From 45d3bcd4340170ebee57bae83cdf224e0f9e110c Mon Sep 17 00:00:00 2001
From: Jianlin Shi <jianlins@users.noreply.github.com>
Date: Fri, 20 Dec 2024 22:37:03 -0700
Subject: [PATCH 01/19] 12.6.2

---
 .github/workflows/custom_build.yml | 177 +++++++++++++++++++++++++++++
 1 file changed, 177 insertions(+)
 create mode 100644 .github/workflows/custom_build.yml

diff --git a/.github/workflows/custom_build.yml b/.github/workflows/custom_build.yml
new file mode 100644
index 0000000000..02237991fc
--- /dev/null
+++ b/.github/workflows/custom_build.yml
@@ -0,0 +1,177 @@
+name: build_win_cuda_full
+
+on:
+  workflow_dispatch: # allows manual triggering
+    inputs:
+      create_release:
+        description: 'Create new release'
+        required: true
+        type: boolean
+      cuda:
+        description: 'cuda version used to build. For available versions, check https://github.com/Jimver/cuda-toolkit/blob/master/src/links/windows-links.ts'
+        required: false
+        type: string
+        default: '12.6.2'
+      gpu_arch:
+        description: 'nvidia gpu arch numbers'
+        required: false
+        type: string
+        default: '70;75;80;86'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+permissions:
+  contents: write 
+
+env:
+  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+
+jobs:
+  
+
+  windows-2019-cmake-cuda:
+    runs-on: windows-2019
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+            fetch-depth: 0
+      
+      - name: Get Python
+        uses: actions/setup-python@v2
+        with:
+          python-version: 3.8.10
+
+      - name: Install python dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install customtkinter==5.2.0 pyinstaller==5.11.0 psutil==5.9.5
+
+      - name: Download and install win64devkit
+        run: |
+          curl -L https://github.com/skeeto/w64devkit/releases/download/v1.22.0/w64devkit-1.22.0.zip --output w64devkit.zip
+          Expand-Archive w64devkit.zip -DestinationPath .
+
+      - name: Add w64devkit to PATH
+        run: |
+          echo "$(Get-Location)\w64devkit\bin" | Out-File -Append -FilePath $env:GITHUB_PATH -Encoding utf8
+
+      - name: Print System Environment Variables
+        id: printvars
+        run: |
+          echo "Number of processors: ${env:NUMBER_OF_PROCESSORS}"
+          echo "Processor Architecture: ${env:PROCESSOR_ARCHITECTURE}"
+          echo "Computer Name: ${env:COMPUTERNAME}"
+          wmic cpu get name
+          wmic os get TotalVisibleMemorySize, FreePhysicalMemory
+      
+      - uses: Jimver/cuda-toolkit@v0.2.19
+        id: cuda-toolkit
+        with:
+          cuda: ${{ github.event.inputs.cuda }}
+      
+
+      - name: Install ccache
+        uses: hendrikmuhs/ccache-action@v1.2
+        with:
+          key: ${{ github.job }}-cuda${{ github.event.inputs.cuda }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES="${{ github.event.inputs.gpu_arch }}" -DCMAKE_SYSTEM_VERSION="10.0.19041.0"
+          cmake --build . --config Release -j 2      
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}-${{ github.event.inputs.cuda }}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=WinFull-${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}-${{ github.event.inputs.cuda }}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip .\build\bin\Release\*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip
+          name: llama-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip
+
+      - name: Copy and pack Cuda runtime
+        run: |
+          echo "Cuda install location: ${{ env.CUDA_PATH }}"
+          $dst='.\build\bin\cudart\'
+          robocopy "${{env.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
+          robocopy "${{env.CUDA_PATH}}\lib" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
+          7z a cudart-llama-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip $dst\*
+
+      - name: Upload Cuda runtime
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          path: cudart-llama-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip
+          name: cudart-llama-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip
+
+      - name: Download artifacts
+        id: download-artifact
+        uses: actions/download-artifact@v4
+        with:
+          path: ./artifact
+
+      - name: Move artifacts
+        id: move_artifacts
+        run: mkdir -p ./artifact/release && mv ./artifact/*/*.zip ./artifact/release          
+
+      - name: Create release
+        id: create_release
+        if: ${{ github.event.inputs.create_release == 'true' }}
+        uses: anzz1/action-create-release@v1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          tag_name: ${{ steps.tag.outputs.name }}
+  
+      - name: Upload release
+        id: upload_release
+        if: ${{ github.event.inputs.create_release == 'true' }}
+        uses: actions/github-script@v3
+        with:
+          github-token: ${{secrets.GITHUB_TOKEN}}
+          script: |
+            const path = require('path');
+            const fs = require('fs');
+            const release_id = '${{ steps.create_release.outputs.id }}';
+            for (let file of await fs.readdirSync('./artifact/release')) {
+              if (path.extname(file) === '.zip') {
+                console.log('uploadReleaseAsset', file);
+                await github.repos.uploadReleaseAsset({
+                  owner: context.repo.owner,
+                  repo: context.repo.repo,
+                  release_id: release_id,
+                  name: file,
+                  data: await fs.readFileSync(`./artifact/release/${file}`)
+                });
+              }
+            }

From 59afb0160e39cfeb8c5c90a9a84d7a4cf45bb4a5 Mon Sep 17 00:00:00 2001
From: Jianlin Shi <jianlins@users.noreply.github.com>
Date: Fri, 20 Dec 2024 22:37:51 -0700
Subject: [PATCH 02/19] custom_build_wincuda

---
 .../workflows/{custom_build.yml => custom_build_wincuda.yml}    | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
 rename .github/workflows/{custom_build.yml => custom_build_wincuda.yml} (96%)

diff --git a/.github/workflows/custom_build.yml b/.github/workflows/custom_build_wincuda.yml
similarity index 96%
rename from .github/workflows/custom_build.yml
rename to .github/workflows/custom_build_wincuda.yml
index 02237991fc..069fa23787 100644
--- a/.github/workflows/custom_build.yml
+++ b/.github/workflows/custom_build_wincuda.yml
@@ -1,4 +1,4 @@
-name: build_win_cuda_full
+name: custom_build_wincuda
 
 on:
   workflow_dispatch: # allows manual triggering

From 9449305e47aaa96af249b10fc6bc341df2a765b7 Mon Sep 17 00:00:00 2001
From: Jianlin Shi <jianlins@users.noreply.github.com>
Date: Sun, 29 Dec 2024 16:08:46 -0700
Subject: [PATCH 03/19] trim trailing whitespaces

---
 .github/workflows/custom_build_wincuda.yml | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)

diff --git a/.github/workflows/custom_build_wincuda.yml b/.github/workflows/custom_build_wincuda.yml
index 069fa23787..27a9e7517f 100644
--- a/.github/workflows/custom_build_wincuda.yml
+++ b/.github/workflows/custom_build_wincuda.yml
@@ -23,7 +23,7 @@ concurrency:
   cancel-in-progress: true
 
 permissions:
-  contents: write 
+  contents: write
 
 env:
   BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
@@ -34,7 +34,7 @@ env:
   LLAMA_LOG_TIMESTAMPS: 1
 
 jobs:
-  
+
 
   windows-2019-cmake-cuda:
     runs-on: windows-2019
@@ -45,7 +45,7 @@ jobs:
         uses: actions/checkout@v4
         with:
             fetch-depth: 0
-      
+
       - name: Get Python
         uses: actions/setup-python@v2
         with:
@@ -73,12 +73,12 @@ jobs:
           echo "Computer Name: ${env:COMPUTERNAME}"
           wmic cpu get name
           wmic os get TotalVisibleMemorySize, FreePhysicalMemory
-      
+
       - uses: Jimver/cuda-toolkit@v0.2.19
         id: cuda-toolkit
         with:
           cuda: ${{ github.event.inputs.cuda }}
-      
+
 
