat(oai_json);
         assert_equals((size_t) 1, tools2.size());
         auto tool2 = tools2[0];
@@ -723,7 +726,7 @@ static void test_tools_oaicompat_json_conversion() {
             "  }\n"
             "]"
         ),
-        common_chat_tools_to_json_oaicompat<json>({special_function_tool}).dump(2));
+        common_chat_tools_to_json_oaicompat({special_function_tool}).dump(2));
 
     {
         auto tools_no_params = common_chat_tools_parse_oaicompat(json::parse(
@@ -750,6 +753,25 @@ static void test_tools_oaicompat_json_conversion() {
     }
 }
 
+// for compat; ref: https://github.com/ggml-org/llama.cpp/pull/18961
+struct test_parser_params {
+    common_chat_format       format                = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    common_reasoning_format  reasoning_format      = COMMON_REASONING_FORMAT_NONE;
+    bool                     reasoning_in_content  = false;
+    bool                     thinking_forced_open  = false;
+    bool                     parse_tool_calls      = true;
+};
+
+static common_chat_msg test_chat_parse(const std::string & input, bool is_partial, const test_parser_params & syntax) {
+    common_chat_parser_params params;
+    params.format               = syntax.format;
+    params.reasoning_format     = syntax.reasoning_format;
+    params.reasoning_in_content = syntax.reasoning_in_content;
+    params.thinking_forced_open = syntax.thinking_forced_open;
+    params.parse_tool_calls     = syntax.parse_tool_calls;
+    return common_chat_parse(input, is_partial, params);
+}
+
 static void test_template_output_parsers() {
     printf("[%s]\n", __func__);
 
@@ -781,17 +803,17 @@ static void test_template_output_parsers() {
         }
 
         assert_msg_equals(message_assist,
-            common_chat_parse(
+            test_chat_parse(
                 "Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_COMMAND_R7B}));
         assert_msg_equals(message_assist,
-            common_chat_parse(
+            test_chat_parse(
                 "<|START_RESPONSE|>Hello, world!\nWhat's up?<|END_RESPONSE|>",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_COMMAND_R7B}));
         assert_msg_equals(message_assist_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<|START_THINKING|>I'm\nthinking<|END_THINKING|>"
                 "<|START_RESPONSE|>Hello, world!\nWhat's up?<|END_RESPONSE|>",
                 /* is_partial= */ false,
@@ -800,7 +822,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
                 }));
         assert_msg_equals(message_assist_thoughts_unparsed_deepseek,
-            common_chat_parse(
+            test_chat_parse(
                 "<|START_THINKING|>I'm\nthinking<|END_THINKING|>"
                 "<|START_RESPONSE|>Hello, world!\nWhat's up?<|END_RESPONSE|>",
                 /* is_partial= */ false,
@@ -811,13 +833,13 @@ static void test_template_output_parsers() {
                     /* .thinking_forced_open = */ false,
                 }));
         assert_msg_equals(message_assist_thoughts_unparsed_r7b,
-            common_chat_parse(
+            test_chat_parse(
                 "<|START_THINKING|>I'm\nthinking<|END_THINKING|>"
                 "<|START_RESPONSE|>Hello, world!\nWhat's up?<|END_RESPONSE|>",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_COMMAND_R7B}));
         assert_msg_equals(message_assist_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<|START_THINKING|>I'm\nthinking<|END_THINKING|>"
                 "<|START_RESPONSE|>Hello, world!\nWhat's up?<|END_RESPONSE|>",
                 /* is_partial= */ false,
@@ -826,7 +848,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
                 }));
         assert_msg_equals(message_assist_thoughts_call_idx,
-            common_chat_parse(
+            test_chat_parse(
                 "<|START_THINKING|>I'm\nthinking<|END_THINKING|>"
                 "<|START_ACTION|>[\n"
                 "    {\"tool_call_id\": \"0\", \"tool_name\": \"special_function\", \"parameters\": {\"arg1\": 1}}\n"
@@ -837,7 +859,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
                 }));
         assert_msg_equals(message_assist_thoughts_no_content,
-            common_chat_parse(
+            test_chat_parse(
                 "<|START_THINKING|>I'm\nthinking<|END_THINKING|>"
                 "<|START_ACTION|>[\n"
                 "    {\"tool_call_id\": \"0\", \"tool_name\": \"special",
@@ -877,7 +899,7 @@ static void test_template_output_parsers() {
 
         assert_equals(
             simple_assist_msg("{ \"tool_call\" : { \"name\" : \"t"),
-            common_chat_parse(
+            test_chat_parse(
                 "{ \"tool_call\" : { \"name\" : \"t",
                 /* is_partial= */ true,
                 {
@@ -889,33 +911,33 @@ static void test_template_output_parsers() {
                 }));
         assert_equals(
             message_assist_empty,
-            common_chat_parse(
+            test_chat_parse(
                 "{ \"tool_call\" : { \"name\" : \"t",
                 /* is_partial= */ true,
                 {COMMON_CHAT_FORMAT_GENERIC}));
 
         assert_equals(
             simple_assist_msg("", "", "puppeteer_screenshot", "{\"name\":\"servethehome_homepage\","),
-            common_chat_parse(
+            test_chat_parse(
                 R"({"tool_call": {"name": "puppeteer_screenshot", "arguments": {"name": "servethehome_homepage",)",
                 /* is_partial= */ true,
                 {COMMON_CHAT_FORMAT_GENERIC}));
 
         assert_equals(
             message_assist_call_empty_args,
-            common_chat_parse(
+            test_chat_parse(
                 "{ \"tool_call\" : { \"name\" : \"special_function\"",
                 /* is_partial= */ true,
                 {COMMON_CHAT_FORMAT_GENERIC}));
         assert_equals(
             message_assist_call_cutoff_args,
-            common_chat_parse(
+            test_chat_parse(
                 "{ \"tool_call\" : { \"name\" : \"special_function\", \"arguments\" : { \"arg",
                 /* is_partial= */ true,
                 {COMMON_CHAT_FORMAT_GENERIC}));
 
         assert_msg_equals(message_assist,
-            common_chat_parse(
+            test_chat_parse(
                 "{\n"
                 "  \"response\": \"Hello, world!\\nWhat's up?\"\n"
                 "}",
@@ -951,7 +973,7 @@ static void test_template_output_parsers() {
     {
         assert_msg_equals(
             simple_assist_msg("Réponse", "raisonnement"),
-            common_chat_parse(
+            test_chat_parse(
                 message_assist_thoughts_unparsed_magistral.content,
                 /* is_partial= */ false,
                 {
@@ -988,14 +1010,14 @@ static void test_template_output_parsers() {
         // Test parsing
         assert_msg_equals(
             simple_assist_msg("", "", "python", ""),
-            common_chat_parse(
+            test_chat_parse(
                 "```json\n"
                 "<function_call> { \"name\" : \"python\"",
                 /* is_partial= */ true,
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(
             simple_assist_msg("Let's call something\n"),
-            common_chat_parse(
+            test_chat_parse(
                 "Let's call something\n"
                 "<tool_call>{\"name\"",
                 /* is_partial= */ true,
@@ -1005,7 +1027,7 @@ static void test_template_output_parsers() {
                 }));
         assert_msg_equals(
             simple_assist_msg("Let's call something\n"),
-            common_chat_parse(
+            test_chat_parse(
                 "Let's call something\n"
                 "<tool_call>{\"name",
                 /* is_partial= */ true,
@@ -1014,7 +1036,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
                 }));
         assert_msg_equals(message_assist_call_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 // QwQ-32B's template adds a trailing <think> if add_generation_prompt
                 "I'm\nthinking</think>\n"
                 "<tool_call>{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}</tool_call>",
@@ -1027,14 +1049,14 @@ static void test_template_output_parsers() {
                 }));
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<tool_call>\n"
                 "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
                 "</tool_call>",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(message_assist_call_content,
-            common_chat_parse(
+            test_chat_parse(
                 "Hello, world!\nWhat's up?<tool_call>\n"
                 "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
                 "</tool_call>",
@@ -1042,13 +1064,13 @@ static void test_template_output_parsers() {
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<function=special_function>{\"arg1\": 1}</function>",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<function name=\"special_function\">\n"
                 "{\"arg1\": 1}\n"
                 "</function>",
@@ -1056,7 +1078,7 @@ static void test_template_output_parsers() {
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<tool>\n"
                 "  {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
                 "</tool>",
@@ -1064,7 +1086,7 @@ static void test_template_output_parsers() {
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<tools>\n"
                 "  {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
                 "</tools>",
@@ -1072,7 +1094,7 @@ static void test_template_output_parsers() {
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<response>\n"
                 "  {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
                 "</response>",
@@ -1080,7 +1102,7 @@ static void test_template_output_parsers() {
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "```xml\n"
                 "<response>\n"
                 "    {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
@@ -1090,7 +1112,7 @@ static void test_template_output_parsers() {
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "```xml\n"
                 "  {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
                 "```",
@@ -1098,7 +1120,7 @@ static void test_template_output_parsers() {
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "```\n"
                 "  {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
                 "```",
@@ -1106,7 +1128,7 @@ static void test_template_output_parsers() {
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "```\n"
                 "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
                 "```",
@@ -1114,7 +1136,7 @@ static void test_template_output_parsers() {
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "```json\n"
                 "  {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
                 "```",
@@ -1122,7 +1144,7 @@ static void test_template_output_parsers() {
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "```json\n"
                 "\n"
                 "                    <function_call> {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}} \n"
@@ -1132,7 +1154,7 @@ static void test_template_output_parsers() {
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<json>\n"
                 "  {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
                 "</json>",
@@ -1140,7 +1162,7 @@ static void test_template_output_parsers() {
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<xml>\n"
                 "  {\n"
                 "    \"name\": \"special_function\", \"arguments\": {\"arg1\": 1}\n"
@@ -1150,7 +1172,7 @@ static void test_template_output_parsers() {
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<JSON>\n"
                 "  {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
                 "</JSON>",
@@ -1158,13 +1180,13 @@ static void test_template_output_parsers() {
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "{\n  \"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
@@ -1178,7 +1200,7 @@ static void test_template_output_parsers() {
 
         assert_msg_equals(
             message_assist_multiple_calls,
-            common_chat_parse(
+            test_chat_parse(
                 "<tool_call>\n"
                 "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
                 "</tool_call>\n"
@@ -1190,7 +1212,7 @@ static void test_template_output_parsers() {
 
         assert_msg_equals(
             message_assist_multiple_calls,
-            common_chat_parse(
+            test_chat_parse(
                 "<function=special_function>{\"arg1\": 1}</function>\n"
                 "<function=python>{\"code\":\"print('hello')\"}</function>",
                 /* is_partial= */ false,
@@ -1202,27 +1224,27 @@ static void test_template_output_parsers() {
                 "",
                 "special_function",
                 "{\"arg1\": 1}"),
-            common_chat_parse(
+            test_chat_parse(
                 "This is not a tool call:\n"
                 "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(message_assist,
-            common_chat_parse(
+            test_chat_parse(
                 "Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         assert_msg_equals(message_assist_thoughts_unparsed_deepseek,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_HERMES_2_PRO}));
         // assert_msg_equals(message_assist_thoughts_unparsed_deepseek,
-        //     common_chat_parse(
+        //     test_chat_parse(
         //         "I'm\nthinking</think>Hello, world!\nWhat's up?",
         //         COMMON_CHAT_FORMAT_HERMES_2_PRO));
         assert_msg_equals(message_assist_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -1230,7 +1252,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
                 }));
         assert_msg_equals(message_assist_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ true,
                 {
@@ -1238,7 +1260,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
                 }));
         assert_msg_equals(message_assist_thoughts_unparsed_md,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think>Hello, world!\nWhat's up?\n```json\n{}```",
                 /* is_partial= */ false,
                 {
@@ -1249,7 +1271,7 @@ static void test_template_output_parsers() {
                     /* .parse_tool_calls = */ false,
                 }));
         assert_msg_equals(message_assist_thoughts_unparsed_md_partial,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think>Hello, world!\nWhat's up?\n```json\n{}```",
                 /* is_partial= */ true,
                 {
@@ -1259,7 +1281,7 @@ static void test_template_output_parsers() {
                     /* .thinking_forced_open = */ false,
                 }));
         assert_msg_equals(message_assist_thoughts_unopened_unparsed,
-            common_chat_parse(
+            test_chat_parse(
                 "I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -1267,7 +1289,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
                 }));
         assert_msg_equals(message_assist_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -1304,7 +1326,7 @@ static void test_template_output_parsers() {
                       "</tool_call>");
         assert_msg_equals(
             simple_assist_msg("", /* reasoning_content= */ "<tool_call>nah uhg</tool_call>"),
-            common_chat_parse(
+            test_chat_parse(
                 "<think><tool_call>nah uhg</tool_call>",
                 /* is_partial= */ false,
                 {
@@ -1328,7 +1350,7 @@ static void test_template_output_parsers() {
 
         assert_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "{\"name\": \"special_function\", \"parameters\": {\"arg1\": 1}}",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_LLAMA_3_X}));
@@ -1366,7 +1388,7 @@ static void test_template_output_parsers() {
         for (auto is_partial : { false, true }) {
             assert_equals(
                 message_assist_call,
-                common_chat_parse(
+                test_chat_parse(
                     "<function=special_function>{\"arg1\": 1}</function>",
                     is_partial,
                     {COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1}));
@@ -1374,7 +1396,7 @@ static void test_template_output_parsers() {
 
         assert_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<function=special_function>{\"arg1\": 1}<",
                 /* is_partial= */ true,
                 {COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1}));
@@ -1396,7 +1418,7 @@ static void test_template_output_parsers() {
                 "",
                 "special_function",
                 "{\"arg1\": 1}"),
-            common_chat_parse(
+            test_chat_parse(
                 "all\n"
                 "Hello, world!\n"
                 "nono\n"
@@ -1405,27 +1427,27 @@ static void test_template_output_parsers() {
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2}));
         assert_msg_equals(message_assist_call_python_lines,
-            common_chat_parse(
+            test_chat_parse(
                 "python\n"
                 "# This is a program:\n"
                 "print('hey')",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2}));
         assert_msg_equals(message_assist_call_python_lines_unclosed,
-            common_chat_parse(
+            test_chat_parse(
                 "python\n"
                 "# This is a program:\n"
                 "print('hey')",
                 /* is_partial= */ true,
                 {COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2}));
         assert_msg_equals(message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "special_function\n"
                 "{\"arg1\": 1} \n                    ",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2}));
         assert_msg_equals(message_assist,
-            common_chat_parse(
+            test_chat_parse(
                 "all\n"
                 "Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
@@ -1466,7 +1488,7 @@ static void test_template_output_parsers() {
         test_templates(tmpls.get(), end_tokens, message_assist_thoughts, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
         assert_msg_equals(
             simple_assist_msg("Hello, world!\nWhat's up?", "<think>I'm\nthinking"),
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -1477,7 +1499,7 @@ static void test_template_output_parsers() {
                 }));
         assert_msg_equals(
             simple_assist_msg("", "I need to remember the correct syntax. It starts with <｜tool▁calls▁begin｜> and ends with"),
-            common_chat_parse(
+            test_chat_parse(
                 "I need to remember the correct syntax. It starts with <｜tool▁calls▁begin｜> and ends with",
                 /* is_partial= */ true,
                 {
@@ -1487,7 +1509,7 @@ static void test_template_output_parsers() {
                     /* .thinking_forced_open = */ true,
                 }));
         assert_msg_equals(message_assist_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -1495,7 +1517,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
                 }));
         assert_msg_equals(message_assist_thoughts_unopened_unparsed,
-            common_chat_parse(
+            test_chat_parse(
                 "I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -1503,7 +1525,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
                 }));
         assert_msg_equals(message_assist_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -1514,7 +1536,7 @@ static void test_template_output_parsers() {
                 }));
         assert_msg_equals(message_assist_thoughts,
             // Latest template update (ast of 20250209) adds a trailing <think>\n if add_generation_prompt is true.
-            common_chat_parse(
+            test_chat_parse(
                 "I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -1543,12 +1565,12 @@ static void test_template_output_parsers() {
         test_templates(tmpls.get(), end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
         test_templates(tmpls.get(), end_tokens, message_assist_thoughts, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
         assert_msg_equals(message_assist_thoughts_unparsed_deepseek,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_DEEPSEEK_R1}));
         assert_msg_equals(message_assist_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -1556,7 +1578,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
                 }));
         assert_msg_equals(message_assist_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -1567,7 +1589,7 @@ static void test_template_output_parsers() {
                 }));
 
         assert_msg_equals(message_assist_call_thoughts_unparsed,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think>\n\n"
                 "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>special_function\n"
                 "```json\n"
@@ -1576,7 +1598,7 @@ static void test_template_output_parsers() {
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_DEEPSEEK_R1}));
         assert_msg_equals(message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<｜tool▁calls｜>function<｜tool▁sep｜>special_function\n"
                 "```json\n"
                 "{\"arg1\": 1}\n"
@@ -1585,7 +1607,7 @@ static void test_template_output_parsers() {
                 {COMMON_CHAT_FORMAT_DEEPSEEK_R1}));
 
         assert_msg_equals(message_assist_call_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think>\n\n"
                 "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>special_function\n"
                 "```json\n"
@@ -1612,20 +1634,20 @@ static void test_template_output_parsers() {
 
         // Test parsing regular content
         assert_msg_equals(message_assist,
-            common_chat_parse(
+            test_chat_parse(
                 "Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_GRANITE}));
         assert_msg_equals(
             message_assist,
-            common_chat_parse(
+            test_chat_parse(
                 "Hello, world!\nWhat's up?",
                 /* is_partial= */ true,
                 {COMMON_CHAT_FORMAT_GRANITE}));
 
         // Test parsing content with thinking
         assert_msg_equals(message_assist_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -1633,12 +1655,12 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
                 }));
         assert_msg_equals(message_assist_thoughts_unparsed_deepseek,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_GRANITE}));
         assert_msg_equals(message_assist_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think><response>Hello, world!\nWhat's up?",
                 /* is_partial= */ true,
                 {
@@ -1646,7 +1668,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
                 }));
         assert_msg_equals(message_assist_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think><response>Hello, world!\nWhat's up?</response>",
                 /* is_partial= */ false,
                 {
@@ -1654,12 +1676,12 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
                 }));
         assert_msg_equals(simple_assist_msg("<think>I'm\nthinking</think><response>Hello, world!\nWhat's up?</response>"),
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think><response>Hello, world!\nWhat's up?</response>",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_GRANITE}));
         assert_msg_equals(message_assist_empty,
-            common_chat_parse(
+            test_chat_parse(
                 "<think",
                 /* is_partial= */ true,
                 {
@@ -1667,12 +1689,12 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
                 }));
         assert_msg_equals(message_assist_empty,
-            common_chat_parse(
+            test_chat_parse(
                 "<think",
                 /* is_partial= */ true,
                 {COMMON_CHAT_FORMAT_GRANITE}));
         assert_msg_equals(message_assist_thoughts_no_content,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking",
                 /* is_partial= */ true,
                 {
@@ -1681,32 +1703,32 @@ static void test_template_output_parsers() {
                 }));
         assert_msg_equals(
             message_assist_empty,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think><response",
                 /* is_partial= */ true,
                 {COMMON_CHAT_FORMAT_GRANITE}));
 
         // Test parsing tool calls
         assert_msg_equals(message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<|tool_call|>[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_GRANITE}));
         assert_msg_equals(
             message_assist_call_empty_args,
-            common_chat_parse(
+            test_chat_parse(
                 "<|tool_call|>[{\"name\": \"special_function\"",
                 /* is_partial= */ true,
                 {COMMON_CHAT_FORMAT_GRANITE}));
         assert_msg_equals(
             message_assist_call_cutoff_args,
-            common_chat_parse(
+            test_chat_parse(
                 "<|tool_call|>[{\"name\": \"special_function\", \"arguments\": {\"arg",
                 /* is_partial= */ true,
                 {COMMON_CHAT_FORMAT_GRANITE}));
         assert_msg_equals(
             message_assist_call_cutoff_args,
-            common_chat_parse(
+            test_chat_parse(
                 "<|tool_call|>[{\"name\": \"special_function\", \"arguments\": {\"arg",
                 /* is_partial= */ true,
                 {
@@ -1717,7 +1739,7 @@ static void test_template_output_parsers() {
         // Test parsing tool calls with thinking
         assert_msg_equals(
             message_assist_call_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think><|tool_call|>[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}, {",
                 /* is_partial= */ true,
                 {
@@ -1757,7 +1779,7 @@ static void test_template_output_parsers() {
         assert_equals(COMMON_CHAT_FORMAT_GPT_OSS, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
 
         assert_msg_equals(simple_assist_msg("", "I'm\nthink"),
-            common_chat_parse(
+            test_chat_parse(
                 "<|channel|>analysis<|message|>I'm\nthink",
                 /* is_partial= */ true,
                 {
@@ -1765,7 +1787,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
                 }));
         assert_msg_equals(simple_assist_msg("", "I'm\nthinking"),
-            common_chat_parse(
+            test_chat_parse(
                 "<|channel|>analysis<|message|>I'm\nthinking<|end|>",
                 /* is_partial= */ true,
                 {
@@ -1773,7 +1795,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
                 }));
         assert_msg_equals(simple_assist_msg("Hello, world!\nWhat's up?", "I'm\nthinking"),
-            common_chat_parse(
+            test_chat_parse(
                 "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
                 "<|start|>assistant<|channel|>final<|message|>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
@@ -1782,7 +1804,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
                 }));
         assert_msg_equals(simple_assist_msg("", "I'm\nthinking", "special_function", "{\"arg1"),
-            common_chat_parse(
+            test_chat_parse(
                 "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
                 "<|start|>assistant<|channel|>commentary to=functions.special_function <|constrain|>json<|message|>{\"arg1",
                 /* is_partial= */ true,
@@ -1791,7 +1813,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
                 }));
         assert_msg_equals(simple_assist_msg("", "I'm\nthinking", "special_function", "{\"arg1"),
-            common_chat_parse(
+            test_chat_parse(
                 "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
                 "<|start|>assistant<|channel|>commentary to=functions.special_function<|message|>{\"arg1",
                 /* is_partial= */ true,
@@ -1800,7 +1822,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
                 }));
         assert_msg_equals(simple_assist_msg("", "I'm\nthinking", "special_function", "{\"arg1\": 1}"),
-            common_chat_parse(
+            test_chat_parse(
                 "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
                 "<|start|>assistant<|channel|>commentary to=functions.special_function <|constrain|>json<|message|>{\"arg1\": 1}",
                 /* is_partial= */ false,
@@ -1809,7 +1831,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
                 }));
         assert_msg_equals(simple_assist_msg("", "I'm\nthinking", "special_function", "{\"arg1\": 1}"),
-            common_chat_parse(
+            test_chat_parse(
                 "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
                 "<|start|>assistant<|channel|>analysis to=functions.special_function <|constrain|>json<|message|>{\"arg1\": 1}",
                 /* is_partial= */ false,
@@ -1818,7 +1840,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
                 }));
         assert_msg_equals(simple_assist_msg("Hello, world!\nWhat's up?", "I'm\nthinking"),
-            common_chat_parse(
+            test_chat_parse(
                 "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
                 "<|start|>assistant<|channel|>commentary<|message|>Hello, world!\nWhat's up?",
                 /* is_partial= */ true,
@@ -1827,7 +1849,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
                 }));
         assert_msg_equals(simple_assist_msg("Hello, world!\nWhat's up?", "I'm\nthinking", "special_function", "{\"arg1\": 1}"),
-            common_chat_parse(
+            test_chat_parse(
                 "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
                 "<|start|>assistant<|channel|>commentary<|message|>Hello, world!\nWhat's up?<|end|>"
                 "<|start|>assistant<|channel|>commentary to=functions.special_function <|constrain|>json<|message|>{\"arg1\": 1}",
@@ -1840,7 +1862,7 @@ static void test_template_output_parsers() {
         // Test parse_tool_calls == false
         assert_msg_equals(
             simple_assist_msg("Hello, world!\nWhat's up?", "I'm\nthinking"),
-            common_chat_parse(
+            test_chat_parse(
                 "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
                 "<|start|>assistant<|channel|>final<|message|>Hello, world!\nWhat's up?",
                 /* is_partial= */ true,
@@ -1853,7 +1875,7 @@ static void test_template_output_parsers() {
                 }));
         assert_msg_equals(
             simple_assist_msg("", "I'm\nthinking"),
-            common_chat_parse(
+            test_chat_parse(
                 "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
                 "<|start|>assistant<|channel|>commentary to=functions.special_function<|message|>{\"arg1",
                 /* is_partial= */ true,
@@ -1866,7 +1888,7 @@ static void test_template_output_parsers() {
                 }));
         assert_msg_equals(
             simple_assist_msg("", "I'm\nthinking"),
-            common_chat_parse(
+            test_chat_parse(
                 "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
                 "<|start|>assistant<|channel|>commentary to=functions.special_function <|constrain|>json<|message|>{\"arg1\": 1}",
                 /* is_partial= */ false,
@@ -1882,7 +1904,7 @@ static void test_template_output_parsers() {
         assert_msg_equals(
             simple_assist_msg(
                 "<|channel|>analysis<|message|>I'm\nthinking<|end|>Hello, world!\nWhat's up?"),
-            common_chat_parse(
+            test_chat_parse(
                 "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
                 "<|start|>assistant<|channel|>final<|message|>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
@@ -1894,7 +1916,7 @@ static void test_template_output_parsers() {
         assert_msg_equals(
             simple_assist_msg(
                 "<|channel|>analysis<|message|>I'm\nthinking<|end|>Hello, world!\nWhat's up?"),
-            common_chat_parse(
+            test_chat_parse(
                 "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
                 "<|start|>assistant<|channel|>final<|message|>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
@@ -1906,7 +1928,7 @@ static void test_template_output_parsers() {
 
         // Test tool calling in role header
         assert_msg_equals(simple_assist_msg("", "", "special_function", "{\"arg1\": 1}"),
-            common_chat_parse(
+            test_chat_parse(
                 " to=functions.special_function<|channel|>commentary <|constrain|>json<|message|>{\"arg1\": 1}",
                 /* is_partial= */ false,
                 {
@@ -1914,7 +1936,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
                 }));
         assert_msg_equals(simple_assist_msg("", "", "special_function", "{\"arg1\": 1}"),
-            common_chat_parse(
+            test_chat_parse(
                 " to=functions.special_function<|channel|>analysis <|constrain|>json<|message|>{\"arg1\": 1}",
                 /* is_partial= */ false,
                 {
@@ -1922,7 +1944,7 @@ static void test_template_output_parsers() {
                     /* .reasoning_format = */ COMMON_REASONING_FORMAT_AUTO,
                 }));
         assert_msg_equals(simple_assist_msg("", "I'm\nthinking", "special_function", "{\"arg1\": 1}"),
-            common_chat_parse(
+            test_chat_parse(
                 "<|channel|>analysis<|message|>I'm\nthinking<|end|>"
                 "<|start|>assistant to=functions.special_function<|channel|>analysis <|constrain|>json<|message|>{\"arg1\": 1}",
                 /* is_partial= */ false,
@@ -1944,7 +1966,7 @@ static void test_template_output_parsers() {
         // Test simple reasoning content
         assert_msg_equals(
             simple_assist_msg("Hello, world!", "I'm thinking about the answer"),
-            common_chat_parse(
+            test_chat_parse(
                 "<seed:think>I'm thinking about the answer</seed:think>Hello, world!",
                 /* is_partial= */ false,
                 {
@@ -1959,7 +1981,7 @@ static void test_template_output_parsers() {
         msg_budget_reflect.reasoning_content = "Token usage: 45/1000\nI should continue thinking to find the best solution.";
         assert_msg_equals(
             msg_budget_reflect,
-            common_chat_parse(
+            test_chat_parse(
                 "<seed:think>Token usage: 45/1000\nI should continue thinking to find the best solution.</seed:think>"
                 "<seed:cot_budget_reflect>Token usage: 45/1000\nI should continue thinking to find the best solution.</seed:cot_budget_reflect>"
                 "I need to calculate this step by step.",
@@ -1975,7 +1997,7 @@ static void test_template_output_parsers() {
         msg_tool_call.tool_calls.push_back({"calculate_sum", "{\"numbers\": [1, 2, 3]}", ""});
         assert_msg_equals(
             msg_tool_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<seed:tool_call>\n"
                 "<function=calculate_sum>\n"
                 "<parameter=numbers>[1, 2, 3]</parameter>\n"
@@ -1992,7 +2014,7 @@ static void test_template_output_parsers() {
         msg_reasoning_tool.tool_calls.push_back({"calculate_sum", "{\"numbers\": [1, 2, 3]}", ""});
         assert_msg_equals(
             msg_reasoning_tool,
-            common_chat_parse(
+            test_chat_parse(
                 "<seed:think>I need to calculate the sum of these numbers</seed:think>"
                 "<seed:tool_call>\n"
                 "<function=calculate_sum>\n"
@@ -2013,7 +2035,7 @@ static void test_template_output_parsers() {
         std::size_t previousToolCalls = 0;
         for (std::size_t i = std::string("<seed:tool_call>").length(); i < tool_msg.length() - 1; i++) {
             auto partial = tool_msg.substr(0, i);
-            auto partial_res = common_chat_parse(partial, true, { COMMON_CHAT_FORMAT_SEED_OSS, COMMON_REASONING_FORMAT_DEEPSEEK });
+            auto partial_res = test_chat_parse(partial, true, { COMMON_CHAT_FORMAT_SEED_OSS, COMMON_REASONING_FORMAT_DEEPSEEK });
             if (partial_res.tool_calls.size() < previousToolCalls) {
                 throw std::runtime_error("Tool call size decreased on partial: " + partial + " from " + std::to_string(previousToolCalls) + " to " + std::to_string(partial_res.tool_calls.size()));
             }
@@ -2026,7 +2048,7 @@ static void test_template_output_parsers() {
         msg_multi_param.tool_calls.push_back({"process_data", "{\"input\": \"test\", \"format\": \"json\"}", ""});
         assert_msg_equals(
             msg_multi_param,
-            common_chat_parse(
+            test_chat_parse(
                 "<seed:tool_call>\n"
                 "<function=process_data>\n"
                 "<parameter=input>test</parameter>\n"
@@ -2039,7 +2061,7 @@ static void test_template_output_parsers() {
         // Test partial parsing for incomplete tool call - don't actually add the call until parsing parameters is done
         assert_msg_equals(
             simple_assist_msg("", "", "calculate_sum", "{\"numbers\":"),
-            common_chat_parse(
+            test_chat_parse(
                 "<seed:tool_call>\n"
                 "<function=calculate_sum>\n"
                 "<parameter=numbers>[1,\n",
@@ -2049,7 +2071,7 @@ static void test_template_output_parsers() {
         // Test incomplete reasoning tag
         assert_msg_equals(
             simple_assist_msg("", "I was thinking"),
-            common_chat_parse(
+            test_chat_parse(
                 "<seed:think>I was thinking",
                 /* is_partial= */ true,
                 {
@@ -2060,7 +2082,7 @@ static void test_template_output_parsers() {
         // Test content without reasoning
         assert_msg_equals(
             simple_assist_msg("This is a simple response without reasoning."),
-            common_chat_parse(
+            test_chat_parse(
                 "This is a simple response without reasoning.",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_SEED_OSS}));
@@ -2074,14 +2096,14 @@ static void test_template_output_parsers() {
 
