diff --git a/.devops/cpu.Dockerfile b/.devops/cpu.Dockerfile index aa2aa03120..9459f08c10 100644 --- a/.devops/cpu.Dockerfile +++ b/.devops/cpu.Dockerfile @@ -14,9 +14,9 @@ WORKDIR /app COPY . . RUN if [ "$TARGETARCH" = "amd64" ]; then \ - cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \ + cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \ elif [ "$TARGETARCH" = "arm64" ]; then \ - cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \ + cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \ else \ echo "Unsupported architecture"; \ exit 1; \ diff --git a/.devops/cuda.Dockerfile b/.devops/cuda.Dockerfile index 8ae57d2e28..94f1433972 100644 --- a/.devops/cuda.Dockerfile +++ b/.devops/cuda.Dockerfile @@ -21,7 +21,7 @@ COPY . . RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \ export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \ fi && \ - cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \ + cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_BUILD_TESTS=OFF ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \ cmake --build build --config Release -j$(nproc) RUN mkdir -p /app/lib && \ diff --git a/.devops/intel.Dockerfile b/.devops/intel.Dockerfile index 091e1dc5d8..c8839fe027 100644 --- a/.devops/intel.Dockerfile +++ b/.devops/intel.Dockerfile @@ -17,7 +17,7 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \ && export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \ fi && \ echo "Building with dynamic libs" && \ - cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${OPT_SYCL_F16} && \ + cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_BUILD_TESTS=OFF ${OPT_SYCL_F16} && \ cmake --build build --config Release -j$(nproc) RUN mkdir -p /app/lib && \ diff --git a/.devops/llama-cli-cann.Dockerfile b/.devops/llama-cli-cann.Dockerfile index 0eb1af87cb..ef43d78cd2 100644 --- a/.devops/llama-cli-cann.Dockerfile +++ b/.devops/llama-cli-cann.Dockerfile @@ -22,7 +22,7 @@ ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH RUN echo "Building with static libs" && \ source /usr/local/Ascend/ascend-toolkit/set_env.sh --force && \ - cmake -B build -DGGML_NATIVE=OFF -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF && \ + cmake -B build -DGGML_NATIVE=OFF -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF -DLLAMA_BUILD_TESTS=OFF && \ cmake --build build --config Release --target llama-cli # TODO: use image with NNRT diff --git a/.devops/musa.Dockerfile b/.devops/musa.Dockerfile index 261a2823a0..e0f1ad9728 100644 --- a/.devops/musa.Dockerfile +++ b/.devops/musa.Dockerfile @@ -35,7 +35,7 @@ COPY . . RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \ export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \ fi && \ - cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \ + cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \ cmake --build build --config Release -j$(nproc) RUN mkdir -p /app/lib && \ diff --git a/.devops/rocm.Dockerfile b/.devops/rocm.Dockerfile index a1b34723a4..1c00f1b9c2 100644 --- a/.devops/rocm.Dockerfile +++ b/.devops/rocm.Dockerfile @@ -40,7 +40,7 @@ WORKDIR /app COPY . . RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \ - cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON \ + cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \ && cmake --build build --config Release -j$(nproc) RUN mkdir -p /app/lib \ diff --git a/.devops/vulkan.Dockerfile b/.devops/vulkan.Dockerfile index f8f3072e95..fcd81ffa1e 100644 --- a/.devops/vulkan.Dockerfile +++ b/.devops/vulkan.Dockerfile @@ -16,7 +16,7 @@ WORKDIR /app COPY . . -RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \ +RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \ cmake --build build --config Release -j$(nproc) RUN mkdir -p /app/lib && \ diff --git a/.editorconfig b/.editorconfig index 5d63d0a51e..1eadda334a 100644 --- a/.editorconfig +++ b/.editorconfig @@ -21,15 +21,15 @@ indent_style = tab [prompts/*.txt] insert_final_newline = unset -[examples/server/public/*] +[tools/server/public/*] indent_size = 2 -[examples/server/public/deps_*] +[tools/server/public/deps_*] trim_trailing_whitespace = unset indent_style = unset indent_size = unset -[examples/server/deps_*] +[tools/server/deps_*] trim_trailing_whitespace = unset indent_style = unset indent_size = unset @@ -37,7 +37,7 @@ indent_size = unset [examples/llama.swiftui/llama.swiftui.xcodeproj/*] indent_style = tab -[examples/cvector-generator/*.txt] +[tools/cvector-generator/*.txt] trim_trailing_whitespace = unset insert_final_newline = unset diff --git a/.flake8 b/.flake8 index d64c2564ac..669d231f1f 100644 --- a/.flake8 +++ b/.flake8 @@ -2,8 +2,9 @@ max-line-length = 125 ignore = E203,E211,E221,E225,E231,E241,E251,E261,E266,E501,E701,E704,W503 exclude = - # Do not traverse examples + # Do not traverse examples and tools examples, + tools, # Do not include package initializers __init__.py, # No need to traverse our git directory diff --git a/.github/actions/get-tag-name/action.yml b/.github/actions/get-tag-name/action.yml new file mode 100644 index 0000000000..7ace23b2a3 --- /dev/null +++ b/.github/actions/get-tag-name/action.yml @@ -0,0 +1,22 @@ +name: "Determine tag name" +description: "Determine the tag name to use for a release" +outputs: + name: + description: "The name of the tag" + value: ${{ steps.tag.outputs.name }} + +runs: + using: "composite" + steps: + - name: Determine tag name + id: tag + shell: bash + run: | + BUILD_NUMBER="$(git rev-list --count HEAD)" + SHORT_HASH="$(git rev-parse --short=7 HEAD)" + if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then + echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT + else + SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') + echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT + fi diff --git a/.github/actions/windows-setup-cuda/action.yml b/.github/actions/windows-setup-cuda/action.yml new file mode 100644 index 0000000000..5575caeca3 --- /dev/null +++ b/.github/actions/windows-setup-cuda/action.yml @@ -0,0 +1,67 @@ +name: "Windows - Setup CUDA Toolkit" +description: "Setup CUDA Toolkit for Windows" +inputs: + cuda_version: + description: "CUDA toolkit version" + required: true + +runs: + using: "composite" + steps: + - name: Install Cuda Toolkit 11.7 + if: ${{ inputs.cuda_version == '11.7' }} + shell: pwsh + run: | + mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" + choco install unzip -y + curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-11.7.99-archive.zip" + curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-11.7.99-archive.zip" + curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-11.7.99-archive.zip" + curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-11.7.4.6-archive.zip" + curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-11.7.91-archive.zip" + curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-11.7.91-archive.zip" + curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-11.7.101-archive.zip" + curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-11.7.91-archive.zip" + unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" + xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_cudart-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y + xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvcc-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y + xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvrtc-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y + xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\libcublas-windows-x86_64-11.7.4.6-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y + xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvtx-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y + xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\visual_studio_integration-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y + xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvprof-windows-x86_64-11.7.101-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y + xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_cccl-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y + echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append + echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append + echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8 + echo "CUDA_PATH_V11_7=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8 + + - name: Install Cuda Toolkit 12.4 + if: ${{ inputs.cuda_version == '12.4' }} + shell: pwsh + run: | + mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" + choco install unzip -y + curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-12.4.127-archive.zip" + curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-12.4.131-archive.zip" + curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-12.4.127-archive.zip" + curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-12.4.5.8-archive.zip" + curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-12.4.127-archive.zip" + curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_profiler_api/windows-x86_64/cuda_profiler_api-windows-x86_64-12.4.127-archive.zip" + curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-12.4.127-archive.zip" + curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-12.4.127-archive.zip" + curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-12.4.127-archive.zip" + unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" + xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_cudart-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y + xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvcc-windows-x86_64-12.4.131-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y + xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvrtc-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y + xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libcublas-windows-x86_64-12.4.5.8-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y + xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvtx-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y + xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_profiler_api-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y + xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\visual_studio_integration-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y + xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvprof-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y + xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_cccl-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y + echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append + echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append + echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8 + echo "CUDA_PATH_V12_4=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8 diff --git a/.github/labeler.yml b/.github/labeler.yml index 1b47bc9688..278032ef2e 100644 --- a/.github/labeler.yml +++ b/.github/labeler.yml @@ -45,7 +45,9 @@ build: - CMakePresets.json examples: - changed-files: - - any-glob-to-any-file: examples/** + - any-glob-to-any-file: + - examples/** + - tools/** devops: - changed-files: - any-glob-to-any-file: @@ -70,7 +72,7 @@ android: server: - changed-files: - any-glob-to-any-file: - - examples/server/** + - tools/server/** ggml: - changed-files: - any-glob-to-any-file: diff --git a/.github/workflows/bench.yml.disabled b/.github/workflows/bench.yml.disabled index 75d2714792..f2d7e16e98 100644 --- a/.github/workflows/bench.yml.disabled +++ b/.github/workflows/bench.yml.disabled @@ -27,10 +27,10 @@ on: push: branches: - master - paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp'] + paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'tools/server/*.h*', 'tools/server/*.cpp'] pull_request_target: types: [opened, synchronize, reopened] - paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp'] + paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'tools/server/*.h*', 'tools/server/*.cpp'] schedule: - cron: '04 2 * * *' @@ -69,7 +69,7 @@ jobs: - name: Install python env id: pipenv run: | - cd examples/server/bench + cd tools/server/bench python3 -m venv venv source venv/bin/activate pip install -r requirements.txt @@ -79,7 +79,7 @@ jobs: run: | wget --quiet https://github.com/prometheus/prometheus/releases/download/v2.51.0/prometheus-2.51.0.linux-amd64.tar.gz tar xzf prometheus*.tar.gz --strip-components=1 - ./prometheus --config.file=examples/server/bench/prometheus.yml & + ./prometheus --config.file=tools/server/bench/prometheus.yml & while ! nc -z localhost 9090; do sleep 0.1 done @@ -92,7 +92,7 @@ jobs: - name: Install k6 and xk6-sse id: k6_installation run: | - cd examples/server/bench + cd tools/server/bench go install go.k6.io/xk6/cmd/xk6@latest xk6 build master \ --with github.com/phymbert/xk6-sse @@ -116,7 +116,7 @@ jobs: - name: Download the dataset id: download_dataset run: | - cd examples/server/bench + cd tools/server/bench wget --quiet https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json - name: Server bench @@ -126,7 +126,7 @@ jobs: run: | set -eux - cd examples/server/bench + cd tools/server/bench source venv/bin/activate python bench.py \ --runner-label ${{ env.RUNNER_LABEL }} \ @@ -157,9 +157,9 @@ jobs: name: bench-server-${{ github.job }}-${{ env.RUNNER_LABEL }}-${{ matrix.model }}-${{ matrix.ftype }} compression-level: 9 path: | - examples/server/bench/*.jpg - examples/server/bench/*.json - examples/server/bench/*.log + tools/server/bench/*.jpg + tools/server/bench/*.json + tools/server/bench/*.log - name: Commit status uses: Sibz/github-status-action@v1 @@ -178,17 +178,17 @@ jobs: with: client_id: ${{secrets.IMGUR_CLIENT_ID}} path: | - examples/server/bench/prompt_tokens_seconds.jpg - examples/server/bench/predicted_tokens_seconds.jpg - examples/server/bench/kv_cache_usage_ratio.jpg - examples/server/bench/requests_processing.jpg + tools/server/bench/prompt_tokens_seconds.jpg + tools/server/bench/predicted_tokens_seconds.jpg + tools/server/bench/kv_cache_usage_ratio.jpg + tools/server/bench/requests_processing.jpg - name: Extract mermaid id: set_mermaid run: | set -eux - cd examples/server/bench + cd tools/server/bench PROMPT_TOKENS_SECONDS=$(cat prompt_tokens_seconds.mermaid) echo "PROMPT_TOKENS_SECONDS<> $GITHUB_ENV echo "$PROMPT_TOKENS_SECONDS" >> $GITHUB_ENV diff --git a/.github/workflows/build-linux-cross.yml b/.github/workflows/build-linux-cross.yml index e8639913ea..1c38d7e11d 100644 --- a/.github/workflows/build-linux-cross.yml +++ b/.github/workflows/build-linux-cross.yml @@ -4,18 +4,25 @@ on: workflow_call: jobs: - ubuntu-latest-riscv64-cpu-cross: - runs-on: ubuntu-latest + ubuntu-24-riscv64-cpu-cross: + runs-on: ubuntu-24.04 steps: - uses: actions/checkout@v4 - name: Setup Riscv run: | sudo dpkg --add-architecture riscv64 - sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \ - /etc/apt/sources.list /etc/apt/apt-mirrors.txt - sudo apt-get clean - sudo apt-get update + + # Add arch-specific repositories for non-amd64 architectures + cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list + deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe + deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe + deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe + deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe + EOF + + sudo apt-get update || true ;# Prevent failure due to missing URLs. + sudo apt-get install -y --no-install-recommends \ build-essential \ gcc-14-riscv64-linux-gnu \ @@ -27,6 +34,7 @@ jobs: cmake -B build -DCMAKE_BUILD_TYPE=Release \ -DGGML_OPENMP=OFF \ -DLLAMA_BUILD_EXAMPLES=ON \ + -DLLAMA_BUILD_TOOLS=ON \ -DLLAMA_BUILD_TESTS=OFF \ -DCMAKE_SYSTEM_NAME=Linux \ -DCMAKE_SYSTEM_PROCESSOR=riscv64 \ @@ -40,21 +48,25 @@ jobs: cmake --build build --config Release -j $(nproc) - ubuntu-latest-riscv64-vulkan-cross: - runs-on: ubuntu-latest + ubuntu-24-riscv64-vulkan-cross: + runs-on: ubuntu-24.04 steps: - uses: actions/checkout@v4 - with: - fetch-depth: 0 - - name: Setup Riscv run: | sudo dpkg --add-architecture riscv64 - sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \ - /etc/apt/sources.list /etc/apt/apt-mirrors.txt - sudo apt-get clean - sudo apt-get update + + # Add arch-specific repositories for non-amd64 architectures + cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list + deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe + deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe + deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe + deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe + EOF + + sudo apt-get update || true ;# Prevent failure due to missing URLs. + sudo apt-get install -y --no-install-recommends \ build-essential \ glslc \ @@ -69,6 +81,7 @@ jobs: -DGGML_VULKAN=ON \ -DGGML_OPENMP=OFF \ -DLLAMA_BUILD_EXAMPLES=ON \ + -DLLAMA_BUILD_TOOLS=ON \ -DLLAMA_BUILD_TESTS=OFF \ -DCMAKE_SYSTEM_NAME=Linux \ -DCMAKE_SYSTEM_PROCESSOR=riscv64 \ @@ -82,21 +95,25 @@ jobs: cmake --build build --config Release -j $(nproc) - ubuntu-latest-arm64-vulkan-cross: - runs-on: ubuntu-latest + ubuntu-24-arm64-vulkan-cross: + runs-on: ubuntu-24.04 steps: - uses: actions/checkout@v4 - with: - fetch-depth: 0 - - name: Setup Arm64 run: | sudo dpkg --add-architecture arm64 - sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \ - /etc/apt/sources.list /etc/apt/apt-mirrors.txt - sudo apt-get clean - sudo apt-get update + + # Add arch-specific repositories for non-amd64 architectures + cat << EOF | sudo tee /etc/apt/sources.list.d/arm64-ports.list + deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe + deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe + deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe + deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe + EOF + + sudo apt-get update || true ;# Prevent failure due to missing URLs. + sudo apt-get install -y --no-install-recommends \ build-essential \ glslc \ @@ -110,6 +127,7 @@ jobs: -DGGML_VULKAN=ON \ -DGGML_OPENMP=OFF \ -DLLAMA_BUILD_EXAMPLES=ON \ + -DLLAMA_BUILD_TOOLS=ON \ -DLLAMA_BUILD_TESTS=OFF \ -DCMAKE_SYSTEM_NAME=Linux \ -DCMAKE_SYSTEM_PROCESSOR=aarch64 \ diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index 32c8b7717f..b62720f308 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -2,30 +2,19 @@ name: CI on: workflow_dispatch: # allows manual triggering - inputs: - create_release: - description: 'Create new release' - required: true - type: boolean push: branches: - master - paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp'] + paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp'] pull_request: types: [opened, synchronize, reopened] - paths: ['.github/workflows/build.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp'] + paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp'] concurrency: group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }} cancel-in-progress: true -# Fine-grant permission -# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token -permissions: - contents: write # for creating release - env: - BRANCH_NAME: ${{ github.head_ref || github.ref_name }} GGML_NLOOP: 3 GGML_N_THREADS: 1 LLAMA_LOG_COLORS: 1 @@ -40,8 +29,6 @@ jobs: - name: Clone id: checkout uses: actions/checkout@v4 - with: - fetch-depth: 0 - name: ccache uses: hendrikmuhs/ccache-action@v1.2.16 @@ -74,33 +61,6 @@ jobs: cd build ctest -L 'main|curl' --verbose --timeout 900 - - name: Determine tag name - id: tag - shell: bash - run: | - BUILD_NUMBER="$(git rev-list --count HEAD)" - SHORT_HASH="$(git rev-parse --short=7 HEAD)" - if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then - echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT - else - SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') - echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT - fi - - - name: Pack artifacts - id: pack_artifacts - if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} - run: | - cp LICENSE ./build/bin/ - zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/* - - - 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-macos-arm64.zip - name: llama-bin-macos-arm64.zip - macOS-latest-cmake-x64: runs-on: macos-13 @@ -108,8 +68,6 @@ jobs: - name: Clone id: checkout uses: actions/checkout@v4 - with: - fetch-depth: 0 - name: ccache uses: hendrikmuhs/ccache-action@v1.2.16 @@ -143,33 +101,6 @@ jobs: cd build ctest -L main --verbose --timeout 900 - - 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}" >> $GITHUB_OUTPUT - else - SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') - echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT - fi - - - name: Pack artifacts - id: pack_artifacts - if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} - run: | - cp LICENSE ./build/bin/ - zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/* - - - 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-macos-x64.zip - name: llama-bin-macos-x64.zip - ubuntu-cpu-cmake: strategy: matrix: @@ -185,8 +116,6 @@ jobs: - name: Clone id: checkout uses: actions/checkout@v4 - with: - fetch-depth: 0 - name: ccache uses: hendrikmuhs/ccache-action@v1.2.16 @@ -225,33 +154,6 @@ jobs: ./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf ./bin/llama-cli -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256 - - 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}" >> $GITHUB_OUTPUT - else - SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') - echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT - fi - - - name: Pack artifacts - id: pack_artifacts - if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} - run: | - cp LICENSE ./build/bin/ - zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip ./build/bin/* - - - 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-ubuntu-${{ matrix.build }}.zip - name: llama-bin-ubuntu-${{ matrix.build }}.zip - ubuntu-latest-cmake-sanitizer: runs-on: ubuntu-latest @@ -378,8 +280,6 @@ jobs: - name: Clone id: checkout uses: actions/checkout@v4 - with: - fetch-depth: 0 - name: ccache uses: hendrikmuhs/ccache-action@v1.2.16 @@ -407,34 +307,7 @@ jobs: run: | cd build # This is using llvmpipe and runs slower than other backends - ctest -L main --verbose --timeout 2700 - - - 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}" >> $GITHUB_OUTPUT - else - SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') - echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT - fi - - - name: Pack artifacts - id: pack_artifacts - if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} - run: | - cp LICENSE ./build/bin/ - zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip ./build/bin/* - - - 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-ubuntu-vulkan-x64.zip - name: llama-bin-ubuntu-vulkan-x64.zip + ctest -L main --verbose --timeout 3600 ubuntu-22-cmake-hip: runs-on: ubuntu-22.04 @@ -601,9 +474,8 @@ jobs: -DGGML_SYCL_F16=ON cmake --build build --config Release -j $(nproc) -# Disabled for now due to sporadic issue syncing. -# build-linux-cross: -# uses: ./.github/workflows/build-linux-cross.yml + build-linux-cross: + uses: ./.github/workflows/build-linux-cross.yml macOS-latest-cmake-ios: runs-on: macos-latest @@ -634,6 +506,7 @@ jobs: -DGGML_METAL_EMBED_LIBRARY=ON \ -DLLAMA_BUILD_COMMON=OFF \ -DLLAMA_BUILD_EXAMPLES=OFF \ + -DLLAMA_BUILD_TOOLS=OFF \ -DLLAMA_BUILD_TESTS=OFF \ -DLLAMA_BUILD_SERVER=OFF \ -DCMAKE_SYSTEM_NAME=iOS \ @@ -670,6 +543,7 @@ jobs: -DGGML_METAL_EMBED_LIBRARY=ON \ -DLLAMA_BUILD_COMMON=OFF \ -DLLAMA_BUILD_EXAMPLES=OFF \ + -DLLAMA_BUILD_TOOLS=OFF \ -DLLAMA_BUILD_TESTS=OFF \ -DLLAMA_BUILD_SERVER=OFF \ -DCMAKE_SYSTEM_NAME=tvOS \ @@ -700,6 +574,7 @@ jobs: -DGGML_METAL_EMBED_LIBRARY=ON \ -DLLAMA_BUILD_COMMON=OFF \ -DLLAMA_BUILD_EXAMPLES=OFF \ + -DLLAMA_BUILD_TOOLS=OFF \ -DLLAMA_BUILD_TESTS=OFF \ -DLLAMA_BUILD_SERVER=OFF \ -DCMAKE_SYSTEM_NAME=visionOS \ @@ -740,6 +615,7 @@ jobs: -DGGML_METAL_EMBED_LIBRARY=ON \ -DLLAMA_CURL=OFF \ -DLLAMA_BUILD_EXAMPLES=OFF \ + -DLLAMA_BUILD_TOOLS=OFF \ -DLLAMA_BUILD_TESTS=OFF \ -DLLAMA_BUILD_SERVER=OFF \ -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" @@ -768,7 +644,7 @@ jobs: uses: hendrikmuhs/ccache-action@v1.2.16 with: key: windows-msys2 - variant: sccache + variant: ccache evict-old-files: 1d - name: Setup ${{ matrix.sys }} @@ -811,39 +687,29 @@ jobs: strategy: matrix: include: - - build: 'noavx-x64' - defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF' - - build: 'avx2-x64' - defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON' - - build: 'avx-x64' - defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF' - - build: 'avx512-x64' - defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON' + - build: 'cpu-x64' + defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF' - build: 'openblas-x64' - defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"' - - build: 'kompute-x64' - defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON' + defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"' - build: 'vulkan-x64' - defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON' + defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON' - build: 'llvm-arm64' defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON' - - build: 'msvc-arm64' - defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON' - build: 'llvm-arm64-opencl-adreno' defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON' + # - build: 'kompute-x64' + # defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON' steps: - name: Clone id: checkout uses: actions/checkout@v4 - with: - fetch-depth: 0 - name: ccache uses: hendrikmuhs/ccache-action@v1.2.16 with: key: windows-latest-cmake-${{ matrix.build }} - variant: sccache + variant: ccache evict-old-files: 1d - name: Clone Kompute submodule @@ -919,68 +785,26 @@ jobs: cp $env:RUNNER_TEMP/openblas/bin/libopenblas.dll ./build/bin/Release/openblas.dll cp $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt ./build/bin/Release/OpenBLAS-${env:OPENBLAS_VERSION}.txt - - name: Check AVX512F support - id: check_avx512f - if: ${{ matrix.build == 'avx512-x64' }} - continue-on-error: true - run: | - cd build - $vcdir = $(vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath) - $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim())) - $cl = $(join-path $msvc 'bin\Hostx64\x64\cl.exe') - echo 'int main(void){unsigned int a[4];__cpuid(a,7);return !(a[1]&65536);}' >> avx512f.c - & $cl /O2 /GS- /kernel avx512f.c /link /nodefaultlib /entry:main - .\avx512f.exe && echo "AVX512F: YES" && ( echo HAS_AVX512F=1 >> $env:GITHUB_ENV ) || echo "AVX512F: NO" - - name: Test id: cmake_test - # not all machines have native AVX-512 - if: ${{ matrix.build != 'msvc-arm64' && matrix.build != 'llvm-arm64' && matrix.build != 'llvm-arm64-opencl-adreno' && matrix.build != 'kompute-x64' && matrix.build != 'vulkan-x64' && (matrix.build != 'avx512-x64' || env.HAS_AVX512F == '1') }} + if: ${{ matrix.build != 'llvm-arm64' && matrix.build != 'llvm-arm64-opencl-adreno' }} run: | cd build ctest -L main -C Release --verbose --timeout 900 - - name: Test (Intel SDE) - id: cmake_test_sde - if: ${{ matrix.build == 'avx512-x64' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation - run: | - curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/813591/sde-external-${env:SDE_VERSION}-win.tar.xz" - # for some weird reason windows tar doesn't like sde tar.xz - 7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar.xz - 7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar - $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe) - cd build - $env:LLAMA_SKIP_TESTS_SLOW_ON_EMULATOR = 1 - & $sde -future -- ctest -L main -C Release --verbose --timeout 900 - - - name: Determine tag name - id: tag - shell: bash - run: | - BUILD_NUMBER="$(git rev-list --count HEAD)" - SHORT_HASH="$(git rev-parse --short=7 HEAD)" - if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then - echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT - else - SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') - echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $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' }} - env: - CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} - run: | - Copy-Item $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll - 7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.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-${{ matrix.build }}.zip - name: llama-bin-win-${{ matrix.build }}.zip + # TODO: disabled for now, consider adding tests for all CPU variants instead + # - name: Test (Intel SDE) + # id: cmake_test_sde + # if: ${{ matrix.build == 'avx512-x64' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation + # run: | + # curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/813591/sde-external-${env:SDE_VERSION}-win.tar.xz" + # # for some weird reason windows tar doesn't like sde tar.xz + # 7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar.xz + # 7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar + # $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe) + # cd build + # $env:LLAMA_SKIP_TESTS_SLOW_ON_EMULATOR = 1 + # & $sde -future -- ctest -L main -C Release --verbose --timeout 900 ubuntu-latest-cmake-cuda: runs-on: ubuntu-latest @@ -990,8 +814,6 @@ jobs: - name: Clone id: checkout uses: actions/checkout@v4 - with: - fetch-depth: 0 - name: Install dependencies env: @@ -1023,77 +845,23 @@ jobs: strategy: matrix: cuda: ['12.4', '11.7'] - build: ['cuda'] steps: - name: Clone id: checkout uses: actions/checkout@v4 - with: - fetch-depth: 0 - name: Install ccache uses: hendrikmuhs/ccache-action@v1.2.16 with: - key: ${{ github.job }}-${{ matrix.cuda }}-${{ matrix.build }} - variant: sccache + key: windows-cuda-${{ matrix.cuda }} + variant: ccache evict-old-files: 1d - - name: Install Cuda Toolkit 11.7 - if: ${{ matrix.cuda == '11.7' }} - run: | - mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" - choco install unzip -y - curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-11.7.99-archive.zip" - curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-11.7.99-archive.zip" - curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-11.7.99-archive.zip" - curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-11.7.4.6-archive.zip" - curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-11.7.91-archive.zip" - curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-11.7.91-archive.zip" - curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-11.7.101-archive.zip" - curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-11.7.91-archive.zip" - unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" - xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_cudart-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y - xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvcc-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y - xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvrtc-windows-x86_64-11.7.99-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y - xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\libcublas-windows-x86_64-11.7.4.6-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y - xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvtx-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y - xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\visual_studio_integration-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y - xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_nvprof-windows-x86_64-11.7.101-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y - xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\cuda_cccl-windows-x86_64-11.7.91-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" /E /I /H /Y - echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append - echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append - echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8 - echo "CUDA_PATH_V11_7=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.7" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8 - - - name: Install Cuda Toolkit 12.4 - if: ${{ matrix.cuda == '12.4' }} - run: | - mkdir -p "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" - choco install unzip -y - curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/windows-x86_64/cuda_cudart-windows-x86_64-12.4.127-archive.zip" - curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/windows-x86_64/cuda_nvcc-windows-x86_64-12.4.131-archive.zip" - curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvrtc/windows-x86_64/cuda_nvrtc-windows-x86_64-12.4.127-archive.zip" - curl -O "https://developer.download.nvidia.com/compute/cuda/redist/libcublas/windows-x86_64/libcublas-windows-x86_64-12.4.5.8-archive.zip" - curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvtx/windows-x86_64/cuda_nvtx-windows-x86_64-12.4.127-archive.zip" - curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_profiler_api/windows-x86_64/cuda_profiler_api-windows-x86_64-12.4.127-archive.zip" - curl -O "https://developer.download.nvidia.com/compute/cuda/redist/visual_studio_integration/windows-x86_64/visual_studio_integration-windows-x86_64-12.4.127-archive.zip" - curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvprof/windows-x86_64/cuda_nvprof-windows-x86_64-12.4.127-archive.zip" - curl -O "https://developer.download.nvidia.com/compute/cuda/redist/cuda_cccl/windows-x86_64/cuda_cccl-windows-x86_64-12.4.127-archive.zip" - unzip '*.zip' -d "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" - xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_cudart-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y - xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvcc-windows-x86_64-12.4.131-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y - xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvrtc-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y - xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libcublas-windows-x86_64-12.4.5.8-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y - xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvtx-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y - xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_profiler_api-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y - xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\visual_studio_integration-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y - xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_nvprof-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y - xcopy "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\cuda_cccl-windows-x86_64-12.4.127-archive\*" "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" /E /I /H /Y - echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\bin" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append - echo "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\libnvvp" | Out-File -FilePath $env:GITHUB_PATH -Encoding utf8 -Append - echo "CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8 - echo "CUDA_PATH_V12_4=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4" | Out-File -FilePath $env:GITHUB_ENV -Append -Encoding utf8 + - name: Install Cuda Toolkit + uses: ./.github/actions/windows-setup-cuda + with: + cuda_version: ${{ matrix.cuda }} - name: Install Ninja id: install_ninja @@ -1114,6 +882,8 @@ jobs: cmake -S . -B build -G "Ninja Multi-Config" ^ -DLLAMA_BUILD_SERVER=ON ^ -DGGML_NATIVE=OFF ^ + -DGGML_BACKEND_DL=ON ^ + -DGGML_CPU_ALL_VARIANTS=ON ^ -DGGML_CUDA=ON ^ -DGGML_RPC=ON ^ -DCURL_LIBRARY="%CURL_PATH%/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="%CURL_PATH%/include" @@ -1121,51 +891,6 @@ jobs: cmake --build build --config Release -j %NINJA_JOBS% -t ggml cmake --build build --config Release - - name: Determine tag name - id: tag - shell: bash - run: | - BUILD_NUMBER="$(git rev-list --count HEAD)" - SHORT_HASH="$(git rev-parse --short=7 HEAD)" - if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then - echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT - else - SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') - echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $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' }} - env: - CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} - run: | - cp $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll - 7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.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-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip - name: llama-bin-win-cu${{ matrix.cuda }}-x64.zip - - - name: Copy and pack Cuda runtime - if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }} - 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" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll - 7z a cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip $dst\* - - - name: Upload Cuda runtime - if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} - uses: actions/upload-artifact@v4 - with: - path: cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip - name: cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip - windows-latest-cmake-sycl: runs-on: windows-latest @@ -1181,14 +906,12 @@ jobs: - name: Clone id: checkout uses: actions/checkout@v4 - with: - fetch-depth: 0 - name: ccache uses: hendrikmuhs/ccache-action@v1.2.16 with: key: windows-latest-cmake-sycl - variant: sccache + variant: ccache evict-old-files: 1d - name: Install @@ -1201,52 +924,6 @@ jobs: id: cmake_build run: examples/sycl/win-build-sycl.bat - - name: Determine tag name - id: tag - shell: bash - run: | - BUILD_NUMBER="$(git rev-list --count HEAD)" - SHORT_HASH="$(git rev-parse --short=7 HEAD)" - if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then - echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT - else - SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') - echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT - fi - - - name: Build the release package - id: pack_artifacts - if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} - run: | - echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin" - - cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.5.dll" ./build/bin - cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin - cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin - - cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin - cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_opencl.dll" ./build/bin - cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_loader.dll" ./build/bin - cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_win_proxy_loader.dll" ./build/bin - - cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl8.dll" ./build/bin - cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin - cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin - cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin - - cp "${{ env.ONEAPI_ROOT }}/dnnl/latest/bin/dnnl.dll" ./build/bin - cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin - - echo "cp oneAPI running time dll files to ./build/bin done" - 7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/* - - - name: Upload the release package - 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-sycl-x64.zip - name: llama-bin-win-sycl-x64.zip - windows-latest-cmake-hip: if: ${{ github.event.inputs.create_release != 'true' }} runs-on: windows-latest @@ -1304,110 +981,12 @@ jobs: -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" cmake --build build -j ${env:NUMBER_OF_PROCESSORS} - # TODO: reuse windows-latest-cmake-hip instead of duplicating this job - windows-latest-cmake-hip-release: - if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} - runs-on: windows-latest - - strategy: - matrix: - gpu_target: [gfx1100, gfx1101, gfx1030] - - steps: - - name: Clone - id: checkout - uses: actions/checkout@v4 - with: - fetch-depth: 0 - - - name: Clone rocWMMA repository - id: clone_rocwmma - run: | - git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1 - - - name: ccache - uses: hendrikmuhs/ccache-action@v1.2.16 - with: - key: windows-latest-cmake-hip-release - evict-old-files: 1d - - - name: Install - id: depends - run: | - $ErrorActionPreference = "Stop" - write-host "Downloading AMD HIP SDK Installer" - Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe" - write-host "Installing AMD HIP SDK" - Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait - write-host "Completed AMD HIP SDK installation" - - - name: Verify ROCm - id: verify - run: | - & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version - - - name: libCURL - id: get_libcurl - uses: ./.github/actions/windows-setup-curl - - - name: Build - id: cmake_build - env: - CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} - run: | - $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path) - $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}" - cmake -G "Unix Makefiles" -B build -S . ` - -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" ` - -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" ` - -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/" ` - -DCMAKE_BUILD_TYPE=Release ` - -DAMDGPU_TARGETS=${{ matrix.gpu_target }} ` - -DGGML_HIP_ROCWMMA_FATTN=ON ` - -DGGML_HIP=ON ` - -DGGML_RPC=ON ` - -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" - cmake --build build -j ${env:NUMBER_OF_PROCESSORS} - md "build\bin\rocblas\library\" - cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\" - cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\" - cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\" - - - 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}" >> $GITHUB_OUTPUT - else - SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') - echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT - fi - - - name: Pack artifacts - id: pack_artifacts - env: - CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} - run: | - cp $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\libcurl-x64.dll - 7z a llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip .\build\bin\* - - - name: Upload artifacts - uses: actions/upload-artifact@v4 - with: - path: llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip - name: llama-bin-win-hip-x64-${{ matrix.gpu_target }}.zip - ios-xcode-build: runs-on: macos-latest steps: - name: Checkout code uses: actions/checkout@v4 - with: - fetch-depth: 0 - name: Build id: cmake_build @@ -1418,6 +997,7 @@ jobs: -DGGML_METAL_EMBED_LIBRARY=ON \ -DLLAMA_CURL=OFF \ -DLLAMA_BUILD_EXAMPLES=OFF \ + -DLLAMA_BUILD_TOOLS=OFF \ -DLLAMA_BUILD_TESTS=OFF \ -DLLAMA_BUILD_SERVER=OFF \ -DCMAKE_SYSTEM_NAME=iOS \ @@ -1433,32 +1013,6 @@ jobs: - name: Build Xcode project run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build - - 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}" >> $GITHUB_OUTPUT - else - SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') - echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT - fi - - - name: Pack artifacts - id: pack_artifacts - if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} - run: | - zip --symlinks -r llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework - - - 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 }}-xcframework.zip - name: llama-${{ steps.tag.outputs.name }}-xcframework - android-build: runs-on: ubuntu-latest @@ -1486,283 +1040,8 @@ jobs: - name: Build run: | cd examples/llama.android - ./gradlew build --no-daemon - release: - if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} - - runs-on: ubuntu-latest - - needs: - - ubuntu-cpu-cmake - - ubuntu-22-cmake-vulkan - - windows-latest-cmake - - windows-2019-cmake-cuda - - windows-latest-cmake-sycl - - windows-latest-cmake-hip-release - - macOS-latest-cmake-arm64 - - macOS-latest-cmake-x64 - - steps: - - name: Clone - id: checkout - uses: actions/checkout@v4 - with: - fetch-depth: 0 - - - name: ccache - uses: hendrikmuhs/ccache-action@v1.2.16 - with: - key: release - evict-old-files: 1d - - - name: Determine tag name - id: tag - shell: bash - run: | - BUILD_NUMBER="$(git rev-list --count HEAD)" - SHORT_HASH="$(git rev-parse --short=7 HEAD)" - if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then - echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT - else - SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') - echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT - fi - - - 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 - uses: ggml-org/action-create-release@v1 - env: - GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }} - with: - tag_name: ${{ steps.tag.outputs.name }} - - - name: Upload release - id: upload_release - 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/${file}`) - }); - } - } - -# ubuntu-latest-gcc: -# runs-on: ubuntu-latest -# -# strategy: -# matrix: -# build: [Debug, Release] -# -# steps: -# - name: Clone -# uses: actions/checkout@v4 -# -# - name: Dependencies -# run: | -# sudo apt-get update -# sudo apt-get install build-essential -# sudo apt-get install cmake -# -# - name: Configure -# run: cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }} -# -# - name: Build -# run: | -# make -# -# ubuntu-latest-clang: -# runs-on: ubuntu-latest -# -# strategy: -# matrix: -# build: [Debug, Release] -# -# steps: -# - name: Clone -# uses: actions/checkout@v4 -# -# - name: Dependencies -# run: | -# sudo apt-get update -# sudo apt-get install build-essential -# sudo apt-get install cmake -# -# - name: Configure -# run: cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang -# -# - name: Build -# run: | -# make -# -# ubuntu-latest-gcc-sanitized: -# runs-on: ubuntu-latest -# -# strategy: -# matrix: -# sanitizer: [ADDRESS, THREAD, UNDEFINED] -# -# steps: -# - name: Clone -# uses: actions/checkout@v4 -# -# - name: Dependencies -# run: | -# sudo apt-get update -# sudo apt-get install build-essential -# sudo apt-get install cmake -# -# - name: Configure -# run: cmake . -DCMAKE_BUILD_TYPE=Debug -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -# -# - name: Build -# run: | -# make -# -# windows: -# runs-on: windows-latest -# -# strategy: -# matrix: -# build: [Release] -# arch: [Win32, x64] -# include: -# - arch: Win32 -# s2arc: x86 -# - arch: x64 -# s2arc: x64 -# -# steps: -# - name: Clone -# uses: actions/checkout@v4 -# -# - name: Add msbuild to PATH -# uses: microsoft/setup-msbuild@v1 -# -# - name: Configure -# run: > -# cmake -S . -B ./build -A ${{ matrix.arch }} -# -DCMAKE_BUILD_TYPE=${{ matrix.build }} -# -# - name: Build -# run: | -# cd ./build -# msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }} -# -# - name: Upload binaries -# uses: actions/upload-artifact@v4 -# with: -# name: llama-bin-${{ matrix.arch }} -# path: build/bin/${{ matrix.build }} -# -# windows-blas: -# runs-on: windows-latest -# -# strategy: -# matrix: -# build: [Release] -# arch: [Win32, x64] -# blas: [ON] -# include: -# - arch: Win32 -# obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-x86.zip -# s2arc: x86 -# - arch: x64 -# obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-x64.zip -# s2arc: x64 -# -# steps: -# - name: Clone -# uses: actions/checkout@v4 -# -# - name: Add msbuild to PATH -# uses: microsoft/setup-msbuild@v1 -# -# - name: Fetch OpenBLAS -# if: matrix.blas == 'ON' -# run: | -# C:/msys64/usr/bin/wget.exe -qO blas.zip ${{ matrix.obzip }} -# 7z x blas.zip -oblas -y -# copy blas/include/cblas.h . -# copy blas/include/openblas_config.h . -# echo "blasdir=$env:GITHUB_WORKSPACE/blas" >> $env:GITHUB_ENV -# -# - name: Configure -# run: > -# cmake -S . -B ./build -A ${{ matrix.arch }} -# -DCMAKE_BUILD_TYPE=${{ matrix.build }} -# -DLLAMA_SUPPORT_OPENBLAS=${{ matrix.blas }} -# -DCMAKE_LIBRARY_PATH="$env:blasdir/lib" -# -# - name: Build -# run: | -# cd ./build -# msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }} -# -# - name: Copy libopenblas.dll -# if: matrix.blas == 'ON' -# run: copy "$env:blasdir/bin/libopenblas.dll" build/bin/${{ matrix.build }} -# -# - name: Upload binaries -# if: matrix.blas == 'ON' -# uses: actions/upload-artifact@v4 -# with: -# name: llama-blas-bin-${{ matrix.arch }} -# path: build/bin/${{ matrix.build }} -# -# emscripten: -# runs-on: ubuntu-latest -# -# strategy: -# matrix: -# build: [Release] -# -# steps: -# - name: Clone -# uses: actions/checkout@v4 -# -# - name: Dependencies -# run: | -# wget -q https://github.com/emscripten-core/emsdk/archive/master.tar.gz -# tar -xvf master.tar.gz -# emsdk-master/emsdk update -# emsdk-master/emsdk install latest -# emsdk-master/emsdk activate latest -# -# - name: Configure -# run: echo "tmp" -# -# - name: Build -# run: | -# pushd emsdk-master -# source ./emsdk_env.sh -# popd -# emcmake cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }} -# make - openEuler-latest-cmake-cann: if: ${{ github.event_name != 'pull_request' || contains(github.event.pull_request.labels.*.name, 'Ascend NPU') }} defaults: diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml index 114cbf83ed..2067927be5 100644 --- a/.github/workflows/docker.yml +++ b/.github/workflows/docker.yml @@ -36,10 +36,13 @@ jobs: matrix: config: # Multi-stage build - - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: false } + # Note: the arm64 images are failing, which prevents the amd64 images from being built + # https://github.com/ggml-org/llama.cpp/issues/11888 + #- { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: false } + - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false } - { tag: "cuda", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false } - { tag: "musa", dockerfile: ".devops/musa.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true } - - { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false } + - { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true } - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false } # Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete #- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: true } diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml new file mode 100644 index 0000000000..5f54909dcb --- /dev/null +++ b/.github/workflows/release.yml @@ -0,0 +1,709 @@ +name: Create Release + +on: + workflow_dispatch: # allows manual triggering + inputs: + create_release: + description: 'Create new release' + required: true + type: boolean + push: + branches: + - master + paths: ['.github/workflows/release.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp'] + +concurrency: + group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }} + cancel-in-progress: true + +env: + BRANCH_NAME: ${{ github.head_ref || github.ref_name }} + CMAKE_ARGS: "-DLLAMA_BUILD_EXAMPLES=OFF -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_TOOLS=ON -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON" + +jobs: + macOS-arm64: + runs-on: macos-14 + + steps: + - name: Clone + id: checkout + uses: actions/checkout@v4 + with: + fetch-depth: 0 + + - name: ccache + uses: hendrikmuhs/ccache-action@v1.2.16 + with: + key: macOS-latest-cmake-arm64 + evict-old-files: 1d + + - name: Dependencies + id: depends + continue-on-error: true + run: | + brew update + brew install curl + + - name: Build + id: cmake_build + run: | + sysctl -a + cmake -B build \ + -DCMAKE_BUILD_RPATH="@loader_path" \ + -DLLAMA_FATAL_WARNINGS=ON \ + -DGGML_METAL_USE_BF16=ON \ + -DGGML_METAL_EMBED_LIBRARY=ON \ + -DGGML_RPC=ON \ + ${{ env.CMAKE_ARGS }} + cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) + + - name: Determine tag name + id: tag + uses: ./.github/actions/get-tag-name + + - name: Pack artifacts + id: pack_artifacts + run: | + cp LICENSE ./build/bin/ + zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/* + + - name: Upload artifacts + uses: actions/upload-artifact@v4 + with: + path: llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip + name: llama-bin-macos-arm64.zip + + macOS-x64: + runs-on: macos-13 + + steps: + - name: Clone + id: checkout + uses: actions/checkout@v4 + with: + fetch-depth: 0 + + - name: ccache + uses: hendrikmuhs/ccache-action@v1.2.16 + with: + key: macOS-latest-cmake-x64 + evict-old-files: 1d + + - name: Dependencies + id: depends + continue-on-error: true + run: | + brew update + brew install curl + + - name: Build + id: cmake_build + run: | + sysctl -a + # Metal is disabled due to intermittent failures with Github runners not having a GPU: + # https://github.com/ggml-org/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313 + cmake -B build \ + -DCMAKE_BUILD_RPATH="@loader_path" \ + -DLLAMA_FATAL_WARNINGS=ON \ + -DGGML_METAL=OFF \ + -DGGML_RPC=ON + cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) + + - name: Determine tag name + id: tag + uses: ./.github/actions/get-tag-name + + - name: Pack artifacts + id: pack_artifacts + run: | + cp LICENSE ./build/bin/ + zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/* + + - name: Upload artifacts + uses: actions/upload-artifact@v4 + with: + path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip + name: llama-bin-macos-x64.zip + + ubuntu-22-cpu: + strategy: + matrix: + include: + - build: 'x64' + os: ubuntu-22.04 + - build: 'arm64' + os: ubuntu-22.04-arm + + runs-on: ${{ matrix.os }} + + steps: + - name: Clone + id: checkout + uses: actions/checkout@v4 + with: + fetch-depth: 0 + + - name: ccache + uses: hendrikmuhs/ccache-action@v1.2.16 + with: + key: ubuntu-cpu-cmake + evict-old-files: 1d + + - name: Dependencies + id: depends + run: | + sudo apt-get update + sudo apt-get install build-essential libcurl4-openssl-dev + + - name: Build + id: cmake_build + run: | + cmake -B build \ + -DLLAMA_FATAL_WARNINGS=ON \ + ${{ env.CMAKE_ARGS }} + cmake --build build --config Release -j $(nproc) + + - name: Determine tag name + id: tag + uses: ./.github/actions/get-tag-name + + - name: Pack artifacts + id: pack_artifacts + run: | + cp LICENSE ./build/bin/ + zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip ./build/bin/* + + - name: Upload artifacts + uses: actions/upload-artifact@v4 + with: + path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip + name: llama-bin-ubuntu-${{ matrix.build }}.zip + + ubuntu-22-vulkan: + runs-on: ubuntu-22.04 + + steps: + - name: Clone + id: checkout + uses: actions/checkout@v4 + with: + fetch-depth: 0 + + - name: ccache + uses: hendrikmuhs/ccache-action@v1.2.16 + with: + key: ubuntu-22-cmake-vulkan + evict-old-files: 1d + + - name: Dependencies + id: depends + run: | + wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add - + sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list + sudo apt-get update -y + sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev + + - name: Build + id: cmake_build + run: | + cmake -B build \ + -DGGML_VULKAN=ON \ + ${{ env.CMAKE_ARGS }} + cmake --build build --config Release -j $(nproc) + + - name: Determine tag name + id: tag + uses: ./.github/actions/get-tag-name + + - name: Pack artifacts + id: pack_artifacts + run: | + cp LICENSE ./build/bin/ + zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip ./build/bin/* + + - name: Upload artifacts + uses: actions/upload-artifact@v4 + with: + path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip + name: llama-bin-ubuntu-vulkan-x64.zip + + windows: + runs-on: windows-latest + + env: + OPENBLAS_VERSION: 0.3.23 + VULKAN_VERSION: 1.4.309.0 + + strategy: + matrix: + include: + - build: 'cpu-x64' + defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF' + #- build: 'openblas-x64' + # defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"' + - build: 'vulkan-x64' + defines: '-DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON' + - build: 'cpu-arm64' + defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF' + - build: 'opencl-adreno-arm64' + defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON' + + steps: + - name: Clone + id: checkout + uses: actions/checkout@v4 + with: + fetch-depth: 0 + + - name: ccache + uses: hendrikmuhs/ccache-action@v1.2.16 + with: + key: windows-latest-cmake-${{ matrix.build }} + variant: ccache + evict-old-files: 1d + + - name: Download OpenBLAS + id: get_openblas + if: ${{ matrix.build == 'openblas-x64' }} + run: | + curl.exe -o $env:RUNNER_TEMP/openblas.zip -L "https://github.com/xianyi/OpenBLAS/releases/download/v${env:OPENBLAS_VERSION}/OpenBLAS-${env:OPENBLAS_VERSION}-x64.zip" + curl.exe -o $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt -L "https://github.com/xianyi/OpenBLAS/raw/v${env:OPENBLAS_VERSION}/LICENSE" + mkdir $env:RUNNER_TEMP/openblas + tar.exe -xvf $env:RUNNER_TEMP/openblas.zip -C $env:RUNNER_TEMP/openblas + $vcdir = $(vswhere -latest -products * -requires Microsoft.VisualStudio.Component.VC.Tools.x86.x64 -property installationPath) + $msvc = $(join-path $vcdir $('VC\Tools\MSVC\'+$(gc -raw $(join-path $vcdir 'VC\Auxiliary\Build\Microsoft.VCToolsVersion.default.txt')).Trim())) + $lib = $(join-path $msvc 'bin\Hostx64\x64\lib.exe') + & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll + + - name: Install Vulkan SDK + id: get_vulkan + if: ${{ matrix.build == 'vulkan-x64' }} + run: | + curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe" + & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install + Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}" + Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin" + + - name: Install Ninja + id: install_ninja + run: | + choco install ninja + + - name: Install OpenCL Headers and Libs + id: install_opencl + if: ${{ matrix.build == 'opencl-adreno-arm64' }} + run: | + git clone https://github.com/KhronosGroup/OpenCL-Headers + cd OpenCL-Headers + cmake -B build ` + -DBUILD_TESTING=OFF ` + -DOPENCL_HEADERS_BUILD_TESTING=OFF ` + -DOPENCL_HEADERS_BUILD_CXX_TESTS=OFF ` + -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release" + cmake --build build --target install + git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader + cd OpenCL-ICD-Loader + cmake -B build-arm64-release ` + -A arm64 ` + -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" ` + -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release" + cmake --build build-arm64-release --target install --config release + + - name: libCURL + id: get_libcurl + uses: ./.github/actions/windows-setup-curl + + - name: Build + id: cmake_build + env: + CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} + run: | + cmake -S . -B build ${{ matrix.defines }} ` + -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" ` + ${{ env.CMAKE_ARGS }} + cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} + + - name: Add libopenblas.dll + id: add_libopenblas_dll + if: ${{ matrix.build == 'openblas-x64' }} + run: | + cp $env:RUNNER_TEMP/openblas/bin/libopenblas.dll ./build/bin/Release/openblas.dll + cp $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt ./build/bin/Release/OpenBLAS-${env:OPENBLAS_VERSION}.txt + + - name: Determine tag name + id: tag + uses: ./.github/actions/get-tag-name + + - name: Pack artifacts + id: pack_artifacts + env: + CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} + run: | + Copy-Item $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll + 7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\* + + - name: Upload artifacts + uses: actions/upload-artifact@v4 + with: + path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip + name: llama-bin-win-${{ matrix.build }}.zip + + windows-cuda: + runs-on: windows-2019 + + strategy: + matrix: + cuda: ['12.4', '11.7'] + + steps: + - name: Clone + id: checkout + uses: actions/checkout@v4 + with: + fetch-depth: 0 + + - name: Install ccache + uses: hendrikmuhs/ccache-action@v1.2.16 + with: + key: windows-cuda-${{ matrix.cuda }} + variant: ccache + evict-old-files: 1d + + - name: Install Cuda Toolkit + uses: ./.github/actions/windows-setup-cuda + with: + cuda_version: ${{ matrix.cuda }} + + - name: Install Ninja + id: install_ninja + run: | + choco install ninja + + - name: libCURL + id: get_libcurl + uses: ./.github/actions/windows-setup-curl + + - name: Build + id: cmake_build + shell: cmd + env: + CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} + run: | + call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat" + cmake -S . -B build -G "Ninja Multi-Config" ^ + -DGGML_NATIVE=OFF ^ + -DGGML_BACKEND_DL=ON ^ + -DGGML_CPU_ALL_VARIANTS=ON ^ + -DGGML_CUDA=ON ^ + -DCURL_LIBRARY="%CURL_PATH%/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="%CURL_PATH%/include" ^ + ${{ env.CMAKE_ARGS }} + set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1 + cmake --build build --config Release -j %NINJA_JOBS% -t ggml + cmake --build build --config Release + + - name: Determine tag name + id: tag + uses: ./.github/actions/get-tag-name + + - name: Pack artifacts + id: pack_artifacts + env: + CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} + run: | + cp $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll + 7z a llama-${{ steps.tag.outputs.name }}-bin-win-cuda${{ matrix.cuda }}-x64.zip .