       - name: Install ccache
         uses: hendrikmuhs/ccache-action@v1.2
@@ -91,7 +91,7 @@ jobs:
           mkdir build
           cd build
           cmake .. -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES="${{ github.event.inputs.gpu_arch }}" -DCMAKE_SYSTEM_VERSION="10.0.19041.0"
-          cmake --build . --config Release -j 2      
+          cmake --build . --config Release -j 2
 
       - name: Determine tag name
         id: tag
@@ -142,7 +142,7 @@ jobs:
 
       - name: Move artifacts
         id: move_artifacts
-        run: mkdir -p ./artifact/release && mv ./artifact/*/*.zip ./artifact/release          
+        run: mkdir -p ./artifact/release && mv ./artifact/*/*.zip ./artifact/release
 
       - name: Create release
         id: create_release
@@ -152,7 +152,7 @@ jobs:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         with:
           tag_name: ${{ steps.tag.outputs.name }}
-  
+
       - name: Upload release
         id: upload_release
         if: ${{ github.event.inputs.create_release == 'true' }}

From 632bc30e7542d906dbd51f3887ab87855b5bc143 Mon Sep 17 00:00:00 2001
From: Jianlin Shi <jianlins@users.noreply.github.com>
Date: Sun, 20 Apr 2025 14:00:06 -0600
Subject: [PATCH 04/19] test auto sync

---
 .github/workflows/syn_upstream.yml | 40 ++++++++++++++++++++++++++++++
 1 file changed, 40 insertions(+)
 create mode 100644 .github/workflows/syn_upstream.yml

diff --git a/.github/workflows/syn_upstream.yml b/.github/workflows/syn_upstream.yml
new file mode 100644
index 0000000000..5925f5cc38
--- /dev/null
+++ b/.github/workflows/syn_upstream.yml
@@ -0,0 +1,40 @@
+name: Sync Upstream
+
+on:
+  schedule:
+    - cron: '0 0 * * 0' # Every Sunday at midnight
+
+jobs:
+  merge:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout master branch
+        uses: actions/checkout@v2
+        with:
+          ref: master
+
+      - name: Merge upstream master
+        id: merge
+        run: |
+          git remote add upstream https://github.com/ggml-org/repository.git
+          git fetch upstream
+          if git merge upstream/master; then
+            echo "merge_success=true" >> $GITHUB_OUTPUT
+          else
+            echo "merge_success=false" >> $GITHUB_OUTPUT
+            echo "Automatic merge failed. Skipping push."
+            exit 0
+          fi
+
+      - name: Push changes
+        if: steps.merge.outputs.merge_success == 'true'
+        run: |
+          # Check if there are any changes
+          if git diff --quiet HEAD@{1} HEAD; then
+            echo "No changes to push"
+          else
+            echo "Changes detected, pushing updates..."
+            git push origin master
+          fi
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
\ No newline at end of file

From d251d891645cd3231d770fcbfe7943ce84fd9421 Mon Sep 17 00:00:00 2001
From: Jianlin Shi <jianlins@users.noreply.github.com>
Date: Wed, 27 Aug 2025 15:27:29 -0600
Subject: [PATCH 05/19] drop deprecated windows 2019 runner

---
 .github/workflows/custom_build_wincuda.yml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/custom_build_wincuda.yml b/.github/workflows/custom_build_wincuda.yml
index 27a9e7517f..30d88c87ba 100644
--- a/.github/workflows/custom_build_wincuda.yml
+++ b/.github/workflows/custom_build_wincuda.yml
@@ -36,8 +36,8 @@ env:
 jobs:
 
 
-  windows-2019-cmake-cuda:
-    runs-on: windows-2019
+  windows-latest-cmake-cuda:
+    runs-on: windows-latest
 
     steps:
       - name: Clone

From 2a1e77436c920973f79b9795dcb34ed55b13bf22 Mon Sep 17 00:00:00 2001
From: Jianlin Shi <jianlins@users.noreply.github.com>
Date: Wed, 11 Mar 2026 20:02:20 -0500
Subject: [PATCH 06/19] build-linux-cuda

---
 .github/workflows/custom_build_rockylinux.yml | 130 ++++++++++++++++++
 1 file changed, 130 insertions(+)
 create mode 100644 .github/workflows/custom_build_rockylinux.yml

diff --git a/.github/workflows/custom_build_rockylinux.yml b/.github/workflows/custom_build_rockylinux.yml
new file mode 100644
index 0000000000..8ba3861eb1
--- /dev/null
+++ b/.github/workflows/custom_build_rockylinux.yml
@@ -0,0 +1,130 @@
+name: build-linux-cuda
+
+on:
+  release:
+    types: [published]
+
+permissions:
+  contents: write
+
+jobs:
+  build-linux-cuda:
+    runs-on: ubuntu-latest
+
+    container:
+      image: nvidia/cuda:12.8.0-devel-rockylinux8
+
+    env:
+      BUILD_TYPE: Release
+      ARCHIVE_BASENAME: llama-bin-linux-cuda-12.8-rockylinux8
+
+    steps:
+      - name: Install build dependencies
+        shell: bash
+        run: |
+          set -euxo pipefail
+          yum -y update
+          yum -y install \
+            git \
+            cmake \
+            gcc \
+            gcc-c++ \
+            make \
+            ninja-build \
+            tar \
+            gzip \
+            zip \
+            findutils \
+            file \
+            which
+          yum clean all
+
+      - name: Checkout source at release tag
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event.release.tag_name }}
+          submodules: recursive
+          fetch-depth: 0
+
+      - name: Show build environment
+        shell: bash
+        run: |
+          set -euxo pipefail
+          cat /etc/os-release || true
+          uname -a
+          gcc --version
+          g++ --version
+          cmake --version
+          nvcc --version
+
+      - name: Configure
+        shell: bash
+        run: |
+          set -euxo pipefail
+          cmake -S . -B build \
+            -G Ninja \
+            -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
+            -DGGML_CUDA=ON \
+            -DCMAKE_CUDA_COMPILER=$(command -v nvcc)
+
+      - name: Build
+        shell: bash
+        run: |
+          set -euxo pipefail
+          cmake --build build --config ${BUILD_TYPE} -j"$(nproc)"
+
+      - name: Prepare release payload
+        shell: bash
+        run: |
+          set -euxo pipefail
+
+          TAG="${{ github.event.release.tag_name }}"
+          OUTDIR="${ARCHIVE_BASENAME}-${TAG}"
+          mkdir -p "${OUTDIR}/bin" "${OUTDIR}/lib" "${OUTDIR}/meta"
+
+          if [[ -d build/bin ]]; then
+            find build/bin -maxdepth 1 -type f \( -name 'llama-*' -o -name 'ggml-*' \) -exec cp -av {} "${OUTDIR}/bin/" \;
+          fi
+
+          find build -maxdepth 3 -type f \( -name 'libllama*' -o -name 'libggml*' \) -exec cp -av {} "${OUTDIR}/lib/" \; || true
+
+          git rev-parse HEAD > "${OUTDIR}/meta/git-commit.txt"
+          git describe --tags --always > "${OUTDIR}/meta/git-describe.txt" || true
+          cat /etc/os-release > "${OUTDIR}/meta/os-release.txt" || true
+          gcc --version > "${OUTDIR}/meta/gcc-version.txt"
+          g++ --version > "${OUTDIR}/meta/gxx-version.txt"
+          cmake --version > "${OUTDIR}/meta/cmake-version.txt"
+          nvcc --version > "${OUTDIR}/meta/nvcc-version.txt"
+
+          if compgen -G "${OUTDIR}/bin/*" > /dev/null; then
+            file "${OUTDIR}"/bin/* > "${OUTDIR}/meta/file-bin.txt" || true
+            ldd "${OUTDIR}"/bin/* > "${OUTDIR}/meta/ldd-bin.txt" || true
+          fi
+
+          if compgen -G "${OUTDIR}/lib/*" > /dev/null; then
+            file "${OUTDIR}"/lib/* > "${OUTDIR}/meta/file-lib.txt" || true
+            ldd "${OUTDIR}"/lib/* > "${OUTDIR}/meta/ldd-lib.txt" || true
+          fi
+
+          tar -czf "${OUTDIR}.tar.gz" "${OUTDIR}"
+          sha256sum "${OUTDIR}.tar.gz" > "${OUTDIR}.tar.gz.sha256"
+
+      - name: Upload tarball to release
+        uses: actions/upload-release-asset@v1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          upload_url: ${{ github.event.release.upload_url }}
+          asset_path: ./llama-bin-linux-cuda-12.8-rockylinux8-${{ github.event.release.tag_name }}.tar.gz
+          asset_name: llama-bin-linux-cuda-12.8-rockylinux8-${{ github.event.release.tag_name }}.tar.gz
+          asset_content_type: application/gzip
+
+      - name: Upload checksum to release
+        uses: actions/upload-release-asset@v1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          upload_url: ${{ github.event.release.upload_url }}
+          asset_path: ./llama-bin-linux-cuda-12.8-rockylinux8-${{ github.event.release.tag_name }}.tar.gz.sha256
+          asset_name: llama-bin-linux-cuda-12.8-rockylinux8-${{ github.event.release.tag_name }}.tar.gz.sha256
+          asset_content_type: text/plain