         // Test parsing regular content
         assert_msg_equals(message_assist,
-            common_chat_parse(
+            test_chat_parse(
                 "Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_NEMOTRON_V2}));
 
         // Test parsing content with thinking
         assert_msg_equals(message_assist_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -2091,14 +2113,14 @@ static void test_template_output_parsers() {
 
         // Test parsing tool calls
         assert_msg_equals(message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<TOOLCALL>[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]</TOOLCALL>",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_NEMOTRON_V2}));
 
         // Test parsing tool calls with thinking
         assert_msg_equals(message_assist_call_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think><TOOLCALL>[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]</TOOLCALL>",
                 /* is_partial= */ false,
                 {
@@ -2108,7 +2130,7 @@ static void test_template_output_parsers() {
 
         // Test tool calls with extra content
         assert_msg_equals(message_assist_call_content,
-            common_chat_parse(
+            test_chat_parse(
                 "<TOOLCALL>[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]</TOOLCALL>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_NEMOTRON_V2}
@@ -2116,7 +2138,7 @@ static void test_template_output_parsers() {
 
         // Test tool calls with extra content AND thinking
         assert_msg_equals(message_assist_call_thoughts_content,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think><TOOLCALL>[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]</TOOLCALL>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -2149,7 +2171,7 @@ static void test_template_output_parsers() {
         test_templates(tmpls.get(), end_tokens, message_assist_thoughts, tools, "</think>Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
         assert_msg_equals(
             simple_assist_msg("Hello, world!\nWhat's up?", "I'm\nthinking"),
-            common_chat_parse(
+            test_chat_parse(
                 "I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -2161,7 +2183,7 @@ static void test_template_output_parsers() {
         // variant: thinking forced open, reasoning_format none
         assert_msg_equals(
             simple_assist_msg("REASONING</think>ok", ""),
-            common_chat_parse(
+            test_chat_parse(
                 "REASONING</think>ok",
                 /* is_partial= */ false,
                 {
@@ -2174,7 +2196,7 @@ static void test_template_output_parsers() {
         // variant: happy path for when it works as the model card says it should
         assert_msg_equals(
             simple_assist_msg("", "", "get_time", "{\"city\":\"Tokyo\"}"),
-            common_chat_parse(
+            test_chat_parse(
                 "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
                 /* is_partial= */ false,
                 {
@@ -2187,7 +2209,7 @@ static void test_template_output_parsers() {
         // variant: simple + thinking open
         assert_msg_equals(
             simple_assist_msg("", "REASONING", "get_time", "{\"city\":\"Tokyo\"}"),
-            common_chat_parse(
+            test_chat_parse(
                 "REASONING</think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
                 /* is_partial= */ false,
                 {
@@ -2205,7 +2227,7 @@ static void test_template_output_parsers() {
         message_assist_multiple_calls.tool_calls.push_back({"get_weather", "{\"city\":\"Paris\"}", ""});
         assert_msg_equals(
             message_assist_multiple_calls,
-            common_chat_parse(
+            test_chat_parse(
                 "CONTENT<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Paris\"}<｜tool▁call▁end｜><｜tool▁call▁begin｜>get_weather<｜tool▁sep｜>{\"city\": \"Paris\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
                 /* is_partial= */ false,
                 {
@@ -2218,7 +2240,7 @@ static void test_template_output_parsers() {
         // variant: thinking forced open + tool call in reasoning content
         assert_msg_equals(
             simple_assist_msg("", "REASONING<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time2<｜tool▁sep｜>{\"city\": \"Tokyo2\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>REASONING", "get_time", "{\"city\":\"Tokyo\"}"),
-            common_chat_parse(
+            test_chat_parse(
                 "REASONING<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time2<｜tool▁sep｜>{\"city\": \"Tokyo2\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>REASONING</think><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
                 /* is_partial= */ false,
                 {
@@ -2234,7 +2256,7 @@ static void test_template_output_parsers() {
         //          add the reasoning content as regular content and parse the tool calls.
         assert_msg_equals(
             simple_assist_msg("REASONING", "", "get_time", "{\"city\":\"Tokyo\"}"),
-            common_chat_parse(
+            test_chat_parse(
                 "REASONING<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
                 /* is_partial= */ false,
                 {
@@ -2247,7 +2269,7 @@ static void test_template_output_parsers() {
         // variant: thinking forced open + tool call in reasoning content + no closing think + partial
         assert_msg_equals(
             simple_assist_msg("", "REASONING<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>", "", ""),
-            common_chat_parse(
+            test_chat_parse(
                 "REASONING<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>get_time<｜tool▁sep｜>{\"city\": \"Tokyo\"}<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
                 /* is_partial= */ true,
                 {
@@ -2260,7 +2282,7 @@ static void test_template_output_parsers() {
         // variant: thinking not forced open + missing reasoning + no tool calls
         assert_msg_equals(
             simple_assist_msg("CONTENT", ""),
-            common_chat_parse(
+            test_chat_parse(
                 "CONTENT",
                 /* is_partial= */ false,
                 {
@@ -2280,14 +2302,14 @@ static void test_template_output_parsers() {
 
         // Test parsing regular content
         assert_msg_equals(message_assist,
-            common_chat_parse(
+            test_chat_parse(
                 "Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_APERTUS}));
 
         // Test parsing content with thinking
         assert_msg_equals(message_assist_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<|inner_prefix|>I'm\nthinking<|inner_suffix|>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -2297,14 +2319,14 @@ static void test_template_output_parsers() {
 
         // Test parsing tool calls
         assert_msg_equals(message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<|tools_prefix|>[{\"special_function\": {\"arg1\": 1}}]<|tools_suffix|>",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_APERTUS}));
 
         // Test parsing tool calls with thinking
         assert_msg_equals(message_assist_call_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<|inner_prefix|>I'm\nthinking<|inner_suffix|><|tools_prefix|>[{\"special_function\": {\"arg1\": 1}}]<|tools_suffix|>",
                 /* is_partial= */ false,
                 {
@@ -2314,7 +2336,7 @@ static void test_template_output_parsers() {
 
         // Test tool calls with extra content
         assert_msg_equals(message_assist_call_content,
-            common_chat_parse(
+            test_chat_parse(
                 "<|tools_prefix|>[{\"special_function\": {\"arg1\": 1}}]<|tools_suffix|>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_APERTUS}
@@ -2322,7 +2344,7 @@ static void test_template_output_parsers() {
 
         // Test tool calls with extra content AND thinking
         assert_msg_equals(message_assist_call_thoughts_content,
-            common_chat_parse(
+            test_chat_parse(
                 "<|inner_prefix|>I'm\nthinking<|inner_suffix|><|tools_prefix|>[{\"special_function\": {\"arg1\": 1}}]<|tools_suffix|>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -2402,7 +2424,7 @@ Hey there!<|im_end|>
 
         // Test parsing regular content
         assert_msg_equals(message_assist,
-            common_chat_parse(
+            test_chat_parse(
                 "Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS}));
@@ -2413,7 +2435,7 @@ Hey there!<|im_end|>
         msg_single_tool_call.tool_calls.push_back({"special_function", "{\"arg1\":1}", ""});
         assert_msg_equals(
             msg_single_tool_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<|tool_call_start|>[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]<|tool_call_end|>",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS}));
@@ -2424,7 +2446,7 @@ Hey there!<|im_end|>
         msg_tool_call_string.tool_calls.push_back({"get_weather", "{\"location\":\"Paris\"}", ""});
         assert_msg_equals(
             msg_tool_call_string,
-            common_chat_parse(
+            test_chat_parse(
                 "<|tool_call_start|>[{\"name\": \"get_weather\", \"arguments\": {\"location\": \"Paris\"}}]<|tool_call_end|>",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS}));
@@ -2435,7 +2457,7 @@ Hey there!<|im_end|>
         msg_multi_args.tool_calls.push_back({"calculate", "{\"x\":10,\"y\":20,\"operation\":\"add\"}", ""});
         assert_msg_equals(
             msg_multi_args,
-            common_chat_parse(
+            test_chat_parse(
                 "<|tool_call_start|>[{\"name\": \"calculate\", \"arguments\": {\"x\": 10, \"y\": 20, \"operation\": \"add\"}}]<|tool_call_end|>",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS}));
@@ -2447,7 +2469,7 @@ Hey there!<|im_end|>
         msg_multiple_tools.tool_calls.push_back({"get_time", "{\"timezone\":\"UTC\"}", ""});
         assert_msg_equals(
             msg_multiple_tools,
-            common_chat_parse(
+            test_chat_parse(
                 "<|tool_call_start|>[{\"name\": \"get_weather\", \"arguments\": {\"location\": \"Paris\"}}, {\"name\": \"get_time\", \"arguments\": {\"timezone\": \"UTC\"}}]<|tool_call_end|>",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS}));
@@ -2459,7 +2481,7 @@ Hey there!<|im_end|>
         msg_content_before_tool.tool_calls.push_back({"get_weather", "{\"location\":\"Paris\"}", ""});
         assert_msg_equals(
             msg_content_before_tool,
-            common_chat_parse(
+            test_chat_parse(
                 "Let me check the weather for you.<|tool_call_start|>[{\"name\": \"get_weather\", \"arguments\": {\"location\": \"Paris\"}}]<|tool_call_end|>",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS}));
@@ -2471,7 +2493,7 @@ Hey there!<|im_end|>
         msg_content_after_tool.tool_calls.push_back({"get_weather", "{\"location\":\"Paris\"}", ""});
         assert_msg_equals(
             msg_content_after_tool,
-            common_chat_parse(
+            test_chat_parse(
                 "<|tool_call_start|>[{\"name\": \"get_weather\", \"arguments\": {\"location\": \"Paris\"}}]<|tool_call_end|>Here's the result.",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS}));
@@ -2482,7 +2504,7 @@ Hey there!<|im_end|>
         msg_tool_call_newlines.tool_calls.push_back({"get_current_time", "{\"location\":\"Paris\"}", ""});
         assert_msg_equals(
             msg_tool_call_newlines,
-            common_chat_parse(
+            test_chat_parse(
                 "<|tool_call_start|>[{\n    \"name\": \"get_current_time\",\n    \"arguments\": {\n        \"location\": \"Paris\"\n    }\n}]<|tool_call_end|>",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS}));
@@ -2502,14 +2524,14 @@ Hey there!<|im_end|>
 
         // Test parsing regular content
         assert_msg_equals(message_assist,
-            common_chat_parse(
+            test_chat_parse(
                 "Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_MINIMAX_M2}));
 
         // Test parsing content with thinking
         assert_msg_equals(message_assist_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -2519,14 +2541,14 @@ Hey there!<|im_end|>
 
         // Test parsing tool calls
         assert_msg_equals(message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<minimax:tool_call><invoke name=\"special_function\"><parameter name=\"arg1\">1</parameter></invoke></minimax:tool_call>",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_MINIMAX_M2}));
 
         // Test parsing tool calls with thinking
         assert_msg_equals(message_assist_call_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think><minimax:tool_call><invoke name=\"special_function\"><parameter name=\"arg1\">1</parameter></invoke></minimax:tool_call>",
                 /* is_partial= */ false,
                 {
@@ -2536,7 +2558,7 @@ Hey there!<|im_end|>
 
         // Test tool calls with extra content
         assert_msg_equals(message_assist_call_content,
-            common_chat_parse(
+            test_chat_parse(
                 "<minimax:tool_call><invoke name=\"special_function\"><parameter name=\"arg1\">1</parameter></invoke></minimax:tool_call>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_MINIMAX_M2}
@@ -2544,7 +2566,7 @@ Hey there!<|im_end|>
 
         // Test tool calls with extra content AND thinking
         assert_msg_equals(message_assist_call_thoughts_content,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think><minimax:tool_call><invoke name=\"special_function\"><parameter name=\"arg1\">1</parameter></invoke></minimax:tool_call>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -2555,25 +2577,25 @@ Hey there!<|im_end|>
         // Test streaming
         test_parser_with_streaming(message_assist_call_thoughts_content,
             "<think>I'm\nthinking\n</think>Hello, world!\nWhat's up?\n<minimax:tool_call><invoke name=\"special_function\"><parameter name=\"arg1\">1</parameter></invoke></minimax:tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     /*  .format = */ COMMON_CHAT_FORMAT_MINIMAX_M2,
                     /*  .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK
             }); });
         test_parser_with_streaming(message_assist_call_thoughts_unparsed,
             "<think>I'm\nthinking</think>\n\n<minimax:tool_call><invoke name=\"special_function\"><parameter name=\"arg1\">1</parameter></invoke></minimax:tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     /*  .format = */ COMMON_CHAT_FORMAT_MINIMAX_M2,
                     /*  .reasoning_format = */ COMMON_REASONING_FORMAT_NONE
             }); });
         test_parser_with_streaming(message_assist_call_thoughts_content,
             "<think>I'm\nthinking\n</think>\n\nHello, world!\nWhat's up?\n\n<minimax:tool_call>\n<invoke name=\"special_function\">\n<parameter name=\"arg1\">1</parameter>\n</invoke>\n</minimax:tool_call>\n",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     /*  .format = */ COMMON_CHAT_FORMAT_MINIMAX_M2,
                     /*  .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK
             }); });
         test_parser_with_streaming(message_assist_call_withopt,
             "<minimax:tool_call>\n<invoke name=\"special_function_with_opt\">\n<parameter name=\"arg1\">1</parameter>\n<parameter name=\"arg2\">2</parameter>\n</invoke>\n</minimax:tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     /*  .format = */ COMMON_CHAT_FORMAT_MINIMAX_M2,
                     /*  .reasoning_format = */ COMMON_REASONING_FORMAT_NONE
             }); });
@@ -2618,14 +2640,14 @@ Hey there!<|im_end|>
 
         // Test parsing regular content
         assert_msg_equals(message_assist,
-            common_chat_parse(
+            test_chat_parse(
                 "Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_GLM_4_5}));
 
         // Test parsing content with thinking
         assert_msg_equals(message_assist_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "\n<think>I'm\nthinking</think>\nHello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -2635,14 +2657,14 @@ Hey there!<|im_end|>
 
         // Test parsing tool calls
         assert_msg_equals(message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "\n<tool_call>special_function\n<arg_key>arg1</arg_key>\n<arg_value>1</arg_value>\n</tool_call>",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_GLM_4_5}), true);
 
         // Test parsing tool calls with thinking
         assert_msg_equals(message_assist_call_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "\n<think>I'm\nthinking</think>\n<tool_call>special_function\n<arg_key>arg1</arg_key>\n<arg_value>1</arg_value>\n</tool_call>",
                 /* is_partial= */ false,
                 {
@@ -2652,7 +2674,7 @@ Hey there!<|im_end|>
 
         // Test tool calls with extra content
         assert_msg_equals(message_assist_call_content,
-            common_chat_parse(
+            test_chat_parse(
                 "\n<tool_call>special_function\n<arg_key>arg1</arg_key>\n<arg_value>1</arg_value>\n</tool_call>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_GLM_4_5}
@@ -2660,7 +2682,7 @@ Hey there!<|im_end|>
 
         // Test tool calls with extra content AND thinking
         assert_msg_equals(message_assist_call_thoughts_content,
-            common_chat_parse(
+            test_chat_parse(
                 "\n<think>I'm\nthinking</think>Hello, world!\nWhat's up?\n<tool_call>special_function\n<arg_key>arg1</arg_key>\n<arg_value>1</arg_value>\n</tool_call>",
                 /* is_partial= */ false,
                 {
@@ -2671,19 +2693,19 @@ Hey there!<|im_end|>
         // Test streaming
         test_parser_with_streaming(message_assist_call_thoughts_content,
             "\n<think>I'm\nthinking</think>Hello, world!\nWhat's up?\n<tool_call>special_function\n<arg_key>arg1</arg_key>\n<arg_value>1</arg_value>\n</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     /*  .format = */ COMMON_CHAT_FORMAT_GLM_4_5,
                     /*  .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK
             }); });
         test_parser_with_streaming(message_assist_call_thoughts_unparsed,
             "\n<think>I'm\nthinking</think>\n\n<tool_call>special_function\n<arg_key>arg1</arg_key>\n<arg_value>1</arg_value>\n</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     /*  .format = */ COMMON_CHAT_FORMAT_GLM_4_5,
                     /*  .reasoning_format = */ COMMON_REASONING_FORMAT_NONE
             }); });
         test_parser_with_streaming(message_assist_call_withopt,
             "\n<think></think>\n<tool_call>special_function_with_opt\n<arg_key>arg1</arg_key>\n<arg_value>1</arg_value>\n<arg_key>arg2</arg_key>\n<arg_value>2</arg_value>\n</tool_call>\n",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     /*  .format = */ COMMON_CHAT_FORMAT_GLM_4_5,
                     /*  .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK
             }); });
@@ -2699,7 +2721,7 @@ Hey there!<|im_end|>
                 "<arg_key>score</arg_key>\n"
                 "<arg_value>95.5</arg_value>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_GLM_4_5}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_GLM_4_5}); });
         test_parser_with_streaming(
                 simple_assist_msg("", "", "web_search", "{\"query\":\"\\\"From Zero\\\" Linkin Park album tracklist complete songs\",\"limit\":3,\"type\":\"text\"}"),
                 "<tool_call>web_search\n"
@@ -2710,18 +2732,18 @@ Hey there!<|im_end|>
                 "<arg_key>type</arg_key>\n"
                 "<arg_value>text</arg_value>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_GLM_4_5}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_GLM_4_5}); });
 
         // Test interleaved thinking
         test_parser_with_streaming(simple_assist_msg("Hello, world!\n\nWhat's up?", "I'm\nthinkingThinking2", "special_function", "{\"arg1\": 1}"),
             "\n<think>I'm\nthinking</think>Hello, world!\n<think>Thinking2</think>What's up?\n<tool_call>special_function\n<arg_key>arg1</arg_key>\n<arg_value>1</arg_value>\n</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     /*  .format = */ COMMON_CHAT_FORMAT_GLM_4_5,
                     /*  .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK
             }); });
         test_parser_with_streaming(simple_assist_msg("\n<think>I'm\nthinking</think>Hello, world!\n<think>Thinking2</think>What's up?", "", "special_function", "{\"arg1\": 1}"),
             "\n<think>I'm\nthinking</think>Hello, world!\n<think>Thinking2</think>What's up?\n<tool_call>special_function\n<arg_key>arg1</arg_key>\n<arg_value>1</arg_value>\n</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     /*  .format = */ COMMON_CHAT_FORMAT_GLM_4_5,
                     /*  .reasoning_format = */ COMMON_REASONING_FORMAT_NONE
             }); });
@@ -2766,14 +2788,14 @@ Hey there!<|im_end|>
 
         // Test parsing regular content
         assert_msg_equals(message_assist,
-            common_chat_parse(
+            test_chat_parse(
                 "Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_KIMI_K2}));
 
         // Test parsing content with thinking
         assert_msg_equals(message_assist_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -2783,14 +2805,14 @@ Hey there!<|im_end|>
 
         // Test parsing tool calls
         assert_msg_equals(message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<|tool_calls_section_begin|><|tool_call_begin|>functions.special_function:0<|tool_call_argument_begin|>{\"arg1\": 1}<|tool_call_end|><|tool_calls_section_end|>",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_KIMI_K2}));
 
         // Test parsing tool calls with thinking
         assert_msg_equals(message_assist_call_thoughts,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think><|tool_calls_section_begin|><|tool_call_begin|>functions.special_function:0<|tool_call_argument_begin|>{\"arg1\": 1}<|tool_call_end|><|tool_calls_section_end|>",
                 /* is_partial= */ false,
                 {
@@ -2800,7 +2822,7 @@ Hey there!<|im_end|>
 
         // Test tool calls with extra content
         assert_msg_equals(message_assist_call_content,
-            common_chat_parse(
+            test_chat_parse(
                 "<|tool_calls_section_begin|><|tool_call_begin|>functions.special_function:0<|tool_call_argument_begin|>{\"arg1\": 1}<|tool_call_end|><|tool_calls_section_end|>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {COMMON_CHAT_FORMAT_KIMI_K2}
@@ -2808,7 +2830,7 @@ Hey there!<|im_end|>
 
         // Test tool calls with extra content AND thinking
         assert_msg_equals(message_assist_call_thoughts_content,
-            common_chat_parse(
+            test_chat_parse(
                 "<think>I'm\nthinking</think><|tool_calls_section_begin|><|tool_call_begin|>functions.special_function:0<|tool_call_argument_begin|>{\"arg1\": 1}<|tool_call_end|><|tool_calls_section_end|>Hello, world!\nWhat's up?",
                 /* is_partial= */ false,
                 {
@@ -2819,43 +2841,43 @@ Hey there!<|im_end|>
         // Test streaming
         test_parser_with_streaming(message_assist_call_thoughts_content,
             "<think>I'm\nthinking\n</think>Hello, world!\nWhat's up?\n<|tool_calls_section_begin|><|tool_call_begin|>functions.special_function:0<|tool_call_argument_begin|>{\"arg1\": 1}<|tool_call_end|><|tool_calls_section_end|>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     /*  .format = */ COMMON_CHAT_FORMAT_KIMI_K2,
                     /*  .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK
             }); });
         test_parser_with_streaming(message_assist_call_thoughts_unparsed,
             "<think>I'm\nthinking</think>\n\n<|tool_calls_section_begin|><|tool_call_begin|>functions.special_function:0<|tool_call_argument_begin|>{\"arg1\": 1}<|tool_call_end|><|tool_calls_section_end|>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     /*  .format = */ COMMON_CHAT_FORMAT_KIMI_K2,
                     /*  .reasoning_format = */ COMMON_REASONING_FORMAT_NONE
             }); });
         test_parser_with_streaming(message_assist_call_thoughts_content,
             "<think>I'm\nthinking\n</think>\n\nHello, world!\nWhat's up?\n\n<|tool_calls_section_begin|><|tool_call_begin|>functions.special_function:0<|tool_call_argument_begin|>{\"arg1\": 1}<|tool_call_end|><|tool_calls_section_end|>\n",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     /*  .format = */ COMMON_CHAT_FORMAT_KIMI_K2,
                     /*  .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK
             }); });
         test_parser_with_streaming(message_assist_call_withopt,
             "<|tool_calls_section_begin|><|tool_call_begin|>functions.special_function_with_opt:0<|tool_call_argument_begin|>{\"arg1\": 1, \"arg2\": 2}<|tool_call_end|><|tool_calls_section_end|>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     /*  .format = */ COMMON_CHAT_FORMAT_KIMI_K2,
                     /*  .reasoning_format = */ COMMON_REASONING_FORMAT_NONE
             }); });
         test_parser_with_streaming(simple_assist_msg("Hello, world!\nWhat's up?", "I'm\nthinking", "special_function", "{\"arg1\": \"123456\"}"),
             "<think>I'm\nthinking</think>Hello, world!\nWhat's up?\n<|tool_calls_section_begin|><|tool_call_begin|>functions.special_function:0<|tool_call_argument_begin|>{\"arg1\": \"123456\"}<|tool_call_end|><|tool_calls_section_end|>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     /*  .format = */ COMMON_CHAT_FORMAT_KIMI_K2,
                     /*  .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK
             }); });
         test_parser_with_streaming(simple_assist_msg("Hello, world!\nWhat's up?", "I'm\nthinking", "special_function", "{\"arg1\": [1, 2, \"345\", 6]}"),
             "<think>I'm\nthinking</think>Hello, world!\nWhat's up?\n<|tool_calls_section_begin|><|tool_call_begin|>functions.special_function:0<|tool_call_argument_begin|>{\"arg1\": [1, 2, \"345\", 6]}<|tool_call_end|><|tool_calls_section_end|>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     /*  .format = */ COMMON_CHAT_FORMAT_KIMI_K2,
                     /*  .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK
             }); });
         test_parser_with_streaming(simple_assist_msg("Hello, world!\nWhat's up?", "I'm\nthinking", "special_function", "{\"arg1\": {\"12\": 34, \"5\": [67, 8], \"9\": \"10\"}}"),
             "<think>I'm\nthinking</think>Hello, world!\nWhat's up?\n<|tool_calls_section_begin|><|tool_call_begin|>functions.special_function:0<|tool_call_argument_begin|>{\"arg1\": {\"12\": 34, \"5\": [67, 8], \"9\": \"10\"}}<|tool_call_end|><|tool_calls_section_end|>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     /*  .format = */ COMMON_CHAT_FORMAT_KIMI_K2,
                     /*  .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK
             }); });
@@ -2864,19 +2886,19 @@ Hey there!<|im_end|>
                 "<|tool_calls_section_begin|><|tool_call_begin|>functions.complex_function:0<|tool_call_argument_begin|>"
                 "{\"name\": \"John Doe\", \"age\": 30, \"active\": true, \"score\": 95.5}"
                 "<|tool_call_end|><|tool_calls_section_end|>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_KIMI_K2}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_KIMI_K2}); });
         test_parser_with_streaming(
                 simple_assist_msg("", "", "web_search", "{\"query\":\"\\\"From Zero\\\" Linkin Park album tracklist complete songs\",\"limit\":3,\"type\":\"text\"}"),
                 "<|tool_calls_section_begin|><|tool_call_begin|>functions.web_search:0<|tool_call_argument_begin|>"
                 "{\"query\":\"\\\"From Zero\\\" Linkin Park album tracklist complete songs\",\"limit\":3,\"type\":\"text\"}"
                 "<|tool_call_end|><|tool_calls_section_end|>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_KIMI_K2}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_KIMI_K2}); });
         test_parser_with_streaming(
                 simple_assist_msg("", "", "read_file", "{\"args\": [{\"path\": \"src/providers/ThemeProvider.tsx\"}, {\"path\": \"src/components/Header.tsx\"}, {\"path\": \"src/components/ThemeToggle.tsx\"}, {\"path\": \"src/app/globals.css\"}, {\"path\": \"src/app/layout.tsx\"}]}"),
                 "<|tool_calls_section_begin|><|tool_call_begin|>functions.read_file:0<|tool_call_argument_begin|>"
                 "{\"args\": [{\"path\": \"src/providers/ThemeProvider.tsx\"}, {\"path\": \"src/components/Header.tsx\"}, {\"path\": \"src/components/ThemeToggle.tsx\"}, {\"path\": \"src/app/globals.css\"}, {\"path\": \"src/app/layout.tsx\"}]}"
                 "<|tool_call_end|><|tool_calls_section_end|>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_KIMI_K2}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_KIMI_K2}); });
         test_parser_with_streaming(
                 simple_assist_msg(
                         "Let me start by examining the relevant files to understand the current implementation.", "",
@@ -2886,7 +2908,7 @@ Hey there!<|im_end|>
                 "<|tool_calls_section_begin|><|tool_call_begin|>functions.read_file:0<|tool_call_argument_begin|>"
                 "{\"files\":[{\"path\":\"src/app/Partners.tsx\",\"line_ranges\":[\"1-100\"]}]}"
                 "<|tool_call_end|><|tool_calls_section_end|>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_KIMI_K2}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_KIMI_K2}); });
         auto multi_tool_msg = simple_assist_msg("Let me call multiple tools.", "I'm thinking.");
         multi_tool_msg.tool_calls.push_back({ "read_file", "{\"files\": [{\"path\": \"src/app/Partners.tsx\", \"line_ranges\": [\"1-100\"]}]}", "" });
         multi_tool_msg.tool_calls.push_back({ "web_search", "{\"query\":\"\\\"From Zero\\\" Linkin Park album tracklist complete songs\",\"limit\":3,\"type\":\"text\"}", "" });
@@ -2908,7 +2930,7 @@ Hey there!<|im_end|>
                 "{\"message\":\"Hello! 👋 🌟 🚀 Testing emojis: 😀😃😄😁 and symbols: ∑∏∆∇\"}"
                 "<|tool_call_end|>"
                 "<|tool_calls_section_end|>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     COMMON_CHAT_FORMAT_KIMI_K2,
                     COMMON_REASONING_FORMAT_DEEPSEEK
             }); });
@@ -2917,7 +2939,7 @@ Hey there!<|im_end|>
                 "<think>I'm thinking<|tool_calls_section_begin|><|tool_call_begin|>functions.complex_function_in_think:0<|tool_call_argument_begin|>"
                 "{\"name\": \"John Doe\", \"age\": 30, \"active\": true, \"score\": 95.5}"
                 "<|tool_call_end|><|tool_calls_section_end|>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     COMMON_CHAT_FORMAT_KIMI_K2,
                     COMMON_REASONING_FORMAT_DEEPSEEK
             }); });
@@ -2926,7 +2948,7 @@ Hey there!<|im_end|>
                 "<think>I'm thinking<|tool_calls_section_begin|><|tool_call_begin|>functions.complex_function_in_think:0<|tool_call_argument_begin|>"
                 "{\"name\": \"John Doe\", \"age\": 30, \"active\": true, \"score\": 95.5}"
                 "<|tool_call_end|><|tool_calls_section_end|>I'm still thinking</think>Hello",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {
                     COMMON_CHAT_FORMAT_KIMI_K2,
                     COMMON_REASONING_FORMAT_DEEPSEEK
             }); });
@@ -3001,7 +3023,7 @@ Hey there!<|im_end|>
         // Basic XML tool call parsing
         assert_msg_equals(
             message_assist_call,
-            common_chat_parse(
+            test_chat_parse(
                 "<tool_call>\n"
                 "  <function=special_function>\n"
                 "    <parameter=arg1>\n"
@@ -3036,7 +3058,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Special characters and Unicode
         common_chat_msg expected_special_chars;
@@ -3053,7 +3075,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Multiline content with newlines and indentation
         common_chat_msg expected_multiline;
@@ -3072,7 +3094,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // JSON object as parameter value
         common_chat_msg expected_json_param;
@@ -3090,7 +3112,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Array as parameter value
         common_chat_msg expected_array_param;
@@ -3108,7 +3130,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Empty parameter
         common_chat_msg expected_empty_param;
@@ -3125,7 +3147,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Boolean values (true/false)
         common_chat_msg expected_boolean;
@@ -3146,7 +3168,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Null value
         common_chat_msg expected_null;
@@ -3164,7 +3186,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Negative numbers and scientific notation
         common_chat_msg expected_numbers;
@@ -3188,7 +3210,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // XML-like content in parameters (should be escaped)
         common_chat_msg expected_xml_content;
@@ -3206,7 +3228,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Quotes and escape characters
         common_chat_msg expected_quotes;
@@ -3224,7 +3246,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Long parameter value (simplified)
         std::string long_text = "This is a long text parameter that should test the parser's ability to handle larger amounts of text data.";
@@ -3244,7 +3266,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Mixed content with text before and after tool call
         common_chat_msg expected_mixed_content;
@@ -3263,7 +3285,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Compact format (no extra whitespace)
         common_chat_msg expected_compact;
@@ -3275,7 +3297,7 @@ Hey there!<|im_end|>
         test_parser_with_streaming(
             expected_compact,
                 "<tool_call><function=compact_function><parameter=param>value</parameter></function></tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Function name with underscores and numbers
         common_chat_msg expected_complex_name;
@@ -3293,7 +3315,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Parameter names with underscores and numbers
         common_chat_msg expected_complex_params;
@@ -3317,7 +3339,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Very deeply nested XML content in parameter
         common_chat_msg expected_deep_xml;
@@ -3335,7 +3357,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Parameter with only whitespace
         common_chat_msg expected_whitespace_param;
@@ -3353,7 +3375,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Parameter with tabs and mixed whitespace
         common_chat_msg expected_mixed_whitespace;
@@ -3373,7 +3395,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Control characters and special Unicode
         common_chat_msg expected_control_chars;
@@ -3391,7 +3413,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Emoji and extended Unicode characters
         common_chat_msg expected_emoji;
@@ -3409,7 +3431,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Mathematical expressions and formulas
         common_chat_msg expected_math;
@@ -3427,7 +3449,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // SQL injection-like content (should be safely escaped)
         common_chat_msg expected_sql;
@@ -3445,7 +3467,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // HTML/XML injection content
         common_chat_msg expected_html;
@@ -3463,7 +3485,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Binary-like content (base64)
         common_chat_msg expected_binary;
@@ -3481,7 +3503,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
 