\build\bin\Release\* + + - name: Upload artifacts + uses: actions/upload-artifact@v4 + with: + path: llama-${{ steps.tag.outputs.name }}-bin-win-cuda${{ matrix.cuda }}-x64.zip + name: llama-bin-win-cuda${{ matrix.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" $dst cudart64_*.dll cublas64_*.dll cublasLt64_*.dll + 7z a cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip $dst\* + + - name: Upload Cuda runtime + uses: actions/upload-artifact@v4 + with: + path: cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip + name: cudart-llama-bin-win-cuda${{ matrix.cuda }}-x64.zip + + windows-sycl: + runs-on: windows-latest + + defaults: + run: + shell: bash + + env: + WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b380d914-366b-4b77-a74a-05e3c38b3514/intel-oneapi-base-toolkit-2025.0.0.882_offline.exe + WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel + ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI" + steps: + - name: Clone + id: checkout + uses: actions/checkout@v4 + with: + fetch-depth: 0 + + - name: ccache + uses: hendrikmuhs/ccache-action@v1.2.16 + with: + key: windows-latest-cmake-sycl + variant: ccache + evict-old-files: 1d + + - name: Install + run: | + scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL + + # TODO: add libcurl support ; we will also need to modify win-build-sycl.bat to accept user-specified args + + - name: Build + id: cmake_build + run: examples/sycl/win-build-sycl.bat + + - name: Determine tag name + id: tag + uses: ./.github/actions/get-tag-name + + - name: Build the release package + id: pack_artifacts + run: | + echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin" + + cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.5.dll" ./build/bin + cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin + cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin + + cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin + cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_opencl.dll" ./build/bin + cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_loader.dll" ./build/bin + cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_win_proxy_loader.dll" ./build/bin + + cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl8.dll" ./build/bin + cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin + cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin + cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin + + cp "${{ env.ONEAPI_ROOT }}/dnnl/latest/bin/dnnl.dll" ./build/bin + cp "${{ env.ONEAPI_ROOT }}/tbb/latest/bin/tbb12.dll" ./build/bin + + echo "cp oneAPI running time dll files to ./build/bin done" + 7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/* + + - name: Upload the release package + uses: actions/upload-artifact@v4 + with: + path: llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip + name: llama-bin-win-sycl-x64.zip + + windows-hip: + runs-on: windows-latest + + strategy: + matrix: + gpu_target: [gfx1100, gfx1101, gfx1030] + + steps: + - name: Clone + id: checkout + uses: actions/checkout@v4 + with: + fetch-depth: 0 + + - name: Clone rocWMMA repository + id: clone_rocwmma + run: | + git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1 + + - name: ccache + uses: hendrikmuhs/ccache-action@v1.2.16 + with: + key: windows-latest-cmake-hip-release + evict-old-files: 1d + + - name: Install + id: depends + run: | + $ErrorActionPreference = "Stop" + write-host "Downloading AMD HIP SDK Installer" + Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe" + write-host "Installing AMD HIP SDK" + Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait + write-host "Completed AMD HIP SDK installation" + + - name: Verify ROCm + id: verify + run: | + & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version + + - name: libCURL + id: get_libcurl + uses: ./.github/actions/windows-setup-curl + + - name: Build + id: cmake_build + env: + CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} + run: | + $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path) + $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}" + cmake -G "Unix Makefiles" -B build -S . ` + -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" ` + -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" ` + -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/" ` + -DCMAKE_BUILD_TYPE=Release ` + -DAMDGPU_TARGETS=${{ matrix.gpu_target }} ` + -DGGML_HIP_ROCWMMA_FATTN=ON ` + -DGGML_HIP=ON ` + -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" ` + ${{ env.CMAKE_ARGS }} + cmake --build build -j ${env:NUMBER_OF_PROCESSORS} + md "build\bin\rocblas\library\" + cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\" + cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\" + cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\" + + - name: Determine tag name + id: tag + uses: ./.github/actions/get-tag-name + + - name: Pack artifacts + id: pack_artifacts + env: + CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} + run: | + cp $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\libcurl-x64.dll + 7z a llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip .\build\bin\* + + - name: Upload artifacts + uses: actions/upload-artifact@v4 + with: + path: llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip + name: llama-bin-win-hip-x64-${{ matrix.gpu_target }}.zip + + ios-xcode-build: + runs-on: macos-latest + + steps: + - name: Checkout code + uses: actions/checkout@v4 + with: + fetch-depth: 0 + + - name: Build + id: cmake_build + run: | + sysctl -a + cmake -B build -G Xcode \ + -DGGML_METAL_USE_BF16=ON \ + -DGGML_METAL_EMBED_LIBRARY=ON \ + -DLLAMA_CURL=OFF \ + -DLLAMA_BUILD_EXAMPLES=OFF \ + -DLLAMA_BUILD_TOOLS=OFF \ + -DLLAMA_BUILD_TESTS=OFF \ + -DLLAMA_BUILD_SERVER=OFF \ + -DCMAKE_SYSTEM_NAME=iOS \ + -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \ + -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml + cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO + + - name: xcodebuild for swift package + id: xcodebuild + run: | + ./build-xcframework.sh + + - name: Build Xcode project + run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build + + - name: Determine tag name + id: tag + uses: ./.github/actions/get-tag-name + + - name: Pack artifacts + id: pack_artifacts + run: | + zip --symlinks -r llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework + + - name: Upload artifacts + uses: actions/upload-artifact@v4 + with: + path: llama-${{ steps.tag.outputs.name }}-xcframework.zip + name: llama-${{ steps.tag.outputs.name }}-xcframework + + release: + if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} + + # Fine-grant permission + # https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token + permissions: + contents: write # for creating release + + runs-on: ubuntu-latest + + needs: + - ubuntu-22-cpu + - ubuntu-22-vulkan + - windows + - windows-cuda + - windows-sycl + - windows-hip + - macOS-arm64 + - macOS-x64 + + steps: + - name: Clone + id: checkout + uses: actions/checkout@v4 + with: + fetch-depth: 0 + + - name: Determine tag name + id: tag + uses: ./.github/actions/get-tag-name + + - 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 + uses: ggml-org/action-create-release@v1 + env: + GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }} + with: + tag_name: ${{ steps.tag.outputs.name }} + + - name: Upload release + id: upload_release + 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/server.yml b/.github/workflows/server.yml index 6c9b513227..4baf6f6c75 100644 --- a/.github/workflows/server.yml +++ b/.github/workflows/server.yml @@ -15,10 +15,10 @@ on: push: branches: - master - paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*'] + paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*'] pull_request: types: [opened, synchronize, reopened] - paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*'] + paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*'] env: LLAMA_LOG_COLORS: 1 @@ -74,7 +74,7 @@ jobs: - name: Tests dependencies id: test_dependencies run: | - pip install -r examples/server/tests/requirements.txt + pip install -r tools/server/tests/requirements.txt # Setup nodejs (to be used for verifying bundled index.html) - uses: actions/setup-node@v4 @@ -84,14 +84,14 @@ jobs: - name: WebUI - Install dependencies id: webui_lint run: | - cd examples/server/webui + cd tools/server/webui npm ci - name: WebUI - Check code format id: webui_format run: | git config --global --add safe.directory $(realpath .) - cd examples/server/webui + cd tools/server/webui git status npm run format @@ -108,7 +108,7 @@ jobs: id: verify_server_index_html run: | git config --global --add safe.directory $(realpath .) - cd examples/server/webui + cd tools/server/webui git status npm run build @@ -161,21 +161,21 @@ jobs: env: GITHUB_ACTIONS: "true" run: | - cd examples/server/tests + cd tools/server/tests ./tests.sh - name: Tests (sanitizers) id: server_integration_tests_sanitizers if: ${{ matrix.sanitizer != '' }} run: | - cd examples/server/tests + cd tools/server/tests LLAMA_SANITIZE=1 ./tests.sh - name: Slow tests id: server_integration_tests_slow if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }} run: | - cd examples/server/tests + cd tools/server/tests SLOW_TESTS=1 ./tests.sh @@ -211,7 +211,7 @@ jobs: - name: Tests dependencies id: test_dependencies run: | - pip install -r examples/server/tests/requirements.txt + pip install -r tools/server/tests/requirements.txt - name: Copy Libcurl id: prepare_libcurl @@ -224,7 +224,7 @@ jobs: id: server_integration_tests if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }} run: | - cd examples/server/tests + cd tools/server/tests $env:PYTHONIOENCODING = ":replace" pytest -v -x -m "not slow" @@ -232,6 +232,6 @@ jobs: id: server_integration_tests_slow if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }} run: | - cd examples/server/tests + cd tools/server/tests $env:SLOW_TESTS = "1" pytest -v -x diff --git a/.gitignore b/.gitignore index 2c67ad7f7c..f8ceb1560a 100644 --- a/.gitignore +++ b/.gitignore @@ -96,11 +96,11 @@ perf-*.txt # Examples examples/jeopardy/results.txt -examples/server/*.css.hpp -examples/server/*.html.hpp -examples/server/*.js.hpp -examples/server/*.mjs.hpp -examples/server/*.gz.hpp +tools/server/*.css.hpp +tools/server/*.html.hpp +tools/server/*.js.hpp +tools/server/*.mjs.hpp +tools/server/*.gz.hpp !build_64.sh !examples/*.bat !examples/*/*.kts @@ -110,7 +110,7 @@ examples/server/*.gz.hpp # Server Web UI temporary files node_modules -examples/server/webui/dist +tools/server/webui/dist # Python diff --git a/CMakeLists.txt b/CMakeLists.txt index de51c0a17b..ac3e909033 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -77,6 +77,7 @@ option(LLAMA_BUILD_COMMON "llama: build common utils library" ${LLAMA_STANDALONE # extra artifacts option(LLAMA_BUILD_TESTS "llama: build tests" ${LLAMA_STANDALONE}) +option(LLAMA_BUILD_TOOLS "llama: build tools" ${LLAMA_STANDALONE}) option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE}) option(LLAMA_BUILD_SERVER "llama: build server example" ${LLAMA_STANDALONE}) @@ -187,6 +188,10 @@ if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_EXAMPLES) add_subdirectory(pocs) endif() +if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TOOLS) + add_subdirectory(tools) +endif() + # # install # @@ -247,20 +252,3 @@ configure_file(cmake/llama.pc.in install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc" DESTINATION ${CMAKE_INSTALL_LIBDIR}/pkgconfig) - -# -# copy the license files -# - -# Check if running in GitHub Actions -if(DEFINED ENV{GITHUB_ACTIONS} AND "$ENV{GITHUB_ACTIONS}" STREQUAL "true") - message(STATUS "Running inside GitHub Actions - copying license files") - - # Copy all files from licenses/ to build/bin/ - file(GLOB LICENSE_FILES "${CMAKE_SOURCE_DIR}/licenses/*") - foreach(LICENSE_FILE ${LICENSE_FILES}) - get_filename_component(FILENAME ${LICENSE_FILE} NAME) - configure_file(${LICENSE_FILE} "${CMAKE_BINARY_DIR}/bin/${FILENAME}" COPYONLY) - endforeach() -endif() - diff --git a/CMakePresets.json b/CMakePresets.json index 13bdd7907a..e984470130 100644 --- a/CMakePresets.json +++ b/CMakePresets.json @@ -38,15 +38,6 @@ } }, - { - "name": "arm64-windows-msvc", "hidden": true, - "architecture": { "value": "arm64", "strategy": "external" }, - "toolset": { "value": "host=x64", "strategy": "external" }, - "cacheVariables": { - "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-windows-msvc.cmake" - } - }, - { "name": "arm64-windows-llvm", "hidden": true, "architecture": { "value": "arm64", "strategy": "external" }, @@ -73,10 +64,6 @@ { "name": "arm64-apple-clang-release", "inherits": [ "base", "arm64-apple-clang", "reldbg" ] }, { "name": "arm64-apple-clang+static-release", "inherits": [ "base", "arm64-apple-clang", "reldbg", "static" ] }, - { "name": "arm64-windows-msvc-debug", "inherits": [ "base", "arm64-windows-msvc", "debug" ] }, - { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc", "reldbg" ] }, - { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc", "reldbg", "static" ] }, - { "name": "x64-windows-llvm-debug", "inherits": [ "base", "x64-windows-llvm", "debug" ] }, { "name": "x64-windows-llvm-release", "inherits": [ "base", "x64-windows-llvm", "release" ] }, { "name": "x64-windows-llvm-reldbg", "inherits": [ "base", "x64-windows-llvm", "reldbg" ] }, diff --git a/CODEOWNERS b/CODEOWNERS index 72d594b46e..3186f8eb1c 100644 --- a/CODEOWNERS +++ b/CODEOWNERS @@ -2,7 +2,7 @@ /ci/ @ggerganov /.devops/*.Dockerfile @ngxson -/examples/server/ @ngxson +/tools/server/ @ngxson /ggml/src/ggml-cuda/fattn* @JohannesGaessler /ggml/src/ggml-cuda/mmq.* @JohannesGaessler /ggml/src/ggml-cuda/mmv.* @JohannesGaessler diff --git a/Makefile b/Makefile index 772993ada2..958ad8f2fc 100644 --- a/Makefile +++ b/Makefile @@ -1156,10 +1156,10 @@ $(LIB_COMMON_S): $(OBJ_COMMON) # Clean generated server assets clean-server-assets: - find examples/server -type f -name "*.js.hpp" -delete - find examples/server -type f -name "*.mjs.hpp" -delete - find examples/server -type f -name "*.css.hpp" -delete - find examples/server -type f -name "*.html.hpp" -delete + find tools/server -type f -name "*.js.hpp" -delete + find tools/server -type f -name "*.mjs.hpp" -delete + find tools/server -type f -name "*.css.hpp" -delete + find tools/server -type f -name "*.html.hpp" -delete # Clean rule clean: clean-server-assets @@ -1179,7 +1179,7 @@ clean: clean-server-assets # Helper function that replaces .c, .cpp, and .cu file endings with .o: GET_OBJ_FILE = $(patsubst %.c,%.o,$(patsubst %.cpp,%.o,$(patsubst %.cu,%.o,$(1)))) -llama-cli: examples/main/main.cpp \ +llama-cli: tools/main/main.cpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) @@ -1187,12 +1187,7 @@ llama-cli: examples/main/main.cpp \ @echo '==== Run ./llama-cli -h for help. ====' @echo -llama-infill: examples/infill/infill.cpp \ - $(OBJ_ALL) - $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) - -llama-run: examples/run/run.cpp \ +llama-run: tools/run/run.cpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) @@ -1207,7 +1202,7 @@ llama-simple-chat: examples/simple-chat/simple-chat.cpp \ $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -llama-tokenize: examples/tokenize/tokenize.cpp \ +llama-tokenize: tools/tokenize/tokenize.cpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) @@ -1217,27 +1212,27 @@ llama-batched: examples/batched/batched.cpp \ $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -llama-batched-bench: examples/batched-bench/batched-bench.cpp \ +llama-batched-bench: tools/batched-bench/batched-bench.cpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -llama-quantize: examples/quantize/quantize.cpp \ +llama-quantize: tools/quantize/quantize.cpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -llama-quantize-stats: examples/quantize-stats/quantize-stats.cpp \ +llama-quantize-stats: tools/quantize-stats/quantize-stats.cpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -llama-perplexity: examples/perplexity/perplexity.cpp \ +llama-perplexity: tools/perplexity/perplexity.cpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -llama-imatrix: examples/imatrix/imatrix.cpp \ +llama-imatrix: tools/imatrix/imatrix.cpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) @@ -1279,7 +1274,7 @@ llama-gguf-hash: examples/gguf-hash/gguf-hash.cpp examples/gguf-hash/deps/sha1/s $(CXX) $(CXXFLAGS) -Iexamples/gguf-hash/deps -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -llama-gguf-split: examples/gguf-split/gguf-split.cpp \ +llama-gguf-split: tools/gguf-split/gguf-split.cpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) @@ -1289,7 +1284,7 @@ llama-eval-callback: examples/eval-callback/eval-callback.cpp \ $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -llama-cvector-generator: examples/cvector-generator/cvector-generator.cpp \ +llama-cvector-generator: tools/cvector-generator/cvector-generator.cpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) @@ -1299,12 +1294,12 @@ llama-convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c- $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -llama-bench: examples/llama-bench/llama-bench.cpp \ +llama-bench: tools/llama-bench/llama-bench.cpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -llama-export-lora: examples/export-lora/export-lora.cpp \ +llama-export-lora: tools/export-lora/export-lora.cpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) @@ -1360,17 +1355,17 @@ llama-gbnf-validator: examples/gbnf-validator/gbnf-validator.cpp \ $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) ifdef GGML_RPC -rpc-server: examples/rpc/rpc-server.cpp \ +rpc-server: tools/rpc/rpc-server.cpp \ $(OBJ_GGML) $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) endif # GGML_RPC llama-server: \ - examples/server/server.cpp \ - examples/server/utils.hpp \ - examples/server/httplib.h \ - examples/server/index.html.hpp \ - examples/server/loading.html.hpp \ + tools/server/server.cpp \ + tools/server/utils.hpp \ + tools/server/httplib.h \ + tools/server/index.html.hpp \ + tools/server/loading.html.hpp \ common/chat.cpp \ common/chat.h \ common/chat-template.hpp \ @@ -1378,10 +1373,10 @@ llama-server: \ common/minja.hpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2) + $(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Itools/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2) -# Portable equivalent of `cd examples/server/public && xxd -i $(notdir $<) ../$(notdir $<).hpp`: -examples/server/%.hpp: examples/server/public/% FORCE Makefile +# Portable equivalent of `cd tools/server/public && xxd -i $(notdir $<) ../$(notdir $<).hpp`: +tools/server/%.hpp: tools/server/public/% FORCE Makefile @( export NAME=$(subst .,_,$(subst -,_,$(notdir $<))) && \ echo "unsigned char $${NAME}[] = {" && \ cat $< | od -v -t x1 -An | sed -E 's/([0-9a-fA-F]+)/0x\1, /g' && \ @@ -1394,36 +1389,36 @@ llama-gen-docs: examples/gen-docs/gen-docs.cpp \ $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -libllava.a: examples/llava/llava.cpp \ - examples/llava/llava.h \ - examples/llava/clip.cpp \ - examples/llava/clip.h \ +libllava.a: tools/mtmd/llava.cpp \ + tools/mtmd/llava.h \ + tools/mtmd/clip.cpp \ + tools/mtmd/clip.h \ common/stb_image.h \ common/base64.hpp \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual -llama-llava-cli: examples/llava/llava-cli.cpp \ - examples/llava/llava.cpp \ - examples/llava/llava.h \ - examples/llava/clip.cpp \ - examples/llava/clip.h \ +llama-llava-cli: tools/mtmd/llava-cli.cpp \ + tools/mtmd/llava.cpp \ + tools/mtmd/llava.h \ + tools/mtmd/clip.cpp \ + tools/mtmd/clip.h \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual -llama-minicpmv-cli: examples/llava/minicpmv-cli.cpp \ - examples/llava/llava.cpp \ - examples/llava/llava.h \ - examples/llava/clip.cpp \ - examples/llava/clip.h \ +llama-minicpmv-cli: tools/mtmd/minicpmv-cli.cpp \ + tools/mtmd/llava.cpp \ + tools/mtmd/llava.h \ + tools/mtmd/clip.cpp \ + tools/mtmd/clip.h \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual -llama-qwen2vl-cli: examples/llava/qwen2vl-cli.cpp \ - examples/llava/llava.cpp \ - examples/llava/llava.h \ - examples/llava/clip.cpp \ - examples/llava/clip.h \ +llama-qwen2vl-cli: tools/mtmd/qwen2vl-cli.cpp \ + tools/mtmd/llava.cpp \ + tools/mtmd/llava.h \ + tools/mtmd/clip.cpp \ + tools/mtmd/clip.h \ $(OBJ_ALL) $(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual @@ -1480,12 +1475,12 @@ tests/test-double-float: tests/test-double-float.cpp tests/test-json-schema-to-grammar: tests/test-json-schema-to-grammar.cpp \ $(OBJ_ALL) - $(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) -Itools/server -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) tests/test-chat: tests/test-chat.cpp \ $(OBJ_ALL) - $(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) -Itools/server -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) tests/test-opt: tests/test-opt.cpp \ diff --git a/README.md b/README.md index a0e7bd2d21..0401723ffc 100644 --- a/README.md +++ b/README.md @@ -16,9 +16,10 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) ## Hot topics -- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli` and `gemma3-cli` https://github.com/ggml-org/llama.cpp/pull/13012, `libllava` will be deprecated -- **How to use [MTLResidencySet](https://developer.apple.com/documentation/metal/mtlresidencyset?language=objc) to keep the GPU memory active?** https://github.com/ggml-org/llama.cpp/pull/11427 -- **VS Code extension for FIM completions:** https://github.com/ggml-org/llama.vscode +- 🔥 Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md) +- **GGML developer experience survey (organized and reviewed by NVIDIA):** [link](https://forms.gle/Gasw3cRgyhNEnrwK9) +- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141](https://github.com/ggml-org/llama.cpp/pull/13141)), `libllava` will be deprecated +- VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode - Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim - Introducing GGUF-my-LoRA https://github.com/ggml-org/llama.cpp/discussions/10123 @@ -242,7 +243,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo | [Vulkan](docs/build.md#vulkan) | GPU | | [CANN](docs/build.md#cann) | Ascend NPU | | [OpenCL](docs/backend/OPENCL.md) | Adreno GPU | -| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/examples/rpc) | All | +| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All | ## Building the project @@ -276,9 +277,9 @@ The Hugging Face platform provides a variety of online tools for converting, qua - Use the [GGUF-editor space](https://huggingface.co/spaces/CISCai/gguf-editor) to edit GGUF meta data in the browser (more info: https://github.com/ggml-org/llama.cpp/discussions/9268) - Use the [Inference Endpoints](https://ui.endpoints.huggingface.co/) to directly host `llama.cpp` in the cloud (more info: https://github.com/ggml-org/llama.cpp/discussions/9669) -To learn more about model quantization, [read this documentation](examples/quantize/README.md) +To learn more about model quantization, [read this documentation](tools/quantize/README.md) -## [`llama-cli`](examples/main) +## [`llama-cli`](tools/main) #### A CLI tool for accessing and experimenting with most of `llama.cpp`'s functionality. @@ -341,7 +342,7 @@ To learn more about model quantization, [read this documentation](examples/quant -## [`llama-server`](examples/server) +## [`llama-server`](tools/server) #### A lightweight, [OpenAI API](https://github.com/openai/openai-openapi) compatible, HTTP server for serving LLMs. @@ -411,7 +412,7 @@ To learn more about model quantization, [read this documentation](examples/quant -## [`llama-perplexity`](examples/perplexity) +## [`llama-perplexity`](tools/perplexity) #### A tool for measuring the perplexity [^1][^2] (and other quality metrics) of a model over a given text. @@ -436,10 +437,10 @@ To learn more about model quantization, [read this documentation](examples/quant -[^1]: [examples/perplexity/README.md](./examples/perplexity/README.md) +[^1]: [tools/perplexity/README.md](./tools/perplexity/README.md) [^2]: [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity) -## [`llama-bench`](examples/llama-bench) +## [`llama-bench`](tools/llama-bench) #### Benchmark the performance of the inference for various parameters. @@ -460,7 +461,7 @@ To learn more about model quantization, [read this documentation](examples/quant -## [`llama-run`](examples/run) +## [`llama-run`](tools/run) #### A comprehensive example for running `llama.cpp` models. Useful for inferencing. Used with RamaLama [^3]. @@ -504,8 +505,8 @@ To learn more about model quantization, [read this documentation](examples/quant ## Other documentation -- [main (cli)](examples/main/README.md) -- [server](examples/server/README.md) +- [main (cli)](tools/main/README.md) +- [server](tools/server/README.md) - [GBNF grammars](grammars/README.md) #### Development documentation diff --git a/SECURITY.md b/SECURITY.md index 9370fb1a88..9749e95b71 100644 --- a/SECURITY.md +++ b/SECURITY.md @@ -40,7 +40,7 @@ To protect sensitive data from potential leaks or unauthorized access, it is cru ### Untrusted environments or networks If you can't run your models in a secure and isolated environment or if it must be exposed to an untrusted network, make sure to take the following security precautions: -* Do not use the RPC backend, [rpc-server](https://github.com/ggml-org/llama.cpp/tree/master/examples/rpc) and [llama-server](https://github.com/ggml-org/llama.cpp/tree/master/examples/server) functionality (see https://github.com/ggml-org/llama.cpp/pull/13061). +* Do not use the RPC backend, [rpc-server](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) and [llama-server](https://github.com/ggml-org/llama.cpp/tree/master/tools/server) functionality (see https://github.com/ggml-org/llama.cpp/pull/13061). * Confirm the hash of any downloaded artifact (e.g. pre-trained model weights) matches a known-good value. * Encrypt your data if sending it over the network. diff --git a/build-xcframework.sh b/build-xcframework.sh index 97001b5f7f..a08419a801 100755 --- a/build-xcframework.sh +++ b/build-xcframework.sh @@ -8,6 +8,7 @@ TVOS_MIN_OS_VERSION=16.4 BUILD_SHARED_LIBS=OFF LLAMA_BUILD_EXAMPLES=OFF +LLAMA_BUILD_TOOLS=OFF LLAMA_BUILD_TESTS=OFF LLAMA_BUILD_SERVER=OFF GGML_METAL=ON @@ -31,6 +32,7 @@ COMMON_CMAKE_ARGS=( -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml -DBUILD_SHARED_LIBS=${BUILD_SHARED_LIBS} -DLLAMA_BUILD_EXAMPLES=${LLAMA_BUILD_EXAMPLES} + -DLLAMA_BUILD_TOOLS=${LLAMA_BUILD_TOOLS} -DLLAMA_BUILD_TESTS=${LLAMA_BUILD_TESTS} -DLLAMA_BUILD_SERVER=${LLAMA_BUILD_SERVER} -DGGML_METAL_EMBED_LIBRARY=${GGML_METAL_EMBED_LIBRARY} @@ -115,6 +117,7 @@ setup_framework_structure() { # Copy all required headers (common for all platforms) cp include/llama.h ${header_path} cp ggml/include/ggml.h ${header_path} + cp ggml/include/ggml-opt.h ${header_path} cp ggml/include/ggml-alloc.h ${header_path} cp ggml/include/ggml-backend.h ${header_path} cp ggml/include/ggml-metal.h ${header_path} diff --git a/ci/run.sh b/ci/run.sh index f463d7a8b2..b49a3a5f82 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -187,8 +187,8 @@ function gg_run_test_scripts_debug { set -e - (cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log - (cd ./examples/quantize && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log + (cd ./tools/gguf-split && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log + (cd ./tools/quantize && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log set +e } @@ -211,8 +211,8 @@ function gg_run_test_scripts_release { set -e - (cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log - (cd ./examples/quantize && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log + (cd ./tools/gguf-split && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log + (cd ./tools/quantize && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log set +e } diff --git a/cmake/arm64-windows-msvc.cmake b/cmake/arm64-windows-msvc.cmake deleted file mode 100644 index c77631420c..0000000000 --- a/cmake/arm64-windows-msvc.cmake +++ /dev/null @@ -1,6 +0,0 @@ -set( CMAKE_SYSTEM_NAME Windows ) -set( CMAKE_SYSTEM_PROCESSOR arm64 ) - -set( target arm64-pc-windows-msvc ) -set( CMAKE_C_COMPILER_TARGET ${target} ) -set( CMAKE_CXX_COMPILER_TARGET ${target} ) diff --git a/cmake/build-info.cmake b/cmake/build-info.cmake index c1a456e179..75c78222f2 100644 --- a/cmake/build-info.cmake +++ b/cmake/build-info.cmake @@ -41,14 +41,20 @@ endif() if(MSVC) set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}") - set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME}) + if (CMAKE_VS_PLATFORM_NAME) + set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME}) + else() + set(BUILD_TARGET "${CMAKE_SYSTEM_NAME} ${CMAKE_SYSTEM_PROCESSOR}") + endif() else() execute_process( - COMMAND sh -c "\"$@\" --version | head -1" _ ${CMAKE_C_COMPILER} + COMMAND ${CMAKE_C_COMPILER} --version OUTPUT_VARIABLE OUT OUTPUT_STRIP_TRAILING_WHITESPACE ) + string(REGEX REPLACE " *\n.*" "" OUT "${OUT}") set(BUILD_COMPILER ${OUT}) + execute_process( COMMAND ${CMAKE_C_COMPILER} -dumpmachine OUTPUT_VARIABLE OUT diff --git a/cmake/x64-windows-llvm.cmake b/cmake/x64-windows-llvm.cmake index 0603d738fb..77e7914079 100644 --- a/cmake/x64-windows-llvm.cmake +++ b/cmake/x64-windows-llvm.cmake @@ -3,9 +3,3 @@ set( CMAKE_SYSTEM_PROCESSOR x86_64 ) set( CMAKE_C_COMPILER clang ) set( CMAKE_CXX_COMPILER clang++ ) - -set( arch_c_flags "-march=native" ) - -set( CMAKE_C_FLAGS_INIT "${arch_c_flags}" ) -set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags}" ) - diff --git a/common/CMakeLists.txt b/common/CMakeLists.txt index 43533fc86a..6b0011e4df 100644 --- a/common/CMakeLists.txt +++ b/common/CMakeLists.txt @@ -39,7 +39,9 @@ add_custom_command( COMMENT "Generating build details from Git" COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION} -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME} - -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake" + -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} + -DCMAKE_SYSTEM_NAME=${CMAKE_SYSTEM_NAME} -DCMAKE_SYSTEM_PROCESSOR=${CMAKE_SYSTEM_PROCESSOR} + -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake" WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.." DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX} VERBATIM @@ -117,8 +119,8 @@ if (LLAMA_LLGUIDANCE) ExternalProject_Add(llguidance_ext GIT_REPOSITORY https://github.com/guidance-ai/llguidance - # v0.7.10: - GIT_TAG 0309d2a6bf40abda35344a362edc71e06d5009f8 + # v0.7.19 (+ fancy-regex build fix): + GIT_TAG b59f98f85269892a7de3d3641ad155366f13daa6 PREFIX ${CMAKE_BINARY_DIR}/llguidance SOURCE_DIR ${LLGUIDANCE_SRC} BUILD_IN_SOURCE TRUE @@ -142,3 +144,27 @@ endif () target_include_directories(${TARGET} PUBLIC .) target_compile_features (${TARGET} PUBLIC cxx_std_17) target_link_libraries (${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads) + + +# +# copy the license files +# + +# Check if running in GitHub Actions +if (DEFINED ENV{GITHUB_ACTIONS} AND "$ENV{GITHUB_ACTIONS}" STREQUAL "true") + message(STATUS "Running inside GitHub Actions - copying license files") + + # Copy all files from licenses/ to build/bin/ + file(GLOB LICENSE_FILES "${CMAKE_SOURCE_DIR}/licenses/*") + foreach(LICENSE_FILE ${LICENSE_FILES}) + get_filename_component(FILENAME ${LICENSE_FILE} NAME) + add_custom_command( + POST_BUILD + TARGET ${TARGET} + COMMAND ${CMAKE_COMMAND} -E copy_if_different + "${LICENSE_FILE}" + "$/${FILENAME}" + COMMENT "Copying ${FILENAME} to ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}") + message(STATUS "Copying ${LICENSE_FILE} to ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${FILENAME}") + endforeach() +endif() diff --git a/common/arg.cpp b/common/arg.cpp index 0657553e4e..a1fd4c9651 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -40,9 +40,28 @@ using json = nlohmann::ordered_json; std::initializer_list mmproj_examples = { LLAMA_EXAMPLE_LLAVA, - // TODO: add LLAMA_EXAMPLE_SERVER when it's ready + LLAMA_EXAMPLE_SERVER, }; +static std::string read_file(const std::string & fname) { + std::ifstream file(fname); + if (!file) { + throw std::runtime_error(string_format("error: failed to open file '%s'\n", fname.c_str())); + } + std::string content((std::istreambuf_iterator(file)), std::istreambuf_iterator()); + file.close(); + return content; +} + +static void write_file(const std::string & fname, const std::string & content) { + std::ofstream file(fname); + if (!file) { + throw std::runtime_error(string_format("error: failed to open file '%s'\n", fname.c_str())); + } + file << content; + file.close(); +} + common_arg & common_arg::set_examples(std::initializer_list examples) { this->examples = std::move(examples); return *this; @@ -162,6 +181,10 @@ struct common_hf_file_res { #ifdef LLAMA_USE_CURL +bool common_has_curl() { + return true; +} + #ifdef __linux__ #include #elif defined(_WIN32) @@ -194,11 +217,11 @@ struct curl_slist_ptr { #define CURL_MAX_RETRY 3 #define CURL_RETRY_DELAY_SECONDS 2 -static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) { +static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds, const char * method_name) { int remaining_attempts = max_attempts; while (remaining_attempts > 0) { - LOG_INF("%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts); + LOG_INF("%s: %s %s (attempt %d of %d)...\n", __func__ , method_name, url.c_str(), max_attempts - remaining_attempts + 1, max_attempts); CURLcode res = curl_easy_perform(curl); if (res == CURLE_OK) { @@ -209,6 +232,7 @@ static bool curl_perform_with_retry(const std::string & url, CURL * curl, int ma LOG_WRN("%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay); remaining_attempts--; + if (remaining_attempts == 0) break; std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay)); } @@ -227,8 +251,6 @@ static bool common_download_file_single(const std::string & url, const std::stri return false; } - bool force_download = false; - // Set the URL, allow to follow http redirection curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str()); curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L); @@ -252,7 +274,7 @@ static bool common_download_file_single(const std::string & url, const std::stri // If the file exists, check its JSON metadata companion file. std::string metadata_path = path + ".json"; - nlohmann::json metadata; + nlohmann::json metadata; // TODO @ngxson : get rid of this json, use regex instead std::string etag; std::string last_modified; @@ -262,14 +284,7 @@ static bool common_download_file_single(const std::string & url, const std::stri if (metadata_in.good()) { try { metadata_in >> metadata; - LOG_INF("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str()); - if (metadata.contains("url") && metadata.at("url").is_string()) { - auto previous_url = metadata.at("url").get(); - if (previous_url != url) { - LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str()); - return false; - } - } + LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str()); if (metadata.contains("etag") && metadata.at("etag").is_string()) { etag = metadata.at("etag"); } @@ -277,10 +292,10 @@ static bool common_download_file_single(const std::string & url, const std::stri last_modified = metadata.at("lastModified"); } } catch (const nlohmann::json::exception & e) { - LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what()); - return false; + LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what()); } } + // if we cannot open the metadata file, we assume that the downloaded file is not valid (etag and last-modified are left empty, so we will download it again) } else { LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str()); } @@ -292,7 +307,10 @@ static bool common_download_file_single(const std::string & url, const std::stri }; common_load_model_from_url_headers headers; + bool head_request_ok = false; + bool should_download = !file_exists; // by default, we should download if the file does not exist + // get ETag to see if the remote file has changed { typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *); auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t { @@ -321,23 +339,28 @@ static bool common_download_file_single(const std::string & url, const std::stri curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast(header_callback)); curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers); - bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS); + // we only allow retrying once for HEAD requests + // this is for the use case of using running offline (no internet), retrying can be annoying + bool was_perform_successful = curl_perform_with_retry(url, curl.get(), 1, 0, "HEAD"); if (!was_perform_successful) { - return false; + head_request_ok = false; } long http_code = 0; curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code); - if (http_code != 200) { - // HEAD not supported, we don't know if the file has changed - // force trigger downloading - force_download = true; - LOG_ERR("%s: HEAD invalid http status code received: %ld\n", __func__, http_code); + if (http_code == 200) { + head_request_ok = true; + } else { + LOG_WRN("%s: HEAD invalid http status code received: %ld\n", __func__, http_code); + head_request_ok = false; } } - bool should_download = !file_exists || force_download; - if (!should_download) { + // if head_request_ok is false, we don't have the etag or last-modified headers + // we leave should_download as-is, which is true if the file does not exist + if (head_request_ok) { + // check if ETag or Last-Modified headers are different + // if it is, we need to download the file again if (!etag.empty() && etag != headers.etag) { LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str()); should_download = true; @@ -346,6 +369,7 @@ static bool common_download_file_single(const std::string & url, const std::stri should_download = true; } } + if (should_download) { std::string path_temporary = path + ".downloadInProgress"; if (file_exists) { @@ -399,7 +423,7 @@ static bool common_download_file_single(const std::string & url, const std::stri // start the download LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__, llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str()); - bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS); + bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS, "GET"); if (!was_perform_successful) { return false; } @@ -420,13 +444,15 @@ static bool common_download_file_single(const std::string & url, const std::stri {"etag", headers.etag}, {"lastModified", headers.last_modified} }); - std::ofstream(metadata_path) << metadata.dump(4); - LOG_INF("%s: file metadata saved: %s\n", __func__, metadata_path.c_str()); + write_file(metadata_path, metadata.dump(4)); + LOG_DBG("%s: file metadata saved: %s\n", __func__, metadata_path.c_str()); if (rename(path_temporary.c_str(), path.c_str()) != 0) { LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str()); return false; } + } else { + LOG_INF("%s: using cached file: %s\n", __func__, path.c_str()); } return true; @@ -527,6 +553,50 @@ static bool common_download_model( return true; } +std::pair> common_remote_get_content(const std::string & url, const common_remote_params & params) { + curl_ptr curl(curl_easy_init(), &curl_easy_cleanup); + curl_slist_ptr http_headers; + std::vector res_buffer; + + curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str()); + curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); + curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L); + typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data); + auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t { + auto data_vec = static_cast *>(data); + data_vec->insert(data_vec->end(), (char *)ptr, (char *)ptr + size * nmemb); + return size * nmemb; + }; + curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast(write_callback)); + curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_buffer); +#if defined(_WIN32) + curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA); +#endif + if (params.timeout > 0) { + curl_easy_setopt(curl.get(), CURLOPT_TIMEOUT, params.timeout); + } + if (params.max_size > 0) { + curl_easy_setopt(curl.get(), CURLOPT_MAXFILESIZE, params.max_size); + } + http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp"); + for (const auto & header : params.headers) { + http_headers.ptr = curl_slist_append(http_headers.ptr, header.c_str()); + } + curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr); + + CURLcode res = curl_easy_perform(curl.get()); + + if (res != CURLE_OK) { + std::string error_msg = curl_easy_strerror(res); + throw std::runtime_error("error: cannot make GET request: " + error_msg); + } + + long res_code; + curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code); + + return { res_code, std::move(res_buffer) }; +} + /** * Allow getting the HF file from the HF repo with tag (like ollama), for example: * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4 @@ -546,46 +616,48 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_ throw std::invalid_argument("error: invalid HF repo format, expected /[:quant]\n"); } - // fetch model info from Hugging Face Hub API - curl_ptr curl(curl_easy_init(), &curl_easy_cleanup); - curl_slist_ptr http_headers; - std::string res_str; + std::string url = get_model_endpoint() + "v2/" + hf_repo + "/manifests/" + tag; - std::string model_endpoint = get_model_endpoint(); - - std::string url = model_endpoint + "v2/" + hf_repo + "/manifests/" + tag; - curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str()); - curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); - typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data); - auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t { - static_cast(data)->append((char * ) ptr, size * nmemb); - return size * nmemb; - }; - curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast(write_callback)); - curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str); -#if defined(_WIN32) - curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA); -#endif + // headers + std::vector headers; + headers.push_back("Accept: application/json"); if (!bearer_token.empty()) { - std::string auth_header = "Authorization: Bearer " + bearer_token; - http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str()); + headers.push_back("Authorization: Bearer " + bearer_token); } // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response - http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp"); - http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json"); - curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr); + // User-Agent header is already set in common_remote_get_content, no need to set it here - CURLcode res = curl_easy_perform(curl.get()); + // we use "=" to avoid clashing with other component, while still being allowed on windows + std::string cached_response_fname = "manifest=" + hf_repo + "=" + tag + ".json"; + string_replace_all(cached_response_fname, "/", "_"); + std::string cached_response_path = fs_get_cache_file(cached_response_fname); - if (res != CURLE_OK) { - throw std::runtime_error("error: cannot make GET request to HF API"); + // make the request + common_remote_params params; + params.headers = headers; + long res_code = 0; + std::string res_str; + bool use_cache = false; + try { + auto res = common_remote_get_content(url, params); + res_code = res.first; + res_str = std::string(res.second.data(), res.second.size()); + } catch (const std::exception & e) { + LOG_WRN("error: failed to get manifest: %s\n", e.what()); + LOG_WRN("try reading from cache\n"); + // try to read from cache + try { + res_str = read_file(cached_response_path); + res_code = 200; + use_cache = true; + } catch (const std::exception & e) { + throw std::runtime_error("error: failed to get manifest (check your internet connection)"); + } } + std::string ggufFile; + std::string mmprojFile; - long res_code; - std::string ggufFile = ""; - std::string mmprojFile = ""; - curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code); - if (res_code == 200) { + if (res_code == 200 || res_code == 304) { // extract ggufFile.rfilename in json, using regex { std::regex pattern("\"ggufFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\""); @@ -602,6 +674,10 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_ mmprojFile = match[1].str(); } } + if (!use_cache) { + // if not using cached response, update the cache file + write_file(cached_response_path, res_str); + } } else if (res_code == 401) { throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token"); } else { @@ -618,6 +694,10 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_ #else +bool common_has_curl() { + return false; +} + static bool common_download_file_single(const std::string &, const std::string &, const std::string &) { LOG_ERR("error: built without CURL, cannot download model from internet\n"); return false; @@ -640,6 +720,14 @@ static struct common_hf_file_res common_get_hf_file(const std::string &, const s return {}; } +std::pair> common_remote_get_content(const std::string & url, const common_remote_params &) { + if (!url.empty()) { + throw std::runtime_error("error: built without CURL, cannot download model from the internet"); + } + + return {}; +} + #endif // LLAMA_USE_CURL // @@ -1101,6 +1189,9 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e fprintf(stderr, "%s\n", ex.what()); ctx_arg.params = params_org; return false; + } catch (std::exception & ex) { + fprintf(stderr, "%s\n", ex.what()); + exit(1); // for other exceptions, we exit with status code 1 } return true; @@ -1192,7 +1283,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params) { params.use_color = true; } - ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL, LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP})); + ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP})); add_opt(common_arg( {"-t", "--threads"}, "N", string_format("number of threads to use during generation (default: %d)", params.cpuparams.n_threads), @@ -1325,7 +1416,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex add_opt(common_arg( {"-n", "--predict", "--n-predict"}, "N", string_format( - ex == LLAMA_EXAMPLE_MAIN || ex == LLAMA_EXAMPLE_INFILL + ex == LLAMA_EXAMPLE_MAIN ? "number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)" : "number of tokens to predict (default: %d, -1 = infinity)", params.n_predict), @@ -1401,13 +1492,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex {"-f", "--file"}, "FNAME", "a file containing the prompt (default: none)", [](common_params & params, const std::string & value) { - std::ifstream file(value); - if (!file) { - throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str())); - } + params.prompt = read_file(value); // store the external file name in params params.prompt_file = value; - std::copy(std::istreambuf_iterator(file), std::istreambuf_iterator(), back_inserter(params.prompt)); if (!params.prompt.empty() && params.prompt.back() == '\n') { params.prompt.pop_back(); } @@ -1417,11 +1504,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex {"-sysf", "--system-prompt-file"}, "FNAME", "a file containing the system prompt (default: none)", [](common_params & params, const std::string & value) { - std::ifstream file(value); - if (!file) { - throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str())); - } - std::copy(std::istreambuf_iterator(file), std::istreambuf_iterator(), back_inserter(params.system_prompt)); + params.system_prompt = read_file(value); if (!params.system_prompt.empty() && params.system_prompt.back() == '\n') { params.system_prompt.pop_back(); } @@ -1572,7 +1655,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.input_prefix = value; params.enable_chat_template = false; } - ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL})); + ).set_examples({LLAMA_EXAMPLE_MAIN})); add_opt(common_arg( {"--in-suffix"}, "STRING", "string to suffix after user inputs with (default: empty)", @@ -1580,7 +1663,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.input_suffix = value; params.enable_chat_template = false; } - ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL})); + ).set_examples({LLAMA_EXAMPLE_MAIN})); add_opt(common_arg( {"--no-warmup"}, "skip warming up the model with an empty run", @@ -1597,7 +1680,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params) { params.spm_infill = true; } - ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_INFILL})); + ).set_examples({LLAMA_EXAMPLE_SERVER})); add_opt(common_arg( {"--samplers"}, "SAMPLERS", string_format("samplers that will be used for generation in the order, separated by \';\'\n(default: %s)", sampler_type_names.c_str()), @@ -1846,15 +1929,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex {"--grammar-file"}, "FNAME", "file to read grammar from", [](common_params & params, const std::string & value) { - std::ifstream file(value); - if (!file) { - throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str())); - } - std::copy( - std::istreambuf_iterator(file), - std::istreambuf_iterator(), - std::back_inserter(params.sampling.grammar) - ); + params.sampling.grammar = read_file(value); } ).set_sparam()); add_opt(common_arg( @@ -1864,6 +1939,23 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.sampling.grammar = json_schema_to_grammar(json::parse(value)); } ).set_sparam()); + add_opt(common_arg( + {"-jf", "--json-schema-file"}, "FILE", + "File containing a JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object\nFor schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead", + [](common_params & params, const std::string & value) { + std::ifstream file(value); + if (!file) { + throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str())); + } + std::string schema; + std::copy( + std::istreambuf_iterator(file), + std::istreambuf_iterator(), + std::back_inserter(schema) + ); + params.sampling.grammar = json_schema_to_grammar(json::parse(schema)); + } + ).set_sparam()); add_opt(common_arg( {"--pooling"}, "{none,mean,cls,last,rank}", "pooling type for embeddings, use model default if unspecified", @@ -2005,13 +2097,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.cache_type_v = kv_cache_type_from_str(value); } ).set_env("LLAMA_ARG_CACHE_TYPE_V")); - add_opt(common_arg( - {"--perplexity", "--all-logits"}, - string_format("return logits for all tokens in the batch (default: %s)", params.logits_all ? "true" : "false"), - [](common_params & params) { - params.logits_all = true; - } - ).set_examples({LLAMA_EXAMPLE_PERPLEXITY})); add_opt(common_arg( {"--hellaswag"}, "compute HellaSwag score over random tasks from datafile supplied with -f", @@ -2119,32 +2204,33 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_CONT_BATCHING")); add_opt(common_arg( {"--mmproj"}, "FILE", - "path to a multimodal projector file. see examples/llava/README.md", + "path to a multimodal projector file. see tools/mtmd/README.md\n" + "note: if -hf is used, this argument can be omitted", [](common_params & params, const std::string & value) { params.mmproj.path = value; } - ).set_examples(mmproj_examples)); + ).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ")); add_opt(common_arg( {"--mmproj-url"}, "URL", - "URL to a multimodal projector file. see examples/llava/README.md", + "URL to a multimodal projector file. see tools/mtmd/README.md", [](common_params & params, const std::string & value) { params.mmproj.url = value; } - ).set_examples(mmproj_examples)); + ).