From 29b3c14619df80335edd68b5902aecb2fd0e9f04 Mon Sep 17 00:00:00 2001
From: jianlins <jianlinshi.cn@gmail.com>
Date: Wed, 11 Mar 2026 21:44:36 -0600
Subject: [PATCH 07/19] workflow: update build configurations for Linux and
 Windows CUDA

---
 .github/workflows/custom_build_rockylinux.yml | 260 ++++++-------
 .github/workflows/custom_build_wincuda.yml    | 354 +++++++++---------
 .github/workflows/syn_upstream.yml            |  80 ++--
 3 files changed, 347 insertions(+), 347 deletions(-)

diff --git a/.github/workflows/custom_build_rockylinux.yml b/.github/workflows/custom_build_rockylinux.yml
index 8ba3861eb1..a025f141ce 100644
--- a/.github/workflows/custom_build_rockylinux.yml
+++ b/.github/workflows/custom_build_rockylinux.yml
@@ -1,130 +1,130 @@
-name: build-linux-cuda
-
-on:
-  release:
-    types: [published]
-
-permissions:
-  contents: write
-
-jobs:
-  build-linux-cuda:
-    runs-on: ubuntu-latest
-
-    container:
-      image: nvidia/cuda:12.8.0-devel-rockylinux8
-
-    env:
-      BUILD_TYPE: Release
-      ARCHIVE_BASENAME: llama-bin-linux-cuda-12.8-rockylinux8
-
-    steps:
-      - name: Install build dependencies
-        shell: bash
-        run: |
-          set -euxo pipefail
-          yum -y update
-          yum -y install \
-            git \
-            cmake \
-            gcc \
-            gcc-c++ \
-            make \
-            ninja-build \
-            tar \
-            gzip \
-            zip \
-            findutils \
-            file \
-            which
-          yum clean all
-
-      - name: Checkout source at release tag
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ github.event.release.tag_name }}
-          submodules: recursive
-          fetch-depth: 0
-
-      - name: Show build environment
-        shell: bash
-        run: |
-          set -euxo pipefail
-          cat /etc/os-release || true
-          uname -a
-          gcc --version
-          g++ --version
-          cmake --version
-          nvcc --version
-
-      - name: Configure
-        shell: bash
-        run: |
-          set -euxo pipefail
-          cmake -S . -B build \
-            -G Ninja \
-            -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
-            -DGGML_CUDA=ON \
-            -DCMAKE_CUDA_COMPILER=$(command -v nvcc)
-
-      - name: Build
-        shell: bash
-        run: |
-          set -euxo pipefail
-          cmake --build build --config ${BUILD_TYPE} -j"$(nproc)"
-
-      - name: Prepare release payload
-        shell: bash
-        run: |
-          set -euxo pipefail
-
-          TAG="${{ github.event.release.tag_name }}"
-          OUTDIR="${ARCHIVE_BASENAME}-${TAG}"
-          mkdir -p "${OUTDIR}/bin" "${OUTDIR}/lib" "${OUTDIR}/meta"
-
-          if [[ -d build/bin ]]; then
-            find build/bin -maxdepth 1 -type f \( -name 'llama-*' -o -name 'ggml-*' \) -exec cp -av {} "${OUTDIR}/bin/" \;
-          fi
-
-          find build -maxdepth 3 -type f \( -name 'libllama*' -o -name 'libggml*' \) -exec cp -av {} "${OUTDIR}/lib/" \; || true
-
-          git rev-parse HEAD > "${OUTDIR}/meta/git-commit.txt"
-          git describe --tags --always > "${OUTDIR}/meta/git-describe.txt" || true
-          cat /etc/os-release > "${OUTDIR}/meta/os-release.txt" || true
-          gcc --version > "${OUTDIR}/meta/gcc-version.txt"
-          g++ --version > "${OUTDIR}/meta/gxx-version.txt"
-          cmake --version > "${OUTDIR}/meta/cmake-version.txt"
-          nvcc --version > "${OUTDIR}/meta/nvcc-version.txt"
-
-          if compgen -G "${OUTDIR}/bin/*" > /dev/null; then
-            file "${OUTDIR}"/bin/* > "${OUTDIR}/meta/file-bin.txt" || true
-            ldd "${OUTDIR}"/bin/* > "${OUTDIR}/meta/ldd-bin.txt" || true
-          fi
-
-          if compgen -G "${OUTDIR}/lib/*" > /dev/null; then
-            file "${OUTDIR}"/lib/* > "${OUTDIR}/meta/file-lib.txt" || true
-            ldd "${OUTDIR}"/lib/* > "${OUTDIR}/meta/ldd-lib.txt" || true
-          fi
-
-          tar -czf "${OUTDIR}.tar.gz" "${OUTDIR}"
-          sha256sum "${OUTDIR}.tar.gz" > "${OUTDIR}.tar.gz.sha256"
-
-      - name: Upload tarball to release
-        uses: actions/upload-release-asset@v1
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        with:
-          upload_url: ${{ github.event.release.upload_url }}
-          asset_path: ./llama-bin-linux-cuda-12.8-rockylinux8-${{ github.event.release.tag_name }}.tar.gz
-          asset_name: llama-bin-linux-cuda-12.8-rockylinux8-${{ github.event.release.tag_name }}.tar.gz
-          asset_content_type: application/gzip
-
-      - name: Upload checksum to release
-        uses: actions/upload-release-asset@v1
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        with:
-          upload_url: ${{ github.event.release.upload_url }}
-          asset_path: ./llama-bin-linux-cuda-12.8-rockylinux8-${{ github.event.release.tag_name }}.tar.gz.sha256
-          asset_name: llama-bin-linux-cuda-12.8-rockylinux8-${{ github.event.release.tag_name }}.tar.gz.sha256
-          asset_content_type: text/plain
+name: build-linux-cuda
+
+on:
+  release:
+    types: [published]
+
+permissions:
+  contents: write
+
+jobs:
+  build-linux-cuda:
+    runs-on: ubuntu-latest
+
+    container:
+      image: nvidia/cuda:12.8.0-devel-rockylinux8
+
+    env:
+      BUILD_TYPE: Release
+      ARCHIVE_BASENAME: llama-bin-linux-cuda-12.8-rockylinux8
+
+    steps:
+      - name: Install build dependencies
+        shell: bash
+        run: |
+          set -euxo pipefail
+          yum -y update
+          yum -y install \
+            git \
+            cmake \
+            gcc \
+            gcc-c++ \
+            make \
+            ninja-build \
+            tar \
+            gzip \
+            zip \
+            findutils \
+            file \
+            which
+          yum clean all
+
+      - name: Checkout source at release tag
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event.release.tag_name }}
+          submodules: recursive
+          fetch-depth: 0
+
+      - name: Show build environment
+        shell: bash
+        run: |
+          set -euxo pipefail
+          cat /etc/os-release || true
+          uname -a
+          gcc --version
+          g++ --version
+          cmake --version
+          nvcc --version
+
+      - name: Configure
+        shell: bash
+        run: |
+          set -euxo pipefail
+          cmake -S . -B build \
+            -G Ninja \
+            -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
+            -DGGML_CUDA=ON \
+            -DCMAKE_CUDA_COMPILER=$(command -v nvcc)
+
+      - name: Build
+        shell: bash
+        run: |
+          set -euxo pipefail
+          cmake --build build --config ${BUILD_TYPE} -j"$(nproc)"
+
+      - name: Prepare release payload
+        shell: bash
+        run: |
+          set -euxo pipefail
+
+          TAG="${{ github.event.release.tag_name }}"
+          OUTDIR="${ARCHIVE_BASENAME}-${TAG}"
+          mkdir -p "${OUTDIR}/bin" "${OUTDIR}/lib" "${OUTDIR}/meta"
+
+          if [[ -d build/bin ]]; then
+            find build/bin -maxdepth 1 -type f \( -name 'llama-*' -o -name 'ggml-*' \) -exec cp -av {} "${OUTDIR}/bin/" \;
+          fi
+
+          find build -maxdepth 3 -type f \( -name 'libllama*' -o -name 'libggml*' \) -exec cp -av {} "${OUTDIR}/lib/" \; || true
+
+          git rev-parse HEAD > "${OUTDIR}/meta/git-commit.txt"
+          git describe --tags --always > "${OUTDIR}/meta/git-describe.txt" || true