         // Very large numbers (should be parsed as scientific notation)
         common_chat_msg expected_large_numbers;
@@ -3499,7 +3521,7 @@ Hey there!<|im_end|>
                 "    </parameter>\n"
                 "  </function>\n"
                 "</tool_call>",
-            [&](const std::string &msg) { return common_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
+            [&](const std::string &msg) { return test_chat_parse(msg, /* is_partial= */ true, {COMMON_CHAT_FORMAT_QWEN3_CODER_XML}); });
     }
 
     {
@@ -3777,6 +3799,134 @@ static void test_template_output_peg_parsers() {
         });
     }
 
+    {
+        // Solar-Open-100B
+        auto tmpls = read_templates("models/templates/upstage-Solar-Open-100B.jinja");
+
+        // Test basic message
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input = "<|content|>Hello, world!\nWhat's up?";
+            t.expect = message_assist;
+        });
+
+        // Test basic message and reasoning
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input = "<|think|>I'm\nthinking<|end|><|begin|>assistant<|content|>Hello, world!\nWhat's up?";
+            t.expect = message_assist_thoughts;
+        });
+
+        // Test basic message and reasoning_effort = low
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input = "<|content|>Hello, world!\nWhat's up?";
+            t.params.chat_template_kwargs["reasoning_effort"] = "\"low\"";
+            t.expect = message_assist;
+        });
+
+        // Test tool call
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input = "<|tool_calls|>"
+                      "<|tool_call:begin|>123456789"
+                      "<|tool_call:name|>special_function"
+                      "<|tool_call:args|>{\"arg1\":1}"
+                      "<|tool_call:end|>";
+
+            t.params.chat_template_kwargs["reasoning_effort"] = "\"low\"";
+            t.params.tools = {special_function_tool};
+            t.expect = message_assist_call_id;
+        });
+
+        // Test tool call with reasoning
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input = "<|think|>I'm\nthinking<|end|>"
+                      "<|begin|>assistant<|tool_calls|>"
+                      "<|tool_call:begin|>0"
+                      "<|tool_call:name|>special_function"
+                      "<|tool_call:args|>{\"arg1\":1}"
+                      "<|tool_call:end|>";
+
+            t.params.tools = {special_function_tool};
+            t.expect = message_assist_thoughts_call_idx;
+        });
+
+        // Test tool call with reasoning and tool_choice = required
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input = "<|think|>I'm\nthinking<|end|>"
+                      "<|begin|>assistant<|tool_calls|>"
+                      "<|tool_call:begin|>0"
+                      "<|tool_call:name|>special_function"
+                      "<|tool_call:args|>{\"arg1\":1}"
+                      "<|tool_call:end|>";
+
+            t.params.tools = {special_function_tool};
+            t.params.tool_choice = COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+            t.expect = message_assist_thoughts_call_idx;
+        });
+
+        // Test tool call without reasoning and tool_choice = required
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input = "<|tool_calls|>"
+                      "<|tool_call:begin|>0"
+                      "<|tool_call:name|>special_function"
+                      "<|tool_call:args|>{\"arg1\":1}"
+                      "<|tool_call:end|>";
+
+            t.params.tools = {special_function_tool};
+            t.params.tool_choice = COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+            t.params.chat_template_kwargs["reasoning_effort"] = "\"low\"";
+            t.expect = message_assist_call_idx;
+        });
+
+        // Test parallel tool calls
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input = "<|think|>I'm\nthinking<|end|>"
+                      "<|begin|>assistant<|tool_calls|>"
+                      "<|tool_call:begin|>0"
+                      "<|tool_call:name|>special_function"
+                      "<|tool_call:args|>{\"arg1\":1}"
+                      "<|tool_call:end|>"
+                      "<|tool_call:begin|>1"
+                      "<|tool_call:name|>special_function_with_opt"
+                      "<|tool_call:args|>{\"arg1\": 1, \"arg2\": 2}"
+                      "<|tool_call:end|>";
+
+            t.params.parallel_tool_calls = true;
+            t.params.tools = {special_function_tool, special_function_tool_with_optional_param};
+
+            t.expect.reasoning_content = "I'm\nthinking";
+            t.expect.tool_calls = {{
+                /* .name = */      "special_function",
+                /* .arguments = */ R"({"arg1": 1})",
+                /* .id = */        "0",
+            }, {
+                /* .name = */      "special_function_with_opt",
+                /* .arguments = */ R"({"arg1": 1, "arg2": 2})",
+                /* .id = */        "1",
+            }};
+        });
+
+        // Test response format
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input = "<|think|>I need to output the invoice details in JSON<|end|>"
+                      "<|begin|>assistant<|content|>"
+                      R"({"amount": 123.45, "date": "2025-12-03"})";
+
+            t.params.json_schema = invoice_schema;
+
+            t.expect.reasoning_content = "I need to output the invoice details in JSON";
+            t.expect.content =R"({"amount": 123.45, "date": "2025-12-03"})";
+        });
+
+        // Test response format no reasoning
+        test_peg_parser(tmpls.get(), [&](auto & t) {
+            t.input = "<|content|>"
+                      R"({"amount": 123.45, "date": "2025-12-03"})";
+
+            t.params.chat_template_kwargs["reasoning_effort"] = "\"low\"";
+            t.params.json_schema = invoice_schema;
+
+            t.expect.content =R"({"amount": 123.45, "date": "2025-12-03"})";
+        });
+    }
 }
 
 static void test_msg_diffs_compute() {
diff --git a/tests/test-gguf.cpp b/tests/test-gguf.cpp
index 3f0c312e2f..84b7f3bc49 100644
--- a/tests/test-gguf.cpp
+++ b/tests/test-gguf.cpp
@@ -1,9 +1,11 @@
 #include "ggml.h"
 #include "ggml-backend.h"
 #include "../ggml/src/ggml-impl.h"
+#include "gguf.h"
 
 #include <algorithm>
 #include <array>
+#include <cmath>
 #include <cstdint>
 #include <cstdio>
 #include <random>
@@ -34,6 +36,7 @@ enum handcrafted_file_type {
     HANDCRAFTED_TENSORS_BAD_N_DIMS         =  20 + offset_has_tensors,
     HANDCRAFTED_TENSORS_BAD_SHAPE          =  30 + offset_has_tensors,
     HANDCRAFTED_TENSORS_NE_TOO_BIG         =  40 + offset_has_tensors,
+    HANDCRAFTED_TENSORS_NBYTES_TOO_BIG     =  45 + offset_has_tensors,
     HANDCRAFTED_TENSORS_BAD_TYPE           =  50 + offset_has_tensors,
     HANDCRAFTED_TENSORS_BAD_OFFSET         =  60 + offset_has_tensors,
     HANDCRAFTED_TENSORS_DUPLICATE_NAME     =  70 + offset_has_tensors,
@@ -69,6 +72,7 @@ static std::string handcrafted_file_type_name(const enum handcrafted_file_type h
         case HANDCRAFTED_TENSORS_BAD_N_DIMS:         return "TENSORS_BAD_N_DIMS";
         case HANDCRAFTED_TENSORS_BAD_SHAPE:          return "TENSORS_BAD_SHAPE";
         case HANDCRAFTED_TENSORS_NE_TOO_BIG:         return "TENSORS_NE_TOO_BIG";
+        case HANDCRAFTED_TENSORS_NBYTES_TOO_BIG:     return "TENSORS_NBYTES_TOO_BIG";
         case HANDCRAFTED_TENSORS_BAD_TYPE:           return "TENSORS_BAD_TYPE";
         case HANDCRAFTED_TENSORS_BAD_OFFSET:         return "TENSORS_BAD_OFFSET";
         case HANDCRAFTED_TENSORS_DUPLICATE_NAME:     return "TENSORS_DUPLICATE_NAME";
@@ -326,7 +330,7 @@ static FILE * get_handcrafted_file(const unsigned int seed, const enum handcraft
 
     uint64_t offset = 0;
     for (int i = 0; i < int(tensor_configs.size()); ++i) {
-        const ggml_type                          type  = tensor_configs[i].first;
+        const ggml_type                          type  = hft == HANDCRAFTED_TENSORS_NBYTES_TOO_BIG ? GGML_TYPE_I64 : tensor_configs[i].first;
         const std::array<int64_t, GGML_MAX_DIMS> shape = tensor_configs[i].second;
 
         std::string name = "my_tensor";
@@ -343,7 +347,7 @@ static FILE * get_handcrafted_file(const unsigned int seed, const enum handcraft
         }
         helper_write(file, name.data(), name.length());
 
-        uint32_t n_dims = hft == HANDCRAFTED_TENSORS_NE_TOO_BIG ? 2 : 1;
+        uint32_t n_dims = (hft == HANDCRAFTED_TENSORS_NE_TOO_BIG || hft == HANDCRAFTED_TENSORS_NBYTES_TOO_BIG) ? 2 : 1;
         for (int i = GGML_MAX_DIMS-1; i >= 1; --i) {
             if (shape[i] != 1) {
                 n_dims = i + 1;
@@ -358,13 +362,19 @@ static FILE * get_handcrafted_file(const unsigned int seed, const enum handcraft
         }
 
         if (hft == HANDCRAFTED_TENSORS_BAD_SHAPE) {
+            const int64_t bad_dim = -1;
             for (uint32_t j = 0; j < n_dims; ++j) {
-                const int64_t bad_dim = -1;
                 helper_write(file, bad_dim);
             }
         } else if (hft == HANDCRAFTED_TENSORS_NE_TOO_BIG){
+            const int64_t big_dim = 4*int64_t(INT32_MAX);
+            for (uint32_t j = 0; j < n_dims; ++j) {
+                helper_write(file, big_dim);
+            }
+        } else if (hft == HANDCRAFTED_TENSORS_NBYTES_TOO_BIG){
+            const size_t  big_ne  = SIZE_MAX/ggml_type_size(type);
+            const int64_t big_dim = GGML_PAD(int64_t(1.01f*std::pow(big_ne, 1.0f/n_dims)) + 1, ggml_blck_size(type));
             for (uint32_t j = 0; j < n_dims; ++j) {
-                const int64_t big_dim = 4*int64_t(INT32_MAX);
                 helper_write(file, big_dim);
             }
         } else {
@@ -682,6 +692,7 @@ static std::pair<int, int> test_handcrafted_file(const unsigned int seed) {
         HANDCRAFTED_TENSORS_BAD_N_DIMS,
         HANDCRAFTED_TENSORS_BAD_SHAPE,
         HANDCRAFTED_TENSORS_NE_TOO_BIG,
+        HANDCRAFTED_TENSORS_NBYTES_TOO_BIG,
         HANDCRAFTED_TENSORS_BAD_TYPE,
         HANDCRAFTED_TENSORS_BAD_OFFSET,
         HANDCRAFTED_TENSORS_DUPLICATE_NAME,
diff --git a/tests/test-jinja.cpp b/tests/test-jinja.cpp
index 7adb302ffb..f6114f1e2f 100644
--- a/tests/test-jinja.cpp
+++ b/tests/test-jinja.cpp
@@ -4,10 +4,12 @@
 #include <cstdlib>
 
 #include <nlohmann/json.hpp>
+#include <sheredom/subprocess.h>
 
 #include "jinja/runtime.h"
 #include "jinja/parser.h"
 #include "jinja/lexer.h"
+#include "jinja/utils.h"
 
 #include "testing.h"
 
@@ -29,14 +31,27 @@ static void test_tests(testing & t);
 static void test_string_methods(testing & t);
 static void test_array_methods(testing & t);
 static void test_object_methods(testing & t);
+static void test_hasher(testing & t);
 static void test_fuzzing(testing & t);
 
+static bool g_python_mode = false;
+
 int main(int argc, char *argv[]) {
     testing t(std::cout);
     t.verbose = true;
 
-    if (argc >= 2) {
-        t.set_filter(argv[1]);
+    // usage: test-jinja [-py] [filter_regex]
+    //  -py : enable python mode (use python jinja2 for rendering expected output)
+    //        only use this for cross-checking, not for correctness
+    //        note: the implementation of this flag is basic, only intented to be used by maintainers
+
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+        if (arg == "-py") {
+            g_python_mode = true;
+        } else {
+            t.set_filter(arg);
+        }
     }
 
     t.test("whitespace control", test_whitespace_control);
@@ -53,7 +68,10 @@ int main(int argc, char *argv[]) {
     t.test("string methods", test_string_methods);
     t.test("array methods", test_array_methods);
     t.test("object methods", test_object_methods);
-    t.test("fuzzing", test_fuzzing);
+    if (!g_python_mode) {
+        t.test("hasher", test_hasher);
+        t.test("fuzzing", test_fuzzing);
+    }
 
     return t.summary();
 }
@@ -141,6 +159,18 @@ static void test_conditionals(testing & t) {
         "big"
     );
 
+    test_template(t, "object comparison",
+        "{% if {0: 1, none: 2, 1.0: 3, '0': 4, true: 5} == {false: 1, none: 2, 1: 5, '0': 4} %}equal{% endif %}",
+        json::object(),
+        "equal"
+    );
+
+    test_template(t, "array comparison",
+        "{% if [0, 1.0, false] == [false, 1, 0.0] %}equal{% endif %}",
+        json::object(),
+        "equal"
+    );
+
     test_template(t, "logical and",
         "{% if a and b %}both{% endif %}",
         {{"a", true}, {"b", true}},
@@ -176,6 +206,84 @@ static void test_conditionals(testing & t) {
         json::object(),
         "yes"
     );
+
+    test_template(t, "is undefined falsy",
+        "{{ 'yes' if not y else 'no' }}",
+        json::object(),
+        "yes"
+    );
+
+    test_template(t, "is undefined attribute falsy",
+        "{{ 'yes' if not y.x else 'no' }}",
+        {{"y", true}},
+        "yes"
+    );
+
+    test_template(t, "is undefined key falsy",
+        "{{ 'yes' if not y['x'] else 'no' }}",
+        {{"y", {{}}}},
+        "yes"
+    );
+
+    test_template(t, "is empty array falsy",
+        "{{ 'yes' if not y else 'no' }}",
+        {{"y", json::array()}},
+        "yes"
+    );
+
+    test_template(t, "is empty object falsy",
+        "{{ 'yes' if not y else 'no' }}",
+        {{"y", json::object()}},
+        "yes"
+    );
+
+    test_template(t, "is empty string falsy",
+        "{{ 'yes' if not y else 'no' }}",
+        {{"y", ""}},
+        "yes"
+    );
+
+    test_template(t, "is 0 falsy",
+        "{{ 'yes' if not y else 'no' }}",
+        {{"y", 0}},
+        "yes"
+    );
+
+    test_template(t, "is 0.0 falsy",
+        "{{ 'yes' if not y else 'no' }}",
+        {{"y", 0.0}},
+        "yes"
+    );
+
+    test_template(t, "is non-empty array truthy",
+        "{{ 'yes' if y else 'no' }}",
+        {{"y", json::array({""})}},
+        "yes"
+    );
+
+    test_template(t, "is non-empty object truthy",
+        "{{ 'yes' if y else 'no' }}",
+        {{"y", {"x", false}}},
+        "yes"
+    );
+
+    test_template(t, "is non-empty string truthy",
+        "{{ 'yes' if y else 'no' }}",
+        {{"y", "0"}},
+        "yes"
+    );
+
+    test_template(t, "is 1 truthy",
+        "{{ 'yes' if y else 'no' }}",
+        {{"y", 1}},
+        "yes"
+    );
+
+    test_template(t, "is 1.0 truthy",
+        "{{ 'yes' if y else 'no' }}",
+        {{"y", 1.0}},
+        "yes"
+    );
 }
 
 static void test_loops(testing & t) {
@@ -221,6 +329,12 @@ static void test_loops(testing & t) {
         "empty"
     );
 
+    test_template(t, "for undefined empty",
+        "{% for i in items %}{{ i }}{% else %}empty{% endfor %}",
+        json::object(),
+        "empty"
+    );
+
     test_template(t, "nested for",
         "{% for i in a %}{% for j in b %}{{ i }}{{ j }}{% endfor %}{% endfor %}",
         {{"a", json::array({1, 2})}, {"b", json::array({"x", "y"})}},
@@ -247,6 +361,12 @@ static void test_expressions(testing & t) {
         "Bob"
     );
 
+    test_template(t, "negative float (not dot notation)",
+        "{{ -1.0 }}",
+        json::object(),
+        "-1.0"
+    );
+
     test_template(t, "bracket notation",
         "{{ user['name'] }}",
         {{"user", {{"name", "Bob"}}}},
@@ -259,6 +379,30 @@ static void test_expressions(testing & t) {
         "b"
     );
 
+    test_template(t, "array negative access",
+        "{{ items[-1] }}",
+        {{"items", json::array({"a", "b", "c"})}},
+        "c"
+    );
+
+    test_template(t, "array slice",
+        "{{ items[1:-1]|string }}",
+        {{"items", json::array({"a", "b", "c"})}},
+        "['b']"
+    );
+
+    test_template(t, "array slice step",
+        "{{ items[::2]|string }}",
+        {{"items", json::array({"a", "b", "c"})}},
+        "['a', 'c']"
+    );
+
+    test_template(t, "tuple slice",
+        "{{ ('a', 'b', 'c')[::-1]|string }}",
+        json::object(),
+        "('c', 'b', 'a')"
+    );
+
     test_template(t, "arithmetic",
         "{{ (a + b) * c }}",
         {{"a", 2}, {"b", 3}, {"c", 4}},
@@ -302,6 +446,36 @@ static void test_set_statement(testing & t) {
         json::object(),
         "1"
     );
+
+    test_template(t, "set dict with mixed type keys",
+        "{% set d = {0: 1, none: 2, 1.0: 3, '0': 4, (0, 0): 5, false: 6, 1: 7} %}{{ d[(0, 0)] + d[0] + d[none] + d['0'] + d[false] + d[1.0] + d[1] }}",
+        json::object(),
+        "37"
+    );
+
+    test_template(t, "print dict with mixed type keys",
+        "{% set d = {0: 1, none: 2, 1.0: 3, '0': 4, (0, 0): 5, true: 6} %}{{ d|string }}",
+        json::object(),
+        "{0: 1, None: 2, 1.0: 6, '0': 4, (0, 0): 5}"
+    );
+
+    test_template(t, "print array with mixed types",
+        "{% set d = [0, none, 1.0, '0', true, (0, 0)] %}{{ d|string }}",
+        json::object(),
+        "[0, None, 1.0, '0', True, (0, 0)]"
+    );
+
+    test_template(t, "object member assignment with mixed key types",
+        "{% set d = namespace() %}{% set d.a = 123 %}{{ d['a'] == 123 }}",
+        json::object(),
+        "True"
+    );
+
+    test_template(t, "tuple unpacking",
+        "{% set t = (1, 2, 3) %}{% set a, b, c = t %}{{ a + b + c }}",
+        json::object(),
+        "6"
+    );
 }
 
 static void test_filters(testing & t) {
@@ -383,6 +557,32 @@ static void test_filters(testing & t) {
         "123"
     );
 
+    test_template(t, "sort reverse",
+        "{% for i in items|sort(true) %}{{ i }}{% endfor %}",
+        {{"items", json::array({3, 1, 2})}},
+        "321"
+    );
+
+    test_template(t, "sort with attribute",
+        "{{ items|sort(attribute='name')|join(attribute='age') }}",
+        {{"items", json::array({
+            json({{"name", "c"}, {"age", 3}}),
+            json({{"name", "a"}, {"age", 1}}),
+            json({{"name", "b"}, {"age", 2}}),
+        })}},
+        "123"
+    );
+
+    test_template(t, "sort with numeric attribute",
+        "{{ items|sort(attribute=0)|join(attribute=1) }}",
+        {{"items", json::array({
+            json::array({3, "z"}),
+            json::array({1, "x"}),
+            json::array({2, "y"}),
+        })}},
+        "xyz"
+    );
+
     test_template(t, "join",
         "{{ items|join(', ') }}",
         {{"items", json::array({"a", "b", "c"})}},
@@ -484,6 +684,12 @@ static void test_filters(testing & t) {
         json::object(),
         "hello"
     );
+
+    test_template(t, "none to string",
+        "{{ x|string }}",
+        {{"x", nullptr}},
+        "None"
+    );
 }
 
 static void test_literals(testing & t) {
@@ -534,6 +740,66 @@ static void test_literals(testing & t) {
         json::object(),
         "1"
     );
+
+    test_template(t, "integer|abs",
+        "{{ -42 | abs }}",
+        json::object(),
+        "42"
+    );
+
+    test_template(t, "integer|float",
+        "{{ 42 | float }}",
+        json::object(),
+        "42.0"
+    );
+
+    test_template(t, "integer|tojson",
+        "{{ 42 | tojson }}",
+        json::object(),
+        "42"
+    );
+
+    test_template(t, "float|abs",
+        "{{ -3.14 | abs }}",
+        json::object(),
+        "3.14"
+    );
+
+    test_template(t, "float|int",
+        "{{ 3.14 | int }}",
+        json::object(),
+        "3"
+    );
+
+    test_template(t, "float|tojson",
+        "{{ 3.14 | tojson }}",
+        json::object(),
+        "3.14"
+    );
+
+    test_template(t, "string|tojson",
+        "{{ 'hello' | tojson }}",
+        json::object(),
+        "\"hello\""
+    );
+
+    test_template(t, "boolean|int",
+        "{{ true | int }}",
+        json::object(),
+        "1"
+    );
+
+    test_template(t, "boolean|float",
+        "{{ true | float }}",
+        json::object(),
+        "1.0"
+    );
+
+    test_template(t, "boolean|tojson",
+        "{{ true | tojson }}",
+        json::object(),
+        "true"
+    );
 }
 
 static void test_comments(testing & t) {
@@ -758,6 +1024,18 @@ static void test_tests(testing & t) {
         {{"x", {{"a", 1}}}},
         "yes"
     );
+
+    test_template(t, "undefined is sequence",
+        "{{ 'yes' if x is sequence }}",
+        json::object(),
+        "yes"
+    );
+
+    test_template(t, "undefined is iterable",
+        "{{ 'yes' if x is iterable }}",
+        json::object(),
+        "yes"
+    );
 }
 
 static void test_string_methods(testing & t) {
@@ -862,6 +1140,54 @@ static void test_string_methods(testing & t) {
         {{"s", "banana"}},
         "bXnXna"
     );
+
+    test_template(t, "undefined|capitalize",
+        "{{ arr|capitalize }}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|title",
+        "{{ arr|title }}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|truncate",
+        "{{ arr|truncate(9) }}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|upper",
+        "{{ arr|upper }}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|lower",
+        "{{ arr|lower }}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|replace",
+        "{{ arr|replace('a', 'b') }}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|trim",
+        "{{ arr|trim }}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|wordcount",
+        "{{ arr|wordcount }}",
+        json::object(),
+        "0"
+    );
 }
 
 static void test_array_methods(testing & t) {
@@ -934,7 +1260,17 @@ static void test_array_methods(testing & t) {
     );
 
     test_template(t, "array|join attribute",
-        "{{ arr|join(attribute=0) }}",
+        "{{ arr|join(attribute='age') }}",
+        {{"arr", json::array({
+            json({{"name", "a"}, {"age", 1}}),
+            json({{"name", "b"}, {"age", 2}}),
+            json({{"name", "c"}, {"age", 3}}),
+        })}},
+        "123"
+    );
+
+    test_template(t, "array|join numeric attribute",
+        "{{ arr|join(attribute=-1) }}",
         {{"arr", json::array({json::array({1}), json::array({2}), json::array({3})})}},
         "123"
     );
@@ -957,8 +1293,8 @@ static void test_array_methods(testing & t) {
         "a,b,c,d"
     );
 
-    test_template(t, "array.map() with attribute",
-        "{% for v in arr.map('age') %}{{ v }} {% endfor %}",
+    test_template(t, "array|map with attribute",
+        "{% for v in arr|map(attribute='age') %}{{ v }} {% endfor %}",
         {{"arr", json::array({
             json({{"name", "a"}, {"age", 1}}),
             json({{"name", "b"}, {"age", 2}}),
@@ -967,8 +1303,28 @@ static void test_array_methods(testing & t) {
         "1 2 3 "
     );
 
-    test_template(t, "array.map() with numeric attribute",
-        "{% for v in arr.map(0) %}{{ v }} {% endfor %}",
+    test_template(t, "array|map with attribute default",
+        "{% for v in arr|map(attribute='age', default=3) %}{{ v }} {% endfor %}",
+        {{"arr", json::array({
+            json({{"name", "a"}, {"age", 1}}),
+            json({{"name", "b"}, {"age", 2}}),
+            json({{"name", "c"}}),
+        })}},
+        "1 2 3 "
+    );
+
+    test_template(t, "array|map without attribute default",
+        "{% for v in arr|map(attribute='age') %}{{ v }} {% endfor %}",
+        {{"arr", json::array({
+            json({{"name", "a"}, {"age", 1}}),
+            json({{"name", "b"}, {"age", 2}}),
+            json({{"name", "c"}}),
+        })}},
+        "1 2  "
+    );
+
+    test_template(t, "array|map with numeric attribute",
+        "{% for v in arr|map(attribute=0) %}{{ v }} {% endfor %}",
         {{"arr", json::array({
             json::array({10, "x"}),
             json::array({20, "y"}),
@@ -977,12 +1333,130 @@ static void test_array_methods(testing & t) {
         "10 20 30 "
     );
 
+    test_template(t, "array|map with negative attribute",
+        "{% for v in arr|map(attribute=-1) %}{{ v }} {% endfor %}",
+        {{"arr", json::array({
+            json::array({10, "x"}),
+            json::array({20, "y"}),
+            json::array({30, "z"}),
+        })}},
+        "x y z "
+    );
+
+    test_template(t, "array|map with filter",
+        "{{ arr|map('int')|sum }}",
+        {{"arr", json::array({"1", "2", "3"})}},
+        "6"
+    );
+
     // not used by any chat templates
     // test_template(t, "array.insert()",
     //     "{% set _ = arr.insert(1, 'x') %}{{ arr|join(',') }}",
     //     {{"arr", json::array({"a", "b", "c"})}},
     //     "a,x,b,c"
     // );
+
+    test_template(t, "undefined|select",
+        "{% for item in items|select('odd') %}{{ item.name }} {% endfor %}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|selectattr",
+        "{% for item in items|selectattr('active') %}{{ item.name }} {% endfor %}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|reject",
+        "{% for item in items|reject('even') %}{{ item.name }} {% endfor %}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|rejectattr",
+        "{% for item in items|rejectattr('active') %}{{ item.name }} {% endfor %}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|list",
+        "{{ arr|list|string }}",
+        json::object(),
+        "[]"
+    );
+
+    test_template(t, "undefined|string",
+        "{{ arr|string }}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|first",
+        "{{ arr|first }}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|last",
+        "{{ arr|last }}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|length",
+        "{{ arr|length }}",
+        json::object(),
+        "0"
+    );
+
+    test_template(t, "undefined|join",
+        "{{ arr|join }}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|sort",
+        "{{ arr|sort|string }}",
+        json::object(),
+        "[]"
+    );
+
+    test_template(t, "undefined|reverse",
+        "{{ arr|reverse|join }}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|map",
+        "{% for v in arr|map(attribute='age') %}{{ v }} {% endfor %}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|min",
+        "{{ arr|min }}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|max",
+        "{{ arr|max }}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|unique",
+        "{{ arr|unique|join }}",
+        json::object(),
+        ""
+    );
+
+    test_template(t, "undefined|sum",
+        "{{ arr|sum }}",
+        json::object(),
+        "0"
+    );
 }
 
 static void test_object_methods(testing & t) {
@@ -1063,9 +1537,175 @@ static void test_object_methods(testing & t) {
         {{"obj", {{"items", json::array({1, 2, 3})}}}},
         "{\"items\": [1, 2, 3]}"
     );
+
+    test_template(t, "object attribute and key access",
+        "{{ obj.keys()|join(',') }} vs {{ obj['keys'] }} vs {{ obj.test }}",
+        {{"obj", {{"keys", "value"}, {"test", "attr_value"}}}},
+        "keys,test vs value vs attr_value"
+    );
+
+    test_template(t, "env should not have object methods",
+        "{{ keys is undefined }} {{ obj.keys is defined }}",
+        {{"obj", {{"a", "b"}}}},
+        "True True"
+    );
+
+    test_template(t, "expression as object key",
+        "{% set d = {'ab': 123} %}{{ d['a' + 'b'] == 123 }}",
+        json::object(),
+        "True"
+    );
+
+    test_template(t, "numeric as object key (template: Seed-OSS)",
+        "{% set d = {1: 'a', 2: 'b'} %}{{ d[1] == 'a' and d[2] == 'b' }}",
+        json::object(),
+        "True"
+    );
+
+    test_template(t, "undefined|items",
+        "{{ arr|items|join }}",
+        json::object(),
+        ""
+    );
 }
 