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ_URL")); add_opt(common_arg( {"--no-mmproj"}, "explicitly disable multimodal projector, useful when using -hf", [](common_params & params) { params.no_mmproj = true; } - ).set_examples(mmproj_examples)); + ).set_examples(mmproj_examples).set_env("LLAMA_ARG_NO_MMPROJ")); add_opt(common_arg( {"--no-mmproj-offload"}, "do not offload multimodal projector to GPU", [](common_params & params) { params.mmproj_use_gpu = false; } - ).set_examples(mmproj_examples)); + ).set_examples(mmproj_examples).set_env("LLAMA_ARG_NO_MMPROJ_OFFLOAD")); add_opt(common_arg( {"--image"}, "FILE", "path to an image file. use with multimodal models. Specify multiple times for batching", @@ -2351,6 +2437,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex } } )); + add_opt(common_arg( + {"--no-op-offload"}, + string_format("disable offloading host tensor operations to device (default: %s)", params.no_op_offload ? "true" : "false"), + [](common_params & params) { + params.no_op_offload = true; + } + )); add_opt(common_arg( {"--lora"}, "FNAME", "path to LoRA adapter (can be repeated to use multiple adapters)", @@ -2542,6 +2635,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.i_chunk = value; } ).set_examples({LLAMA_EXAMPLE_IMATRIX})); + add_opt(common_arg( + {"--parse-special"}, + string_format("prase special tokens (chat, tool, etc) (default: %s)", params.parse_special ? "true" : "false"), + [](common_params & params) { + params.parse_special = true; + } + ).set_examples({LLAMA_EXAMPLE_IMATRIX})); add_opt(common_arg( {"-pps"}, string_format("is the prompt shared across parallel sequences (default: %s)", params.is_pp_shared ? "true" : "false"), @@ -2691,7 +2791,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_THREADS_HTTP")); add_opt(common_arg( {"--cache-reuse"}, "N", - string_format("min chunk size to attempt reusing from the cache via KV shifting (default: %d)", params.n_cache_reuse), + string_format( + "min chunk size to attempt reusing from the cache via KV shifting (default: %d)\n" + "[(card)](https://ggml.ai/f0.png)", params.n_cache_reuse + ), [](common_params & params, int value) { params.n_cache_reuse = value; } @@ -2774,14 +2877,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex "list of built-in templates:\n%s", list_builtin_chat_templates().c_str() ), [](common_params & params, const std::string & value) { - std::ifstream file(value); - if (!file) { - throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str())); - } - std::copy( - std::istreambuf_iterator(file), - std::istreambuf_iterator(), - std::back_inserter(params.chat_template)); + params.chat_template = read_file(value); } ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE_FILE")); add_opt(common_arg( @@ -2804,7 +2900,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params) { params.simple_io = true; } - ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_INFILL})); + ).set_examples({LLAMA_EXAMPLE_MAIN})); add_opt(common_arg( {"--positive-file"}, "FNAME", string_format("positive prompts file, one prompt per line (default: '%s')", params.cvector_positive_file.c_str()), diff --git a/common/arg.h b/common/arg.h index 49ab8667b1..70bea100fd 100644 --- a/common/arg.h +++ b/common/arg.h @@ -78,3 +78,12 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e // function to be used by test-arg-parser common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr); +bool common_has_curl(); + +struct common_remote_params { + std::vector headers; + long timeout = 0; // CURLOPT_TIMEOUT, in seconds ; 0 means no timeout + long max_size = 0; // max size of the response ; unlimited if 0 ; max is 2GB +}; +// get remote file content, returns +std::pair> common_remote_get_content(const std::string & url, const common_remote_params & params); diff --git a/common/chat.cpp b/common/chat.cpp index bbc5f087cd..ad3d4aa99a 100644 --- a/common/chat.cpp +++ b/common/chat.cpp @@ -125,7 +125,9 @@ std::vector common_chat_msgs_parse_oaicompat(const json & messa msgs.push_back(msg); } } catch (const std::exception & e) { - throw std::runtime_error("Failed to parse messages: " + std::string(e.what()) + "; messages = " + messages.dump(2)); + // @ngxson : disable otherwise it's bloating the API response + // printf("%s\n", std::string("; messages = ") + messages.dump(2)); + throw std::runtime_error("Failed to parse messages: " + std::string(e.what())); } return msgs; diff --git a/common/common.cpp b/common/common.cpp index 94f545f815..2b1d8da592 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1096,7 +1096,6 @@ struct llama_context_params common_context_params_to_llama(const common_params & cparams.n_threads = params.cpuparams.n_threads; cparams.n_threads_batch = params.cpuparams_batch.n_threads == -1 ? params.cpuparams.n_threads : params.cpuparams_batch.n_threads; - cparams.logits_all = params.logits_all; cparams.embeddings = params.embedding; cparams.rope_scaling_type = params.rope_scaling_type; cparams.rope_freq_base = params.rope_freq_base; @@ -1114,6 +1113,7 @@ struct llama_context_params common_context_params_to_llama(const common_params & cparams.offload_kqv = !params.no_kv_offload; cparams.flash_attn = params.flash_attn; cparams.no_perf = params.no_perf; + cparams.op_offload = !params.no_op_offload; if (params.reranking) { cparams.embeddings = true; @@ -1565,3 +1565,20 @@ common_control_vector_data common_control_vector_load(const std::vector & tokens, int64_t stride) { + const int64_t ne_datapoint = llama_n_ctx(ctx); + const int64_t ndata = (tokens.size() - ne_datapoint - 1) / stride; + ggml_opt_dataset_t result = ggml_opt_dataset_init( + GGML_TYPE_I32, GGML_TYPE_I32, ne_datapoint, ne_datapoint, ndata, /*ndata_shard =*/ 1); + + llama_token * data = (llama_token *) ggml_opt_dataset_data(result)->data; + llama_token * labels = (llama_token *) ggml_opt_dataset_labels(result)->data; + + for (int64_t idata = 0; idata < ndata; ++idata) { + memcpy(data + idata*ne_datapoint, tokens.data() + idata*stride + 0, ne_datapoint*sizeof(llama_token)); + memcpy(labels + idata*ne_datapoint, tokens.data() + idata*stride + 1, ne_datapoint*sizeof(llama_token)); + } + + return result; +} diff --git a/common/common.h b/common/common.h index 0a9dc0599f..dea34267c8 100644 --- a/common/common.h +++ b/common/common.h @@ -66,7 +66,6 @@ enum llama_example { LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_MAIN, - LLAMA_EXAMPLE_INFILL, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_RETRIEVAL, @@ -96,6 +95,7 @@ enum common_sampler_type { COMMON_SAMPLER_TYPE_XTC = 8, COMMON_SAMPLER_TYPE_INFILL = 9, COMMON_SAMPLER_TYPE_PENALTIES = 10, + COMMON_SAMPLER_TYPE_TOP_N_SIGMA = 11, }; // dimensionality reduction methods, used by cvector-generator @@ -161,6 +161,7 @@ struct common_params_sampling { std::vector samplers = { COMMON_SAMPLER_TYPE_PENALTIES, COMMON_SAMPLER_TYPE_DRY, + COMMON_SAMPLER_TYPE_TOP_N_SIGMA, COMMON_SAMPLER_TYPE_TOP_K, COMMON_SAMPLER_TYPE_TYPICAL_P, COMMON_SAMPLER_TYPE_TOP_P, @@ -323,7 +324,6 @@ struct common_params { bool ctx_shift = true; // context shift on inifinite text generation bool input_prefix_bos = false; // prefix BOS to user inputs, preceding input_prefix - bool logits_all = false; // return logits for all tokens in the batch bool use_mmap = true; // use mmap for faster loads bool use_mlock = false; // use mlock to keep model in memory bool verbose_prompt = false; // print prompt tokens before generation @@ -332,6 +332,7 @@ struct common_params { bool no_kv_offload = false; // disable KV offloading bool warmup = true; // warmup run bool check_tensors = false; // validate tensor data + bool no_op_offload = false; // globally disable offload host tensor operations to device bool single_turn = false; // single turn chat conversation @@ -340,7 +341,7 @@ struct common_params { common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO; - // multimodal models (see examples/llava) + // multimodal models (see tools/mtmd) struct common_params_model mmproj; bool mmproj_use_gpu = true; // use GPU for multimodal model bool no_mmproj = false; // explicitly disable multimodal model @@ -409,13 +410,14 @@ struct common_params { bool process_output = false; // collect data for the output tensor bool compute_ppl = true; // whether to compute perplexity + bool parse_special = false; // whether to parse special tokens during imatrix tokenization // cvector-generator params int n_pca_batch = 100; int n_pca_iterations = 1000; dimre_method cvector_dimre_method = DIMRE_METHOD_PCA; - std::string cvector_positive_file = "examples/cvector-generator/positive.txt"; - std::string cvector_negative_file = "examples/cvector-generator/negative.txt"; + std::string cvector_positive_file = "tools/cvector-generator/positive.txt"; + std::string cvector_negative_file = "tools/cvector-generator/negative.txt"; bool spm_infill = false; // suffix/prefix/middle pattern for infill @@ -664,3 +666,9 @@ const char * const LLM_KV_SPLIT_COUNT = "split.count"; const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count"; } + +// +// training utils +// + +ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector & tokens, int64_t stride); diff --git a/common/json-schema-to-grammar.cpp b/common/json-schema-to-grammar.cpp index 9067982257..5b3059c2f7 100644 --- a/common/json-schema-to-grammar.cpp +++ b/common/json-schema-to-grammar.cpp @@ -16,6 +16,9 @@ using json = nlohmann::ordered_json; static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") { auto has_max = max_items != std::numeric_limits::max(); + if (max_items == 0) { + return ""; + } if (min_items == 0 && max_items == 1) { return item_rule + "?"; } diff --git a/common/llguidance.cpp b/common/llguidance.cpp index 8bff89ea4a..adce620e4d 100644 --- a/common/llguidance.cpp +++ b/common/llguidance.cpp @@ -189,6 +189,7 @@ static LlgTokenizer * llama_sampler_llg_new_tokenizer(const llama_vocab * vocab) /* .tokenize_fn = */ llama_sampler_llg_tokenize_fn, /* .use_approximate_greedy_tokenize_fn = */ false, /* .tokenize_user_data = */ vocab, + /* .slices = */ nullptr, }; char error_buffer[1024]; diff --git a/common/sampling.cpp b/common/sampling.cpp index 1735b65018..28705e24c0 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -1,6 +1,7 @@ #include "sampling.h" #include "common.h" +#include "log.h" #include #include @@ -229,51 +230,48 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co params.logit_bias.data())); if (params.mirostat == 0) { - if (params.top_n_sigma >= 0) { - llama_sampler_chain_add(result->chain, llama_sampler_init_top_k (params.top_k)); - llama_sampler_chain_add(result->chain, llama_sampler_init_temp (params.temp)); - llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma (params.top_n_sigma)); - } else { - for (const auto & cnstr : params.samplers) { - switch (cnstr) { - case COMMON_SAMPLER_TYPE_DRY: - { - std::vector c_breakers; - c_breakers.reserve(params.dry_sequence_breakers.size()); - for (const auto & str : params.dry_sequence_breakers) { - c_breakers.push_back(str.c_str()); - } - - llama_sampler_chain_add(result->chain, llama_sampler_init_dry (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size())); + for (const auto & cnstr : params.samplers) { + switch (cnstr) { + case COMMON_SAMPLER_TYPE_DRY: + { + std::vector c_breakers; + c_breakers.reserve(params.dry_sequence_breakers.size()); + for (const auto & str : params.dry_sequence_breakers) { + c_breakers.push_back(str.c_str()); } - break; - case COMMON_SAMPLER_TYPE_TOP_K: - llama_sampler_chain_add(result->chain, llama_sampler_init_top_k (params.top_k)); - break; - case COMMON_SAMPLER_TYPE_TOP_P: - llama_sampler_chain_add(result->chain, llama_sampler_init_top_p (params.top_p, params.min_keep)); - break; - case COMMON_SAMPLER_TYPE_MIN_P: - llama_sampler_chain_add(result->chain, llama_sampler_init_min_p (params.min_p, params.min_keep)); - break; - case COMMON_SAMPLER_TYPE_XTC: - llama_sampler_chain_add(result->chain, llama_sampler_init_xtc (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed)); - break; - case COMMON_SAMPLER_TYPE_TYPICAL_P: - llama_sampler_chain_add(result->chain, llama_sampler_init_typical (params.typ_p, params.min_keep)); - break; - case COMMON_SAMPLER_TYPE_TEMPERATURE: - llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent)); - break; - case COMMON_SAMPLER_TYPE_INFILL: - llama_sampler_chain_add(result->chain, llama_sampler_init_infill (vocab)); - break; - case COMMON_SAMPLER_TYPE_PENALTIES: - llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present)); - break; - default: - GGML_ASSERT(false && "unknown sampler type"); - } + + llama_sampler_chain_add(result->chain, llama_sampler_init_dry (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size())); + } + break; + case COMMON_SAMPLER_TYPE_TOP_K: + llama_sampler_chain_add(result->chain, llama_sampler_init_top_k (params.top_k)); + break; + case COMMON_SAMPLER_TYPE_TOP_P: + llama_sampler_chain_add(result->chain, llama_sampler_init_top_p (params.top_p, params.min_keep)); + break; + case COMMON_SAMPLER_TYPE_TOP_N_SIGMA: + llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma (params.top_n_sigma)); + break; + case COMMON_SAMPLER_TYPE_MIN_P: + llama_sampler_chain_add(result->chain, llama_sampler_init_min_p (params.min_p, params.min_keep)); + break; + case COMMON_SAMPLER_TYPE_XTC: + llama_sampler_chain_add(result->chain, llama_sampler_init_xtc (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed)); + break; + case COMMON_SAMPLER_TYPE_TYPICAL_P: + llama_sampler_chain_add(result->chain, llama_sampler_init_typical (params.typ_p, params.min_keep)); + break; + case COMMON_SAMPLER_TYPE_TEMPERATURE: + llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent)); + break; + case COMMON_SAMPLER_TYPE_INFILL: + llama_sampler_chain_add(result->chain, llama_sampler_init_infill (vocab)); + break; + case COMMON_SAMPLER_TYPE_PENALTIES: + llama_sampler_chain_add(result->chain, llama_sampler_init_penalties (params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present)); + break; + default: + GGML_ASSERT(false && "unknown sampler type"); } } llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed)); @@ -475,6 +473,7 @@ char common_sampler_type_to_chr(enum common_sampler_type cnstr) { case COMMON_SAMPLER_TYPE_TOP_K: return 'k'; case COMMON_SAMPLER_TYPE_TYPICAL_P: return 'y'; case COMMON_SAMPLER_TYPE_TOP_P: return 'p'; + case COMMON_SAMPLER_TYPE_TOP_N_SIGMA: return 's'; case COMMON_SAMPLER_TYPE_MIN_P: return 'm'; case COMMON_SAMPLER_TYPE_TEMPERATURE: return 't'; case COMMON_SAMPLER_TYPE_XTC: return 'x'; @@ -490,6 +489,7 @@ std::string common_sampler_type_to_str(enum common_sampler_type cnstr) { case COMMON_SAMPLER_TYPE_TOP_K: return "top_k"; case COMMON_SAMPLER_TYPE_TYPICAL_P: return "typ_p"; case COMMON_SAMPLER_TYPE_TOP_P: return "top_p"; + case COMMON_SAMPLER_TYPE_TOP_N_SIGMA: return "top_n_sigma"; case COMMON_SAMPLER_TYPE_MIN_P: return "min_p"; case COMMON_SAMPLER_TYPE_TEMPERATURE: return "temperature"; case COMMON_SAMPLER_TYPE_XTC: return "xtc"; @@ -504,6 +504,7 @@ std::vector common_sampler_types_from_names(const std::vect { "dry", COMMON_SAMPLER_TYPE_DRY }, { "top_k", COMMON_SAMPLER_TYPE_TOP_K }, { "top_p", COMMON_SAMPLER_TYPE_TOP_P }, + { "top_n_sigma", COMMON_SAMPLER_TYPE_TOP_N_SIGMA }, { "typ_p", COMMON_SAMPLER_TYPE_TYPICAL_P }, { "min_p", COMMON_SAMPLER_TYPE_MIN_P }, { "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE }, @@ -517,6 +518,7 @@ std::vector common_sampler_types_from_names(const std::vect std::unordered_map sampler_alt_name_map { { "top-k", COMMON_SAMPLER_TYPE_TOP_K }, { "top-p", COMMON_SAMPLER_TYPE_TOP_P }, + { "top-n-sigma", COMMON_SAMPLER_TYPE_TOP_N_SIGMA }, { "nucleus", COMMON_SAMPLER_TYPE_TOP_P }, { "typical-p", COMMON_SAMPLER_TYPE_TYPICAL_P }, { "typical", COMMON_SAMPLER_TYPE_TYPICAL_P }, @@ -533,14 +535,16 @@ std::vector common_sampler_types_from_names(const std::vect auto sampler = sampler_canonical_name_map.find(name); if (sampler != sampler_canonical_name_map.end()) { samplers.push_back(sampler->second); - } else { - if (allow_alt_names) { - sampler = sampler_alt_name_map.find(name); - if (sampler != sampler_alt_name_map.end()) { - samplers.push_back(sampler->second); - } + continue; + } + if (allow_alt_names) { + sampler = sampler_alt_name_map.find(name); + if (sampler != sampler_alt_name_map.end()) { + samplers.push_back(sampler->second); + continue; } } + LOG_WRN("%s: unable to match sampler by name '%s'\n", __func__, name.c_str()); } return samplers; @@ -552,6 +556,7 @@ std::vector common_sampler_types_from_chars(const std::stri { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_K), COMMON_SAMPLER_TYPE_TOP_K }, { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TYPICAL_P), COMMON_SAMPLER_TYPE_TYPICAL_P }, { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_P), COMMON_SAMPLER_TYPE_TOP_P }, + { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_N_SIGMA), COMMON_SAMPLER_TYPE_TOP_N_SIGMA }, { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_MIN_P), COMMON_SAMPLER_TYPE_MIN_P }, { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TEMPERATURE), COMMON_SAMPLER_TYPE_TEMPERATURE }, { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_XTC), COMMON_SAMPLER_TYPE_XTC }, @@ -566,6 +571,8 @@ std::vector common_sampler_types_from_chars(const std::stri const auto sampler = sampler_name_map.find(c); if (sampler != sampler_name_map.end()) { samplers.push_back(sampler->second); + } else { + LOG_WRN("%s: unable to match sampler by char '%c'\n", __func__, c); } } diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index cf35fb86ec..a34ba29882 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -16,6 +16,7 @@ from pathlib import Path from hashlib import sha256 from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Literal, Sequence, TypeVar, cast from itertools import chain +from transformers import AutoConfig import math import numpy as np @@ -66,8 +67,6 @@ class ModelBase: part_names: list[str] is_safetensors: bool hparams: dict[str, Any] - block_count: int - tensor_map: gguf.TensorNameMap tensor_names: set[str] | None gguf_writer: gguf.GGUFWriter model_name: str | None @@ -78,7 +77,11 @@ class ModelBase: # subclasses should define this! model_arch: gguf.MODEL_ARCH - def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False, + # subclasses should initialize this! + block_count: int + tensor_map: gguf.TensorNameMap + + def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, *, is_big_endian: bool = False, use_temp_file: bool = False, eager: bool = False, metadata_override: Path | None = None, model_name: str | None = None, split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, @@ -113,8 +116,6 @@ class ModelBase: if not self.is_safetensors: self.part_names = ModelBase.get_model_part_names(self.dir_model, "pytorch_model", ".bin") self.hparams = ModelBase.load_hparams(self.dir_model) if hparams is None else hparams - self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"]) - self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count) self.tensor_names = None self.metadata_override = metadata_override self.model_name = model_name @@ -417,15 +418,19 @@ class ModelBase: @staticmethod def load_hparams(dir_model: Path): - with open(dir_model / "config.json", "r", encoding="utf-8") as f: - hparams = json.load(f) - architectures = hparams.get("architectures") - if "text_config" in hparams: - hparams = {**hparams, **hparams["text_config"]} - if architectures is not None: - # preserve "architectures" from root level config - hparams["architectures"] = architectures - return hparams + try: + # for security reason, we don't allow loading remote code by default + # if a model need remote code, we will fallback to config.json + return AutoConfig.from_pretrained(dir_model, trust_remote_code=False).to_dict() + except Exception as e: + logger.warning(f"Failed to load model config from {dir_model}: {e}") + logger.warning("Trying to load config.json instead") + with open(dir_model / "config.json", "r", encoding="utf-8") as f: + config = json.load(f) + if "llm_config" in config: + # rename for InternVL + config["text_config"] = config["llm_config"] + return config @classmethod def register(cls, *names: str) -> Callable[[AnyModel], AnyModel]: @@ -454,6 +459,20 @@ class ModelBase: class TextModel(ModelBase): + model_type = ModelType.TEXT + hf_arch: str + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self.hf_arch = get_model_architecture(self.hparams, self.model_type) + + if "text_config" in self.hparams: + # move the text_config to the root level + self.hparams = {**self.hparams, **self.hparams["text_config"]} + + self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"]) + self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count) + @classmethod def __init_subclass__(cls): # can't use an abstract property, because overriding it without type errors @@ -495,7 +514,7 @@ class TextModel(ModelBase): def set_gguf_parameters(self): self.gguf_writer.add_block_count(self.block_count) - if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None: + if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx", "n_positions"], optional=True)) is not None: self.gguf_writer.add_context_length(n_ctx) logger.info(f"gguf: context length = {n_ctx}") @@ -779,6 +798,9 @@ class TextModel(ModelBase): if chkhsh == "0e9433cbbb161f89e264eb32e8e64bfe69e834973ffca5d41d3948a604a3e2a3": # ref: https://huggingface.co/mistral-community/pixtral-12b res = "pixtral" + if chkhsh == "d5f1dd6f980fec569fb218a81a7658ac45fc56b38c5a0adeb1c232fbe04ef5ec": + # ref: https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base + res = "seed-coder" if res is None: logger.warning("\n") @@ -1064,10 +1086,36 @@ class TextModel(ModelBase): if (field := vocab_reader.get_field(gguf.Keys.Tokenizer.ADD_EOS)) is not None: self.gguf_writer.add_add_eos_token(field.parts[-1].tolist()[0]) + def _try_set_pooling_type(self) -> None: + # get pooling path + pooling_path = None + module_path = self.dir_model / "modules.json" + if module_path.is_file(): + with open(module_path, encoding="utf-8") as f: + modules = json.load(f) + for mod in modules: + if mod["type"] == "sentence_transformers.models.Pooling": + pooling_path = mod["path"] + break + + # get pooling type + if pooling_path is not None: + with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f: + pooling = json.load(f) + if pooling["pooling_mode_mean_tokens"]: + pooling_type = gguf.PoolingType.MEAN + elif pooling["pooling_mode_cls_token"]: + pooling_type = gguf.PoolingType.CLS + elif pooling["pooling_mode_lasttoken"]: + pooling_type = gguf.PoolingType.LAST + else: + raise NotImplementedError("Only MEAN, CLS, and LAST pooling types supported") + self.gguf_writer.add_pooling_type(pooling_type) + class VisionModel(ModelBase): + model_type = ModelType.VISION model_arch = gguf.MODEL_ARCH.CLIP_VISION - n_text_embd = 0 preprocessor_config: dict[str, Any] global_config: dict[str, Any] @@ -1077,9 +1125,11 @@ class VisionModel(ModelBase): if self.model_arch != gguf.MODEL_ARCH.CLIP_VISION: raise TypeError("VisionModel must be subclassed with model_arch = gguf.MODEL_ARCH.CLIP_VISION") - # small hack to correct the number of layers - self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.CLIP_VISION, 128) - self.n_embd_text = self.find_hparam(["hidden_size", "n_embd"]) + # get n_embd of the text model + if "text_config" not in self.hparams: + self.hparams["text_config"] = {} + text_config = {**self.hparams, **self.hparams["text_config"]} + self.n_embd_text = text_config.get("hidden_size", text_config.get("n_embd", 0)) assert self.n_embd_text > 0, "n_embd not found in hparams" if "vision_config" not in self.hparams: @@ -1088,6 +1138,9 @@ class VisionModel(ModelBase): self.global_config = self.hparams self.hparams = self.hparams["vision_config"] + self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers", "depth"]) + self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.CLIP_VISION, self.block_count) + # load preprocessor config with open(self.dir_model / "preprocessor_config.json", "r", encoding="utf-8") as f: self.preprocessor_config = json.load(f) @@ -1105,12 +1158,12 @@ class VisionModel(ModelBase): self.gguf_writer.add_vision_patch_size(self.find_hparam(["patch_size"])) self.gguf_writer.add_vision_embedding_length(self.find_hparam(["hidden_size"])) self.gguf_writer.add_vision_feed_forward_length(self.find_hparam(["intermediate_size"])) - self.gguf_writer.add_vision_block_count(self.find_hparam(["num_hidden_layers"])) + self.gguf_writer.add_vision_block_count(self.block_count) self.gguf_writer.add_vision_head_count(self.find_hparam(["num_attention_heads"])) # preprocessor config self.gguf_writer.add_vision_image_mean(self.preprocessor_config["image_mean"]) - self.gguf_writer.add_vision_image_std(self.preprocessor_config["image_mean"]) + self.gguf_writer.add_vision_image_std(self.preprocessor_config["image_std"]) def write_vocab(self): raise ValueError("VisionModel does not support vocab writing") @@ -1342,10 +1395,10 @@ class BaichuanModel(TextModel): self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"]) self.gguf_writer.add_file_type(self.ftype) - if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: - if self.hparams["rope_scaling"].get("type") == "linear": - self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) - self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) + rope_scaling = self.hparams.get("rope_scaling") or {} + if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling: + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"]) def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: head_count = self.hparams["num_attention_heads"] @@ -1466,10 +1519,10 @@ class XverseModel(TextModel): self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"]) self.gguf_writer.add_file_type(self.ftype) - if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: - if self.hparams["rope_scaling"].get("type") == "linear": - self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) - self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) + rope_scaling = self.hparams.get("rope_scaling") or {} + if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling: + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"]) def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: del bid # unused @@ -1726,8 +1779,7 @@ class StableLMModel(TextModel): "LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM", - "Idefics3ForConditionalGeneration", - "SmolVLMForConditionalGeneration", + "VLlama3ForCausalLM", "LlavaForConditionalGeneration") class LlamaModel(TextModel): model_arch = gguf.MODEL_ARCH.LLAMA @@ -1736,11 +1788,7 @@ class LlamaModel(TextModel): def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) # fix for SmolVLM2, missing `num_attention_heads` in config.json - if self.hparams["architectures"][0] == "SmolVLMForConditionalGeneration": - self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32) - # fix for Pixtral, missing `num_attention_heads` in config.json - if self.hparams["architectures"][0] == "LlavaForConditionalGeneration" \ - and self.hparams.get("model_type") == "mistral": + if self.hf_arch == "VLlama3ForCausalLM": self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32) def set_vocab(self): @@ -1787,10 +1835,10 @@ class LlamaModel(TextModel): rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"] self.gguf_writer.add_rope_dimension_count(rope_dim) - if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: - if self.hparams["rope_scaling"].get("type") == "linear": - self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) - self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) + rope_scaling = self.hparams.get("rope_scaling") or {} + if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling: + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"]) @staticmethod def permute(weights: Tensor, n_head: int, n_head_kv: int | None): @@ -1898,31 +1946,50 @@ class LlamaModel(TextModel): raise ValueError(f"Unprocessed experts: {experts}") -@ModelBase.register("LlavaForConditionalGeneration") +@ModelBase.register( + "LlavaForConditionalGeneration", # pixtral + "Mistral3ForConditionalGeneration", # mistral small 3.1 +) class LlavaVisionModel(VisionModel): img_break_tok_id = -1 def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) if self.hparams["model_type"] == "pixtral": - # fix missing config.json values - self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 16) - self.hparams["num_hidden_layers"] = self.hparams.get("num_hidden_layers", 24) - self.hparams["intermediate_size"] = self.hparams.get("intermediate_size", 4096) - self.hparams["hidden_size"] = self.hparams.get("hidden_size", 1024) + # layer_norm_eps is not in config.json, it is hard-coded in modeling_pixtral.py self.hparams["layer_norm_eps"] = self.hparams.get("layer_norm_eps", 1e-5) - self.img_break_tok_id = 12 # see tokenizer_config.json + self.img_break_tok_id = self.get_token_id("[IMG_BREAK]") + logger.info(f"Image break token id: {self.img_break_tok_id}") else: raise ValueError(f"Unsupported model type: {self.hparams['model_type']}") + def get_token_id(self, token: str) -> int: + tokenizer_config_file = self.dir_model / 'tokenizer_config.json' + with open(tokenizer_config_file, "r", encoding="utf-8") as f: + added_tokens_decoder = json.load(f)['added_tokens_decoder'] + for id_, token_data in added_tokens_decoder.items(): + if token_data["content"] == token: + return int(id_) + raise ValueError(f"Token '{token}' not found in tokenizer config.") + def set_gguf_parameters(self): super().set_gguf_parameters() hparams = self.hparams if hparams["model_type"] == "pixtral": self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.PIXTRAL) - # default values below are taken from HF tranformers code self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"]) - self.gguf_writer.add_vision_use_silu(True) + + # hidden_act + if hparams["hidden_act"] == "silu": + self.gguf_writer.add_vision_use_silu(True) + elif hparams["hidden_act"] == "gelu": + self.gguf_writer.add_vision_use_gelu(True) + else: + raise ValueError(f"Unsupported hidden_act: {hparams['hidden_act']}") + + # spatial_merge_size + if "spatial_merge_size" in self.global_config: + 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 @@ -1951,13 +2018,12 @@ class LlavaVisionModel(VisionModel): class SmolVLMModel(VisionModel): def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) - # fix for SmolVLM2, missing some keys in config.json - # default values are taken from transformers code if self.hparams["model_type"] == "smolvlm_vision": + # fix for SmolVLM2, missing some keys in config.json + # default values are taken from transformers code self.hparams["hidden_size"] = self.hparams.get("hidden_size", 1152) self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 16) self.hparams["intermediate_size"] = self.hparams.get("intermediate_size", 3072) - self.hparams["num_hidden_layers"] = self.hparams.get("num_hidden_layers", 12) def set_gguf_parameters(self): super().set_gguf_parameters() @@ -2070,6 +2136,9 @@ class DeciModel(TextModel): # if n_heads_in_group is not None, then # _num_kv_heads[il] is num_attention_head // n_heads_in_group and # _num_heads[il] is num_attention_head + # ***dummy layer*** for nemotron 253B + # if n_heads_in_group is None and ffn_mult is None + # then _num_kv_heads[il] is 0 and _num_heads[il] is 0 and _ffn_dims is 0 for il in range(len(_block_configs)): if _block_configs[il]["attention"]["n_heads_in_group"] is None: if _block_configs[il]["attention"]["replace_with_linear"] is True: @@ -2081,7 +2150,10 @@ class DeciModel(TextModel): else: self._num_kv_heads.append(self.hparams["num_attention_heads"] // _block_configs[il]["attention"]["n_heads_in_group"]) self._num_heads.append(self.hparams["num_attention_heads"]) - _ffn_multipliers.append(_block_configs[il]["ffn"]["ffn_mult"]) + if _block_configs[il]["ffn"]["ffn_mult"] is None: # dummy layer + _ffn_multipliers.append(0.0) + else: + _ffn_multipliers.append(_block_configs[il]["ffn"]["ffn_mult"]) assert self.block_count == len(self._num_kv_heads) assert self.block_count == len(self._num_heads) assert self.block_count == len(_ffn_multipliers) @@ -2141,10 +2213,10 @@ class DeciModel(TextModel): rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"] self.gguf_writer.add_rope_dimension_count(rope_dim) - if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: - if self.hparams["rope_scaling"].get("type") == "linear": - self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) - self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) + rope_scaling = self.hparams.get("rope_scaling") or {} + if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling: + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"]) @staticmethod def permute(weights: Tensor, n_head: int, n_head_kv: int | None): @@ -2384,10 +2456,10 @@ class MiniCPMModel(TextModel): logit_scale = self.hparams["hidden_size"] / self.hparams["dim_model_base"] self.gguf_writer.add_logit_scale(logit_scale) logger.info(f"gguf: (minicpm) logit_scale = {logit_scale}") - if self.hparams.get("rope_scaling") is not None: - if self.hparams["rope_scaling"].get("type") == "longrope": - self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LONGROPE) - logger.info(f"gguf: (minicpm) rope_scaling_type = {gguf.RopeScalingType.LONGROPE}") + rope_scaling = self.hparams.get("rope_scaling") or {} + if rope_scaling.get("rope_type", rope_scaling.get("type")) == "longrope": + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LONGROPE) + logger.info(f"gguf: (minicpm) rope_scaling_type = {gguf.RopeScalingType.LONGROPE}") def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: rope_dims = self.hparams["hidden_size"] // self.hparams["num_attention_heads"] @@ -2519,7 +2591,7 @@ class QwenModel(TextModel): self.gguf_writer.add_file_type(self.ftype) -@ModelBase.register("Qwen2ForCausalLM") +@ModelBase.register("Qwen2Model", "Qwen2ForCausalLM") class Qwen2Model(TextModel): model_arch = gguf.MODEL_ARCH.QWEN2 @@ -2531,11 +2603,22 @@ class Qwen2Model(TextModel): def set_gguf_parameters(self): super().set_gguf_parameters() - if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: - if self.hparams["rope_scaling"].get("type") == "yarn": - self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN) - self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) - self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"]) + self._try_set_pooling_type() + rope_scaling = self.hparams.get("rope_scaling") or {} + if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling: + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN) + self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"]) + self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"]) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + if self.hf_arch == "Qwen2Model": + name = f"model.{name}" # map to Qwen2ForCausalLM tensors + if "language_model." in name: + name = name.replace("language_model.", "") # for InternVL + if name.startswith("mlp") or name.startswith("vision_model"): + # skip visual tensors + return [] + yield from super().modify_tensors(data_torch, name, bid) @ModelBase.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration") @@ -2554,11 +2637,144 @@ class Qwen2VLModel(TextModel): except FileNotFoundError: self._set_vocab_gpt2() - def get_tensors(self) -> Iterator[tuple[str, Tensor]]: - for name, data in super().get_tensors(): - if name.startswith("visual."): - continue - yield name, data + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + del bid # unused + if name.startswith("visual."): + # skip visual tensors + return [] + return [(self.map_tensor_name(name), data_torch)] + + +@ModelBase.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration") +class Qwen2VLVisionModel(VisionModel): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self.hparams["image_size"] = self.hparams.get("image_size", 560) + # rename config.json values + self.hparams["num_attention_heads"] = self.hparams.get("num_heads") + self.hparams["num_hidden_layers"] = self.hparams.get("depth") + if "embed_dim" in self.hparams: # qwen2vl + self.hparams["intermediate_size"] = self.hparams.get("hidden_size") + self.hparams["hidden_size"] = self.hparams.get("embed_dim") + + def set_gguf_parameters(self): + super().set_gguf_parameters() + hparams = self.hparams + if self.global_config['model_type'] == 'qwen2_vl': + self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.QWEN2VL) + elif self.global_config['model_type'] == 'qwen2_5_vl': + self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.QWEN25VL) + self.gguf_writer.add_vision_use_silu(True) + # find n_wa_pattern (window attention pattern) + fullatt_block_indexes = hparams.get("fullatt_block_indexes") + assert fullatt_block_indexes is not None, "fullatt_block_indexes is required for qwen2_5_vl" + n_wa_pattern = fullatt_block_indexes[0] + 1 + # validate n_wa_pattern + for i in range(1, len(fullatt_block_indexes)): + if fullatt_block_indexes[i] - fullatt_block_indexes[i - 1] != n_wa_pattern: + raise ValueError(f"Invalid fullatt_block_indexes: {fullatt_block_indexes}") + self.gguf_writer.add_vision_n_wa_pattern(n_wa_pattern) + else: + raise ValueError(f"Unknown QwenVL model type: {self.global_config['model_type']}") + # default values below are taken from HF tranformers code + self.gguf_writer.add_vision_attention_layernorm_eps(self.global_config.get("rms_norm_eps", 1e-6)) + + def tensor_force_quant(self, name, new_name, bid, n_dims): + del bid, name, n_dims # unused + if ".patch_embd." in new_name: + return gguf.GGMLQuantizationType.F16 + if ".position_embd." in new_name: + return gguf.GGMLQuantizationType.F32 + return False + + 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 + if ".qkv." in name: + if data_torch.ndim == 2: # weight + c3, _ = data_torch.shape + else: # bias + c3 = data_torch.shape[0] + assert c3 % 3 == 0 + c = c3 // 3 + 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), + ] + 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, ...]), + ] + else: + return [(self.map_tensor_name(name), data_torch)] + return [] # skip other tensors + + +@ModelBase.register("InternVisionModel") +class InternVisionModel(VisionModel): + def set_gguf_parameters(self): + super().set_gguf_parameters() + hparams = self.hparams + self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.INTERNVL) + self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"]) + # hidden_act + if hparams["hidden_act"] == "silu": + self.gguf_writer.add_vision_use_silu(True) + elif hparams["hidden_act"] == "gelu": + self.gguf_writer.add_vision_use_gelu(True) + else: + raise ValueError(f"Unsupported hidden_act: {hparams['hidden_act']}") + # downsample_ratio + downsample_ratio = self.global_config.get("downsample_ratio") + assert downsample_ratio is not None + self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio)) + + def tensor_force_quant(self, name, new_name, bid, n_dims): + del bid, name, n_dims # unused + if ".patch_embd." in new_name: + return gguf.GGMLQuantizationType.F16 + if ".position_embd." in new_name: + return gguf.GGMLQuantizationType.F32 + return False + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + del bid # unused + if name.startswith("vision_model") or name.startswith("mlp"): + # process visual tensors + # correct name + if name.startswith("vision_model"): + name = "vision_tower." + name + if (".ls" in name or "position_embedding" in name) and not name.endswith(".weight"): + name += ".weight" + # split QKV tensors if needed + if ".qkv." in name: + if data_torch.ndim == 2: # weight + c3, _ = data_torch.shape + else: # bias + c3 = data_torch.shape[0] + assert c3 % 3 == 0 + c = c3 // 3 + 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 @ModelBase.register("WavTokenizerDec") @@ -2613,6 +2829,13 @@ class Qwen2MoeModel(TextModel): if (shared_expert_intermediate_size := self.hparams.get('shared_expert_intermediate_size')) is not None: self.gguf_writer.add_expert_shared_feed_forward_length(shared_expert_intermediate_size) logger.info(f"gguf: expert shared feed forward length = {shared_expert_intermediate_size}") + # YaRN is not enabled by default + # To enable it, please refer to this guide: https://huggingface.co/Qwen/Qwen3-30B-A3B#processing-long-texts + rope_scaling = self.hparams.get("rope_scaling") or {} + if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling: + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN) + self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"]) + self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"]) _experts: list[dict[str, Tensor]] | None = None @@ -2880,7 +3103,7 @@ class Phi3MiniModel(TextModel): scale = max_pos_embds / orig_max_pos_embds - rope_scaling_type = rope_scaling.get('type', '').lower() + rope_scaling_type = rope_scaling.get('rope_type', rope_scaling.get('type', '')).lower() if len(rope_scaling_type) == 0: raise KeyError('Missing the required key rope_scaling.type') @@ -3192,10 +3415,10 @@ class InternLM2Model(TextModel): self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"]) self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"]) self.gguf_writer.add_file_type(self.ftype) - if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: - if self.hparams["rope_scaling"].get("type") == "linear": - self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) - self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) + rope_scaling = self.hparams.get("rope_scaling") or {} + if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling: + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"]) def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: num_heads = self.hparams["num_attention_heads"] @@ -3205,6 +3428,11 @@ class InternLM2Model(TextModel): head_dim = n_embd // num_heads num_groups = num_heads // q_per_kv + name = name.replace("language_model.", "") # InternVL + if name.startswith("mlp") or name.startswith("vision_model"): + # skip visual tensors + return [] + if bid is not None and f"model.layers.{bid}.attention.wqkv" in name: qkv = data_torch @@ -3270,14 +3498,18 @@ class InternLM3Model(TextModel): rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"] self.gguf_writer.add_rope_dimension_count(rope_dim) - if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: - if self.hparams["rope_scaling"].get("type") == "linear" or self.hparams["rope_scaling"].get("rope_type") == "linear": - self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) - self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) + rope_scaling = self.hparams.get("rope_scaling") or {} + if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling: + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"]) def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: n_head = self.hparams["num_attention_heads"] n_kv_head = self.hparams.get("num_key_value_heads") + name = name.replace("language_model.", "") # InternVL + if name.startswith("mlp") or name.startswith("vision_model"): + # skip visual tensors + 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")): @@ -3296,29 +3528,7 @@ class BertModel(TextModel): def set_gguf_parameters(self): super().set_gguf_parameters() self.gguf_writer.add_causal_attention(False) - - # get pooling path - pooling_path = None - module_path = self.dir_model / "modules.json" - if module_path.is_file(): - with open(module_path, encoding="utf-8") as f: - modules = json.load(f) - for mod in modules: - if mod["type"] == "sentence_transformers.models.Pooling": - pooling_path = mod["path"] - break - - # get pooling type - if pooling_path is not None: - with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f: - pooling = json.load(f) - if pooling["pooling_mode_mean_tokens"]: - pooling_type = gguf.PoolingType.MEAN - elif pooling["pooling_mode_cls_token"]: - pooling_type = gguf.PoolingType.CLS - else: - raise NotImplementedError("Only MEAN and CLS pooling types supported") - self.gguf_writer.add_pooling_type(pooling_type) + self._try_set_pooling_type() def set_vocab(self): tokens, toktypes, tokpre = self.get_vocab_base() @@ -3372,14 +3582,7 @@ class BertModel(TextModel): return [(self.map_tensor_name(name), data_torch)] - -@ModelBase.register("RobertaModel") -class RobertaModel(BertModel): - model_arch = gguf.MODEL_ARCH.BERT - - def __init__(self, *args, **kwargs): - super().__init__(*args, **kwargs) - + def _xlmroberta_tokenizer_init(self) -> None: # we need the pad_token_id to know how to chop down position_embd matrix if (pad_token_id := self.hparams.get("pad_token_id")) is not None: self._position_offset = 1 + pad_token_id @@ -3388,82 +3591,7 @@ class RobertaModel(BertModel): else: self._position_offset = None - def set_vocab(self): - """Support BPE tokenizers for roberta models""" - bpe_tok_path = self.dir_model / "tokenizer.json" - if bpe_tok_path.exists(): - self._set_vocab_gpt2() - self.gguf_writer.add_add_bos_token(True) - self.gguf_writer.add_add_eos_token(True) - - # we need this to validate the size of the token_type embeddings - # though currently we are passing all zeros to the token_type embeddings - # "Sequence A" or "Sequence B" - self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1)) - - else: - return super().set_vocab() - - def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: - # if name starts with "roberta.", remove the prefix - # e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main - if name.startswith("roberta."): - name = name[8:] - - # position embeddings start at pad_token_id + 1, so just chop down the weight tensor - if name == "embeddings.position_embeddings.weight": - if self._position_offset is not None: - data_torch = data_torch[self._position_offset:,:] - - return super().modify_tensors(data_torch, name, bid) - - -@ModelBase.register("NomicBertModel") -class NomicBertModel(BertModel): - model_arch = gguf.MODEL_ARCH.NOMIC_BERT - - def __init__(self, *args, **kwargs): - super().__init__(*args, **kwargs) - - # the HF config claims n_ctx=8192, but it uses RoPE scaling - self.hparams["n_ctx"] = 2048 - - # SwigLU activation - assert self.hparams["activation_function"] == "swiglu" - # this doesn't do anything in the HF version - assert self.hparams["causal"] is False - # no bias tensors - assert self.hparams["qkv_proj_bias"] is False - assert self.hparams["mlp_fc1_bias"] is False - assert self.hparams["mlp_fc2_bias"] is False - # norm at end of layer - assert self.hparams["prenorm"] is False - # standard RoPE - assert self.hparams["rotary_emb_fraction"] == 1.0 - assert self.hparams["rotary_emb_interleaved"] is False - assert self.hparams["rotary_emb_scale_base"] is None - - def set_gguf_parameters(self): - super().set_gguf_parameters() - self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"]) - - -@ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification") -class XLMRobertaModel(BertModel): - model_arch = gguf.MODEL_ARCH.BERT - - def __init__(self, *args, **kwargs): - super().__init__(*args, **kwargs) - - # we need the pad_token_id to know how to chop down position_embd matrix - if (pad_token_id := self.hparams.get("pad_token_id")) is not None: - self._position_offset = 1 + pad_token_id - if "max_position_embeddings" in self.hparams: - self.hparams["max_position_embeddings"] -= self._position_offset - else: - self._position_offset = None - - def set_vocab(self): + def _xlmroberta_set_vocab(self) -> None: # to avoid TypeError: Descriptors cannot be created directly # exception when importing sentencepiece_model_pb2 os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python" @@ -3545,6 +3673,145 @@ class XLMRobertaModel(BertModel): self.gguf_writer.add_add_bos_token(True) self.gguf_writer.add_add_eos_token(True) + +@ModelBase.register("RobertaModel") +class RobertaModel(BertModel): + model_arch = gguf.MODEL_ARCH.BERT + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + # we need the pad_token_id to know how to chop down position_embd matrix + if (pad_token_id := self.hparams.get("pad_token_id")) is not None: + self._position_offset = 1 + pad_token_id + if "max_position_embeddings" in self.hparams: + self.hparams["max_position_embeddings"] -= self._position_offset + else: + self._position_offset = None + + def set_vocab(self): + """Support BPE tokenizers for roberta models""" + bpe_tok_path = self.dir_model / "tokenizer.json" + if bpe_tok_path.exists(): + self._set_vocab_gpt2() + self.gguf_writer.add_add_bos_token(True) + self.gguf_writer.add_add_eos_token(True) + + # we need this to validate the size of the token_type embeddings + # though currently we are passing all zeros to the token_type embeddings + # "Sequence A" or "Sequence B" + self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1)) + + else: + return super().set_vocab() + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + # if name starts with "roberta.", remove the prefix + # e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main + if name.startswith("roberta."): + name = name[8:] + + # position embeddings start at pad_token_id + 1, so just chop down the weight tensor + if name == "embeddings.position_embeddings.weight": + if self._position_offset is not None: + data_torch = data_torch[self._position_offset:,:] + + return super().modify_tensors(data_torch, name, bid) + + +@ModelBase.register("NomicBertModel") +class NomicBertModel(BertModel): + model_arch = gguf.MODEL_ARCH.BERT + + def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, **kwargs: Any): + hparams = kwargs.pop("hparams", None) + if hparams is None: + hparams = ModelBase.load_hparams(dir_model) + + self.is_moe = bool(hparams.get("moe_every_n_layers")) + self.model_arch = gguf.MODEL_ARCH.NOMIC_BERT_MOE if self.is_moe else gguf.MODEL_ARCH.NOMIC_BERT + + super().__init__(dir_model, ftype, fname_out, hparams=hparams, **kwargs) + + self._tokenizer_is_xlmroberta = self._is_tokenizer_xlmroberta() + if self._tokenizer_is_xlmroberta: + self._xlmroberta_tokenizer_init() + + npos, mtp = self.hparams["n_positions"], self.hparams.get("max_trained_positions", 2048) + if npos == 8192 and mtp == 2048: + self.hparams["n_positions"] = 2048 # nomic-embed-text v1 and v1.5 are trained for 2048 tokens. + elif npos == 2048 and mtp == 2048: + self.hparams["n_positions"] = 512 # nomic-embed-text-v2-moe is trained for 512 tokens. + else: + raise ValueError(f"unrecognized parameters: n_positions={npos}, max_trained_positions={mtp}") + + assert self.hparams["activation_function"] == "gelu" if self.is_moe else "swiglu" + + # this doesn't do anything in the HF version + assert self.hparams["causal"] is False + # no bias tensors unless MoE + assert self.hparams["qkv_proj_bias"] == self.is_moe + assert self.hparams["mlp_fc1_bias"] == self.is_moe + assert self.hparams["mlp_fc2_bias"] == self.is_moe + + # norm at end of layer + assert self.hparams["prenorm"] is False + # standard RoPE + assert self.hparams["rotary_emb_fraction"] == 1.0 + assert self.hparams["rotary_emb_interleaved"] is False + assert self.hparams["rotary_emb_scale_base"] is None + + def set_vocab(self) -> None: + if self._tokenizer_is_xlmroberta: + return self._xlmroberta_set_vocab() + return super().set_vocab() + + 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. + + if "mlp.experts.mlp.w1" in name: + data_torch = data_torch.view(self.hparams["num_experts"], self.hparams["n_inner"], self.hparams["n_embd"]) + name += ".weight" + + if "mlp.experts.mlp.w2" in name: + data_torch = data_torch.view(self.hparams["num_experts"], self.hparams["n_inner"], self.hparams["n_embd"]) + data_torch = data_torch.transpose(1, 2) + name += ".weight" + + return [(self.map_tensor_name(name), data_torch)] + + def set_gguf_parameters(self): + super().set_gguf_parameters() + self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"]) + if self.is_moe: + self.gguf_writer.add_moe_every_n_layers(self.hparams["moe_every_n_layers"]) + self.gguf_writer.add_expert_count(self.hparams["num_experts"]) + self.gguf_writer.add_expert_used_count(self.hparams["moe_top_k"]) + + def _is_tokenizer_xlmroberta(self) -> bool: + with open(self.dir_model / "tokenizer.json") as f: + tokenizer_json = json.load(f) + toktyp = tokenizer_json["model"]["type"] + if toktyp == "Unigram": + return True + if toktyp == "WordPiece": + return False + raise ValueError(f"unknown tokenizer: {toktyp}") + + +@ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification") +class XLMRobertaModel(BertModel): + model_arch = gguf.MODEL_ARCH.BERT + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self._xlmroberta_tokenizer_init() + + def set_vocab(self): + self._xlmroberta_set_vocab() + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: # if name starts with "roberta.", remove the prefix # e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main @@ -3725,6 +3992,16 @@ class Gemma3VisionModel(VisionModel): # default values below are taken from HF tranformers code self.gguf_writer.add_vision_attention_layernorm_eps(hparams.get("layer_norm_eps", 1e-6)) self.gguf_writer.add_vision_use_gelu(True) + # calculate proj_scale_factor (used by tinygemma3 test model) + image_seq_length = self.preprocessor_config.get("image_seq_length", 256) + n_per_side = int(image_seq_length ** 0.5) + image_size = self.hparams["image_size"] + patch_size = self.hparams["patch_size"] + proj_scale_factor = (image_size // patch_size) // n_per_side + if proj_scale_factor > 0 and proj_scale_factor != 4: + # we only need to write this if it's not the default value + # in this case, we are converting a test model + self.gguf_writer.add_vision_projector_scale_factor(proj_scale_factor) def tensor_force_quant(self, name, new_name, bid, n_dims): del bid, new_name, n_dims # unused @@ -3738,6 +4015,9 @@ class Gemma3VisionModel(VisionModel): 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 + if name.startswith("multi_modal_projector.") or name.startswith("vision_tower.") \ or name.startswith("multimodal_projector.") or name.startswith("vision_model."): # process vision tensors @@ -4663,12 +4943,12 @@ class DeepseekV2Model(TextModel): self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"]) - if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: - if self.hparams["rope_scaling"].get("type") == "yarn": - self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN) - self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) - self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"]) - self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1 * hparams["rope_scaling"]["mscale_all_dim"]) + rope_scaling = self.hparams.get("rope_scaling") or {} + if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling: + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN) + self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"]) + self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"]) + self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1 * rope_scaling["mscale_all_dim"]) _experts: list[dict[str, Tensor]] | None = None @@ -5153,18 +5433,18 @@ class Glm4Model(TextModel): special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"]) special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"]) special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"]) - special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["[gMASK]"]) + special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|endoftext|>"]) special_vocab.add_to_gguf(self.gguf_writer) def set_gguf_parameters(self): super().set_gguf_parameters() rope_dim = self.hparams["head_dim"] self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5))) - if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: - if self.hparams["rope_scaling"].get("type") == "yarn": - self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN) - self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) - self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"]) + rope_scaling = self.hparams.get("rope_scaling") or {} + if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling: + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN) + self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"]) + self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"]) @ModelBase.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration") @@ -5397,10 +5677,10 @@ class ExaoneModel(TextModel): rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"], optional=True) rotary_factor = rotary_factor if rotary_factor is not None else 1.0 self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"]))) - if hparams.get("rope_scaling") is not None and "factor" in hparams["rope_scaling"]: - if hparams["rope_scaling"].get("type") == "linear": - self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) - self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"]) + rope_scaling = self.hparams.get("rope_scaling") or {} + if rope_scaling.get("rope_type", rope_scaling.get("type")) == "linear" and "factor" in rope_scaling: + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"]) def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]: if rope_scaling := self.find_hparam(["rope_scaling"], optional=True): @@ -5503,7 +5783,13 @@ class BailingMoeModel(TextModel): rope_dim = hparams.get("head_dim") or hparams["hidden_size"] // hparams["num_attention_heads"] self.gguf_writer.add_rope_dimension_count(rope_dim) - self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE) + rope_scaling = self.hparams.get("rope_scaling") or {} + if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling: + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN) + self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"]) + self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"]) + else: + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE) self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"]) self.gguf_writer.add_vocab_size(hparams["vocab_size"]) self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"]) @@ -5805,6 +6091,18 @@ def split_str_to_n_bytes(split_str: str) -> int: return n +def get_model_architecture(hparams: dict[str, Any], model_type: ModelType) -> str: + text_config = hparams.get("text_config", {}) + vision_config = hparams.get("vision_config", {}) + arch = hparams["architectures"][0] + # if "architectures" is found in the sub-config, use that instead + if model_type == ModelType.TEXT and text_config.get("architectures") is not None: + arch = text_config["architectures"][0] + elif model_type == ModelType.VISION and vision_config.get("architectures") is not None: + arch = vision_config["architectures"][0] + return arch + + def main() -> None: args = parse_args() @@ -5857,16 +6155,16 @@ def main() -> None: logger.info(f"Loading model: {dir_model.name}") - hparams = ModelBase.load_hparams(dir_model) - if args.mmproj: if "mmproj" not in fname_out.name: fname_out = ModelBase.add_prefix_to_filename(fname_out, "mmproj-") with torch.inference_mode(): output_type = ftype_map[args.outtype] - model_architecture = hparams["architectures"][0] model_type = ModelType.VISION if args.mmproj else ModelType.TEXT + hparams = ModelBase.load_hparams(dir_model) + model_architecture = get_model_architecture(hparams, model_type) + logger.info(f"Model architecture: {model_architecture}") try: model_class = ModelBase.from_model_architecture(model_architecture, model_type=model_type) except NotImplementedError: diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py index 03a1d8d8c9..5993a4c983 100755 --- a/convert_hf_to_gguf_update.py +++ b/convert_hf_to_gguf_update.py @@ -116,6 +116,7 @@ models = [ {"name": "llama4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", }, {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", }, {"name": "pixtral", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistral-community/pixtral-12b", }, + {"name": "seed-coder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base", }, ] diff --git a/docs/development/HOWTO-add-model.md b/docs/development/HOWTO-add-model.md index 78c6f76077..7f71e0247d 100644 --- a/docs/development/HOWTO-add-model.md +++ b/docs/development/HOWTO-add-model.md @@ -9,10 +9,10 @@ Adding a model requires few steps: After following these steps, you can open PR. Also, it is important to check that the examples and main ggml backends (CUDA, METAL, CPU) are working with the new architecture, especially: -- [main](/examples/main/) -- [imatrix](/examples/imatrix/) -- [quantize](/examples/quantize/) -- [server](/examples/server/) +- [main](/tools/main/) +- [imatrix](/tools/imatrix/) +- [quantize](/tools/quantize/) +- [server](/tools/server/) ### 1. Convert the model to GGUF diff --git a/docs/multimodal.md b/docs/multimodal.md new file mode 100644 index 0000000000..6a5d2b342b --- /dev/null +++ b/docs/multimodal.md @@ -0,0 +1,77 @@ +# Multimodal + +llama.cpp supports multimodal input via `libmtmd`. Currently, there are 2 tools support this feature: +- [llama-mtmd-cli](../tools/mtmd/README.md) +- [llama-server](../tools/server/README.md) via OpenAI-compatible `/chat/completions` API + +To enable it, can use use one of the 2 methods below: + +- Use `-hf` option with a supported model (see a list of pre-quantized model below) + - To load a model using `-hf` while disabling multimodal, use `--no-mmproj` + - To load a model using `-hf` while using a custom mmproj file, use `--mmproj local_file.gguf` +- Use `-m model.gguf` option with `--mmproj file.gguf` to specify text and multimodal projector respectively + +By default, multimodal projector will be offloaded to GPU. To disable this, add `--no-mmproj-offload` + +For example: + +```sh +# simple usage with CLI +llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF + +# simple usage with server +llama-server -hf ggml-org/gemma-3-4b-it-GGUF + +# using local file +llama-server -m gemma-3-4b-it-Q4_K_M.gguf --mmproj mmproj-gemma-3-4b-it-Q4_K_M.gguf + +# no GPU offload +llama-server -hf ggml-org/gemma-3-4b-it-GGUF --no-mmproj-offload +``` + +## Pre-quantized models + +These are ready-to-use models, most of them come with `Q4_K_M` quantization by default. + +Replaces the `(tool_name)` with the name of binary you want to use. For example, `llama-mtmd-cli` or `llama-server` + +NOTE: some models may require large context window, for example: `-c 8192` + +```sh +# Gemma 3 +(tool_name) -hf ggml-org/gemma-3-4b-it-GGUF +(tool_name) -hf ggml-org/gemma-3-12b-it-GGUF +(tool_name) -hf ggml-org/gemma-3-27b-it-GGUF + +# SmolVLM +(tool_name) -hf ggml-org/SmolVLM-Instruct-GGUF +(tool_name) -hf ggml-org/SmolVLM-256M-Instruct-GGUF +(tool_name) -hf ggml-org/SmolVLM-500M-Instruct-GGUF +(tool_name) -hf ggml-org/SmolVLM2-2.2B-Instruct-GGUF +(tool_name) -hf ggml-org/SmolVLM2-256M-Video-Instruct-GGUF +(tool_name) -hf ggml-org/SmolVLM2-500M-Video-Instruct-GGUF + +# Pixtral 12B +(tool_name) -hf ggml-org/pixtral-12b-GGUF + +# Qwen 2 VL +(tool_name) -hf ggml-org/Qwen2-VL-2B-Instruct-GGUF +(tool_name) -hf ggml-org/Qwen2-VL-7B-Instruct-GGUF + +# Qwen 2.5 VL +(tool_name) -hf ggml-org/Qwen2.5-VL-3B-Instruct-GGUF +(tool_name) -hf ggml-org/Qwen2.5-VL-7B-Instruct-GGUF +(tool_name) -hf ggml-org/Qwen2.5-VL-32B-Instruct-GGUF +(tool_name) -hf ggml-org/Qwen2.5-VL-72B-Instruct-GGUF + +# Mistral Small 3.1 24B (IQ2_M quantization) +(tool_name) -hf ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF + +# InternVL 2.5 and 3 +(tool_name) -hf ggml-org/InternVL2_5-1B-GGUF +(tool_name) -hf ggml-org/InternVL2_5-4B-GGUF +(tool_name) -hf ggml-org/InternVL3-1B-Instruct-GGUF +(tool_name) -hf ggml-org/InternVL3-2B-Instruct-GGUF +(tool_name) -hf ggml-org/InternVL3-8B-Instruct-GGUF +(tool_name) -hf ggml-org/InternVL3-14B-Instruct-GGUF +``` diff --git a/docs/multimodal/MobileVLM.md b/docs/multimodal/MobileVLM.md index 20ac02f7a8..4f5eca6190 100644 --- a/docs/multimodal/MobileVLM.md +++ b/docs/multimodal/MobileVLM.md @@ -33,13 +33,13 @@ git clone https://huggingface.co/openai/clip-vit-large-patch14-336 2. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents: ```sh -python ./examples/llava/llava_surgery.py -m path/to/MobileVLM-1.7B +python ./tools/mtmd/llava_surgery.py -m path/to/MobileVLM-1.7B ``` 3. Use `convert_image_encoder_to_gguf.py` with `--projector-type ldp` (for **V2** please use `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF: ```sh -python ./examples/llava/convert_image_encoder_to_gguf.py \ +python ./tools/mtmd/convert_image_encoder_to_gguf.py \ -m path/to/clip-vit-large-patch14-336 \ --llava-projector path/to/MobileVLM-1.7B/llava.projector \ --output-dir path/to/MobileVLM-1.7B \ @@ -47,7 +47,7 @@ python ./examples/llava/convert_image_encoder_to_gguf.py \ ``` ```sh -python ./examples/llava/convert_image_encoder_to_gguf.py \ +python ./tools/mtmd/convert_image_encoder_to_gguf.py \ -m path/to/clip-vit-large-patch14-336 \ --llava-projector path/to/MobileVLM-1.7B_V2/llava.projector \ --output-dir path/to/MobileVLM-1.7B_V2 \ @@ -69,10 +69,10 @@ Now both the LLaMA part and the image encoder is in the `MobileVLM-1.7B` directo ## Android compile and run ### compile -refer to `examples/llava/android/build_64.sh` +refer to `tools/mtmd/android/build_64.sh` ```sh -mkdir examples/llava/android/build_64 -cd examples/llava/android/build_64 +mkdir tools/mtmd/android/build_64 +cd tools/mtmd/android/build_64 ../build_64.sh ``` ### run on Android diff --git a/docs/multimodal/glmedge.md b/docs/multimodal/glmedge.md index af6b696a8a..7bae831505 100644 --- a/docs/multimodal/glmedge.md +++ b/docs/multimodal/glmedge.md @@ -25,13 +25,13 @@ git clone https://huggingface.co/THUDM/glm-edge-v-5b or https://huggingface.co/T 2. Use `glmedge-surgery.py` to split the GLMV-EDGE model to LLM and multimodel projector constituents: ```sh -python ./examples/llava/glmedge-surgery.py -m ../model_path +python ./tools/mtmd/glmedge-surgery.py -m ../model_path ``` 4. Use `glmedge-convert-image-encoder-to-gguf.py` to convert the GLMV-EDGE image encoder to GGUF: ```sh -python ./examples/llava/glmedge-convert-image-encoder-to-gguf.py -m ../model_path --llava-projector ../model_path/glm.projector --output-dir ../model_path +python ./tools/mtmd/glmedge-convert-image-encoder-to-gguf.py -m ../model_path --llava-projector ../model_path/glm.projector --output-dir ../model_path ``` 5. Use `examples/convert_hf_to_gguf.py` to convert the LLM part of GLMV-EDGE to GGUF: diff --git a/docs/multimodal/llava.md b/docs/multimodal/llava.md index c5bdc82158..12354ab60a 100644 --- a/docs/multimodal/llava.md +++ b/docs/multimodal/llava.md @@ -37,19 +37,19 @@ git clone https://huggingface.co/openai/clip-vit-large-patch14-336 2. Install the required Python packages: ```sh -pip install -r examples/llava/requirements.txt +pip install -r tools/mtmd/requirements.txt ``` 3. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents: ```sh -python ./examples/llava/llava_surgery.py -m ../llava-v1.5-7b +python ./tools/mtmd/llava_surgery.py -m ../llava-v1.5-7b ``` 4. Use `convert_image_encoder_to_gguf.py` to convert the LLaVA image encoder to GGUF: ```sh -python ./examples/llava/convert_image_encoder_to_gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b +python ./tools/mtmd/convert_image_encoder_to_gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b ``` 5. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF: @@ -69,12 +69,12 @@ git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b 2) Install the required Python packages: ```sh -pip install -r examples/llava/requirements.txt +pip install -r tools/mtmd/requirements.txt ``` 3) Use `llava_surgery_v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models: ```console -python examples/llava/llava_surgery_v2.py -C -m ../llava-v1.6-vicuna-7b/ +python tools/mtmd/llava_surgery_v2.py -C -m ../llava-v1.6-vicuna-7b/ ``` - you will find a llava.projector and a llava.clip file in your model directory @@ -88,7 +88,7 @@ curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.jso 5) Create the visual gguf model: ```console -python ./examples/llava/convert_image_encoder_to_gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision +python ./tools/mtmd/convert_image_encoder_to_gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision ``` - This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP diff --git a/docs/multimodal/minicpmo2.6.md b/docs/multimodal/minicpmo2.6.md index de470d8a82..8c6db8efe5 100644 --- a/docs/multimodal/minicpmo2.6.md +++ b/docs/multimodal/minicpmo2.6.md @@ -29,8 +29,8 @@ cmake --build build --config Release Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-o-2_6-gguf) by us) ```bash -python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-o-2_6 -python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-2_6 --minicpmv-projector ../MiniCPM-o-2_6/minicpmv.projector --output-dir ../MiniCPM-o-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 4 +python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-o-2_6 +python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-2_6 --minicpmv-projector ../MiniCPM-o-2_6/minicpmv.projector --output-dir ../MiniCPM-o-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 4 python ./convert_hf_to_gguf.py ../MiniCPM-o-2_6/model # quantize int4 version diff --git a/docs/multimodal/minicpmv2.5.md b/docs/multimodal/minicpmv2.5.md index 7a6879d395..19b439607d 100644 --- a/docs/multimodal/minicpmv2.5.md +++ b/docs/multimodal/minicpmv2.5.md @@ -28,8 +28,8 @@ cmake --build build --config Release Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5-gguf) by us) ```bash -python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5 -python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2 +python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5 +python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2 python ./convert_hf_to_gguf.py ../MiniCPM-Llama3-V-2_5/model # quantize int4 version diff --git a/docs/multimodal/minicpmv2.6.md b/docs/multimodal/minicpmv2.6.md index 410a5dd177..15c1bbd12e 100644 --- a/docs/multimodal/minicpmv2.6.md +++ b/docs/multimodal/minicpmv2.6.md @@ -28,8 +28,8 @@ cmake --build build --config Release Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) by us) ```bash -python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-V-2_6 -python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 3 +python ./tools/mtmd/minicpmv-surgery.py -m ../MiniCPM-V-2_6 +python ./tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 3 python ./convert_hf_to_gguf.py ../MiniCPM-V-2_6/model # quantize int4 version diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt index 37476f9043..49e4d2cf8c 100644 --- a/examples/CMakeLists.txt +++ b/examples/CMakeLists.txt @@ -12,51 +12,30 @@ llama_add_compile_flags() # examples -include_directories(${CMAKE_CURRENT_SOURCE_DIR}) - if (EMSCRIPTEN) else() - add_subdirectory(batched-bench) add_subdirectory(batched) add_subdirectory(embedding) add_subdirectory(eval-callback) add_subdirectory(gguf-hash) - add_subdirectory(gguf-split) add_subdirectory(gguf) add_subdirectory(gritlm) - add_subdirectory(imatrix) - add_subdirectory(infill) - add_subdirectory(llama-bench) add_subdirectory(lookahead) add_subdirectory(lookup) - add_subdirectory(main) add_subdirectory(parallel) add_subdirectory(passkey) - add_subdirectory(perplexity) - add_subdirectory(quantize) add_subdirectory(retrieval) - if (LLAMA_BUILD_SERVER) - add_subdirectory(server) - endif() add_subdirectory(save-load-state) - add_subdirectory(run) add_subdirectory(simple) add_subdirectory(simple-chat) add_subdirectory(speculative) add_subdirectory(speculative-simple) - add_subdirectory(tokenize) - add_subdirectory(tts) add_subdirectory(gen-docs) + add_subdirectory(training) if (NOT GGML_BACKEND_DL) - # these examples use the backends directly and cannot be built with dynamic loading add_subdirectory(convert-llama2c-to-ggml) - add_subdirectory(cvector-generator) - add_subdirectory(export-lora) - add_subdirectory(llava) - if (GGML_RPC) - add_subdirectory(rpc) - endif() + # these examples use the backends directly and cannot be built with dynamic loading if (GGML_SYCL) add_subdirectory(sycl) endif() diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index 06fce236e2..01ff6763ff 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -35,23 +35,14 @@ static void batch_add_seq(llama_batch & batch, const std::vector & toke static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd, int embd_norm) { const enum llama_pooling_type pooling_type = llama_pooling_type(ctx); - const struct llama_model * model = llama_get_model(ctx); // clear previous kv_cache values (irrelevant for embeddings) llama_kv_self_clear(ctx); // run model LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq); - if (llama_model_has_encoder(model) && !llama_model_has_decoder(model)) { - // encoder-only model - if (llama_encode(ctx, batch) < 0) { - LOG_ERR("%s : failed to encode\n", __func__); - } - } else if (!llama_model_has_encoder(model) && llama_model_has_decoder(model)) { - // decoder-only model - if (llama_decode(ctx, batch) < 0) { - LOG_ERR("%s : failed to decode\n", __func__); - } + if (llama_encode(ctx, batch) < 0) { + LOG_ERR("%s : failed to encode\n", __func__); } for (int i = 0; i < batch.n_tokens; i++) { diff --git a/examples/infill/CMakeLists.txt b/examples/infill/CMakeLists.txt deleted file mode 100644 index fb26628d82..0000000000 --- a/examples/infill/CMakeLists.txt +++ /dev/null @@ -1,5 +0,0 @@ -set(TARGET llama-infill) -add_executable(${TARGET} infill.cpp) -install(TARGETS ${TARGET} RUNTIME) -target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) -target_compile_features(${TARGET} PRIVATE cxx_std_17) diff --git a/examples/infill/README.md b/examples/infill/README.md deleted file mode 100644 index df4d976f2b..0000000000 --- a/examples/infill/README.md +++ /dev/null @@ -1,47 +0,0 @@ -# llama.cpp/example/infill - -This example shows how to use the infill mode with Code Llama models supporting infill mode. -Currently the 7B and 13B models support infill mode. - -Infill supports most of the options available in the main example. - -For further information have a look at the main README.md in llama.cpp/example/main/README.md - -## Common Options - -In this section, we cover the most commonly used options for running the `infill` program with the LLaMA models: - -- `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`). -- `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses. -- `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text. -- `-c N, --ctx-size N`: Set the size of the prompt context. The default is 4096, but if a LLaMA model was built with a longer context, increasing this value will provide better results for longer input/inference. -- `--spm-infill`: Use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. - -## Input Prompts - -The `infill` program provides several ways to interact with the LLaMA models using input prompts: - -- `--in-prefix PROMPT_BEFORE_CURSOR`: Provide the prefix directly as a command-line option. -- `--in-suffix PROMPT_AFTER_CURSOR`: Provide the suffix directly as a command-line option. -- `--interactive-first`: Run the program in interactive mode and wait for input right away. (More on this below.) - -## Interaction - -The `infill` program offers a seamless way to interact with LLaMA models, allowing users to receive real-time infill suggestions. The interactive mode can be triggered using `--interactive`, and `--interactive-first` - -### Interaction Options - -- `-i, --interactive`: Run the program in interactive mode, allowing users to get real time code suggestions from model. -- `--interactive-first`: Run the program in interactive mode and immediately wait for user input before starting the text generation. -- `--color`: Enable colorized output to differentiate visually distinguishing between prompts, user input, and generated text. - -### Example - -Download a model that supports infill, for example CodeLlama: -```console -scripts/hf.sh --repo TheBloke/CodeLlama-13B-GGUF --file codellama-13b.Q5_K_S.gguf --outdir models -``` - -```bash -./llama-infill -t 10 -ngl 0 -m models/codellama-13b.Q5_K_S.gguf -c 4096 --temp 0.7 --repeat_penalty 1.1 -n 20 --in-prefix "def helloworld():\n print(\"hell" --in-suffix "\n print(\"goodbye world\")\n " -``` diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp deleted file mode 100644 index 4e2f7b7270..0000000000 --- a/examples/infill/infill.cpp +++ /dev/null @@ -1,590 +0,0 @@ -#include "arg.h" -#include "common.h" -#include "console.h" -#include "sampling.h" -#include "log.h" -#include "llama.h" - -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include - -#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) -#include -#include -#elif defined (_WIN32) -#define WIN32_LEAN_AND_MEAN -#ifndef NOMINMAX -#define NOMINMAX -#endif -#include -#include -#endif - -#if defined(_MSC_VER) -#pragma warning(disable: 4244 4267) // possible loss of data -#endif - -static llama_context ** g_ctx; -static llama_model ** g_model; -static common_sampler ** g_smpl; -static common_params * g_params; -static std::vector * g_input_tokens; -static std::ostringstream * g_output_ss; -static std::vector * g_output_tokens; - -static bool is_interacting = false; - -#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32) -static void sigint_handler(int signo) { - if (signo == SIGINT) { - if (!is_interacting) { - is_interacting = true; - } else { - console::cleanup(); - LOG("\n"); - common_perf_print(*g_ctx, *g_smpl); - - // make sure all logs are flushed - LOG("Interrupted by user\n"); - common_log_pause(common_log_main()); - - _exit(130); - } - } -} -#endif - -int main(int argc, char ** argv) { - common_params params; - g_params = ¶ms; - - if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_INFILL)) { - return 1; - } - - common_init(); - - auto & sparams = params.sampling; - - console::init(params.simple_io, params.use_color); - atexit([]() { console::cleanup(); }); - - if (params.logits_all) { - LOG_ERR("\n************\n"); - LOG_ERR("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__); - LOG_ERR("************\n\n"); - - return 0; - } - - if (params.embedding) { - LOG_ERR("\n************\n"); - LOG_ERR("%s: please use the 'embedding' tool for embedding calculations\n", __func__); - LOG_ERR("************\n\n"); - - return 0; - } - - if (params.n_ctx != 0 && params.n_ctx < 8) { - LOG_WRN("%s: minimum context size is 8, using minimum size.\n", __func__); - params.n_ctx = 8; - } - - if (!params.interactive_first && (params.input_prefix.empty() && params.input_suffix.empty())) { - LOG_ERR("\n************\n"); - LOG_ERR("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__); - LOG_ERR("************\n\n"); - - return 0; - } - - if (params.rope_freq_base != 0.0) { - LOG_WRN("%s: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base); - } - - if (params.rope_freq_scale != 0.0) { - LOG_WRN("%s: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale); - } - - LOG_INF("%s: llama backend init\n", __func__); - llama_backend_init(); - llama_numa_init(params.numa); - - llama_model * model = nullptr; - llama_context * ctx = nullptr; - common_sampler * smpl = nullptr; - - g_model = &model; - g_ctx = &ctx; - g_smpl = &smpl; - - // load the model and apply lora adapter, if any - LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__); - common_init_result llama_init = common_init_from_params(params); - - model = llama_init.model.get(); - ctx = llama_init.context.get(); - - if (model == NULL) { - LOG_ERR("%s: unable to load model\n", __func__); - return 1; - } - - const llama_vocab * vocab = llama_model_get_vocab(model); - - const int n_ctx_train = llama_model_n_ctx_train(model); - const int n_ctx = llama_n_ctx(ctx); - LOG_DBG("n_ctx: %d\n", n_ctx); - - if (n_ctx > n_ctx_train) { - LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n", __func__, n_ctx_train, n_ctx); - } - - // print system information - { - LOG_INF("\n"); - LOG_INF("%s\n", common_params_get_system_info(params).c_str()); - } - const bool add_bos = llama_vocab_get_add_bos(vocab); - GGML_ASSERT(!llama_vocab_get_add_eos(vocab)); - - std::vector embd_inp; - std::vector embd_end; - std::vector inp_pfx = common_tokenize(ctx, params.input_prefix, false); - std::vector inp_sfx = common_tokenize(ctx, params.input_suffix, false); - - GGML_ASSERT(llama_vocab_fim_pre(vocab) >= 0); - GGML_ASSERT(llama_vocab_fim_suf(vocab) >= 0); - - inp_pfx.insert(inp_pfx.begin(), llama_vocab_fim_pre(vocab)); - inp_sfx.insert(inp_sfx.begin(), llama_vocab_fim_suf(vocab)); - - embd_inp = params.spm_infill ? inp_sfx : inp_pfx; - embd_end = params.spm_infill ? inp_pfx : inp_sfx; - if (add_bos) { - embd_inp.insert(embd_inp.begin(), llama_vocab_bos(vocab)); - } - embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end()); - - const llama_token middle_token = llama_vocab_fim_mid(vocab); - if (middle_token >= 0) { - embd_inp.push_back(middle_token); - } - - LOG_DBG("add_bos: %d\n", add_bos); - LOG_DBG("prefix: \"%s\"\n", params.input_prefix.c_str()); - LOG_DBG("suffix: \"%s\"\n", params.input_suffix.c_str()); - LOG_DBG("tokens: %s\n", string_from(ctx, embd_inp).c_str()); - - // Should not run without any tokens - if (embd_inp.empty()) { - embd_inp.push_back(llama_vocab_bos(vocab)); - LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str()); - } - - if ((int) embd_inp.size() > n_ctx - 4) { - LOG_ERR("%s: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4); - return 1; - } - - // number of tokens to keep when resetting context - if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size()) { - params.n_keep = (int)embd_inp.size(); - } - - LOG_INF("inp_pfx: %s\n", string_from(ctx, inp_pfx).c_str()); - LOG_INF("inp_sfx: %s\n", string_from(ctx, inp_sfx).c_str()); - - // enable interactive mode if interactive start is specified - if (params.interactive_first) { - params.interactive = true; - } - - if (params.verbose_prompt) { - LOG_INF("\n"); - LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str()); - LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size()); - for (int i = 0; i < (int) embd_inp.size(); i++) { - LOG_INF("%6d -> '%s'\n", embd_inp[i], common_token_to_piece(ctx, embd_inp[i]).c_str()); - } - - if (params.n_keep > 0) { - LOG_INF("%s: static prompt based on n_keep: '", __func__); - for (int i = 0; i < params.n_keep; i++) { - LOG_CNT("%s", common_token_to_piece(ctx, embd_inp[i]).c_str()); - } - LOG_CNT("'\n"); - } - LOG_INF("\n"); - } - - if (params.interactive) { -#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) - struct sigaction sigint_action; - sigint_action.sa_handler = sigint_handler; - sigemptyset (&sigint_action.sa_mask); - sigint_action.sa_flags = 0; - sigaction(SIGINT, &sigint_action, NULL); -#elif defined (_WIN32) - auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL { - return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false; - }; - SetConsoleCtrlHandler(reinterpret_cast(console_ctrl_handler), true); -#endif - - LOG_INF("%s: interactive mode on.\n", __func__); - - if (params.input_prefix_bos) { - LOG_INF("Input prefix with BOS\n"); - } - - if (!params.input_prefix.empty()) { - LOG_INF("Input prefix: '%s'\n", params.input_prefix.c_str()); - } - - if (!params.input_suffix.empty()) { - LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str()); - } - } - smpl = common_sampler_init(model, sparams); - - LOG_INF("sampler seed: %u\n", common_sampler_get_seed(smpl)); - LOG_INF("sampler params: \n%s\n", sparams.print().c_str()); - LOG_INF("sampler chain: %s\n", common_sampler_print(smpl).c_str()); - - LOG_INF("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep); - - LOG_INF("\n"); - LOG_INF("\n##### Infill mode #####\n\n"); - if (params.interactive) { - const char *control_message; - if (params.multiline_input) { - control_message = " - To return control to LLaMA, end your input with '\\'.\n" - " - To return control without starting a new line, end your input with '/'.\n"; - } else { - control_message = " - Press Return to return control to LLaMA.\n" - " - To return control without starting a new line, end your input with '/'.\n" - " - If you want to submit another line, end your input with '\\'.\n"; - } - LOG_INF("== Running in interactive mode. ==\n"); -#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32) - LOG_INF( " - Press Ctrl+C to interject at any time.\n"); -#endif - LOG_INF( "%s\n", control_message); - - is_interacting = params.interactive_first; - } - - bool input_echo = true; - - int n_past = 0; - int n_remain = params.n_predict; - int n_consumed = 0; - - std::vector input_tokens; g_input_tokens = &input_tokens; - std::vector output_tokens; g_output_tokens = &output_tokens; - std::ostringstream output_ss; g_output_ss = &output_ss; - - // the first thing we will do is to output the prompt, so set color accordingly - console::set_display(console::prompt); - - std::vector embd; - - while (n_remain != 0 || params.interactive) { - // predict - if (!embd.empty()) { - // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via - // --prompt or --file which uses the same value. - int max_embd_size = n_ctx - 4; - - // Ensure the input doesn't exceed the context size by truncating embd if necessary. - if ((int) embd.size() > max_embd_size) { - const int skipped_tokens = (int) embd.size() - max_embd_size; - embd.resize(max_embd_size); - - console::set_display(console::error); - LOG_WRN("<>", skipped_tokens, skipped_tokens != 1 ? "s" : ""); - console::set_display(console::reset); - } - - // infinite text generation via context swapping - // if we run out of context: - // - take the n_keep first tokens from the original prompt (via n_past) - // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches - if (n_past + (int) embd.size() > n_ctx) { - if (params.n_predict == -2) { - LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict); - break; - } - - const int n_left = n_past - params.n_keep - 1; - const int n_discard = n_left/2; - - LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n", - n_past, n_left, n_ctx, params.n_keep, n_discard); - - llama_kv_self_seq_rm (ctx, 0, params.n_keep + 1 , params.n_keep + n_discard + 1); - llama_kv_self_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard); - - n_past -= n_discard; - - LOG_DBG("after swap: n_past = %d\n", n_past); - - LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str()); - - } - - // evaluate tokens in batches - // embd is typically prepared beforehand to fit within a batch, but not always - for (int i = 0; i < (int) embd.size(); i += params.n_batch) { - int n_eval = (int) embd.size() - i; - if (n_eval > params.n_batch) { - n_eval = params.n_batch; - } - - LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str()); - - if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) { - LOG_ERR("%s : failed to eval\n", __func__); - return 1; - } - - n_past += n_eval; - - LOG_DBG("n_past = %d\n", n_past); - } - - } - - embd.clear(); - - if ((int) embd_inp.size() <= n_consumed && !is_interacting) { - const llama_token id = common_sampler_sample(smpl, ctx, -1); - - common_sampler_accept(smpl, id, true); - - // LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str()); - - embd.push_back(id); - - // echo this to console - input_echo = true; - - // decrement remaining sampling budget - --n_remain; - - LOG_DBG("n_remain: %d\n", n_remain); - } else { - // some user input remains from prompt or interaction, forward it to processing - LOG_DBG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed); - while ((int) embd_inp.size() > n_consumed) { - embd.push_back(embd_inp[n_consumed]); - - // push the prompt in the sampling context in order to apply repetition penalties later - // for the prompt, we don't apply grammar rules - common_sampler_accept(smpl, embd_inp[n_consumed], false); - - ++n_consumed; - if ((int) embd.size() >= params.n_batch) { - break; - } - } - } - - // display text - if (input_echo) { - for (auto id : embd) { - const std::string token_str = common_token_to_piece(ctx, id); - LOG("%s", token_str.c_str()); - - if (embd.size() > 1) { - input_tokens.push_back(id); - } else { - output_tokens.push_back(id); - output_ss << token_str; - } - } - } - // reset color to default if we there is no pending user input - if (input_echo && (int) embd_inp.size() == n_consumed) { - console::set_display(console::reset); - } - - // if not currently processing queued inputs; - if ((int) embd_inp.size() <= n_consumed) { - // deal with eot token in infill mode - if ((common_sampler_last(smpl) == llama_vocab_eot(vocab) || is_interacting) && params.interactive){ - if (is_interacting && !params.interactive_first) { - // print an eot token - LOG("%s", common_token_to_piece(ctx, llama_vocab_eot(vocab)).c_str()); - } - LOG("\n"); - console::set_display(console::user_input); - std::string buffer; - std::string line; - bool another_line=true; - // set a new prefix via stdin - do { - another_line = console::readline(line, params.multiline_input); - buffer += line; - } while (another_line); - // check if we got an empty line, if so we use the old input - if (!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) { - params.input_prefix = buffer; - } - buffer.clear(); - // set a new suffix via stdin - do { - another_line = console::readline(line, params.multiline_input); - buffer += line; - } while (another_line); - // check if we got an empty line - if (!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) { - params.input_suffix = buffer; - } - buffer.clear(); - // done taking input, reset color - console::set_display(console::reset); - - if (params.escape) { - //process escape sequences, for the initial prompt this is done in common.cpp when we load the params, but for the interactive mode we need to do it here - string_process_escapes(params.input_prefix); - string_process_escapes(params.input_suffix); - } - - // tokenize new prefix and suffix - std::vector inp_pfx = common_tokenize(ctx, params.input_prefix, false); - std::vector inp_sfx = common_tokenize(ctx, params.input_suffix, false); - - inp_pfx.insert(inp_pfx.begin(), llama_vocab_fim_pre(vocab)); - inp_sfx.insert(inp_sfx.begin(), llama_vocab_fim_suf(vocab)); - - embd_inp = params.spm_infill ? inp_sfx : inp_pfx; - embd_end = params.spm_infill ? inp_pfx : inp_sfx; - if (add_bos) { - embd_inp.insert(embd_inp.begin(), llama_vocab_bos(vocab)); - } - embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end()); - - if (middle_token >= 0) { - embd_inp.push_back(middle_token); - } - - embd.clear(); - n_remain = params.n_predict; - n_past = 0; - n_consumed = 0; - is_interacting = false; - } - // deal with end of generation tokens in interactive mode - else if (llama_vocab_is_eog(vocab, common_sampler_last(smpl))) { - LOG_DBG("found EOS token\n"); - - if (params.interactive) { - - is_interacting = true; - LOG("\n"); - console::set_display(console::user_input); - } - } - - if (n_past > 0 && is_interacting && !params.interactive) { - LOG_DBG("waiting for user input\n"); - - if (params.input_prefix_bos) { - LOG_DBG("adding input prefix BOS token\n"); - embd_inp.push_back(llama_vocab_bos(vocab)); - } - - std::string buffer; - if (!params.input_prefix.empty()) { - LOG_DBG("appending input prefix: '%s'\n", params.input_prefix.c_str()); - buffer += params.input_prefix; - LOG("%s", buffer.c_str()); - } - - std::string line; - bool another_line = true; - do { - another_line = console::readline(line, params.multiline_input); - buffer += line; - } while (another_line); - - // done taking input, reset color - console::set_display(console::reset); - - // Add tokens to embd only if the input buffer is non-empty - // Entering a empty line lets the user pass control back - if (buffer.length() > 1) { - // append input suffix if any - if (!params.input_suffix.empty()) { - LOG_DBG("appending input suffix: '%s'\n", params.input_suffix.c_str()); - buffer += params.input_suffix; - LOG("%s", params.input_suffix.c_str()); - } - - LOG_DBG("buffer: '%s'\n", buffer.c_str()); - - const size_t original_size = embd_inp.size(); - - const auto line_inp = common_tokenize(ctx, buffer, false); - LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str()); - - embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end()); - - for (size_t i = original_size; i < embd_inp.size(); ++i) { - const llama_token token = embd_inp[i]; - output_tokens.push_back(token); - output_ss << common_token_to_piece(ctx, token); - } - - n_remain -= line_inp.size(); - LOG_DBG("n_remain: %d\n", n_remain); - } else { - LOG_DBG("empty line, passing control back\n"); - } - - input_echo = false; // do not echo this again - } - - if (n_past > 0) { - if (is_interacting) { - common_sampler_reset(smpl); - } - is_interacting = false; - } - } - - // end of generation - if (!embd.empty() && llama_vocab_is_eog(vocab, embd.back()) && !params.interactive) { - break; - } - - // In interactive mode, respect the maximum number of tokens and drop back to user input when reached. - // We skip this logic when n_predict == -1 (infinite) or -2 (stop at context size). - if (params.interactive && n_remain <= 0 && params.n_predict >= 0) { - n_remain = params.n_predict; - is_interacting = true; - } - } - if (!params.interactive && n_remain <= 0) { - LOG("%s", common_token_to_piece(ctx, llama_vocab_eot(vocab)).c_str()); - } - - LOG("\n"); - common_perf_print(ctx, smpl); - - common_sampler_free(smpl); - llama_backend_free(); - - return 0; -} diff --git a/examples/json_schema_to_grammar.py b/examples/json_schema_to_grammar.py index 55f94c0b0a..ed37958554 100755 --- a/examples/json_schema_to_grammar.py +++ b/examples/json_schema_to_grammar.py @@ -10,6 +10,9 @@ from typing import Any, List, Optional, Set, Tuple, Union def _build_repetition(item_rule, min_items, max_items, separator_rule=None): + if max_items == 0: + return "" + if min_items == 0 and max_items == 1: return f'{item_rule}?' diff --git a/examples/llava/mtmd.h b/examples/llava/mtmd.h deleted file mode 100644 index 78be192dd6..0000000000 --- a/examples/llava/mtmd.h +++ /dev/null @@ -1,161 +0,0 @@ -#ifndef MTMD_H -#define MTMD_H - -#include "ggml.h" -#include "llama.h" -#include "clip.h" - -#include -#include -#include - -#ifdef LLAMA_SHARED -# if defined(_WIN32) && !defined(__MINGW32__) -# ifdef LLAMA_BUILD -# define MTMD_API __declspec(dllexport) -# else -# define MTMD_API __declspec(dllimport) -# endif -# else -# define MTMD_API __attribute__ ((visibility ("default"))) -# endif -#else -# define MTMD_API -#endif - -#ifdef __cplusplus - -enum mtmd_input_chunk_type { - MTMD_INPUT_CHUNK_TYPE_TEXT, - MTMD_INPUT_CHUNK_TYPE_IMAGE, -}; - -struct mtmd_context; -struct mtmd_image_tokens; - -// represents raw image data, layout is RGBRGBRGB... -// length of data must be nx * ny * 3 -struct mtmd_bitmap { - uint32_t nx; - uint32_t ny; - std::vector data; - std::string id; // optional user-defined id, for ex: can be set to image hash, useful for KV cache tracking -}; - -struct mtmd_image_tokens_deleter { - void operator()(mtmd_image_tokens * val); // forward declaration -}; -using mtmd_image_tokens_ptr = std::unique_ptr; - -struct mtmd_input_chunk { - mtmd_input_chunk_type type; - std::vector tokens_text; - mtmd_image_tokens_ptr tokens_image; -}; - -using mtmd_input_chunks = std::vector; - -struct mtmd_context_params { - bool use_gpu = true; - bool print_timings = true; - int n_threads = 4; - enum ggml_log_level verbosity = GGML_LOG_LEVEL_INFO; - const char * image_marker = "<__image__>"; -}; - -struct mtmd_input_text { - std::string text; - bool add_special; - bool parse_special; -}; - -// initialize the mtmd context -// return nullptr on failure -MTMD_API mtmd_context * mtmd_init_from_file(const char * mmproj_fname, - const llama_model * text_model, - const mtmd_context_params ctx_params); - -MTMD_API void mtmd_free(mtmd_context * ctx); - -// tokenize an input text prompt and an image -// the prompt must have the input image marker (default: "<__image__>") in it -// the marker will be replaced with the image tokens -// for example: -// "here is an image: <__image__>\ndescribe it in detail." -// this will gives 3 chunks: -// 1. "here is an image: " -// 2. (image tokens) -// 3. "\ndescribe it in detail." -// number of bitmaps must be equal to the number of image markers in the prompt -// this function is thread-safe (shared ctx) -// return values: -// 0 on success -// 1 on number of images not matching the number of markers -// 2 on image preprocessing error -MTMD_API int32_t mtmd_tokenize(mtmd_context * ctx, - std::vector & output, - const mtmd_input_text & text, - const std::vector & bitmaps); - -// access mtmd_image_tokens -MTMD_API size_t mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens); -MTMD_API size_t mtmd_image_tokens_get_nx(const mtmd_image_tokens * image_tokens); -MTMD_API size_t mtmd_image_tokens_get_ny(const mtmd_image_tokens * image_tokens); -MTMD_API std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens); -MTMD_API void mtmd_image_tokens_free(mtmd_image_tokens * image_tokens); - -// returns 0 on success -MTMD_API int32_t mtmd_encode(mtmd_context * ctx, - const mtmd_image_tokens * image_tokens); - -// get output embeddings from the last encode pass -MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx); - -// whether we need to set non-causal mask before llama_decode -MTMD_API bool mtmd_decode_use_non_causal(mtmd_context * ctx); - - - -// -// helper functions (can be implemented based on other functions) -// - -// helper to count the total number of tokens from a list of chunks, useful to keep track of n_past -MTMD_API size_t mtmd_helper_get_n_tokens(mtmd_input_chunks & chunks); - -// helper function that automatically: -// 1. run llama_decode() on text chunks -// 2. run mtmd_encode() on image chunks, then mtmd_get_output_embd() and then llama_decode() -// if any of the mtmd_encode() or llama_decode() calls return non-zero, stop and forward the error -// otherwise, returns 0 on success -MTMD_API int32_t mtmd_helper_eval(mtmd_context * ctx, - llama_context * lctx, - mtmd_input_chunks & chunks, - llama_pos pos0, - llama_seq_id seq_id, - int32_t n_batch); - -// helper function to construct a mtmd_bitmap from a file -// returns 0 on success -// this function is thread-safe -MTMD_API int32_t mtmd_helper_bitmap_init_from_file(const char * fname, mtmd_bitmap & output); - -// helper function to construct a mtmd_bitmap from a buffer -// the buffer must be an image in format supported by stb_image (jpg, png, bmp, gif, etc.) -// returns 0 on success -// this function is thread-safe -MTMD_API int32_t mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len, mtmd_bitmap & output); - -// convenient unique_ptr wrappers -struct mtmd_context_deleter { - void operator()(mtmd_context * val) { mtmd_free(val); } -}; -using mtmd_context_ptr = std::unique_ptr; - -#else - -static_assert(false && "C header is not yet supported by this library"); - -#endif - -#endif diff --git a/examples/llava/qwen2_vl_surgery.py b/examples/llava/qwen2_vl_surgery.py deleted file mode 100644 index c87606b4fd..0000000000 --- a/examples/llava/qwen2_vl_surgery.py +++ /dev/null @@ -1,165 +0,0 @@ -import argparse -from typing import Dict - -import torch -import numpy as np -from gguf import * -from transformers import ( - Qwen2VLForConditionalGeneration, - Qwen2VLProcessor, - AutoProcessor, - Qwen2VLConfig -) - - -VISION = "clip.vision" - - -def k(raw_key: str, arch: str) -> str: - return raw_key.format(arch=arch) - - -def to_gguf_name(name: str) -> str: - og = name - name = name.replace("text_model", "t").replace("vision_model", "v") - name = name.replace("blocks", "blk").replace("embeddings.", "") - name = name.replace("attn.", "attn_") - name = name.replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("proj.", "out.") - # name = name.replace("layrnorm", "ln").replace("layer_norm", "ln").replace("layernorm", "ln") - name = name.replace("norm1", "ln1").replace("norm2", "ln2") - name = name.replace("merger.mlp", 'mm') - print(f"[to_gguf_name] {og} --> {name}") - return name - - -def find_vision_tensors(qwen2vl, dtype) -> Dict[str, np.ndarray]: - vision_model = qwen2vl.visual - tensor_map = {} - for name, ten in vision_model.state_dict().items(): - ten = ten.numpy() - if 'qkv' in name: - if ten.ndim == 2: # weight - c3, _ = ten.shape - else: # bias - c3 = ten.shape[0] - assert c3 % 3 == 0 - c = c3 // 3 - wq = ten[:c] - wk = ten[c: c * 2] - wv = ten[c * 2:] - tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "q")] = wq - tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "k")] = wk - tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "v")] = wv - elif 'merger' in name: - if name.endswith("ln_q.weight"): - tensor_map['v.post_ln.weight'] = ten - elif name.endswith("ln_q.bias"): - tensor_map['v.post_ln.bias'] = ten - else: - # "merger.mlp.%d.weight/bias" --> "mm.%d.weight/bias" - tensor_map[to_gguf_name(name)] = ten - elif 'patch_embed.proj.weight' in name: - # NOTE: split Conv3D into Conv2Ds - c1, c2, kt, kh, kw = ten.shape - assert kt == 2, "Current implmentation only support temporal_patch_size of 2" - tensor_map["v.patch_embd.weight"] = ten[:, :, 0, ...] - tensor_map["v.patch_embd.weight.1"] = ten[:, :, 1, ...] - else: - tensor_map[to_gguf_name(f"vision_model.{name}")] = ten - - for new_name, ten in tensor_map.items(): - if ten.ndim <= 1 or new_name.endswith("_norm.weight"): - tensor_map[new_name] = ten.astype(np.float32) - else: - tensor_map[new_name] = ten.astype(dtype) - tensor_map["v.position_embd.weight"] = np.zeros([10, 10], dtype=np.float32) # dummy tensor, just here as a placeholder - return tensor_map - - -def main(args): - if args.data_type == 'fp32': - dtype = torch.float32 - np_dtype = np.float32 - ftype = 0 - elif args.data_type == 'fp16': - dtype = torch.float32 - np_dtype = np.float16 - ftype = 1 - else: - raise ValueError() - - local_model = False - model_path = "" - model_name = args.model_name - print("model_name: ", model_name) - qwen2vl = Qwen2VLForConditionalGeneration.from_pretrained( - model_name, torch_dtype=dtype, device_map="cpu" - ) - cfg: Qwen2VLConfig = qwen2vl.config # type: ignore[reportAssignmentType] - vcfg = cfg.vision_config - - if os.path.isdir(model_name): - local_model = True - if model_name.endswith(os.sep): - model_name = model_name[:-1] - model_path = model_name - model_name = os.path.basename(model_name) - fname_out = f"{model_name.replace('/', '-').lower()}-vision.gguf" - - fout = GGUFWriter(path=fname_out, arch="clip") - fout.add_description("image encoder for Qwen2VL") - - fout.add_file_type(ftype) - fout.add_bool("clip.has_text_encoder", False) - fout.add_bool("clip.has_vision_encoder", True) - fout.add_bool("clip.has_qwen2vl_merger", True) - fout.add_string("clip.projector_type", "qwen2vl_merger") - - print(cfg.vision_config) - if 'silu' in cfg.vision_config.hidden_act.lower(): - fout.add_bool("clip.use_silu", True) - fout.add_bool("clip.use_gelu", False) - elif 'gelu' in cfg.vision_config.hidden_act.lower(): - fout.add_bool("clip.use_silu", False) - fout.add_bool("clip.use_gelu", 'quick' not in cfg.vision_config.hidden_act.lower()) - else: - raise ValueError() - - tensor_map = find_vision_tensors(qwen2vl, np_dtype) - for name, data in tensor_map.items(): - fout.add_tensor(name, data) - - fout.add_uint32("clip.vision.patch_size", vcfg.patch_size) - fout.add_uint32("clip.vision.image_size", 14 * 40) # some reasonable size that is divable by (14*2) - fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.embed_dim) - fout.add_uint32("clip.vision.projection_dim", vcfg.hidden_size) - fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), vcfg.num_heads) - fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6) - fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), vcfg.depth) - fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), 0) # not sure what this does, put 0 here as a placeholder - fout.add_name(model_name) - """ - HACK: Since vision rope related parameter aren't stored in the `Qwen2VLConfig, - it will be hardcoded in the `clip_image_build_graph` from `clip.cpp`. - """ - - if local_model: - processor: Qwen2VLProcessor = AutoProcessor.from_pretrained(model_path) - else: - processor: Qwen2VLProcessor = AutoProcessor.from_pretrained(model_name) - fout.add_array("clip.vision.image_mean", processor.image_processor.image_mean) # type: ignore[reportAttributeAccessIssue] - fout.add_array("clip.vision.image_std", processor.image_processor.image_std) # type: ignore[reportAttributeAccessIssue] - - fout.write_header_to_file() - fout.write_kv_data_to_file() - fout.write_tensors_to_file() - fout.close() - print("save model as: ", fname_out) - - -if __name__ == "__main__": - parser = argparse.ArgumentParser() - parser.add_argument("model_name", nargs='?', default="Qwen/Qwen2-VL-2B-Instruct") - parser.add_argument("--data_type", nargs='?', choices=['fp32', 'fp16'], default="fp32") - args = parser.parse_args() - main(args) diff --git a/examples/pydantic_models_to_grammar_examples.py b/examples/pydantic_models_to_grammar_examples.py index f94b82ca47..6dadb7f3fa 100755 --- a/examples/pydantic_models_to_grammar_examples.py +++ b/examples/pydantic_models_to_grammar_examples.py @@ -23,7 +23,7 @@ def create_completion(host, prompt, gbnf_grammar): """Calls the /completion API on llama-server. See - https://github.com/ggml-org/llama.cpp/tree/HEAD/examples/server#api-endpoints + https://github.com/ggml-org/llama.cpp/tree/HEAD/tools/server#api-endpoints """ print(f" Request:\n Grammar:\n{textwrap.indent(gbnf_grammar, ' ')}\n Prompt:\n{textwrap.indent(prompt.rstrip(), ' ')}") headers = {"Content-Type": "application/json"} diff --git a/examples/server/public/index.html.gz b/examples/server/public/index.html.gz deleted file mode 100644 index 674e227571..0000000000 Binary files a/examples/server/public/index.html.gz and /dev/null differ diff --git a/examples/server/webui/src/components/Header.tsx b/examples/server/webui/src/components/Header.tsx deleted file mode 100644 index 4c6b291e61..0000000000 --- a/examples/server/webui/src/components/Header.tsx +++ /dev/null @@ -1,178 +0,0 @@ -import { useEffect, useState } from 'react'; -import StorageUtils from '../utils/storage'; -import { useAppContext } from '../utils/app.context'; -import { classNames } from '../utils/misc'; -import daisyuiThemes from 'daisyui/theme/object'; -import { THEMES } from '../Config'; -import { useNavigate } from 'react-router'; - -export default function Header() { - const navigate = useNavigate(); - const [selectedTheme, setSelectedTheme] = useState(StorageUtils.getTheme()); - const { setShowSettings } = useAppContext(); - - const setTheme = (theme: string) => { - StorageUtils.setTheme(theme); - setSelectedTheme(theme); - }; - - useEffect(() => { - document.body.setAttribute('data-theme', selectedTheme); - document.body.setAttribute( - 'data-color-scheme', - daisyuiThemes[selectedTheme]?.['color-scheme'] ?? 'auto' - ); - }, [selectedTheme]); - - const { isGenerating, viewingChat } = useAppContext(); - const isCurrConvGenerating = isGenerating(viewingChat?.conv.id ?? ''); - - const removeConversation = () => { - if (isCurrConvGenerating || !viewingChat) return; - const convId = viewingChat?.conv.id; - if (window.confirm('Are you sure to delete this conversation?')) { - StorageUtils.remove(convId); - navigate('/'); - } - }; - - const downloadConversation = () => { - if (isCurrConvGenerating || !viewingChat) return; - const convId = viewingChat?.conv.id; - const conversationJson = JSON.stringify(viewingChat, null, 2); - const blob = new Blob([conversationJson], { type: 'application/json' }); - const url = URL.createObjectURL(blob); - const a = document.createElement('a'); - a.href = url; - a.download = `conversation_${convId}.json`; - document.body.appendChild(a); - a.click(); - document.body.removeChild(a); - URL.revokeObjectURL(url); - }; - - return ( -
- {/* open sidebar button */} - - -
llama.cpp
- - {/* action buttons (top right) */} -
- {viewingChat && ( -
- {/* "..." button */} - - {/* dropdown menu */} - -
- )} - -
- -
- - {/* theme controller is copied from https://daisyui.com/components/theme-controller/ */} -
-
-
- - - -
-
    -
  • - -
    • - {THEMES.map((theme) => ( -
  • - e.target.checked && setTheme(theme)} - /> -
    • - ))} -
-
-
-
-
- ); -} diff --git a/examples/server/webui/src/components/Sidebar.tsx b/examples/server/webui/src/components/Sidebar.tsx deleted file mode 100644 index 34727c6231..0000000000 --- a/examples/server/webui/src/components/Sidebar.tsx +++ /dev/null @@ -1,96 +0,0 @@ -import { useEffect, useState } from 'react'; -import { classNames } from '../utils/misc'; -import { Conversation } from '../utils/types'; -import StorageUtils from '../utils/storage'; -import { useNavigate, useParams } from 'react-router'; - -export default function Sidebar() { - const params = useParams(); - const navigate = useNavigate(); - - const [conversations, setConversations] = useState([]); - const [currConv, setCurrConv] = useState(null); - - useEffect(() => { - StorageUtils.getOneConversation(params.convId ?? '').then(setCurrConv); - }, [params.convId]); - - useEffect(() => { - const handleConversationChange = async () => { - setConversations(await StorageUtils.getAllConversations()); - }; - StorageUtils.onConversationChanged(handleConversationChange); - handleConversationChange(); - return () => { - StorageUtils.offConversationChanged(handleConversationChange); - }; - }, []); - - return ( - <> - - -
- -
-
-

Conversations

- - {/* close sidebar button */} - -
- - {/* list of conversations */} -
navigate('/')} - > - + New conversation -
- {conversations.map((conv) => ( -
navigate(`/chat/${conv.id}`)} - dir="auto" - > - {conv.name} -
- ))} -
- Conversations are saved to browser's IndexedDB -
-
-
- - ); -} diff --git a/examples/training/CMakeLists.txt b/examples/training/CMakeLists.txt new file mode 100644 index 0000000000..64afe6ddc6 --- /dev/null +++ b/examples/training/CMakeLists.txt @@ -0,0 +1,5 @@ +set(TARGET llama-finetune) +add_executable(${TARGET} finetune.cpp) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/training/README.md b/examples/training/README.md new file mode 100644 index 0000000000..ecdf398f81 --- /dev/null +++ b/examples/training/README.md @@ -0,0 +1,17 @@ +# llama.cpp/examples/training + +This directory contains examples related to language model training using llama.cpp/GGML. +So far finetuning is technically functional (for FP32 models and limited hardware setups) but the code is very much WIP. +Finetuning of Stories 260K and LLaMA 3.2 1b seems to work with 24 GB of memory. +**For CPU training, compile llama.cpp without any additional backends such as CUDA.** +**For CUDA training, use the maximum number of GPU layers.** + +Proof of concept: + +``` sh +export model_name=llama_3.2-1b && export quantization=f32 +./build/bin/finetune --file wikitext-2-raw/wiki.test.raw -ngl 999 --model models/${model_name}-${quantization}.gguf -c 512 -b 512 -ub 512 +./build/bin/perplexity --file wikitext-2-raw/wiki.test.raw -ngl 999 --model finetuned-model.gguf +``` + +The perplexity value of the finetuned model should be lower after training on the test set for 2 epochs. diff --git a/examples/training/finetune.cpp b/examples/training/finetune.cpp new file mode 100644 index 0000000000..23bede49b1 --- /dev/null +++ b/examples/training/finetune.cpp @@ -0,0 +1,96 @@ +#include "arg.h" +#include "common.h" +#include "log.h" +#include "llama.h" + +#include +#include +#include +#include +#include + +#if defined(_MSC_VER) +#pragma warning(disable: 4244 4267) // possible loss of data +#endif + +int main(int argc, char ** argv) { + common_params params; + + params.escape = false; + + if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) { + return 1; + } + + if (params.use_mmap) { + LOG_INF("%s: force disabling memory mapping because it would result in-read-only pointers to the weights\n", __func__); + params.use_mmap = false; + } + if (params.cache_type_k != GGML_TYPE_F32) { + LOG_INF("%s: force changing k cache type to f32 due to a lack of f16 support for OUT_PROD\n", __func__); + params.cache_type_k = GGML_TYPE_F32; + } + if (params.cache_type_v != GGML_TYPE_F32) { + LOG_INF("%s: force changing v cache type to f32 due to a lack of f16 support for OUT_PROD\n", __func__); + params.cache_type_v = GGML_TYPE_F32; + } + + common_init(); + llama_backend_init(); + llama_numa_init(params.numa); + + // load the model and apply lora adapter, if any + common_init_result llama_init = common_init_from_params(params); + llama_model_ptr & model = llama_init.model; + llama_context_ptr & ctx = llama_init.context; + + if (model == NULL) { + LOG_ERR("%s: unable to load model\n", __func__); + return 1; + } + + // print system information + { + LOG_INF("\n"); + LOG_INF("%s\n", common_params_get_system_info(params).c_str()); + } + + constexpr float val_split = 0.05f; + + std::vector tokens = common_tokenize(ctx.get(), params.prompt, true); + ggml_opt_dataset_t dataset = common_opt_dataset_init(ctx.get(), tokens, llama_n_ctx(ctx.get())/2); + + struct ggml_opt_optimizer_params optimizer_params = ggml_opt_get_default_optimizer_params(nullptr); + optimizer_params.adamw.alpha = 1e-7f; // learning rate + + struct llama_opt_params lopt_params { + /*n_ctx_train =*/ 0, + /*param_filter =*/ llama_opt_param_filter_all, + /*param_filter_ud =*/ nullptr, + /*get_opt_pars =*/ ggml_opt_get_constant_optimizer_params, + /*get_opt_pars_ud =*/ &optimizer_params, + }; + llama_opt_init(ctx.get(), model.get(), lopt_params); + + const int64_t idata_split = ggml_opt_dataset_ndata(dataset) * (1.0f - val_split); + + ggml_opt_result_t result_train = ggml_opt_result_init(); + ggml_opt_result_t result_eval = ggml_opt_result_init(); + + for (int epoch = 0; epoch < 2; ++epoch) { + llama_opt_epoch(ctx.get(), dataset, result_train, result_eval, idata_split, + ggml_opt_epoch_callback_progress_bar, ggml_opt_epoch_callback_progress_bar); + fprintf(stderr, "\n"); + + ggml_opt_result_reset(result_train); + ggml_opt_result_reset(result_eval); + } + ggml_opt_result_free(result_train); + ggml_opt_result_free(result_eval); + + llama_model_save_to_file(model.get(), "finetuned-model.gguf"); + + llama_backend_free(); + + return 0; +} diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt index 61fe15a15f..a8300e16d8 100644 --- a/ggml/CMakeLists.txt +++ b/ggml/CMakeLists.txt @@ -360,3 +360,29 @@ write_basic_package_version_file( install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake ${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml) + +if (MSVC) + set(MSVC_WARNING_FLAGS + /wd4005 # Macro redefinition + /wd4244 # Conversion from one type to another type, possible loss of data + /wd4267 # Conversion from 'size_t' to a smaller type, possible loss of data + /wd4996 # Disable POSIX deprecation warnings + /wd4702 # Unreachable code warnings + ) + function(disable_msvc_warnings target_name) + if(TARGET ${target_name}) + target_compile_options(${target_name} PRIVATE ${MSVC_WARNING_FLAGS}) + endif() + endfunction() + + disable_msvc_warnings(ggml-base) + disable_msvc_warnings(ggml) + disable_msvc_warnings(ggml-cpu) + disable_msvc_warnings(ggml-cpu-x64) + disable_msvc_warnings(ggml-cpu-sse42) + disable_msvc_warnings(ggml-cpu-sandybridge) + disable_msvc_warnings(ggml-cpu-haswell) + disable_msvc_warnings(ggml-cpu-skylakex) + disable_msvc_warnings(ggml-cpu-icelake) + disable_msvc_warnings(ggml-cpu-alderlake) +endif() diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h index 64671495b3..778927f682 100644 --- a/ggml/include/ggml-backend.h +++ b/ggml/include/ggml-backend.h @@ -38,7 +38,7 @@ extern "C" { GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer (ggml_backend_buffer_type_t buft, size_t size); GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft); GGML_API size_t ggml_backend_buft_get_max_size (ggml_backend_buffer_type_t buft); - GGML_API size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); + GGML_API size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); GGML_API bool ggml_backend_buft_is_host (ggml_backend_buffer_type_t buft); GGML_API ggml_backend_dev_t ggml_backend_buft_get_device (ggml_backend_buffer_type_t buft); @@ -59,7 +59,7 @@ extern "C" { GGML_API enum ggml_status ggml_backend_buffer_init_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer); GGML_API size_t ggml_backend_buffer_get_max_size (ggml_backend_buffer_t buffer); - GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); + GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor); GGML_API void ggml_backend_buffer_clear (ggml_backend_buffer_t buffer, uint8_t value); GGML_API bool ggml_backend_buffer_is_host (ggml_backend_buffer_t buffer); GGML_API void ggml_backend_buffer_set_usage (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage); @@ -248,7 +248,7 @@ extern "C" { // preferrably to run on the same backend as the buffer ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS); - sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false); + sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false, true); // initialize buffers from a max size graph (optional) reserve_graph = build_graph(sched, max_batch_size); @@ -289,7 +289,7 @@ extern "C" { typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data); // Initialize a backend scheduler, backends with low index are given priority over backends with high index - GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel); + GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel, bool op_offload); GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched); // Initialize backend buffers from a measure graph diff --git a/ggml/include/ggml-cpp.h b/ggml/include/ggml-cpp.h index a12342c25d..48aa79682b 100644 --- a/ggml/include/ggml-cpp.h +++ b/ggml/include/ggml-cpp.h @@ -24,7 +24,7 @@ typedef std::unique_ptr gguf_context_ptr; struct ggml_gallocr_deleter { void operator()(ggml_gallocr_t galloc) { ggml_gallocr_free(galloc); } }; -typedef std::unique_ptr ggml_gallocr_ptr; +typedef std::unique_ptr ggml_gallocr_ptr; // ggml-backend diff --git a/ggml/include/ggml-cpu.h b/ggml/include/ggml-cpu.h index f5e11f1e10..de77a875ec 100644 --- a/ggml/include/ggml-cpu.h +++ b/ggml/include/ggml-cpu.h @@ -133,6 +133,11 @@ extern "C" { GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void); + GGML_BACKEND_API void ggml_cpu_fp32_to_fp16(const float *, ggml_fp16_t *, int64_t); + GGML_BACKEND_API void ggml_cpu_fp16_to_fp32(const ggml_fp16_t *, float *, int64_t); + GGML_BACKEND_API void ggml_cpu_fp32_to_bf16(const float *, ggml_bf16_t *, int64_t); + GGML_BACKEND_API void ggml_cpu_bf16_to_fp32(const ggml_bf16_t *, float *, int64_t); + #ifdef __cplusplus } #endif diff --git a/ggml/include/ggml-opt.h b/ggml/include/ggml-opt.h index eb5eab9de6..da0c24b46f 100644 --- a/ggml/include/ggml-opt.h +++ b/ggml/include/ggml-opt.h @@ -37,13 +37,16 @@ extern "C" { // ====== Dataset ====== GGML_API ggml_opt_dataset_t ggml_opt_dataset_init( - int64_t ne_datapoint, // number of elements per datapoint - int64_t ne_label, // number of elements per label - int64_t ndata, // total number of datapoints/labels - int64_t ndata_shard); // number of datapoints/labels per shard (unit at which the dataset is shuffled/copied) + enum ggml_type type_data, // the type for the internal data tensor + enum ggml_type type_label, // the type for the internal labels tensor + int64_t ne_datapoint, // number of elements per datapoint + int64_t ne_label, // number of elements per label + int64_t ndata, // total number of datapoints/labels + int64_t ndata_shard); // number of datapoints/labels per shard (unit at which the dataset is shuffled/copied) GGML_API void ggml_opt_dataset_free(ggml_opt_dataset_t dataset); // get underlying tensors that store the data + GGML_API int64_t ggml_opt_dataset_ndata (ggml_opt_dataset_t dataset); GGML_API struct ggml_tensor * ggml_opt_dataset_data (ggml_opt_dataset_t dataset); // shape = [ne_datapoint, ndata] GGML_API struct ggml_tensor * ggml_opt_dataset_labels(ggml_opt_dataset_t dataset); // shape = [nd_label, ndata] @@ -56,13 +59,19 @@ extern "C" { struct ggml_tensor * data_batch, // shape = [ne_datapoint, ndata_batch] struct ggml_tensor * labels_batch, // shape = [ne_label, ndata_batch] int64_t ibatch); + GGML_API void ggml_opt_dataset_get_batch_host( + ggml_opt_dataset_t dataset, + void * data_batch, + size_t nb_data_batch, + void * labels_batch, + int64_t ibatch); // ====== Model / Context ====== enum ggml_opt_build_type { - GGML_OPT_BUILD_TYPE_FORWARD, - GGML_OPT_BUILD_TYPE_GRAD, - GGML_OPT_BUILD_TYPE_OPT, + GGML_OPT_BUILD_TYPE_FORWARD = 10, + GGML_OPT_BUILD_TYPE_GRAD = 20, + GGML_OPT_BUILD_TYPE_OPT = 30, }; // parameters that control which optimizer is used and how said optimizer tries to find the minimal loss @@ -81,20 +90,22 @@ extern "C" { // userdata can be used to pass arbitrary data typedef struct ggml_opt_optimizer_params (*ggml_opt_get_optimizer_params)(void * userdata); - // returns the default optimizer params (constant) + // returns the default optimizer params (constant, hard-coded values) // userdata is not used GGML_API struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata); + // casts userdata to ggml_opt_optimizer_params and returns it + GGML_API struct ggml_opt_optimizer_params ggml_opt_get_constant_optimizer_params(void * userdata); + // parameters for initializing a new optimization context struct ggml_opt_params { ggml_backend_sched_t backend_sched; // defines which backends are used to construct the compute graphs - struct ggml_context * ctx_compute; // created in user code, holds non-static tensors - - // the forward graph is defined by inputs and outputs - // those tensors and all tensors inbetween are not intended to be reusable between multiple optimization contexts - struct ggml_tensor * inputs; - struct ggml_tensor * outputs; + // by default the forward graph needs to be reconstructed for each eval + // if ctx_compute, inputs, and outputs are set the graphs are instead allocated statically + struct ggml_context * ctx_compute; + struct ggml_tensor * inputs; + struct ggml_tensor * outputs; enum ggml_opt_loss_type loss_type; enum ggml_opt_build_type build_type; @@ -107,12 +118,9 @@ extern "C" { // get parameters for an optimization context with defaults set where possible // parameters for which no sensible defaults exist are supplied as arguments to this function - GGML_API ggml_opt_params ggml_opt_default_params( - ggml_backend_sched_t backend_sched, - struct ggml_context * ctx_compute, - struct ggml_tensor * inputs, - struct ggml_tensor * outputs, - enum ggml_opt_loss_type loss_type); + GGML_API struct ggml_opt_params ggml_opt_default_params( + ggml_backend_sched_t backend_sched, + enum ggml_opt_loss_type loss_type); GGML_API ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params); GGML_API void ggml_opt_free(ggml_opt_context_t opt_ctx); @@ -121,6 +129,7 @@ extern "C" { GGML_API void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer); // get underlying tensors that store data + // if not using static graphs these pointers become invalid with the next call to ggml_opt_alloc GGML_API struct ggml_tensor * ggml_opt_inputs( ggml_opt_context_t opt_ctx); // forward graph input tensor GGML_API struct ggml_tensor * ggml_opt_outputs( ggml_opt_context_t opt_ctx); // forward graph output tensor GGML_API struct ggml_tensor * ggml_opt_labels( ggml_opt_context_t opt_ctx); // labels to compare outputs against @@ -128,11 +137,12 @@ extern "C" { GGML_API struct ggml_tensor * ggml_opt_pred( ggml_opt_context_t opt_ctx); // predictions made by outputs GGML_API struct ggml_tensor * ggml_opt_ncorrect(ggml_opt_context_t opt_ctx); // number of matching predictions between outputs and labels + // get the gradient accumulator for a node from the forward graph GGML_API struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node); // ====== Optimization Result ====== - GGML_API ggml_opt_result_t ggml_opt_result_init(); + GGML_API ggml_opt_result_t ggml_opt_result_init(void); GGML_API void ggml_opt_result_free(ggml_opt_result_t result); GGML_API void ggml_opt_result_reset(ggml_opt_result_t result); @@ -144,11 +154,20 @@ extern "C" { // ====== Computation ====== - // do forward pass, increment result if not NULL - GGML_API void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result); + // if not using static graphs, this function must be called prior to ggml_opt_alloc + GGML_API void ggml_opt_prepare_alloc( + ggml_opt_context_t opt_ctx, + struct ggml_context * ctx_compute, + struct ggml_cgraph * gf, + struct ggml_tensor * inputs, + struct ggml_tensor * outputs); - // do forward pass, increment result if not NULL, do backward pass - GGML_API void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result); + // allocate the next graph