+          cat /etc/os-release > "${OUTDIR}/meta/os-release.txt" || true
+          gcc --version > "${OUTDIR}/meta/gcc-version.txt"
+          g++ --version > "${OUTDIR}/meta/gxx-version.txt"
+          cmake --version > "${OUTDIR}/meta/cmake-version.txt"
+          nvcc --version > "${OUTDIR}/meta/nvcc-version.txt"
+
+          if compgen -G "${OUTDIR}/bin/*" > /dev/null; then
+            file "${OUTDIR}"/bin/* > "${OUTDIR}/meta/file-bin.txt" || true
+            ldd "${OUTDIR}"/bin/* > "${OUTDIR}/meta/ldd-bin.txt" || true
+          fi
+
+          if compgen -G "${OUTDIR}/lib/*" > /dev/null; then
+            file "${OUTDIR}"/lib/* > "${OUTDIR}/meta/file-lib.txt" || true
+            ldd "${OUTDIR}"/lib/* > "${OUTDIR}/meta/ldd-lib.txt" || true
+          fi
+
+          tar -czf "${OUTDIR}.tar.gz" "${OUTDIR}"
+          sha256sum "${OUTDIR}.tar.gz" > "${OUTDIR}.tar.gz.sha256"
+
+      - name: Upload tarball to release
+        uses: actions/upload-release-asset@v1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          upload_url: ${{ github.event.release.upload_url }}
+          asset_path: ./llama-bin-linux-cuda-12.8-rockylinux8-${{ github.event.release.tag_name }}.tar.gz
+          asset_name: llama-bin-linux-cuda-12.8-rockylinux8-${{ github.event.release.tag_name }}.tar.gz
+          asset_content_type: application/gzip
+
+      - name: Upload checksum to release
+        uses: actions/upload-release-asset@v1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          upload_url: ${{ github.event.release.upload_url }}
+          asset_path: ./llama-bin-linux-cuda-12.8-rockylinux8-${{ github.event.release.tag_name }}.tar.gz.sha256
+          asset_name: llama-bin-linux-cuda-12.8-rockylinux8-${{ github.event.release.tag_name }}.tar.gz.sha256
+          asset_content_type: text/plain
diff --git a/.github/workflows/custom_build_wincuda.yml b/.github/workflows/custom_build_wincuda.yml
index 30d88c87ba..9791726399 100644
--- a/.github/workflows/custom_build_wincuda.yml
+++ b/.github/workflows/custom_build_wincuda.yml
@@ -1,177 +1,177 @@
-name: custom_build_wincuda
-
-on:
-  workflow_dispatch: # allows manual triggering
-    inputs:
-      create_release:
-        description: 'Create new release'
-        required: true
-        type: boolean
-      cuda:
-        description: 'cuda version used to build. For available versions, check https://github.com/Jimver/cuda-toolkit/blob/master/src/links/windows-links.ts'
-        required: false
-        type: string
-        default: '12.6.2'
-      gpu_arch:
-        description: 'nvidia gpu arch numbers'
-        required: false
-        type: string
-        default: '70;75;80;86'
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-permissions:
-  contents: write
-
-env:
-  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
-  GGML_NLOOP: 3
-  GGML_N_THREADS: 1
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
-
-jobs:
-
-
-  windows-latest-cmake-cuda:
-    runs-on: windows-latest
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-        with:
-            fetch-depth: 0
-
-      - name: Get Python
-        uses: actions/setup-python@v2
-        with:
-          python-version: 3.8.10
-
-      - name: Install python dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install customtkinter==5.2.0 pyinstaller==5.11.0 psutil==5.9.5
-
-      - name: Download and install win64devkit
-        run: |
-          curl -L https://github.com/skeeto/w64devkit/releases/download/v1.22.0/w64devkit-1.22.0.zip --output w64devkit.zip
-          Expand-Archive w64devkit.zip -DestinationPath .
-
-      - name: Add w64devkit to PATH
-        run: |
-          echo "$(Get-Location)\w64devkit\bin" | Out-File -Append -FilePath $env:GITHUB_PATH -Encoding utf8
-
-      - name: Print System Environment Variables
-        id: printvars
-        run: |
-          echo "Number of processors: ${env:NUMBER_OF_PROCESSORS}"
-          echo "Processor Architecture: ${env:PROCESSOR_ARCHITECTURE}"
-          echo "Computer Name: ${env:COMPUTERNAME}"
-          wmic cpu get name
-          wmic os get TotalVisibleMemorySize, FreePhysicalMemory
-
-      - uses: Jimver/cuda-toolkit@v0.2.19
-        id: cuda-toolkit
-        with:
-          cuda: ${{ github.event.inputs.cuda }}
-
-
-      - name: Install ccache
-        uses: hendrikmuhs/ccache-action@v1.2
-        with:
-          key: ${{ github.job }}-cuda${{ github.event.inputs.cuda }}
-
-      - name: Build
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake .. -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES="${{ github.event.inputs.gpu_arch }}" -DCMAKE_SYSTEM_VERSION="10.0.19041.0"
-          cmake --build . --config Release -j 2
-
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}-${{ github.event.inputs.cuda }}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=WinFull-${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}-${{ github.event.inputs.cuda }}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip .\build\bin\Release\*
-
-      - name: Upload artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip
-          name: llama-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip
-
-      - name: Copy and pack Cuda runtime
-        run: |
-          echo "Cuda install location: ${{ env.CUDA_PATH }}"
-          $dst='.\build\bin\cudart\'
-          robocopy "${{env.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          robocopy "${{env.CUDA_PATH}}\lib" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
-          7z a cudart-llama-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip $dst\*
-
-      - name: Upload Cuda runtime
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v4
-        with:
-          path: cudart-llama-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip
-          name: cudart-llama-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip
-
-      - name: Download artifacts
-        id: download-artifact
-        uses: actions/download-artifact@v4
-        with:
-          path: ./artifact
-
-      - name: Move artifacts
-        id: move_artifacts
-        run: mkdir -p ./artifact/release && mv ./artifact/*/*.zip ./artifact/release
-
-      - name: Create release
-        id: create_release
-        if: ${{ github.event.inputs.create_release == 'true' }}
-        uses: anzz1/action-create-release@v1
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        with:
-          tag_name: ${{ steps.tag.outputs.name }}
-
-      - name: Upload release
-        id: upload_release
-        if: ${{ github.event.inputs.create_release == 'true' }}
-        uses: actions/github-script@v3
-        with:
-          github-token: ${{secrets.GITHUB_TOKEN}}
-          script: |
-            const path = require('path');
-            const fs = require('fs');
-            const release_id = '${{ steps.create_release.outputs.id }}';
-            for (let file of await fs.readdirSync('./artifact/release')) {
-              if (path.extname(file) === '.zip') {
-                console.log('uploadReleaseAsset', file);
-                await github.repos.uploadReleaseAsset({
-                  owner: context.repo.owner,
-                  repo: context.repo.repo,
-                  release_id: release_id,
-                  name: file,
-                  data: await fs.readFileSync(`./artifact/release/${file}`)
-                });
-              }
-            }
+name: custom_build_wincuda
+
+on:
+  workflow_dispatch: # allows manual triggering