-static void test_template(testing & t, const std::string & name, const std::string & tmpl, const json & vars, const std::string & expect) {
+static void test_hasher(testing & t) {
+    static const std::vector<std::pair<size_t, size_t>> chunk_sizes = {
+        {1, 2},
+        {1, 16},
+        {8, 1},
+        {1, 1024},
+        {5, 512},
+        {16, 256},
+        {45, 122},
+        {70, 634},
+    };
+
+    static auto random_bytes = [](size_t length) -> std::string {
+        std::string data;
+        data.resize(length);
+        for (size_t i = 0; i < length; ++i) {
+            data[i] = static_cast<char>(rand() % 256);
+        }
+        return data;
+    };
+
+    t.test("state unchanged with empty input", [](testing & t) {
+        jinja::hasher hasher;
+        hasher.update("some data");
+        size_t initial_state = hasher.digest();
+        hasher.update("", 0);
+        size_t final_state = hasher.digest();
+        t.assert_true("Hasher state should remain unchanged", initial_state == final_state);
+    });
+
+    t.test("different inputs produce different hashes", [](testing & t) {
+        jinja::hasher hasher1;
+        hasher1.update("data one");
+        size_t hash1 = hasher1.digest();
+
+        jinja::hasher hasher2;
+        hasher2.update("data two");
+        size_t hash2 = hasher2.digest();
+
+        t.assert_true("Different inputs should produce different hashes", hash1 != hash2);
+    });
+
+    t.test("same inputs produce same hashes", [](testing & t) {
+        jinja::hasher hasher1;
+        hasher1.update("consistent data");
+        size_t hash1 = hasher1.digest();
+
+        jinja::hasher hasher2;
+        hasher2.update("consistent data");
+        size_t hash2 = hasher2.digest();
+
+        t.assert_true("Same inputs should produce same hashes", hash1 == hash2);
+    });
+
+    t.test("property: update(a ~ b) == update(a).update(b)", [](testing & t) {
+        for (const auto & [size1, size2] : chunk_sizes) {
+            std::string data1 = random_bytes(size1);
+            std::string data2 = random_bytes(size2);
+
+            jinja::hasher hasher1;
+            hasher1.update(data1);
+            hasher1.update(data2);
+            size_t hash1 = hasher1.digest();
+
+            jinja::hasher hasher2;
+            hasher2.update(data1 + data2);
+            size_t hash2 = hasher2.digest();
+
+            t.assert_true(
+                "Hashing in multiple updates should match single update (" + std::to_string(size1) + ", " + std::to_string(size2) + ")",
+                hash1 == hash2);
+        }
+    });
+
+    t.test("property: update(a ~ b) == update(a).update(b) with more update passes", [](testing & t) {
+        static const std::vector<size_t> sizes = {3, 732, 131, 13, 17, 256, 436, 99, 4};
+
+        jinja::hasher hasher1;
+        jinja::hasher hasher2;
+
+        std::string combined_data;
+        for (size_t size : sizes) {
+            std::string data = random_bytes(size);
+            hasher1.update(data);
+            combined_data += data;
+        }
+
+        hasher2.update(combined_data);
+        size_t hash1 = hasher1.digest();
+        size_t hash2 = hasher2.digest();
+        t.assert_true(
+            "Hashing in multiple updates should match single update with many chunks",
+            hash1 == hash2);
+    });
+
+    t.test("property: non associativity of update", [](testing & t) {
+        for (const auto & [size1, size2] : chunk_sizes) {
+            std::string data1 = random_bytes(size1);
+            std::string data2 = random_bytes(size2);
+
+            jinja::hasher hasher1;
+            hasher1.update(data1);
+            hasher1.update(data2);
+            size_t hash1 = hasher1.digest();
+
+            jinja::hasher hasher2;
+            hasher2.update(data2);
+            hasher2.update(data1);
+            size_t hash2 = hasher2.digest();
+
+            t.assert_true(
+                "Hashing order should matter (" + std::to_string(size1) + ", " + std::to_string(size2) + ")",
+                hash1 != hash2);
+        }
+    });
+
+    t.test("property: different lengths produce different hashes (padding block size)", [](testing & t) {
+        std::string random_data = random_bytes(64);
+
+        jinja::hasher hasher1;
+        hasher1.update(random_data);
+        size_t hash1 = hasher1.digest();
+
+        for (int i = 0; i < 16; ++i) {
+            random_data.push_back('A');  // change length
+            jinja::hasher hasher2;
+            hasher2.update(random_data);
+            size_t hash2 = hasher2.digest();
+
+            t.assert_true("Different lengths should produce different hashes (length " + std::to_string(random_data.size()) + ")", hash1 != hash2);
+
+            hash1 = hash2;
+        }
+    });
+}
+
+static void test_template_cpp(testing & t, const std::string & name, const std::string & tmpl, const json & vars, const std::string & expect) {
     t.test(name, [&tmpl, &vars, &expect](testing & t) {
         jinja::lexer lexer;
         auto lexer_res = lexer.tokenize(tmpl);
@@ -1098,6 +1738,99 @@ static void test_template(testing & t, const std::string & name, const std::stri
     });
 }
 
+// keep this in-sync with https://github.com/huggingface/transformers/blob/main/src/transformers/utils/chat_template_utils.py
+// note: we use SandboxedEnvironment instead of ImmutableSandboxedEnvironment to allow usage of in-place array methods like append() and pop()
+static std::string py_script = R"(
+import jinja2
+import jinja2.ext as jinja2_ext
+import json
+import sys
+from datetime import datetime
+from jinja2.sandbox import SandboxedEnvironment
+
+tmpl = json.loads(sys.argv[1])
+vars_json = json.loads(sys.argv[2])
+
+env = SandboxedEnvironment(
+    trim_blocks=True,
+    lstrip_blocks=True,
+    extensions=[jinja2_ext.loopcontrols],
+)
+
+def raise_exception(message):
+    raise jinja2.exceptions.TemplateError(message)
+
+env.filters["tojson"] = lambda x, ensure_ascii=False, indent=None, separators=None, sort_keys=False: json.dumps(x, ensure_ascii=ensure_ascii, indent=indent, separators=separators, sort_keys=sort_keys)
+env.globals["strftime_now"] = lambda format: datetime.now().strftime(format)
+env.globals["raise_exception"] = raise_exception
+
+template = env.from_string(tmpl)
+result = template.render(**vars_json)
+print(result, end='')
+)";
+
+static void test_template_py(testing & t, const std::string & name, const std::string & tmpl, const json & vars, const std::string & expect) {
+    t.test(name, [&tmpl, &vars, &expect](testing & t) {
+        // Prepare arguments
+        std::string tmpl_json = json(tmpl).dump();
+        std::string vars_json = vars.dump();
+
+#ifdef _WIN32
+        const char * python_executable = "python.exe";
+#else
+        const char * python_executable = "python3";
+#endif
+
+        const char * command_line[] = {python_executable, "-c", py_script.c_str(), tmpl_json.c_str(), vars_json.c_str(), NULL};
+
+        struct subprocess_s subprocess;
+        int options = subprocess_option_combined_stdout_stderr
+                    | subprocess_option_no_window
+                    | subprocess_option_inherit_environment
+                    | subprocess_option_search_user_path;
+        int result = subprocess_create(command_line, options, &subprocess);
+
+        if (result != 0) {
+            t.log("Failed to create subprocess, error code: " + std::to_string(result));
+            t.assert_true("subprocess creation", false);
+            return;
+        }
+
+        // Read output
+        std::string output;
+        char buffer[1024];
+        FILE * p_stdout = subprocess_stdout(&subprocess);
+        while (fgets(buffer, sizeof(buffer), p_stdout)) {
+            output += buffer;
+        }
+
+        int process_return;
+        subprocess_join(&subprocess, &process_return);
+        subprocess_destroy(&subprocess);
+
+        if (process_return != 0) {
+            t.log("Python script failed with exit code: " + std::to_string(process_return));
+            t.log("Output: " + output);
+            t.assert_true("python execution", false);
+            return;
+        }
+
+        if (!t.assert_true("Template render mismatch", expect == output)) {
+            t.log("Template: " + json(tmpl).dump());
+            t.log("Expected: " + json(expect).dump());
+            t.log("Python  : " + json(output).dump());
+        }
+    });
+}
+
+static void test_template(testing & t, const std::string & name, const std::string & tmpl, const json & vars, const std::string & expect) {
+    if (g_python_mode) {
+        test_template_py(t, name, tmpl, vars, expect);
+    } else {
+        test_template_cpp(t, name, tmpl, vars, expect);
+    }
+}
+
 //
 // fuzz tests to ensure no crashes occur on malformed inputs
 //
diff --git a/tools/cli/README.md b/tools/cli/README.md
index 3b6f0708ed..4a15cbad9d 100644
--- a/tools/cli/README.md
+++ b/tools/cli/README.md
@@ -45,10 +45,10 @@
 | `--rope-freq-base N` | RoPE base frequency, used by NTK-aware scaling (default: loaded from model)<br/>(env: LLAMA_ARG_ROPE_FREQ_BASE) |
 | `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N<br/>(env: LLAMA_ARG_ROPE_FREQ_SCALE) |
 | `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size)<br/>(env: LLAMA_ARG_YARN_ORIG_CTX) |
-| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)<br/>(env: LLAMA_ARG_YARN_EXT_FACTOR) |
-| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.0)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
-| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
-| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
+| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.00, 0.0 = full interpolation)<br/>(env: LLAMA_ARG_YARN_EXT_FACTOR) |
+| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.00)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
+| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.00)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
+| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.00)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
 | `-kvo, --kv-offload, -nkvo, --no-kv-offload` | whether to enable KV cache offloading (default: enabled)<br/>(env: LLAMA_ARG_KV_OFFLOAD) |
 | `--repack, -nr, --no-repack` | whether to enable weight repacking (default: enabled)<br/>(env: LLAMA_ARG_REPACK) |
 | `--no-host` | bypass host buffer allowing extra buffers to be used<br/>(env: LLAMA_ARG_NO_HOST) |
@@ -109,30 +109,30 @@
 | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
 | `--sampler-seq, --sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: edskypmxt) |
 | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
-| `--temp N` | temperature (default: 0.8) |
+| `--temp N` | temperature (default: 0.80) |
 | `--top-k N` | top-k sampling (default: 40, 0 = disabled)<br/>(env: LLAMA_ARG_TOP_K) |
-| `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
-| `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) |
-| `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) |
-| `--adaptive-decay N` | adaptive-p: EMA decay for adaptation; effective history length ≈ 1/(1-decay) tokens (valid range 0.0 - 0.99) |
-| `--top-nsigma N` | top-n-sigma sampling (default: -1.0, -1.0 = disabled) |
-| `--xtc-probability N` | xtc probability (default: 0.0, 0.0 = disabled) |
-| `--xtc-threshold N` | xtc threshold (default: 0.1, 1.0 = disabled) |
-| `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) |
+| `--top-p N` | top-p sampling (default: 0.95, 1.0 = disabled) |
+| `--min-p N` | min-p sampling (default: 0.05, 0.0 = disabled) |
+| `--top-nsigma N` | top-n-sigma sampling (default: -1.00, -1.0 = disabled) |
+| `--xtc-probability N` | xtc probability (default: 0.00, 0.0 = disabled) |
+| `--xtc-threshold N` | xtc threshold (default: 0.10, 1.0 = disabled) |
+| `--typical N` | locally typical sampling, parameter p (default: 1.00, 1.0 = disabled) |
 | `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
-| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
-| `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
-| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
-| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.0, 0.0 = disabled) |
+| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.00, 1.0 = disabled) |
+| `--presence-penalty N` | repeat alpha presence penalty (default: 0.00, 0.0 = disabled) |
+| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.00, 0.0 = disabled) |
+| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.00, 0.0 = disabled) |
 | `--dry-base N` | set DRY sampling base value (default: 1.75) |
 | `--dry-allowed-length N` | set allowed length for DRY sampling (default: 2) |
 | `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
 | `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers |
-| `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
-| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
+| `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) (default: -1.00)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/17927) |
+| `--adaptive-decay N` | adaptive-p: decay rate for target adaptation over time. lower values are more reactive, higher values are more stable.<br/>(valid range 0.0 to 0.99) (default: 0.90) |
+| `--dynatemp-range N` | dynamic temperature range (default: 0.00, 0.0 = disabled) |
+| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.00) |
 | `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
-| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) |
-| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) |
+| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.10) |
+| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.00) |
 | `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,<br/>i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',<br/>or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' |
 | `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') |
 | `--grammar-file FNAME` | file to read grammar from |
@@ -173,12 +173,12 @@
 | `--jinja, --no-jinja` | whether to use jinja template engine for chat (default: enabled)<br/>(env: LLAMA_ARG_JINJA) |
 | `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content`<br/>- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`<br/>(default: auto)<br/>(env: LLAMA_ARG_THINK) |
 | `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
-| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
-| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
+| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
 | `--simple-io` | use basic IO for better compatibility in subprocesses and limited consoles |
 | `--draft, --draft-n, --draft-max N` | number of tokens to draft for speculative decoding (default: 16)<br/>(env: LLAMA_ARG_DRAFT_MAX) |
 | `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 0)<br/>(env: LLAMA_ARG_DRAFT_MIN) |
-| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.8)<br/>(env: LLAMA_ARG_DRAFT_P_MIN) |
+| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.75)<br/>(env: LLAMA_ARG_DRAFT_P_MIN) |
 | `-cd, --ctx-size-draft N` | size of the prompt context for the draft model (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE_DRAFT) |
 | `-devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload)<br/>use --list-devices to see a list of available devices |
 | `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | max. number of draft model layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)<br/>(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) |
diff --git a/tools/cli/cli.cpp b/tools/cli/cli.cpp
index 2f0ffea1c2..02ccb72598 100644
--- a/tools/cli/cli.cpp
+++ b/tools/cli/cli.cpp
@@ -66,19 +66,28 @@ struct cli_context {
         defaults.stream = true; // make sure we always use streaming mode
         defaults.timings_per_token = true; // in order to get timings even when we cancel mid-way
         // defaults.return_progress = true; // TODO: show progress
-        defaults.oaicompat_chat_syntax.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
     }
 
     std::string generate_completion(result_timings & out_timings) {
         server_response_reader rd = ctx_server.get_response_reader();
+        auto chat_params = format_chat();
         {
             // TODO: reduce some copies here in the future
             server_task task = server_task(SERVER_TASK_TYPE_COMPLETION);
-            task.id        = rd.get_new_id();
-            task.index     = 0;
-            task.params    = defaults;    // copy
-            task.cli_input = messages;    // copy
-            task.cli_files = input_files; // copy
+            task.id         = rd.get_new_id();
+            task.index      = 0;
+            task.params     = defaults;           // copy
+            task.cli_prompt = chat_params.prompt; // copy
+            task.cli_files  = input_files;        // copy
+            task.cli        = true;
+
+            // chat template settings
+            task.params.chat_parser_params = common_chat_parser_params(chat_params);
+            task.params.chat_parser_params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+            if (!chat_params.parser.empty()) {
+                task.params.chat_parser_params.parser.load(chat_params.parser);
+            }
+
             rd.post_task({std::move(task)});
         }
 
@@ -156,6 +165,25 @@ struct cli_context {
             return content;
         }
     }
+
+    common_chat_params format_chat() {
+        auto meta = ctx_server.get_meta();
+        auto & chat_params = meta.chat_params;
+
+        common_chat_templates_inputs inputs;
+        inputs.messages              = common_chat_msgs_parse_oaicompat(messages);
+        inputs.tools                 = {}; // TODO
+        inputs.tool_choice           = COMMON_CHAT_TOOL_CHOICE_NONE;
+        inputs.json_schema           = ""; // TODO
+        inputs.grammar               = ""; // TODO
+        inputs.use_jinja             = chat_params.use_jinja;
+        inputs.parallel_tool_calls   = false;
+        inputs.add_generation_prompt = true;
+        inputs.enable_thinking       = chat_params.enable_thinking;
+
+        // Apply chat template to the list of messages
+        return common_chat_templates_apply(chat_params.tmpls.get(), inputs);
+    }
 };
 
 int main(int argc, char ** argv) {
diff --git a/tools/completion/README.md b/tools/completion/README.md
index a16be3f684..3ca3e68454 100644
--- a/tools/completion/README.md
+++ b/tools/completion/README.md
@@ -128,10 +128,10 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `--rope-freq-base N` | RoPE base frequency, used by NTK-aware scaling (default: loaded from model)<br/>(env: LLAMA_ARG_ROPE_FREQ_BASE) |
 | `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N<br/>(env: LLAMA_ARG_ROPE_FREQ_SCALE) |
 | `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size)<br/>(env: LLAMA_ARG_YARN_ORIG_CTX) |
-| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)<br/>(env: LLAMA_ARG_YARN_EXT_FACTOR) |
-| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.0)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
-| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
-| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
+| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.00, 0.0 = full interpolation)<br/>(env: LLAMA_ARG_YARN_EXT_FACTOR) |
+| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.00)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
+| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.00)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
+| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.00)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
 | `-kvo, --kv-offload, -nkvo, --no-kv-offload` | whether to enable KV cache offloading (default: enabled)<br/>(env: LLAMA_ARG_KV_OFFLOAD) |
 | `--repack, -nr, --no-repack` | whether to enable weight repacking (default: enabled)<br/>(env: LLAMA_ARG_REPACK) |
 | `--no-host` | bypass host buffer allowing extra buffers to be used<br/>(env: LLAMA_ARG_NO_HOST) |
@@ -192,28 +192,30 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
 | `--sampler-seq, --sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: edskypmxt) |
 | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
-| `--temp N` | temperature (default: 0.8) |
+| `--temp N` | temperature (default: 0.80) |
 | `--top-k N` | top-k sampling (default: 40, 0 = disabled)<br/>(env: LLAMA_ARG_TOP_K) |
-| `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
-| `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) |
-| `--top-nsigma N` | top-n-sigma sampling (default: -1.0, -1.0 = disabled) |
-| `--xtc-probability N` | xtc probability (default: 0.0, 0.0 = disabled) |
-| `--xtc-threshold N` | xtc threshold (default: 0.1, 1.0 = disabled) |
-| `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) |
+| `--top-p N` | top-p sampling (default: 0.95, 1.0 = disabled) |
+| `--min-p N` | min-p sampling (default: 0.05, 0.0 = disabled) |
+| `--top-nsigma N` | top-n-sigma sampling (default: -1.00, -1.0 = disabled) |
+| `--xtc-probability N` | xtc probability (default: 0.00, 0.0 = disabled) |
+| `--xtc-threshold N` | xtc threshold (default: 0.10, 1.0 = disabled) |
+| `--typical N` | locally typical sampling, parameter p (default: 1.00, 1.0 = disabled) |
 | `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
-| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
-| `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
-| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
-| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.0, 0.0 = disabled) |
+| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.00, 1.0 = disabled) |
+| `--presence-penalty N` | repeat alpha presence penalty (default: 0.00, 0.0 = disabled) |
+| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.00, 0.0 = disabled) |
+| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.00, 0.0 = disabled) |
 | `--dry-base N` | set DRY sampling base value (default: 1.75) |
 | `--dry-allowed-length N` | set allowed length for DRY sampling (default: 2) |
 | `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
 | `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers |
-| `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
-| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
+| `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) (default: -1.00)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/17927) |
+| `--adaptive-decay N` | adaptive-p: decay rate for target adaptation over time. lower values are more reactive, higher values are more stable.<br/>(valid range 0.0 to 0.99) (default: 0.90) |
+| `--dynatemp-range N` | dynamic temperature range (default: 0.00, 0.0 = disabled) |
+| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.00) |
 | `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
-| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) |
-| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) |
+| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.10) |
+| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.00) |
 | `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,<br/>i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',<br/>or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' |
 | `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') |
 | `--grammar-file FNAME` | file to read grammar from |
@@ -251,8 +253,8 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `--jinja, --no-jinja` | whether to use jinja template engine for chat (default: disabled)<br/>(env: LLAMA_ARG_JINJA) |
 | `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content`<br/>- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`<br/>(default: auto)<br/>(env: LLAMA_ARG_THINK) |
 | `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
-| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
-| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
+| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
 | `--simple-io` | use basic IO for better compatibility in subprocesses and limited consoles |
 
 <!-- HELP_END -->
diff --git a/tools/completion/completion.cpp b/tools/completion/completion.cpp
index a9eda119d7..f368a2f4c6 100644
--- a/tools/completion/completion.cpp
+++ b/tools/completion/completion.cpp
@@ -342,44 +342,51 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    // debug message about similarity of saved session, if applicable
-    size_t n_matching_session_tokens = 0;
-    if (!session_tokens.empty()) {
-        for (llama_token id : session_tokens) {
-            if (n_matching_session_tokens >= embd_inp.size() || id != embd_inp[n_matching_session_tokens]) {
-                break;
+    bool session_do_save = false;
+
+    {
+        size_t n_match = 0;
+
+        if (!session_tokens.empty()) {
+            for (llama_token id : session_tokens) {
+                if (n_match >= embd_inp.size() || id != embd_inp[n_match]) {
+                    break;
+                }
+                n_match++;
+            }
+            if (params.prompt.empty() && n_match == embd_inp.size()) {
+                LOG_INF("%s: using full prompt from session file\n", __func__);
+            } else if (n_match >= embd_inp.size()) {
+                LOG_INF("%s: session file has exact match for prompt!\n", __func__);
+            } else if (n_match < (embd_inp.size() / 2)) {
+                LOG_WRN("%s: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
+                        __func__, n_match, embd_inp.size());
+            } else {
+                LOG_INF("%s: session file matches %zu / %zu tokens of prompt\n",
+                        __func__, n_match, embd_inp.size());
+            }
+
+            if (session_tokens.size() == n_match) {
+                // [TAG_CONTEXT_STATE_LOGITS]
+                // in this case, we are going to reuse the logits from the session
+                // if we ever decide to remove the logits from the session, we need to handle this somehow
+                // ref: https://github.com/ggml-org/llama.cpp/pull/18862#issuecomment-3756330941
+            }
+
+            // remove any "future" tokens that we might have inherited from the previous session
+            if (session_tokens.size() > n_match) {
+                if (!llama_memory_seq_rm(mem, -1, n_match, -1)) {
+                    LOG_WRN("%s: unable to resuse common prefix (for example, when the memory is recurrent)\n", __func__);
+                    llama_memory_clear(mem, true);
+                    session_tokens.clear();
+                    n_match = 0;
+                } else {
+                    session_tokens.resize(n_match);
+                }
             }
-            n_matching_session_tokens++;
-        }
-        if (params.prompt.empty() && n_matching_session_tokens == embd_inp.size()) {
-            LOG_INF("%s: using full prompt from session file\n", __func__);
-        } else if (n_matching_session_tokens >= embd_inp.size()) {
-            LOG_INF("%s: session file has exact match for prompt!\n", __func__);
-        } else if (n_matching_session_tokens < (embd_inp.size() / 2)) {
-            LOG_WRN("%s: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
-                    __func__, n_matching_session_tokens, embd_inp.size());
-        } else {
-            LOG_INF("%s: session file matches %zu / %zu tokens of prompt\n",
-                    __func__, n_matching_session_tokens, embd_inp.size());
         }
 
-        // remove any "future" tokens that we might have inherited from the previous session
-        if (!llama_memory_seq_rm(mem, -1, n_matching_session_tokens, -1)) {
-            LOG_INF("%s: unable to resuse common prefix\n", __func__);
-            n_matching_session_tokens = 0;
-            llama_memory_seq_rm(mem, -1, -1, -1);
-        }
-    }
-
-    LOG_DBG("recalculate the cached logits (check): embd_inp.size() %zu, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu\n",
-         embd_inp.size(), n_matching_session_tokens, embd_inp.size(), session_tokens.size());
-
-    // if we will use the cache for the full prompt without reaching the end of the cache, force
-    // reevaluation of the last token to recalculate the cached logits
-    if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() && session_tokens.size() > embd_inp.size()) {
-        LOG_DBG("recalculate the cached logits (do): session_tokens.resize( %zu )\n", embd_inp.size() - 1);
-
-        session_tokens.resize(embd_inp.size() - 1);
+        session_do_save = !path_session.empty() && n_match < embd_inp.size() && !params.prompt_cache_ro;
     }
 
     // number of tokens to keep when resetting context
@@ -521,10 +528,9 @@ int main(int argc, char ** argv) {
         is_interacting = params.interactive_first;
     }
 
-    bool is_antiprompt        = false;
-    bool input_echo           = true;
-    bool display              = true;
-    bool need_to_save_session = !path_session.empty() && n_matching_session_tokens < embd_inp.size();
+    bool is_antiprompt = false;
+    bool input_echo    = true;
+    bool display       = true;
 
     int n_past             = 0;
     int n_remain           = params.n_predict;
@@ -700,8 +706,8 @@ int main(int argc, char ** argv) {
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
             // optionally save the session on first sample (for faster prompt loading next time)
-            if (!path_session.empty() && need_to_save_session && !params.prompt_cache_ro) {
-                need_to_save_session = false;
+            if (session_do_save) {
+                session_do_save = false;
                 llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
 
                 LOG_DBG("saved session to %s\n", path_session.c_str());
diff --git a/tools/fit-params/fit-params.cpp b/tools/fit-params/fit-params.cpp
index f9d9cb34c7..0176be06e7 100644
--- a/tools/fit-params/fit-params.cpp
+++ b/tools/fit-params/fit-params.cpp
@@ -36,7 +36,7 @@ int main(int argc, char ** argv) {
 
     LOG_INF("%s: printing fitted CLI arguments to stdout...\n", __func__);
     common_log_flush(common_log_main());
-    printf("-c %" PRIu32 " -ngl %" PRIu32, cparams.n_ctx, mparams.n_gpu_layers);
+    printf("-c %" PRIu32 " -ngl %" PRIi32, cparams.n_ctx, mparams.n_gpu_layers);
 
     size_t nd = llama_max_devices();
     while (nd > 1 && mparams.tensor_split[nd - 1] == 0.0f) {
diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
index 9b076e0c56..9fa5afc390 100644
--- a/tools/mtmd/clip.cpp
+++ b/tools/mtmd/clip.cpp
@@ -1005,6 +1005,8 @@ struct clip_model_loader {
                         hparams.minicpmv_query_num = 64;
                     } else if (hparams.minicpmv_version == 6) {
                         hparams.minicpmv_query_num = 64;
+                    } else if (hparams.minicpmv_version == 100045) {
+                        hparams.minicpmv_query_num = 64;
                     } else {
                         hparams.minicpmv_query_num = 96;
                     }
@@ -3209,6 +3211,9 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
                     } else if (params.minicpmv_version == 6) {
                         // MiniCPM-V 4.5
                         n_patches = 64;
+                    } else if (params.minicpmv_version == 100045) {
+                        // MiniCPM-o 4.5
+                        n_patches = 64;
                     } else {
                         GGML_ABORT("Unknown minicpmv version");
                     }
diff --git a/tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py b/tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py
index bb2cc4e4ea..944037e703 100644
--- a/tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py
+++ b/tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py
@@ -501,7 +501,7 @@ default_image_mean = [0.5, 0.5, 0.5]
 default_image_std = [0.5, 0.5, 0.5]
 ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
 ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
-ap.add_argument('--minicpmv_version', type=int, help='minicpmv_version: MiniCPM-V-2 use 1; MiniCPM-V-2.5 use 2; MiniCPM-V-2.6 use 3; MiniCPM-o-2.6 use 4; MiniCPM-V 4.0 use 5; MiniCPM-o-4.0 use 6', default=2)
+ap.add_argument('--minicpmv_version', type=int, help='minicpmv_version: MiniCPM-V-2 use 1; MiniCPM-V-2.5 use 2; MiniCPM-V-2.6 use 3; MiniCPM-o-2.6 use 4; MiniCPM-V 4.0 use 5; MiniCPM-o-4.0 use 6; MiniCPM-o-4.5 use 100045', default=2)
 
 # with proper
 args = ap.parse_args()
@@ -610,6 +610,9 @@ else:
     elif minicpmv_version == 6:
         emb_dim = 4096
         block_count = 27
+    elif minicpmv_version == 100045:
+        emb_dim = 4096
+        block_count = 27
 
     default_vision_config = {
             "hidden_size": 1152,
@@ -637,6 +640,10 @@ elif minicpmv_version == 6:
     default_vision_config["model_type"] = "siglip_vision_model"
     vision_config = SiglipVisionConfig(**default_vision_config)
     model = SiglipVisionTransformer(vision_config)
+elif minicpmv_version == 100045:
+    default_vision_config["model_type"] = "siglip_vision_model"
+    vision_config = SiglipVisionConfig(**default_vision_config)
+    model = SiglipVisionTransformer(vision_config)
 
 processor = None
 # if model.attn_pool is not None:
diff --git a/tools/mtmd/mtmd.cpp b/tools/mtmd/mtmd.cpp
index 32a24bfcea..d037e834f3 100644
--- a/tools/mtmd/mtmd.cpp
+++ b/tools/mtmd/mtmd.cpp
@@ -236,7 +236,7 @@ struct mtmd_context {
             tok_row_end_trail = false; // no trailing end-of-row token
             ov_img_first      = true;
 