for evaluation, either forward or forward + backward + // must be called exactly once prior to calling ggml_opt_eval + GGML_API void ggml_opt_alloc(ggml_opt_context_t opt_ctx, bool backward); + + // do forward pass, increment result if not NULL, do backward pass if allocated + GGML_API void ggml_opt_eval(ggml_opt_context_t opt_ctx, ggml_opt_result_t result); // ############################################################################ // ## The high-level functions start here. They do not depend on any private ## @@ -200,9 +219,9 @@ extern "C" { // fit model defined by inputs and outputs to dataset GGML_API void ggml_opt_fit( ggml_backend_sched_t backend_sched, // backend scheduler for constructing the compute graphs - ggml_context * ctx_compute, // context with temporarily allocated tensors to calculate the outputs - ggml_tensor * inputs, // input tensor with shape [ne_datapoint, ndata_batch] - ggml_tensor * outputs, // output tensor, must have shape [ne_label, ndata_batch] if labels are used + struct ggml_context * ctx_compute, // context with temporarily allocated tensors to calculate the outputs + struct ggml_tensor * inputs, // input tensor with shape [ne_datapoint, ndata_batch] + struct ggml_tensor * outputs, // output tensor, must have shape [ne_label, ndata_batch] if labels are used ggml_opt_dataset_t dataset, // dataset with data and optionally also labels enum ggml_opt_loss_type loss_type, // loss to minimize ggml_opt_get_optimizer_params get_opt_pars, // callback to get optimizer params, userdata is pointer to epoch (of type int64_t) diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index 51aa5b3a0a..e91dedf14a 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -393,8 +393,8 @@ extern "C" { // precision enum ggml_prec { - GGML_PREC_DEFAULT, - GGML_PREC_F32, + GGML_PREC_DEFAULT = 0, // stored as ggml_tensor.op_params, 0 by default + GGML_PREC_F32 = 10, }; // model file types @@ -673,11 +673,15 @@ extern "C" { GGML_API bool ggml_is_3d (const struct ggml_tensor * tensor); GGML_API int ggml_n_dims (const struct ggml_tensor * tensor); // returns 1 for scalars + // returns whether the tensor elements can be iterated over with a flattened index (no gaps, no permutation) GGML_API bool ggml_is_contiguous (const struct ggml_tensor * tensor); GGML_API bool ggml_is_contiguous_0(const struct ggml_tensor * tensor); // same as ggml_is_contiguous() GGML_API bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1 GGML_API bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2 + // returns whether the tensor elements are allocated as one contiguous block of memory (no gaps, but permutation ok) + GGML_API bool ggml_is_contiguously_allocated(const struct ggml_tensor * tensor); + // true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN GGML_API bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor); @@ -764,7 +768,7 @@ extern "C" { // Tensor flags GGML_API void ggml_set_input(struct ggml_tensor * tensor); GGML_API void ggml_set_output(struct ggml_tensor * tensor); - GGML_API void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor); + GGML_API void ggml_set_param(struct ggml_tensor * tensor); GGML_API void ggml_set_loss(struct ggml_tensor * tensor); // @@ -934,7 +938,7 @@ extern "C" { GGML_API struct ggml_tensor * ggml_repeat_back( struct ggml_context * ctx, struct ggml_tensor * a, - struct ggml_tensor * b); + struct ggml_tensor * b); // sum up values that are adjacent in dims > 0 instead of repeated with same stride // concat a and b along dim // used in stable-diffusion @@ -2045,15 +2049,14 @@ extern "C" { 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_static, // context for static gradients (loss + gradient accumulation) - struct ggml_context * ctx_compute, // context for gradient computation - struct ggml_cgraph * cgraph, - bool accumulate); // whether or not gradients should be accumulated, requires static allocation of tensors in ctx_static + struct ggml_context * ctx, // context for gradient computation + struct ggml_cgraph * cgraph, + struct ggml_tensor ** grad_accs); // graph allocation in a context GGML_API struct ggml_cgraph * ggml_new_graph (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false GGML_API struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t size, bool grads); - GGML_API struct ggml_cgraph * ggml_graph_dup (struct ggml_context * ctx, struct ggml_cgraph * cgraph); + GGML_API struct ggml_cgraph * ggml_graph_dup (struct ggml_context * ctx, struct ggml_cgraph * cgraph, bool force_grads); GGML_API void ggml_graph_cpy (struct ggml_cgraph * src, struct ggml_cgraph * dst); GGML_API void ggml_graph_reset (struct ggml_cgraph * cgraph); // set regular grads + optimizer momenta to 0, set loss grad to 1 GGML_API void ggml_graph_clear (struct ggml_cgraph * cgraph); diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index 43d9fc4fe2..ddea5ad389 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -214,7 +214,7 @@ add_library(ggml target_link_libraries(ggml PUBLIC ggml-base) if (CMAKE_SYSTEM_NAME MATCHES "Linux") - target_link_libraries(ggml PRIVATE dl stdc++fs) + target_link_libraries(ggml PRIVATE dl) endif() function(ggml_add_backend_library backend) diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c index a3d3f69013..5fd379f6a9 100644 --- a/ggml/src/ggml-alloc.c +++ b/ggml/src/ggml-alloc.c @@ -816,7 +816,10 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) { size_t node_size = 0; if (!node->data && !node->view_src) { - GGML_ASSERT(talloc->buffer_id >= 0); // prevent segfault when misusing the API + // If we previously had data but don't now then reallocate + if (talloc->buffer_id < 0) { + return false; + } node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node); } return talloc->size_max >= node_size; diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp index 273075f4e5..b30b4cb386 100644 --- a/ggml/src/ggml-backend.cpp +++ b/ggml/src/ggml-backend.cpp @@ -56,7 +56,7 @@ size_t ggml_backend_buft_get_max_size(ggml_backend_buffer_type_t buft) { return SIZE_MAX; } -size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) { +size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor) { // get_alloc_size is optional, defaults to ggml_nbytes if (buft->iface.get_alloc_size) { size_t size = buft->iface.get_alloc_size(buft, tensor); @@ -152,7 +152,7 @@ size_t ggml_backend_buffer_get_max_size(ggml_backend_buffer_t buffer) { return ggml_backend_buft_get_max_size(ggml_backend_buffer_get_type(buffer)); } -size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { +size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor) { return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_get_type(buffer), tensor); } @@ -674,6 +674,8 @@ struct ggml_backend_sched { char * context_buffer; size_t context_buffer_size; + bool op_offload; + int debug; }; @@ -766,7 +768,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st if (tensor->op != GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) { int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor); // check if a backend with higher prio wants to offload the op - if (src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) { + if (sched->op_offload && src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) { for (int b = 0; b < src_backend_id; b++) { if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) { SET_CAUSE(tensor, "1.off"); @@ -1109,7 +1111,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg const int node_backend_id = tensor_backend_id(node); - assert(node_backend_id != -1); // all nodes should be assigned by now + assert(node_backend_id != -1); // all nodes should be assigned by now, this can happen if there is no CPU fallback // check if we should start a new split based on the sources of the current node bool need_new_split = false; @@ -1452,7 +1454,8 @@ ggml_backend_sched_t ggml_backend_sched_new( ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, - bool parallel) { + bool parallel, + bool op_offload) { GGML_ASSERT(n_backends > 0); GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS); GGML_ASSERT(ggml_backend_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU); @@ -1497,6 +1500,7 @@ ggml_backend_sched_t ggml_backend_sched_new( } sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends); + sched->op_offload = op_offload; ggml_backend_sched_reset(sched); diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt index 6a652738c1..bdaec2881d 100644 --- a/ggml/src/ggml-cpu/CMakeLists.txt +++ b/ggml/src/ggml-cpu/CMakeLists.txt @@ -352,10 +352,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name) # TODO: Separation to determine activation of VX/VXE/VXE2 if (${S390X_M} MATCHES "8561|8562") message(STATUS "z15 target") - list(APPEND ARCH_FLAGS -march=z15 -mtune=z15) + list(APPEND ARCH_FLAGS -march=z15) elseif (${S390X_M} MATCHES "3931") message(STATUS "z16 target") - list(APPEND ARCH_FLAGS -march=z16 -mtune=z16) + list(APPEND ARCH_FLAGS -march=z16) + elseif (${S390X_M} MATCHES "9175|9176") + # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version. + message(STATUS "z17 target") + list(APPEND ARCH_FLAGS -march=z17) else() message(STATUS "Unknown target") message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.") @@ -424,6 +428,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name) ${KLEIDIAI_SRC}/kai/ukernels/ ${KLEIDIAI_SRC}/kai/ukernels/matmul/ ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/ + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/ ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/) set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}") @@ -434,17 +439,19 @@ function(ggml_add_cpu_backend_variant_impl tag_name) string(FIND "${ARCH_FLAGS_TEMP}" "+i8mm" I8MM_ENABLED) string(FIND "${ARCH_FLAGS_TEMP}" "+sme" SME_ENABLED) - set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS}) + set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS_TEMP}) - list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c) - list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c) - list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c) - list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c) + list(APPEND GGML_KLEIDIAI_SOURCES + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c) if (NOT DOTPROD_ENABLED MATCHES -1) - list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c) - list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c) - list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c) + list(APPEND GGML_KLEIDIAI_SOURCES + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c) endif() if (NOT I8MM_ENABLED MATCHES -1) @@ -452,9 +459,13 @@ function(ggml_add_cpu_backend_variant_impl tag_name) endif() if (NOT SME_ENABLED MATCHES -1) - list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c) - list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c) - set(PRIVATE_ARCH_FLAGS "${PRIVATE_ARCH_FLAGS}+sve+sve2") + list(APPEND GGML_KLEIDIAI_SOURCES + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c) + set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2") endif() set_source_files_properties(${GGML_KLEIDIAI_SOURCES} PROPERTIES COMPILE_OPTIONS "${PRIVATE_ARCH_FLAGS}") diff --git a/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp b/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp index 175cba329b..8ff6d64a4d 100644 --- a/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp +++ b/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp @@ -72,8 +72,6 @@ static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wro #if defined(__GNUC__) #pragma GCC diagnostic ignored "-Woverlength-strings" -#elif defined(_MSC_VER) -#pragma warning(disable: 4244 4267) // possible loss of data #endif #define UNUSED GGML_UNUSED diff --git a/ggml/src/ggml-cpu/ggml-cpu-quants.c b/ggml/src/ggml-cpu/ggml-cpu-quants.c index 91a81bdc3c..ccd0651ebc 100644 --- a/ggml/src/ggml-cpu/ggml-cpu-quants.c +++ b/ggml/src/ggml-cpu/ggml-cpu-quants.c @@ -20,12 +20,6 @@ #define GROUP_MAX_EPS_IQ1_M 1e-7f #define GROUP_MAX_EPS_IQ1_S 1e-12f -#if defined(_MSC_VER) -// disable "possible loss of data" to avoid warnings for hundreds of casts -// we should just be careful :) -#pragma warning(disable: 4244 4267) -#endif - #define UNUSED GGML_UNUSED // some compilers don't provide _mm256_set_m128i, e.g. gcc 7 @@ -6596,7 +6590,118 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi } *s = hsum_float_8(acc); +#elif defined(__VXE__) || defined(__VXE2__) + uint32_t aux[3]; + uint32_t utmp[4]; + const int32x4_t v_z = vec_splat_s32(0); + const uint8x16_t v_3m = vec_splat_u8(0x03); + + const uint8x16_t v_0c = vec_splat_u8(1); + const uint8x16_t v_1c = vec_sl(v_0c, 1); + const uint8x16_t v_2c = vec_sl(v_0c, 2); + const uint8x16_t v_3c = vec_sl(v_0c, 3); + + uint8x16_t q3h[4]; + uint8x16_t q3b[2]; + int8x16_t q3bytes[4]; + int8x16_t q8bytes[4]; + uint8x16_t qhbits[2]; + + float sum = 0; + + for (int i = 0; i < nb; ++i) { + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + + const uint8_t * restrict x0l = x[i].qs; + const uint8_t * restrict x0h = x[i].hmask; + const int8_t * restrict y0 = y[i].qs; + + qhbits[0] = vec_xl(0 , x0h); + qhbits[1] = vec_xl(16, x0h); + + int32_t isum = 0; + + memcpy(aux, x[i].scales, 12); + utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4); + utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4); + utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4); + utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4); + + int8_t * scale = (int8_t *)utmp; + for (int j = 0; j < 16; ++j) scale[j] -= 32; + + for (int j = 0; j < QK_K/128; ++j) { + int32x4_t isum0, isum1, isum2, isum3; + + q3b[0] = vec_xl(0 , x0l); + q3b[1] = vec_xl(16, x0l); + x0l += 32; + + q8bytes[0] = vec_xl(0 , y0); + q8bytes[1] = vec_xl(16 , y0); + q8bytes[2] = vec_xl(32 , y0); + q8bytes[3] = vec_xl(48 , y0); + q8bytes[4] = vec_xl(64 , y0); + q8bytes[5] = vec_xl(80 , y0); + q8bytes[6] = vec_xl(96 , y0); + q8bytes[7] = vec_xl(112, y0); + y0 += 128; + + q3h[0] = vec_sl(vec_andc(v_0c, qhbits[0]), 2); + q3h[1] = vec_sl(vec_andc(v_0c, qhbits[1]), 2); + q3h[2] = vec_sl(vec_andc(v_1c, qhbits[0]), 1); + q3h[3] = vec_sl(vec_andc(v_1c, qhbits[1]), 1); + + q3bytes[0] = vec_sub((int8x16_t)vec_and(q3b[0], v_3m), (int8x16_t)q3h[0]); + q3bytes[1] = vec_sub((int8x16_t)vec_and(q3b[1], v_3m), (int8x16_t)q3h[1]); + q3bytes[2] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 2), v_3m), (int8x16_t)q3h[2]); + q3bytes[3] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 2), v_3m), (int8x16_t)q3h[3]); + + isum0 = ggml_vec_dot(v_z, q3bytes[0], q8bytes[0]); + isum1 = ggml_vec_dot(v_z, q3bytes[1], q8bytes[1]); + isum2 = ggml_vec_dot(v_z, q3bytes[2], q8bytes[2]); + isum3 = ggml_vec_dot(v_z, q3bytes[3], q8bytes[3]); + + isum += (isum0[0] + isum0[1] + isum0[2] + isum0[3]) * scale[0]; + isum += (isum1[0] + isum1[1] + isum1[2] + isum1[3]) * scale[1]; + isum += (isum2[0] + isum2[1] + isum2[2] + isum2[3]) * scale[2]; + isum += (isum3[0] + isum3[1] + isum3[2] + isum3[3]) * scale[3]; + + scale += 4; + + q3h[0] = vec_andc(v_2c, qhbits[0]); + q3h[1] = vec_andc(v_2c, qhbits[1]); + q3h[2] = vec_sr(vec_andc(v_3c, qhbits[0]), 1); + q3h[3] = vec_sr(vec_andc(v_3c, qhbits[1]), 1); + + q3bytes[0] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 4), v_3m), (int8x16_t)q3h[0]); + q3bytes[1] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 4), v_3m), (int8x16_t)q3h[1]); + q3bytes[2] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 6), v_3m), (int8x16_t)q3h[2]); + q3bytes[3] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 6), v_3m), (int8x16_t)q3h[3]); + + isum0 = ggml_vec_dot(v_z, q3bytes[0], q8bytes[4]); + isum1 = ggml_vec_dot(v_z, q3bytes[1], q8bytes[5]); + isum2 = ggml_vec_dot(v_z, q3bytes[2], q8bytes[6]); + isum3 = ggml_vec_dot(v_z, q3bytes[3], q8bytes[7]); + + isum += (isum0[0] + isum0[1] + isum0[2] + isum0[3]) * scale[0]; + isum += (isum1[0] + isum1[1] + isum1[2] + isum1[3]) * scale[1]; + isum += (isum2[0] + isum2[1] + isum2[2] + isum2[3]) * scale[2]; + isum += (isum3[0] + isum3[1] + isum3[2] + isum3[3]) * scale[3]; + + scale += 4; + + if (j == 0) { + qhbits[0] = vec_sr(qhbits[0], 4); + qhbits[1] = vec_sr(qhbits[1], 4); + } + } + + sum += d * isum; + } + + *s = sum; #else // scalar version // This function is written like this so the compiler can manage to vectorize most of it diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c index dbad8f61a1..a30e67f227 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.c +++ b/ggml/src/ggml-cpu/ggml-cpu.c @@ -50,19 +50,6 @@ #include "llamafile/sgemm.h" #endif -#if defined(_MSC_VER) -// disable "possible loss of data" to avoid hundreds of casts -// we should just be careful :) -#pragma warning(disable: 4244 4267) - -// disable POSIX deprecation warnings -// these functions are never going away, anyway -#pragma warning(disable: 4996) - -// unreachable code because of multiple instances of code after GGML_ABORT -#pragma warning(disable: 4702) -#endif - // Note: once we move threading into a separate C++ file // will use std::hardware_destructive_interference_size instead of hardcoding it here // and we'll use C++ attribute syntax. @@ -215,7 +202,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = { .nrows = 1, }, [GGML_TYPE_F16] = { - .from_float = (ggml_from_float_t) ggml_fp32_to_fp16_row, + .from_float = (ggml_from_float_t) ggml_cpu_fp32_to_fp16, .vec_dot = (ggml_vec_dot_t) ggml_vec_dot_f16, .vec_dot_type = GGML_TYPE_F16, .nrows = 1, @@ -356,7 +343,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = { .from_float = quantize_row_q8_K, }, [GGML_TYPE_BF16] = { - .from_float = (ggml_from_float_t) ggml_fp32_to_bf16_row, + .from_float = (ggml_from_float_t) ggml_cpu_fp32_to_bf16, .vec_dot = (ggml_vec_dot_t) ggml_vec_dot_bf16, .vec_dot_type = GGML_TYPE_BF16, .nrows = 1, @@ -3166,6 +3153,93 @@ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct g return ggml_graph_compute(cgraph, &cplan); } +void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) { + int64_t i = 0; +#if defined(__F16C__) +#if defined(__AVX512F__) + for (; i + 15 < n; i += 16) { + __m512 x_vec = _mm512_loadu_ps(x + i); + __m256i y_vec = _mm512_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT); + _mm256_storeu_si256((__m256i *)(y + i), y_vec); + } +#endif + for (; i + 7 < n; i += 8) { + __m256 x_vec = _mm256_loadu_ps(x + i); + __m128i y_vec = _mm256_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT); + _mm_storeu_si128((__m128i *)(y + i), y_vec); + } + for (; i + 3 < n; i += 4) { + __m128 x_vec = _mm_loadu_ps(x + i); + __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT); + _mm_storel_epi64((__m128i *)(y + i), y_vec); + } +#endif + for (; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(x[i]); + } +} + +void ggml_cpu_fp16_to_fp32(const ggml_fp16_t * x, float * y, int64_t n) { + int64_t i = 0; +#if defined(__F16C__) +#if defined(__AVX512F__) + for (; i + 15 < n; i += 16) { + __m256i x_vec = _mm256_loadu_si256((const __m256i *)(x + i)); + __m512 y_vec = _mm512_cvtph_ps(x_vec); + _mm512_storeu_ps(y + i, y_vec); + } +#endif + for (; i + 7 < n; i += 8) { + __m128i x_vec = _mm_loadu_si128((const __m128i *)(x + i)); + __m256 y_vec = _mm256_cvtph_ps(x_vec); + _mm256_storeu_ps(y + i, y_vec); + } + for (; i + 3 < n; i += 4) { + __m128i x_vec = _mm_loadl_epi64((const __m128i *)(x + i)); + __m128 y_vec = _mm_cvtph_ps(x_vec); + _mm_storeu_ps(y + i, y_vec); + } +#endif + for (; i < n; ++i) { + y[i] = GGML_FP16_TO_FP32(x[i]); + } +} + +void ggml_cpu_fp32_to_bf16(const float * x, ggml_bf16_t * y, int64_t n) { + int64_t i = 0; + for (; i < n; ++i) { + y[i] = GGML_FP32_TO_BF16(x[i]); + } +} + +void ggml_cpu_bf16_to_fp32(const ggml_bf16_t * x, float * y, int64_t n) { + int64_t i = 0; +#if defined(__AVX2__) +#if defined(__AVX512F__) + for (; i + 15 < n; i += 16) { + _mm512_storeu_ps(y + i, + _mm512_castsi512_ps( + _mm512_slli_epi32( + _mm512_cvtepu16_epi32( + _mm256_loadu_si256( + (const __m256i *)(x + i))), + 16))); + } +#endif + for (; i + 7 < n; i += 8) { + _mm256_storeu_ps(y + i, + _mm256_castsi256_ps( + _mm256_slli_epi32( + _mm256_cvtepu16_epi32( + _mm_loadu_si128( + (const __m128i *)(x + i))), + 16))); + } +#endif + for (; i < n; i++) { + y[i] = GGML_BF16_TO_FP32(x[i]); + } +} int ggml_cpu_has_avx(void) { #if defined(__AVX__) diff --git a/ggml/src/ggml-cpu/ggml-cpu.cpp b/ggml/src/ggml-cpu/ggml-cpu.cpp index 4b688a67eb..e013e8b416 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.cpp +++ b/ggml/src/ggml-cpu/ggml-cpu.cpp @@ -11,24 +11,26 @@ #include #ifdef GGML_USE_CPU_HBM -#include "ggml-cpu-hbm.h" +# include "ggml-cpu-hbm.h" #endif #ifdef GGML_USE_CPU_KLEIDIAI -#include "kleidiai/kleidiai.h" -#endif - -#if defined(__APPLE__) -#include -#include +# include "kleidiai/kleidiai.h" #endif #if defined(_WIN32) -#define WIN32_LEAN_AND_MEAN -#ifndef NOMINMAX - #define NOMINMAX +# define WIN32_LEAN_AND_MEAN +# ifndef NOMINMAX +# define NOMINMAX +# endif +# include +#else +# include #endif -#include + +#if defined(__APPLE__) +# include +# include #endif // ggml-backend interface @@ -70,8 +72,10 @@ static ggml_backend_buffer_type_t * ggml_backend_cpu_device_get_extra_buffers_ty } static bool ggml_backend_cpu_is_extra_buffer_type(ggml_backend_buffer_type_t buft) { - for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) { - if (extra && extra == buft) return true; + for (auto * extra : ggml_backend_cpu_get_extra_buffers_type()) { + if (extra && extra == buft) { + return true; + } } return false; } @@ -330,9 +334,18 @@ static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t d } static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) { - // TODO - *free = 0; - *total = 0; +#ifdef _WIN32 + MEMORYSTATUSEX status; + status.dwLength = sizeof(status); + GlobalMemoryStatusEx(&status); + *total = status.ullTotalPhys; + *free = status.ullAvailPhys; +#else + long pages = sysconf(_SC_PHYS_PAGES); + long page_size = sysconf(_SC_PAGE_SIZE); + *total = pages * page_size; + *free = *total; +#endif GGML_UNUSED(dev); } diff --git a/ggml/src/ggml-cpu/kleidiai/kernels.cpp b/ggml/src/ggml-cpu/kleidiai/kernels.cpp index aacc2bb5ee..910fd0ee4e 100644 --- a/ggml/src/ggml-cpu/kleidiai/kernels.cpp +++ b/ggml/src/ggml-cpu/kleidiai/kernels.cpp @@ -4,16 +4,22 @@ // KleidiAI micro-kernels #include "kai_matmul_clamp_f32_qsi8d32p_qsi4c32p_interface.h" -#include "kai_lhs_quant_pack_qsi8d32p_f32.h" -#include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h" -#include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h" -#include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h" #include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.h" #include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.h" #include "kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.h" #include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.h" #include "kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.h" #include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.h" +#include "kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.h" + +#include "kai_lhs_pack_bf16p2vlx2_f32_sme.h" +#include "kai_lhs_quant_pack_qsi8d32p_f32.h" +#include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h" + +#include "kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.h" +#include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h" +#include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h" + #include "kai_common.h" #include "kernels.h" @@ -61,6 +67,53 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = { /* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon, }, /* .required_cpu = */ CPU_FEATURE_SME, + /* .lhs_type = */ GGML_TYPE_F32, + /* .rhs_type = */ GGML_TYPE_Q4_0, + /* .op_type = */ GGML_TYPE_F32, + }, + { + /* SME GEMM */ + /* .kern_info = */ { + /* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_mr = */ kai_get_mr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_nr = */ kai_get_nr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_kr = */ kai_get_kr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_sr = */ kai_get_sr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .run_kernel = */ kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + }, + /* SME GEMV */ + /* .kern_info = */ { + /* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_mr = */ kai_get_mr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_nr = */ kai_get_nr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_kr = */ kai_get_kr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_sr = */ kai_get_sr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + /* .run_kernel = */ kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa, + }, + /* .lhs_info = */ { + /* .get_offset = */ kai_get_lhs_offset_lhs_pack_bf16p2vlx2_f32_sme, + /* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_pack_bf16p2vlx2_f32_sme, + /* .packed_size = */ kai_get_lhs_packed_size_lhs_pack_bf16p2vlx2_f32_sme, + /* .pack_func = */ kai_run_lhs_pack_bf16p2vlx2_f32_sme, + }, + /* .rhs_info = */ { + /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme, + /* .pack_func = */ kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme, + }, + /* .required_cpu = */ CPU_FEATURE_SME, + /* .lhs_type = */ GGML_TYPE_F32, + /* .rhs_type = */ GGML_TYPE_F16, + /* .op_type = */ GGML_TYPE_F32, }, #endif #if defined(__APPLE__) @@ -105,6 +158,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = { /* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0, }, /* .required_cpu = */ CPU_FEATURE_DOTPROD, + /* .lhs_type = */ GGML_TYPE_F32, + /* .rhs_type = */ GGML_TYPE_Q4_0, + /* .op_type = */ GGML_TYPE_F32, }, #endif #if defined(__ARM_FEATURE_MATMUL_INT8) @@ -148,6 +204,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = { /* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0, }, /* .required_cpu = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM, + /* .lhs_type = */ GGML_TYPE_F32, + /* .rhs_type = */ GGML_TYPE_Q4_0, + /* .op_type = */ GGML_TYPE_F32, }, #endif #else @@ -192,6 +251,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = { /* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0, }, /* .required_cpu = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM, + /* .lhs_type = */ GGML_TYPE_F32, + /* .rhs_type = */ GGML_TYPE_Q4_0, + /* .op_type = */ GGML_TYPE_F32, }, #endif #if defined(__ARM_FEATURE_DOTPROD) @@ -235,12 +297,33 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = { /* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0, }, /* .required_cpu = */ CPU_FEATURE_DOTPROD, + /* .lhs_type = */ GGML_TYPE_F32, + /* .rhs_type = */ GGML_TYPE_Q4_0, + /* .op_type = */ GGML_TYPE_F32, }, #endif #endif }; -ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature features) { +ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor) { + ggml_kleidiai_kernels * kernel = nullptr; + + if (tensor->op == GGML_OP_MUL_MAT && tensor->src[0] != nullptr && tensor->src[1] != nullptr) { + for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) { + if ((cpu_features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu && + gemm_gemv_kernels[i].lhs_type == tensor->src[1]->type && + gemm_gemv_kernels[i].rhs_type == tensor->src[0]->type && + gemm_gemv_kernels[i].op_type == tensor->type) { + kernel = &gemm_gemv_kernels[i]; + break; + } + } + } + + return kernel; +} + +ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features) { ggml_kleidiai_kernels * kernels = nullptr; for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) { diff --git a/ggml/src/ggml-cpu/kleidiai/kernels.h b/ggml/src/ggml-cpu/kleidiai/kernels.h index 2ffe97eb42..5ac02bda7c 100644 --- a/ggml/src/ggml-cpu/kleidiai/kernels.h +++ b/ggml/src/ggml-cpu/kleidiai/kernels.h @@ -4,6 +4,9 @@ #pragma once +#include +#include "ggml.h" + enum cpu_feature { CPU_FEATURE_NONE = 0, CPU_FEATURE_DOTPROD = 1, @@ -26,26 +29,53 @@ struct kernel_info { size_t (*get_nr)(void); size_t (*get_kr)(void); size_t (*get_sr)(void); - size_t (*get_lhs_offset)(size_t m_idx, size_t k, size_t bl); - size_t (*get_rhs_packed_offset)(size_t n_idx, size_t k, size_t bl); + std::variant< + std::function, + std::function + > get_lhs_offset; + std::variant< + std::function, + std::function + > get_rhs_packed_offset; size_t (*get_dst_offset)(size_t m_idx, size_t n_idx, size_t stride); size_t (*get_dst_size)(size_t m, size_t n); - void (*run_kernel)(size_t m, size_t n, size_t k, size_t bl, const void* lhs_packed, const void* rhs_packed, - float* dst, size_t dst_stride_row, size_t dst_stride_col, float scalar_min, float scalar_max); + std::variant< + std::function, + std::function + > run_kernel; }; struct lhs_packing_info { size_t (*get_offset)(size_t m_idx, size_t lhs_stride); - size_t (*get_packed_offset)(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr); - size_t (*packed_size)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr); - void (*pack_func)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const float* lhs, - size_t lhs_stride, void* lhs_packed); + std::variant< + std::function, + std::function + > get_packed_offset; + std::variant< + std::function, + std::function + > packed_size; + std::variant< + std::function, + std::function + > pack_func; }; struct rhs_packing_info { - size_t (*packed_size)(size_t n, size_t k, size_t nr, size_t kr, size_t bl); - void (*pack_func)(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl, const uint8_t* rhs, - const float* bias, void* rhs_packed, size_t extra_bytes, const struct kai_rhs_pack_qs4cxs1s0_param* params); + std::variant< + std::function, + std::function + > packed_size; + std::variant< + std::function, + std::function + > pack_func; }; struct ggml_kleidiai_kernels { @@ -55,6 +85,10 @@ struct ggml_kleidiai_kernels { rhs_packing_info rhs_info; cpu_feature required_cpu; + ggml_type lhs_type; + ggml_type rhs_type; + ggml_type op_type; }; -ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features); +ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor); +ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features); diff --git a/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp b/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp index 4e89ca0faa..f3dffdd6bf 100644 --- a/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp +++ b/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp @@ -34,8 +34,9 @@ #include "ggml-common.h" struct ggml_kleidiai_context { + cpu_feature features; ggml_kleidiai_kernels * kernels; -} static ctx = { NULL }; +} static ctx = { CPU_FEATURE_NONE, NULL }; static void init_kleidiai_context(void) { @@ -47,18 +48,18 @@ static void init_kleidiai_context(void) { const char *env_var = getenv("GGML_KLEIDIAI_SME"); int sme_enabled = 0; - cpu_feature features = (ggml_cpu_has_dotprod() ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) | - (ggml_cpu_has_matmul_int8() ? CPU_FEATURE_I8MM : CPU_FEATURE_NONE) | - (ggml_cpu_has_sve() ? CPU_FEATURE_SVE : CPU_FEATURE_NONE); + ctx.features = (ggml_cpu_has_dotprod() ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) | + (ggml_cpu_has_matmul_int8() ? CPU_FEATURE_I8MM : CPU_FEATURE_NONE) | + (ggml_cpu_has_sve() ? CPU_FEATURE_SVE : CPU_FEATURE_NONE); if (env_var) { sme_enabled = atoi(env_var); } if (sme_enabled != 0) { - features |= ggml_cpu_has_sme() ? CPU_FEATURE_SME : CPU_FEATURE_NONE; + ctx.features |= ggml_cpu_has_sme() ? CPU_FEATURE_SME : CPU_FEATURE_NONE; } - ctx.kernels = ggml_kleidiai_select_kernels(features); + ctx.kernels = ggml_kleidiai_select_kernels_q4_0(ctx.features); } ggml_critical_section_end(); } @@ -68,95 +69,275 @@ static inline int64_t ggml_ne(const ggml_tensor * tensor, int dim) { return tensor->ne[dim]; } +template +static Ret variant_call(const Variant & var, Args&&... args) { + return std::visit([&](auto&& func) -> Ret { + if constexpr (std::is_invocable_r_v) { + return func(std::forward(args)...); + } else { + throw std::runtime_error("Invalid function type in variant_call"); + } + }, var); +} + namespace ggml::cpu::kleidiai { + +static size_t round_down(size_t x, size_t y) { + return y == 0 ? x : x - (x % y); +} + +static void transpose_f32kxn_f16nxk(size_t n, size_t k, float * dst, const uint16_t * src, size_t rhs_stride) { + size_t src_stride = rhs_stride / sizeof(uint16_t); + size_t dst_stride = n; + + for (size_t k_idx = 0; k_idx < k; ++k_idx) { + for (size_t n_idx = 0; n_idx < n; ++n_idx) { + uint16_t v = *(src + k_idx + n_idx * src_stride); + *(dst + n_idx + k_idx * dst_stride) = kai_cast_f32_f16(v); + } + } +} + class tensor_traits : public ggml::cpu::tensor_traits { bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override { - GGML_ASSERT(ctx.kernels); - kernel_info * kernel = op->src[1]->ne[1] == 1 ? &ctx.kernels->gemv : &ctx.kernels->gemm; + ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, op); + GGML_ASSERT(kernels); + kernel_info * kernel = op->src[1]->ne[1] == 1 ? &kernels->gemv : &kernels->gemm; size_t k = op->src[0]->ne[0]; + size_t n = op->src[0]->ne[1]; size_t m = op->src[1]->ne[1]; size_t mr = kernel->get_mr(); size_t kr = kernel->get_kr(); size_t sr = kernel->get_sr(); - size = ctx.kernels->lhs_info.packed_size(m, k, QK4_0, mr, kr, sr); + if (kernels->rhs_type == GGML_TYPE_Q4_0) { + size = variant_call(kernels->lhs_info.packed_size, m, k, QK4_0, mr, kr, sr); + } else if (kernels->rhs_type == GGML_TYPE_F16) { + size = variant_call(kernels->lhs_info.packed_size, m, k, mr, kr, sr) + + variant_call(kernels->rhs_info.packed_size, n, k) + + k * n * sizeof(float) + n * sizeof(float); + } else { + GGML_ASSERT(false); + } return true; } + bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * dst) override { if (dst->op == GGML_OP_MUL_MAT) { - const ggml_tensor * src0 = dst->src[0]; - const ggml_tensor * src1 = dst->src[1]; + if (dst->src[0]->type == GGML_TYPE_Q4_0) { + return compute_forward_q4_0(params, dst); + } else if (dst->src[0]->type == GGML_TYPE_F16) { + return compute_forward_kv_cache(params, dst); + } + } + return false; + } - GGML_TENSOR_BINARY_OP_LOCALS + bool compute_forward_kv_cache(ggml_compute_params * params, struct ggml_tensor * dst) { + static std::atomic_flag first_to_arrive = ATOMIC_FLAG_INIT; - GGML_ASSERT(ctx.kernels); - kernel_info * kernel = src1->ne[1] == 1 ? &ctx.kernels->gemv : &ctx.kernels->gemm; - lhs_packing_info * lhs_info = &ctx.kernels->lhs_info; + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; - GGML_ASSERT(kernel); + GGML_TENSOR_BINARY_OP_LOCALS - const int ith = params->ith; - const int nth = params->nth; + ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst); + GGML_ASSERT(kernels); - const size_t k = ne00; - const size_t m = ne11; - const size_t n = ne01; + kernel_info * kernel = src1->ne[1] == 1 ? &kernels->gemv : &kernels->gemm; + GGML_ASSERT(kernel); - const size_t n_step = kernel->get_n_step(); - const size_t num_n_per_thread = kai_roundup(kai_roundup(n, nth) / nth, n_step); - const size_t n_start = ith * num_n_per_thread; + const int nth = params->nth; + const int ith = params->ith; - size_t n_to_process = num_n_per_thread; - if ((n_start + n_to_process) > n) { - n_to_process = n - n_start; + const int64_t lhs_batch_size0 = ne12; + const int64_t rhs_batch_size0 = ne02; + const int64_t batch_size = rhs_batch_size0; + + const int64_t r = lhs_batch_size0 / rhs_batch_size0; + + const int64_t m = ne11 * r; + const int64_t n = ne01; + const int64_t k = ne00; + + const size_t lhs_stride = src1->nb[1]; + const size_t rhs_stride = src0->nb[1]; + const size_t dst_stride = dst->nb[1]; + + const int64_t mr = static_cast(kernel->get_mr()); + const int64_t nr = static_cast(kernel->get_nr()); + const int64_t kr = static_cast(kernel->get_kr()); + const int64_t sr = static_cast(kernel->get_sr()); + + const size_t lhs_packed_size = variant_call(kernels->lhs_info.packed_size, m, k, mr, kr, sr); + const size_t rhs_packed_size = variant_call(kernels->rhs_info.packed_size, n, k); + const size_t kxn_size = k * n * sizeof(float); + const size_t bias_size = n * sizeof(float); + + const size_t wsize_required = lhs_packed_size + rhs_packed_size + kxn_size + bias_size; + GGML_ASSERT(wsize_required <= params->wsize); + + uint8_t * lhs_packed = static_cast(params->wdata); + uint8_t * rhs_packed = lhs_packed + lhs_packed_size; + uint8_t * rhs_kxn = rhs_packed + rhs_packed_size; + uint8_t * bias = rhs_kxn + kxn_size; + + for (int64_t batch_idx = 0; batch_idx < batch_size; ++batch_idx) { + const uint8_t * lhs_batch = static_cast(src1->data) + batch_idx * m * lhs_stride; + const uint8_t * rhs_batch = static_cast(src0->data) + batch_idx * n * rhs_stride; + uint8_t * dst_batch = static_cast(dst->data) + batch_idx * m * dst_stride; + + // LHS packing + { + const int64_t m_roundup_mr = kai_roundup(m, mr); + const int64_t num_threads = KAI_MIN(m_roundup_mr / mr, nth); + + if (ith < num_threads) { + const int64_t num_m_per_thread0 = round_down(m_roundup_mr / num_threads, mr); + const int64_t num_m_per_threadN_1 = m - (num_threads - 1) * num_m_per_thread0; + + const int64_t m_start = ith * num_m_per_thread0; + const int64_t num_m_per_thread = (ith == num_threads - 1) ? num_m_per_threadN_1 : num_m_per_thread0; + + const size_t lhs_offset = variant_call(kernels->gemm.get_lhs_offset, m_start, lhs_stride); + const size_t lhs_packed_offset = variant_call(kernels->lhs_info.get_packed_offset, m_start, k, mr, kr, sr); + + const void * src_ptr = static_cast(lhs_batch) + lhs_offset; + void * dst_ptr = static_cast(lhs_packed) + lhs_packed_offset; + + variant_call(kernels->lhs_info.pack_func, num_m_per_thread, k, mr, kr, sr, 0, src_ptr, lhs_stride, dst_ptr); + } } - const uint8_t * lhs = static_cast(src1->data); - uint8_t * lhs_packed = (uint8_t*)params->wdata; - const uint8_t * rhs_packed = static_cast(src0->data); + // RHS packing + if (first_to_arrive.test_and_set(std::memory_order_acquire) == false) { + // First thread to reach this point handles RHS packing + memset(bias, 0, n * sizeof(float)); + transpose_f32kxn_f16nxk(n, k, reinterpret_cast(rhs_kxn), + reinterpret_cast(rhs_batch), rhs_stride); - size_t mr = kernel->get_mr(); - size_t kr = kernel->get_kr(); - size_t sr = kernel->get_sr(); - - // Calculate number of columns to be processed per thread - const size_t num_m_per_thread = kai_roundup(m, mr * nth) / nth; - const size_t m_start = ith * num_m_per_thread; - size_t m_to_process = num_m_per_thread; - if ((m_start + m_to_process) > m) { - m_to_process = m - m_start; - } - - if(m_start < m) { - // Transform LHS - const size_t src_stride = src1->nb[1]; - const float * src_ptr = reinterpret_cast(lhs + lhs_info->get_offset(m_start, dst->src[1]->nb[1])); - const size_t lhs_packed_offset = lhs_info->get_packed_offset(m_start, k, QK4_0, mr, kr, sr); - void * lhs_packed_ptr = static_cast(lhs_packed + lhs_packed_offset); - - lhs_info->pack_func(m_to_process, k, QK4_0, mr, kr, sr, 0, src_ptr, src_stride, lhs_packed_ptr); + variant_call(kernels->rhs_info.pack_func, 1, n, k, nr, kr, sr, n * sizeof(float), + rhs_kxn, bias, nullptr, rhs_packed, 0, nullptr); } ggml_barrier(params->threadpool); - // Perform the operation - const size_t dst_stride = dst->nb[1]; - const size_t lhs_packed_offset = lhs_info->get_packed_offset(0, k, QK4_0, mr, kr, sr); - const size_t rhs_packed_offset = kernel->get_rhs_packed_offset(n_start, k, QK4_0); - const size_t dst_offset = kernel->get_dst_offset(0, n_start, dst_stride); - const void * rhs_ptr = static_cast(rhs_packed + rhs_packed_offset); - const void* lhs_ptr = (const void*)((const char *)lhs_packed + lhs_packed_offset); - float *dst_ptr = reinterpret_cast(static_cast(dst->data) + dst_offset); + first_to_arrive.clear(std::memory_order_release); - kernel->run_kernel(m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, - dst_stride, sizeof(float), -FLT_MAX, FLT_MAX); - return true; + // Perform the matmul + { + const int64_t m_to_process = m; + const int64_t m_start = 0; + + const int64_t n_step = static_cast(kernel->get_n_step()); + const int64_t num_threads = KAI_MIN(n / n_step, nth); + + if (ith < num_threads) { + const int64_t num_n_per_thread0 = round_down(n / num_threads, n_step); + const int64_t num_n_per_threadN_1 = n - (num_threads - 1) * num_n_per_thread0; + + const int64_t n_start = ith * num_n_per_thread0; + const int64_t n_to_process = (ith == num_threads - 1) ? num_n_per_threadN_1 : num_n_per_thread0; + + const size_t lhs_packed_offset = variant_call(kernel->get_lhs_offset, m_start, k); + const size_t rhs_packed_offset = variant_call(kernel->get_rhs_packed_offset, n_start, k); + const size_t dst_offset = kernel->get_dst_offset(m_start, n_start, dst_stride); + + const void * lhs_ptr = lhs_packed + lhs_packed_offset; + const void * rhs_ptr = rhs_packed + rhs_packed_offset; + float * dst_ptr = reinterpret_cast(dst_batch + dst_offset); + + variant_call(kernel->run_kernel, m_to_process, n_to_process, k, lhs_ptr, rhs_ptr, dst_ptr, dst_stride, sizeof(float), -FLT_MAX, FLT_MAX); + } + } + + if (batch_idx != batch_size - 1) { + // This barrier is necessary when the batch size is larger than 1. While processing a batch, + // the work data buffer (params->wdata) is used as temporary storage which means that only + // a single batch can be processed at any given time. No barrier is needed for the last + // batch since GGML inserts a barrier between the execution of every operator. + ggml_barrier(params->threadpool); + } } - return false; + + return true; + } + + bool compute_forward_q4_0(struct ggml_compute_params * params, struct ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS + + ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst); + GGML_ASSERT(kernels); + + kernel_info * kernel = src1->ne[1] == 1 ? &kernels->gemv : &kernels->gemm; + lhs_packing_info * lhs_info = &kernels->lhs_info; + + GGML_ASSERT(kernel); + + const int ith = params->ith; + const int nth = params->nth; + + const size_t k = ne00; + const size_t m = ne11; + const size_t n = ne01; + + size_t mr = kernel->get_mr(); + size_t kr = kernel->get_kr(); + size_t sr = kernel->get_sr(); + + const uint8_t * lhs = static_cast(src1->data); + uint8_t * lhs_packed = (uint8_t*)params->wdata; + const uint8_t * rhs_packed = static_cast(src0->data); + + const size_t n_step = kernel->get_n_step(); + const size_t num_n_per_thread = kai_roundup(kai_roundup(n, nth) / nth, n_step); + const size_t n_start = ith * num_n_per_thread; + + size_t n_to_process = num_n_per_thread; + if ((n_start + n_to_process) > n) { + n_to_process = n - n_start; + } + + // Calculate number of columns to be processed per thread + const size_t num_m_per_thread = kai_roundup(m, mr * nth) / nth; + const size_t m_start = ith * num_m_per_thread; + size_t m_to_process = num_m_per_thread; + if ((m_start + m_to_process) > m) { + m_to_process = m - m_start; + } + + if (m_start < m) { + // Transform LHS + const size_t src_stride = src1->nb[1]; + const float * src_ptr = reinterpret_cast(lhs + lhs_info->get_offset(m_start, dst->src[1]->nb[1])); + const size_t lhs_packed_offset = variant_call(lhs_info->get_packed_offset, m_start, k, QK4_0, mr, kr, sr); + void * lhs_packed_ptr = static_cast(lhs_packed + lhs_packed_offset); + + variant_call(lhs_info->pack_func, m_to_process, k, QK4_0, mr, kr, sr, 0, src_ptr, src_stride, lhs_packed_ptr); + } + + ggml_barrier(params->threadpool); + + // Perform the operation + const size_t dst_stride = dst->nb[1]; + const size_t lhs_packed_offset = variant_call(lhs_info->get_packed_offset, 0, k, QK4_0, mr, kr, sr); + const size_t rhs_packed_offset = variant_call(kernel->get_rhs_packed_offset, n_start, k, QK4_0); + const size_t dst_offset = kernel->get_dst_offset(0, n_start, dst_stride); + const void * rhs_ptr = static_cast(rhs_packed + rhs_packed_offset); + const void* lhs_ptr = (const void*)((const char *)lhs_packed + lhs_packed_offset); + float *dst_ptr = reinterpret_cast(static_cast(dst->data) + dst_offset); + + variant_call(kernel->run_kernel, m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride, + sizeof(float), -FLT_MAX, FLT_MAX); + + return true; } public: @@ -169,13 +350,13 @@ public: size_t sr = ctx.kernels->gemm.get_sr(); #ifndef NDEBUG - const size_t repacked_size = ctx.kernels->rhs_info.packed_size(n, k, nr, kr, QK4_0); + const size_t repacked_size = variant_call(ctx.kernels->rhs_info.packed_size, n, k, nr, kr, QK4_0); GGML_ASSERT(repacked_size <= data_size && "repacked size larger than the packed size!"); #endif struct kai_rhs_pack_qs4cxs1s0_param params; params.lhs_zero_point = 1; params.rhs_zero_point = 8; - ctx.kernels->rhs_info.pack_func(1, n, k, nr, kr, sr, QK4_0, (const uint8_t *)data, NULL, tensor->data, 0, ¶ms); + variant_call(ctx.kernels->rhs_info.pack_func, 1, n, k, nr, kr, sr, QK4_0, (const uint8_t*)data, nullptr, tensor->data, 0, ¶ms); return 0; @@ -189,7 +370,7 @@ static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struc } } // namespace ggml::cpu::kleidiai -GGML_API enum ggml_status ggml_backend_cpu_kleidiai_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { +static enum ggml_status ggml_backend_cpu_kleidiai_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { tensor->extra = (void *) ggml::cpu::kleidiai::get_tensor_traits(buffer, tensor); GGML_UNUSED(buffer); @@ -238,12 +419,11 @@ static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alignment(ggml_backend_b namespace ggml::cpu::kleidiai { class extra_buffer_type : ggml::cpu::extra_buffer_type { bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override { - if ( op->op == GGML_OP_MUL_MAT && - op->src[0]->type == GGML_TYPE_Q4_0 && - op->src[0]->buffer && - (ggml_n_dims(op->src[0]) == 2) && - op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() && ctx.kernels - ) { + if (op->op == GGML_OP_MUL_MAT && + op->src[0]->type == GGML_TYPE_Q4_0 && + op->src[0]->buffer && + (ggml_n_dims(op->src[0]) == 2) && + op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() && ctx.kernels) { if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) { return false; } @@ -260,6 +440,19 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type { if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type()) { return (ggml::cpu::tensor_traits *) op->src[0]->extra; } + else if (ggml_kleidiai_select_kernels(ctx.features, op) && + op->src[0]->op == GGML_OP_VIEW && + (op->src[1]->op == GGML_OP_PERMUTE || op->src[1]->op == GGML_OP_SOFT_MAX) && + op->src[1]->ne[1] > 1) { + if ((op->src[0]->nb[0] != 2) || + (op->src[1]->nb[0] != 4) || + (op->src[0]->nb[1] * op->src[0]->ne[1] != op->src[0]->nb[2]) || + (op->src[1]->nb[1] * op->src[1]->ne[1] != op->src[1]->nb[2])) { + return nullptr; + } + + return ggml::cpu::kleidiai::get_tensor_traits(NULL, NULL); + } } return nullptr; } diff --git a/ggml/src/ggml-cpu/llamafile/sgemm.cpp b/ggml/src/ggml-cpu/llamafile/sgemm.cpp index f6374f7894..1d46158f92 100644 --- a/ggml/src/ggml-cpu/llamafile/sgemm.cpp +++ b/ggml/src/ggml-cpu/llamafile/sgemm.cpp @@ -1054,6 +1054,493 @@ class tinyBLAS_Q0_AVX { } \ } \ +template +class tinyBLAS_BF16_PPC { + public: + tinyBLAS_BF16_PPC(int64_t k, + const TA *A, int64_t lda, + const TB *B, int64_t ldb, + TC *C, int64_t ldc, + int ith, int nth) + : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) { + } + + void matmul(int64_t m, int64_t n) { + mnpack(0, m, 0, n); + } + + private: + void vector_permute_store(vec_t *c, int numVec, unsigned char *vecOffset) { + vec_t t[8], s[8]; + vec_t swiz1 = {0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23}; + vec_t swiz2 = {8, 9, 10, 11, 24, 25, 26, 27, 12, 13, 14, 15, 28, 29, 30, 31}; + vec_t swiz3 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23}; + vec_t swiz4 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31}; + + if (numVec == 2) { + t[0] = vec_perm(c[0], c[1], swiz1); + t[1] = vec_perm(c[2], c[3], swiz1); + s[0] = vec_perm(t[0], t[1], swiz3); + s[1] = vec_perm(t[0], t[1], swiz4); + vec_xst(s[0], 0, (vec_t*)vecOffset); + vec_xst(s[1], 0, (vec_t*)(vecOffset + 16)); + } else if (numVec == 4) { + t[0] = vec_perm(c[0], c[1], swiz1); + t[1] = vec_perm(c[0], c[1], swiz2); + t[2] = vec_perm(c[2], c[3], swiz1); + t[3] = vec_perm(c[2], c[3], swiz2); + s[0] = vec_perm(t[0], t[2], swiz3); + s[1] = vec_perm(t[0], t[2], swiz4); + s[2] = vec_perm(t[1], t[3], swiz3); + s[3] = vec_perm(t[1], t[3], swiz4); + for (int i = 0; i < 4; ++i) + vec_xst(s[i], 0, (vec_t*)(vecOffset + i * 16)); + } else if (numVec == 8) { + for (int i = 0; i < 4; i += 2) { + t[i+0] = vec_perm(c[i+0], c[i+1], swiz1); + t[i+1] = vec_perm(c[i+0], c[i+1], swiz2); + } + for (int i = 4; i < 8; i += 2) { + t[i+0] = vec_perm(c[i+0], c[i+1], swiz1); + t[i+1] = vec_perm(c[i+0], c[i+1], swiz2); + } + s[0] = vec_perm(t[0], t[2], swiz3); + s[1] = vec_perm(t[0], t[2], swiz4); + s[2] = vec_perm(t[1], t[3], swiz3); + s[3] = vec_perm(t[1], t[3], swiz4); + s[4] = vec_perm(t[4], t[6], swiz3); + s[5] = vec_perm(t[4], t[6], swiz4); + s[6] = vec_perm(t[5], t[7], swiz3); + s[7] = vec_perm(t[5], t[7], swiz4); + for (int i = 0; i < 8; ++i) + vec_xst(s[i], 0, (vec_t*)(vecOffset + i * 16)); + } + } + + void packNormal(const TA* a, int64_t lda, int rows, int cols, unsigned char* vec) { + int64_t i, j; + TA *aoffset = NULL; + unsigned char *vecOffset = NULL; + TA * aoffsets[8]; + vector unsigned char c_arr[8]; + aoffset = const_cast(a); + vecOffset = vec; + j = (rows >> 3); + if (j > 0) { + do { + if (cols == 4) { + aoffsets[0] = aoffset; + for (int it = 1; it < 4; ++it) + aoffsets[it] = aoffsets[it-1] + lda; + aoffset += 4 * lda; + for (int i = 0; i < 4; ++i) + c_arr[i] = vec_xl(0, (vector unsigned char*)aoffsets[i]); + vector_permute_store(c_arr, 4, vecOffset); + for (int i = 0; i<4; i++) + aoffsets[i] = aoffsets[i]+lda; + vecOffset +=64; + } + i = (cols >> 3); + if (i > 0) { + aoffsets[0] = aoffset; + for (int it = 1; it < 8; ++it) { + aoffsets[it] = aoffsets[it-1] + lda; + } + aoffset += 8 * lda; + do { + for (int it = 0; it < 8; ++it) + c_arr[it] = vec_xl(0, (vector unsigned char*)aoffsets[it]); + vector_permute_store(c_arr, 8, vecOffset); + for (int it = 0; it < 8; ++it) + aoffsets[it] = aoffsets[it] + 8*lda; + vecOffset += 128; + i--; + } while(i > 0); + } + j--; + } while(j > 0); + } + if (rows & 4) { + aoffsets[0] = aoffset; + for (int it = 1; it < 4; ++it) + aoffsets[it] = aoffsets[it-1] + lda; + aoffset += 4 * lda; + if (cols == 4) { + for (int it = 0; it < 4; ++it) + c_arr[it] = vec_xl(0, (vector unsigned char*)aoffsets[it]); + vector_permute_store(c_arr, 2, vecOffset); + for (int it = 0; it< 4; it++) + aoffsets[it] = aoffsets[it] + lda; + vecOffset += 32; + } + i = (cols >> 3); + if (i > 0) { + do { + for (int it = 0; it < 4; ++it) + c_arr[it] = vec_xl(0, (vector unsigned char*)aoffsets[it]); + vector_permute_store(c_arr, 4, vecOffset); + for (int it = 0; it< 4; it++) + aoffsets[it] = aoffsets[it] + 8*lda; + vecOffset += 64; + i--; + } while(i > 0); + } + } + if (rows & 3) { + aoffsets[0] = aoffset; + for (int it = 1; it < 4; ++it) + aoffsets[it] = aoffsets[it-1] + lda; + if (cols == 4) { + switch(rows) { + case 3: c_arr[2] = vec_xl(0, (vector unsigned char*)aoffsets[2]); + case 2: c_arr[1] = vec_xl(0, (vector unsigned char*)aoffsets[1]); + case 1: c_arr[0] = vec_xl(0, (vector unsigned char*)aoffsets[0]); + break; + } + vector_permute_store(c_arr, 2, vecOffset); + for (int it = 0; it< 4; it++) + aoffsets[it] = aoffsets[it] + lda; + vecOffset += 32; + } + i = (cols >> 3); + if (i > 0) { + do { + switch(rows) { + case 3: c_arr[2] = vec_xl(0, (vector unsigned char*)aoffsets[2]); + case 2: c_arr[1] = vec_xl(0, (vector unsigned char*)aoffsets[1]); + case 1: c_arr[0] = vec_xl(0, (vector unsigned char*)aoffsets[0]); + break; + } + vector_permute_store(c_arr, 4, vecOffset); + for (int it = 0; it <4; it++) + aoffsets[it] = aoffsets[it] + 8* lda; + vecOffset += 64; + i--; + } while(i > 0); + } + } + } + + void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) { + int64_t mc, nc, mp, np; + int m_rem = MIN(m - m0, 8); + int n_rem = MIN(n - n0, 8); + + if (m_rem >= 8 && n_rem >= 8) { + mc = 8; + nc = 8; + gemm<8,8>(m0, m, n0, n); + } else if (m_rem >= 4 && n_rem >= 8) { + mc = 4; + nc = 8; + gemm<4,8>(m0, m, n0, n); + } else if (m_rem >=8 && n_rem >=4){ + mc = 8; + nc = 4; + gemm<8,4>(m0, m, n0, n); + } else if ((m_rem < 4) && (n_rem >= 8)) { + nc = 8; + switch(m_rem) { + case 1: + mc = 1; + gemm_Mx8<1>(m0, m, n0, n); + break; + case 2: + mc = 2; + gemm_Mx8<2>(m0, m, n0, n); + break; + case 3: + mc = 3; + gemm_Mx8<3>(m0, m, n0, n); + break; + default: + return; + } + } else if (m_rem >= 4 && n_rem >= 4) { + mc = 4; + nc = 4; + gemm_small<4, 4>(m0, m, n0, n); + } else if ((m_rem > 4) && (n_rem < 4)) { + mc = 4; + switch(n_rem) { + case 1: + nc = 1; + gemm_small<4, 1>(m0, m, n0, n); + break; + case 2: + nc = 2; + gemm_small<4, 2>(m0, m, n0, n); + break; + case 3: + nc = 3; + gemm_small<4, 3>(m0, m, n0, n); + break; + + default: + return; + } + } else { + switch((m_rem << 4) | n_rem) { + case 0x43: + mc = 4; + nc = 3; + gemm_small<4, 3>(m0, m, n0, n); + break; + case 0x42: + mc = 4; + nc = 2; + gemm_small<4, 2>(m0, m, n0, n); + break; + case 0x41: + mc = 4; + nc = 1; + gemm_small<4, 1>(m0, m, n0, n); + break; + case 0x34: + mc = 3; + nc = 4; + gemm_small<3, 4>(m0, m, n0, n); + break; + case 0x33: + mc = 3; + nc = 3; + gemm_small<3, 3>(m0, m, n0, n); + break; + case 0x32: + mc = 3; + nc = 2; + gemm_small<3, 2>(m0, m, n0, n); + break; + case 0x31: + mc = 3; + nc = 1; + gemm_small<3, 1>(m0, m, n0, n); + break; + case 0x24: + mc = 2; + nc = 4; + gemm_small<2,4>(m0, m, n0, n); + break; + case 0x23: + mc = 2; + nc = 3; + gemm_small<2, 3>(m0, m, n0, n); + break; + case 0x22: + mc = 2; + nc = 2; + gemm_small<2, 2>(m0, m, n0, n); + break; + case 0x21: + mc = 2; + nc = 1; + gemm_small<2, 1>(m0, m, n0, n); + break; + case 0x14: + mc = 1; + nc = 4; + gemm_small<1, 4>(m0, m, n0, n); + break; + case 0x13: + mc = 1; + nc = 3; + gemm_small<1, 3>(m0, m, n0, n); + break; + case 0x12: + mc = 1; + nc = 2; + gemm_small<1, 2>(m0, m, n0, n); + break; + case 0x11: + mc = 1; + nc = 1; + gemm_small<1, 1>(m0, m, n0, n); + break; + default: + return; + } + } + mp = m0 + (m - m0) / mc * mc; + np = n0 + (n - n0) / nc * nc; + mnpack(mp, m, n0, np); + mnpack(m0, m, np, n); + } + + void KERNEL_4x8(int64_t ii, int64_t jj) { + vec_t vec_A[4], vec_B[8] , vec_C[4]; + acc_t acc_0, acc_1; + __builtin_mma_xxsetaccz(&acc_0); + __builtin_mma_xxsetaccz(&acc_1); + for (int l = 0; l < k; l+=8) { + 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]); + } + } + SAVE_ACC(&acc_0, ii, jj); + SAVE_ACC(&acc_1, ii, jj+4); + } + + void KERNEL_8x4(int64_t ii, int64_t jj) { + vec_t vec_A[8], vec_B[4] , vec_C[4]; + acc_t acc_0, acc_1; + __builtin_mma_xxsetaccz(&acc_0); + __builtin_mma_xxsetaccz(&acc_1); + for (int l = 0; l < k; l+=8) { + 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]); + } + } + SAVE_ACC(&acc_0, ii, jj); + SAVE_ACC(&acc_1, ii+4, jj); + } + + + void KERNEL_8x8(int64_t ii, int64_t jj) { + vec_t vec_A[8], vec_B[8], vec_C[4]; + acc_t acc_0, acc_1, acc_2, acc_3; + __builtin_mma_xxsetaccz(&acc_0); + __builtin_mma_xxsetaccz(&acc_1); + __builtin_mma_xxsetaccz(&acc_2); + __builtin_mma_xxsetaccz(&acc_3); + for (int l = 0; l < k; l+=8) { + 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]); + } + } + + SAVE_ACC(&acc_0, ii, jj); + SAVE_ACC(&acc_1, ii, jj+4); + SAVE_ACC(&acc_2, ii+4, jj); + SAVE_ACC(&acc_3, ii+4, jj+4); + } + + template + void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n) { + int64_t ytiles = (m - m0) / RM; + int64_t xtiles = (n - n0) / RN; + int64_t tiles = xtiles * ytiles; + int64_t duty = (tiles + nth - 1) / nth; + int64_t start = duty * ith; + int64_t end = start + duty; + if (end > tiles) + end = tiles; + for (int64_t job = start; job < end; ++job) { + int64_t ii = m0 + job / xtiles * RM; + int64_t jj = n0 + job % xtiles * RN; + vec_t vec_C[4]; + acc_t acc_0; + __builtin_mma_xxsetaccz(&acc_0); + vec_t vec_A[2], vec_B[2]; + for (int l=0; l + void gemm_Mx8(int64_t m0, int64_t m, int64_t n0, int64_t n) { + int RN = 8; + int64_t ytiles = (m - m0) / RM; + int64_t xtiles = (n - n0) / RN; + int64_t tiles = xtiles * ytiles; + int64_t duty = (tiles + nth - 1) / nth; + int64_t start = duty * ith; + int64_t end = start + duty; + if (end > tiles) + end = tiles; + for (int64_t job = start; job < end; ++job) { + int64_t ii = m0 + job / xtiles * RM; + int64_t jj = n0 + job % xtiles * RN; + vec_t vec_C[4]; + acc_t acc_0, acc_1; + __builtin_mma_xxsetaccz(&acc_0); + __builtin_mma_xxsetaccz(&acc_1); + vec_t vec_A[4], vec_B[8]; + for (int l=0; l + inline void kernel(int64_t ii, int64_t jj) { + if constexpr(RM == 4 && RN == 8) { + KERNEL_4x8(ii,jj); + } else if constexpr(RM == 8 && RN == 8) { + KERNEL_8x8(ii,jj); + } else if constexpr(RM == 8 && RN == 4) { + KERNEL_8x4(ii,jj); + } else { + static_assert(false, "RN/RM values not supported"); + } + } + + template + NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) { + int64_t ytiles = (m - m0) / RM; + int64_t xtiles = (n - n0) / RN; + int64_t tiles = xtiles * ytiles; + int64_t duty = (tiles + nth - 1) / nth; + int64_t start = duty * ith; + int64_t end = start + duty; + if (end > tiles) + end = tiles; + for (int64_t job = start; job < end; ++job) { + int64_t ii = m0 + job / xtiles * RM; + int64_t jj = n0 + job % xtiles * RN; + kernel(ii, jj); + } + } + + const TA *const A; + const TB *const B; + TC *C; + const int64_t k; + const int64_t lda; + const int64_t ldb; + const int64_t ldc; + const int ith; + const int nth; +}; + template class tinyBLAS_Q0_PPC { public: @@ -2202,6 +2689,7 @@ class tinyBLAS_PPC { boffset = vec; j = (rows >> 3); if (j > 0) { + do { aoffset1 = aoffset; aoffset2 = aoffset1 + lda; @@ -2875,9 +3363,22 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64 (float *)C, ldc}; return tb.matmul(m, n); } +#elif defined(__MMA__) + if ((k % 8)) + return false; + if(Btype == GGML_TYPE_BF16) { + tinyBLAS_BF16_PPC 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; + } #endif return false; } + case GGML_TYPE_F16: { #if defined(__AVX512F__) if (Btype == GGML_TYPE_F16) { diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp index 3c2adb2172..955fec59a6 100644 --- a/ggml/src/ggml-cpu/ops.cpp +++ b/ggml/src/ggml-cpu/ops.cpp @@ -8,19 +8,6 @@ #include -#if defined(_MSC_VER) -// disable "possible loss of data" to avoid hundreds of casts -// we should just be careful :) -#pragma warning(disable: 4244 4267) - -// disable POSIX deprecation warnings -// these functions are never going away, anyway -#pragma warning(disable: 4996) - -// unreachable code because of multiple instances of code after GGML_ABORT -#pragma warning(disable: 4702) -#endif - // ggml_compute_forward_dup static void ggml_compute_forward_dup_same_cont( @@ -4222,7 +4209,7 @@ static void ggml_compute_forward_get_rows_f16( GGML_ASSERT(i01 >= 0 && i01 < ne01); - ggml_fp16_to_fp32_row( + ggml_cpu_fp16_to_fp32( (const ggml_fp16_t*) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03), (float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3), nc); } @@ -4263,7 +4250,7 @@ static void ggml_compute_forward_get_rows_bf16( GGML_ASSERT(i01 >= 0 && i01 < ne01); - ggml_bf16_to_fp32_row( + ggml_cpu_bf16_to_fp32( (const ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03), (float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3), nc); } diff --git a/ggml/src/ggml-cpu/simd-mappings.h b/ggml/src/ggml-cpu/simd-mappings.h index 04d10cec26..45c31cf1fa 100644 --- a/ggml/src/ggml-cpu/simd-mappings.h +++ b/ggml/src/ggml-cpu/simd-mappings.h @@ -341,7 +341,7 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) { #define GGML_F32_EPR 4 #define GGML_F32x4 vector float -#define GGML_F32x4_ZERO 0.0f +#define GGML_F32x4_ZERO {0.0f} #define GGML_F32x4_SET1 vec_splats #define GGML_F32x4_LOAD(p) vec_xl(0, p) #define GGML_F32x4_STORE(p, r) vec_xst(r, 0, p) diff --git a/ggml/src/ggml-cpu/vec.cpp b/ggml/src/ggml-cpu/vec.cpp index dfe2218e30..02d4061822 100644 --- a/ggml/src/ggml-cpu/vec.cpp +++ b/ggml/src/ggml-cpu/vec.cpp @@ -2,12 +2,6 @@ #include -#if defined(_MSC_VER) -// disable "possible loss of data" to avoid hundreds of casts -// we should just be careful :) -#pragma warning(disable: 4244 4267) -#endif - // precomputed gelu table for f16 (128 KB) ggml_fp16_t ggml_table_gelu_f16[1 << 16]; diff --git a/ggml/src/ggml-cuda/CMakeLists.txt b/ggml/src/ggml-cuda/CMakeLists.txt index 8623214c78..c9ff4aa321 100644 --- a/ggml/src/ggml-cuda/CMakeLists.txt +++ b/ggml/src/ggml-cuda/CMakeLists.txt @@ -12,12 +12,30 @@ if (CUDAToolkit_FOUND) # 61 == Pascal, __dp4a instruction (per-byte integer dot product) # 70 == V100, FP16 tensor cores # 75 == Turing, int8 tensor cores + # 80 == Ampere, asynchronous data loading, faster tensor core instructions + # 86 == RTX 3000, needs CUDA v11.1 + # 89 == RTX 4000, needs CUDA v11.8 + # + # XX-virtual == compile CUDA code as PTX, do JIT compilation to binary code on first run + # XX-real == compile CUDA code as device code for this specific architecture + # no suffix == compile as both PTX and device code + # + # The default behavior for a non-native is to build virtual architectures as needed to cover all features needed + # for best performance and to also build real architectures for the most commonly used GPUs. if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24") set(CMAKE_CUDA_ARCHITECTURES "native") elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16) - set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75;80") + if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8") + set(CMAKE_CUDA_ARCHITECTURES "60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real;89-real") + else() + set(CMAKE_CUDA_ARCHITECTURES "60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real") + endif() else() - set(CMAKE_CUDA_ARCHITECTURES "50;61;70;75;80") + if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8") + set(CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real;89-real") + else() + set(CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real") + endif() endif() endif() message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}") @@ -100,7 +118,7 @@ if (CUDAToolkit_FOUND) set(CUDA_CXX_FLAGS "") - set(CUDA_FLAGS -use_fast_math) + set(CUDA_FLAGS -use_fast_math -extended-lambda) if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8") # Options are: @@ -133,6 +151,7 @@ if (CUDAToolkit_FOUND) COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion" OUTPUT_VARIABLE CUDA_CCVER ERROR_QUIET + OUTPUT_STRIP_TRAILING_WHITESPACE ) else() if (CUDA_CCFULLVER MATCHES Apple) @@ -143,7 +162,7 @@ if (CUDAToolkit_FOUND) string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER}) endif() - message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}") + message(STATUS "CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}") ggml_get_flags(${CUDA_CCID} ${CUDA_CCVER}) list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS}) # This is passed to -Xcompiler later diff --git a/ggml/src/ggml-cuda/acc.cu b/ggml/src/ggml-cuda/acc.cu index 96bfe1c9d8..e084607c02 100644 --- a/ggml/src/ggml-cuda/acc.cu +++ b/ggml/src/ggml-cuda/acc.cu @@ -1,47 +1,61 @@ #include "acc.cuh" -static __global__ void acc_f32(const float * x, const float * y, float * dst, const int ne, - const int ne10, const int ne11, const int ne12, - const int nb1, const int nb2, int offset) { - const int i = blockDim.x * blockIdx.x + threadIdx.x; +static __global__ void acc_f32(const float * x, const float * y, float * dst, const int64_t ne, + const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13, + const int64_t s11, const int64_t s12, const int64_t s13, const int64_t offset) { + const int64_t i = blockDim.x * blockIdx.x + threadIdx.x; + if (i >= ne) { return; } - int src1_idx = i - offset; - int oz = src1_idx / nb2; - int oy = (src1_idx - (oz * nb2)) / nb1; - int ox = src1_idx % nb1; - if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) { - dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11]; - } else { - dst[i] = x[i]; + + int64_t src1_idx = i - offset; + + int64_t tmp = src1_idx; + const int64_t i13 = tmp / s13; + tmp -= i13 * s13; + const int64_t i12 = tmp / s12; + tmp -= i12 * s12; + const int64_t i11 = tmp / s11; + tmp -= i11 * s11; + const int64_t i10 = tmp; + + float val = x[i]; + if (src1_idx >= 0 && i10 < ne10 && i11 < ne11 && i12 < ne12 && i13 < ne13) { + val += y[((i13*ne12 + i12) * ne11 + i11) * ne10 + i10]; } + dst[i] = val; } -static void acc_f32_cuda(const float * x, const float * y, float * dst, const int n_elements, - const int ne10, const int ne11, const int ne12, - const int nb1, const int nb2, const int offset, cudaStream_t stream) { - int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE; - acc_f32<<>>(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset); +static void acc_f32_cuda(const float * x, const float * y, float * dst, const int64_t n_elements, + const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13, + const int64_t s1, const int64_t s2, const int64_t s3, const int64_t offset, cudaStream_t stream) { + const int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE; + acc_f32<<>>(x, y, dst, n_elements, ne10, ne11, ne12, ne13, s1, s2, s3, offset); } void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * src0 = dst->src[0]; const ggml_tensor * src1 = dst->src[1]; - const float * src0_d = (const float *)src0->data; - const float * src1_d = (const float *)src1->data; - float * dst_d = (float *)dst->data; + + const float * src0_d = (const float *) src0->data; + const float * src1_d = (const float *) src1->data; + float * dst_d = (float *) dst->data; + cudaStream_t stream = ctx.stream(); GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT(src1->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); - GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported - int nb1 = dst->op_params[0] / 4; // 4 bytes of float32 - int nb2 = dst->op_params[1] / 4; // 4 bytes of float32 - // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused - int offset = dst->op_params[3] / 4; // offset in bytes + GGML_ASSERT(ggml_is_contiguous(src1)); + GGML_ASSERT(dst->nb[0] == ggml_element_size(dst)); + GGML_ASSERT(ggml_is_contiguously_allocated(dst)); - acc_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, stream); + const int64_t s1 = dst->op_params[0] / sizeof(float); + const int64_t s2 = dst->op_params[1] / sizeof(float); + const int64_t s3 = dst->op_params[2] / sizeof(float); + const int64_t offset = dst->op_params[3] / sizeof(float); + + acc_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3], s1, s2, s3, offset, stream); } diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index 8284a0017d..64fb4ff4ce 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -78,13 +78,13 @@ // Moore Threads #define GGML_CUDA_MUSA_ARCH_IS_QY1 (__MUSA_ARCH__ <= 210) -#define GGML_CUDA_CC_QY1 (GGML_MUSA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000 -#define GGML_CUDA_CC_QY2 (GGML_MUSA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000 -#define GGML_CUDA_CC_NG (GGML_MUSA_CC_OFFSET_MTHREADS + 0x310) // TBD +#define GGML_CUDA_CC_QY1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000 +#define GGML_CUDA_CC_QY2 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000 +#define GGML_CUDA_CC_NG (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD #define GGML_CUDA_CC_IS_MTHREADS(cc) (cc >= GGML_CUDA_CC_OFFSET_MTHREADS && cc < GGML_CUDA_CC_OFFSET_AMD) #define GGML_CUDA_CC_IS_QY1(cc) (cc >= GGML_CUDA_CC_QY1 && cc < GGML_CUDA_CC_QY2) -#define GGML_CUDA_CC_IS_QY2(cc) (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NEXT) +#define GGML_CUDA_CC_IS_QY2(cc) (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NG) #define GGML_CUDA_CC_IS_NG(cc) (cc >= GGML_CUDA_CC_NG) #ifdef __CUDA_ARCH_LIST__ @@ -130,10 +130,6 @@ static int ggml_cuda_highest_compiled_arch(const int arch) { #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses -#if defined(_MSC_VER) -#pragma warning(disable: 4244 4267) // possible loss of data -#endif - #define GGML_CUDA_MAX_STREAMS 8 [[noreturn]] @@ -300,6 +296,25 @@ static __device__ void no_device_code( #define NO_DEVICE_CODE //GGML_ABORT("NO_DEVICE_CODE not valid in host code.") #endif // __CUDA_ARCH__ +// The compiler is always able to unroll loops if they contain continue expressions. +// In such cases loop unrolling can still be achieved via recursion: +template +struct ggml_cuda_unroll { + template + __device__ void operator()(const Func & f, Args... args) const { + f(n - 1, args...); + ggml_cuda_unroll{}(f, args...); + } +}; + +template <> +struct ggml_cuda_unroll<1> { + template + __device__ void operator()(const Func & f, Args... args) const { + f(0, args...); + } +}; + template static __device__ __forceinline__ int warp_reduce_sum(int x) { #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE diff --git a/ggml/src/ggml-cuda/convert.cu b/ggml/src/ggml-cuda/convert.cu index a224ec0e12..c6dec4276b 100644 --- a/ggml/src/ggml-cuda/convert.cu +++ b/ggml/src/ggml-cuda/convert.cu @@ -1,6 +1,8 @@ #include "convert.cuh" #include "dequantize.cuh" +#include + #define CUDA_Q8_0_NE_ALIGN 2048 template @@ -570,30 +572,46 @@ static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t } template -static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) { - const int64_t i = (int64_t)blockDim.x*blockIdx.x + threadIdx.x; +static __global__ void convert_unary( + const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01, const int64_t ne02, + const int64_t s01, const int64_t s02, const int64_t s03) { + const int64_t i00 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x; - if (i >= k) { + if (i00 >= ne00) { return; } + const int64_t i01 = blockIdx.y; + const int64_t i02 = blockIdx.z % ne02; + const int64_t i03 = blockIdx.z / ne02; + const src_t * x = (const src_t *) vx; - y[i] = float(x[i]); + const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00; + const int64_t iy = ((i03*ne02 + i02)*ne01 + i01)*ne00 + i00; + y[iy] = float(x[ix]); } template -static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE; - convert_unary<<>>(vx, y, k); +static void convert_unary_cuda(const void * vx, dst_t * y, + const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03, + const int64_t s01, const int64_t s02, const int64_t s03, cudaStream_t stream) { + const dim3 num_blocks((ne00 + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE, ne01, ne02*ne03); + convert_unary<<>> + (vx, y, ne00, ne01, ne02, s01, s02, s03); +} + +template +static void convert_unary_cont_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) { + convert_unary_cuda(vx, y, k, 1, 1, 1, k, k, k, stream); } to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) { switch (type) { case GGML_TYPE_F32: - return convert_unary_cuda; + return convert_unary_cont_cuda; case GGML_TYPE_F16: - return convert_unary_cuda; + return convert_unary_cont_cuda; default: return nullptr; } @@ -643,9 +661,9 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) { case GGML_TYPE_IQ3_S: return dequantize_row_iq3_s_cuda; case GGML_TYPE_F32: - return convert_unary_cuda; + return convert_unary_cont_cuda; case GGML_TYPE_BF16: - return convert_unary_cuda; + return convert_unary_cont_cuda; default: return nullptr; } @@ -692,7 +710,18 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) { case GGML_TYPE_IQ3_S: return dequantize_row_iq3_s_cuda; case GGML_TYPE_F16: - return convert_unary_cuda; + return convert_unary_cont_cuda; + case GGML_TYPE_BF16: + return convert_unary_cont_cuda; + default: + return nullptr; + } +} + +to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type) { + switch (type) { + case GGML_TYPE_F32: + return convert_unary_cuda; case GGML_TYPE_BF16: return convert_unary_cuda; default: diff --git a/ggml/src/ggml-cuda/convert.cuh b/ggml/src/ggml-cuda/convert.cuh index 411a13cf12..b65b98e08e 100644 --- a/ggml/src/ggml-cuda/convert.cuh +++ b/ggml/src/ggml-cuda/convert.cuh @@ -3,7 +3,7 @@ #define CUDA_DEQUANTIZE_BLOCK_SIZE 256 template -using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int64_t k, cudaStream_t stream); +using to_t_cuda_t = void (*)(const void * x, T * y, int64_t k, cudaStream_t stream); typedef to_t_cuda_t to_fp32_cuda_t; typedef to_t_cuda_t to_fp16_cuda_t; @@ -14,3 +14,13 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type); to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type); to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type); + +// TODO more general support for non-contiguous inputs + +template +using to_t_nc_cuda_t = void (*)(const void * x, T * y, + int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne03, + int64_t s01, int64_t s02, int64_t s03, cudaStream_t stream); + +typedef to_t_nc_cuda_t to_fp16_nc_cuda_t; +to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type); diff --git a/ggml/src/ggml-cuda/cp-async.cuh b/ggml/src/ggml-cuda/cp-async.cuh index ecb659997b..63d0c482ff 100644 --- a/ggml/src/ggml-cuda/cp-async.cuh +++ b/ggml/src/ggml-cuda/cp-async.cuh @@ -2,6 +2,17 @@ #include "common.cuh" + +static __device__ __forceinline__ unsigned int ggml_cuda_cvta_generic_to_shared(void * generic_ptr) { +#ifdef CP_ASYNC_AVAILABLE + return __cvta_generic_to_shared(generic_ptr); +#else + GGML_UNUSED(generic_ptr); + NO_DEVICE_CODE; + return 0; +#endif // CP_ASYNC_AVAILABLE +} + // Copies data from global to shared memory, cg == cache global. // Both the src and dst pointers must be aligned to 16 bit. // Shared memory uses 32 bit addressing, the pointer is passed as unsigned int. diff --git a/ggml/src/ggml-cuda/cpy.cu b/ggml/src/ggml-cuda/cpy.cu index ed25646e8e..d027271fcd 100644 --- a/ggml/src/ggml-cuda/cpy.cu +++ b/ggml/src/ggml-cuda/cpy.cu @@ -592,6 +592,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg dest_ptrs_d = ctx.cuda_graph->dest_ptrs_d; graph_cpynode_index = ctx.cuda_graph->graph_cpynode_index; } +#else + GGML_UNUSED(disable_indirection_for_this_node); #endif if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) { GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1)); @@ -639,6 +641,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) { ctx.cuda_graph->graph_cpynode_index = graph_cpynode_index; } +#else + GGML_UNUSED(disable_indirection_for_this_node); #endif } diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh index 56121705bd..b7180d5955 100644 --- a/ggml/src/ggml-cuda/fattn-common.cuh +++ b/ggml/src/ggml-cuda/fattn-common.cuh @@ -516,7 +516,7 @@ constexpr __device__ dequantize_1_f32_t get_dequantize_1_f32(ggml_type type_V) { nullptr; } -template // D == head size +template // 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 ne11) { @@ -665,13 +665,13 @@ static void on_no_fattn_vec_case(const int D) { fprintf(stderr, "Compile with GGML_CUDA_FA_ALL_QUANTS for all combinations of q4_0, q4_1, q5_0, q5_1, q8_0, and f16.\n"); GGML_ABORT("fatal error"); } else { - fprintf(stderr, "Unsupported KV type combination for head_size 256.\n"); + fprintf(stderr, "Unsupported KV type combination for head_size %d.\n", D); fprintf(stderr, "Only f16 is supported.\n"); GGML_ABORT("fatal error"); } } -template +template void launch_fattn( ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, const int nwarps, const size_t nbytes_shared, const int KQ_row_granularity, const bool need_f16_K, const bool need_f16_V, const bool stream_k, const int warp_size = WARP_SIZE @@ -691,7 +691,7 @@ void launch_fattn( GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16); GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) && - "the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big"); + "the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big"); GGML_ASSERT(K->ne[1] % FATTN_KQ_STRIDE == 0 && "Incorrect KV cache padding."); @@ -719,6 +719,7 @@ void launch_fattn( size_t nb23 = V->nb[3]; if (need_f16_K && K->type != GGML_TYPE_F16) { + GGML_ASSERT(ggml_is_contiguously_allocated(K)); K_f16.alloc(ggml_nelements(K)); to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(K->type); to_fp16(K_data, K_f16.ptr, ggml_nelements(K), main_stream); @@ -733,6 +734,7 @@ void launch_fattn( } if (need_f16_V && V->type != GGML_TYPE_F16) { + GGML_ASSERT(ggml_is_contiguously_allocated(V)); V_f16.alloc(ggml_nelements(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); @@ -752,10 +754,13 @@ void launch_fattn( const int ntiles_total = ntiles_x * (Q->ne[2] / ncols2) * Q->ne[3]; const dim3 block_dim(warp_size, nwarps, 1); + int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy. + CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared)); + dim3 blocks_num; if (stream_k) { // For short contexts it can be faster to have the SMs work on whole tiles because this lets us skip the fixup. - const int max_blocks = 2*nsm; + const int max_blocks = max_blocks_per_sm*nsm; const int tiles_nwaves = (ntiles_total + max_blocks - 1) / max_blocks; const int tiles_efficiency_percent = 100 * ntiles_total / (max_blocks*tiles_nwaves); @@ -767,14 +772,11 @@ void launch_fattn( blocks_num.y = 1; blocks_num.z = 1; - dst_tmp_meta.alloc(blocks_num.x*ncols * (2*2 + D) * sizeof(float)); + dst_tmp_meta.alloc(blocks_num.x*ncols * (2*2 + DV) * sizeof(float)); } else { GGML_ASSERT(K->ne[1] % KQ_row_granularity == 0); const int ntiles_KQ = K->ne[1] / KQ_row_granularity; // Max. number of parallel blocks limited by tensor size. - int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy. - CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared)); - // parallel_blocks should be at least large enough to achieve max. occupancy for a single wave: parallel_blocks = std::max((nsm * max_blocks_per_sm) / ntiles_total, 1); @@ -851,19 +853,19 @@ void launch_fattn( if (stream_k) { if (ntiles_total % blocks_num.x != 0) { // Fixup is only needed if the SMs work on fractional tiles. - const dim3 block_dim_combine(D, 1, 1); + const dim3 block_dim_combine(DV, 1, 1); const dim3 blocks_num_combine = {blocks_num.x, ncols1, ncols2}; - flash_attn_stream_k_fixup + flash_attn_stream_k_fixup <<>> ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], K->ne[1]); } } else if (parallel_blocks > 1) { - const dim3 block_dim_combine(D, 1, 1); + const dim3 block_dim_combine(DV, 1, 1); const dim3 blocks_num_combine(Q->ne[1], 1, blocks_num.z); const size_t nbytes_shared_combine = parallel_blocks*sizeof(float2); - flash_attn_combine_results + flash_attn_combine_results <<>> (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data, parallel_blocks); } diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh index 04804a15c9..491780abd4 100644 --- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh +++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh @@ -13,104 +13,217 @@ typedef tile<16, 16, float> tile_C_KQ_16; typedef tile<16, 4, half2> tile_C_VKQ; typedef tile<16, 8, half2> tile_C_VKQ_16; -template +// Config options for specific head sizes. +// Should not affect results, only speed/register pressure/shared memory use. +// +// nbatch_fa: number of KV rows per softmax rescaling of KQ rowsums and VKQ accumulators. +// nwarps_max: maximum number of warps per CUDA block, up to 8 warps in total can run per SM (given enough shared memory). +// Q_in_reg: whether the Q values should be kept permanently in registers. +// nstages_target: targeted number of pipeline stages for cp_async (if available), 0 means synchronous data loading. +// nbatch_K2: number of K half2 values in direction of DKQ to load in parallel. +// nbatch_V2: number of V half2 values in direction of DV to load in parallel. +// nbatch_combine: number of VKQ half2 values in direction of DV to combine in parallel. + +template +struct fattn_mma_f16_config; + +template <> +struct fattn_mma_f16_config< 64, 64> { + static constexpr int nbatch_fa = 64; + static constexpr int nwarps_max = 4; + static constexpr bool Q_in_reg = true; + static constexpr int nstages_target = 2; + static constexpr int nbatch_K2 = 32; + static constexpr int nbatch_V2 = 32; + static constexpr int nbatch_combine = 32; +}; + +template <> +struct fattn_mma_f16_config< 80, 80> { + static constexpr int nbatch_fa = 64; + static constexpr int nwarps_max = 4; + static constexpr bool Q_in_reg = true; + static constexpr int nstages_target = 2; + static constexpr int nbatch_K2 = 40; + static constexpr int nbatch_V2 = 40; + static constexpr int nbatch_combine = 40; +}; + +template <> +struct fattn_mma_f16_config< 96, 96> { + static constexpr int nbatch_fa = 64; + static constexpr int nwarps_max = 4; + static constexpr bool Q_in_reg = true; + static constexpr int nstages_target = 2; + static constexpr int nbatch_K2 = 48; + static constexpr int nbatch_V2 = 48; + static constexpr int nbatch_combine = 48; +}; + +template <> +struct fattn_mma_f16_config<112, 112> { + static constexpr int nbatch_fa = 64; + static constexpr int nwarps_max = 4; + static constexpr bool Q_in_reg = true; + static constexpr int nstages_target = 2; + static constexpr int nbatch_K2 = 56; + static constexpr int nbatch_V2 = 56; + static constexpr int nbatch_combine = 56; +}; + +template <> +struct fattn_mma_f16_config<128, 128> { + static constexpr int nbatch_fa = 64; + static constexpr int nwarps_max = 4; + static constexpr bool Q_in_reg = true; + static constexpr int nstages_target = 2; + static constexpr int nbatch_K2 = 64; + static constexpr int nbatch_V2 = 64; + static constexpr int nbatch_combine = 64; +}; + +template <> +struct fattn_mma_f16_config<256, 256> { + static constexpr int nbatch_fa = 32; + static constexpr int nwarps_max = 4; + static constexpr bool Q_in_reg = true; + static constexpr int nstages_target = 2; + static constexpr int nbatch_K2 = 128; + static constexpr int nbatch_V2 = 128; + static constexpr int nbatch_combine = 128; +}; + +template <> +struct fattn_mma_f16_config<576, 512> { + static constexpr int nbatch_fa = 32; + static constexpr int nwarps_max = 8; + static constexpr bool Q_in_reg = false; + static constexpr int nstages_target = 1; + static constexpr int nbatch_K2 = 160; + static constexpr int nbatch_V2 = 128; + static constexpr int nbatch_combine = 128; +}; + +// ------------------------------------------------------------------------------------------------------------------ + +template static __device__ __forceinline__ void flash_attn_ext_f16_load_tile( - const half2 * const __restrict__ KV, half2 * const __restrict__ tile_KV, const int stride_KV) { - constexpr int D2_padded = D/2 + 4; // Size of D in half2, padded to avoid shared memory bank conflicts. + const half2 * const __restrict__ KV, half2 * const __restrict__ tile_KV, const int D2, const int stride_KV) { - // If cp.async is available, load up to the highest power of 2 in D asynchronously: -#ifdef CP_ASYNC_AVAILABLE - static_assert(D >= 64 && D < 512, "bad D"); - constexpr int k0_sync_start = D/2 < 64 ? 32 : (D/2 < 128 ? 64 : 128); - - const unsigned int tile_KV_32 = __cvta_generic_to_shared(tile_KV); - - constexpr int preload = 64; - constexpr int h2_per_chunk = 16/sizeof(half2); - constexpr int chunks_per_row = k0_sync_start / h2_per_chunk; - constexpr int stride_i = WARP_SIZE / chunks_per_row; -#pragma unroll - for (int i0 = 0; i0 < KQ_per_iter; i0 += nwarps*stride_i) { - const int i = i0 + threadIdx.y*stride_i + (chunks_per_row == WARP_SIZE ? 0 : threadIdx.x / chunks_per_row); - const int k = (chunks_per_row == WARP_SIZE ? threadIdx.x : threadIdx.x % chunks_per_row)*h2_per_chunk; - - cp_async_cg_16(tile_KV_32 + (i*D2_padded + k)*sizeof(half2), KV + i*stride_KV + k); - } -#else - constexpr int k0_sync_start = 0; -#endif // CP_ASYNC_AVAILABLE - static_assert(k0_sync_start % WARP_SIZE == 0, "bad k0_sync_start"); - - // If D is not a power of 2, the rest is loaded synchronously. // K/V data is loaded with decreasing granularity for D for better memory bandwidth. - static_assert(KQ_per_iter % (4*nwarps) == 0, "out of bounds"); -#pragma unroll - for (int stride_k : {WARP_SIZE, WARP_SIZE/2, WARP_SIZE/4}) { - const int k0_start = stride_k == WARP_SIZE ? k0_sync_start : D/2 - (D/2) % (2*stride_k); - const int k0_stop = D/2 - (D/2) % (1*stride_k); - const int stride_i = WARP_SIZE / stride_k; + // The minimum granularity with cp.async is 16 bytes, with synchronous data loading it's 4 bytes. - if (k0_start == k0_stop || k0_stop <= k0_sync_start) { - continue; - } + if (use_cp_async) { + constexpr int preload = 64; + constexpr int h2_per_chunk = 16/sizeof(half2); + const int chunks_per_row = D2 / h2_per_chunk; -#pragma unroll - for (int i0 = 0; i0 < KQ_per_iter; i0 += nwarps*stride_i) { - const int i = i0 + threadIdx.y*stride_i + (stride_k == WARP_SIZE ? 0 : threadIdx.x / stride_k); + const unsigned int tile_KV_32 = ggml_cuda_cvta_generic_to_shared(tile_KV); -#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); + auto load = [&] __device__ (const int n) { + const int stride_k = WARP_SIZE >> n; + const int k0_start = stride_k == WARP_SIZE ? 0 : chunks_per_row - chunks_per_row % (2*stride_k); + const int k0_stop = chunks_per_row - chunks_per_row % (1*stride_k); + const int stride_i = WARP_SIZE / stride_k; - tile_KV[i*D2_padded + k] = KV[i*stride_KV + k]; + if (k0_start == k0_stop) { + return; } - } + +#pragma unroll + for (int i0 = 0; i0 < nbatch_fa; i0 += nwarps*stride_i) { + const int i = i0 + threadIdx.y*stride_i + (stride_k == WARP_SIZE ? 0 : threadIdx.x / stride_k); + + if (i0 + nwarps*stride_i > nbatch_fa && i >= nbatch_fa) { + break; + } + +#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); + + cp_async_cg_16(tile_KV_32 + i*(stride_tile*sizeof(half2)) + k*16, KV + i*stride_KV + k*h2_per_chunk); + } + } + }; + ggml_cuda_unroll<5>{}(load); + } else { + static_assert(nbatch_fa % (4*nwarps) == 0, "out of bounds"); + auto load = [&] __device__ (const int n) { + const int stride_k = WARP_SIZE >> n; + const int k0_start = stride_k == WARP_SIZE ? 0 : D2 - D2 % (2*stride_k); + const int k0_stop = D2 - D2 % (1*stride_k); + const int stride_i = WARP_SIZE / stride_k; + + if (k0_start == k0_stop) { + return; + } + +#pragma unroll + for (int i0 = 0; i0 < nbatch_fa; i0 += nwarps*stride_i) { + const int i = i0 + threadIdx.y*stride_i + (stride_k == WARP_SIZE ? 0 : threadIdx.x / stride_k); + + if (i0 + nwarps*stride_i > nbatch_fa && i >= nbatch_fa) { + break; + } + +#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); + + tile_KV[i*stride_tile + k] = KV[i*stride_KV + k]; + } + } + }; + ggml_cuda_unroll<3>{}(load); } } -template +template static __device__ __forceinline__ void flash_attn_ext_f16_load_mask( const half2 * const __restrict__ mask_h2, half2 * const __restrict__ tile_mask, const int stride_mask) { - static_assert(KQ_per_iter == 2*WARP_SIZE || KQ_per_iter == WARP_SIZE, "bad KQ_per_iter"); -#ifdef CP_ASYNC_AVAILABLE - constexpr int preload = KQ_per_iter * sizeof(half); - constexpr int cols_per_warp = 8*WARP_SIZE/KQ_per_iter; - constexpr int stride_j = nwarps * cols_per_warp; + static_assert(nbatch_fa == 2*WARP_SIZE || WARP_SIZE % nbatch_fa == 0, "bad KQ_per_iter"); - const unsigned int tile_mask_32 = __cvta_generic_to_shared(tile_mask); + if (use_cp_async) { + constexpr int preload = nbatch_fa >= 32 ? nbatch_fa * sizeof(half) : 64; + constexpr int cols_per_warp = 8*WARP_SIZE/nbatch_fa; + constexpr int stride_j = nwarps * cols_per_warp; + + const unsigned int tile_mask_32 = ggml_cuda_cvta_generic_to_shared(tile_mask); +#pragma unroll + for (int j0 = 0; j0 < ncols1; j0 += stride_j) { + const int j = j0 + threadIdx.y*cols_per_warp + + (nbatch_fa == 2*WARP_SIZE ? threadIdx.x / (WARP_SIZE/4) : threadIdx.x / (WARP_SIZE/cols_per_warp)); + + if (j0 + stride_j > ncols1 && j >= ncols1) { + break; + } + + const int i = 4 * (threadIdx.x % (nbatch_fa/8)); + + cp_async_cg_16(tile_mask_32 + j*(nbatch_fa*sizeof(half) + 16) + i*sizeof(half2), mask_h2 + j*stride_mask + i); + } + return; + } + + constexpr int cols_per_warp = 2*WARP_SIZE/nbatch_fa; + constexpr int stride_j = nwarps * cols_per_warp; #pragma unroll for (int j0 = 0; j0 < ncols1; j0 += stride_j) { - const int j = j0 + threadIdx.y*cols_per_warp + - (KQ_per_iter == 2*WARP_SIZE ? threadIdx.x / (WARP_SIZE/4) : threadIdx.x / (WARP_SIZE/8)); + const int j = j0 + threadIdx.y*cols_per_warp + (nbatch_fa == 2*WARP_SIZE ? 0 : threadIdx.x / (WARP_SIZE/cols_per_warp)); if (j0 + stride_j > ncols1 && j >= ncols1) { break; } - const int i = 4 * (KQ_per_iter == 2*WARP_SIZE ? threadIdx.x % (WARP_SIZE/4) : threadIdx.x % (WARP_SIZE/8)); + const int i = nbatch_fa == 2*WARP_SIZE ? threadIdx.x : threadIdx.x % (WARP_SIZE/cols_per_warp); - cp_async_cg_16(tile_mask_32 + j*(KQ_per_iter*sizeof(half) + 16) + i*sizeof(half2), mask_h2 + j*stride_mask + i); + tile_mask[j*(nbatch_fa/2 + 4) + i] = mask_h2[j*stride_mask + i]; } -#else - constexpr int cols_per_warp = 2*WARP_SIZE/KQ_per_iter; - constexpr int stride_j = nwarps * cols_per_warp; -#pragma unroll - for (int j0 = 0; j0 < ncols1; j0 += stride_j) { - const int j = j0 + threadIdx.y*cols_per_warp + (KQ_per_iter == 2*WARP_SIZE ? 0 : threadIdx.x / (WARP_SIZE/2)); - - if (j0 + stride_j > ncols1 && j >= ncols1) { - break; - } - - const int i = KQ_per_iter == 2*WARP_SIZE ? threadIdx.x : threadIdx.x % (WARP_SIZE/2); - - tile_mask[j*(KQ_per_iter/2 + 4) + i] = mask_h2[j*stride_mask + i]; - } -#endif // CP_ASYNC_AVAILABLE } -template +template static __device__ __forceinline__ void flash_attn_ext_f16_iter( const float2 * const __restrict__ Q_f2, const half2 * const __restrict__ K_h2, @@ -123,9 +236,11 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( const float logit_softcap, const int ne01, const int ne02, - const int stride_KV, + const int stride_K, + const int stride_V, const int stride_mask, const int jt, + half2 * const __restrict__ tile_Q, half2 * const __restrict__ tile_K, half2 * const __restrict__ tile_V, half2 * const __restrict__ tile_mask, @@ -135,59 +250,107 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( float * const __restrict__ KQ_rowsum, const int kb0) { #ifdef NEW_MMA_AVAILABLE + typedef fattn_mma_f16_config c; + +#ifdef CP_ASYNC_AVAILABLE + constexpr int nstages = c::nstages_target; +#else + constexpr int nstages = 0; +#endif // CP_ASYNC_AVAILABLE + constexpr int cols_per_warp = ntiles * tile_B::I; constexpr int cols_per_thread = ntiles == 1 ? 2 : ntiles; constexpr int np = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column. - constexpr int D2_padded = D/2 + 4; // Size of D in half2, padded to avoid shared memory bank conflicts. - const int k_VKQ_0 = kb0 * KQ_per_iter; - tile_C_KQ KQ_C[KQ_per_iter/(np*tile_C_KQ::I) * ntiles]; + constexpr int stride_tile_Q = DKQ/2 + 4; + constexpr int stride_tile_K = c::nbatch_K2 + 4; + constexpr int stride_tile_V = c::nbatch_V2 + 4; + + const int k_VKQ_0 = kb0 * c::nbatch_fa; + tile_C_KQ KQ_C[c::nbatch_fa/(np*tile_C_KQ::I) * ntiles]; // Use wide variants of tiles if ntiles >= 2. tile_B_16 * Q_B_16 = (tile_B_16 *) Q_B; tile_C_VKQ_16 * VKQ_C_16 = (tile_C_VKQ_16 *) VKQ_C; tile_C_KQ_16 * KQ_C_16 = (tile_C_KQ_16 *) KQ_C; -#ifdef CP_ASYNC_AVAILABLE - cp_async_wait_all(); - __syncthreads(); - flash_attn_ext_f16_load_tile(V_h2 + k_VKQ_0*stride_KV, tile_V, stride_KV); -#else - if (ncols2 > 1 || mask_h2) { - flash_attn_ext_f16_load_mask(mask_h2 + k_VKQ_0/2, tile_mask, stride_mask); - } - flash_attn_ext_f16_load_tile(K_h2 + k_VKQ_0*stride_KV, tile_K, stride_KV); - __syncthreads(); -#endif // CP_ASYNC_AVAILABLE - - // Calculate tile of KQ: -#pragma unroll - for (int i_KQ_00 = 0; i_KQ_00 < KQ_per_iter; i_KQ_00 += np*tile_A::I) { - const int i_KQ_0 = i_KQ_00 + (threadIdx.y % np)*tile_A::I; -#pragma unroll - for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += tile_A::J) { - tile_A K_A; - load_ldmatrix(K_A, tile_K + i_KQ_0*D2_padded + k_KQ_0, D2_padded); - if (ntiles == 1) { - mma(KQ_C[i_KQ_00/(np*tile_A::I)], K_A, Q_B[k_KQ_0/tile_A::J]); - } else { -#pragma unroll - for (int t = 0; t < ntiles/2; ++t) { - // Wide version of KQ_C is column-major => swap A and B. - mma(KQ_C_16[i_KQ_00/(np*tile_A::I) * ntiles/2 + t], Q_B_16[k_KQ_0/tile_A::J * ntiles/2 + t], K_A); - } - } + if constexpr (nstages > 1) { + static_assert(c::nbatch_K2 == DKQ/2, "batching not implemented for multi stage loading"); + constexpr bool use_cp_async = true; + cp_async_wait_all(); + __syncthreads(); + flash_attn_ext_f16_load_tile + (V_h2 + k_VKQ_0*stride_V, tile_V, c::nbatch_V2, stride_V); + } else { + constexpr bool use_cp_async = nstages == 1; + if (ncols2 > 1 || mask_h2) { + flash_attn_ext_f16_load_mask(mask_h2 + k_VKQ_0/2, tile_mask, stride_mask); } } -#ifndef CP_ASYNC_AVAILABLE - __syncthreads(); // Only needed if tile_K == tile_V. -#endif // CP_ASYNC_AVAILABLE +#pragma unroll + for (int k0_start = 0; k0_start < DKQ/2; k0_start += c::nbatch_K2) { + const int k0_stop = k0_start + c::nbatch_K2 < DKQ/2 ? k0_start + c::nbatch_K2 : DKQ/2; + const int k0_diff = k0_stop - k0_start; + + if (nstages <= 1) { + constexpr bool use_cp_async = nstages == 1; + flash_attn_ext_f16_load_tile + (K_h2 + k_VKQ_0*stride_K + k0_start, tile_K, k0_diff, stride_K); + if (use_cp_async) { + cp_async_wait_all(); + } + __syncthreads(); + } + + // Calculate tile of KQ: + if constexpr (c::Q_in_reg) { +#pragma unroll + for (int i_KQ_00 = 0; i_KQ_00 < c::nbatch_fa; i_KQ_00 += np*tile_A::I) { + const int i_KQ_0 = i_KQ_00 + (threadIdx.y % np)*tile_A::I; +#pragma unroll + for (int k_KQ_0 = k0_start; k_KQ_0 < k0_stop; k_KQ_0 += tile_A::J) { + tile_A K_A; + load_ldmatrix(K_A, tile_K + i_KQ_0*stride_tile_K + (k_KQ_0 - k0_start), stride_tile_K); + if (ntiles == 1) { + mma(KQ_C[i_KQ_00/(np*tile_A::I)], K_A, Q_B[k_KQ_0/tile_A::J]); + } else { +#pragma unroll + for (int t = 0; t < ntiles/2; ++t) { + // Wide version of KQ_C is column-major => swap A and B. + mma(KQ_C_16[i_KQ_00/(np*tile_A::I) * ntiles/2 + t], Q_B_16[k_KQ_0/tile_A::J * ntiles/2 + t], K_A); + } + } + } + } + } else { + static_assert(ntiles == 2, "ntiles != 2 not implemented"); +#pragma unroll + for (int k_KQ_0 = k0_start; k_KQ_0 < k0_stop; k_KQ_0 += tile_A::J) { + load_ldmatrix(Q_B_16[0], tile_Q + (threadIdx.y / np)*(tile_B_16::I*stride_tile_Q) + k_KQ_0, stride_tile_Q); + +#pragma unroll + for (int i_KQ_00 = 0; i_KQ_00 < c::nbatch_fa; i_KQ_00 += np*tile_A::I) { + const int i_KQ_0 = i_KQ_00 + (threadIdx.y % np)*tile_A::I; + + tile_A 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 => swap A and B. + mma(KQ_C_16[i_KQ_00/(np*tile_A::I)], Q_B_16[0], K_A); + } + } + } + + if (nstages <= 1) { + __syncthreads(); // Only needed if tile_K == tile_V. + } + } if (use_logit_softcap) { - static_assert(KQ_per_iter % (np*tile_C_KQ::I) == 0, "bad loop size"); + static_assert(c::nbatch_fa % (np*tile_C_KQ::I) == 0, "bad loop size"); #pragma unroll - for (int i = 0; i < KQ_per_iter/(np*tile_C_KQ::I) * ntiles; ++i) { + for (int i = 0; i < c::nbatch_fa/(np*tile_C_KQ::I) * ntiles; ++i) { #pragma unroll for (int l = 0; l < tile_C_KQ::ne; ++l) { KQ_C[i].x[l] = logit_softcap*tanhf(KQ_C[i].x[l]); @@ -205,7 +368,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( if (ntiles == 1) { if (ncols2 > 1 || mask_h2) { #pragma unroll - for (int i00 = 0; i00 < KQ_per_iter; i00 += np*tile_C_KQ::I) { + for (int i00 = 0; i00 < c::nbatch_fa; i00 += np*tile_C_KQ::I) { const int i0 = i00 + (threadIdx.y % np)*tile_C_KQ::I; #pragma unroll for (int l = 0; l < tile_C_KQ::ne; ++l) { @@ -213,16 +376,16 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( const int j = ((threadIdx.y / np)*tile_C_KQ::J + tile_C_KQ::get_j(l)) / ncols2; KQ_C[i00/(np*tile_C_KQ::I)].x[l] += slope * - __half2float(((const half *) tile_mask)[j*(KQ_per_iter + 8) + i]); + __half2float(((const half *) tile_mask)[j*(c::nbatch_fa + 8) + i]); } } } // Calculate softmax for each KQ column using the current max. value. // The divisor is stored in KQ_rowsum and will be applied at the end. - static_assert(KQ_per_iter % (np*tile_C_KQ::I) == 0, "bad loop size"); + static_assert(c::nbatch_fa % (np*tile_C_KQ::I) == 0, "bad loop size"); #pragma unroll - for (int k = 0; k < KQ_per_iter/(np*tile_C_KQ::I); ++k) { + for (int k = 0; k < c::nbatch_fa/(np*tile_C_KQ::I); ++k) { #pragma unroll for (int l = 0; l < tile_C_KQ::ne; ++l) { KQ_max_new[l % 2] = fmaxf(KQ_max_new[l % 2], KQ_C[k].x[l]); @@ -238,10 +401,9 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( } } - static_assert(KQ_per_iter % (np*tile_C_KQ::I) == 0, "bad loop size"); - + static_assert(c::nbatch_fa % (np*tile_C_KQ::I) == 0, "bad loop size"); #pragma unroll - for (int k = 0; k < KQ_per_iter/(np*tile_C_KQ::I); ++k) { + for (int k = 0; k < c::nbatch_fa/(np*tile_C_KQ::I); ++k) { #pragma unroll for (int l = 0; l < tile_C_KQ::ne; ++l) { KQ_C[k].x[l] = expf(KQ_C[k].x[l] - KQ_max_new[l % 2]); @@ -252,7 +414,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( } else { // ntiles > 1 if (ncols2 > 1 || mask_h2) { #pragma unroll - for (int i00 = 0; i00 < KQ_per_iter; i00 += np*tile_C_KQ_16::J) { + for (int i00 = 0; i00 < c::nbatch_fa; i00 += np*tile_C_KQ_16::J) { const int i0 = i00 + (threadIdx.y % np)*tile_C_KQ_16::J; #pragma unroll for (int t = 0; t < ntiles/2; ++t) { @@ -261,7 +423,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( const int i = (i0 + tile_C_KQ_16::get_j(l0)) / 2; const int j = ((threadIdx.y / np)*cols_per_warp + t*tile_C_KQ_16::I + tile_C_KQ_16::get_i(l0)) / ncols2; - const float2 tmp = __half22float2(tile_mask[j*(KQ_per_iter/2 + 4) + i]); + const float2 tmp = __half22float2(tile_mask[j*(c::nbatch_fa/2 + 4) + i]); const int KQ_index = i00/(np*tile_C_KQ_16::J) * ntiles/2 + t; KQ_C_16[KQ_index].x[l0 + 0] += slope*tmp.x; KQ_C_16[KQ_index].x[l0 + 1] += slope*tmp.y; @@ -272,9 +434,9 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( // Calculate softmax for each KQ column using the current max. value. // The divisor is stored in KQ_rowsum and will be applied at the end. - static_assert(KQ_per_iter % (np*tile_C_KQ::I) == 0, "bad loop size"); + static_assert(c::nbatch_fa % (np*tile_C_KQ::I) == 0, "bad loop size"); #pragma unroll - for (int k = 0; k < KQ_per_iter/(np*tile_C_KQ_16::J); ++k) { + for (int k = 0; k < c::nbatch_fa/(np*tile_C_KQ_16::J); ++k) { #pragma unroll for (int t = 0; t < ntiles/2; ++t) { #pragma unroll @@ -294,9 +456,9 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( } } - static_assert(KQ_per_iter % (np*tile_C_KQ_16::J) == 0, "bad loop size"); + static_assert(c::nbatch_fa % (np*tile_C_KQ_16::J) == 0, "bad loop size"); #pragma unroll - for (int k = 0; k < KQ_per_iter/(np*tile_C_KQ_16::J); ++k) { + for (int k = 0; k < c::nbatch_fa/(np*tile_C_KQ_16::J); ++k) { #pragma unroll for (int t = 0; t < ntiles/2; ++t) { #pragma unroll @@ -325,7 +487,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( if (ntiles == 1) { const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[1]); #pragma unroll - for (int i = 0; i < D/tile_C_VKQ::I; ++i) { + for (int i = 0; i < DV/tile_C_VKQ::I; ++i) { #pragma unroll for (int l = 0; l < tile_C_VKQ::ne; ++l) { VKQ_C[i].x[l] *= KQ_max_scale_h2; @@ -336,7 +498,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( for (int col = 0; col < cols_per_thread; ++col) { const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[col], KQ_max_scale[col]); #pragma unroll - for (int i = 0; i < D/tile_C_VKQ_16::J; ++i) { + for (int i = 0; i < DV/tile_C_VKQ_16::J; ++i) { #pragma unroll for (int l0 = 0; l0 < tile_C_VKQ_16::ne; l0 += 2) { VKQ_C_16[i*ntiles/2 + col/2].x[l0 + col % 2] *= KQ_max_scale_h2; @@ -347,16 +509,16 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( } // Convert KQ C tiles into B tiles for VKQ calculation: - tile_B B[KQ_per_iter/(np*2*tile_B::J) * ntiles]; + tile_B B[c::nbatch_fa/(np*2*tile_B::J) * ntiles]; tile_B_16 * B_16 = (tile_B_16 *) B; - static_assert(KQ_per_iter % (np*2*tile_B::J) == 0, "bad loop size"); + static_assert(c::nbatch_fa % (np*2*tile_B::J) == 0, "bad loop size"); if (ntiles == 1) { #pragma unroll - for (int k = 0; k < KQ_per_iter/(np*2*tile_B::J); ++k) { + for (int k = 0; k < c::nbatch_fa/(np*2*tile_B::J); ++k) { B[k] = get_transposed(get_half2(KQ_C[k])); } } else { - for (int k = 0; k < KQ_per_iter/(np*2*tile_B_16::J); ++k) { + for (int k = 0; k < c::nbatch_fa/(np*2*tile_B_16::J); ++k) { #pragma unroll for (int t = 0; t < ntiles/2; ++t) { B_16[k*ntiles/2 + t] = get_half2(KQ_C_16[k*ntiles/2 + t]); @@ -364,52 +526,67 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( } } -#ifdef CP_ASYNC_AVAILABLE - // Preload K tile for next iteration: - cp_async_wait_all(); - __syncthreads(); - if (!last_iter) { - if (ncols2 > 1 || mask_h2) { - flash_attn_ext_f16_load_mask(mask_h2 + (k_VKQ_0 + KQ_per_iter)/2, tile_mask, stride_mask); + if (nstages > 1) { + // Preload K tile for next iteration: + constexpr bool use_cp_async = true; + cp_async_wait_all(); + __syncthreads(); + if (!last_iter) { + if (ncols2 > 1 || mask_h2) { + flash_attn_ext_f16_load_mask + (mask_h2 + (k_VKQ_0 + c::nbatch_fa)/2, tile_mask, stride_mask); + } + flash_attn_ext_f16_load_tile + (K_h2 + (k_VKQ_0 + c::nbatch_fa)*stride_K, tile_K, c::nbatch_K2, stride_K); } - flash_attn_ext_f16_load_tile(K_h2 + (k_VKQ_0 + KQ_per_iter)*stride_KV, tile_K, stride_KV); } -#else - flash_attn_ext_f16_load_tile(V_h2 + k_VKQ_0*stride_KV, tile_V, stride_KV); - __syncthreads(); -#endif // CP_ASYNC_AVAILABLE - // Calculate VKQ tile: #pragma unroll - for (int i_VKQ_0 = 0; i_VKQ_0 < D; i_VKQ_0 += tile_C_VKQ::I) { - static_assert((KQ_per_iter/2) % (np*tile_A::J) == 0, "bad loop size"); -#pragma unroll - for (int k00 = 0; k00 < KQ_per_iter/2; k00 += np*tile_A::J) { - const int k0 = k00 + (threadIdx.y % np)*tile_A::J; + for (int i0_start = 0; i0_start < DV; i0_start += 2*c::nbatch_V2) { + const int i0_stop = i0_start + 2*c::nbatch_V2 < DV ? i0_start + 2*c::nbatch_V2 : DV; + const int i0_diff = i0_stop - i0_start; - tile_A A; - load_ldmatrix_trans(A, tile_V + 2*k0*D2_padded + i_VKQ_0/2, D2_padded); - if (ntiles == 1) { - mma(VKQ_C[i_VKQ_0/tile_C_VKQ::I], A, B[k00/(np*tile_A::J)]); - } else { + if (nstages <= 1) { + constexpr bool use_cp_async = nstages == 1; + flash_attn_ext_f16_load_tile + (V_h2 + k_VKQ_0*stride_V + i0_start/2, tile_V, i0_diff/2, stride_V); + if (use_cp_async) { + cp_async_wait_all(); + } + __syncthreads(); + } + + // Calculate VKQ tile: #pragma unroll - for (int t = 0; t < ntiles/2; ++t) { - // Wide version of VKQ_C is column-major => swap A and B. - mma(VKQ_C_16[i_VKQ_0/tile_C_VKQ::I * ntiles/2 + t], B_16[k00/(np*tile_A::J) * ntiles/2 + t], A); + for (int i_VKQ_0 = i0_start; i_VKQ_0 < i0_stop; i_VKQ_0 += tile_C_VKQ::I) { + static_assert((c::nbatch_fa/2) % (np*tile_A::J) == 0, "bad loop size"); +#pragma unroll + for (int k00 = 0; k00 < c::nbatch_fa/2; k00 += np*tile_A::J) { + const int k0 = k00 + (threadIdx.y % np)*tile_A::J; + + tile_A A; + load_ldmatrix_trans(A, tile_V + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V); + if (ntiles == 1) { + mma(VKQ_C[i_VKQ_0/tile_C_VKQ::I], A, B[k00/(np*tile_A::J)]); + } else { +#pragma unroll + for (int t = 0; t < ntiles/2; ++t) { + // Wide version of VKQ_C is column-major => swap A and B. + mma(VKQ_C_16[i_VKQ_0/tile_C_VKQ::I * ntiles/2 + t], B_16[k00/(np*tile_A::J) * ntiles/2 + t], A); + } } } } + + if (nstages <= 1) { + __syncthreads(); // Only needed if tile_K == tile_V. + } } - -#ifndef CP_ASYNC_AVAILABLE - __syncthreads(); // Only needed if tile_K == tile_V. -#endif // CP_ASYNC_AVAILABLE - #else GGML_UNUSED(Q_f2); GGML_UNUSED(K_h2); GGML_UNUSED(V_h2); GGML_UNUSED(mask_h2); GGML_UNUSED(dstk); GGML_UNUSED(dstk_fixup); GGML_UNUSED(scale); GGML_UNUSED(slope); GGML_UNUSED(logit_softcap); - GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(stride_KV); + GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(stride_K); GGML_UNUSED(stride_V); GGML_UNUSED(stride_mask); GGML_UNUSED(jt); GGML_UNUSED(tile_K); GGML_UNUSED(stride_mask); GGML_UNUSED(jt); GGML_UNUSED(tile_K); GGML_UNUSED(tile_V); GGML_UNUSED(tile_mask); GGML_UNUSED(Q_B); @@ -419,7 +596,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( #endif // NEW_MMA_AVAILABLE } -template +template static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( const float2 * const __restrict__ Q_f2, const half2 * const __restrict__ K_h2, @@ -434,7 +611,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( const int ne02, const int stride_Q1, const int stride_Q2, - const int stride_KV, + const int stride_K, + const int stride_V, const int stride_mask, const int jt, const int kb0_start, @@ -442,6 +620,14 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( #ifdef NEW_MMA_AVAILABLE //In this kernel Q, K, V are matrices while i, j, k are matrix indices. + typedef fattn_mma_f16_config c; + +#ifdef CP_ASYNC_AVAILABLE + constexpr int nstages = c::nstages_target; +#else + constexpr int nstages = 0; +#endif // CP_ASYNC_AVAILABLE + constexpr int ncols = ncols1 * ncols2; constexpr int cols_per_warp = ntiles * tile_B::I; constexpr int cols_per_thread = ntiles == 1 ? 2 : ntiles; @@ -449,22 +635,19 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( static_assert(nwarps * (cols_per_warp/ncols2) % ncols1 == 0, "bad nwarps"); - static_assert(D % nwarps == 0, "bad D"); - static_assert(KQ_per_iter % nwarps == 0, "bad KQ_per_iter"); + constexpr int stride_tile_Q = DKQ/2 + 4; + constexpr int stride_tile_K = c::nbatch_K2 + 4; + constexpr int stride_tile_V = c::nbatch_V2 + 4; - constexpr int D2_padded = D/2 + 4; // Size of D in half2, padded to avoid shared memory bank conflicts. + constexpr int stride_tile_KV_max = stride_tile_K > stride_tile_V ? stride_tile_K : stride_tile_V; - // Temporary shared buffer for loading K/V data with KQ_per_iter*D logical elements: - extern __shared__ half2 tile_K[]; -#ifdef CP_ASYNC_AVAILABLE - half2 * tile_V = tile_K + KQ_per_iter*D2_padded; -#else - half2 * tile_V = tile_K; -#endif // CP_ASYNC_AVAILABLE - half2 * tile_mask = tile_V + KQ_per_iter*D2_padded; + extern __shared__ half2 tile_Q[]; + half2 * tile_K = c::Q_in_reg ? tile_Q : tile_Q + ncols * stride_tile_Q; + half2 * tile_V = nstages > 1 ? tile_K + c::nbatch_fa * stride_tile_K : tile_K; + half2 * tile_mask = nstages > 1 ? tile_V + c::nbatch_fa * stride_tile_V : tile_V + c::nbatch_fa * stride_tile_KV_max; - tile_B Q_B[D/(2*tile_B::J) * ntiles]; - tile_C_VKQ VKQ_C[D/tile_C_VKQ::I * ntiles]; + tile_B Q_B[(c::Q_in_reg ? DKQ/(2*tile_B::J) : 1) * ntiles]; + tile_C_VKQ VKQ_C[DV/tile_C_VKQ::I * ntiles]; tile_B_16 * Q_B_16 = (tile_B_16 *) Q_B; tile_C_VKQ_16 * VKQ_C_16 = (tile_C_VKQ_16 *) VKQ_C; @@ -476,13 +659,14 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( KQ_max[col] = -FLT_MAX/2.0f; } - // Temporarily load Q data into tile_K, will be loaded into registers afterwards. + // Load Q data into tile_Q, either temporarily or permanently. + // Q in registers is faster, but register pressure is the biggest bottleneck. // The loading is done with decreasing granularity for D for better memory bandwidth. const half2 scale_h2 = make_half2(scale, scale); #pragma unroll for (int stride_k : {WARP_SIZE, WARP_SIZE/2, WARP_SIZE/4}) { - const int k0_start = stride_k == WARP_SIZE ? 0 : D/2 - (D/2) % (2*stride_k); - const int k0_stop = D/2 - (D/2) % (1*stride_k); + const int k0_start = stride_k == WARP_SIZE ? 0 : DKQ/2 - (DKQ/2) % (2*stride_k); + const int k0_stop = DKQ/2 - (DKQ/2) % (1*stride_k); const int stride_jc = WARP_SIZE / stride_k; if (k0_start == k0_stop) { @@ -506,14 +690,14 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( const int k = k0 + (stride_k == WARP_SIZE ? threadIdx.x : threadIdx.x % stride_k); const float2 tmp = Q_f2[(jt*ncols1 + j)*stride_Q1 + c*stride_Q2 + k]; - tile_K[jc*D2_padded + k] = scale_h2 * make_half2(tmp.x, tmp.y); + tile_Q[jc*stride_tile_Q + k] = scale_h2 * make_half2(tmp.x, tmp.y); } } else { #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); - tile_K[jc*D2_padded + k] = make_half2(0.0f, 0.0f); + tile_Q[jc*stride_tile_Q + k] = make_half2(0.0f, 0.0f); } } } @@ -521,18 +705,18 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( __syncthreads(); - { + if (c::Q_in_reg) { const int j0 = (threadIdx.y / np) * cols_per_warp; #pragma unroll - for (int k0 = 0; k0 < D/2; k0 += tile_B::J) { + for (int k0 = 0; k0 < DKQ/2; k0 += tile_B::J) { if (ntiles == 1) { - load_ldmatrix(Q_B[k0/tile_B::J], tile_K + j0*D2_padded + k0, D2_padded); + load_ldmatrix(Q_B[k0/tile_B::J], tile_Q + j0*stride_tile_Q + k0, stride_tile_Q); } else { #pragma unroll for (int t = 0; t < ntiles/2; ++t) { load_ldmatrix(Q_B_16[k0/tile_B_16::J * ntiles/2 + t], - tile_K + (j0 + t*tile_B_16::I)*D2_padded + k0, D2_padded); + tile_Q + (j0 + t*tile_B_16::I)*stride_tile_Q + k0, stride_tile_Q); } } } @@ -540,35 +724,37 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( __syncthreads(); - // Preload mask and K data for first iteration when using cp_async: -#ifdef CP_ASYNC_AVAILABLE - if (ncols2 > 1 || mask_h2) { - flash_attn_ext_f16_load_mask(mask_h2 + kb0_start*KQ_per_iter/2, tile_mask, stride_mask); + // Preload mask and K data for first iteration when using cp_async with multiple stages: + if constexpr (nstages > 1) { + static_assert(c::nbatch_K2 == DKQ/2, "batching not implemented for multi-stage pipeline"); + constexpr bool use_cp_async = true; + if (ncols2 > 1 || mask_h2) { + flash_attn_ext_f16_load_mask + (mask_h2 + kb0_start*c::nbatch_fa/2, tile_mask, stride_mask); + } + flash_attn_ext_f16_load_tile + (K_h2 + kb0_start*c::nbatch_fa*stride_K, tile_K, c::nbatch_K2, stride_K); } - flash_attn_ext_f16_load_tile(K_h2 + kb0_start*KQ_per_iter*stride_KV, tile_K, stride_KV); -#endif // CP_ASYNC_AVAILABLE // Iterate over ne11 == previous tokens: for (int kb0 = kb0_start; kb0 < kb0_stop-1; ++kb0) { constexpr bool last_iter = false; - flash_attn_ext_f16_iter + flash_attn_ext_f16_iter (Q_f2, K_h2, V_h2, mask_h2, dstk, dstk_fixup, scale, slope, logit_softcap, - ne01, ne02, stride_KV, stride_mask, jt, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0); + ne01, ne02, stride_K, stride_V, stride_mask, jt, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0); } { // kb0_start is always < kb0_stop so the last iter can be executed unconditionally. constexpr bool last_iter = true; - flash_attn_ext_f16_iter + flash_attn_ext_f16_iter (Q_f2, K_h2, V_h2, mask_h2, dstk, dstk_fixup, scale, slope, logit_softcap, - ne01, ne02, stride_KV, stride_mask, jt, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0_stop-1); + ne01, ne02, stride_K, stride_V, stride_mask, jt, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0_stop-1); } - // With cp_async there is no __syncthreads at the end of the iter, + // With multi-stage loading there is no __syncthreads at the end of the iter, // there can be a race condition on shared memory access for combining/writing back results. -#ifdef CP_ASYNC_AVAILABLE - if (nwarps*cols_per_warp > KQ_per_iter) { + if (nstages > 1 && nwarps*cols_per_warp > c::nbatch_fa) { __syncthreads(); } -#endif // CP_ASYNC_AVAILABLE // Finally, sum up partial KQ rowsums. // The partial sums are spread across 8/4 threads each, does not need full reduce. @@ -584,38 +770,13 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( } } - // Write VKQ accumulators to shared memory in column-major format. - // It's faster to do small writes to shared memory, then large write to VRAM than to do small writes to VRAM. - // Also for np > 1 the combination is done via these values in shared memory. - if (ntiles == 1) { - const int jc_cwd = threadIdx.y*tile_B::I + tile_B::get_i(-1); // jc combine write data -#pragma unroll - for (int k0 = 0; k0 < D/2; k0 += tile_B::J) { - const tile_B B = get_transposed(VKQ_C[k0/tile_B::J]); // Conversion of C to B matrix puts it in column-major format. + // Combine VKQ accumulator values if np > 1. + // It's also faster to do small writes to shared memory, then large write to VRAM than to do small writes to VRAM. + // So also write VKQ accumulators to shared memory in column-major format if np == 1. -#pragma unroll - for (int l = 0; l < tile_B::ne; ++l) { - const int k = k0 + tile_B::get_j(l); - - tile_K[jc_cwd*D2_padded + k] = B.x[l]; - } - } - } else { -#pragma unroll - for (int t = 0; t < ntiles/2; ++t) { - const int j0 = threadIdx.y*cols_per_warp + t*tile_C_VKQ_16::I; -#pragma unroll - for (int k0 = 0; k0 < D/2; k0 += tile_C_VKQ_16::J) { -#pragma unroll - for (int l = 0; l < tile_C_VKQ_16::ne; ++l) { - const int j = j0 + tile_C_VKQ_16::get_i(l); - const int k = k0 + tile_C_VKQ_16::get_j(l); - - tile_K[j*D2_padded + k] = VKQ_C_16[k0/tile_C_VKQ_16::J * ntiles/2 + t].x[l]; - } - } - } - } + constexpr int nbatch_combine = c::Q_in_reg ? DV/2 : DV/4; + constexpr int tile_stride = nbatch_combine + 4; + static_assert((DV/2) % nbatch_combine == 0, "bad nbatch_combine"); if constexpr (ntiles == 1) { const int jc_cwmo = (threadIdx.x % (2*tile_C_VKQ::J)) / tile_C_VKQ::J; // jc combine write meta offset @@ -624,7 +785,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( if (((!needs_fixup && !is_fixup) || np > 1) && threadIdx.x < 2*tile_C_VKQ::J) { // Use the 16 bytes of padding in each row to store the meta data: KQ max, KQ rowsum, KQ max scale. - ((float2 *) tile_K)[jc_cwm*(D2_padded/2) + D/4] = KQ_cmr; + ((float2 *) tile_Q)[jc_cwm*(tile_stride/2) + nbatch_combine/2] = KQ_cmr; } __syncthreads(); @@ -649,7 +810,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( if (((!needs_fixup && !is_fixup) || np > 1) && (ntiles == 4 || threadIdx.x % 4 < cols_per_thread)) { // Use the 16 bytes of padding in each row to store the meta data: KQ max, KQ rowsum, KQ max scale. - ((float2 *) tile_K)[jc_cwm*(D2_padded/2) + D/4] = KQ_cmr; + ((float2 *) tile_Q)[jc_cwm*(tile_stride/2) + nbatch_combine/2] = KQ_cmr; } __syncthreads(); @@ -676,11 +837,11 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( constexpr int nmeta = np*cols_per_warp >= WARP_SIZE ? np*cols_per_warp/WARP_SIZE : 1; const int jc_meta = threadIdx.y*cols_per_warp + (np*cols_per_warp < WARP_SIZE ? threadIdx.x % (np*cols_per_warp) : threadIdx.x); - float2 * const meta_ptr = ((float2 *) tile_K) + jc_meta*(D2_padded/2) + D/4; + float2 * const meta_ptr = ((float2 *) tile_Q) + jc_meta*(tile_stride/2) + nbatch_combine/2; float2 meta[nmeta]; #pragma unroll for (int imeta = 0; imeta < nmeta; ++imeta) { - meta[imeta] = meta_ptr[imeta * WARP_SIZE * D2_padded/2]; + meta[imeta] = meta_ptr[imeta * WARP_SIZE * tile_stride/2]; } float KQ_cmn = meta[0].x; // KQ combine max new, max between all parallel warps. @@ -690,10 +851,9 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( } #pragma unroll for (int offset = np*cols_per_warp/2; offset >= cols_per_warp; offset >>= 1) { - if (offset >= WARP_SIZE) { - continue; + if (offset < WARP_SIZE) { + KQ_cmn = fmaxf(KQ_cmn, __shfl_xor_sync(0xFFFFFFFF, KQ_cmn, offset, WARP_SIZE)); } - KQ_cmn = fmaxf(KQ_cmn, __shfl_xor_sync(0xFFFFFFFF, KQ_cmn, offset, WARP_SIZE)); } float KQ_cms[nmeta]; // KQ combine max scale per warp. @@ -709,18 +869,19 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( } #pragma unroll for (int offset = np*cols_per_warp/2; offset >= cols_per_warp; offset >>= 1) { - if (offset >= WARP_SIZE) { - continue; + if (offset < WARP_SIZE) { + KQ_crs += __shfl_xor_sync(0xFFFFFFFF, KQ_crs, offset, WARP_SIZE); } - KQ_crs += __shfl_xor_sync(0xFFFFFFFF, KQ_crs, offset, WARP_SIZE); } + __syncthreads(); + // Write back combined meta data: #pragma unroll for (int imeta = 0; imeta < nmeta; ++imeta) { if (np*cols_per_warp >= WARP_SIZE || threadIdx.x < np*cols_per_warp) { // Combined KQ max scale + rowsum. - meta_ptr[imeta * WARP_SIZE * D2_padded/2] = make_float2(KQ_cms[imeta], KQ_crs); + meta_ptr[imeta * WARP_SIZE * tile_stride/2] = make_float2(KQ_cms[imeta], KQ_crs); } } @@ -734,90 +895,125 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( float2 * dstk_fixup_meta = dstk_fixup + (gridDim.x + blockIdx.x)*ncols; dstk_fixup_meta[(threadIdx.y/np)*cols_per_warp + threadIdx.x] = make_float2(KQ_cmn, KQ_crs); } - } - - if (np > 1) { + } else if (np > 1) { + // Warps with threadIdx.y % np == 0 execute a __syncthreads() in the if branch. + // Therefore, all other warps also need to execute a __syncthreads(). + // Otherwise the points at which warps synchronize with each other would become misaligned. __syncthreads(); } - if (np == 1 || threadIdx.y % np == 0) { - // The first 2*2*gridDim.x*ncols floats in dstk_fixup are for storing max. values and row sums. - // The values after that are for the partial results of the individual blocks. - float2 * dstk_fixup_data = dstk_fixup + gridDim.x*(2*ncols) + blockIdx.x*(ncols*(D/2)); +#pragma unroll + for (int k00 = 0; k00 < DV/2; k00 += nbatch_combine) { + if (ntiles == 1) { + const int jc_cwd = threadIdx.y*tile_B::I + tile_B::get_i(-1); // jc combine write data +#pragma unroll + for (int k0 = 0; k0 < nbatch_combine; k0 += tile_B::J) { + const tile_B B = get_transposed(VKQ_C[(k00 + k0)/tile_B::J]); // Conversion of C to B matrix puts it in column-major format. #pragma unroll - for (int stride_k : {WARP_SIZE, WARP_SIZE/2, WARP_SIZE/4}) { - const int k0_start = stride_k == WARP_SIZE ? 0 : D/2 - (D/2) % (2*stride_k); - const int k0_stop = D/2 - (D/2) % (1*stride_k); - const int stride_jc = WARP_SIZE / stride_k; + for (int l = 0; l < tile_B::ne; ++l) { + const int k = k0 + tile_B::get_j(l); - if (k0_start == k0_stop) { - continue; + tile_Q[jc_cwd*tile_stride + k] = B.x[l]; + } } - + } else { #pragma unroll - for (int jc0_dst = 0; jc0_dst < ncols; jc0_dst += (nwarps/np)*stride_jc) { - const int jc_dst = jc0_dst + (threadIdx.y/np)*stride_jc + (stride_k == WARP_SIZE ? 0 : threadIdx.x / stride_k); - - if (jc0_dst + (nwarps/np)*stride_jc > ncols && jc_dst >= ncols) { - break; - } - - const int jc_tile_K = (jc_dst/cols_per_warp)*(np*cols_per_warp) + jc_dst % cols_per_warp; - - const int j_dst = jc_dst / ncols2; - const int c_dst = jc_dst % ncols2; - - if (!is_fixup && jt*ncols1 + j_dst >= ne01) { - continue; - } - - const float * meta_j = (const float *) tile_K + jc_tile_K*D2_padded + D/2; + for (int t = 0; t < ntiles/2; ++t) { + const int j0 = threadIdx.y*cols_per_warp + t*tile_C_VKQ_16::I; #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); - - float2 dstk_val = make_float2(0.0f, 0.0f); + for (int k0 = 0; k0 < nbatch_combine; k0 += tile_C_VKQ_16::J) { #pragma unroll - for (int ip = 0; ip < np; ++ip) { - const float KQ_crs = np == 1 ? 1.0f : meta_j[ip*cols_per_warp * D2_padded + 0]; - const float2 dstk_val_add = __half22float2(tile_K[(jc_tile_K + ip*cols_per_warp) * D2_padded + k]); - dstk_val.x += dstk_val_add.x*KQ_crs; - dstk_val.y += dstk_val_add.y*KQ_crs; - } + for (int l = 0; l < tile_C_VKQ_16::ne; ++l) { + const int j = j0 + tile_C_VKQ_16::get_i(l); + const int k = k0 + tile_C_VKQ_16::get_j(l); - if (!needs_fixup && !is_fixup) { - const float KQ_rowsum_j = meta_j[1]; - dstk_val.x /= KQ_rowsum_j; - dstk_val.y /= KQ_rowsum_j; - } - - if (is_fixup) { - dstk_fixup_data[jc_dst*(D/2) + k] = dstk_val; - } else { - dstk[((jt*ncols1 + j_dst)*ne02 + c_dst)*(D/2) + k] = dstk_val; + tile_Q[j*tile_stride + k] = VKQ_C_16[(k00 + k0)/tile_C_VKQ_16::J * ntiles/2 + t].x[l]; } } } } - } - if (np > 1) { __syncthreads(); + + if (np == 1 || threadIdx.y % np == 0) { + // The first 2*2*gridDim.x*ncols floats in dstk_fixup are for storing max. values and row sums. + // The values after that are for the partial results of the individual blocks. + float2 * dstk_fixup_data = dstk_fixup + gridDim.x*(2*ncols) + blockIdx.x*(ncols*(DV/2)); + +#pragma unroll + for (int stride_k : {WARP_SIZE, WARP_SIZE/2, WARP_SIZE/4}) { + const int k0_start = stride_k == WARP_SIZE ? 