+    inputs:
+      create_release:
+        description: 'Create new release'
+        required: true
+        type: boolean
+      cuda:
+        description: 'cuda version used to build. For available versions, check https://github.com/Jimver/cuda-toolkit/blob/master/src/links/windows-links.ts'
+        required: false
+        type: string
+        default: '12.6.2'
+      gpu_arch:
+        description: 'nvidia gpu arch numbers'
+        required: false
+        type: string
+        default: '70;75;80;86'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+permissions:
+  contents: write
+
+env:
+  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+
+jobs:
+
+
+  windows-latest-cmake-cuda:
+    runs-on: windows-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+            fetch-depth: 0
+
+      - name: Get Python
+        uses: actions/setup-python@v2
+        with:
+          python-version: 3.8.10
+
+      - name: Install python dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install customtkinter==5.2.0 pyinstaller==5.11.0 psutil==5.9.5
+
+      - name: Download and install win64devkit
+        run: |
+          curl -L https://github.com/skeeto/w64devkit/releases/download/v1.22.0/w64devkit-1.22.0.zip --output w64devkit.zip
+          Expand-Archive w64devkit.zip -DestinationPath .
+
+      - name: Add w64devkit to PATH
+        run: |
+          echo "$(Get-Location)\w64devkit\bin" | Out-File -Append -FilePath $env:GITHUB_PATH -Encoding utf8
+
+      - name: Print System Environment Variables
+        id: printvars
+        run: |
+          echo "Number of processors: ${env:NUMBER_OF_PROCESSORS}"
+          echo "Processor Architecture: ${env:PROCESSOR_ARCHITECTURE}"
+          echo "Computer Name: ${env:COMPUTERNAME}"
+          wmic cpu get name
+          wmic os get TotalVisibleMemorySize, FreePhysicalMemory
+
+      - uses: Jimver/cuda-toolkit@v0.2.19
+        id: cuda-toolkit
+        with:
+          cuda: ${{ github.event.inputs.cuda }}
+
+
+      - name: Install ccache
+        uses: hendrikmuhs/ccache-action@v1.2
+        with:
+          key: ${{ github.job }}-cuda${{ github.event.inputs.cuda }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES="${{ github.event.inputs.gpu_arch }}" -DCMAKE_SYSTEM_VERSION="10.0.19041.0"
+          cmake --build . --config Release -j 2
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}-${{ github.event.inputs.cuda }}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=WinFull-${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}-${{ github.event.inputs.cuda }}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip .\build\bin\Release\*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip
+          name: llama-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip
+
+      - name: Copy and pack Cuda runtime
+        run: |
+          echo "Cuda install location: ${{ env.CUDA_PATH }}"
+          $dst='.\build\bin\cudart\'
+          robocopy "${{env.CUDA_PATH}}\bin" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
+          robocopy "${{env.CUDA_PATH}}\lib" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll
+          7z a cudart-llama-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip $dst\*
+
+      - name: Upload Cuda runtime
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          path: cudart-llama-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip
+          name: cudart-llama-bin-win-cu${{ github.event.inputs.cuda }}-x64.zip
+
+      - name: Download artifacts
+        id: download-artifact
+        uses: actions/download-artifact@v4
+        with:
+          path: ./artifact
+
+      - name: Move artifacts
+        id: move_artifacts
+        run: mkdir -p ./artifact/release && mv ./artifact/*/*.zip ./artifact/release
+
+      - name: Create release
+        id: create_release
+        if: ${{ github.event.inputs.create_release == 'true' }}
+        uses: anzz1/action-create-release@v1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          tag_name: ${{ steps.tag.outputs.name }}
+
+      - name: Upload release
+        id: upload_release
+        if: ${{ github.event.inputs.create_release == 'true' }}
+        uses: actions/github-script@v3
+        with:
+          github-token: ${{secrets.GITHUB_TOKEN}}
+          script: |
+            const path = require('path');
+            const fs = require('fs');
+            const release_id = '${{ steps.create_release.outputs.id }}';
+            for (let file of await fs.readdirSync('./artifact/release')) {
+              if (path.extname(file) === '.zip') {
+                console.log('uploadReleaseAsset', file);
+                await github.repos.uploadReleaseAsset({
+                  owner: context.repo.owner,
+                  repo: context.repo.repo,
+                  release_id: release_id,
+                  name: file,
+                  data: await fs.readFileSync(`./artifact/release/${file}`)
+                });
+              }
+            }
diff --git a/.github/workflows/syn_upstream.yml b/.github/workflows/syn_upstream.yml
index 5925f5cc38..683e6d0a0f 100644
--- a/.github/workflows/syn_upstream.yml
+++ b/.github/workflows/syn_upstream.yml
@@ -1,40 +1,40 @@
-name: Sync Upstream
-
-on:
-  schedule:
-    - cron: '0 0 * * 0' # Every Sunday at midnight
-
-jobs:
-  merge:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout master branch
-        uses: actions/checkout@v2
-        with:
-          ref: master
-
-      - name: Merge upstream master
-        id: merge
-        run: |
-          git remote add upstream https://github.com/ggml-org/repository.git
-          git fetch upstream
-          if git merge upstream/master; then
-            echo "merge_success=true" >> $GITHUB_OUTPUT
-          else
-            echo "merge_success=false" >> $GITHUB_OUTPUT
-            echo "Automatic merge failed. Skipping push."
-            exit 0
-          fi
-
-      - name: Push changes
-        if: steps.merge.outputs.merge_success == 'true'
-        run: |
-          # Check if there are any changes
-          if git diff --quiet HEAD@{1} HEAD; then
-            echo "No changes to push"
-          else
-            echo "Changes detected, pushing updates..."
-            git push origin master
-          fi
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
\ No newline at end of file
+name: Sync Upstream
+
+on:
+  schedule:
+    - cron: '0 0 * * 0' # Every Sunday at midnight
+
+jobs:
+  merge:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout master branch
+        uses: actions/checkout@v2
+        with:
+          ref: master
+
+      - name: Merge upstream master
+        id: merge
+        run: |
+          git remote add upstream https://github.com/ggml-org/repository.git
+          git fetch upstream
+          if git merge upstream/master; then
+            echo "merge_success=true" >> $GITHUB_OUTPUT
+          else
+            echo "merge_success=false" >> $GITHUB_OUTPUT
+            echo "Automatic merge failed. Skipping push."
+            exit 0
+          fi
+
+      - name: Push changes
+        if: steps.merge.outputs.merge_success == 'true'
+        run: |
+          # Check if there are any changes
+          if git diff --quiet HEAD@{1} HEAD; then
+            echo "No changes to push"
+          else
+            echo "Changes detected, pushing updates..."
+            git push origin master
+          fi
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