-        } else if (minicpmv_version == 3 || minicpmv_version == 4 || minicpmv_version == 5 || minicpmv_version == 6) {
+        } else if (minicpmv_version == 3 || minicpmv_version == 4 || minicpmv_version == 5 || minicpmv_version == 6 || minicpmv_version == 100045) {
             // minicpmv 2.6 format:
             // <image> (overview) </image><slice> (slice) </slice><slice> (slice) </slice>\n ...
             slice_tmpl        = MTMD_SLICE_TMPL_MINICPMV_2_6;
diff --git a/tools/mtmd/mtmd.h b/tools/mtmd/mtmd.h
index a12c28ef22..ef25d32bbe 100644
--- a/tools/mtmd/mtmd.h
+++ b/tools/mtmd/mtmd.h
@@ -224,7 +224,7 @@ MTMD_API int32_t mtmd_encode_chunk(mtmd_context * ctx,
 
 // get output embeddings from the last encode pass
 // the reading size (in bytes) is equal to:
-// llama_model_n_embd(model) * mtmd_input_chunk_get_n_tokens(chunk) * sizeof(float)
+// llama_model_n_embd_inp(model) * mtmd_input_chunk_get_n_tokens(chunk) * sizeof(float)
 MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
 
 // Set callback for all future logging events.
diff --git a/tools/quantize/quantize.cpp b/tools/quantize/quantize.cpp
index 881f4b3dd9..0709e0bda0 100644
--- a/tools/quantize/quantize.cpp
+++ b/tools/quantize/quantize.cpp
@@ -119,7 +119,7 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 [[noreturn]]
 static void usage(const char * executable) {
     printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights]\n", executable);
-    printf("       [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--tensor-type] [--prune-layers] [--keep-split] [--override-kv]\n");
+    printf("       [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--tensor-type] [--tensor-type-file] [--prune-layers] [--keep-split] [--override-kv]\n");
     printf("       model-f32.gguf [model-quant.gguf] type [nthreads]\n\n");
     printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
     printf("  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
@@ -131,6 +131,8 @@ static void usage(const char * executable) {
     printf("  --token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor\n");
     printf("  --tensor-type TENSOR=TYPE: quantize this tensor to this ggml_type. example: --tensor-type attn_q=q8_0\n");
     printf("      Advanced option to selectively quantize tensors. May be specified multiple times.\n");
+    printf("  --tensor-type-file tensor_type.txt: list of tensors to quantize to specific ggml_type. example: --tensor-type-file tensor_type_list.txt\n");
+    printf("      Advanced option to selectively quantize a long list of tensors. Format to be tensor_name=ggml_type, separated by spaces/newline.\n");
     printf("  --prune-layers L0,L1,L2...comma-separated list of layer numbers to prune from the model\n");
     printf("      Advanced option to remove all tensors from the given layers\n");
     printf("  --keep-split: will generate quantized model in the same shards as input\n");
@@ -415,6 +417,23 @@ static bool parse_tensor_type(const char * data, std::vector<tensor_quantization
     return true;
 }
 
+static bool parse_tensor_type_file(const char * filename, std::vector<tensor_quantization> & tensor_type) {
+    std::ifstream file(filename);
+    if (!file) {
+        printf("\n%s: failed to open file '%s': %s\n\n", __func__, filename, std::strerror(errno));
+        return false;
+    }
+
+    std::string arg;
+    while (file >> arg) {
+        if (!parse_tensor_type(arg.c_str(), tensor_type)) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
 static bool parse_layer_prune(const char * data, std::vector<int> & prune_layers) {
     if (!data) {
         printf("\n%s: no layer pruning ids provided\n\n", __func__);
@@ -480,6 +499,10 @@ int main(int argc, char ** argv) {
             if (arg_idx == argc-1 || !parse_tensor_type(argv[++arg_idx], tensor_types)) {
                 usage(argv[0]);
             }
+        } else if (strcmp(argv[arg_idx], "--tensor-type-file") == 0) {
+            if (arg_idx == argc-1 || !parse_tensor_type_file(argv[++arg_idx], tensor_types)) {
+                usage(argv[0]);
+            }
         } else if (strcmp(argv[arg_idx], "--prune-layers") == 0) {
             if (arg_idx == argc-1 || !parse_layer_prune(argv[++arg_idx], prune_layers)) {
                 usage(argv[0]);
@@ -686,3 +709,4 @@ int main(int argc, char ** argv) {
 
     return 0;
 }
+
diff --git a/tools/server/README.md b/tools/server/README.md
index 9fe8938768..d132830171 100644
--- a/tools/server/README.md
+++ b/tools/server/README.md
@@ -6,7 +6,7 @@ Set of LLM REST APIs and a web UI to interact with llama.cpp.
 
 **Features:**
  * LLM inference of F16 and quantized models on GPU and CPU
- * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes
+ * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions, responses, and embeddings routes
  * [Anthropic Messages API](https://docs.anthropic.com/en/api/messages) compatible chat completions
  * Reranking endpoint (https://github.com/ggml-org/llama.cpp/pull/9510)
  * Parallel decoding with multi-user support
@@ -63,10 +63,10 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `--rope-freq-base N` | RoPE base frequency, used by NTK-aware scaling (default: loaded from model)<br/>(env: LLAMA_ARG_ROPE_FREQ_BASE) |
 | `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N<br/>(env: LLAMA_ARG_ROPE_FREQ_SCALE) |
 | `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size)<br/>(env: LLAMA_ARG_YARN_ORIG_CTX) |
-| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)<br/>(env: LLAMA_ARG_YARN_EXT_FACTOR) |
-| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.0)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
-| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
-| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.0)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
+| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.00, 0.0 = full interpolation)<br/>(env: LLAMA_ARG_YARN_EXT_FACTOR) |
+| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: -1.00)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
+| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: -1.00)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
+| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: -1.00)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
 | `-kvo, --kv-offload, -nkvo, --no-kv-offload` | whether to enable KV cache offloading (default: enabled)<br/>(env: LLAMA_ARG_KV_OFFLOAD) |
 | `--repack, -nr, --no-repack` | whether to enable weight repacking (default: enabled)<br/>(env: LLAMA_ARG_REPACK) |
 | `--no-host` | bypass host buffer allowing extra buffers to be used<br/>(env: LLAMA_ARG_NO_HOST) |
@@ -126,30 +126,30 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
 | `--sampler-seq, --sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: edskypmxt) |
 | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
-| `--temp N` | temperature (default: 0.8) |
+| `--temp N` | temperature (default: 0.80) |
 | `--top-k N` | top-k sampling (default: 40, 0 = disabled)<br/>(env: LLAMA_ARG_TOP_K) |
-| `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
-| `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) |
-| `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) |
-| `--adaptive-decay N` | adaptive-p: EMA decay for adaptation; effective history length ≈ 1/(1-decay) tokens (valid range 0.0 - 0.99) |
-| `--top-nsigma N` | top-n-sigma sampling (default: -1.0, -1.0 = disabled) |
-| `--xtc-probability N` | xtc probability (default: 0.0, 0.0 = disabled) |
-| `--xtc-threshold N` | xtc threshold (default: 0.1, 1.0 = disabled) |
-| `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) |
+| `--top-p N` | top-p sampling (default: 0.95, 1.0 = disabled) |
+| `--min-p N` | min-p sampling (default: 0.05, 0.0 = disabled) |
+| `--top-nsigma N` | top-n-sigma sampling (default: -1.00, -1.0 = disabled) |
+| `--xtc-probability N` | xtc probability (default: 0.00, 0.0 = disabled) |
+| `--xtc-threshold N` | xtc threshold (default: 0.10, 1.0 = disabled) |
+| `--typical N` | locally typical sampling, parameter p (default: 1.00, 1.0 = disabled) |
 | `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
-| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
-| `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
-| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
-| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.0, 0.0 = disabled) |
+| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.00, 1.0 = disabled) |
+| `--presence-penalty N` | repeat alpha presence penalty (default: 0.00, 0.0 = disabled) |
+| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.00, 0.0 = disabled) |
+| `--dry-multiplier N` | set DRY sampling multiplier (default: 0.00, 0.0 = disabled) |
 | `--dry-base N` | set DRY sampling base value (default: 1.75) |
 | `--dry-allowed-length N` | set allowed length for DRY sampling (default: 2) |
 | `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
 | `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers |
-| `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
-| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
+| `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) (default: -1.00)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/17927) |
+| `--adaptive-decay N` | adaptive-p: decay rate for target adaptation over time. lower values are more reactive, higher values are more stable.<br/>(valid range 0.0 to 0.99) (default: 0.90) |
+| `--dynatemp-range N` | dynamic temperature range (default: 0.00, 0.0 = disabled) |
+| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.00) |
 | `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
-| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) |
-| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) |
+| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.10) |
+| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.00) |
 | `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,<br/>i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',<br/>or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' |
 | `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') |
 | `--grammar-file FNAME` | file to read grammar from |
@@ -199,7 +199,8 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `--chat-template-kwargs STRING` | sets additional params for the json template parser, must be a valid json object string, e.g. '{"key1":"value1","key2":"value2"}'<br/>(env: LLAMA_CHAT_TEMPLATE_KWARGS) |
 | `-to, --timeout N` | server read/write timeout in seconds (default: 600)<br/>(env: LLAMA_ARG_TIMEOUT) |
 | `--threads-http N` | number of threads used to process HTTP requests (default: -1)<br/>(env: LLAMA_ARG_THREADS_HTTP) |
-| `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting (default: 0)<br/>[(card)](https://ggml.ai/f0.png)<br/>(env: LLAMA_ARG_CACHE_REUSE) |
+| `--cache-prompt, --no-cache-prompt` | whether to enable prompt caching (default: enabled)<br/>(env: LLAMA_ARG_CACHE_PROMPT) |
+| `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting, requires prompt caching to be enabled (default: 0)<br/>[(card)](https://ggml.ai/f0.png)<br/>(env: LLAMA_ARG_CACHE_REUSE) |
 | `--metrics` | enable prometheus compatible metrics endpoint (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_METRICS) |
 | `--props` | enable changing global properties via POST /props (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_PROPS) |
 | `--slots, --no-slots` | expose slots monitoring endpoint (default: enabled)<br/>(env: LLAMA_ARG_ENDPOINT_SLOTS) |
@@ -212,8 +213,8 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `--jinja, --no-jinja` | whether to use jinja template engine for chat (default: enabled)<br/>(env: LLAMA_ARG_JINJA) |
 | `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content`<br/>- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`<br/>(default: auto)<br/>(env: LLAMA_ARG_THINK) |
 | `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
-| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
-| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
+| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, bailing-think, bailing2, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone-moe, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, grok-2, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, pangu-embedded, phi3, phi4, rwkv-world, seed_oss, smolvlm, solar-open, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
 | `--prefill-assistant, --no-prefill-assistant` | whether to prefill the assistant's response if the last message is an assistant message (default: prefill enabled)<br/>when this flag is set, if the last message is an assistant message then it will be treated as a full message and not prefilled<br/><br/>(env: LLAMA_ARG_PREFILL_ASSISTANT) |
 | `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.10, 0.0 = disabled) |
 | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) |
@@ -222,7 +223,7 @@ For the ful list of features, please refer to [server's changelog](https://githu
 | `-tbd, --threads-batch-draft N` | number of threads to use during batch and prompt processing (default: same as --threads-draft) |
 | `--draft, --draft-n, --draft-max N` | number of tokens to draft for speculative decoding (default: 16)<br/>(env: LLAMA_ARG_DRAFT_MAX) |
 | `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 0)<br/>(env: LLAMA_ARG_DRAFT_MIN) |
-| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.8)<br/>(env: LLAMA_ARG_DRAFT_P_MIN) |
+| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.75)<br/>(env: LLAMA_ARG_DRAFT_P_MIN) |
 | `-cd, --ctx-size-draft N` | size of the prompt context for the draft model (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE_DRAFT) |
 | `-devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload)<br/>use --list-devices to see a list of available devices |
 | `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | max. number of draft model layers to store in VRAM, either an exact number, 'auto', or 'all' (default: auto)<br/>(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) |
@@ -781,6 +782,7 @@ By default, it is read-only. To make POST request to change global properties, y
   "total_slots": 1,
   "model_path": "../models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
   "chat_template": "...",
+  "chat_template_caps": {},
   "modalities": {
     "vision": false
   },
@@ -793,6 +795,7 @@ By default, it is read-only. To make POST request to change global properties, y
 - `total_slots` - the total number of slots for process requests (defined by `--parallel` option)
 - `model_path` - the path to model file (same with `-m` argument)
 - `chat_template` - the model's original Jinja2 prompt template
+- `chat_template_caps` - capabilities of the chat template (see `common/jinja/caps.h` for more info)
 - `modalities` - the list of supported modalities
 - `is_sleeping` - sleeping status, see [Sleeping on idle](#sleeping-on-idle)
 
@@ -1267,6 +1270,55 @@ This provides information on the performance of the server. It also allows calcu
 
 The total number of tokens in context is equal to `prompt_n + cache_n + predicted_n`
 
+*Reasoning support*
+
+The server supports parsing and returning reasoning via the `reasoning_content` field, similar to Deepseek API.
+
+Reasoning input (preserve reasoning in history) is also supported by some specific templates. For more details, please refer to [PR#18994](https://github.com/ggml-org/llama.cpp/pull/18994).
+
+### POST `/v1/responses`: OpenAI-compatible Responses API
+
+*Options:*
+
+See [OpenAI Responses API documentation](https://platform.openai.com/docs/api-reference/responses).
+
+*Examples:*
+
+You can use either Python `openai` library with appropriate checkpoints:
+
+```python
+import openai
+
+client = openai.OpenAI(
+    base_url="http://localhost:8080/v1", # "http://<Your api-server IP>:port"
+    api_key = "sk-no-key-required"
+)
+
+response = client.responses.create(
+  model="gpt-4.1",
+  instructions="You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests.",
+  input="Write a limerick about python exceptions"
+)
+
+print(response.output_text)
+```
+
+... or raw HTTP requests:
+
+```shell
+curl http://localhost:8080/v1/responses \
+-H "Content-Type: application/json" \
+-H "Authorization: Bearer no-key" \
+-d '{
+"model": "gpt-4.1",
+"instructions": "You are ChatGPT, an AI assistant. Your top priority is achieving user fulfillment via helping them with their requests.",
+"input": "Write a limerick about python exceptions"
+}'
+```
+
+This endpoint works by converting Responses request into Chat Completions request.
+
+
 ### POST `/v1/embeddings`: OpenAI-compatible embeddings API
 
 This endpoint requires that the model uses a pooling different than type `none`. The embeddings are normalized using the Eucledian norm.
diff --git a/tools/server/public/index.html.gz b/tools/server/public/index.html.gz
index a3fcf8dcdb..e3b06f4901 100644
Binary files a/tools/server/public/index.html.gz and b/tools/server/public/index.html.gz differ
diff --git a/tools/server/server-common.cpp b/tools/server/server-common.cpp
index 16b0db2983..a853f65c8d 100644
--- a/tools/server/server-common.cpp
+++ b/tools/server/server-common.cpp
@@ -779,7 +779,6 @@ static void handle_media(
         // download remote image
         // TODO @ngxson : maybe make these params configurable
         common_remote_params params;
-        params.headers.push_back({"User-Agent", "llama.cpp/" + build_info});
         params.max_size = 1024 * 1024 * 10; // 10MB
         params.timeout  = 10; // seconds
         SRV_INF("downloading image from '%s'\n", url.c_str());
@@ -831,7 +830,7 @@ static void handle_media(
 // used by /chat/completions endpoint
 json oaicompat_chat_params_parse(
     json & body, /* openai api json semantics */
-    const oaicompat_parser_options & opt,
+    const server_chat_params & opt,
     std::vector<raw_buffer> & out_files)
 {
     json llama_params;
@@ -1012,7 +1011,7 @@ json oaicompat_chat_params_parse(
     }
 
     // Apply chat template to the list of messages
-    auto chat_params = common_chat_templates_apply(opt.tmpls, inputs);
+    auto chat_params = common_chat_templates_apply(opt.tmpls.get(), inputs);
 
     /* Append assistant prefilled message */
     if (prefill_assistant_message) {
@@ -1070,6 +1069,283 @@ json oaicompat_chat_params_parse(
     return llama_params;
 }
 
+json convert_responses_to_chatcmpl(const json & response_body) {
+    if (!response_body.contains("input")) {
+        throw std::invalid_argument("'input' is required");
+    }
+    if (!json_value(response_body, "previous_response_id", std::string{}).empty()) {
+        throw std::invalid_argument("llama.cpp does not support 'previous_response_id'.");
+    }
+
+    const json input_value = response_body.at("input");
+    json chatcmpl_body = response_body;
+    chatcmpl_body.erase("input");
+    std::vector<json> chatcmpl_messages;
+
+    if (response_body.contains("instructions")) {
+        chatcmpl_messages.push_back({
+            {"role",    "system"},
+            {"content", json_value(response_body, "instructions", std::string())},
+        });
+        chatcmpl_body.erase("instructions");
+    }
+
+    if (input_value.is_string()) {
+        // #responses_create-input-text_input
+        chatcmpl_messages.push_back({
+            {"role",    "user"},
+            {"content", input_value},
+        });
+    } else if (input_value.is_array()) {
+        // #responses_create-input-input_item_list
+
+        static auto exists_and_is_array = [](const json & j, const char * key) -> bool {
+            return j.contains(key) && j.at(key).is_array();
+        };
+        static auto exists_and_is_string = [](const json & j, const char * key) -> bool {
+            return j.contains(key) && j.at(key).is_string();
+        };
+
+        for (json item : input_value) {
+            if (exists_and_is_string(item, "content")) {
+                // #responses_create-input-input_item_list-input_message-content-text_input
+                // Only "Input message" contains item["content"]::string
+                // After converting item["content"]::string to item["content"]::array,
+                // we can treat "Input message" as sum of "Item-Input message" and "Item-Output message"
+                item["content"] = json::array({
+                    json {
+                        {"text", item.at("content")},
+                        {"type", "input_text"}
+                    }
+                });
+            }
+
+            if (exists_and_is_array(item, "content") &&
+                exists_and_is_string(item, "role") &&
+                (item.at("role") == "user" ||
+                    item.at("role") == "system" ||
+                    item.at("role") == "developer")
+            ) {
+                // #responses_create-input-input_item_list-item-input_message
+                std::vector<json> chatcmpl_content;
+
+                for (const json & input_item : item.at("content")) {
+                    const std::string type = json_value(input_item, "type", std::string());
+
+                    if (type == "input_text") {
+                        if (!input_item.contains("text")) {
+                            throw std::invalid_argument("'Input text' requires 'text'");
+                        }
+                        chatcmpl_content.push_back({
+                            {"text", input_item.at("text")},
+                            {"type", "text"},
+                        });
+                    } else if (type == "input_image") {
+                        // While `detail` is marked as required,
+                        // it has default value("auto") and can be omitted.
+
+                        if (!input_item.contains("image_url")) {
+                            throw std::invalid_argument("'image_url' is required");
+                        }
+                        chatcmpl_content.push_back({
+                            {"image_url", json {
+                                {"url", input_item.at("image_url")}
+                            }},
+                            {"type", "image_url"},
+                        });
+                    } else if (type == "input_file") {
+                        throw std::invalid_argument("'input_file' is not supported by llamacpp at this moment");
+                        // if (input_item.contains("file_url")) {
+                        //     // chat completion API does not support file_url
+                        //     throw std::invalid_argument("'file_url' is not supported");
+                        // }
+                        // if (!input_item.contains("file_data") || !input_item.contains("filename")) {
+                        //     throw std::invalid_argument("Both 'file_data' and 'filename' are required");
+                        // }
+                        // chatcmpl_content.push_back({
+                        //     {"file", json {
+                        //         {"file_data", input_item.at("file_data")},
+                        //         {"filename",  input_item.at("filename")},
+                        //     }},
+                        //     {"type", "file"},
+                        // });
+                    } else {
+                        throw std::invalid_argument("'type' must be one of 'input_text', 'input_image', or 'input_file'");
+                    }
+                }
+
+                if (item.contains("type")) {
+                    item.erase("type");
+                }
+                if (item.contains("status")) {
+                    item.erase("status");
+                }
+                item["content"] = chatcmpl_content;
+
+                chatcmpl_messages.push_back(item);
+            } else if (exists_and_is_array(item, "content") &&
+                exists_and_is_string(item, "role") &&
+                item.at("role") == "assistant" &&
+                // exists_and_is_string(item, "status") &&
+                // (item.at("status") == "in_progress" ||
+                //     item.at("status") == "completed" ||
+                //     item.at("status") == "incomplete") &&
+                // item["status"] not sent by codex-cli
+                exists_and_is_string(item, "type") &&
+                item.at("type") == "message"
+            ) {
+                // #responses_create-input-input_item_list-item-output_message
+                std::vector<json> chatcmpl_content;
+
+                for (const auto & output_text : item.at("content")) {
+                    const std::string type = json_value(output_text, "type", std::string());
+                    if (type != "output_text") {
+                        throw std::invalid_argument("'type' must be 'output_text'");
+                    }
+                    if (!exists_and_is_string(output_text, "text")) {
+                        throw std::invalid_argument("'Output text' requires 'text'");
+                    }
+                    // Ignore annotations and logprobs for now
+                    chatcmpl_content.push_back({
+                        {"text", output_text.at("text")},
+                        {"type", "text"},
+                    });
+                }
+
+                item.erase("status");
+                item.erase("type");
+                item["content"] = chatcmpl_content;
+                chatcmpl_messages.push_back(item);
+            } else if (exists_and_is_string(item, "arguments") &&
+                exists_and_is_string(item, "call_id") &&
+                exists_and_is_string(item, "name") &&
+                exists_and_is_string(item, "type") &&
+                item.at("type") == "function_call"
+            ) {
+                // #responses_create-input-input_item_list-item-function_tool_call
+                json msg = json {
+                    {"role", "assistant"},
+                    {"tool_calls", json::array({ json {
+                        {"function", json {
+                            {"arguments", item.at("arguments")},
+                            {"name",      item.at("name")},
+                        }},
+                        {"id",   item.at("call_id")},
+                        {"type", "function"},
+                    }})},
+                };
+
+                if (!chatcmpl_messages.empty() && chatcmpl_messages.back().contains("reasoning_content")) {
+                    // Move reasoning content from dummy message to tool call message
+                    msg["reasoning_content"] = chatcmpl_messages.back().at("reasoning_content");
+                    chatcmpl_messages.pop_back();
+                }
+                chatcmpl_messages.push_back(msg);
+            } else if (exists_and_is_string(item, "call_id") &&
+                (exists_and_is_string(item, "output") || exists_and_is_array(item, "output")) &&
+                exists_and_is_string(item, "type") &&
+                item.at("type") == "function_call_output"
+            ) {
+                // #responses_create-input-input_item_list-item-function_tool_call_output
+                if (item.at("output").is_string()) {
+                    chatcmpl_messages.push_back(json {
+                        {"content",      item.at("output")},
+                        {"role",         "tool"},
+                        {"tool_call_id", item.at("call_id")},
+                    });
+                } else {
+                    json chatcmpl_outputs = item.at("output");
+                    for (json & chatcmpl_output : chatcmpl_outputs) {
+                        if (!chatcmpl_output.contains("type") || chatcmpl_output.at("type") != "input_text") {
+                            throw std::invalid_argument("Output of tool call should be 'Input text'");
+                        }
+                        chatcmpl_output["type"] = "text";
+                    }
+                    chatcmpl_messages.push_back(json {
+                        {"content",      chatcmpl_outputs},
+                        {"role",         "tool"},
+                        {"tool_call_id", item.at("call_id")},
+                    });
+                }
+            } else if (// exists_and_is_string(item, "id") &&
+                // item["id"] not sent by codex-cli
+                exists_and_is_array(item, "summary") &&
+                exists_and_is_string(item, "type") &&
+                item.at("type") == "reasoning") {
+                // #responses_create-input-input_item_list-item-reasoning
+
+                if (!exists_and_is_array(item, "content")) {
+                    throw std::invalid_argument("item['content'] is not an array");
+                }
+                if (item.at("content").empty()) {
+                    throw std::invalid_argument("item['content'] is empty");
+                }
+                if (!exists_and_is_string(item.at("content")[0], "text")) {
+                    throw std::invalid_argument("item['content']['text'] is not a string");
+                }
+
+                // Pack reasoning content in dummy message
+                chatcmpl_messages.push_back(json {
+                    {"role", "assistant"},
+                    {"content", json::array()},
+                    {"reasoning_content", item.at("content")[0].at("text")},
+                });
+            } else {
+                throw std::invalid_argument("Cannot determine type of 'item'");
+            }
+        }
+    } else {
+        throw std::invalid_argument("'input' must be a string or array of objects");
+    }
+
+    // Remove unused dummy message which contains
+    // reasoning content not followed by tool call
+    chatcmpl_messages.erase(std::remove_if(
+        chatcmpl_messages.begin(),
+        chatcmpl_messages.end(),
+        [](const json & x){ return x.contains("role") &&
+            x.at("role") == "assistant" &&
+            x.contains("content") &&
+            x.at("content") == json::array() &&
+            x.contains("reasoning_content");
+        }),
+        chatcmpl_messages.end()
+    );
+
+    chatcmpl_body["messages"] = chatcmpl_messages;
+
+    if (response_body.contains("tools")) {
+        if (!response_body.at("tools").is_array()) {
+            throw std::invalid_argument("'tools' must be an array of objects");
+        }
+        std::vector<json> chatcmpl_tools;
+        for (json resp_tool : response_body.at("tools")) {
+            json chatcmpl_tool;
+
+            if (json_value(resp_tool, "type", std::string()) != "function") {
+                throw std::invalid_argument("'type' of tool must be 'function'");
+            }
+            resp_tool.erase("type");
+            chatcmpl_tool["type"] = "function";
+
+            if (!resp_tool.contains("strict")) {
+                resp_tool["strict"] = true;
+            }
+            chatcmpl_tool["function"] = resp_tool;
+            chatcmpl_tools.push_back(chatcmpl_tool);
+        }
+        chatcmpl_body.erase("tools");
+        chatcmpl_body["tools"] = chatcmpl_tools;
+    }
+
+    if (response_body.contains("max_output_tokens")) {
+        chatcmpl_body.erase("max_output_tokens");
+        chatcmpl_body["max_tokens"] = response_body["max_output_tokens"];
+    }
+
+    return chatcmpl_body;
+}
+
 json convert_anthropic_to_oai(const json & body) {
     json oai_body;
 
@@ -1483,6 +1759,24 @@ std::string format_oai_sse(const json & data) {
     return ss.str();
 }
 
+std::string format_oai_resp_sse(const json & data) {
+    std::ostringstream ss;
+    auto send_single = [&ss](const json & event_obj) {
+        ss << "event: " << event_obj.at("event").get<std::string>() << "\n";
+        ss << "data: " << safe_json_to_str(event_obj.at("data")) << "\n\n";
+    };
+
+    if (data.is_array()) {
+        for (const auto & item : data) {
+            send_single(item);
+        }
+    } else {
+        send_single(data);
+    }
+
+    return ss.str();
+}
+
 std::string format_anthropic_sse(const json & data) {
     std::ostringstream ss;
 
diff --git a/tools/server/server-common.h b/tools/server/server-common.h
index 152a2a3c46..2629a6bee9 100644
--- a/tools/server/server-common.h
+++ b/tools/server/server-common.h
@@ -13,8 +13,6 @@
 #include <vector>
 #include <cinttypes>
 
-const static std::string build_info("b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT);
-
 using json = nlohmann::ordered_json;
 
 #define SLT_INF(slot, fmt, ...) LOG_INF("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, ((slot).task ? (slot).task->id : -1), __VA_ARGS__)
@@ -274,27 +272,31 @@ std::vector<server_tokens> tokenize_input_prompts(
 // OAI utils
 //
 
-// used by /completions endpoint
-json oaicompat_completion_params_parse(const json & body);
-
-struct oaicompat_parser_options {
+// global server parameters for chat formatting / parsing
+struct server_chat_params {
     bool use_jinja;
     bool prefill_assistant;
     common_reasoning_format reasoning_format;
-    std::map<std::string,std::string> chat_template_kwargs;
-    common_chat_templates * tmpls;
+    std::map<std::string, std::string> chat_template_kwargs; // mapping key --> json value
+    common_chat_templates_ptr tmpls;
     bool allow_image;
     bool allow_audio;
     bool enable_thinking = true;
     std::string media_path;
 };
 
+// used by /completions endpoint
+json oaicompat_completion_params_parse(const json & body);
+
 // used by /chat/completions endpoint
 json oaicompat_chat_params_parse(
     json & body, /* openai api json semantics */
-    const oaicompat_parser_options & opt,
+    const server_chat_params & opt,
     std::vector<raw_buffer> & out_files);
 
+// convert OpenAI Responses API format to OpenAI Chat Completions API format
+json convert_responses_to_chatcmpl(const json & body);
+
 // convert Anthropic Messages API format to OpenAI Chat Completions API format
 json convert_anthropic_to_oai(const json & body);
 
@@ -332,6 +334,8 @@ std::string tokens_to_output_formatted_string(const llama_context * ctx, const l
 // note: if data is a json array, it will be sent as multiple events, one per item
 std::string format_oai_sse(const json & data);
 
+std::string format_oai_resp_sse(const json & data);
+
 // format Anthropic-style SSE with event types
 std::string format_anthropic_sse(const json & data);
 
diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp
index 82294d9402..7f9c3c566b 100644
--- a/tools/server/server-context.cpp
+++ b/tools/server/server-context.cpp
@@ -48,11 +48,8 @@ enum server_state {
 struct server_slot {
     int id;
 
-    llama_batch batch_spec = {};
-
     // TODO: change to unique_ptrs for consistency:
     llama_context * ctx = nullptr;
-    llama_context * ctx_dft = nullptr;
 