0 : nbatch_combine - nbatch_combine % (2*stride_k); + const int k0_stop = nbatch_combine - nbatch_combine % (1*stride_k); + const int stride_jc = WARP_SIZE / stride_k; + + if (k0_start == k0_stop) { + continue; + } + +#pragma unroll + for (int jc0_dst = 0; jc0_dst < ncols; jc0_dst += (nwarps/np)*stride_jc) { + const int jc_dst = jc0_dst + (threadIdx.y/np)*stride_jc + (stride_k == WARP_SIZE ? 0 : threadIdx.x / stride_k); + + if (jc0_dst + (nwarps/np)*stride_jc > ncols && jc_dst >= ncols) { + break; + } + + const int jc_tile_K = (jc_dst/cols_per_warp)*(np*cols_per_warp) + jc_dst % cols_per_warp; + + const int j_dst = jc_dst / ncols2; + const int c_dst = jc_dst % ncols2; + + if (!is_fixup && jt*ncols1 + j_dst >= ne01) { + continue; + } + + const float * meta_j = (const float *) tile_Q + jc_tile_K*tile_stride + nbatch_combine; +#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); + + float2 dstk_val = make_float2(0.0f, 0.0f); +#pragma unroll + for (int ip = 0; ip < np; ++ip) { + const float KQ_crs = np == 1 ? 1.0f : meta_j[ip*cols_per_warp * tile_stride + 0]; + const float2 dstk_val_add = __half22float2(tile_Q[(jc_tile_K + ip*cols_per_warp) * tile_stride + k]); + dstk_val.x += dstk_val_add.x*KQ_crs; + dstk_val.y += dstk_val_add.y*KQ_crs; + } + + if (!needs_fixup && !is_fixup) { + const float KQ_rowsum_j = meta_j[1]; + dstk_val.x /= KQ_rowsum_j; + dstk_val.y /= KQ_rowsum_j; + } + + if (is_fixup) { + dstk_fixup_data[jc_dst*(DV/2) + k00 + k] = dstk_val; + } else { + dstk[((jt*ncols1 + j_dst)*ne02 + c_dst)*(DV/2) + k00 + k] = dstk_val; + } + } + } + } + } + if (np > 1) { + __syncthreads(); + } } #else GGML_UNUSED(Q_f2); GGML_UNUSED(K_h2); GGML_UNUSED(V_h2); GGML_UNUSED(mask_h2); GGML_UNUSED(dstk); GGML_UNUSED(dstk_fixup); GGML_UNUSED(scale); GGML_UNUSED(slope); GGML_UNUSED(logit_softcap); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(stride_Q1); - GGML_UNUSED(stride_Q2); GGML_UNUSED(stride_KV); GGML_UNUSED(stride_mask); + GGML_UNUSED(stride_Q2); GGML_UNUSED(stride_K); GGML_UNUSED(stride_V); GGML_UNUSED(stride_mask); GGML_UNUSED(jt); GGML_UNUSED(kb0_start); GGML_UNUSED(kb0_stop); NO_DEVICE_CODE; #endif // NEW_MMA_AVAILABLE } -template -__launch_bounds__(nwarps*WARP_SIZE, 2) +template +__launch_bounds__(nwarps*WARP_SIZE, 1) static __global__ void flash_attn_ext_f16( const char * __restrict__ Q, const char * __restrict__ K, @@ -857,24 +1053,27 @@ static __global__ void flash_attn_ext_f16( #if defined(FLASH_ATTN_AVAILABLE) && defined(NEW_MMA_AVAILABLE) // Skip unused kernel variants for faster compilation: - if (use_logit_softcap && !(D == 128 || D == 256)) { + if (use_logit_softcap && !(DKQ == 128 || DKQ == 256)) { NO_DEVICE_CODE; return; } - static_assert(FATTN_KQ_STRIDE % KQ_per_iter == 0, "bad KQ_per_iter"); + typedef fattn_mma_f16_config c; + + static_assert(FATTN_KQ_STRIDE % fattn_mma_f16_config::nbatch_fa == 0, "bad nbatch_fa"); const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix. const int stride_Q1 = nb01 / sizeof(float2); const int stride_Q2 = nb02 / sizeof(float2); - const int stride_KV = nb11 / sizeof(half2); + const int stride_K = nb11 / sizeof(half2); + const int stride_V = nb21 / sizeof(half2); const int stride_mask = nb31 / sizeof(half2); const int iter_k = ne11 / FATTN_KQ_STRIDE; const int iter_j = (ne01 + (ncols1 - 1)) / ncols1; - constexpr int kb_niter = FATTN_KQ_STRIDE / KQ_per_iter; // Number of kernel iterations per assigned KQ slice. + constexpr int kb_niter = FATTN_KQ_STRIDE / c::nbatch_fa; // Number of kernel iterations per assigned KQ slice. // kbc == k block continuous, current index in continuous ijk space. int kbc = (blockIdx.x + 0)*iter_k*iter_j*(ne02/ncols2) / gridDim.x; @@ -893,9 +1092,9 @@ static __global__ void flash_attn_ext_f16( const float2 * Q_f2 = (const float2 *) (Q + nb02* channel*ncols2); const half2 * K_h2 = (const half2 *) (K + nb12*(channel*ncols2 / gqa_ratio)); - const half2 * V_h2 = (const half2 *) (V + nb12*(channel*ncols2 / gqa_ratio)); // K and V have same shape + const half2 * V_h2 = (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio)); const half2 * mask_h2 = ncols2 > 1 || mask ? (const half2 *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr; - float2 * dstk = ((float2 *) dst) + channel*(ncols2 * D/2); + float2 * dstk = ((float2 *) dst) + channel*(ncols2 * DV/2); const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, channel, n_head_log2, m0, m1) : 1.0f; @@ -905,14 +1104,14 @@ 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 + flash_attn_ext_f16_process_tile (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap, - ne01, ne02, stride_Q1, stride_Q2, stride_KV, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel); + ne01, ne02, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel); } else { constexpr bool needs_fixup = true; // CUDA block is working on the beginning of a tile. - flash_attn_ext_f16_process_tile + flash_attn_ext_f16_process_tile (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap, - ne01, ne02, stride_Q1, stride_Q2, stride_KV, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel); + ne01, ne02, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel); } kbc += iter_k; @@ -931,9 +1130,9 @@ static __global__ void flash_attn_ext_f16( const float2 * Q_f2 = (const float2 *) (Q + nb02* channel*ncols2); const half2 * K_h2 = (const half2 *) (K + nb12*(channel*ncols2 / gqa_ratio)); - const half2 * V_h2 = (const half2 *) (V + nb12*(channel*ncols2 / gqa_ratio)); // K and V have same shape + const half2 * V_h2 = (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio)); // K and V have same shape const half2 * mask_h2 = ncols2 > 1 || mask ? (const half2 *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr; - float2 * dstk = ((float2 *) dst) + channel*(ncols2 * D/2); + float2 * dstk = ((float2 *) dst) + channel*(ncols2 * DV/2); const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, channel, n_head_log2, m0, m1) : 1.0f; @@ -942,9 +1141,9 @@ 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 + flash_attn_ext_f16_process_tile (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap, - ne01, ne02, stride_Q1, stride_Q2, stride_KV, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel); + ne01, ne02, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel); #else GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale); @@ -960,28 +1159,42 @@ static __global__ void flash_attn_ext_f16( #endif // defined(FLASH_ATTN_AVAILABLE) && defined(NEW_MMA_AVAILABLE) } -template +template void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - constexpr int ncols = ncols1 * ncols2; - constexpr int KQ_per_iter = D <= 128 && ncols1 <= 64 ? 64 : 32; - constexpr int nwarps = (KQ_per_iter == 32 && ncols <= 16) ? 2 : 4; - constexpr int ntiles = ncols <= 8 ? 1 : (ncols <= 64 ? 2 : 4); - constexpr int cols_per_warp = ntiles * tile_B::I; + const ggml_tensor * KQV = dst; + const int id = ggml_cuda_get_device(); + const int cc = ggml_cuda_info().devices[id].cc; - static_assert(D % tile_B::J == 0, "bad D"); + typedef fattn_mma_f16_config c; + + constexpr int nbatch_K2 = c::nbatch_K2 < 1 ? DKQ/2 : c::nbatch_K2; + constexpr int nbatch_V2 = c::nbatch_V2 < 1 ? DV /2 : c::nbatch_V2; + constexpr int nbatch_combine = c::nbatch_combine < 1 ? DV /2 : c::nbatch_combine; + + const int nstages = cp_async_available(cc) ? c::nstages_target : 0; + + constexpr int ncols = ncols1 * ncols2; + constexpr int ntiles = ncols <= 8 ? 1 : 2; // Number of tiles per warp. + constexpr int cols_per_warp = ntiles * tile_B::I; + constexpr int nwarps_max_x = ncols / cols_per_warp; + constexpr int nwarps_max_y = c::nbatch_fa / tile_A::I; + constexpr int nwarps = nwarps_max_x*nwarps_max_y <= c::nwarps_max ? nwarps_max_x*nwarps_max_y : c::nwarps_max; + + static_assert(DKQ % tile_B::J == 0, "bad DKQ"); + static_assert(DV % tile_A::J == 0, "bad DV"); static_assert(ncols % cols_per_warp == 0, "bad ncols"); - const ggml_tensor * KQV = dst; - const int id = ggml_cuda_get_device(); - const int cc = ggml_cuda_info().devices[id].cc; + const size_t nbytes_shared_KV_1stage = c::nbatch_fa * std::max(c::nbatch_K2 + 4, c::nbatch_V2 + 4) * sizeof(half2); + const size_t nbytes_shared_KV_2stage = c::nbatch_fa * (c::nbatch_K2 + 4 + c::nbatch_V2 + 4) * sizeof(half2); + const size_t nbytes_shared_Q = ncols * (DKQ/2 + 4) * sizeof(half2); + const size_t nbytes_shared_mask = ncols1 * (c::nbatch_fa/2 + 4) * sizeof(half2); + const size_t nbytes_shared_combine = nwarps*cols_per_warp * (nbatch_combine + 4) * sizeof(half2); - const int KQ_shared_rows = cp_async_available(cc) ? 2*KQ_per_iter : KQ_per_iter; + const size_t nbytes_shared_KV = nstages <= 1 ? nbytes_shared_KV_1stage : nbytes_shared_KV_2stage; - const size_t nbytes_shared_KV = KQ_shared_rows * (D + 8) * sizeof(half); - const size_t nbytes_shared_mask = ncols1 * (KQ_per_iter + 8) * sizeof(half); - const size_t nbytes_shared_combine = nwarps*cols_per_warp * (D + 8) * sizeof(half); - - const size_t nbytes_shared_total = std::max(nbytes_shared_KV + nbytes_shared_mask, nbytes_shared_combine); + const size_t nbytes_shared_total = std::max(nbytes_shared_combine, c::Q_in_reg ? + std::max(nbytes_shared_Q, nbytes_shared_KV + nbytes_shared_mask) : + nbytes_shared_Q + nbytes_shared_KV + nbytes_shared_mask); float logit_softcap; memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float)); @@ -989,59 +1202,73 @@ 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; + fattn_kernel = flash_attn_ext_f16; + +#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA) + static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false}; + if (!shared_memory_limit_raised[id]) { + CUDA_CHECK(cudaFuncSetAttribute(fattn_kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total)); + shared_memory_limit_raised[id] = true; + } +#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA) } else { constexpr bool use_logit_softcap = true; - fattn_kernel = flash_attn_ext_f16; + fattn_kernel = flash_attn_ext_f16; + +#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA) + static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false}; + if (!shared_memory_limit_raised[id]) { + CUDA_CHECK(cudaFuncSetAttribute(fattn_kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total)); + shared_memory_limit_raised[id] = true; + } +#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA) } - launch_fattn + launch_fattn (ctx, dst, fattn_kernel, nwarps, nbytes_shared_total, FATTN_KQ_STRIDE, true, true, true); } -#define DECL_FATTN_MMA_F16_CASE(D, ncols1, ncols2) \ - template void ggml_cuda_flash_attn_ext_mma_f16_case \ - (ggml_backend_cuda_context & ctx, ggml_tensor * dst) \ +#define DECL_FATTN_MMA_F16_CASE(DKQ, DV, ncols1, ncols2) \ + template void ggml_cuda_flash_attn_ext_mma_f16_case \ + (ggml_backend_cuda_context & ctx, ggml_tensor * dst) \ -#define DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(D, ncols) \ - extern DECL_FATTN_MMA_F16_CASE(D, (ncols)/1, 1); \ - extern DECL_FATTN_MMA_F16_CASE(D, (ncols)/2, 2); \ - extern DECL_FATTN_MMA_F16_CASE(D, (ncols)/4, 4); \ - extern DECL_FATTN_MMA_F16_CASE(D, (ncols)/8, 8); \ +#define DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(DKQ, DV, ncols) \ + extern DECL_FATTN_MMA_F16_CASE(DKQ, DV, (ncols)/ 1, 1); \ + extern DECL_FATTN_MMA_F16_CASE(DKQ, DV, (ncols)/ 2, 2); \ + extern DECL_FATTN_MMA_F16_CASE(DKQ, DV, (ncols)/ 4, 4); \ + extern DECL_FATTN_MMA_F16_CASE(DKQ, DV, (ncols)/ 8, 8); \ + extern DECL_FATTN_MMA_F16_CASE(DKQ, DV, (ncols)/16, 16); \ -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 8) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 8) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 8) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 8) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 8) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 8) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 64, 8) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 80, 8) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 96, 8) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 112, 8) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 128, 8) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 256, 8) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 16) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 16) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 16) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 16) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 16) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 16) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 64, 16) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 80, 16) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 96, 16) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 112, 16) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 128, 16) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 256, 16) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 32) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 32) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 32) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 32) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 32) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 32) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 64, 32) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 80, 32) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 96, 32) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 112, 32) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 128, 32) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 256, 32) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 64) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 64) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 64) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 64) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 64) -DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 64) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 64, 64) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 80, 64) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 96, 64) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 112, 64) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 128, 64) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 256, 64) -// Kernels with ncols == 128 are only 4% faster due to register pressure. -// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 128) -// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 128) -// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 128) -// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 128) -// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 128) -// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 128) // Needs too much shared memory. +// The number of viable configurations for Deepseek is very limited: +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); diff --git a/ggml/src/ggml-cuda/fattn-tile-f16.cu b/ggml/src/ggml-cuda/fattn-tile-f16.cu index e0039e1755..9283560d5c 100644 --- a/ggml/src/ggml-cuda/fattn-tile-f16.cu +++ b/ggml/src/ggml-cuda/fattn-tile-f16.cu @@ -307,7 +307,7 @@ void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * constexpr int nwarps = 8; constexpr size_t nbytes_shared = 0; fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16; - launch_fattn + launch_fattn (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false); } break; case 128: { @@ -315,7 +315,7 @@ void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * constexpr int nwarps = 8; constexpr size_t nbytes_shared = 0; fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16; - launch_fattn + launch_fattn (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false); } break; default: { diff --git a/ggml/src/ggml-cuda/fattn-tile-f32.cu b/ggml/src/ggml-cuda/fattn-tile-f32.cu index fcb6f848fe..32673adb57 100644 --- a/ggml/src/ggml-cuda/fattn-tile-f32.cu +++ b/ggml/src/ggml-cuda/fattn-tile-f32.cu @@ -318,7 +318,7 @@ void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * constexpr int nwarps = 8; constexpr size_t nbytes_shared = 0; fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32; - launch_fattn + launch_fattn (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F32, true, true, false); } break; case 128: { @@ -326,7 +326,7 @@ void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * constexpr int nwarps = 8; constexpr size_t nbytes_shared = 0; fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32; - launch_fattn + launch_fattn (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F32, true, true, false); } break; default: { diff --git a/ggml/src/ggml-cuda/fattn-vec-f16.cuh b/ggml/src/ggml-cuda/fattn-vec-f16.cuh index e17d2d0e4f..d96e392129 100644 --- a/ggml/src/ggml-cuda/fattn-vec-f16.cuh +++ b/ggml/src/ggml-cuda/fattn-vec-f16.cuh @@ -168,6 +168,7 @@ static __global__ void flash_attn_vec_ext_f16( for (int j = 0; j < ncols; ++j) { KQ[j*D + tid] = -HALF_MAX_HALF; } + __syncthreads(); half2 VKQ[ncols] = {{0.0f, 0.0f}}; @@ -315,7 +316,7 @@ void ggml_cuda_flash_attn_ext_vec_f16_case_impl(ggml_backend_cuda_context & ctx, constexpr bool need_f16_K = D != 128; constexpr bool need_f16_V = D != 128 && D != 64; constexpr size_t nbytes_shared = 0; - launch_fattn(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false); + launch_fattn(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false); } template diff --git a/ggml/src/ggml-cuda/fattn-vec-f32.cuh b/ggml/src/ggml-cuda/fattn-vec-f32.cuh index d42ddca49f..7064675d5a 100644 --- a/ggml/src/ggml-cuda/fattn-vec-f32.cuh +++ b/ggml/src/ggml-cuda/fattn-vec-f32.cuh @@ -310,7 +310,7 @@ void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx, constexpr bool need_f16_K = D != 128; constexpr bool need_f16_V = D != 128 && D != 64; constexpr size_t nbytes_shared = 0; - launch_fattn(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false); + launch_fattn(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false); } template diff --git a/ggml/src/ggml-cuda/fattn-wmma-f16.cu b/ggml/src/ggml-cuda/fattn-wmma-f16.cu index bc21b27a0c..c5668adb15 100644 --- a/ggml/src/ggml-cuda/fattn-wmma-f16.cu +++ b/ggml/src/ggml-cuda/fattn-wmma-f16.cu @@ -490,7 +490,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggm fattn_kernel = flash_attn_ext_f16< D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), KQ_acc_t, use_logit_softcap>; } - launch_fattn(ctx, dst, fattn_kernel, nwarps, 0, FATTN_KQ_STRIDE, true, true, false, warp_size); + launch_fattn(ctx, dst, fattn_kernel, nwarps, 0, FATTN_KQ_STRIDE, true, true, false, warp_size); } void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu index 7a2d1e4536..9c5c803d02 100644 --- a/ggml/src/ggml-cuda/fattn.cu +++ b/ggml/src/ggml-cuda/fattn.cu @@ -8,58 +8,32 @@ #include "fattn-wmma-f16.cuh" #include "fattn.cuh" -template +template static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * Q = dst->src[0]; - if (Q->ne[1] <= 8/ncols2) { - ggml_cuda_flash_attn_ext_mma_f16_case(ctx, dst); - return; + if constexpr (ncols2 <= 8) { + if (Q->ne[1] <= 8/ncols2) { + ggml_cuda_flash_attn_ext_mma_f16_case(ctx, dst); + return; + } } if (Q->ne[1] <= 16/ncols2) { - ggml_cuda_flash_attn_ext_mma_f16_case(ctx, dst); + ggml_cuda_flash_attn_ext_mma_f16_case(ctx, dst); return; } if (Q->ne[1] <= 32/ncols2) { - ggml_cuda_flash_attn_ext_mma_f16_case(ctx, dst); + ggml_cuda_flash_attn_ext_mma_f16_case(ctx, dst); return; } - ggml_cuda_flash_attn_ext_mma_f16_case(ctx, dst); + ggml_cuda_flash_attn_ext_mma_f16_case(ctx, dst); } -template -static void ggml_cuda_flash_attn_ext_mma_f16_switch_hs(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * Q = dst->src[0]; - - switch (Q->ne[0]) { - case 64: - ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1< 64, ncols2>(ctx, dst); - break; - case 80: - ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1< 80, ncols2>(ctx, dst); - break; - case 96: - ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1< 96, ncols2>(ctx, dst); - break; - case 112: - ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<112, ncols2>(ctx, dst); - break; - case 128: - ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<128, ncols2>(ctx, dst); - break; - case 256: - ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<256, ncols2>(ctx, dst); - break; - default: - GGML_ABORT("fatal error"); - break; - } -} - -static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { +template +static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * KQV = dst; const ggml_tensor * Q = dst->src[0]; const ggml_tensor * K = dst->src[1]; @@ -68,27 +42,79 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg float max_bias = 0.0f; memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float)); - const float use_gqa_opt = mask && max_bias == 0.0f; + const bool use_gqa_opt = mask && max_bias == 0.0f; 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) { - ggml_cuda_flash_attn_ext_mma_f16_switch_hs<8>(ctx, dst); + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ctx, dst); return; } - if (use_gqa_opt && gqa_ratio == 4) { - ggml_cuda_flash_attn_ext_mma_f16_switch_hs<4>(ctx, dst); + if (use_gqa_opt && gqa_ratio % 4 == 0) { + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ctx, dst); return; } - if (use_gqa_opt && gqa_ratio == 2) { - ggml_cuda_flash_attn_ext_mma_f16_switch_hs<2>(ctx, dst); + if (use_gqa_opt && gqa_ratio % 2 == 0) { + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ctx, dst); return; } - ggml_cuda_flash_attn_ext_mma_f16_switch_hs<1>(ctx, dst); + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ctx, dst); +} + +static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * KQV = dst; + 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]; + + switch (Q->ne[0]) { + case 64: + GGML_ASSERT(V->ne[0] == 64); + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2< 64, 64>(ctx, dst); + break; + case 80: + GGML_ASSERT(V->ne[0] == 80); + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2< 80, 80>(ctx, dst); + break; + case 96: + GGML_ASSERT(V->ne[0] == 96); + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2< 96, 96>(ctx, dst); + break; + case 112: + GGML_ASSERT(V->ne[0] == 112); + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<112, 112>(ctx, dst); + break; + case 128: + GGML_ASSERT(V->ne[0] == 128); + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<128, 128>(ctx, dst); + break; + case 256: + GGML_ASSERT(V->ne[0] == 256); + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<256, 256>(ctx, dst); + break; + case 576: { + // For Deepseek, go straight to the ncols1 switch to avoid compiling unnecessary kernels. + GGML_ASSERT(V->ne[0] == 512); + float max_bias = 0.0f; + memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float)); + + const bool use_gqa_opt = mask && max_bias == 0.0f; + GGML_ASSERT(use_gqa_opt); + + 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); + } break; + default: + GGML_ABORT("fatal error"); + break; + } } #define FATTN_VEC_F16_CASE(D, type_K, type_V) \ @@ -299,7 +325,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst const bool gqa_opt_applies = ((Q->ne[2] / K->ne[2]) % 2 == 0) && mask; // The mma-based kernels have GQA-specific optimizations const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16; const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && cc < GGML_CUDA_CC_ADA_LOVELACE && !mma_needs_data_conversion; - const bool can_use_vector_kernel = Q->ne[0] % (2*warp_size) == 0; + const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % (2*warp_size) == 0; if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) { if (prec == GGML_PREC_DEFAULT) { ggml_cuda_flash_attn_ext_vec_f16(ctx, dst); diff --git a/ggml/src/ggml-cuda/getrows.cu b/ggml/src/ggml-cuda/getrows.cu index 4cef53a98c..963e4d03dd 100644 --- a/ggml/src/ggml-cuda/getrows.cu +++ b/ggml/src/ggml-cuda/getrows.cu @@ -10,10 +10,11 @@ static __global__ void k_get_rows( /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03, const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) { - const int i00 = (blockIdx.x*blockDim.x + threadIdx.x)*2; - const int i10 = blockDim.y*blockIdx.y + threadIdx.y; - const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12; - const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12; + // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher. + const int i00 = (blockIdx.y * blockDim.x + threadIdx.x)*2; + const int i10 = blockIdx.x; + const int i11 = blockIdx.z / ne12; + const int i12 = blockIdx.z % ne12; if (i00 >= ne00) { return; @@ -33,8 +34,8 @@ static __global__ void k_get_rows( dfloat2 v; dequantize_kernel(src0_row, ib, iqs, v); - dst_row[iybs + iqs + 0] = v.x; - dst_row[iybs + iqs + y_offset] = v.y; + dst_row[iybs + iqs + 0] = float(v.x); + dst_row[iybs + iqs + y_offset] = float(v.y); } template @@ -46,10 +47,11 @@ static __global__ void k_get_rows_float( /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03, const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) { - const int i00 = blockIdx.x*blockDim.x + threadIdx.x; - const int i10 = blockDim.y*blockIdx.y + threadIdx.y; - const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12; - const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12; + // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher. + const int i00 = blockIdx.y * blockDim.x + threadIdx.x; + const int i10 = blockIdx.x; + const int i11 = blockIdx.z / ne12; + const int i12 = blockIdx.z % ne12; if (i00 >= ne00) { return; @@ -60,7 +62,7 @@ static __global__ void k_get_rows_float( dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3; const src0_t * src0_row = (const src0_t *)((const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03); - dst_row[i00] = src0_row[i00]; + dst_row[i00] = float(src0_row[i00]); } template @@ -86,120 +88,159 @@ static __global__ void k_get_rows_back_float( dst[dst_row*ncols + col] = sum; } -template -static void get_rows_cuda( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const void * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) { - - GGML_TENSOR_BINARY_OP_LOCALS - +template +static void get_rows_cuda_q( + const void * src0_d, const int32_t * src1_d, dst_t * dst_d, + const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03, + const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12, + const size_t nb1, const size_t nb2, const size_t nb3, + cudaStream_t stream) { const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1); - const int block_num_x = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE); - const dim3 block_nums(block_num_x, ne10, ne11*ne12); + const int block_num_y = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE); + const dim3 block_nums(ne10, block_num_y, ne11*ne12); // strides in elements - //const size_t s0 = nb0 / ggml_element_size(dst); - const size_t s1 = nb1 / ggml_element_size(dst); - const size_t s2 = nb2 / ggml_element_size(dst); - const size_t s3 = nb3 / ggml_element_size(dst); + // const size_t s0 = nb0 / sizeof(dst_t); + const size_t s1 = nb1 / sizeof(dst_t); + const size_t s2 = nb2 / sizeof(dst_t); + const size_t s3 = nb3 / sizeof(dst_t); - const size_t s10 = nb10 / ggml_element_size(src1); - const size_t s11 = nb11 / ggml_element_size(src1); - const size_t s12 = nb12 / ggml_element_size(src1); - //const size_t s13 = nb13 / ggml_element_size(src1); + const size_t s10 = nb10 / sizeof(int32_t); + const size_t s11 = nb11 / sizeof(int32_t); + const size_t s12 = nb12 / sizeof(int32_t); + // const size_t s13 = nb13 / sizeof(int32_t); GGML_ASSERT(ne00 % 2 == 0); k_get_rows<<>>( - src0_dd, src1_dd, dst_dd, + src0_d, src1_d, dst_d, ne00, /*ne01, ne02, ne03,*/ /*ne10, ne11,*/ ne12, /*ne13,*/ /* s0,*/ s1, s2, s3, /* nb00,*/ nb01, nb02, nb03, s10, s11, s12/*, s13*/); - - GGML_UNUSED(dst); } -template +template static void get_rows_cuda_float( - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, - const src0_t * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) { - - GGML_TENSOR_BINARY_OP_LOCALS - - GGML_ASSERT(ne13 == 1); - + const src0_t * src0_d, const int32_t * src1_d, dst_t * dst_d, + const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03, + const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12, + const size_t nb1, const size_t nb2, const size_t nb3, + cudaStream_t stream) { const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1); - const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE; - const dim3 block_nums(block_num_x, ne10, ne11*ne12); + const int block_num_y = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE; + const dim3 block_nums(ne10, block_num_y, ne11*ne12); // strides in elements - //const size_t s0 = nb0 / ggml_element_size(dst); - const size_t s1 = nb1 / ggml_element_size(dst); - const size_t s2 = nb2 / ggml_element_size(dst); - const size_t s3 = nb3 / ggml_element_size(dst); + // const size_t s0 = nb0 / sizeof(dst_t); + const size_t s1 = nb1 / sizeof(dst_t); + const size_t s2 = nb2 / sizeof(dst_t); + const size_t s3 = nb3 / sizeof(dst_t); - const size_t s10 = nb10 / ggml_element_size(src1); - const size_t s11 = nb11 / ggml_element_size(src1); - const size_t s12 = nb12 / ggml_element_size(src1); - //const size_t s13 = nb13 / ggml_element_size(src1); + const size_t s10 = nb10 / sizeof(int32_t); + const size_t s11 = nb11 / sizeof(int32_t); + const size_t s12 = nb12 / sizeof(int32_t); + // const size_t s13 = nb13 / sizeof(int32_t); k_get_rows_float<<>>( - src0_dd, src1_dd, dst_dd, + src0_d, src1_d, dst_d, ne00, /*ne01, ne02, ne03,*/ /*ne10, ne11,*/ ne12, /*ne13,*/ /* s0,*/ s1, s2, s3, /* nb00,*/ nb01, nb02, nb03, s10, s11, s12/*, s13*/); +} - GGML_UNUSED(dst); +template +static void ggml_cuda_get_rows_switch_src0_type( + const void * src0_d, const ggml_type src0_type, const int32_t * src1_d, dst_t * dst_d, + const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03, + const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12, + const size_t nb1, const size_t nb2, const size_t nb3, + cudaStream_t stream) { + switch (src0_type) { + case GGML_TYPE_F16: + get_rows_cuda_float((const half *) src0_d, src1_d, dst_d, + ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream); + break; + case GGML_TYPE_F32: + get_rows_cuda_float((const float *) src0_d, src1_d, dst_d, + ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream); + break; + case GGML_TYPE_BF16: + get_rows_cuda_float((const nv_bfloat16 *) src0_d, src1_d, dst_d, + ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream); + break; + case GGML_TYPE_Q4_0: + get_rows_cuda_q(src0_d, src1_d, dst_d, + ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream); + break; + case GGML_TYPE_Q4_1: + get_rows_cuda_q(src0_d, src1_d, dst_d, + ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream); + break; + case GGML_TYPE_Q5_0: + get_rows_cuda_q(src0_d, src1_d, dst_d, + ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream); + break; + case GGML_TYPE_Q5_1: + get_rows_cuda_q(src0_d, src1_d, dst_d, + ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream); + break; + case GGML_TYPE_Q8_0: + get_rows_cuda_q(src0_d, src1_d, dst_d, + ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream); + break; + default: + // TODO: k-quants + GGML_ABORT("%s: unsupported src0 type: %s\n", __func__, ggml_type_name(src0_type)); + break; + } +} + +void get_rows_cuda( + const void * src0_d, ggml_type src0_type, const int32_t * src1_d, void * dst_d, ggml_type dst_type, + int64_t ne00, size_t nb01, size_t nb02, size_t nb03, + int64_t ne10, int64_t ne11, int64_t ne12, size_t nb10, size_t nb11, size_t nb12, + size_t nb1, size_t nb2, size_t nb3, + cudaStream_t stream) { + switch (dst_type) { + case GGML_TYPE_F32: + ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (float *) dst_d, + ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream); + break; + case GGML_TYPE_F16: + ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (half *) dst_d, + ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream); + break; + case GGML_TYPE_BF16: + ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (nv_bfloat16 *) dst_d, + ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream); + break; + default: + GGML_ABORT("%s: unsupported dst type: %s\n", __func__, ggml_type_name(dst_type)); + break; + } } void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * src0 = dst->src[0]; const ggml_tensor * src1 = dst->src[1]; - const void * src0_d = (const void *) src0->data; - const int32_t * src1_d = (const int32_t *) src1->data; - float * dst_d = (float *) dst->data; - cudaStream_t stream = ctx.stream(); + GGML_TENSOR_BINARY_OP_LOCALS + GGML_ASSERT(src1->type == GGML_TYPE_I32); - GGML_ASSERT(dst->type == GGML_TYPE_F32); + GGML_ASSERT(ne13 == 1); GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type)); GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type)); GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type)); - switch (src0->type) { - case GGML_TYPE_F16: - get_rows_cuda_float(src0, src1, dst, (const half *) src0_d, src1_d, dst_d, stream); - break; - case GGML_TYPE_F32: - get_rows_cuda_float(src0, src1, dst, (const float *) src0_d, src1_d, dst_d, stream); - break; - case GGML_TYPE_Q4_0: - get_rows_cuda(src0, src1, dst, src0_d, src1_d, dst_d, stream); - break; - case GGML_TYPE_Q4_1: - get_rows_cuda(src0, src1, dst, src0_d, src1_d, dst_d, stream); - break; - case GGML_TYPE_Q5_0: - get_rows_cuda(src0, src1, dst, src0_d, src1_d, dst_d, stream); - break; - case GGML_TYPE_Q5_1: - get_rows_cuda(src0, src1, dst, src0_d, src1_d, dst_d, stream); - break; - case GGML_TYPE_Q8_0: - get_rows_cuda(src0, src1, dst, src0_d, src1_d, dst_d, stream); - break; - default: - // TODO: k-quants - GGML_ABORT("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type)); - break; - } + get_rows_cuda(src0->data, src0->type, (const int32_t *) src1->data, dst->data, dst->type, + ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream); } void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { diff --git a/ggml/src/ggml-cuda/getrows.cuh b/ggml/src/ggml-cuda/getrows.cuh index a1ca643f1c..3c5bea5f48 100644 --- a/ggml/src/ggml-cuda/getrows.cuh +++ b/ggml/src/ggml-cuda/getrows.cuh @@ -3,6 +3,13 @@ #define CUDA_GET_ROWS_BLOCK_SIZE 256 #define CUDA_GET_ROWS_BACK_BLOCK_SIZE 256 +void get_rows_cuda( + const void * src0_d, ggml_type src0_type, const int32_t * src1_d, void * dst_d, ggml_type dst_type, + int64_t ne00, size_t nb01, size_t nb02, size_t nb03, + int64_t ne10, int64_t ne11, int64_t ne12, size_t nb10, size_t nb11, size_t nb12, + size_t nb1, size_t nb2, size_t nb3, + cudaStream_t stream); + void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst); void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index e0e0d2137f..b4b85abcda 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -555,8 +555,8 @@ static enum ggml_status ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer if (ggml_is_quantized(tensor->type) && tensor->view_src == nullptr && ggml_backend_buffer_get_usage(buffer) != GGML_BACKEND_BUFFER_USAGE_COMPUTE) { // initialize padding to 0 to avoid possible NaN values - size_t original_size = ggml_nbytes(tensor); - size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor); + const size_t original_size = ggml_nbytes(tensor); + const size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor); if (padded_size > original_size) { ggml_cuda_set_device(ctx->device); @@ -679,6 +679,7 @@ static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_t if (ggml_is_quantized(tensor->type)) { if (ne0 % MATRIX_ROW_PADDING != 0) { + GGML_ASSERT(tensor->nb[0] == ggml_element_size(tensor)); size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); } } @@ -800,6 +801,7 @@ static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buff static enum ggml_status ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported + GGML_ASSERT(ggml_is_contiguous(tensor) && "split buffers only supported for contiguous tensors"); ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context; ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context; @@ -851,6 +853,7 @@ static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buffer_t buff // split tensors must always be set in their entirety at once GGML_ASSERT(offset == 0); GGML_ASSERT(size == ggml_nbytes(tensor)); + GGML_ASSERT(ggml_is_contiguous(tensor) && "split buffers only supported for contiguous tensors"); ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context; @@ -889,6 +892,7 @@ static void ggml_backend_cuda_split_buffer_get_tensor(ggml_backend_buffer_t buff // split tensors must always be set in their entirety at once GGML_ASSERT(offset == 0); GGML_ASSERT(size == ggml_nbytes(tensor)); + GGML_ASSERT(ggml_is_contiguous(tensor) && "split buffers only supported for contiguous tensors"); ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context; @@ -970,6 +974,7 @@ static size_t ggml_backend_cuda_split_buffer_type_get_alignment(ggml_backend_buf static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { ggml_backend_cuda_split_buffer_type_context * ctx = (ggml_backend_cuda_split_buffer_type_context *)buft->context; + GGML_ASSERT(ggml_is_contiguous(tensor) && "split buffers only supported for contiguous tensors"); size_t total_size = 0; @@ -1531,6 +1536,8 @@ static void ggml_cuda_op_mul_mat( // If src0 is on a temporary compute buffer (partial offloading) there may be some padding that needs to be cleared: if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) { + GGML_ASSERT(ggml_is_contiguously_allocated(src0)); + GGML_ASSERT(!src0->view_src); const size_t nbytes_data = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00); const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING); CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data, 0, nbytes_padding, stream)); @@ -1551,7 +1558,7 @@ static void ggml_cuda_op_mul_mat( if (src1_on_device && src1_is_contiguous) { quantize_src1( - dev[id].src1_ddf, dev[id].src1_ddq, src0->type, ne10, + dev[id].src1_ddf, nullptr, dev[id].src1_ddq, src0->type, ne10, nb11/sizeof(float), nb12/sizeof(float), nb13/sizeof(float), src1_padded_col_size, ne11, ne12, ne13, stream); CUDA_CHECK(cudaGetLastError()); @@ -1649,7 +1656,7 @@ static void ggml_cuda_op_mul_mat( if (quantize_src1 && !src1_is_contiguous) { quantize_src1( - src1_ddf_i, src1_ddq_i, src0->type, ne10, ne10, ne11*ne10, ne12*ne11*ne10, + src1_ddf_i, nullptr, src1_ddq_i, src0->type, ne10, ne10, ne11*ne10, ne12*ne11*ne10, src1_padded_col_size, src1_ncols, 1, 1, stream); CUDA_CHECK(cudaGetLastError()); } @@ -1720,15 +1727,15 @@ static __global__ void k_compute_batched_ptrs( size_t nb12, size_t nb13, size_t nbd2, size_t nbd3, int64_t r2, int64_t r3) { - int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x; - int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y; + const int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x; + const int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y; if (i13 >= ne13 || i12 >= ne12) { return; } - int64_t i03 = i13 / r3; - int64_t i02 = i12 / r2; + const int64_t i03 = i13 / r3; + const int64_t i02 = i12 / r2; ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03; ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12 + i13*nb13; @@ -1742,6 +1749,10 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer)); GGML_ASSERT(src0->type == GGML_TYPE_F16); + // Byte offsets and tensor dimensions are currently used in an inconsistent way for dst. + // As long as dst is contiguous this does not matter though. + GGML_ASSERT(ggml_is_contiguous(dst)); + GGML_TENSOR_BINARY_OP_LOCALS const int64_t ne_dst = ggml_nelements(dst); @@ -1750,21 +1761,31 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream)); - void * src0_ddq = src0->data; - half * src0_f16 = (half *) src0_ddq; - float * src1_ddf = (float *) src1->data; - float * dst_ddf = (float *) dst->data; + const half * src0_f16 = (const half *) src0->data; + float * dst_ddf = (float *) dst->data; + + const half * src1_f16 = (const half *) src1->data; + const size_t ts_src1 = ggml_type_size(src1->type); + GGML_ASSERT(nb10 == ts_src1); + int64_t s11 = nb11 / ts_src1; + int64_t s12 = nb12 / ts_src1; + int64_t s13 = nb13 / ts_src1; + ggml_cuda_pool_alloc src1_f16_alloc(ctx.pool()); // convert src1 to fp16 - ggml_cuda_pool_alloc src1_f16_alloc(ctx.pool()); if (src1->type != GGML_TYPE_F16) { - const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type); + const to_fp16_nc_cuda_t to_fp16_cuda = ggml_get_to_fp16_nc_cuda(src1->type); const int64_t ne_src1 = ggml_nelements(src1); src1_f16_alloc.alloc(ne_src1); GGML_ASSERT(to_fp16_cuda != nullptr); - to_fp16_cuda(src1_ddf, src1_f16_alloc.get(), ne_src1, main_stream); + + to_fp16_cuda(src1_f16, src1_f16_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream); + + src1_f16 = src1_f16_alloc.get(); + s11 = ne10; + s12 = ne11*s11; + s13 = ne12*s12; } - half * src1_f16 = src1->type == GGML_TYPE_F16 ? (half *) src1_ddf : src1_f16_alloc.get(); ggml_cuda_pool_alloc dst_f16(ctx.pool()); char * dst_t; @@ -1824,13 +1845,13 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co int i02 = i12 / r2; CUBLAS_CHECK( - cublasGemmEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N, - ne01, ne11, ne10, - alpha, (const char *) src0_as_f16 + i02*src0->nb[2] + i03*src0->nb[3] , CUDA_R_16F, nb01/sizeof(half), - (const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F, nb11/sizeof(float), - beta, ( char *) dst_t + i12*nbd2 + i13*nbd3, cu_data_type, ne01, - cu_compute_type, - CUBLAS_GEMM_DEFAULT_TENSOR_OP)); + cublasGemmEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N, + ne01, ne11, ne10, + alpha, (const char *) src0_f16 + i03*nb03 + i02*nb02, CUDA_R_16F, nb01/sizeof(half), + src1_f16 + i13*s13 + i12*s12, CUDA_R_16F, s11, + beta, ( char *) dst_t + i13*nbd3 + i12*nbd2, cu_data_type, ne0, + cu_compute_type, + CUBLAS_GEMM_DEFAULT_TENSOR_OP)); } } } @@ -1841,15 +1862,15 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co CUBLAS_CHECK( cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N, ne01, ne11, ne10, - alpha, (const char *) src0_f16, CUDA_R_16F, nb01/nb00, nb02/nb00, // strideA - (const char *) src1_f16, CUDA_R_16F, nb11/nb10, nb12/nb10, // strideB - beta, ( char *) dst_t, cu_data_type, ne01, nb2/nb0, // strideC + alpha, src0_f16, CUDA_R_16F, nb01/nb00, nb02/nb00, // strideA + src1_f16, CUDA_R_16F, s11, s12, // strideB + beta, dst_t, cu_data_type, ne0, ne1*ne0, // strideC ne12*ne13, cu_compute_type, CUBLAS_GEMM_DEFAULT_TENSOR_OP)); } else { // use cublasGemmBatchedEx - const int ne23 = ne12*ne13; + const int64_t ne23 = ne12*ne13; ggml_cuda_pool_alloc ptrs_src(ctx.pool(), 2*ne23); ggml_cuda_pool_alloc< void *> ptrs_dst(ctx.pool(), 1*ne23); @@ -1861,8 +1882,8 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co ne12, ne13, ne23, nb02, nb03, - src1->type == GGML_TYPE_F16 ? nb12 : nb12/2, - src1->type == GGML_TYPE_F16 ? nb13 : nb13/2, + src1->type == GGML_TYPE_F16 ? nb12 : s12*sizeof(half), + src1->type == GGML_TYPE_F16 ? nb13 : s13*sizeof(half), nbd2, nbd3, r2, r3); CUDA_CHECK(cudaGetLastError()); @@ -1871,8 +1892,8 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N, ne01, ne11, ne10, alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F, nb01/nb00, - (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F, nb11/nb10, - beta, ( void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne01, + (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F, s11, + beta, ( void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne0, ne23, cu_compute_type, CUBLAS_GEMM_DEFAULT_TENSOR_OP)); @@ -1888,13 +1909,19 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft); + // If src0 is a temporary compute buffer it may have some padding that needs to be cleared for mul_mat_vec_q or mul_mat_q. + // But if src0 is also a view of another tensor then this cannot be done safely because it may overwrite valid tensor data. + // Therefore, in such cases use cuBLAS. + const bool bad_padding_clear = ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE + && ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) && src0->view_src; + bool use_mul_mat_vec = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16) && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 && src0->ne[0] % 2 == 0 && src1->ne[1] == 1; - bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) + bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE; - bool use_mul_mat_q = ggml_is_quantized(src0->type) + bool use_mul_mat_q = ggml_is_quantized(src0->type) && !bad_padding_clear && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32; bool any_gpus_with_slow_fp16 = false; @@ -1935,8 +1962,10 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor ggml_cuda_mul_mat_vec(ctx, src0, src1, nullptr, dst); } else if (!split && use_mul_mat_vec_q) { ggml_cuda_mul_mat_vec_q(ctx, src0, src1, nullptr, dst); - } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16) - && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) { + } else if (!split && use_mul_mat_q) { + ggml_cuda_mul_mat_q(ctx, src0, src1, nullptr, dst); + } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16) && + !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) { // general KQ + KQV multi-batch without FlashAttention ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst); } else if (use_mul_mat_vec) { @@ -1950,183 +1979,147 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor } } -struct mmid_row_mapping { - int32_t i1; - int32_t i2; -}; - -static __global__ void k_copy_src1_to_contiguous(const char * __restrict__ src1_original, char * __restrict__ src1_contiguous, - int * __restrict__ cur_src1_row, mmid_row_mapping * __restrict__ row_mapping, - const char * __restrict ids, int64_t i02, size_t ids_nb1, size_t ids_nb0, - int64_t ne11, int64_t ne10, - size_t nb11, size_t nb12) { - int32_t iid1 = blockIdx.x; - int32_t id = blockIdx.y; - - const int32_t row_id_i = *(const int32_t *) (ids + iid1*ids_nb1 + id*ids_nb0); - - if (row_id_i != i02) { - return; - } - - const int64_t i11 = id % ne11; - const int64_t i12 = iid1; - - __shared__ int src1_row; - if (threadIdx.x == 0) { - src1_row = atomicAdd(cur_src1_row, 1); - row_mapping[src1_row] = {id, iid1}; - } - __syncthreads(); - - const float * src1_row_original = (const float *)(src1_original + i11*nb11 + i12*nb12); - float * src1_row_contiguous = (float *)(src1_contiguous + src1_row*nb11); - - for (int i = threadIdx.x; i < ne10; i += blockDim.x) { - src1_row_contiguous[i] = src1_row_original[i]; - } -} - -static __global__ void k_copy_dst_from_contiguous(char * __restrict__ dst_original, const char * __restrict__ dst_contiguous, - const mmid_row_mapping * __restrict__ row_mapping, - int64_t ne0, - size_t nb1, size_t nb2) { - int32_t i = blockIdx.x; - - const int32_t i1 = row_mapping[i].i1; - const int32_t i2 = row_mapping[i].i2; - - const float * dst_row_contiguous = (const float *)(dst_contiguous + i*nb1); - float * dst_row_original = (float *)(dst_original + i1*nb1 + i2*nb2); - - for (int j = threadIdx.x; j < ne0; j += blockDim.x) { - dst_row_original[j] = dst_row_contiguous[j]; - } -} - static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * src0 = dst->src[0]; const ggml_tensor * src1 = dst->src[1]; const ggml_tensor * ids = dst->src[2]; + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft) && "mul_mat_id does not support split buffers"); + GGML_TENSOR_BINARY_OP_LOCALS - if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 && ne2 == 1) { - if (ggml_is_quantized(src0->type)) { - ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst); - } else { - ggml_cuda_mul_mat_vec(ctx, src0, src1, ids, dst); - } - return; - } + const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc; - GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft) && "mul_mat_id does not support split buffers"); + if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + if (ne2 == 1) { + if (ggml_is_quantized(src0->type)) { + ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst); + } else { + ggml_cuda_mul_mat_vec(ctx, src0, src1, ids, dst); + } + return; + } + + if (ggml_cuda_should_use_mmq(src0->type, cc, ne12)) { + ggml_cuda_mul_mat_q(ctx, src0, src1, ids, dst); + return; + } + } cudaStream_t stream = ctx.stream(); - const int64_t n_as = ne02; - const int64_t n_ids = ids->ne[0]; + GGML_ASSERT(nb12 % nb11 == 0); + GGML_ASSERT(nb2 % nb1 == 0); + + const ggml_type type_src1_sorted = (src0->type == GGML_TYPE_F16 && !fast_fp16_hardware_available(cc)) + || ggml_is_quantized(src0->type) ? GGML_TYPE_F32 : src0->type; + const ggml_type type_dst_sorted = GGML_TYPE_F32; + const size_t ts_src1_sorted = ggml_type_size(type_src1_sorted); + const size_t ts_dst_sorted = ggml_type_size(type_dst_sorted); + + const int64_t n_expert_used = ids->ne[0]; + const int64_t ne_get_rows = ne12 * n_expert_used; + + std::vector ids_to_sorted_host; + ids_to_sorted_host.reserve(2*ne_get_rows); + std::vector ids_from_sorted_host(ne_get_rows); + + ggml_cuda_pool_alloc ids_buf_dev(ctx.pool(), 2*ne_get_rows); + + std::vector tokens_per_expert(ne02); + + ggml_cuda_pool_alloc src1_sorted(ctx.pool(), ne12*n_expert_used*ne10*ts_src1_sorted); + ggml_cuda_pool_alloc dst_sorted(ctx.pool(), ne2 *n_expert_used* ne0*ts_dst_sorted); std::vector ids_host(ggml_nbytes(ids)); - const char * ids_dev = (const char *) ids->data; - CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids_dev, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream)); + CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids->data, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream)); CUDA_CHECK(cudaStreamSynchronize(stream)); - ggml_tensor src0_row = *src0; - ggml_tensor src1_row = *src1; - ggml_tensor dst_row = *dst; - - char * src0_original = (char *) src0->data; - char * src1_original = (char *) src1->data; - char * dst_original = (char *) dst->data; - - src0_row.ne[2] = 1; - src0_row.ne[3] = 1; - src0_row.nb[3] = nb02; - - src1_row.ne[1] = 1; - src1_row.ne[2] = 1; - src1_row.ne[3] = 1; - src1_row.nb[2] = nb11; - src1_row.nb[3] = nb11; - - dst_row.ne[1] = 1; - dst_row.ne[2] = 1; - dst_row.ne[3] = 1; - dst_row.nb[2] = nb1; - dst_row.nb[3] = nb1; - - ggml_cuda_pool_alloc src1_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(src1)); - ggml_cuda_pool_alloc dst_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(dst)); - - src1_row.data = src1_contiguous.get(); - dst_row.data = dst_contiguous.get(); - - for (int64_t i02 = 0; i02 < n_as; i02++) { - int64_t num_src1_rows = 0; - - for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) { - for (int64_t id = 0; id < n_ids; id++) { - const int32_t row_id_i = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]); - - GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as); - - if (row_id_i != i02) { - continue; + for (int64_t i02 = 0; i02 < ne02; ++i02) { // expert matrices + for (int64_t i12 = 0; i12 < ne12; ++i12) { // tokens + for (int64_t iex = 0; iex < n_expert_used; ++iex) { + const int32_t expert_to_use = *(const int32_t *)(ids_host.data() + i12*ids->nb[1] + iex*ids->nb[0]); + assert(expert_to_use >= 0 && expert_to_use < ne02); + if (expert_to_use == i02) { + ids_from_sorted_host[i12*n_expert_used + iex] = ids_to_sorted_host.size(); + ids_to_sorted_host.push_back(i12*ne11 + iex % ne11); + tokens_per_expert[i02]++; + break; } - - num_src1_rows++; } } + } + GGML_ASSERT(ids_to_sorted_host.size() == size_t(ne_get_rows)); - if (num_src1_rows == 0) { + ids_to_sorted_host.insert(ids_to_sorted_host.end(), ids_from_sorted_host.begin(), ids_from_sorted_host.end()); + + CUDA_CHECK(cudaMemcpyAsync(ids_buf_dev.ptr, ids_to_sorted_host.data(), 2*ne_get_rows*sizeof(int32_t), cudaMemcpyHostToDevice, stream)); + CUDA_CHECK(cudaStreamSynchronize(stream)); + + const int32_t * ids_to_sorted = ids_buf_dev.ptr + 0*ne_get_rows; + const int32_t * ids_from_sorted = ids_buf_dev.ptr + 1*ne_get_rows; + + get_rows_cuda(src1->data, src1->type, ids_to_sorted, src1_sorted.ptr, type_src1_sorted, + ne10, nb11, nb12, nb13, + ne_get_rows, 1, 1, sizeof(int32_t), ne_get_rows*sizeof(int32_t), ne_get_rows*sizeof(int32_t), + ne10*ts_src1_sorted, ne_get_rows*ne10*ts_src1_sorted, ne_get_rows*ne10*ts_src1_sorted, stream); + CUDA_CHECK(cudaGetLastError()); + + char * src1_data_cur = (char *) src1_sorted.ptr; + char * dst_data_cur = (char *) dst_sorted.ptr; + for (int64_t i02 = 0; i02 < ne02; ++i02) { + if (tokens_per_expert[i02] == 0) { continue; } - ggml_cuda_pool_alloc dev_cur_src1_row(ctx.pool(), 1); - ggml_cuda_pool_alloc dev_row_mapping(ctx.pool(), num_src1_rows); - CUDA_CHECK(cudaMemsetAsync(dev_cur_src1_row.get(), 0, sizeof(int), stream)); + ggml_tensor src0_slice = *src0; + src0_slice.ne[2] = 1; + src0_slice.nb[3] = src0_slice.nb[2]; + src0_slice.op = GGML_OP_VIEW; + src0_slice.view_src = dst->src[0]; // non-const pointer to src0 + src0_slice.data = (char *) src0->data + i02*nb02; - { - dim3 block_dims(std::min((unsigned int)ne10, 768u)); - dim3 grid_dims(ids->ne[1], n_ids); - k_copy_src1_to_contiguous<<>>( - src1_original, src1_contiguous.get(), - dev_cur_src1_row.get(), dev_row_mapping.get(), - ids_dev, i02, ids->nb[1], ids->nb[0], - ne11, ne10, - nb11, nb12); - CUDA_CHECK(cudaGetLastError()); - } + ggml_tensor src1_slice; + memset(&src1_slice, 0, sizeof(src1_slice)); + src1_slice.buffer = src1->buffer; + src1_slice.type = type_src1_sorted; + src1_slice.ne[0] = ne10; + src1_slice.ne[1] = tokens_per_expert[i02]; + src1_slice.ne[2] = 1; + src1_slice.ne[3] = 1; + src1_slice.nb[0] = ts_src1_sorted; + src1_slice.nb[1] = src1_slice.ne[0] * src1_slice.nb[0]; + src1_slice.nb[2] = src1_slice.ne[1] * src1_slice.nb[1]; + src1_slice.nb[3] = src1_slice.ne[2] * src1_slice.nb[2]; + src1_slice.data = src1_data_cur; - src0_row.data = src0_original + i02*nb02; + ggml_tensor dst_slice; + memset(&dst_slice, 0, sizeof(dst_slice)); + dst_slice.buffer = dst->buffer; + dst_slice.type = type_dst_sorted; + dst_slice.ne[0] = ne0; + dst_slice.ne[1] = tokens_per_expert[i02]; + dst_slice.ne[2] = 1; + dst_slice.ne[3] = 1; + dst_slice.nb[0] = ts_dst_sorted; + dst_slice.nb[1] = dst_slice.ne[0] * dst_slice.nb[0]; + dst_slice.nb[2] = dst_slice.ne[1] * dst_slice.nb[1]; + dst_slice.nb[3] = dst_slice.ne[2] * dst_slice.nb[2]; + dst_slice.data = dst_data_cur; - GGML_ASSERT(nb11 == sizeof(float)*ne10); - GGML_ASSERT(nb1 == sizeof(float)*ne0); + ggml_cuda_mul_mat(ctx, &src0_slice, &src1_slice, &dst_slice); + CUDA_CHECK(cudaGetLastError()); - src1_row.ne[1] = num_src1_rows; - src1_row.nb[1] = nb11; - src1_row.nb[2] = num_src1_rows*nb11; - src1_row.nb[3] = num_src1_rows*nb11; - - dst_row.ne[1] = num_src1_rows; - dst_row.nb[1] = nb1; - dst_row.nb[2] = num_src1_rows*nb1; - dst_row.nb[3] = num_src1_rows*nb1; - - ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row); - - { - dim3 block_dims(std::min((unsigned int)ne0, 768u)); - dim3 grid_dims(num_src1_rows); - k_copy_dst_from_contiguous<<>>( - dst_original, dst_contiguous.get(), - dev_row_mapping.get(), - ne0, - nb1, nb2); - CUDA_CHECK(cudaGetLastError()); - } + src1_data_cur += src1_slice.nb[2]; + dst_data_cur += dst_slice.nb[2]; } + + get_rows_cuda(dst_sorted.ptr, type_dst_sorted, ids_from_sorted, dst->data, dst->type, + ne0, ne0*ts_dst_sorted, ne_get_rows*ne0*ts_dst_sorted, ne_get_rows*ne0*ts_dst_sorted, + ne_get_rows, 1, 1, sizeof(int32_t), ne_get_rows*sizeof(int32_t), ne_get_rows*sizeof(int32_t), + nb1, nb2, nb3, stream); } static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct ggml_tensor * dst) { @@ -3228,16 +3221,16 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g return false; #endif // FLASH_ATTN_AVAILABLE if (op->src[1]->ne[0] != op->src[2]->ne[0]) { - // different head sizes of K and V are not supported yet - return false; + const int cc = ggml_cuda_info().devices[dev_ctx->device].cc; + if (!new_mma_available(cc) || cc < GGML_CUDA_CC_AMPERE) { + return false; + } + const int gqa_ratio = op->src[0]->ne[2] / op->src[1]->ne[2]; + return op->src[1]->ne[0] == 576 && op->src[2]->ne[0] == 512 && op->src[3] && gqa_ratio % 16 == 0; } if (op->src[0]->ne[0] == 192) { return false; } - if (op->src[0]->ne[0] == 576) { - // DeepSeek MLA - return false; - } if (op->src[0]->ne[3] != 1) { return false; } diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu index b36b43d541..e1cf843de1 100644 --- a/ggml/src/ggml-cuda/mmq.cu +++ b/ggml/src/ggml-cuda/mmq.cu @@ -1,37 +1,10 @@ #include "mmq.cuh" +#include "quantize.cuh" -void ggml_cuda_op_mul_mat_q( - ggml_backend_cuda_context & ctx, - const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, - const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, - const int64_t src1_padded_row_size, cudaStream_t stream) { +#include - const int64_t ne00 = src0->ne[0]; - - const int64_t ne10 = src1->ne[0]; - const int64_t ne11 = src1->ne[1]; - GGML_ASSERT(ne10 % QK8_1 == 0); - - const int64_t ne0 = dst->ne[0]; - - const int64_t row_diff = row_high - row_low; - const int64_t stride00 = ne00 / ggml_blck_size(src0->type); - - int id = ggml_cuda_get_device(); - const int cc = ggml_cuda_info().devices[id].cc; - - // the main device has a larger memory buffer to hold the results from all GPUs - // nrows_dst == nrows of the matrix that the kernel writes into - const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff; - - // The stream-k decomposition is only faster for recent NVIDIA GPUs. - // Also its fixup needs to allocate a temporary buffer in the memory pool. - // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer. - const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && - ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11; - const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k}; - - switch (src0->type) { +static void ggml_cuda_mul_mat_q_switch_type(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) { + switch (args.type_x) { case GGML_TYPE_Q4_0: mul_mat_q_case(ctx, args, stream); break; @@ -90,10 +63,206 @@ void ggml_cuda_op_mul_mat_q( GGML_ABORT("fatal error"); break; } +} + +void ggml_cuda_mul_mat_q( + 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); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + GGML_ASSERT(!ids || ids->type == GGML_TYPE_I32); // Optional, used for batched GGML_MUL_MAT_ID. + + GGML_TENSOR_BINARY_OP_LOCALS; + + cudaStream_t stream = ctx.stream(); + const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc; + + const size_t ts_src0 = ggml_type_size(src0->type); + const size_t ts_src1 = ggml_type_size(src1->type); + const size_t ts_dst = ggml_type_size(dst->type); + + GGML_ASSERT( nb00 == ts_src0); + GGML_ASSERT( nb10 == ts_src1); + GGML_ASSERT( nb0 == ts_dst); + GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type)); + + const char * src0_d = (const char *) src0->data; + const float * src1_d = (const float *) src1->data; + float * dst_d = (float *) dst->data; + + // If src0 is a temporary compute buffer, clear any potential padding. + if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) { + const