From a3cc7bf99247a3eac06b9853dd6baa2d4a9628cf Mon Sep 17 00:00:00 2001
From: jianlins <jianlinshi.cn@gmail.com>
Date: Wed, 11 Mar 2026 21:46:17 -0600
Subject: [PATCH 08/19] workflow: add manual trigger for release creation in
 Rocky Linux build

---
 .github/workflows/custom_build_rockylinux.yml | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/.github/workflows/custom_build_rockylinux.yml b/.github/workflows/custom_build_rockylinux.yml
index a025f141ce..e1549b51d2 100644
--- a/.github/workflows/custom_build_rockylinux.yml
+++ b/.github/workflows/custom_build_rockylinux.yml
@@ -1,6 +1,12 @@
 name: build-linux-cuda
 
 on:
+  workflow_dispatch: # allows manual triggering
+    inputs:
+      create_release:
+        description: 'Create new release'
+        required: true
+        type: boolean
   release:
     types: [published]
 

From 38c82a6db77cea411f252452f3a522002e408242 Mon Sep 17 00:00:00 2001
From: jianlins <jianlinshi.cn@gmail.com>
Date: Wed, 11 Mar 2026 21:55:44 -0600
Subject: [PATCH 09/19] workflow: install epel-release as a build dependency
 for Rocky Linux

---
 .github/workflows/custom_build_rockylinux.yml | 1 +
 1 file changed, 1 insertion(+)

diff --git a/.github/workflows/custom_build_rockylinux.yml b/.github/workflows/custom_build_rockylinux.yml
index e1549b51d2..d638205b7d 100644
--- a/.github/workflows/custom_build_rockylinux.yml
+++ b/.github/workflows/custom_build_rockylinux.yml
@@ -30,6 +30,7 @@ jobs:
         run: |
           set -euxo pipefail
           yum -y update
+          yum -y install epel-release
           yum -y install \
             git \
             cmake \

From ddb3fcd924d69b3cf50e26d37d0be1e844f8e207 Mon Sep 17 00:00:00 2001
From: jianlins <jianlinshi.cn@gmail.com>
Date: Wed, 11 Mar 2026 22:01:02 -0600
Subject: [PATCH 10/19] workflow: comment out yum update in Rocky Linux build
 script

---
 .github/workflows/custom_build_rockylinux.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/custom_build_rockylinux.yml b/.github/workflows/custom_build_rockylinux.yml
index d638205b7d..0d8f34b583 100644
--- a/.github/workflows/custom_build_rockylinux.yml
+++ b/.github/workflows/custom_build_rockylinux.yml
@@ -29,7 +29,7 @@ jobs:
         shell: bash
         run: |
           set -euxo pipefail
-          yum -y update
+          # yum -y update
           yum -y install epel-release
           yum -y install \
             git \

From 65f59acdde54b34566c4788883e5a2044d8afcd1 Mon Sep 17 00:00:00 2001
From: jianlins <jianlinshi.cn@gmail.com>
Date: Wed, 11 Mar 2026 22:05:15 -0600
Subject: [PATCH 11/19] workflow: update package manager commands from yum to
 dnf in Rocky Linux build script

---
 .github/workflows/custom_build_rockylinux.yml | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/.github/workflows/custom_build_rockylinux.yml b/.github/workflows/custom_build_rockylinux.yml
index 0d8f34b583..02ef65e514 100644
--- a/.github/workflows/custom_build_rockylinux.yml
+++ b/.github/workflows/custom_build_rockylinux.yml
@@ -29,9 +29,11 @@ jobs:
         shell: bash
         run: |
           set -euxo pipefail
-          # yum -y update
-          yum -y install epel-release
-          yum -y install \
+          # dnf -y update
+          dnf -y install epel-release
+          dnf config-manager --set-enabled powertools
+          dnf config-manager --set-enabled crb
+          dnf -y install \
             git \
             cmake \
             gcc \
@@ -44,7 +46,7 @@ jobs:
             findutils \
             file \
             which
-          yum clean all
+          dnf clean all
 
       - name: Checkout source at release tag
         uses: actions/checkout@v4

From 200905df28f4a90063207d7ac9966fdc9db0da77 Mon Sep 17 00:00:00 2001
From: jianlins <jianlinshi.cn@gmail.com>
Date: Wed, 11 Mar 2026 22:10:13 -0600
Subject: [PATCH 12/19] workflow: remove crb repository enablement from Rocky
 Linux build script

---
 .github/workflows/custom_build_rockylinux.yml | 1 -
 1 file changed, 1 deletion(-)

diff --git a/.github/workflows/custom_build_rockylinux.yml b/.github/workflows/custom_build_rockylinux.yml
index 02ef65e514..809ee85ba6 100644
--- a/.github/workflows/custom_build_rockylinux.yml
+++ b/.github/workflows/custom_build_rockylinux.yml
@@ -32,7 +32,6 @@ jobs:
           # dnf -y update
           dnf -y install epel-release
           dnf config-manager --set-enabled powertools
-          dnf config-manager --set-enabled crb
           dnf -y install \
             git \
             cmake \