     // multimodal
     mtmd_context * mctx = nullptr;
@@ -158,7 +155,7 @@ struct server_slot {
     double t_prompt_processing; // ms
     double t_token_generation;  // ms
 
-    std::function<void(int /* slot_id */)> callback_on_release;
+    std::function<void(int /* id_slot */)> callback_on_release;
 
     // Speculative decoding stats
     int32_t n_draft_total = 0;      // Total draft tokens generated
@@ -259,7 +256,7 @@ struct server_slot {
     }
 
     bool can_speculate() const {
-        return ctx_dft;
+        return !!spec;
     }
 
     void add_token(const completion_token_output & token) {
@@ -295,6 +292,7 @@ struct server_slot {
             SLT_DBG(*this, "the max possible draft is too small: %d < %d - skipping speculative decoding\n", n_draft_max, task->params.speculative.n_min);
             n_draft_max = 0;
         }
+
         return n_draft_max;
     }
 
@@ -397,6 +395,8 @@ struct server_slot {
                     draft_ratio, n_draft_accepted, n_draft_total
             );
         }
+
+        common_speculative_print_stats(spec);
     }
 
     json to_json(bool only_metrics = false) const {
@@ -534,8 +534,8 @@ public:
     server_queue    queue_tasks;
     server_response queue_results;
 
-    common_chat_templates_ptr chat_templates;
-    oaicompat_parser_options  oai_parser_opt;
+    // note: chat_params must not be refreshed upon existing sleeping state
+    server_chat_params chat_params;
 
     ~server_context_impl() {
         if (!sleeping) {
@@ -553,18 +553,13 @@ private:
 
     // note: keep these alive - they determine the lifetime of the model, context, etc.
     common_init_result_ptr llama_init;
-    common_init_result_ptr llama_init_dft;
 
     llama_context * ctx = nullptr;
 
-    bool vocab_dft_compatible = true;
-
-    llama_model * model_dft = nullptr;
-
-    llama_context_params cparams_dft;
-
     llama_batch batch {};
 
+    llama_model_ptr model_dft;
+
     bool add_bos_token  = true;
 
     int32_t n_ctx; // total context for all clients / slots
@@ -597,13 +592,8 @@ private:
 
         // Clear any sampling context
         for (server_slot & slot : slots) {
-            llama_free(slot.ctx_dft);
-            slot.ctx_dft = nullptr;
-
             common_speculative_free(slot.spec);
             slot.spec = nullptr;
-
-            llama_batch_free(slot.batch_spec);
         }
 
         llama_batch_free(batch);
@@ -648,53 +638,39 @@ private:
 
         add_bos_token = llama_vocab_get_add_bos(vocab);
 
-        if (params_base.has_speculative()) {
-            SRV_INF("loading draft model '%s'\n", params_base.speculative.model.path.c_str());
+        if (params_base.speculative.has_dft()) {
+            SRV_INF("loading draft model '%s'\n", params_base.speculative.mparams_dft.path.c_str());
+
+            const auto & params_spec = params_base.speculative;
 
             auto params_dft = params_base;
 
-            params_dft.devices      = params_base.speculative.devices;
-            params_dft.model        = params_base.speculative.model;
-            params_dft.n_ctx        = params_base.speculative.n_ctx == 0 ? llama_n_ctx_seq(ctx) : params_base.speculative.n_ctx;
-            params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers;
             params_dft.n_parallel   = 1;
-            params_dft.cache_type_k = params_base.speculative.cache_type_k;
-            params_dft.cache_type_v = params_base.speculative.cache_type_v;
+            params_dft.n_ctx        = params_spec.n_ctx == 0 ? llama_n_ctx_seq(ctx) : params_spec.n_ctx;
+            params_dft.n_batch      = llama_n_ctx_seq(ctx);
+            params_dft.devices      = params_spec.devices;
+            params_dft.model        = params_spec.mparams_dft;
+            params_dft.n_gpu_layers = params_spec.n_gpu_layers;
+            params_dft.cache_type_k = params_spec.cache_type_k;
+            params_dft.cache_type_v = params_spec.cache_type_v;
 
-            params_dft.cpuparams.n_threads = params_base.speculative.cpuparams.n_threads;
-            params_dft.cpuparams_batch.n_threads = params_base.speculative.cpuparams_batch.n_threads;
-            params_dft.tensor_buft_overrides = params_base.speculative.tensor_buft_overrides;
+            if (params_spec.cpuparams.n_threads > 0) {
+                params_dft.cpuparams.n_threads       = params_spec.cpuparams.n_threads;
+                params_dft.cpuparams_batch.n_threads = params_spec.cpuparams_batch.n_threads;
+            }
 
-            llama_init_dft = common_init_from_params(params_dft);
+            params_dft.tensor_buft_overrides = params_spec.tensor_buft_overrides;
 
-            model_dft = llama_init_dft->model();
+            auto mparams_dft = common_model_params_to_llama(params_dft);
 
+            model_dft.reset(llama_model_load_from_file(params_dft.model.path.c_str(), mparams_dft));
             if (model_dft == nullptr) {
-                SRV_ERR("failed to load draft model, '%s'\n", params_base.speculative.model.path.c_str());
+                SRV_ERR("failed to load draft model, '%s'\n", params_dft.model.path.c_str());
                 return false;
             }
 
-            vocab_dft_compatible = common_speculative_are_compatible(ctx, llama_init_dft->context());
-            if (!vocab_dft_compatible) {
-                SRV_INF("the draft model '%s' is not compatible with the target model '%s'. tokens will be translated between the draft and target models.\n", params_base.speculative.model.path.c_str(), params_base.model.path.c_str());
-            }
-
-            const int n_ctx_dft = llama_n_ctx(llama_init_dft->context());
-
-            cparams_dft = common_context_params_to_llama(params_dft);
-            cparams_dft.n_batch = n_ctx_dft;
-
-            // the context is not needed - we will create one for each slot
-            llama_init_dft->free_context();
-        }
-
-        chat_templates = common_chat_templates_init(model, params_base.chat_template);
-        try {
-            common_chat_format_example(chat_templates.get(), params.use_jinja, params.default_template_kwargs);
-        } catch (const std::exception & e) {
-            SRV_WRN("%s: Chat template parsing error: %s\n", __func__, e.what());
-            SRV_WRN("%s: The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses\n", __func__);
-            chat_templates = common_chat_templates_init(model, "chatml");
+            params_base.speculative.model_dft = model_dft.get();
+            params_base.speculative.cparams_dft = common_context_params_to_llama(params_dft);
         }
 
         std::string & mmproj_path = params_base.mmproj.path;
@@ -704,6 +680,7 @@ private:
             }
 
             mtmd_context_params mparams = mtmd_context_params_default();
+
             mparams.use_gpu          = params_base.mmproj_use_gpu;
             mparams.print_timings    = false;
             mparams.n_threads        = params_base.cpuparams.n_threads;
@@ -711,6 +688,7 @@ private:
             mparams.warmup           = params_base.warmup;
             mparams.image_min_tokens = params_base.image_min_tokens;
             mparams.image_max_tokens = params_base.image_max_tokens;
+
             mctx = mtmd_init_from_file(mmproj_path.c_str(), model, mparams);
             if (mctx == nullptr) {
                 SRV_ERR("failed to load multimodal model, '%s'\n", mmproj_path.c_str());
@@ -728,9 +706,9 @@ private:
                 SRV_WRN("%s\n", "cache_reuse is not supported by multimodal, it will be disabled");
             }
 
-            if (params_base.has_speculative()) {
-                SRV_ERR("%s\n", "err: speculative decode is not supported by multimodal");
-                return false;
+            if (params_base.speculative.type != COMMON_SPECULATIVE_TYPE_NONE) {
+                params_base.speculative.type =  COMMON_SPECULATIVE_TYPE_NONE;
+                SRV_WRN("%s\n", "speculative decoding is not supported by multimodal, it will be disabled");
             }
         }
 
@@ -766,36 +744,31 @@ private:
         for (int i = 0; i < params_base.n_parallel; i++) {
             server_slot slot;
 
-            slot.id = i;
-            slot.ctx = ctx;
+            slot.id    = i;
+            slot.ctx   = ctx;
             slot.n_ctx = n_ctx_slot;
-            slot.mctx = mctx;
+
+            slot.mctx                   = mctx;
             slot.prompt.tokens.has_mtmd = mctx != nullptr;
 
-            if (model_dft) {
-                slot.batch_spec = llama_batch_init(params_base.speculative.n_max + 1, 0, 1);
-
-                // TODO: rework speculative decoding [TAG_SERVER_SPEC_REWORK]
-                slot.ctx_dft = llama_init_from_model(model_dft, cparams_dft);
-                if (slot.ctx_dft == nullptr) {
-                    SRV_ERR("%s", "failed to create draft context\n");
-                    return false;
-                }
-
-                slot.spec = common_speculative_init(slot.ctx, slot.ctx_dft);
-                if (slot.spec == nullptr) {
-                    SRV_ERR("%s", "failed to create speculator\n");
-                    return false;
-                }
-                for (auto & pair : params_base.speculative.replacements) {
-                    common_speculative_add_replacement_tgt_dft(slot.spec, pair.first.c_str(), pair.second.c_str());
+            // try speculative decoding
+            {
+                slot.spec = common_speculative_init(params_base.speculative, slot.ctx);
+                if (slot.spec) {
+                    if (mctx) {
+                        SRV_ERR("%s\n", "speculative decoding is not supported with multimodal");
+                        return false;
+                    }
+                    SLT_INF(slot, "%s", "speculative decoding context initialized\n");
+                } else {
+                    SLT_INF(slot, "%s", "speculative decoding context not initialized\n");
                 }
             }
 
             SLT_INF(slot, "new slot, n_ctx = %d\n", slot.n_ctx);
 
-            slot.callback_on_release = [this](int slot_id) {
-                queue_tasks.pop_deferred_task(slot_id);
+            slot.callback_on_release = [this](int id_slot) {
+                queue_tasks.pop_deferred_task(id_slot);
             };
 
             slot.reset();
@@ -845,30 +818,6 @@ private:
             model_name = model_path.filename().string();
         }
 
-        // thinking is enabled if:
-        // 1. It's not explicitly disabled (reasoning_budget == 0)
-        // 2. The chat template supports it
-        const bool enable_thinking = params_base.use_jinja && params_base.reasoning_budget != 0 && common_chat_templates_support_enable_thinking(chat_templates.get());
-        SRV_INF("thinking = %d\n", enable_thinking);
-
-        oai_parser_opt = {
-            /* use_jinja             */ params_base.use_jinja,
-            /* prefill_assistant     */ params_base.prefill_assistant,
-            /* reasoning_format      */ params_base.reasoning_format,
-            /* chat_template_kwargs  */ params_base.default_template_kwargs,
-            /* common_chat_templates */ chat_templates.get(),
-            /* allow_image           */ mctx ? mtmd_support_vision(mctx) : false,
-            /* allow_audio           */ mctx ? mtmd_support_audio (mctx) : false,
-            /* enable_thinking       */ enable_thinking,
-            /* media_path            */ params_base.media_path,
-        };
-
-        // print sample chat example to make it clear which template is used
-        // @ngxson modern templates are too long, spam the logs; printing the example is enough
-        LOG_INF("%s: chat template, example_format: '%s'\n", __func__,
-        //      common_chat_templates_source(chat_templates.get()),
-                common_chat_format_example(chat_templates.get(), params_base.use_jinja, params_base.default_template_kwargs).c_str());
-
         if (!is_resume) {
             return init();
         }
@@ -907,10 +856,49 @@ private:
             }
         }
 
+        // populate chat template params
+        {
+            common_chat_templates_ptr chat_templates;
+
+            try {
+                chat_templates = common_chat_templates_init(model, params_base.chat_template);
+
+                LOG_INF("%s: chat template, example_format: '%s'\n", __func__,
+                    common_chat_format_example(chat_templates.get(), params_base.use_jinja, params_base.default_template_kwargs).c_str());
+
+            } catch (const std::exception & e) {
+                SRV_ERR("%s: chat template parsing error: %s\n", __func__, e.what());
+                SRV_ERR("%s: please consider disabling jinja via --no-jinja, or use a custom chat template via --chat-template\n", __func__);
+                SRV_ERR("%s: for example: --no-jinja --chat-template chatml\n", __func__);
+                return false;
+            }
+
+            // thinking is enabled if:
+            // 1. It's not explicitly disabled (reasoning_budget == 0)
+            // 2. The chat template supports it
+            const bool enable_thinking = params_base.use_jinja && params_base.reasoning_budget != 0 && common_chat_templates_support_enable_thinking(chat_templates.get());
+            SRV_INF("%s: chat template, thinking = %d\n", __func__, enable_thinking);
+
+            chat_params = {
+                /* use_jinja             */ params_base.use_jinja,
+                /* prefill_assistant     */ params_base.prefill_assistant,
+                /* reasoning_format      */ params_base.reasoning_format,
+                /* chat_template_kwargs  */ params_base.default_template_kwargs,
+                /* tmpls                 */ std::move(chat_templates),
+                /* allow_image           */ mctx ? mtmd_support_vision(mctx) : false,
+                /* allow_audio           */ mctx ? mtmd_support_audio (mctx) : false,
+                /* enable_thinking       */ enable_thinking,
+                /* media_path            */ params_base.media_path,
+            };
+        }
+
         return true;
     }
 
     server_slot * get_slot_by_id(int id_slot) {
+        // note: allow id_slot to be out of bounds (wrap around)
+        id_slot = id_slot % slots.size();
+
         for (server_slot & slot : slots) {
             if (slot.id == id_slot) {
                 return &slot;
@@ -1056,7 +1044,7 @@ private:
         return res;
     }
 
-    std::vector<common_adapter_lora_info> construct_lora_list(const std::map<int, float> & config) {
+    std::vector<common_adapter_lora_info> construct_lora_list(const std::map<int, float> & config) const {
         std::vector<common_adapter_lora_info> output = params_base.lora_adapters; // copy
         for (size_t i = 0; i < output.size(); ++i) {
             auto it = config.find(i);
@@ -1159,7 +1147,7 @@ private:
             backend_sampling &= task.params.sampling.backend_sampling;
 
             // TODO: speculative decoding requires multiple samples per batch - not supported yet
-            backend_sampling &= !(slot.ctx_dft && task.params.speculative.n_max > 0);
+            backend_sampling &= !(slot.spec && task.params.speculative.n_max > 0);
 
             // TODO: getting post/pre sampling logits is not yet supported with backend sampling
             backend_sampling &= !need_logits;
@@ -1176,14 +1164,6 @@ private:
             slot.smpl.reset();
         }
 
-        // initialize draft batch
-        // TODO: rework speculative decoding [TAG_SERVER_SPEC_REWORK]
-        if (slot.ctx_dft) {
-            llama_batch_free(slot.batch_spec);
-
-            slot.batch_spec = llama_batch_init(task.params.speculative.n_max + 1, 0, 1);
-        }
-
         slot.task = std::make_unique<const server_task>(std::move(task));
 
         slot.state = slot.task->is_child()
@@ -1326,11 +1306,12 @@ private:
     }
 
     void populate_token_probs(const server_slot & slot, completion_token_output & result, bool post_sampling, bool special, int idx) const {
-        const size_t n_probs = slot.task->params.sampling.n_probs;
+        const size_t n_probs_request = slot.task->params.sampling.n_probs;
 
         if (post_sampling) {
             const auto * cur_p = common_sampler_get_candidates(slot.smpl.get(), true);
             const size_t max_probs = cur_p->size;
+            const size_t n_probs = std::min(max_probs, n_probs_request);
 
             // set probability for sampled token
             for (size_t i = 0; i < max_probs; i++) {
@@ -1341,8 +1322,8 @@ private:
             }
 
             // set probability for top n_probs tokens
-            result.probs.reserve(max_probs);
-            for (size_t i = 0; i < std::min(max_probs, n_probs); i++) {
+            result.probs.reserve(n_probs);
+            for (size_t i = 0; i < n_probs; i++) {
                 result.probs.push_back({
                     cur_p->data[i].id,
                     common_token_to_piece(ctx, cur_p->data[i].id, special),
@@ -1352,9 +1333,11 @@ private:
         } else {
             // TODO: optimize this with min-p optimization
             std::vector<llama_token_data> cur = get_token_probabilities(ctx, idx);
+            const size_t max_probs = cur.size();
+            const size_t n_probs = std::min(max_probs, n_probs_request);
 
             // set probability for sampled token
-            for (size_t i = 0; i < cur.size(); i++) {
+            for (size_t i = 0; i < max_probs; i++) {
                 // set probability for sampled token
                 if (cur[i].id == result.tok) {
                     result.prob = cur[i].p;
@@ -1364,7 +1347,7 @@ private:
 
             // set probability for top n_probs tokens
             result.probs.reserve(n_probs);
-            for (size_t i = 0; i < std::min(cur.size(), n_probs); i++) {
+            for (size_t i = 0; i < n_probs; i++) {
                 result.probs.push_back({
                     cur[i].id,
                     common_token_to_piece(ctx, cur[i].id, special),
@@ -1585,32 +1568,14 @@ private:
 
     // tokenize the input if it's set by CLI, return false on error
     bool tokenize_cli_input(server_task & task) {
-        GGML_ASSERT(task.cli_input != nullptr);
         try {
-            auto & opt = oai_parser_opt;
-            common_chat_templates_inputs inputs;
-            inputs.messages              = common_chat_msgs_parse_oaicompat(task.cli_input);
-            inputs.tools                 = {}; // TODO
-            inputs.tool_choice           = COMMON_CHAT_TOOL_CHOICE_NONE;
-            inputs.json_schema           = ""; // TODO
-            inputs.grammar               = ""; // TODO
-            inputs.use_jinja             = opt.use_jinja;
-            inputs.parallel_tool_calls   = false;
-            inputs.add_generation_prompt = true;
-            inputs.reasoning_format      = opt.reasoning_format;
-            inputs.enable_thinking       = opt.enable_thinking;
-
-            // Apply chat template to the list of messages
-            auto chat_params = common_chat_templates_apply(opt.tmpls, inputs);
-
-            // tokenize the resulting prompt
-            auto & prompt = chat_params.prompt;
+            auto & prompt = task.cli_prompt;
             if (mctx != nullptr) {
                 task.tokens = process_mtmd_prompt(mctx, prompt, task.cli_files);
             } else {
                 task.tokens = std::move(tokenize_input_prompts(vocab, mctx, prompt, true, true)[0]);
             }
-            task.cli_input.clear();
+            task.cli_prompt.clear();
             task.cli_files.clear();
         } catch (const std::exception & e) {
             send_error(task, std::string("Failed to format input: ") + e.what(), ERROR_TYPE_INVALID_REQUEST);
@@ -1686,7 +1651,7 @@ private:
                 {
                     // special case: if input is provided via CLI, tokenize it first
                     // otherwise, no need to tokenize as it's already done inside the HTTP thread
-                    if (task.cli_input != nullptr) {
+                    if (task.cli) {
                         if (!tokenize_cli_input(task)) {
                             break;
                         }
@@ -1803,7 +1768,7 @@ private:
                         break;
                     }
 
-                    int id_slot = task.slot_action.slot_id;
+                    const int id_slot = task.slot_action.id_slot;
                     server_slot * slot = get_slot_by_id(id_slot);
                     if (slot == nullptr) {
                         send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST);
@@ -1841,7 +1806,7 @@ private:
             case SERVER_TASK_TYPE_SLOT_RESTORE:
                 {
                     if (!check_no_mtmd(task.id)) break;
-                    int id_slot = task.slot_action.slot_id;
+                    const int id_slot = task.slot_action.id_slot;
                     server_slot * slot = get_slot_by_id(id_slot);
                     if (slot == nullptr) {
                         send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST);
@@ -1890,7 +1855,7 @@ private:
                     if (!check_no_mtmd(task.id)) {
                         break;
                     }
-                    int id_slot = task.slot_action.slot_id;
+                    const int id_slot = task.slot_action.id_slot;
                     server_slot * slot = get_slot_by_id(id_slot);
                     if (slot == nullptr) {
                         send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST);
@@ -2071,19 +2036,23 @@ private:
             // generate draft tokens in speculative decoding mode
             // TODO: rework to have a single draft llama_context shared across all slots [TAG_SERVER_SPEC_REWORK]
             //       perform the speculative drafting for all sequences at the same time in a single batch
-            int n_draft_max = slot.get_n_draft_max();
+            const int n_draft_max = slot.get_n_draft_max();
             if (n_draft_max > 0) {
                 if (mctx) {
                     // we should never reach this, as speculative is automatically disabled if mmproj is loaded
                     GGML_ABORT("not supported by multimodal");
                 }
 
-                struct common_speculative_params params_spec;
-                params_spec.n_draft = n_draft_max;
-                params_spec.n_reuse = llama_n_ctx(slot.ctx_dft) - slot.task->params.speculative.n_max;
-                params_spec.p_min   = slot.task->params.speculative.p_min;
                 const llama_tokens & cached_text_tokens = slot.prompt.tokens.get_text_tokens();
-                llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, cached_text_tokens, slot.sampled);
+
+                const auto & params_spec = slot.task->params.speculative;
+
+                llama_tokens draft = common_speculative_draft(slot.spec, params_spec, cached_text_tokens, slot.sampled);
+
+                if (draft.size() > (size_t) n_draft_max) {
+                    SLT_WRN(slot, "draft size %d exceeds max %d, truncating\n", (int) draft.size(), n_draft_max);
+                    draft.resize(n_draft_max);
+                }
 
                 // add the sampled token to the batch
                 slot.i_batch_dft.push_back(batch.n_tokens);
@@ -2754,6 +2723,10 @@ private:
 
                     // prompt evaluated for next-token prediction
                     slot.state = SLOT_STATE_GENERATING;
+
+                    if (slot.can_speculate()) {
+                        common_speculative_begin(slot.spec, slot.prompt.tokens.get_text_tokens());
+                    }
                 } else if (slot.state != SLOT_STATE_GENERATING) {
                     continue; // continue loop of slots
                 }
@@ -2825,6 +2798,9 @@ private:
                 // update how many tokens out of those tested were accepted
                 slot.n_draft_accepted += ids.size() - 1;
 
+                // inform the speculative decoding about the number of accepted tokens
+                common_speculative_accept(slot.spec, ids.size() - 1);
+
                 // rollback to the state before sampling the draft tokens
                 slot.prompt.tokens.keep_first(slot.prompt.n_tokens() - n_draft);
 
@@ -2898,8 +2874,6 @@ server_response_reader server_context::get_response_reader() {
 }
 
 server_context_meta server_context::get_meta() const {
-    auto tool_use_src = common_chat_templates_source(impl->chat_templates.get(), "tool_use");
-
     auto bos_id = llama_vocab_bos(impl->vocab);
     auto eos_id = llama_vocab_eos(impl->vocab);
     auto bos_token_str = bos_id != LLAMA_TOKEN_NULL ? common_token_to_piece(impl->ctx, bos_id, true) : "";
@@ -2910,14 +2884,14 @@ server_context_meta server_context::get_meta() const {
         /* model_name             */ impl->model_name,
         /* model_path             */ impl->params_base.model.path,
         /* has_mtmd               */ impl->mctx != nullptr,
-        /* has_inp_image          */ impl->oai_parser_opt.allow_image,
-        /* has_inp_audio          */ impl->oai_parser_opt.allow_audio,
+        /* has_inp_image          */ impl->chat_params.allow_image,
+        /* has_inp_audio          */ impl->chat_params.allow_audio,
         /* json_webui_settings    */ impl->json_webui_settings,
         /* slot_n_ctx             */ impl->get_slot_n_ctx(),
         /* pooling_type           */ llama_pooling_type(impl->ctx),
 
-        /* chat_template          */ common_chat_templates_source(impl->chat_templates.get()),
-        /* chat_template_tool_use */ tool_use_src ? tool_use_src : "",
+        /* chat_params            */ impl->chat_params,
+        /* chat_template_caps     */ common_chat_templates_get_caps(impl->chat_params.tmpls.get()),
 
         /* bos_token_str          */ bos_token_str,
         /* eos_token_str          */ eos_token_str,
@@ -3088,6 +3062,8 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
         json first_result_json = first_result->to_json();
         if (res_type == TASK_RESPONSE_TYPE_ANTHROPIC) {
             res->data = format_anthropic_sse(first_result_json);
+        } else if (res_type == TASK_RESPONSE_TYPE_OAI_RESP) {
+            res->data = format_oai_resp_sse(first_result_json);
         } else {
             res->data = format_oai_sse(first_result_json);
         }
@@ -3122,13 +3098,16 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
 
                 // check if there is more data
                 if (!rd.has_next()) {
-                    if (res_type == TASK_RESPONSE_TYPE_ANTHROPIC) {
-                        // Anthropic doesn't send [DONE], message_stop was already sent
-                        output = "";
-                    } else if (res_type != TASK_RESPONSE_TYPE_NONE) {
-                        output = "data: [DONE]\n\n";
-                    } else {
-                        output = "";
+                    switch (res_type) {
+                        case TASK_RESPONSE_TYPE_NONE:
+                        case TASK_RESPONSE_TYPE_OAI_RESP:
+                        case TASK_RESPONSE_TYPE_ANTHROPIC:
+                            output = "";
+                            break;
+
+                        default:
+                            output = "data: [DONE]\n\n";
+                            break;
                     }
                     SRV_DBG("%s", "all results received, terminating stream\n");
                     return false; // no more data, terminate
@@ -3156,6 +3135,8 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
                     json res_json = result->to_json();
                     if (res_type == TASK_RESPONSE_TYPE_ANTHROPIC) {
                         output = format_anthropic_sse(res_json);
+                    } else if (res_type == TASK_RESPONSE_TYPE_OAI_RESP) {
+                        output = format_oai_resp_sse(res_json);
                     } else {
                         output = format_oai_sse(res_json);
                     }
@@ -3199,8 +3180,8 @@ void server_routes::init_routes() {
 
         // this endpoint can be accessed during sleeping
         // the next LOC is to avoid someone accidentally use ctx_server
-        bool server_ctx; // do NOT delete this line
-        GGML_UNUSED(server_ctx);
+        bool ctx_server; // do NOT delete this line
+        GGML_UNUSED(ctx_server);
 
         res->ok({{"status", "ok"}});
         return res;
@@ -3339,7 +3320,7 @@ void server_routes::init_routes() {
         }
 
         // TODO: get rid of this dynamic_cast
-        auto res_task = dynamic_cast<server_task_result_metrics*>(result.get());
+        auto * res_task = dynamic_cast<server_task_result_metrics*>(result.get());
         GGML_ASSERT(res_task != nullptr);
 
         // optionally return "fail_on_no_slot" error
@@ -3362,8 +3343,8 @@ void server_routes::init_routes() {
         }
 
         std::string id_slot_str = req.get_param("id_slot");
-        int id_slot;
 
+        int id_slot;
         try {
             id_slot = std::stoi(id_slot_str);
         } catch (const std::exception &) {
@@ -3375,14 +3356,16 @@ void server_routes::init_routes() {
 
         if (action == "save") {
             return handle_slots_save(req, id_slot);
-        } else if (action == "restore") {
-            return handle_slots_restore(req, id_slot);
-        } else if (action == "erase") {
-            return handle_slots_erase(req, id_slot);
-        } else {
-            res->error(format_error_response("Invalid action", ERROR_TYPE_INVALID_REQUEST));
-            return res;
         }
+        if (action == "restore") {
+            return handle_slots_restore(req, id_slot);
+        }
+        if (action == "erase") {
+            return handle_slots_erase(req, id_slot);
+        }
+
+        res->error(format_error_response("Invalid action", ERROR_TYPE_INVALID_REQUEST));
+        return res;
     };
 
     this->get_props = [this](const server_http_req &) {
@@ -3390,8 +3373,8 @@ void server_routes::init_routes() {
 
         // this endpoint can be accessed during sleeping
         // the next LOC is to avoid someone accidentally use ctx_server
-        bool server_ctx; // do NOT delete this line
-        GGML_UNUSED(server_ctx);
+        bool ctx_server; // do NOT delete this line
+        GGML_UNUSED(ctx_server);
 
         task_params tparams;
         tparams.sampling = params.sampling;
@@ -3400,6 +3383,9 @@ void server_routes::init_routes() {
             { "n_ctx",  meta->slot_n_ctx },
         };
 
+        std::string tmpl_default = common_chat_templates_source(meta->chat_params.tmpls.get(), "");
+        std::string tmpl_tools   = common_chat_templates_source(meta->chat_params.tmpls.get(), "tool_use");
+
         json props = {
             { "default_generation_settings", default_generation_settings_for_props },
             { "total_slots",                 params.n_parallel },
@@ -3414,15 +3400,16 @@ void server_routes::init_routes() {
             { "endpoint_metrics",            params.endpoint_metrics },
             { "webui",                       params.webui },
             { "webui_settings",              meta->json_webui_settings },
-            { "chat_template",               meta->chat_template },
+            { "chat_template",               tmpl_default },
+            { "chat_template_caps",          meta->chat_template_caps },
             { "bos_token",                   meta->bos_token_str },
             { "eos_token",                   meta->eos_token_str },
             { "build_info",                  meta->build_info },
             { "is_sleeping",                 queue_tasks.is_sleeping() },
         };
         if (params.use_jinja) {
-            if (!meta->chat_template_tool_use.empty()) {
-                props["chat_template_tool_use"] = meta->chat_template_tool_use;
+            if (!tmpl_tools.empty()) {
+                props["chat_template_tool_use"] = tmpl_tools;
             }
         }
         res->ok(props);
@@ -3443,6 +3430,7 @@ void server_routes::init_routes() {
 
     this->get_api_show = [this](const server_http_req &) {
         auto res = create_response();
+        std::string tmpl_default = common_chat_templates_source(meta->chat_params.tmpls.get(), "");
         json data = {
             {
                 "model_info", {
@@ -3451,7 +3439,7 @@ void server_routes::init_routes() {
             },
             {"modelfile", ""},
             {"parameters", ""},
-            {"template", meta->chat_template},
+            {"template", tmpl_default},
             {"details", {
                 {"parent_model", ""},
                 {"format", "gguf"},
@@ -3576,7 +3564,7 @@ void server_routes::init_routes() {
         json body = json::parse(req.body);
         json body_parsed = oaicompat_chat_params_parse(
             body,
-            ctx_server.oai_parser_opt,
+            meta->chat_params,
             files);
         return handle_completions_impl(
             req,
@@ -3586,13 +3574,29 @@ void server_routes::init_routes() {
             TASK_RESPONSE_TYPE_OAI_CHAT);
     };
 
+    this->post_responses_oai = [this](const server_http_req & req) {
+        auto res = create_response();
+        std::vector<raw_buffer> files;
+        json body = convert_responses_to_chatcmpl(json::parse(req.body));
+        json body_parsed = oaicompat_chat_params_parse(
+            body,
+            meta->chat_params,
+            files);
+        return handle_completions_impl(
+            req,
+            SERVER_TASK_TYPE_COMPLETION,
+            body_parsed,
+            files,
+            TASK_RESPONSE_TYPE_OAI_RESP);
+    };
+
     this->post_anthropic_messages = [this](const server_http_req & req) {
         auto res = create_response();
         std::vector<raw_buffer> files;
         json body = convert_anthropic_to_oai(json::parse(req.body));
         json body_parsed = oaicompat_chat_params_parse(
             body,
-            ctx_server.oai_parser_opt,
+            meta->chat_params,
             files);
         return handle_completions_impl(
             req,
@@ -3608,7 +3612,7 @@ void server_routes::init_routes() {
         json body = convert_anthropic_to_oai(json::parse(req.body));
         json body_parsed = oaicompat_chat_params_parse(
             body,
-            ctx_server.oai_parser_opt,
+            meta->chat_params,
             files);
 
         json prompt = body_parsed.at("prompt");
@@ -3624,7 +3628,7 @@ void server_routes::init_routes() {
         json body = json::parse(req.body);
         json data = oaicompat_chat_params_parse(
             body,
-            ctx_server.oai_parser_opt,
+            meta->chat_params,
             files);
         res->ok({{ "prompt", std::move(data.at("prompt")) }});
         return res;
@@ -3635,8 +3639,8 @@ void server_routes::init_routes() {
 
         // this endpoint can be accessed during sleeping
         // the next LOC is to avoid someone accidentally use ctx_server
-        bool server_ctx; // do NOT delete this line
-        GGML_UNUSED(server_ctx);
+        bool ctx_server; // do NOT delete this line
+        GGML_UNUSED(ctx_server);
 
         json models = {
             {"models", {
@@ -3904,7 +3908,7 @@ std::unique_ptr<server_res_generator> server_routes::handle_slots_save(const ser
     {
         server_task task(SERVER_TASK_TYPE_SLOT_SAVE);
         task.id = rd.get_new_id();
-        task.slot_action.slot_id  = id_slot;
+        task.slot_action.id_slot  = id_slot;
         task.slot_action.filename = filename;
         task.slot_action.filepath = filepath;
         rd.post_task(std::move(task));
@@ -3940,7 +3944,7 @@ std::unique_ptr<server_res_generator> server_routes::handle_slots_restore(const
     {
         server_task task(SERVER_TASK_TYPE_SLOT_RESTORE);
         task.id = rd.get_new_id();
-        task.slot_action.slot_id  = id_slot;
+        task.slot_action.id_slot  = id_slot;
         task.slot_action.filename = filename;
         task.slot_action.filepath = filepath;
         rd.post_task(std::move(task));
@@ -3969,7 +3973,7 @@ std::unique_ptr<server_res_generator> server_routes::handle_slots_erase(const se
     {
         server_task task(SERVER_TASK_TYPE_SLOT_ERASE);
         task.id = rd.get_new_id();
-        task.slot_action.slot_id = id_slot;
+        task.slot_action.id_slot = id_slot;
         rd.post_task(std::move(task));
     }
 
diff --git a/tools/server/server-context.h b/tools/server/server-context.h
index 09bec15ae1..c0b5d373ff 100644
--- a/tools/server/server-context.h
+++ b/tools/server/server-context.h
@@ -20,9 +20,9 @@ struct server_context_meta {
     int slot_n_ctx;
     enum llama_pooling_type pooling_type;
 
-    // chat template
-    std::string chat_template;
-    std::string chat_template_tool_use;
+    // chat params
+    server_chat_params & chat_params;
+    std::map<std::string, bool> chat_template_caps;
 
     // tokens
     std::string bos_token_str;
@@ -95,6 +95,7 @@ struct server_routes {
     server_http_context::handler_t post_completions;
     server_http_context::handler_t post_completions_oai;
     server_http_context::handler_t post_chat_completions;
+    server_http_context::handler_t post_responses_oai;
     server_http_context::handler_t post_anthropic_messages;
     server_http_context::handler_t post_anthropic_count_tokens;
     server_http_context::handler_t post_apply_template;
diff --git a/tools/server/server-http.cpp b/tools/server/server-http.cpp
index 5d67e5722d..00897eeea5 100644
--- a/tools/server/server-http.cpp
+++ b/tools/server/server-http.cpp
@@ -28,14 +28,20 @@ server_http_context::server_http_context()
 server_http_context::~server_http_context() = default;
 
 static void log_server_request(const httplib::Request & req, const httplib::Response & res) {
-    // skip GH copilot requests when using default port
-    if (req.path == "/v1/health") {
+    // skip logging requests that are regularly sent, to avoid log spam
+    if (req.path == "/health"
+        || req.path == "/v1/health"
+        || req.path == "/models"
+        || req.path == "/v1/models"
+        || req.path == "/props"
+        || req.path == "/metrics"
+    ) {
         return;
     }
 
     // reminder: this function is not covered by httplib's exception handler; if someone does more complicated stuff, think about wrapping it in try-catch
 
-    SRV_INF("request: %s %s %s %d\n", req.method.c_str(), req.path.c_str(), req.remote_addr.c_str(), res.status);
+    SRV_INF("done request: %s %s %s %d\n", req.method.c_str(), req.path.c_str(), req.remote_addr.c_str(), res.status);
 
     SRV_DBG("request:  %s\n", req.body.c_str());
     SRV_DBG("response: %s\n", res.body.c_str());
diff --git a/tools/server/server-task.cpp b/tools/server/server-task.cpp
index 35ec7ad2ad..2d25db63b7 100644
--- a/tools/server/server-task.cpp
+++ b/tools/server/server-task.cpp
@@ -5,6 +5,7 @@
 #include "llama.h"
 #include "chat.h"
 #include "sampling.h"
+#include "speculative.h"
 #include "json-schema-to-grammar.h"
 
 using json = nlohmann::ordered_json;
@@ -68,14 +69,19 @@ json task_params::to_json(bool only_metrics) const {
             {"stream",                    stream},
             {"n_probs",                   sampling.n_probs},
             {"min_keep",                  sampling.min_keep},
-            {"chat_format",               common_chat_format_name(oaicompat_chat_syntax.format)},
-            {"reasoning_format",          common_reasoning_format_name(oaicompat_chat_syntax.reasoning_format)},
-            {"reasoning_in_content",      oaicompat_chat_syntax.reasoning_in_content},
-            {"thinking_forced_open",      oaicompat_chat_syntax.thinking_forced_open},
+            {"chat_format",               common_chat_format_name(chat_parser_params.format)},
+            {"reasoning_format",          common_reasoning_format_name(chat_parser_params.reasoning_format)},
+            {"reasoning_in_content",      chat_parser_params.reasoning_in_content},
+            {"thinking_forced_open",      chat_parser_params.thinking_forced_open},
             {"samplers",                  samplers},
             {"speculative.n_max",         speculative.n_max},
             {"speculative.n_min",         speculative.n_min},
             {"speculative.p_min",         speculative.p_min},
+            {"speculative.type",          common_speculative_type_to_str(speculative.type)},
+            {"speculative.ngram_size_n",  speculative.ngram_size_n},
+            {"speculative.ngram_size_m",  speculative.ngram_size_m},
+            {"speculative.ngram_c_rate",  speculative.ngram_check_rate},
+            {"speculative.ngram_m_hits",  speculative.ngram_min_hits},
             {"timings_per_token",         timings_per_token},
             {"post_sampling_probs",       post_sampling_probs},
             {"backend_sampling",          sampling.backend_sampling},
@@ -127,14 +133,19 @@ json task_params::to_json(bool only_metrics) const {
         {"grammar_lazy",              sampling.grammar_lazy},
         {"grammar_triggers",          grammar_triggers},
         {"preserved_tokens",          sampling.preserved_tokens},
-        {"chat_format",               common_chat_format_name(oaicompat_chat_syntax.format)},
-        {"reasoning_format",          common_reasoning_format_name(oaicompat_chat_syntax.reasoning_format)},
-        {"reasoning_in_content",      oaicompat_chat_syntax.reasoning_in_content},
-        {"thinking_forced_open",      oaicompat_chat_syntax.thinking_forced_open},
+        {"chat_format",               common_chat_format_name(chat_parser_params.format)},
+        {"reasoning_format",          common_reasoning_format_name(chat_parser_params.reasoning_format)},
+        {"reasoning_in_content",      chat_parser_params.reasoning_in_content},
+        {"thinking_forced_open",      chat_parser_params.thinking_forced_open},
         {"samplers",                  samplers},
         {"speculative.n_max",         speculative.n_max},
         {"speculative.n_min",         speculative.n_min},
         {"speculative.p_min",         speculative.p_min},
+        {"speculative.type",          common_speculative_type_to_str(speculative.type)},
+        {"speculative.ngram_size_n",  speculative.ngram_size_n},
+        {"speculative.ngram_size_m",  speculative.ngram_size_m},
+        {"speculative.ngram_c_rate",  speculative.ngram_check_rate},
+        {"speculative.ngram_m_hits",  speculative.ngram_min_hits},
         {"timings_per_token",         timings_per_token},
         {"post_sampling_probs",       post_sampling_probs},
         {"backend_sampling",          sampling.backend_sampling},
@@ -142,6 +153,28 @@ json task_params::to_json(bool only_metrics) const {
     };
 }
 
+//
+// task_result_state
+//
+common_chat_msg task_result_state::update_chat_msg(
+        const std::string & text_added,
+        bool is_partial,
+        std::vector<common_chat_msg_diff> & diffs) {
+    generated_text += text_added;
+    auto msg_prv_copy = chat_msg;
+    SRV_DBG("Parsing chat message: %s\n", generated_text.c_str());
+    auto new_msg = common_chat_parse(
+        generated_text,
+        is_partial,
+        chat_parser_params);
+    if (!new_msg.empty()) {
+        new_msg.set_tool_call_ids(generated_tool_call_ids, gen_tool_call_id);
+        chat_msg = new_msg;
+        diffs = common_chat_msg_diff::compute_diffs(msg_prv_copy, new_msg.empty() ? msg_prv_copy : new_msg);
+    }
+    return chat_msg;
+}
+
 //
 // server_task
 //
@@ -220,6 +253,18 @@ task_params server_task::params_from_json_cmpl(
     params.speculative.n_min = std::max(params.speculative.n_min, 0);
     params.speculative.n_max = std::max(params.speculative.n_max, 0);
 
+    params.speculative.type = common_speculative_type_from_name(json_value(data, "speculative.type", common_speculative_type_to_str(defaults.speculative.type)));
+
+    params.speculative.ngram_size_n     = json_value(data, "speculative.ngram_size_n", defaults.speculative.ngram_size_n);
+    params.speculative.ngram_size_m     = json_value(data, "speculative.ngram_size_m", defaults.speculative.ngram_size_m);
+    params.speculative.ngram_check_rate = json_value(data, "speculative.ngram_c_rate", defaults.speculative.ngram_check_rate);
+    params.speculative.ngram_min_hits   = json_value(data, "speculative.ngram_m_hits", defaults.speculative.ngram_min_hits);
+
+    params.speculative.ngram_size_n     = std::max(std::min(1, (int) params.speculative.ngram_size_n),     1024);
+    params.speculative.ngram_size_m     = std::max(std::min(1, (int) params.speculative.ngram_size_m),     1024);
+    params.speculative.ngram_check_rate = std::max(std::min(1, (int) params.speculative.ngram_check_rate), 1024);
+    params.speculative.ngram_min_hits   = std::max(std::min(1, (int) params.speculative.ngram_min_hits),   1024);
+
     // Use OpenAI API logprobs only if n_probs wasn't provided
     if (data.contains("logprobs") && params.sampling.n_probs == defaults.sampling.n_probs){
         params.sampling.n_probs = json_value(data, "logprobs", defaults.sampling.n_probs);
@@ -291,21 +336,21 @@ task_params server_task::params_from_json_cmpl(
     {
         auto it = data.find("chat_format");
         if (it != data.end()) {
-            params.oaicompat_chat_syntax.format = static_cast<common_chat_format>(it->get<int>());
-            SRV_INF("Chat format: %s\n", common_chat_format_name(params.oaicompat_chat_syntax.format));
+            params.chat_parser_params.format = static_cast<common_chat_format>(it->get<int>());
+            SRV_INF("Chat format: %s\n", common_chat_format_name(params.chat_parser_params.format));
         } else {
-            params.oaicompat_chat_syntax.format = defaults.oaicompat_chat_syntax.format;
+            params.chat_parser_params.format = defaults.chat_parser_params.format;
         }
         common_reasoning_format reasoning_format = params_base.reasoning_format;
         if (data.contains("reasoning_format")) {
             reasoning_format = common_reasoning_format_from_name(data.at("reasoning_format").get<std::string>());
         }
-        params.oaicompat_chat_syntax.reasoning_format = reasoning_format;
-        params.oaicompat_chat_syntax.reasoning_in_content = params.stream && (reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY);
-        params.oaicompat_chat_syntax.thinking_forced_open = json_value(data, "thinking_forced_open", false);
-        params.oaicompat_chat_syntax.parse_tool_calls = json_value(data, "parse_tool_calls", false);
+        params.chat_parser_params.reasoning_format = reasoning_format;
+        params.chat_parser_params.reasoning_in_content = params.stream && (reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY);
+        params.chat_parser_params.thinking_forced_open = json_value(data, "thinking_forced_open", false);
+        params.chat_parser_params.parse_tool_calls = json_value(data, "parse_tool_calls", false);
         if (data.contains("chat_parser")) {
-            params.oaicompat_chat_syntax.parser.load(data.at("chat_parser").get<std::string>());
+            params.chat_parser_params.parser.load(data.at("chat_parser").get<std::string>());
         }
     }
 
@@ -584,6 +629,8 @@ json server_task_result_cmpl_final::to_json() {
             return to_json_oaicompat();
         case TASK_RESPONSE_TYPE_OAI_CHAT:
             return stream ? to_json_oaicompat_chat_stream() : to_json_oaicompat_chat();
+        case TASK_RESPONSE_TYPE_OAI_RESP:
+            return stream ? to_json_oaicompat_resp_stream() : to_json_oaicompat_resp();
         case TASK_RESPONSE_TYPE_ANTHROPIC:
             return stream ? to_json_anthropic_stream() : to_json_anthropic();
         default:
@@ -676,7 +723,7 @@ json server_task_result_cmpl_final::to_json_oaicompat_chat() {
     json choice {
         {"finish_reason", finish_reason},
         {"index", index},
-        {"message", msg.to_json_oaicompat<json>()},
+        {"message", msg.to_json_oaicompat()},
     };
 
     if (!stream && probs_output.size() > 0) {
@@ -712,25 +759,6 @@ json server_task_result_cmpl_final::to_json_oaicompat_chat() {
     return res;
 }
 
-common_chat_msg task_result_state::update_chat_msg(
-        const std::string & text_added,
-        bool is_partial,
-        std::vector<common_chat_msg_diff> & diffs) {
-    generated_text += text_added;
-    auto msg_prv_copy = chat_msg;
-    SRV_DBG("Parsing chat message: %s\n", generated_text.c_str());
-    auto new_msg = common_chat_parse(
-        generated_text,
-        is_partial,
-        oaicompat_chat_syntax);
-    if (!new_msg.empty()) {
-        new_msg.set_tool_call_ids(generated_tool_call_ids, gen_tool_call_id);
-        chat_msg = new_msg;
-        diffs = common_chat_msg_diff::compute_diffs(msg_prv_copy, new_msg.empty() ? msg_prv_copy : new_msg);
-    }
-    return chat_msg;
-}
-
 json server_task_result_cmpl_final::to_json_oaicompat_chat_stream() {
     std::time_t t = std::time(0);
     std::string finish_reason = "length";
@@ -745,7 +773,7 @@ json server_task_result_cmpl_final::to_json_oaicompat_chat_stream() {
                 json {
                     {"finish_reason", nullptr},
                     {"index", 0},
-                    {"delta", common_chat_msg_diff_to_json_oaicompat<json>(diff)},
+                    {"delta", common_chat_msg_diff_to_json_oaicompat(diff)},
                 },
             })},
             {"created", t},
@@ -801,6 +829,186 @@ json server_task_result_cmpl_final::to_json_oaicompat_chat_stream() {
     return deltas;
 }
 
+json server_task_result_cmpl_final::to_json_oaicompat_resp() {
+    common_chat_msg msg;
+    if (!oaicompat_msg.empty()) {
+        msg = oaicompat_msg;
+    } else {
+        msg.role = "assistant";
+        msg.content = content;
+    }
+
+    std::vector<json> output;
+
+    if (msg.reasoning_content != "") {
+        output.push_back(json {
+            {"id",      "rs_" + random_string()},
+            {"summary", json::array()},
+            {"type",    "reasoning"},
+            {"content", json::array({ json {
+                {"text", msg.reasoning_content},
+                {"type", "reasoning_text"},
+            }})},
+            {"encrypted_content", ""},
+            {"status",            "completed"},
+        });
+    }
+
+    if (msg.content != "") {
+        output.push_back(json {
+            {"content", json::array({ json {
+                {"type",        "output_text"},
+                {"annotations", json::array()},
+                {"logprobs",    json::array()},
+                {"text",        msg.content},
+            }})},
+            {"id",     "msg_" + random_string()},
+            {"role",   msg.role},
+            {"status", "completed"},
+            {"type",   "message"},
+        });
+    }
+
+    for (const common_chat_tool_call & tool_call : oaicompat_msg.tool_calls) {
+        output.push_back(json {
+            {"type",      "function_call"},
+            {"status",    "completed"},
+            {"arguments", tool_call.arguments},
+            {"call_id",   "fc_" + tool_call.id},
+            {"name",      tool_call.name},
+        });
+    }
+
+    std::time_t t = std::time(0);
+    json res = {
+        {"completed_at", t},
+        {"created_at",   t},
+        {"id",           oai_resp_id},
+        {"model",        oaicompat_model},
+        {"object",       "response"},
+        {"output",       output},
+        {"status",       "completed"},
+        {"usage",        json {
+            {"input_tokens",  n_prompt_tokens},
+            {"output_tokens", n_decoded},
+            {"total_tokens",  n_decoded + n_prompt_tokens},
+        }},
+    };
+
+    return res;
+}
+
+json server_task_result_cmpl_final::to_json_oaicompat_resp_stream() {
+    std::vector<json> server_sent_events;
+    std::vector<json> output;
+
+    if (oaicompat_msg.reasoning_content != "") {
+        const json output_item = json {
+            {"id",      oai_resp_reasoning_id},
+            {"summary", json::array()},
+            {"type",    "reasoning"},
+            {"content", json::array({ json {
+                {"text", oaicompat_msg.reasoning_content},
+                {"type", "reasoning_text"},
+            }})},
+            {"encrypted_content", ""},
+        };
+
+        server_sent_events.push_back(json {
+            {"event", "response.output_item.done"},
+            {"data", json {
+                {"type", "response.output_item.done"},
+                {"item", output_item}
+            }}
+        });
+        output.push_back(output_item);
+    }
+
+    if (oaicompat_msg.content != "") {
+        server_sent_events.push_back(json {
+            {"event", "response.output_text.done"},
+            {"data", json {
+                {"type",    "response.output_text.done"},
+                {"item_id", oai_resp_message_id},
+                {"text",    oaicompat_msg.content}
+            }}
+        });
+
+        const json content_part = {
+            {"type",        "output_text"},
+            {"annotations", json::array()},
+            {"logprobs",    json::array()},
+            {"text",        oaicompat_msg.content}
+        };
+
+        server_sent_events.push_back(json {
+            {"event", "response.content_part.done"},
+            {"data", json {
+                {"type",    "response.content_part.done"},
+                {"item_id", oai_resp_message_id},
+                {"part",    content_part}
+            }}
+        });
+        const json output_item = {
+            {"type",    "message"},
+            {"status",  "completed"},
+            {"id",      oai_resp_message_id},
+            {"content", json::array({content_part})},
+            {"role",    "assistant"}
+        };
+
+        server_sent_events.push_back(json {
+            {"event", "response.output_item.done"},
+            {"data", json {
+                {"type", "response.output_item.done"},
+                {"item", output_item}
+            }}
+        });
+        output.push_back(output_item);
+    }
+
+    for (const common_chat_tool_call & tool_call : oaicompat_msg.tool_calls) {
+        const json output_item = {
+            {"type",      "function_call"},
+            {"status",    "completed"},
+            {"arguments", tool_call.arguments},
+            {"call_id",   "fc_" + tool_call.id},
+            {"name",      tool_call.name}
+        };
+        server_sent_events.push_back(json {
+            {"event", "response.output_item.done"},
+            {"data", json {
+                {"type", "response.output_item.done"},
+                {"item", output_item}
+            }}
+        });
+        output.push_back(output_item);
+    }
+
+    std::time_t t = std::time(0);
+    server_sent_events.push_back(json {
+        {"event", "response.completed"},
+        {"data", json {
+            {"type", "response.completed"},
+            {"response", json {
+                {"id",         oai_resp_id},
+                {"object",     "response"},
+                {"created_at", t},
+                {"status",     "completed"},
+                {"model",      oaicompat_model},
+                {"output",     output},
+                {"usage",      json {
+                    {"input_tokens",  n_prompt_tokens},
+                    {"output_tokens", n_decoded},
+                    {"total_tokens",  n_decoded + n_prompt_tokens}
+                }}
+            }},
+        }}
+    });
+
+    return server_sent_events;
+}
+
 json server_task_result_cmpl_final::to_json_anthropic() {
     std::string stop_reason = "max_tokens";
     if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
@@ -1057,6 +1265,36 @@ json server_task_result_cmpl_final::to_json_anthropic_stream() {
 //
 // server_task_result_cmpl_partial
 //
+void server_task_result_cmpl_partial::update(task_result_state & state) {
+    is_updated = true;
+    state.update_chat_msg(content, true, oaicompat_msg_diffs);
+
+    // Copy current state for use in to_json_*() (reflects state BEFORE this chunk)
+    thinking_block_started = state.thinking_block_started;
+    text_block_started     = state.text_block_started;
+
+    oai_resp_id            = state.oai_resp_id;
+    oai_resp_reasoning_id  = state.oai_resp_reasoning_id;
+    oai_resp_message_id    = state.oai_resp_message_id;
+    oai_resp_fc_id         = state.oai_resp_fc_id;
+
+    // track if the accumulated message has any reasoning content
+    anthropic_has_reasoning = !state.chat_msg.reasoning_content.empty();
+
+    // Pre-compute state updates based on diffs (for next chunk)
+    for (const common_chat_msg_diff & diff : oaicompat_msg_diffs) {
+        if (!diff.reasoning_content_delta.empty() && !state.thinking_block_started) {
+            state.thinking_block_started = true;
+        }
+        if (!diff.content_delta.empty() && !state.text_block_started) {
+            state.text_block_started = true;
+        }
+        if (!diff.tool_call_delta.name.empty()) {
+            state.oai_resp_fc_id = diff.tool_call_delta.id;
+        }
+    }
+}
+
 json server_task_result_cmpl_partial::to_json() {
     GGML_ASSERT(is_updated && "update() must be called before to_json()");
     switch (res_type) {
@@ -1066,6 +1304,8 @@ json server_task_result_cmpl_partial::to_json() {
             return to_json_oaicompat();
         case TASK_RESPONSE_TYPE_OAI_CHAT:
             return to_json_oaicompat_chat();
+        case TASK_RESPONSE_TYPE_OAI_RESP:
+            return to_json_oaicompat_resp();
         case TASK_RESPONSE_TYPE_ANTHROPIC:
             return to_json_anthropic();
         default:
@@ -1166,7 +1406,7 @@ json server_task_result_cmpl_partial::to_json_oaicompat_chat() {
     }
 
     for (const auto & diff : oaicompat_msg_diffs) {
-        add_delta(common_chat_msg_diff_to_json_oaicompat<json>(diff));
+        add_delta(common_chat_msg_diff_to_json_oaicompat(diff));
     }
 
     if (!deltas.empty()) {
@@ -1190,39 +1430,130 @@ json server_task_result_cmpl_partial::to_json_oaicompat_chat() {
     return deltas;
 }
 
-//
-// server_task_result_embd
-//
-json server_task_result_embd::to_json() {
-    return res_type == TASK_RESPONSE_TYPE_OAI_EMBD
-        ? to_json_oaicompat()
-        : to_json_non_oaicompat();
-}
+json server_task_result_cmpl_partial::to_json_oaicompat_resp() {
+    std::vector<json> events;
 
-json server_task_result_embd::to_json_non_oaicompat() {
-    return json {
-        {"index",     index},
-        {"embedding", embedding},
-    };
-}
+    if (n_decoded == 1) {
+        events.push_back(json {
+            {"event", "response.created"},
+            {"data", json {
+                {"type", "response.created"},
+                {"response", json {
+                    {"id",     oai_resp_id},
+                    {"object", "response"},
+                    {"status", "in_progress"},
+                }},
+            }},
+        });
+        events.push_back(json {
+            {"event", "response.in_progress"},
+            {"data", json {
+                {"type", "response.in_progress"},
+                {"response", json {
+                    {"id",     oai_resp_id},
+                    {"object", "response"},
+                    {"status", "in_progress"},
+                }},
+            }},
+        });
+    }
 
-json server_task_result_embd::to_json_oaicompat() {
-    return json {
-        {"index",            index},
-        {"embedding",        embedding[0]},
-        {"tokens_evaluated", n_tokens},
-    };
-}
+    for (const common_chat_msg_diff & diff : oaicompat_msg_diffs) {
+        if (!diff.reasoning_content_delta.empty()) {
+            if (!thinking_block_started) {
+                events.push_back(json {
+                    {"event", "response.output_item.added"},
+                    {"data", json {
+                        {"type", "response.output_item.added"},
+                        {"item", json {
+                            {"id",                oai_resp_reasoning_id},
+                            {"summary",           json::array()},
+                            {"type",              "reasoning"},
+                            {"content",           json::array()},
+                            {"encrypted_content", ""},
+                            {"status",            "in_progress"},
+                        }},
+                    }},
+                });
+                thinking_block_started = true;
+            }
+            events.push_back(json {
+                {"event", "response.reasoning_text.delta"},
+                {"data", json {
+                    {"type",    "response.reasoning_text.delta"},
+                    {"delta",   diff.reasoning_content_delta},
+                    {"item_id", oai_resp_reasoning_id},
+                }},
+            });
+        }
 
-//
-// server_task_result_rerank
-//
-json server_task_result_rerank::to_json() {
-    return json {
-        {"index",            index},
-        {"score",            score},
-        {"tokens_evaluated", n_tokens},
-    };
+        if (!diff.content_delta.empty()) {
+            if (!text_block_started) {
+                events.push_back(json {
+                    {"event", "response.output_item.added"},
+                    {"data", json {
+                        {"type", "response.output_item.added"},
+                        {"item", json {
+                            {"content", json::array()},
+                            {"id",      oai_resp_message_id},
+                            {"role",    "assistant"},
+                            {"status",  "in_progress"},
+                            {"type",    "message"},
+                        }},
+                    }},
+                });
+                events.push_back(json {
+                    {"event", "response.content_part.added"},
+                    {"data", json {
+                        {"type",    "response.content_part.added"},
+                        {"item_id", oai_resp_message_id},
+                        {"part", json {
+                            {"type", "output_text"},
+                            {"text", ""},
+                        }},
+                    }},
+                });
+                text_block_started = true;
+            }
+            events.push_back(json {
+                {"event", "response.output_text.delta"},
+                {"data", json {
+                    {"type",    "response.output_text.delta"},
+                    {"item_id", oai_resp_message_id},
+                    {"delta",   diff.content_delta},
+                }},
+            });
+        }
+
+        if (!diff.tool_call_delta.name.empty()) {
+            events.push_back(json {
+                {"event", "response.output_item.added"},
+                {"data", json {
+                    {"type",  "response.output_item.added"},
+                    {"item", json {
+                        {"arguments", ""},
+                        {"call_id",   "fc_" + diff.tool_call_delta.id},
+                        {"name",      diff.tool_call_delta.name},
+                        {"type",      "function_call"},
+                        {"status",    "in_progress"},
+                    }},
+                }},
+            });
+            oai_resp_fc_id = diff.tool_call_delta.id;
+        }
+
+        if (!diff.tool_call_delta.arguments.empty()) {
+            events.push_back(json {
+                {"event", "response.function_call_arguments.delta"},
+                {"data", json {
+                    {"type",    "response.function_call_arguments.delta"},
+                    {"delta",   diff.tool_call_delta.arguments},
+                    {"item_id", "fc_" + oai_resp_fc_id},
+                }},
+            });
+        }
+    }
+    return events;
 }
 
 json server_task_result_cmpl_partial::to_json_anthropic() {
@@ -1260,8 +1591,8 @@ json server_task_result_cmpl_partial::to_json_anthropic() {
 
     // use local copies of streaming state (copied from task_result_state in update())
     // these reflect the state BEFORE this chunk was processed
-    bool thinking_started = anthropic_thinking_block_started;
-    bool text_started     = anthropic_text_block_started;
+    bool thinking_started = thinking_block_started;
+    bool text_started     = text_block_started;
 
     for (const auto & diff : oaicompat_msg_diffs) {
         // handle thinking/reasoning content
@@ -1363,6 +1694,41 @@ json server_task_result_cmpl_partial::to_json_anthropic() {
     return events;
 }
 