From b1e28fa5110891a6d25a8b90e07b9d8684b6e561 Mon Sep 17 00:00:00 2001
From: jianlins <jianlinshi.cn@gmail.com>
Date: Wed, 11 Mar 2026 22:28:21 -0600
Subject: [PATCH 13/19] workflow: add gpu_arch input for customizable CUDA
 architecture in build script

---
 .github/workflows/custom_build_rockylinux.yml | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/.github/workflows/custom_build_rockylinux.yml b/.github/workflows/custom_build_rockylinux.yml
index 809ee85ba6..5a82b28cee 100644
--- a/.github/workflows/custom_build_rockylinux.yml
+++ b/.github/workflows/custom_build_rockylinux.yml
@@ -7,6 +7,11 @@ on:
         description: 'Create new release'
         required: true
         type: boolean
+      gpu_arch:
+        description: 'nvidia gpu arch numbers (semicolon-separated, e.g. 70;75;80;86)'
+        required: false
+        type: string
+        default: '70;75;80;86;89'
   release:
     types: [published]
 
@@ -73,7 +78,8 @@ jobs:
             -G Ninja \
             -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
             -DGGML_CUDA=ON \
-            -DCMAKE_CUDA_COMPILER=$(command -v nvcc)
+            -DCMAKE_CUDA_COMPILER=$(command -v nvcc) \
+            -DCMAKE_CUDA_ARCHITECTURES="${{ github.event.inputs.gpu_arch }}"
 
       - name: Build
         shell: bash

From 460a53595658fe48c1549e51909eaeb43c6e6012 Mon Sep 17 00:00:00 2001
From: jianlins <jianlinshi.cn@gmail.com>
Date: Thu, 12 Mar 2026 00:01:18 -0600
Subject: [PATCH 14/19] workflow: update build script to use gcc-toolset-12 and
 add linker flags

---
 .github/workflows/custom_build_rockylinux.yml | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/.github/workflows/custom_build_rockylinux.yml b/.github/workflows/custom_build_rockylinux.yml
index 5a82b28cee..dd7332d9fc 100644
--- a/.github/workflows/custom_build_rockylinux.yml
+++ b/.github/workflows/custom_build_rockylinux.yml
@@ -40,8 +40,7 @@ jobs:
           dnf -y install \
             git \
             cmake \
-            gcc \
-            gcc-c++ \
+            gcc-toolset-12 \
             make \
             ninja-build \
             tar \
@@ -63,6 +62,7 @@ jobs:
         shell: bash
         run: |
           set -euxo pipefail
+          source scl_source enable gcc-toolset-12
           cat /etc/os-release || true
           uname -a
           gcc --version
@@ -74,17 +74,20 @@ jobs:
         shell: bash
         run: |
           set -euxo pipefail
+          source scl_source enable gcc-toolset-12
           cmake -S . -B build \
             -G Ninja \
             -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
             -DGGML_CUDA=ON \
             -DCMAKE_CUDA_COMPILER=$(command -v nvcc) \
-            -DCMAKE_CUDA_ARCHITECTURES="${{ github.event.inputs.gpu_arch }}"
+            -DCMAKE_CUDA_ARCHITECTURES="${{ github.event.inputs.gpu_arch }}" \
+            -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined
 
       - name: Build
         shell: bash
         run: |
           set -euxo pipefail
+          source scl_source enable gcc-toolset-12
           cmake --build build --config ${BUILD_TYPE} -j"$(nproc)"
 
       - name: Prepare release payload

From eb7a304948d31cc06342e53cc99d1678403f42ee Mon Sep 17 00:00:00 2001
From: jianlins <jianlinshi.cn@gmail.com>
Date: Thu, 12 Mar 2026 00:08:04 -0600
Subject: [PATCH 15/19] workflow: update build steps to handle gcc-toolset-12
 sourcing more safely

---
 .github/workflows/custom_build_rockylinux.yml | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/.github/workflows/custom_build_rockylinux.yml b/.github/workflows/custom_build_rockylinux.yml
index dd7332d9fc..9a1b895198 100644
--- a/.github/workflows/custom_build_rockylinux.yml
+++ b/.github/workflows/custom_build_rockylinux.yml
@@ -62,7 +62,7 @@ jobs:
         shell: bash
         run: |
           set -euxo pipefail
-          source scl_source enable gcc-toolset-12
+          set +u; source scl_source enable gcc-toolset-12; set -u
           cat /etc/os-release || true
           uname -a
           gcc --version
@@ -74,7 +74,7 @@ jobs:
         shell: bash
         run: |
           set -euxo pipefail
-          source scl_source enable gcc-toolset-12
+          set +u; source scl_source enable gcc-toolset-12; set -u
           cmake -S . -B build \
             -G Ninja \
             -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
@@ -87,7 +87,7 @@ jobs:
         shell: bash
         run: |
           set -euxo pipefail
-          source scl_source enable gcc-toolset-12
+          set +u; source scl_source enable gcc-toolset-12; set -u
           cmake --build build --config ${BUILD_TYPE} -j"$(nproc)"
 
       - name: Prepare release payload

From a23f3fb7fdd6c7c5c8bc3e52a3becac48be322b2 Mon Sep 17 00:00:00 2001
From: jianlins <jianlinshi.cn@gmail.com>
Date: Thu, 12 Mar 2026 20:12:53 -0600
Subject: [PATCH 16/19] workflow: add safe.directory configuration for Git in
 build script

---
 .github/workflows/custom_build_rockylinux.yml | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/.github/workflows/custom_build_rockylinux.yml b/.github/workflows/custom_build_rockylinux.yml
index 9a1b895198..3231aaacc0 100644
--- a/.github/workflows/custom_build_rockylinux.yml
+++ b/.github/workflows/custom_build_rockylinux.yml
@@ -105,6 +105,8 @@ jobs:
 
           find build -maxdepth 3 -type f \( -name 'libllama*' -o -name 'libggml*' \) -exec cp -av {} "${OUTDIR}/lib/" \; || true
 
+          git config --global --add safe.directory "${GITHUB_WORKSPACE:-/__w/llama.cpp/llama.cpp}"
+
           git rev-parse HEAD > "${OUTDIR}/meta/git-commit.txt"
           git describe --tags --always > "${OUTDIR}/meta/git-describe.txt" || true
           cat /etc/os-release > "${OUTDIR}/meta/os-release.txt" || true

From 25581cbb35e9a0dd6859e2a8d278abc2cdf51d08 Mon Sep 17 00:00:00 2001
From: jianlins <jianlinshi.cn@gmail.com>
Date: Thu, 12 Mar 2026 22:05:23 -0600
Subject: [PATCH 17/19] workflow: require tag_name input for manual release
 trigger and update asset upload process