+//
+// server_task_result_embd
+//
+json server_task_result_embd::to_json() {
+    return res_type == TASK_RESPONSE_TYPE_OAI_EMBD
+        ? to_json_oaicompat()
+        : to_json_non_oaicompat();
+}
+
+json server_task_result_embd::to_json_non_oaicompat() {
+    return json {
+        {"index",     index},
+        {"embedding", embedding},
+    };
+}
+
+json server_task_result_embd::to_json_oaicompat() {
+    return json {
+        {"index",            index},
+        {"embedding",        embedding[0]},
+        {"tokens_evaluated", n_tokens},
+    };
+}
+
+//
+// server_task_result_rerank
+//
+json server_task_result_rerank::to_json() {
+    return json {
+        {"index",            index},
+        {"score",            score},
+        {"tokens_evaluated", n_tokens},
+    };
+}
+
 //
 // server_task_result_error
 //
diff --git a/tools/server/server-task.h b/tools/server/server-task.h
index 11943ee4f8..a69e8f1a3d 100644
--- a/tools/server/server-task.h
+++ b/tools/server/server-task.h
@@ -33,6 +33,7 @@ enum task_response_type {
     TASK_RESPONSE_TYPE_NONE, // llama.cpp native format
     TASK_RESPONSE_TYPE_OAI_CHAT,
     TASK_RESPONSE_TYPE_OAI_CMPL,
+    TASK_RESPONSE_TYPE_OAI_RESP,
     TASK_RESPONSE_TYPE_OAI_EMBD,
     TASK_RESPONSE_TYPE_ANTHROPIC,
 };
@@ -78,7 +79,9 @@ struct task_params {
     task_response_type res_type = TASK_RESPONSE_TYPE_NONE;
     std::string        oaicompat_model;
     std::string        oaicompat_cmpl_id;
-    common_chat_syntax oaicompat_chat_syntax;
+
+    // per-request parameters for chat parsing
+    common_chat_parser_params chat_parser_params;
 
     // Embeddings
     int32_t embd_normalize = 2; // (-1=none, 0=max absolute int16, 1=taxicab, 2=Euclidean/L2, >2=p-norm)
@@ -91,17 +94,27 @@ struct task_params {
 struct task_result_state {
     // tracking diffs for partial tool calls
     std::vector<common_chat_msg_diff> diffs;
-    common_chat_syntax oaicompat_chat_syntax;
+    common_chat_parser_params chat_parser_params;
     common_chat_msg chat_msg;
     std::string generated_text; // append new chunks of generated text here
     std::vector<std::string> generated_tool_call_ids;
 
-    // for Anthropic API streaming: track content block state across chunks
-    bool anthropic_thinking_block_started = false;
-    bool anthropic_text_block_started = false;
+    // for OpenAI Responses and Anthropic streaming API:
+    // track output item / content block state across chunks
+    bool thinking_block_started = false;
+    bool text_block_started = false;
 
-    task_result_state(const common_chat_syntax & oaicompat_chat_syntax)
-        : oaicompat_chat_syntax(oaicompat_chat_syntax) {}
+    // for OpenAI Responses streaming API
+    const std::string oai_resp_id;
+    const std::string oai_resp_reasoning_id;
+    const std::string oai_resp_message_id;
+    std::string oai_resp_fc_id; // function call ID for current args delta
+
+    task_result_state(const common_chat_parser_params & chat_parser_params)
+        : chat_parser_params(chat_parser_params)
+        , oai_resp_id("resp_" + random_string())
+        , oai_resp_reasoning_id("rs_" + random_string())
+        , oai_resp_message_id("msg_" + random_string()) {}
 
     // parse partial tool calls and update the internal state
     common_chat_msg update_chat_msg(
@@ -130,15 +143,17 @@ struct server_task {
     task_params   params;
     server_tokens tokens;
 
-    // only used by CLI, this delegates the tokenization to the server
-    json                    cli_input = nullptr;
+    // only used by CLI, this allow tokenizing CLI inputs on server side
+    // we need this because mtmd_context and vocab are not accessible outside of server_context
+    bool                    cli = false;
+    std::string             cli_prompt;
     std::vector<raw_buffer> cli_files;
 
     server_task_type type;
 
     // used by SERVER_TASK_TYPE_SLOT_SAVE, SERVER_TASK_TYPE_SLOT_RESTORE, SERVER_TASK_TYPE_SLOT_ERASE
     struct slot_action {
-        int slot_id;
+        int id_slot;
         std::string filename;
         std::string filepath;
     };
@@ -228,7 +243,7 @@ struct server_task {
     // the task will be moved into queue, then onto slots
     // however, the state must be kept by caller (e.g., HTTP thread)
     task_result_state create_state() const {
-        return task_result_state(params.oaicompat_chat_syntax);
+        return task_result_state(params.chat_parser_params);
     }
 
     bool is_parent() const {
@@ -348,6 +363,11 @@ struct server_task_result_cmpl_final : server_task_result {
     std::vector<common_chat_msg_diff> oaicompat_msg_diffs; // to be populated by update()
     bool is_updated = false;
 
+    // for OpenAI Responses API
+    std::string oai_resp_id;
+    std::string oai_resp_reasoning_id;
+    std::string oai_resp_message_id;
+
     virtual bool is_stop() override {
         return true; // in stream mode, final responses are considered stop
     }
@@ -357,6 +377,10 @@ struct server_task_result_cmpl_final : server_task_result {
     virtual void update(task_result_state & state) override {
         is_updated = true;
         oaicompat_msg = state.update_chat_msg(content, false, oaicompat_msg_diffs);
+
+        oai_resp_id = state.oai_resp_id;
+        oai_resp_reasoning_id = state.oai_resp_reasoning_id;
+        oai_resp_message_id = state.oai_resp_message_id;
     }
 
     json to_json_non_oaicompat();
@@ -367,6 +391,10 @@ struct server_task_result_cmpl_final : server_task_result {
 
     json to_json_oaicompat_chat_stream();
 
+    json to_json_oaicompat_resp();
+
+    json to_json_oaicompat_resp_stream();
+
     json to_json_anthropic();
 
     json to_json_anthropic_stream();
@@ -393,45 +421,35 @@ struct server_task_result_cmpl_partial : server_task_result {
     std::vector<common_chat_msg_diff> oaicompat_msg_diffs; // to be populated by update()
     bool is_updated = false;
 
+    // Streaming state copied from task_result_state for this chunk
+    bool thinking_block_started = false;
+    bool text_block_started     = false;
+
+    // for OpenAI Responses API
+    std::string oai_resp_id;
+    std::string oai_resp_reasoning_id;
+    std::string oai_resp_message_id;
+    std::string oai_resp_fc_id;
+
     // for Anthropic API: track if any reasoning content has been generated
     bool anthropic_has_reasoning = false;
-    // Streaming state copied from task_result_state for this chunk
-    bool anthropic_thinking_block_started = false;
-    bool anthropic_text_block_started = false;
 
     virtual bool is_stop() override {
         return false; // in stream mode, partial responses are not considered stop
     }
 
+    virtual void update(task_result_state & state) override;
+
     virtual json to_json() override;
 
-    virtual void update(task_result_state & state) override {
-        is_updated = true;
-        state.update_chat_msg(content, true, oaicompat_msg_diffs);
-        // track if the accumulated message has any reasoning content
-        anthropic_has_reasoning = !state.chat_msg.reasoning_content.empty();
-
-        // Copy current state for use in to_json_anthropic() (reflects state BEFORE this chunk)
-        anthropic_thinking_block_started = state.anthropic_thinking_block_started;
-        anthropic_text_block_started = state.anthropic_text_block_started;
-
-        // Pre-compute state updates based on diffs (for next chunk)
-        for (const auto & diff : oaicompat_msg_diffs) {
-            if (!diff.reasoning_content_delta.empty() && !state.anthropic_thinking_block_started) {
-                state.anthropic_thinking_block_started = true;
-            }
-            if (!diff.content_delta.empty() && !state.anthropic_text_block_started) {
-                state.anthropic_text_block_started = true;
-            }
-        }
-    }
-
     json to_json_non_oaicompat();
 
     json to_json_oaicompat();
 
     json to_json_oaicompat_chat();
 
+    json to_json_oaicompat_resp();
+
     json to_json_anthropic();
 };
 
diff --git a/tools/server/server.cpp b/tools/server/server.cpp
index 1d9abf6055..d3d4316026 100644
--- a/tools/server/server.cpp
+++ b/tools/server/server.cpp
@@ -140,6 +140,7 @@ int main(int argc, char ** argv) {
         routes.post_completions            = models_routes->proxy_post;
         routes.post_completions_oai        = models_routes->proxy_post;
         routes.post_chat_completions       = models_routes->proxy_post;
+        routes.post_responses_oai          = models_routes->proxy_post;
         routes.post_anthropic_messages     = models_routes->proxy_post;
         routes.post_anthropic_count_tokens = models_routes->proxy_post;
         routes.post_infill                 = models_routes->proxy_post;
@@ -176,6 +177,7 @@ int main(int argc, char ** argv) {
     ctx_http.post("/chat/completions",    ex_wrapper(routes.post_chat_completions));
     ctx_http.post("/v1/chat/completions", ex_wrapper(routes.post_chat_completions));
     ctx_http.post("/api/chat",            ex_wrapper(routes.post_chat_completions)); // ollama specific endpoint
+    ctx_http.post("/v1/responses",        ex_wrapper(routes.post_responses_oai));
     ctx_http.post("/v1/messages",         ex_wrapper(routes.post_anthropic_messages)); // anthropic messages API
     ctx_http.post("/v1/messages/count_tokens", ex_wrapper(routes.post_anthropic_count_tokens)); // anthropic token counting
     ctx_http.post("/infill",              ex_wrapper(routes.post_infill));
diff --git a/tools/server/tests/requirements.txt b/tools/server/tests/requirements.txt
index 4ea7f19f77..ca79d025ed 100644
--- a/tools/server/tests/requirements.txt
+++ b/tools/server/tests/requirements.txt
@@ -2,7 +2,7 @@ aiohttp~=3.9.3
 pytest~=8.3.3
 huggingface_hub>=0.34.0,<1.0
 numpy~=1.26.4
-openai~=1.55.3
+openai~=2.14.0
 prometheus-client~=0.20.0
 requests~=2.32.3
 wget~=3.2
diff --git a/tools/server/tests/unit/test_compat_oai_responses.py b/tools/server/tests/unit/test_compat_oai_responses.py
new file mode 100644
index 0000000000..7aab4a8ba6
--- /dev/null
+++ b/tools/server/tests/unit/test_compat_oai_responses.py
@@ -0,0 +1,73 @@
+import pytest
+from openai import OpenAI
+from utils import *
+
+server: ServerProcess
+
+@pytest.fixture(autouse=True)
+def create_server():
+    global server
+    server = ServerPreset.tinyllama2()
+
+def test_responses_with_openai_library():
+    global server
+    server.start()
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
+    res = client.responses.create(
+        model="gpt-4.1",
+        input=[
+            {"role": "system", "content": "Book"},
+            {"role": "user", "content": "What is the best book"},
+        ],
+        max_output_tokens=8,
+        temperature=0.8,
+    )
+    assert res.id.startswith("resp_")
+    assert res.output[0].id is not None
+    assert res.output[0].id.startswith("msg_")
+    assert match_regex("(Suddenly)+", res.output_text)
+
+def test_responses_stream_with_openai_library():
+    global server
+    server.start()
+    client = OpenAI(api_key="dummy", base_url=f"http://{server.server_host}:{server.server_port}/v1")
+    stream = client.responses.create(
+        model="gpt-4.1",
+        input=[
+            {"role": "system", "content": "Book"},
+            {"role": "user", "content": "What is the best book"},
+        ],
+        max_output_tokens=8,
+        temperature=0.8,
+        stream=True,
+    )
+
+    gathered_text = ''
+    resp_id = ''
+    msg_id = ''
+    for r in stream:
+        if r.type == "response.created":
+            assert r.response.id.startswith("resp_")
+            resp_id = r.response.id
+        if r.type == "response.in_progress":
+            assert r.response.id == resp_id
+        if r.type == "response.output_item.added":
+            assert r.item.id is not None
+            assert r.item.id.startswith("msg_")
+            msg_id = r.item.id
+        if (r.type == "response.content_part.added" or
+            r.type == "response.output_text.delta" or
+            r.type == "response.output_text.done" or
+            r.type == "response.content_part.done"):
+            assert r.item_id == msg_id
+        if r.type == "response.output_item.done":
+            assert r.item.id == msg_id
+
+        if r.type == "response.output_text.delta":
+            gathered_text += r.delta
+        if r.type == "response.completed":
+            assert r.response.id.startswith("resp_")
+            assert r.response.output[0].id is not None
+            assert r.response.output[0].id.startswith("msg_")
+            assert gathered_text == r.response.output_text
+            assert match_regex("(Suddenly)+", r.response.output_text)
diff --git a/tools/server/tests/unit/test_speculative.py b/tools/server/tests/unit/test_speculative.py
index d2f3fba5fe..eebd3cc8fa 100644
--- a/tools/server/tests/unit/test_speculative.py
+++ b/tools/server/tests/unit/test_speculative.py
@@ -30,6 +30,7 @@ def test_with_and_without_draft():
         "prompt": "I believe the meaning of life is",
         "temperature": 0.0,
         "top_k": 1,
+        "n_predict": 16,
     })
     assert res.status_code == 200
     content_no_draft = res.body["content"]
@@ -42,6 +43,7 @@ def test_with_and_without_draft():
         "prompt": "I believe the meaning of life is",
         "temperature": 0.0,
         "top_k": 1,
+        "n_predict": 16,
     })
     assert res.status_code == 200
     content_draft = res.body["content"]
@@ -68,6 +70,7 @@ def test_different_draft_min_draft_max():
             "prompt": "I believe the meaning of life is",
             "temperature": 0.0,
             "top_k": 1,
+            "n_predict": 16,
         })
         assert res.status_code == 200
         if last_content is not None:
diff --git a/tools/server/webui/package-lock.json b/tools/server/webui/package-lock.json
index 6fa9d39c71..6834416824 100644
--- a/tools/server/webui/package-lock.json
+++ b/tools/server/webui/package-lock.json
@@ -61,7 +61,7 @@
 				"remark-math": "^6.0.0",
 				"sass": "^1.93.3",
 				"storybook": "^10.0.7",
-				"svelte": "^5.0.0",
+				"svelte": "^5.38.2",
 				"svelte-check": "^4.0.0",
 				"tailwind-merge": "^3.3.1",
 				"tailwind-variants": "^3.2.2",
@@ -88,6 +88,7 @@
 			"version": "2.3.0",
 			"resolved": "https://registry.npmjs.org/@ampproject/remapping/-/remapping-2.3.0.tgz",
 			"integrity": "sha512-30iZtAPgz+LTIYoeivqYo853f02jBYSd5uGnGpkFV0M3xOt9aN73erkgYAmZU43x4VfqcnLxW9Kpg3R5LC4YYw==",
+			"dev": true,
 			"license": "Apache-2.0",
 			"dependencies": {
 				"@jridgewell/gen-mapping": "^0.3.5",
@@ -867,6 +868,7 @@
 			"integrity": "sha512-oJrXtQiAXLvT9clCf1K4kxp3eKsQhIaZqxEyowkBcsvZDdZkbWrVmnGknxs5flTD0VGsxrxKgBCZty1EzoiMzA==",
 			"dev": true,
 			"license": "Apache-2.0",
+			"peer": true,
 			"dependencies": {
 				"@swc/helpers": "^0.5.0"
 			}
@@ -898,7 +900,6 @@
 			"version": "2.3.5",
 			"resolved": "https://registry.npmjs.org/@jridgewell/remapping/-/remapping-2.3.5.tgz",
 			"integrity": "sha512-LI9u/+laYG4Ds1TDKSJW2YPrIlcVYOwi2fUC6xB43lueCjgxV4lffOCZCtYFiH6TNOX+tQKXx97T4IKHbhyHEQ==",
-			"dev": true,
 			"license": "MIT",
 			"dependencies": {
 				"@jridgewell/gen-mapping": "^0.3.5",
@@ -2031,6 +2032,7 @@
 			"integrity": "sha512-rO+YQhHucy47Vh67z318pALmd6x+K1Kj30Fb4a6oOEw4xn4zCo9KTmkMWs24c4oduEXD/eJu3badlRmsVXzyfA==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"ts-dedent": "^2.0.0",
 				"type-fest": "~2.19"
@@ -2114,6 +2116,7 @@
 			"integrity": "sha512-Vp3zX/qlwerQmHMP6x0Ry1oY7eKKRcOWGc2P59srOp4zcqyn+etJyQpELgOi4+ZSUgteX8Y387NuwruLgGXLUQ==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"@standard-schema/spec": "^1.0.0",
 				"@sveltejs/acorn-typescript": "^1.0.5",
@@ -2153,6 +2156,7 @@
 			"integrity": "sha512-YZs/OSKOQAQCnJvM/P+F1URotNnYNeU3P2s4oIpzm1uFaqUEqRxUB0g5ejMjEb5Gjb9/PiBI5Ktrq4rUUF8UVQ==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"@sveltejs/vite-plugin-svelte-inspector": "^5.0.0",
 				"debug": "^4.4.1",
@@ -2568,6 +2572,7 @@
 			"integrity": "sha512-pemlzrSESWbdAloYml3bAJMEfNh1Z7EduzqPKprCH5S341frlpYnUEW0H72dLxa6IsYr+mPno20GiSm+h9dEdQ==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"@babel/code-frame": "^7.10.4",
 				"@babel/runtime": "^7.12.5",
@@ -2735,6 +2740,7 @@
 			"integrity": "sha512-bJFoMATwIGaxxx8VJPeM8TonI8t579oRvgAuT8zFugJsJZgzqv0Fu8Mhp68iecjzG7cnN3mO2dJQ5uUM2EFrgQ==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"undici-types": "~6.21.0"
 			}
@@ -2802,6 +2808,7 @@
 			"integrity": "sha512-kVIaQE9vrN9RLCQMQ3iyRlVJpTiDUY6woHGb30JDkfJErqrQEmtdWH3gV0PBAfGZgQXoqzXOO0T3K6ioApbbAA==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"@typescript-eslint/scope-manager": "8.37.0",
 				"@typescript-eslint/types": "8.37.0",
@@ -3026,6 +3033,7 @@
 			"integrity": "sha512-tJxiPrWmzH8a+w9nLKlQMzAKX/7VjFs50MWgcAj7p9XQ7AQ9/35fByFYptgPELyLw+0aixTnC4pUWV+APcZ/kw==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"@testing-library/dom": "^10.4.0",
 				"@testing-library/user-event": "^14.6.1",
@@ -3129,6 +3137,7 @@
 			"integrity": "sha512-oukfKT9Mk41LreEW09vt45f8wx7DordoWUZMYdY/cyAk7w5TWkTRCNZYF7sX7n2wB7jyGAl74OxgwhPgKaqDMQ==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"@vitest/utils": "3.2.4",
 				"pathe": "^2.0.3",
@@ -3186,6 +3195,7 @@
 			"resolved": "https://registry.npmjs.org/acorn/-/acorn-8.15.0.tgz",
 			"integrity": "sha512-NZyJarBfL7nWwIq+FDL6Zp/yHEhePMNnnJ0y3qfieCrmNvYct8uvtiV41UvlSe6apAfk0fY1FbWx+NwfmpvtTg==",
 			"license": "MIT",
+			"peer": true,
 			"bin": {
 				"acorn": "bin/acorn"
 			},
@@ -3738,8 +3748,7 @@
 			"resolved": "https://registry.npmjs.org/csstype/-/csstype-3.1.3.tgz",
 			"integrity": "sha512-M1uQkMl8rQK/szD0LNhtqxIPLpimGm8sOBwU7lLnCpSbTyY3yeU1Vc7l4KT5zT4s/yOxHH5O7tIuuLOCnLADRw==",
 			"dev": true,
-			"license": "MIT",
-			"peer": true
+			"license": "MIT"
 		},
 		"node_modules/debug": {
 			"version": "4.4.1",
@@ -3840,10 +3849,9 @@
 			}
 		},
 		"node_modules/devalue": {
-			"version": "5.3.2",
-			"resolved": "https://registry.npmjs.org/devalue/-/devalue-5.3.2.tgz",
-			"integrity": "sha512-UDsjUbpQn9kvm68slnrs+mfxwFkIflOhkanmyabZ8zOYk8SMEIbJ3TK+88g70hSIeytu4y18f0z/hYHMTrXIWw==",
-			"dev": true,
+			"version": "5.6.2",
+			"resolved": "https://registry.npmjs.org/devalue/-/devalue-5.6.2.tgz",
+			"integrity": "sha512-nPRkjWzzDQlsejL1WVifk5rvcFi/y1onBRxjaFMjZeR9mFpqu2gmAZ9xUB9/IEanEP/vBtGeGganC/GO1fmufg==",
 			"license": "MIT"
 		},
 		"node_modules/devlop": {
@@ -3973,6 +3981,7 @@
 			"dev": true,
 			"hasInstallScript": true,
 			"license": "MIT",
+			"peer": true,
 			"bin": {
 				"esbuild": "bin/esbuild"
 			},
@@ -4027,6 +4036,7 @@
 			"integrity": "sha512-QldCVh/ztyKJJZLr4jXNUByx3gR+TDYZCRXEktiZoUR3PGy4qCmSbkxcIle8GEwGpb5JBZazlaJ/CxLidXdEbQ==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"@eslint-community/eslint-utils": "^4.2.0",
 				"@eslint-community/regexpp": "^4.12.1",
@@ -6939,6 +6949,7 @@
 				}
 			],
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"nanoid": "^3.3.11",
 				"picocolors": "^1.1.1",
@@ -7072,6 +7083,7 @@
 			"integrity": "sha512-I7AIg5boAr5R0FFtJ6rCfD+LFsWHp81dolrFD8S79U9tb8Az2nGrJncnMSnys+bpQJfRUzqs9hnA81OAA3hCuQ==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"bin": {
 				"prettier": "bin/prettier.cjs"
 			},
@@ -7088,6 +7100,7 @@
 			"integrity": "sha512-pn1ra/0mPObzqoIQn/vUTR3ZZI6UuZ0sHqMK5x2jMLGrs53h0sXhkVuDcrlssHwIMk7FYrMjHBPoUSyyEEDlBQ==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"peerDependencies": {
 				"prettier": "^3.0.0",
 				"svelte": "^3.2.0 || ^4.0.0-next.0 || ^5.0.0-next.0"
@@ -7312,6 +7325,7 @@
 			"integrity": "sha512-FS+XFBNvn3GTAWq26joslQgWNoFu08F4kl0J4CgdNKADkdSGXQyTCnKteIAJy96Br6YbpEU1LSzV5dYtjMkMDg==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"engines": {
 				"node": ">=0.10.0"
 			}
@@ -7322,6 +7336,7 @@
 			"integrity": "sha512-Xs1hdnE+DyKgeHJeJznQmYMIBG3TKIHJJT95Q58nHLSrElKlGQqDTR2HQ9fx5CN/Gk6Vh/kupBTDLU11/nDk/g==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"scheduler": "^0.26.0"
 			},
@@ -7598,6 +7613,7 @@
 			"integrity": "sha512-4iya7Jb76fVpQyLoiVpzUrsjQ12r3dM7fIVz+4NwoYvZOShknRmiv+iu9CClZml5ZLGb0XMcYLutK6w9tgxHDw==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"@types/estree": "1.0.8"
 			},
@@ -7704,6 +7720,7 @@
 			"integrity": "sha512-elOcIZRTM76dvxNAjqYrucTSI0teAF/L2Lv0s6f6b7FOwcwIuA357bIE871580AjHJuSvLIRUosgV+lIWx6Rgg==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"chokidar": "^4.0.0",
 				"immutable": "^5.0.2",
@@ -7938,6 +7955,7 @@
 			"integrity": "sha512-7smAu0o+kdm378Q2uIddk32pn0UdIbrtTVU+rXRVtTVTCrK/P2cCui2y4JH+Bl3NgEq1bbBQpCAF/HKrDjk2Qw==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"@storybook/global": "^5.0.0",
 				"@storybook/icons": "^1.6.0",
@@ -8079,12 +8097,13 @@
 			}
 		},
 		"node_modules/svelte": {
-			"version": "5.36.12",
-			"resolved": "https://registry.npmjs.org/svelte/-/svelte-5.36.12.tgz",
-			"integrity": "sha512-c3mWT+b0yBLl3gPGSHiy4pdSQCsPNTjLC0tVoOhrGJ6PPfCzD/RQpAmAfJtQZ304CAae2ph+L3C4aqds3R3seQ==",
+			"version": "5.48.3",
+			"resolved": "https://registry.npmjs.org/svelte/-/svelte-5.48.3.tgz",
+			"integrity": "sha512-w7QZ398cdNherTdiQ/v3SYLLGOO4948Jgjh04PYqtTYVohmBvbmFwLmo7pp8gp4/1tceRWfSTjHgjtfpCVNJmQ==",
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
-				"@ampproject/remapping": "^2.3.0",
+				"@jridgewell/remapping": "^2.3.4",
 				"@jridgewell/sourcemap-codec": "^1.5.0",
 				"@sveltejs/acorn-typescript": "^1.0.5",
 				"@types/estree": "^1.0.5",
@@ -8092,8 +8111,9 @@
 				"aria-query": "^5.3.1",
 				"axobject-query": "^4.1.0",
 				"clsx": "^2.1.1",
+				"devalue": "^5.6.2",
 				"esm-env": "^1.2.1",
-				"esrap": "^2.1.0",
+				"esrap": "^2.2.1",
 				"is-reference": "^3.0.3",
 				"locate-character": "^3.0.0",
 				"magic-string": "^0.30.11",
@@ -8281,9 +8301,9 @@
 			}
 		},
 		"node_modules/svelte/node_modules/esrap": {
-			"version": "2.1.0",
-			"resolved": "https://registry.npmjs.org/esrap/-/esrap-2.1.0.tgz",
-			"integrity": "sha512-yzmPNpl7TBbMRC5Lj2JlJZNPml0tzqoqP5B1JXycNUwtqma9AKCO0M2wHrdgsHcy1WRW7S9rJknAMtByg3usgA==",
+			"version": "2.2.2",
+			"resolved": "https://registry.npmjs.org/esrap/-/esrap-2.2.2.tgz",
+			"integrity": "sha512-zA6497ha+qKvoWIK+WM9NAh5ni17sKZKhbS5B3PoYbBvaYHZWoS33zmFybmyqpn07RLUxSmn+RCls2/XF+d0oQ==",
 			"license": "MIT",
 			"dependencies": {
 				"@jridgewell/sourcemap-codec": "^1.4.15"
@@ -8326,6 +8346,7 @@
 			"integrity": "sha512-gBXpgUm/3rp1lMZZrM/w7D8GKqshif0zAymAhbCyIt8KMe+0v9DQ7cdYLR4FHH/cKpdTXb+A/tKKU3eolfsI+g==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"funding": {
 				"type": "github",
 				"url": "https://github.com/sponsors/dcastil"
@@ -8356,7 +8377,8 @@
 			"resolved": "https://registry.npmjs.org/tailwindcss/-/tailwindcss-4.1.11.tgz",
 			"integrity": "sha512-2E9TBm6MDD/xKYe+dvJZAmg3yxIEDNRc0jwlNyDg/4Fil2QcSLjFKGVff0lAf1jjeaArlG/M75Ey/EYr/OJtBA==",
 			"dev": true,
-			"license": "MIT"
+			"license": "MIT",
+			"peer": true
 		},
 		"node_modules/tapable": {
 			"version": "2.2.2",
@@ -8569,6 +8591,7 @@
 			"integrity": "sha512-p1diW6TqL9L07nNxvRMM7hMMw4c5XOo/1ibL4aAIGmSAt9slTE1Xgw5KWuof2uTOvCg9BY7ZRi+GaF+7sfgPeQ==",
 			"dev": true,
 			"license": "Apache-2.0",
+			"peer": true,
 			"bin": {
 				"tsc": "bin/tsc",
 				"tsserver": "bin/tsserver"
@@ -8934,6 +8957,7 @@
 			"integrity": "sha512-BxAKBWmIbrDgrokdGZH1IgkIk/5mMHDreLDmCJ0qpyJaAteP8NvMhkwr/ZCQNqNH97bw/dANTE9PDzqwJghfMQ==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"esbuild": "^0.25.0",
 				"fdir": "^6.5.0",
@@ -9094,6 +9118,7 @@
 			"integrity": "sha512-LUCP5ev3GURDysTWiP47wRRUpLKMOfPh+yKTx3kVIEiu5KOMeqzpnYNsKyOoVrULivR8tLcks4+lga33Whn90A==",
 			"dev": true,
 			"license": "MIT",
+			"peer": true,
 			"dependencies": {
 				"@types/chai": "^5.2.2",
 				"@vitest/expect": "3.2.4",
diff --git a/tools/server/webui/package.json b/tools/server/webui/package.json
index 1a8c273749..a361ce76e3 100644
--- a/tools/server/webui/package.json
+++ b/tools/server/webui/package.json
@@ -62,7 +62,7 @@
 		"remark-math": "^6.0.0",
 		"sass": "^1.93.3",
 		"storybook": "^10.0.7",
-		"svelte": "^5.0.0",
+		"svelte": "^5.38.2",
 		"svelte-check": "^4.0.0",
 		"tailwind-merge": "^3.3.1",
 		"tailwind-variants": "^3.2.2",
diff --git a/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageAssistant.svelte b/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageAssistant.svelte
index c1ef4dfd0f..2b34b1c20a 100644
--- a/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageAssistant.svelte
+++ b/tools/server/webui/src/lib/components/app/chat/ChatMessages/ChatMessageAssistant.svelte
@@ -249,7 +249,7 @@
 		</div>
 	{/if}
 
-	<div class="info my-6 grid gap-4">
+	<div class="info my-6 grid gap-4 tabular-nums">
 		{#if displayedModel()}
 			<div class="inline-flex flex-wrap items-start gap-2 text-xs text-muted-foreground">
 				{#if isRouter}
diff --git a/vendor/cpp-httplib/CMakeLists.txt b/vendor/cpp-httplib/CMakeLists.txt
index 172b925453..3d938d9f36 100644
--- a/vendor/cpp-httplib/CMakeLists.txt
+++ b/vendor/cpp-httplib/CMakeLists.txt
@@ -142,7 +142,7 @@ elseif (LLAMA_OPENSSL)
             target_link_libraries(${TARGET} PUBLIC OpenSSL::SSL OpenSSL::Crypto)
         endif()
     else()
-        message(STATUS "OpenSSL not found, SSL support disabled")
+        message(WARNING "OpenSSL not found, HTTPS support disabled")
     endif()
 endif()