---
 .github/workflows/custom_build_rockylinux.yml | 34 ++++++++-----------
 1 file changed, 14 insertions(+), 20 deletions(-)

diff --git a/.github/workflows/custom_build_rockylinux.yml b/.github/workflows/custom_build_rockylinux.yml
index 3231aaacc0..3b45c5b444 100644
--- a/.github/workflows/custom_build_rockylinux.yml
+++ b/.github/workflows/custom_build_rockylinux.yml
@@ -7,6 +7,10 @@ on:
         description: 'Create new release'
         required: true
         type: boolean
+      tag_name:
+        description: 'Tag name for the release (required for manual trigger)'
+        required: true
+        type: string
       gpu_arch:
         description: 'nvidia gpu arch numbers (semicolon-separated, e.g. 70;75;80;86)'
         required: false
@@ -28,6 +32,7 @@ jobs:
     env:
       BUILD_TYPE: Release
       ARCHIVE_BASENAME: llama-bin-linux-cuda-12.8-rockylinux8
+      TAG_NAME: ${{ github.event.release.tag_name || github.event.inputs.tag_name }}
 
     steps:
       - name: Install build dependencies
@@ -54,7 +59,7 @@ jobs:
       - name: Checkout source at release tag
         uses: actions/checkout@v4
         with:
-          ref: ${{ github.event.release.tag_name }}
+          ref: ${{ env.TAG_NAME }}
           submodules: recursive
           fetch-depth: 0
 
@@ -95,7 +100,7 @@ jobs:
         run: |
           set -euxo pipefail
 
-          TAG="${{ github.event.release.tag_name }}"
+          TAG="${TAG_NAME}"
           OUTDIR="${ARCHIVE_BASENAME}-${TAG}"
           mkdir -p "${OUTDIR}/bin" "${OUTDIR}/lib" "${OUTDIR}/meta"
 
@@ -128,22 +133,11 @@ jobs:
           tar -czf "${OUTDIR}.tar.gz" "${OUTDIR}"
           sha256sum "${OUTDIR}.tar.gz" > "${OUTDIR}.tar.gz.sha256"
 
-      - name: Upload tarball to release
-        uses: actions/upload-release-asset@v1
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+      - name: Upload release assets
+        if: github.event_name == 'release' || github.event.inputs.create_release == 'true'
+        uses: softprops/action-gh-release@v2
         with:
-          upload_url: ${{ github.event.release.upload_url }}
-          asset_path: ./llama-bin-linux-cuda-12.8-rockylinux8-${{ github.event.release.tag_name }}.tar.gz
-          asset_name: llama-bin-linux-cuda-12.8-rockylinux8-${{ github.event.release.tag_name }}.tar.gz
-          asset_content_type: application/gzip
-
-      - name: Upload checksum to release
-        uses: actions/upload-release-asset@v1
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        with:
-          upload_url: ${{ github.event.release.upload_url }}
-          asset_path: ./llama-bin-linux-cuda-12.8-rockylinux8-${{ github.event.release.tag_name }}.tar.gz.sha256
-          asset_name: llama-bin-linux-cuda-12.8-rockylinux8-${{ github.event.release.tag_name }}.tar.gz.sha256
-          asset_content_type: text/plain
+          tag_name: ${{ env.TAG_NAME }}
+          files: |
+            ./${{ env.ARCHIVE_BASENAME }}-${{ env.TAG_NAME }}.tar.gz
+            ./${{ env.ARCHIVE_BASENAME }}-${{ env.TAG_NAME }}.tar.gz.sha256

From cd9d8771c2b743de97d672edbbf5587acbcbaba3 Mon Sep 17 00:00:00 2001
From: jianlins <jianlinshi.cn@gmail.com>
Date: Thu, 12 Mar 2026 23:02:38 -0600
Subject: [PATCH 18/19] workflow: update source checkout step and add tag
 creation logic in build script

---
 .github/workflows/custom_build_rockylinux.yml | 19 ++++++++++++++++---
 1 file changed, 16 insertions(+), 3 deletions(-)

diff --git a/.github/workflows/custom_build_rockylinux.yml b/.github/workflows/custom_build_rockylinux.yml
index 3b45c5b444..6a53f84dc0 100644
--- a/.github/workflows/custom_build_rockylinux.yml
+++ b/.github/workflows/custom_build_rockylinux.yml
@@ -56,13 +56,26 @@ jobs:
             which
           dnf clean all
 
-      - name: Checkout source at release tag
+      - name: Checkout source
         uses: actions/checkout@v4
         with:
-          ref: ${{ env.TAG_NAME }}
           submodules: recursive
           fetch-depth: 0
 
+      - name: Create tag if it does not exist
+        shell: bash
+        run: |
+          set -euxo pipefail
+          git config --global --add safe.directory "${GITHUB_WORKSPACE:-/__w/llama.cpp/llama.cpp}"
+          if git rev-parse "refs/tags/${TAG_NAME}" >/dev/null 2>&1; then
+            echo "Tag ${TAG_NAME} already exists, checking out."
+            git checkout "${TAG_NAME}"
+          else
+            echo "Tag ${TAG_NAME} does not exist, creating it on HEAD."
+            git tag "${TAG_NAME}"
+            git push origin "refs/tags/${TAG_NAME}"
+          fi
+
       - name: Show build environment
         shell: bash
         run: |
@@ -140,4 +153,4 @@ jobs:
           tag_name: ${{ env.TAG_NAME }}
           files: |
             ./${{ env.ARCHIVE_BASENAME }}-${{ env.TAG_NAME }}.tar.gz
-            ./${{ env.ARCHIVE_BASENAME }}-${{ env.TAG_NAME }}.tar.gz.sha256
+            ./${{ env.ARCHIVE_BASENAME }}-${{ env.TAG_NAME }}.tar.gz.sha256
\ No newline at end of file

From 48c561fcc9af58a1eaf24f52b39df3acced4db15 Mon Sep 17 00:00:00 2001
From: jianlins <jianlinshi.cn@gmail.com>
Date: Fri, 13 Mar 2026 11:57:51 -0600
Subject: [PATCH 19/19] workflow: enhance library packaging by preserving
 symlinks and adding runtime checks

---
 .github/workflows/custom_build_rockylinux.yml | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/custom_build_rockylinux.yml b/.github/workflows/custom_build_rockylinux.yml
index 6a53f84dc0..1abc520c3b 100644
--- a/.github/workflows/custom_build_rockylinux.yml
+++ b/.github/workflows/custom_build_rockylinux.yml
@@ -121,7 +121,12 @@ jobs:
             find build/bin -maxdepth 1 -type f \( -name 'llama-*' -o -name 'ggml-*' \) -exec cp -av {} "${OUTDIR}/bin/" \;
           fi
 
-          find build -maxdepth 3 -type f \( -name 'libllama*' -o -name 'libggml*' \) -exec cp -av {} "${OUTDIR}/lib/" \; || true
+          # Keep both files and symlinks so SONAME links (for example libmtmd.so.0) survive packaging.
+          find build -maxdepth 3 \( -type f -o -type l \) \( -name 'libllama*' -o -name 'libggml*' -o -name 'libmtmd*' \) -exec cp -av {} "${OUTDIR}/lib/" \; || true
+
+          # Fail fast if required runtime library links are missing from the payload.
+          test -e "${OUTDIR}/lib/libmtmd.so.0"
+          test -e "${OUTDIR}/lib/libllama.so.0"
 
           git config --global --add safe.directory "${GITHUB_WORKSPACE:-/__w/llama.cpp/llama.cpp}"
 
@@ -153,4 +158,4 @@ jobs:
           tag_name: ${{ env.TAG_NAME }}
           files: |
             ./${{ env.ARCHIVE_BASENAME }}-${{ env.TAG_NAME }}.tar.gz
-            ./${{ env.ARCHIVE_BASENAME }}-${{ env.TAG_NAME }}.tar.gz.sha256
\ No newline at end of file
+            ./${{ env.ARCHIVE_BASENAME }}-${{ env.TAG_NAME }}.tar.gz.sha256