size_t size_data = ggml_nbytes(src0); + const size_t size_alloc = ggml_backend_buffer_get_alloc_size(src0->buffer, src0); + if (size_alloc > size_data) { + GGML_ASSERT(ggml_is_contiguously_allocated(src0)); + GGML_ASSERT(!src0->view_src); + CUDA_CHECK(cudaMemsetAsync((char *) src0->data + size_data, 0, size_alloc - size_data, stream)); + } + } + + const int64_t ne10_padded = GGML_PAD(ne10, MATRIX_ROW_PADDING); + + const int64_t s01 = src0->nb[1] / ts_src0; + const int64_t s1 = dst->nb[1] / ts_dst; + const int64_t s02 = src0->nb[2] / ts_src0; + const int64_t s2 = dst->nb[2] / ts_dst; + const int64_t s03 = src0->nb[3] / ts_src0; + const int64_t s3 = dst->nb[3] / ts_dst; + + const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA; + + if (!ids) { + const size_t nbytes_src1_q8_1 = ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1 + + get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq); + ggml_cuda_pool_alloc src1_q8_1(ctx.pool(), nbytes_src1_q8_1); + + { + const int64_t s11 = src1->nb[1] / ts_src1; + const int64_t s12 = src1->nb[2] / ts_src1; + const int64_t s13 = src1->nb[3] / ts_src1; + quantize_mmq_q8_1_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type, + ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream); + } + + const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int)); + const int64_t s13 = ne12*s12; + + const mmq_args args = { + src0_d, src0->type, (const int *) src1_q8_1.ptr, nullptr, nullptr, dst_d, + ne00, ne01, ne1, s01, ne11, s1, + ne02, ne12, s02, s12, s2, + ne03, ne13, s03, s13, s3, + use_stream_k}; + ggml_cuda_mul_mat_q_switch_type(ctx, args, stream); + return; + } + + GGML_ASSERT(ne13 == 1); + GGML_ASSERT(nb12 % nb11 == 0); + GGML_ASSERT(nb2 % nb1 == 0); + + const int64_t n_expert_used = ids->ne[0]; + const int64_t ne_get_rows = ne12 * n_expert_used; + + std::vector ids_host(ggml_nbytes(ids)); + std::vector ids_src1_host; + ids_src1_host.reserve(ne_get_rows); + std::vector ids_dst_host; + ids_dst_host.reserve(ne_get_rows); + std::vector tokens_per_expert_host(ne02); + std::vector expert_bounds_host(ne02 + 1); + ggml_cuda_pool_alloc ids_buf_dev(ctx.pool()); + + CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids->data, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream)); + CUDA_CHECK(cudaStreamSynchronize(stream)); + + for (int64_t i02 = 0; i02 < ne02; ++i02) { // expert matrices + for (int64_t i12 = 0; i12 < ne12; ++i12) { // tokens + for (int64_t iex = 0; iex < n_expert_used; ++iex) { + const int32_t expert_to_use = *(const int32_t *)(ids_host.data() + i12*ids->nb[1] + iex*ids->nb[0]); + assert(expert_to_use >= 0 && expert_to_use < ne02); + if (expert_to_use == i02) { + ids_src1_host.push_back(i12*(nb12/nb11) + iex % ne11); + ids_dst_host.push_back(i12*ne1 + iex); + tokens_per_expert_host[i02]++; + break; + } + } + } + } + + int32_t cumsum = 0; + for (int64_t i = 0; i < ne02; ++i) { + expert_bounds_host[i] = cumsum; + cumsum += tokens_per_expert_host[i]; + } + expert_bounds_host[ne02] = cumsum; + + std::vector ids_buf_host; + ids_buf_host.reserve(ids_src1_host.size() + ids_dst_host.size() + expert_bounds_host.size()); + ids_buf_host.insert(ids_buf_host.end(), ids_src1_host.begin(), ids_src1_host.end()); + ids_buf_host.insert(ids_buf_host.end(), ids_dst_host.begin(), ids_dst_host.end()); + ids_buf_host.insert(ids_buf_host.end(), expert_bounds_host.begin(), expert_bounds_host.end()); + ids_buf_dev.alloc(ids_buf_host.size() + get_mmq_x_max_host(cc)); // Expert bounds are padded on device. + CUDA_CHECK(cudaMemcpyAsync(ids_buf_dev.ptr, ids_buf_host.data(), ids_buf_host.size()*sizeof(int32_t), cudaMemcpyHostToDevice, stream)); + CUDA_CHECK(cudaStreamSynchronize(stream)); + + const int32_t * ids_src1_dev = ids_buf_dev.ptr; + const int32_t * ids_dst_dev = ids_src1_dev + ids_src1_host.size(); + const int32_t * expert_bounds_dev = ids_dst_dev + ids_dst_host.size(); + + const size_t nbytes_src1_q8_1 = ne12*n_expert_used*ne10_padded * sizeof(block_q8_1)/QK8_1 + + get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq); + ggml_cuda_pool_alloc src1_q8_1(ctx.pool(), nbytes_src1_q8_1); + + const int64_t ne11_flat = ne12*n_expert_used; + const int64_t ne12_flat = 1; + const int64_t ne13_flat = 1; + + { + const int64_t s11 = src1->nb[1] / ts_src1; + const int64_t s12 = src1->nb[2] / ts_src1; + const int64_t s13 = src1->nb[2] / ts_src1; + quantize_mmq_q8_1_cuda(src1_d, ids_src1_dev, src1_q8_1.get(), src0->type, + ne10, s11, s12, s13, ne10_padded, ne11_flat, ne12_flat, ne13_flat, stream); + } + + const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int)); + const int64_t s13 = ne12*s12; + + // Note that ne02 is used instead of ne12 because the number of y channels determines the z dimension of the CUDA grid. + const mmq_args args = { + src0_d, src0->type, (const int *) src1_q8_1.ptr, ids_dst_dev, expert_bounds_dev, dst_d, + ne00, ne01, ne_get_rows, s01, ne_get_rows, s1, + ne02, ne02, s02, s12, s2, + ne03, ne13, s03, s13, s3, + use_stream_k}; + + ggml_cuda_mul_mat_q_switch_type(ctx, args, stream); +} + +void ggml_cuda_op_mul_mat_q( + ggml_backend_cuda_context & ctx, + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, + const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols, + const int64_t src1_padded_row_size, cudaStream_t stream) { + + const int64_t ne00 = src0->ne[0]; + + const int64_t ne10 = src1->ne[0]; + const int64_t ne11 = src1->ne[1]; + GGML_ASSERT(ne10 % QK8_1 == 0); + + const int64_t ne0 = dst->ne[0]; + + const int64_t row_diff = row_high - row_low; + const int64_t stride01 = ne00 / ggml_blck_size(src0->type); + + const int id = ggml_cuda_get_device(); + const int cc = ggml_cuda_info().devices[id].cc; + + // the main device has a larger memory buffer to hold the results from all GPUs + // nrows_dst == nrows of the matrix that the kernel writes into + const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff; + + // The stream-k decomposition is only faster for recent NVIDIA GPUs. + // Also its fixup needs to allocate a temporary buffer in the memory pool. + // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer. + const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && + ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11; + const mmq_args args = { + src0_dd_i, src0->type, (const int *) src1_ddq_i, nullptr, nullptr, dst_dd_i, + ne00, row_diff, src1_ncols, stride01, ne11, nrows_dst, + 1, 1, 0, 0, 0, + 1, 1, 0, 0, 0, + use_stream_k}; + + ggml_cuda_mul_mat_q_switch_type(ctx, args, stream); GGML_UNUSED(src1); GGML_UNUSED(dst); GGML_UNUSED(src1_ddf_i); + GGML_UNUSED(src1_padded_row_size); } bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) { diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh index 3cb2015520..80baf459c1 100644 --- a/ggml/src/ggml-cuda/mmq.cuh +++ b/ggml/src/ggml-cuda/mmq.cuh @@ -13,9 +13,10 @@ using namespace ggml_cuda_mma; #define MMQ_ITER_K 256 #define MMQ_NWARPS 8 -typedef void (*load_tiles_mmq_t)(const char * __restrict__ x, int * x_tile, const int & kbx0, const int & i_max, const int & stride); -typedef void (*vec_dot_mmq_t)(const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00); -typedef void (*mmq_write_back_t)(const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max); +typedef void (*load_tiles_mmq_t)(const char * __restrict__ x, int * x_tile, const int kbx0, const int i_max, const int stride); +typedef void (*vec_dot_mmq_t)(const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00); +typedef void (*mmq_write_back_t)(const float * __restrict__ sum, const int32_t * __restrict__ get_rows_to_sorted, + float * __restrict__ dst, const int stride, const int i_max, const int j_max); enum mmq_q8_1_ds_layout { MMQ_Q8_1_DS_LAYOUT_D4, @@ -233,7 +234,7 @@ static constexpr __device__ int mmq_get_granularity_device(const int /* mmq_x */ // ------------------------------------------------------------ template static __device__ __forceinline__ void load_tiles_q4_0( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -289,7 +290,7 @@ template static __device__ __forceinlin template static __device__ __forceinline__ void vec_dot_q4_0_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y); const int * x_qs = (const int *) x; @@ -328,7 +329,7 @@ static __device__ __forceinline__ void vec_dot_q4_0_q8_1_dp4a( } template static __device__ __forceinline__ void load_tiles_q4_1( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -384,7 +385,7 @@ template static __device__ __forceinlin template static __device__ __forceinline__ void vec_dot_q4_1_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, mmq_y); const int * x_qs = (const int *) x; @@ -423,7 +424,7 @@ static __device__ __forceinline__ void vec_dot_q4_1_q8_1_dp4a( } template static __device__ __forceinline__ void load_tiles_q5_0( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -495,7 +496,7 @@ template static __device__ __forceinlin } template static __device__ __forceinline__ void load_tiles_q5_1( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -565,7 +566,7 @@ template static __device__ __forceinlin } template static __device__ __forceinline__ void load_tiles_q8_0( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -621,7 +622,7 @@ template static __device__ __forceinlin template static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y); const int * x_qs = (const int *) x; @@ -651,7 +652,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a( template static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { typedef tile<16, 8, int> tile_A; typedef tile< 8, 8, int> tile_B; @@ -732,7 +733,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma( template static __device__ __forceinline__ void vec_dot_q8_1_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, mmq_y); const int * x_qs = (const int *) x; @@ -762,7 +763,7 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_dp4a( template static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { typedef tile<16, 8, int> tile_A; typedef tile< 8, 8, int> tile_B; @@ -839,7 +840,7 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma( template static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { constexpr tile_x_sizes txs = MMQ_DP4A_TXS_Q8_0_16; const int * x_qs = (const int *) x; @@ -871,7 +872,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_dp4a( template static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { #ifdef NEW_MMA_AVAILABLE typedef tile<16, 4, int> tile_A; @@ -955,7 +956,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma( } template static __device__ __forceinline__ void load_tiles_q2_K( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -1011,7 +1012,7 @@ template static __device__ __forceinlin template static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, mmq_y); const int * x_qs = (const int *) x; @@ -1074,7 +1075,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a( template static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { #ifdef NEW_MMA_AVAILABLE typedef tile<16, 4, int> tile_A; @@ -1201,7 +1202,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma( } template static __device__ __forceinline__ void load_tiles_q3_K( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -1298,7 +1299,7 @@ template static __device__ __forceinlin template static __device__ __forceinline__ void vec_dot_q3_K_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, mmq_y); const int * x_qs = (const int *) x; @@ -1340,7 +1341,7 @@ static __device__ __forceinline__ int unpack_scales_q45_K(const int * scales, co } template static __device__ __forceinline__ void load_tiles_q4_K( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -1437,7 +1438,7 @@ template static __device__ __forceinlin template static __device__ __forceinline__ void vec_dot_q4_K_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, mmq_y); const int * x_qs = (const int *) x; @@ -1469,7 +1470,7 @@ static __device__ __forceinline__ void vec_dot_q4_K_q8_1_dp4a( } template static __device__ __forceinline__ void load_tiles_q5_K( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -1578,7 +1579,7 @@ template static __device__ __forceinlin template static __device__ __forceinline__ void vec_dot_q5_K_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, mmq_y); const int * x_qs = (const int *) x; @@ -1610,7 +1611,7 @@ static __device__ __forceinline__ void vec_dot_q5_K_q8_1_dp4a( } template static __device__ __forceinline__ void load_tiles_q6_K( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -1693,7 +1694,7 @@ template static __device__ __forceinlin template static __device__ __forceinline__ void vec_dot_q6_K_q8_1_dp4a( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, mmq_y); const int * x_qs = (const int *) x; @@ -1726,7 +1727,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_dp4a( template static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma( - const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { + const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { #ifdef NEW_MMA_AVAILABLE typedef tile<16, 4, int> tile_A; @@ -1835,7 +1836,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma( } template static __device__ __forceinline__ void load_tiles_iq4_nl( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -1893,7 +1894,7 @@ template static __device__ __forceinlin } template static __device__ __forceinline__ void load_tiles_iq2_xxs( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -1951,7 +1952,7 @@ template static __device__ __forceinlin } template static __device__ __forceinline__ void load_tiles_iq2_xs( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -2007,7 +2008,7 @@ template static __device__ __forceinlin } template static __device__ __forceinline__ void load_tiles_iq2_s( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -2070,7 +2071,7 @@ template static __device__ __forceinlin } template static __device__ __forceinline__ void load_tiles_iq3_xxs( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -2126,7 +2127,7 @@ template static __device__ __forceinlin } template static __device__ __forceinline__ void load_tiles_iq3_s( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -2189,7 +2190,7 @@ template static __device__ __forceinlin } template static __device__ __forceinline__ void load_tiles_iq1_s( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -2245,7 +2246,7 @@ template static __device__ __forceinlin } template static __device__ __forceinline__ void load_tiles_iq4_xs( - const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) { + const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { #ifdef NEW_MMA_AVAILABLE int * x_qs = (int *) x_tile; @@ -2306,8 +2307,8 @@ template static __device__ __forceinlin template static __device__ __forceinline__ void mmq_write_back_dp4a( - const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max) { - + const float * __restrict__ sum, const int32_t * __restrict__ ids_dst, float * __restrict__ dst, + const int stride, const int i_max, const int j_max) { #pragma unroll for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { const int j = j0 + threadIdx.y; @@ -2324,15 +2325,15 @@ static __device__ __forceinline__ void mmq_write_back_dp4a( continue; } - dst[j*stride + i] = sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE]; + dst[ids_dst[j]*stride + i] = sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE]; } } } template static __device__ __forceinline__ void mmq_write_back_mma( - const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max) { - + const float * __restrict__ sum, const int * __restrict__ ids_dst, float * __restrict__ dst, + const int stride, const int i_max, const int j_max) { typedef tile<16, 8, int> tile_C; constexpr int granularity = mmq_get_granularity_device(mmq_x); @@ -2362,7 +2363,7 @@ static __device__ __forceinline__ void mmq_write_back_mma( continue; } - dst[j*stride + i] = sum[(j0/tile_C::J + n)*tile_C::ne + l]; + dst[ids_dst[j]*stride + i] = sum[(j0/tile_C::J + n)*tile_C::ne + l]; } } } @@ -2518,17 +2519,18 @@ struct mmq_type_traits { }; template -static __device__ void mul_mat_q_process_tile( - const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup, - const int & ne00, const int & ne01, const int & stride01, const int & ne10, const int & ne11, const int & stride11, const int & ne0, - const int & it, const int & jt, const int & kb0_start, const int & kb0_stop) { +static __device__ __forceinline__ void mul_mat_q_process_tile( + const char * __restrict__ x, const int offset_x, const int * __restrict__ y, + const int * __restrict__ ids_dst, float * __restrict__ dst, float * __restrict__ tmp_fixup, + const int stride_row_x, const int ncols_y, const int stride_col_dst, + const int tile_x_max_i, const int tile_y_max_j, const int kb0_start, const int kb0_stop) { constexpr int qk = ggml_cuda_type_traits::qk; constexpr int mmq_y = get_mmq_y_device(); constexpr load_tiles_mmq_t load_tiles = mmq_type_traits::load_tiles; - extern __shared__ char data_mul_mat_q[]; - int * tile_y = (int *) data_mul_mat_q; + extern __shared__ int data_mul_mat_q[]; + int * tile_y = data_mul_mat_q + mmq_x; int * tile_x = tile_y + GGML_PAD(mmq_x*(WARP_SIZE + WARP_SIZE/QI8_1), nwarps*WARP_SIZE); #ifdef NEW_MMA_AVAILABLE @@ -2543,16 +2545,11 @@ static __device__ void mul_mat_q_process_tile( float sum[mmq_x*mmq_y / (nwarps*WARP_SIZE)] = {0.0f}; - const int tile_x_max_i = ne01 - it*mmq_y - 1; - const int tile_y_max_j = ne11 - jt*mmq_x - 1; - - const int * y = (const int *) yc + jt*(mmq_x*sizeof(block_q8_1_mmq)/sizeof(int)); - for (int kb0 = kb0_start; kb0 < kb0_stop; kb0 += blocks_per_iter) { - load_tiles(x, tile_x, stride01*it*mmq_y + kb0, tile_x_max_i, stride01); + load_tiles(x, tile_x, offset_x + kb0, tile_x_max_i, stride_row_x); { - const int * by0 = y + stride11*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 0*sizeof(block_q8_1_mmq)/sizeof(int)); + const int * by0 = y + ncols_y*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 0*sizeof(block_q8_1_mmq)/sizeof(int)); #pragma unroll for (int l0 = 0; l0 < mmq_x*MMQ_TILE_Y_K; l0 += nwarps*WARP_SIZE) { int l = l0 + threadIdx.y*WARP_SIZE + threadIdx.x; @@ -2568,7 +2565,7 @@ static __device__ void mul_mat_q_process_tile( __syncthreads(); { - const int * by0 = y + stride11*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 1*sizeof(block_q8_1_mmq)/sizeof(int)); + const int * by0 = y + ncols_y*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 1*sizeof(block_q8_1_mmq)/sizeof(int)); #pragma unroll for (int l0 = 0; l0 < mmq_x*MMQ_TILE_Y_K; l0 += nwarps*WARP_SIZE) { int l = l0 + threadIdx.y*WARP_SIZE + threadIdx.x; @@ -2585,12 +2582,10 @@ static __device__ void mul_mat_q_process_tile( } if (fixup) { - write_back(sum, tmp_fixup + blockIdx.x*(mmq_x*mmq_y), mmq_y, mmq_y, mmq_x); + write_back(sum, ids_dst, tmp_fixup + blockIdx.x*(mmq_x*mmq_y), mmq_y, mmq_y, mmq_x); } else { - write_back(sum, dst + jt*mmq_x*ne0 + it*mmq_y, ne0, tile_x_max_i, tile_y_max_j); + write_back(sum, ids_dst, dst, stride_col_dst, tile_x_max_i, tile_y_max_j); } - - GGML_UNUSED(ne00); GGML_UNUSED(ne10); } @@ -2609,8 +2604,11 @@ template #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) static __global__ void mul_mat_q( - const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup, - const int ne00, const int ne01, const int stride01, const int ne10, const int ne11, const int stride11, const int ne0) { + const char * __restrict__ x, const int * __restrict__ y, const int32_t * __restrict__ ids_dst, + const int32_t * __restrict__ expert_bounds, float * __restrict__ dst, float * __restrict__ tmp_fixup, + const int ncols_x, const int nrows_x, const int ncols_dst, const int stride_row_x, const int ncols_y, const int stride_col_dst, + const int channel_ratio, const int nchannels_y, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst, + const int sample_ratio, const int nsamples_y, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) { // Skip unused template specializations for faster compilation: if (mmq_x > get_mmq_x_max_device() || mmq_x % mmq_get_granularity_device(mmq_x) != 0) { @@ -2621,26 +2619,88 @@ static __global__ void mul_mat_q( constexpr int qk = ggml_cuda_type_traits::qk; constexpr int mmq_y = get_mmq_y_device(); + const int ntx = (ncols_dst + mmq_x - 1) / mmq_x; // Number of tiles x + const int nty = (nrows_x + mmq_y - 1) / mmq_y; // Number of tiles y + + // Initialize the ids for writing back data with just the index. + // For regular matrix multiplications this is never changed. + // For MoE the correct indices are loaded from ids_dst. + extern __shared__ int ids_dst_shared[]; // Stored at beginning of shared memory. +#pragma unroll + for (int j0 = 0; j0 < mmq_x; j0 += nwarps*WARP_SIZE) { + const int j = j0 + threadIdx.y*WARP_SIZE + threadIdx.x; + + if (j0 + nwarps*WARP_SIZE > mmq_x && j >= mmq_x) { + break; + } + + ids_dst_shared[j] = j; + } + __syncthreads(); + // On AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead: #if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA { + const int wt = blockIdx.z / nchannels_y; + const int zt = blockIdx.z - wt*nchannels_y; + const int jt = blockIdx.y; + const int it = blockIdx.x; + + // Defaults for regular matrix multiplication: + int col_low = 0; + int col_high = ncols_dst; + int col_diff = ncols_dst; + int offset_y = wt*stride_sample_y + zt*stride_channel_y; + int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst; + + if (ids_dst) { + col_low = expert_bounds[zt + 0]; + col_high = expert_bounds[zt + 1]; + col_diff = col_high - col_low; + + offset_y = 0; + offset_dst = 0; + + if (jt*mmq_x >= col_diff) { + return; + } + + // __syncthreads(); // There is no previous tile that could cause a race condition. +#pragma unroll + for (int j0 = 0; j0 < mmq_x; j0 += nwarps*WARP_SIZE) { + const int j = j0 + threadIdx.y*WARP_SIZE + threadIdx.x; + + if (j0 + nwarps*WARP_SIZE > mmq_x && j >= mmq_x) { + break; + } + + ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j]; + } + __syncthreads(); + } + + offset_y += (col_low + jt*mmq_x)*(sizeof(block_q8_1_mmq)/sizeof(int)); + offset_dst += it*mmq_y; + + const int tile_x_max_i = nrows_x - it*mmq_y - 1; + const int tile_y_max_j = col_diff - jt*mmq_x - 1; + + const int offset_x = (wt/sample_ratio)*stride_sample_x + (zt/channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x; + constexpr bool fixup = false; mul_mat_q_process_tile - (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0, - blockIdx.x, blockIdx.y, 0, ne00/qk); + (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, stride_row_x, ncols_y, stride_col_dst, + tile_x_max_i, tile_y_max_j, 0, ncols_x/qk); return; } #endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA - const int64_t blocks_per_ne00 = ne00 / qk; + const int64_t blocks_per_ne00 = ncols_x / qk; constexpr int blocks_per_iter = MMQ_ITER_K / qk; - const int ntx = (ne11 + mmq_x - 1) / mmq_x; // Number of tiles x - const int nty = (ne01 + mmq_y - 1) / mmq_y; // Number of tiles y - // kbc == k block continuous, current index in continuous ijk space. - int64_t kbc = (int64_t) blockIdx.x *blocks_per_ne00*ntx*nty / gridDim.x; - int64_t kbc_stop = (int64_t)(blockIdx.x + 1)*blocks_per_ne00*ntx*nty / gridDim.x; + int64_t kbc = (int64_t) blockIdx.x *nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x; + int64_t kbc_stop = (int64_t)(blockIdx.x + 1)*nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x; kbc -= (kbc % blocks_per_ne00) % blocks_per_iter; kbc_stop -= (kbc_stop % blocks_per_ne00) % blocks_per_iter; @@ -2649,13 +2709,66 @@ static __global__ void mul_mat_q( int kb0_start = kbc % blocks_per_ne00; int kb0_stop = min(blocks_per_ne00, kb0_start + kbc_stop - kbc); while (kbc < kbc_stop && kb0_stop == blocks_per_ne00) { - const int jt = kbc / (blocks_per_ne00*nty); // j index of current tile. - const int it = (kbc - jt*(blocks_per_ne00*nty)) / blocks_per_ne00; // i index of current tile. + int tmp = kbc; + const int it = tmp / (nsamples_y*nchannels_y*ntx*blocks_per_ne00); + tmp -= it * (nsamples_y*nchannels_y*ntx*blocks_per_ne00); + const int wt = tmp / (nchannels_y*ntx*blocks_per_ne00); + tmp -= wt * (nchannels_y*ntx*blocks_per_ne00); + const int zt = tmp / (ntx*blocks_per_ne00); + tmp -= zt * (ntx*blocks_per_ne00); + const int jt = tmp / blocks_per_ne00; + + // Defaults for regular matrix multiplication: + int col_low = 0; + int col_high = ncols_dst; + int col_diff = ncols_dst; + int offset_y = wt*stride_sample_y + zt*stride_channel_y; + int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst; + + if (ids_dst) { + col_low = expert_bounds[zt + 0]; + col_high = expert_bounds[zt + 1]; + col_diff = col_high - col_low; + + offset_y = 0; + offset_dst = 0; + + if (jt*mmq_x >= col_diff) { + kbc += blocks_per_ne00; + kbc -= kbc % blocks_per_ne00; + + kb0_start = 0; + kb0_stop = min(blocks_per_ne00, kbc_stop - kbc); + + continue; + } + + __syncthreads(); +#pragma unroll + for (int j0 = 0; j0 < mmq_x; j0 += nwarps*WARP_SIZE) { + const int j = j0 + threadIdx.y*WARP_SIZE + threadIdx.x; + + if (j0 + nwarps*WARP_SIZE > mmq_x && j >= mmq_x) { + break; + } + + ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j]; + } + __syncthreads(); + } + + offset_y += (col_low + jt*mmq_x)*(sizeof(block_q8_1_mmq)/sizeof(int)); + offset_dst += it*mmq_y; + + const int tile_x_max_i = nrows_x - it*mmq_y - 1; + const int tile_y_max_j = col_diff - jt*mmq_x - 1; + + const int offset_x = (wt/sample_ratio)*stride_sample_x + (zt/channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x; constexpr bool fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer. mul_mat_q_process_tile - (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0, - it, jt, kb0_start, kb0_stop); + (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, stride_row_x, ncols_y, stride_col_dst, + tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop); kbc += blocks_per_ne00; kbc -= kbc % blocks_per_ne00; @@ -2668,55 +2781,108 @@ static __global__ void mul_mat_q( return; } - const int jt = kbc / (blocks_per_ne00*nty); - const int it = (kbc - jt*(blocks_per_ne00*nty)) / blocks_per_ne00; + int tmp = kbc; + const int it = tmp / (nsamples_y*nchannels_y*ntx*blocks_per_ne00); + tmp -= it * (nsamples_y*nchannels_y*ntx*blocks_per_ne00); + const int wt = tmp / (nchannels_y*ntx*blocks_per_ne00); + tmp -= wt * (nchannels_y*ntx*blocks_per_ne00); + const int zt = tmp / (ntx*blocks_per_ne00); + tmp -= zt * (ntx*blocks_per_ne00); + const int jt = tmp / blocks_per_ne00; + + // Defaults for regular matrix multiplication: + int col_low = 0; + int col_high = ncols_dst; + int col_diff = ncols_dst; + int offset_y = wt*stride_sample_y + zt*stride_channel_y; + int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst; + + if (ids_dst) { + col_low = expert_bounds[zt + 0]; + col_high = expert_bounds[zt + 1]; + col_diff = col_high - col_low; + + offset_y = 0; + offset_dst = 0; + + if (jt*mmq_x >= col_diff) { + return; + } + + // The memory layout for the fixup buffer is always contiguous, therefore reset ids: + __syncthreads(); +#pragma unroll + for (int j0 = 0; j0 < mmq_x; j0 += nwarps*WARP_SIZE) { + const int j = j0 + threadIdx.y*WARP_SIZE + threadIdx.x; + + if (j0 + nwarps*WARP_SIZE > mmq_x && j >= mmq_x) { + break; + } + + ids_dst_shared[j] = j; + } + __syncthreads(); + } + + offset_y += (col_low + jt*mmq_x)*(sizeof(block_q8_1_mmq)/sizeof(int)); + offset_dst += it*mmq_y; + + const int tile_x_max_i = nrows_x - it*mmq_y - 1; + const int tile_y_max_j = col_diff - jt*mmq_x - 1; + + const int offset_x = (wt/sample_ratio)*stride_sample_x + (zt/channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x; constexpr bool fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks. mul_mat_q_process_tile - (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0, - it, jt, kb0_start, kb0_stop); + (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, stride_row_x, ncols_y, stride_col_dst, + tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop); } template static __global__ void mul_mat_q_stream_k_fixup( - float * __restrict__ dst, const float * __restrict__ tmp_last_tile, const int ne00, const int ne01, const int ne11, const int ne0, const int block_num_mmq) { - + const int32_t * ids_dst, const int32_t * expert_bounds, float * __restrict__ dst, const float * __restrict__ tmp_last_tile, + const int ncols_x, const int nrows_x, const int ncols_dst, const int stride_col_dst, + const int nchannels_y, const int stride_channel_dst, const int nsamples_y, const int stride_sample_dst) { constexpr int mmq_y = get_mmq_y_device(); constexpr int qk = ggml_cuda_type_traits::qk; constexpr int blocks_per_iter = MMQ_ITER_K / qk; - const int64_t blocks_per_ne00 = ne00 / qk; + const int64_t blocks_per_ne00 = ncols_x / qk; float sum[mmq_x*mmq_y / (nwarps*WARP_SIZE)] = {0.0f}; - const int ntx = (ne11 + mmq_x - 1) / mmq_x; - const int nty = (ne01 + mmq_y - 1) / mmq_y; + const int ntx = (ncols_dst + mmq_x - 1) / mmq_x; + const int nty = (nrows_x + mmq_y - 1) / mmq_y; + + const int bidx0 = blockIdx.x; + + // kbc == k block continuous, current index in continuous ijk space. + int64_t kbc0 = (int64_t) bidx0 *nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x; + int64_t kbc0_stop = (int64_t)(bidx0 + 1)*nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x; + + kbc0 -= (kbc0 % blocks_per_ne00) % blocks_per_iter; + kbc0_stop -= (kbc0_stop % blocks_per_ne00) % blocks_per_iter; + + const bool did_not_have_any_data = kbc0 == kbc0_stop; + const bool wrote_beginning_of_tile = kbc0 % blocks_per_ne00 == 0; + const bool did_not_write_last = kbc0/blocks_per_ne00 == kbc0_stop/blocks_per_ne00 && kbc0_stop % blocks_per_ne00 != 0; + if (did_not_have_any_data || wrote_beginning_of_tile || did_not_write_last) { + return; + } bool any_fixup = false; - const int bidx_start = ((blockIdx.y*nty + blockIdx.x) * block_num_mmq) / (gridDim.y*gridDim.x); - const int bidx_stop = ((blockIdx.y*nty + blockIdx.x + 1) * block_num_mmq + gridDim.y*gridDim.x - 1) / (gridDim.y*gridDim.x); + // Iterate over previous blocks and sum up partial sums written to fixup buffer. + // All CUDA blocks that get here must have a previous block that needs a fixup. + int64_t bidx = bidx0 - 1; + int64_t kbc_stop = kbc0; + while(true) { + int64_t kbc = bidx*nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x; + kbc -= (kbc % blocks_per_ne00) % blocks_per_iter; - int64_t kbc_0; - int64_t kbc_stop_0 = (int64_t) bidx_start*blocks_per_ne00*ntx*nty / block_num_mmq; - - for (int bidx = bidx_start; bidx < bidx_stop; ++bidx) { - kbc_0 = kbc_stop_0; - kbc_stop_0 = (int64_t) (bidx + 1)*blocks_per_ne00*ntx*nty / block_num_mmq; - - const int64_t kbc = kbc_0 - (kbc_0 % blocks_per_ne00) % blocks_per_iter; - const int64_t kbc_stop = kbc_stop_0 - (kbc_stop_0 % blocks_per_ne00) % blocks_per_iter; - - // Skip fixup tile if the MMQ CUDA block never wrote anything to it: - if (kbc == kbc_stop || kbc_stop % blocks_per_ne00 == 0) { - continue; - } - - const int jt = kbc_stop / (blocks_per_ne00*nty); - const int it = (kbc_stop - jt*(blocks_per_ne00*nty)) / blocks_per_ne00; - - // Skip fixup tile if it's unrelated to the output tile assigned to this CUDA block: - if ((unsigned)it != blockIdx.x || (unsigned)jt != blockIdx.y) { + if (kbc == kbc_stop) { // Did not have any data. + bidx--; + kbc_stop = kbc; continue; } @@ -2733,16 +2899,72 @@ static __global__ void mul_mat_q_stream_k_fixup( sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE] += tmp_last_tile[bidx*(mmq_x*mmq_y) + j*mmq_y + i]; } } + + // If this block started in a previous tile we are done and don't need to combine additional partial results. + if (kbc % blocks_per_ne00 == 0 || kbc/blocks_per_ne00 < kbc0/blocks_per_ne00) { + break; + } + bidx--; + kbc_stop = kbc; } if (!any_fixup) { return; } - dst += blockIdx.y*mmq_x*ne0 + blockIdx.x*mmq_y; + int tmp = kbc0; + const int it = tmp / (nsamples_y*nchannels_y*ntx*blocks_per_ne00); + tmp -= it * (nsamples_y*nchannels_y*ntx*blocks_per_ne00); + const int wt = tmp / (nchannels_y*ntx*blocks_per_ne00); + tmp -= wt * (nchannels_y*ntx*blocks_per_ne00); + const int zt = tmp / (ntx*blocks_per_ne00); + tmp -= zt * (ntx*blocks_per_ne00); + const int jt = tmp / blocks_per_ne00; - const int i_max = ne01 - blockIdx.x*mmq_y - 1; - const int j_max = ne11 - blockIdx.y*mmq_x - 1; + if (!ids_dst) { + const int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst + it*mmq_y; + dst += offset_dst; + + const int i_max = nrows_x - it*mmq_y - 1; + const int j_max = ncols_dst - jt*mmq_x - 1; + +#pragma unroll + for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { + const int j = j0 + threadIdx.y; + + if (j > j_max) { + return; + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) { + const int i = i0 + threadIdx.x; + + if (need_check && i > i_max) { + continue; + } + + dst[j*stride_col_dst + i] += sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE]; + } + } + return; + } + + __shared__ int ids_dst_shared[mmq_x]; + const int col_low = expert_bounds[zt + 0]; + const int col_high = expert_bounds[zt + 1]; + const int col_diff = col_high - col_low; + + for (int j = threadIdx.y*WARP_SIZE + threadIdx.x; j < mmq_x; j += nwarps*WARP_SIZE) { + ids_dst_shared[j] = ids_dst[col_low + j]; + } + __syncthreads(); + + const int offset_dst = it*mmq_y; + dst += offset_dst; + + const int i_max = nrows_x - it*mmq_y - 1; + const int j_max = col_diff - jt*mmq_x - 1; #pragma unroll for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { @@ -2760,26 +2982,27 @@ static __global__ void mul_mat_q_stream_k_fixup( continue; } - dst[j*ne0 + i] += sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE]; + dst[ids_dst_shared[j]*stride_col_dst + i] += sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE]; } } } struct mmq_args { - const char * x; const char * y; float * dst; - int64_t ne00; int64_t ne01; int64_t stride01; - int64_t ne10; int64_t ne11; int64_t stride11; - int64_t ne0; + const char * x; ggml_type type_x; const int * y; const int32_t * ids_dst; const int32_t * expert_bounds; float * dst; + int64_t ncols_x; int64_t nrows_x; int64_t ncols_dst; int64_t stride_row_x; int64_t ncols_y; int64_t nrows_dst; + int64_t nchannels_x; int64_t nchannels_y; int64_t stride_channel_x; int64_t stride_channel_y; int64_t stride_channel_dst; + int64_t nsamples_x; int64_t nsamples_y; int64_t stride_sample_x; int64_t stride_sample_y; int64_t stride_sample_dst; bool use_stream_k; }; template -static int mmq_get_shmem(const int mmq_x, const int mmq_y, const int cc) { +static size_t mmq_get_nbytes_shared(const int mmq_x, const int mmq_y, const int cc) { const tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(type, mmq_y); const int mmq_tile_x_k = mmq_get_mma_tile_x_k(type); - const int shmem_x = new_mma_available(cc) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int); - const int shmem_y = mmq_x*sizeof(block_q8_1_mmq); - return shmem_x + GGML_PAD(shmem_y, MMQ_NWARPS*WARP_SIZE*sizeof(int)); + const size_t nbs_ids = mmq_x*sizeof(int); + const size_t nbs_x = new_mma_available(cc) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int); + const size_t nbs_y = mmq_x*sizeof(block_q8_1_mmq); + return nbs_ids + nbs_x + GGML_PAD(nbs_y, MMQ_NWARPS*WARP_SIZE*sizeof(int)); } template @@ -2791,86 +3014,114 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a const dim3 block_dims(WARP_SIZE, MMQ_NWARPS, 1); - const int shmem = mmq_get_shmem(mmq_x, mmq_y, cc); + const int nbytes_shared = mmq_get_nbytes_shared(mmq_x, mmq_y, cc); #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA) - static bool shmem_limit_raised[GGML_CUDA_MAX_DEVICES] = {false}; - if (!shmem_limit_raised[id]) { - CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem)); - CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem)); - shmem_limit_raised[id] = true; + static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false}; + if (!shared_memory_limit_raised[id]) { + CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared)); + CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared)); + shared_memory_limit_raised[id] = true; } #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA) - const int nty = (args.ne01 + mmq_y - 1) / mmq_y; - const int ntx = (args.ne11 + mmq_x - 1) / mmq_x; - const dim3 block_nums_xy_tiling(nty, ntx, 1); + const int nty = (args.nrows_x + mmq_y - 1) / mmq_y; + const int ntx = (args.ncols_dst + mmq_x - 1) / mmq_x; + const int ntzw = args.nchannels_y * args.nsamples_y; + const dim3 block_nums_xy_tiling(nty, ntx, ntzw); + + GGML_ASSERT(args.nchannels_y % args.nchannels_x == 0); + GGML_ASSERT(args.nsamples_y % args.nsamples_x == 0); + const int channel_ratio = args.nchannels_y / args.nchannels_x; + const int sample_ratio = args.nsamples_y / args.nsamples_x; if (!args.use_stream_k) { - if (args.ne01 % mmq_y == 0) { + if (args.nrows_x % mmq_y == 0) { constexpr bool need_check = false; - mul_mat_q<<>> - (args.x, args.y, args.dst, nullptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0); + mul_mat_q<<>> + (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr, + args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst, + channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst, + sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst); } else { constexpr bool need_check = true; - mul_mat_q<<>> - (args.x, args.y, args.dst, nullptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0); + mul_mat_q<<>> + (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr, + args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst, + channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst, + sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst); } return; } - const dim3 block_nums_mmq(nsm, 1, 1); + const dim3 block_nums_stream_k(nsm, 1, 1); + const bool fixup_needed = ntx*nty*ntzw % nsm != 0; ggml_cuda_pool & pool = ctx.pool(id); - ggml_cuda_pool_alloc tmp_fixup(pool, block_nums_mmq.x * mmq_x*mmq_y); + ggml_cuda_pool_alloc tmp_fixup(pool); + if (fixup_needed) { + tmp_fixup.alloc(block_nums_stream_k.x * mmq_x*mmq_y); + } - if (args.ne01 % mmq_y == 0) { + if (args.nrows_x % mmq_y == 0) { constexpr bool need_check = false; - mul_mat_q<<>> - (args.x, args.y, args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0); + mul_mat_q<<>> + (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, + args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst, + channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst, + sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst); - mul_mat_q_stream_k_fixup<<>> - (args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.ne11, args.ne0, block_nums_mmq.x); + if (!fixup_needed) { + return; + } + + mul_mat_q_stream_k_fixup<<>> + (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_dst, + args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst); } else { constexpr bool need_check = true; - mul_mat_q<<>> - (args.x, args.y, args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0); + mul_mat_q<<>> + (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, + args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst, + channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst, + sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst); - mul_mat_q_stream_k_fixup<<>> - (args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.ne11, args.ne0, block_nums_mmq.x); + if (!fixup_needed) { + return; + } + + mul_mat_q_stream_k_fixup<<>> + (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_dst, + args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst); } } template void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) { - const int id = ggml_cuda_get_device(); - const int cc = ggml_cuda_info().devices[id].cc; - const int smpbo = ggml_cuda_info().devices[id].smpbo; + const int id = ggml_cuda_get_device(); + const int cc = ggml_cuda_info().devices[id].cc; + const size_t smpbo = ggml_cuda_info().devices[id].smpbo; const int mmq_x_max = get_mmq_x_max_host(cc); const int mmq_y = get_mmq_y_host(cc); - const int block_num_y = (args.ne01 + mmq_y - 1) / mmq_y; - const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA; int mmq_x_best = 0; - int nparts_best = INT_MAX; + int ntiles_x_best = INT_MAX; - for (int mmq_x = 8; mmq_x <= mmq_x_max && nparts_best > 1; mmq_x += 8) { + for (int mmq_x = 8; mmq_x <= mmq_x_max && ntiles_x_best > 1; mmq_x += 8) { const int granularity = mmq_get_granularity_host(mmq_x, cc); - if (mmq_x % granularity != 0 || mmq_get_shmem(mmq_x, mmq_y, cc) > smpbo) { + if (mmq_x % granularity != 0 || mmq_get_nbytes_shared(mmq_x, mmq_y, cc) > smpbo) { continue; } - const int ntiles_x = (args.ne11 + mmq_x - 1) / mmq_x; - const int nwaves_xy_tiling = ntiles_x*block_num_y; - const int nparts = use_stream_k ? ntiles_x : nwaves_xy_tiling; + const int ntiles_x = (args.ncols_y + mmq_x - 1) / mmq_x; - if (nparts < nparts_best) { - mmq_x_best = mmq_x; - nparts_best = nparts; + if (ntiles_x < ntiles_x_best) { + mmq_x_best = mmq_x; + ntiles_x_best = ntiles_x; } } @@ -2954,6 +3205,9 @@ extern DECL_MMQ_CASE(GGML_TYPE_IQ4_XS); // ------------------------------------------------------------------------------------------------------------------------- +void ggml_cuda_mul_mat_q( + ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst); + void ggml_cuda_op_mul_mat_q( ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu index d846e35a6a..dc7adf509f 100644 --- a/ggml/src/ggml-cuda/mmvq.cu +++ b/ggml/src/ggml-cuda/mmvq.cu @@ -158,7 +158,7 @@ static __global__ void mul_mat_vec_q( const int blocks_per_row_x = ncols_x / qk; constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi; - // The MUL_MAT_ID code path with ids != nullptr is only implemetned for ncols_dst == 1. + // The MUL_MAT_ID code path with ids != nullptr is only implemented for ncols_dst == 1. const int channel_dst = blockIdx.y; const int channel_x = ncols_dst == 1 && ids ? ids[channel_dst] : channel_dst / channel_ratio; const int channel_y = ncols_dst == 1 && ids ? channel_dst % nchannels_y : channel_dst; @@ -507,19 +507,30 @@ void ggml_cuda_mul_mat_vec_q( GGML_ASSERT( nb0 == ts_dst); GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type)); - GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for batch size 1. + GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for batch size 1. const float * src1_d = (const float *) src1->data; const int32_t * ids_d = ids ? (const int32_t *) ids->data : nullptr; float * dst_d = (float *) dst->data; + // If src0 is a temporary compute buffer, clear any potential padding. + if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) { + const size_t size_data = ggml_nbytes(src0); + const size_t size_alloc = ggml_backend_buffer_get_alloc_size(src0->buffer, src0); + if (size_alloc > size_data) { + GGML_ASSERT(ggml_is_contiguously_allocated(src0)); + GGML_ASSERT(!src0->view_src); + CUDA_CHECK(cudaMemsetAsync((char *) src0->data + size_data, 0, size_alloc - size_data, stream)); + } + } + const int64_t ne10_padded = GGML_PAD(ne10, MATRIX_ROW_PADDING); ggml_cuda_pool_alloc src1_q8_1(ctx.pool(), ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1); { const int64_t s11 = src1->nb[1] / ts_src1; const int64_t s12 = src1->nb[2] / ts_src1; const int64_t s13 = src1->nb[3] / ts_src1; - quantize_row_q8_1_cuda(src1_d, src1_q8_1.get(), src0->type, ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream); + quantize_row_q8_1_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type, ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream); } const int64_t s01 = src0->nb[1] / ts_src0; diff --git a/ggml/src/ggml-cuda/quantize.cu b/ggml/src/ggml-cuda/quantize.cu index 3bab47d56a..cb93181455 100644 --- a/ggml/src/ggml-cuda/quantize.cu +++ b/ggml/src/ggml-cuda/quantize.cu @@ -49,29 +49,38 @@ static __global__ void quantize_q8_1( template static __global__ void quantize_mmq_q8_1( - const float * __restrict__ x, void * __restrict__ vy, const int64_t kx0, const int64_t kx1, const int64_t kx0_padded) { + const float * __restrict__ x, const int32_t * __restrict__ ids, void * __restrict__ vy, + const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03, + const int64_t ne0, const int ne1, const int ne2) { constexpr int vals_per_scale = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 64 : 32; constexpr int vals_per_sum = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 16 : 32; - const int64_t ix0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4; + const int64_t i0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4; - if (ix0 >= kx0_padded) { + if (i0 >= ne0) { return; } - const float4 * x4 = (const float4 *) x; + const int64_t i1 = blockIdx.y; + const int64_t i2 = blockIdx.z % ne2; + const int64_t i3 = blockIdx.z / ne2; - const int64_t ix1 = kx1*blockIdx.z + blockIdx.y; + const int64_t i00 = i0; + const int64_t i01 = ids ? ids[i1] : i1; + const int64_t i02 = i2; + const int64_t i03 = i3; + + const float4 * x4 = (const float4 *) x; block_q8_1_mmq * y = (block_q8_1_mmq *) vy; const int64_t ib0 = blockIdx.z*((int64_t)gridDim.y*gridDim.x*blockDim.x/QK8_1); // first block of channel - const int64_t ib = ib0 + (ix0 / (4*QK8_1))*kx1 + blockIdx.y; // block index in channel - const int64_t iqs = ix0 % (4*QK8_1); // quant index in block + const int64_t ib = ib0 + (i0 / (4*QK8_1))*ne1 + blockIdx.y; // block index in channel + const int64_t iqs = i0 % (4*QK8_1); // quant index in block // Load 4 floats per thread and calculate max. abs. value between them: - const float4 xi = ix0 < kx0 ? x4[(ix1*kx0 + ix0)/4] : make_float4(0.0f, 0.0f, 0.0f, 0.0f); + const float4 xi = i0 < ne00 ? x4[(i03*s03 + i02*s02 + i01*s01 + i00)/4] : make_float4(0.0f, 0.0f, 0.0f, 0.0f); float amax = fabsf(xi.x); amax = fmaxf(amax, fabsf(xi.y)); amax = fmaxf(amax, fabsf(xi.z)); @@ -87,7 +96,7 @@ static __global__ void quantize_mmq_q8_1( if (ds_layout != MMQ_Q8_1_DS_LAYOUT_D4) { sum = xi.x + xi.y + xi.z + xi.w; - // Exchange calculate sum across vals_per_sum/4 threads. + // Calculate sums across vals_per_sum/4 threads. #pragma unroll for (int offset = vals_per_sum/8; offset > 0; offset >>= 1) { sum += __shfl_xor_sync(0xFFFFFFFF, sum, offset, WARP_SIZE); @@ -137,9 +146,10 @@ static __global__ void quantize_mmq_q8_1( } void quantize_row_q8_1_cuda( - const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03, - const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) { - + const float * x, const int32_t * ids, void * vy, const ggml_type type_src0, + const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03, + const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) { + GGML_ASSERT(!ids); GGML_ASSERT(ne0 % QK8_1 == 0); const int64_t block_num_x = (ne0 + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE; @@ -150,9 +160,10 @@ void quantize_row_q8_1_cuda( } void quantize_mmq_q8_1_cuda( - const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03, - const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) { - + const float * x, const int32_t * ids, void * vy, const ggml_type type_src0, + const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03, + const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) { + GGML_ASSERT(ne00 % 4 == 0); GGML_ASSERT(ne0 % (4*QK8_1) == 0); const int64_t block_num_x = (ne0 + 4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ - 1) / (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ); @@ -161,21 +172,18 @@ void quantize_mmq_q8_1_cuda( switch (mmq_get_q8_1_ds_layout(type_src0)) { case MMQ_Q8_1_DS_LAYOUT_D4: quantize_mmq_q8_1 - <<>>(x, vy, ne00, ne1, ne0); + <<>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2); break; case MMQ_Q8_1_DS_LAYOUT_DS4: quantize_mmq_q8_1 - <<>>(x, vy, ne00, ne1, ne0); + <<>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2); break; case MMQ_Q8_1_DS_LAYOUT_D2S6: quantize_mmq_q8_1 - <<>>(x, vy, ne00, ne1, ne0); + <<>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2); break; default: GGML_ABORT("fatal error"); break; } - GGML_UNUSED(s01); - GGML_UNUSED(s02); - GGML_UNUSED(s03); } diff --git a/ggml/src/ggml-cuda/quantize.cuh b/ggml/src/ggml-cuda/quantize.cuh index b627c4e400..725ab52443 100644 --- a/ggml/src/ggml-cuda/quantize.cuh +++ b/ggml/src/ggml-cuda/quantize.cuh @@ -12,13 +12,16 @@ static_assert(MATRIX_ROW_PADDING % CUDA_QUANTIZE_BLOCK_SIZE == 0, "Risk static_assert(MATRIX_ROW_PADDING % (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ) == 0, "Risk of out-of-bounds access."); typedef void (*quantize_cuda_t)( - const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03, - const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream); + const float * x, const int32_t * ids, void * vy, + ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03, + int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream); void quantize_row_q8_1_cuda( - const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03, - const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream); + const float * x, const int32_t * ids, void * vy, + ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03, + int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream); void quantize_mmq_q8_1_cuda( - const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03, - const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream); + const float * x, const int32_t * ids, void * vy, + ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03, + int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream); diff --git a/ggml/src/ggml-cuda/sum.cu b/ggml/src/ggml-cuda/sum.cu index f9589080a0..eb3d7cdba9 100644 --- a/ggml/src/ggml-cuda/sum.cu +++ b/ggml/src/ggml-cuda/sum.cu @@ -31,7 +31,7 @@ void ggml_cuda_op_sum(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); - GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguously_allocated(src0)); const float * src0_d = (const float *) src0->data; float * dst_d = (float *) dst->data; diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.cu new file mode 100644 index 0000000000..fb26abeb0d --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_16.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, 16); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu index 80108615ac..dc16829021 100644 --- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu @@ -2,9 +2,9 @@ #include "../fattn-mma-f16.cuh" -DECL_FATTN_MMA_F16_CASE(64, 1, 8); -DECL_FATTN_MMA_F16_CASE(80, 1, 8); -DECL_FATTN_MMA_F16_CASE(96, 1, 8); -DECL_FATTN_MMA_F16_CASE(112, 1, 8); -DECL_FATTN_MMA_F16_CASE(128, 1, 8); -DECL_FATTN_MMA_F16_CASE(256, 1, 8); +DECL_FATTN_MMA_F16_CASE(64, 64, 1, 8); +DECL_FATTN_MMA_F16_CASE(80, 80, 1, 8); +DECL_FATTN_MMA_F16_CASE(96, 96, 1, 8); +DECL_FATTN_MMA_F16_CASE(112, 112, 1, 8); +DECL_FATTN_MMA_F16_CASE(128, 128, 1, 8); +DECL_FATTN_MMA_F16_CASE(256, 256, 1, 8); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu index 66161c0abe..9d3cfd8edf 100644 --- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu @@ -2,9 +2,9 @@ #include "../fattn-mma-f16.cuh" -DECL_FATTN_MMA_F16_CASE(64, 16, 1); -DECL_FATTN_MMA_F16_CASE(80, 16, 1); -DECL_FATTN_MMA_F16_CASE(96, 16, 1); -DECL_FATTN_MMA_F16_CASE(112, 16, 1); -DECL_FATTN_MMA_F16_CASE(128, 16, 1); -DECL_FATTN_MMA_F16_CASE(256, 16, 1); +DECL_FATTN_MMA_F16_CASE(64, 64, 16, 1); +DECL_FATTN_MMA_F16_CASE(80, 80, 16, 1); +DECL_FATTN_MMA_F16_CASE(96, 96, 16, 1); +DECL_FATTN_MMA_F16_CASE(112, 112, 16, 1); +DECL_FATTN_MMA_F16_CASE(128, 128, 16, 1); +DECL_FATTN_MMA_F16_CASE(256, 256, 16, 1); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu index ee88c72aa0..2e1883af40 100644 --- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu @@ -2,9 +2,9 @@ #include "../fattn-mma-f16.cuh" -DECL_FATTN_MMA_F16_CASE(64, 16, 2); -DECL_FATTN_MMA_F16_CASE(80, 16, 2); -DECL_FATTN_MMA_F16_CASE(96, 16, 2); -DECL_FATTN_MMA_F16_CASE(112, 16, 2); -DECL_FATTN_MMA_F16_CASE(128, 16, 2); -DECL_FATTN_MMA_F16_CASE(256, 16, 2); +DECL_FATTN_MMA_F16_CASE(64, 64, 16, 2); +DECL_FATTN_MMA_F16_CASE(80, 80, 16, 2); +DECL_FATTN_MMA_F16_CASE(96, 96, 16, 2); +DECL_FATTN_MMA_F16_CASE(112, 112, 16, 2); +DECL_FATTN_MMA_F16_CASE(128, 128, 16, 2); +DECL_FATTN_MMA_F16_CASE(256, 256, 16, 2); 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 d888a5a423..2074e954a3 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 @@ -2,9 +2,9 @@ #include "../fattn-mma-f16.cuh" -DECL_FATTN_MMA_F16_CASE(64, 16, 4); -DECL_FATTN_MMA_F16_CASE(80, 16, 4); -DECL_FATTN_MMA_F16_CASE(96, 16, 4); -DECL_FATTN_MMA_F16_CASE(112, 16, 4); -DECL_FATTN_MMA_F16_CASE(128, 16, 4); -DECL_FATTN_MMA_F16_CASE(256, 16, 4); +DECL_FATTN_MMA_F16_CASE(64, 64, 16, 4); +DECL_FATTN_MMA_F16_CASE(80, 80, 16, 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); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.cu new file mode 100644 index 0000000000..f011a208cd --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_16.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, 16); 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 d93a2d08ed..24c64cf000 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 @@ -2,9 +2,9 @@ #include "../fattn-mma-f16.cuh" -DECL_FATTN_MMA_F16_CASE(64, 2, 4); -DECL_FATTN_MMA_F16_CASE(80, 2, 4); -DECL_FATTN_MMA_F16_CASE(96, 2, 4); -DECL_FATTN_MMA_F16_CASE(112, 2, 4); -DECL_FATTN_MMA_F16_CASE(128, 2, 4); -DECL_FATTN_MMA_F16_CASE(256, 2, 4); +DECL_FATTN_MMA_F16_CASE(64, 64, 2, 4); +DECL_FATTN_MMA_F16_CASE(80, 80, 2, 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); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu index 617464c945..163b1d939e 100644 --- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu @@ -2,9 +2,9 @@ #include "../fattn-mma-f16.cuh" -DECL_FATTN_MMA_F16_CASE(64, 2, 8); -DECL_FATTN_MMA_F16_CASE(80, 2, 8); -DECL_FATTN_MMA_F16_CASE(96, 2, 8); -DECL_FATTN_MMA_F16_CASE(112, 2, 8); -DECL_FATTN_MMA_F16_CASE(128, 2, 8); -DECL_FATTN_MMA_F16_CASE(256, 2, 8); +DECL_FATTN_MMA_F16_CASE(64, 64, 2, 8); +DECL_FATTN_MMA_F16_CASE(80, 80, 2, 8); +DECL_FATTN_MMA_F16_CASE(96, 96, 2, 8); +DECL_FATTN_MMA_F16_CASE(112, 112, 2, 8); +DECL_FATTN_MMA_F16_CASE(128, 128, 2, 8); +DECL_FATTN_MMA_F16_CASE(256, 256, 2, 8); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu index 970d2b6869..0543532ea3 100644 --- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu @@ -2,9 +2,9 @@ #include "../fattn-mma-f16.cuh" -DECL_FATTN_MMA_F16_CASE(64, 32, 1); -DECL_FATTN_MMA_F16_CASE(80, 32, 1); -DECL_FATTN_MMA_F16_CASE(96, 32, 1); -DECL_FATTN_MMA_F16_CASE(112, 32, 1); -DECL_FATTN_MMA_F16_CASE(128, 32, 1); -DECL_FATTN_MMA_F16_CASE(256, 32, 1); +DECL_FATTN_MMA_F16_CASE(64, 64, 32, 1); +DECL_FATTN_MMA_F16_CASE(80, 80, 32, 1); +DECL_FATTN_MMA_F16_CASE(96, 96, 32, 1); +DECL_FATTN_MMA_F16_CASE(112, 112, 32, 1); +DECL_FATTN_MMA_F16_CASE(128, 128, 32, 1); +DECL_FATTN_MMA_F16_CASE(256, 256, 32, 1); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu index 65cd377c39..407b6cf4c7 100644 --- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu @@ -2,9 +2,9 @@ #include "../fattn-mma-f16.cuh" -DECL_FATTN_MMA_F16_CASE(64, 32, 2); -DECL_FATTN_MMA_F16_CASE(80, 32, 2); -DECL_FATTN_MMA_F16_CASE(96, 32, 2); -DECL_FATTN_MMA_F16_CASE(112, 32, 2); -DECL_FATTN_MMA_F16_CASE(128, 32, 2); -DECL_FATTN_MMA_F16_CASE(256, 32, 2); +DECL_FATTN_MMA_F16_CASE(64, 64, 32, 2); +DECL_FATTN_MMA_F16_CASE(80, 80, 32, 2); +DECL_FATTN_MMA_F16_CASE(96, 96, 32, 2); +DECL_FATTN_MMA_F16_CASE(112, 112, 32, 2); +DECL_FATTN_MMA_F16_CASE(128, 128, 32, 2); +DECL_FATTN_MMA_F16_CASE(256, 256, 32, 2); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.cu new file mode 100644 index 0000000000..f5fd0e2369 --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_16.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, 4, 16); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu index f4a8bf3489..5e46685024 100644 --- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu @@ -2,9 +2,9 @@ #include "../fattn-mma-f16.cuh" -DECL_FATTN_MMA_F16_CASE(64, 4, 2); -DECL_FATTN_MMA_F16_CASE(80, 4, 2); -DECL_FATTN_MMA_F16_CASE(96, 4, 2); -DECL_FATTN_MMA_F16_CASE(112, 4, 2); -DECL_FATTN_MMA_F16_CASE(128, 4, 2); -DECL_FATTN_MMA_F16_CASE(256, 4, 2); +DECL_FATTN_MMA_F16_CASE(64, 64, 4, 2); +DECL_FATTN_MMA_F16_CASE(80, 80, 4, 2); +DECL_FATTN_MMA_F16_CASE(96, 96, 4, 2); +DECL_FATTN_MMA_F16_CASE(112, 112, 4, 2); +DECL_FATTN_MMA_F16_CASE(128, 128, 4, 2); +DECL_FATTN_MMA_F16_CASE(256, 256, 4, 2); 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 de191a8ab6..1ada657f19 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 @@ -2,9 +2,9 @@ #include "../fattn-mma-f16.cuh" -DECL_FATTN_MMA_F16_CASE(64, 4, 4); -DECL_FATTN_MMA_F16_CASE(80, 4, 4); -DECL_FATTN_MMA_F16_CASE(96, 4, 4); -DECL_FATTN_MMA_F16_CASE(112, 4, 4); -DECL_FATTN_MMA_F16_CASE(128, 4, 4); -DECL_FATTN_MMA_F16_CASE(256, 4, 4); +DECL_FATTN_MMA_F16_CASE(64, 64, 4, 4); +DECL_FATTN_MMA_F16_CASE(80, 80, 4, 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); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu index e8cb0e1b31..bad296b414 100644 --- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu @@ -2,9 +2,9 @@ #include "../fattn-mma-f16.cuh" -DECL_FATTN_MMA_F16_CASE(64, 4, 8); -DECL_FATTN_MMA_F16_CASE(80, 4, 8); -DECL_FATTN_MMA_F16_CASE(96, 4, 8); -DECL_FATTN_MMA_F16_CASE(112, 4, 8); -DECL_FATTN_MMA_F16_CASE(128, 4, 8); -DECL_FATTN_MMA_F16_CASE(256, 4, 8); +DECL_FATTN_MMA_F16_CASE(64, 64, 4, 8); +DECL_FATTN_MMA_F16_CASE(80, 80, 4, 8); +DECL_FATTN_MMA_F16_CASE(96, 96, 4, 8); +DECL_FATTN_MMA_F16_CASE(112, 112, 4, 8); +DECL_FATTN_MMA_F16_CASE(128, 128, 4, 8); +DECL_FATTN_MMA_F16_CASE(256, 256, 4, 8); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu index a532e96296..0d7a9c7285 100644 --- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu @@ -2,9 +2,9 @@ #include "../fattn-mma-f16.cuh" -DECL_FATTN_MMA_F16_CASE(64, 64, 1); -DECL_FATTN_MMA_F16_CASE(80, 64, 1); -DECL_FATTN_MMA_F16_CASE(96, 64, 1); -DECL_FATTN_MMA_F16_CASE(112, 64, 1); -DECL_FATTN_MMA_F16_CASE(128, 64, 1); -DECL_FATTN_MMA_F16_CASE(256, 64, 1); +DECL_FATTN_MMA_F16_CASE(64, 64, 64, 1); +DECL_FATTN_MMA_F16_CASE(80, 80, 64, 1); +DECL_FATTN_MMA_F16_CASE(96, 96, 64, 1); +DECL_FATTN_MMA_F16_CASE(112, 112, 64, 1); +DECL_FATTN_MMA_F16_CASE(128, 128, 64, 1); +DECL_FATTN_MMA_F16_CASE(256, 256, 64, 1); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu index bf25181aa7..9d5a9976f0 100644 --- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu @@ -2,9 +2,9 @@ #include "../fattn-mma-f16.cuh" -DECL_FATTN_MMA_F16_CASE(64, 8, 1); -DECL_FATTN_MMA_F16_CASE(80, 8, 1); -DECL_FATTN_MMA_F16_CASE(96, 8, 1); -DECL_FATTN_MMA_F16_CASE(112, 8, 1); -DECL_FATTN_MMA_F16_CASE(128, 8, 1); -DECL_FATTN_MMA_F16_CASE(256, 8, 1); +DECL_FATTN_MMA_F16_CASE(64, 64, 8, 1); +DECL_FATTN_MMA_F16_CASE(80, 80, 8, 1); +DECL_FATTN_MMA_F16_CASE(96, 96, 8, 1); +DECL_FATTN_MMA_F16_CASE(112, 112, 8, 1); +DECL_FATTN_MMA_F16_CASE(128, 128, 8, 1); +DECL_FATTN_MMA_F16_CASE(256, 256, 8, 1); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu index 378c132e65..a6e6f093dc 100644 --- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu @@ -2,9 +2,9 @@ #include "../fattn-mma-f16.cuh" -DECL_FATTN_MMA_F16_CASE(64, 8, 2); -DECL_FATTN_MMA_F16_CASE(80, 8, 2); -DECL_FATTN_MMA_F16_CASE(96, 8, 2); -DECL_FATTN_MMA_F16_CASE(112, 8, 2); -DECL_FATTN_MMA_F16_CASE(128, 8, 2); -DECL_FATTN_MMA_F16_CASE(256, 8, 2); +DECL_FATTN_MMA_F16_CASE(64, 64, 8, 2); +DECL_FATTN_MMA_F16_CASE(80, 80, 8, 2); +DECL_FATTN_MMA_F16_CASE(96, 96, 8, 2); +DECL_FATTN_MMA_F16_CASE(112, 112, 8, 2); +DECL_FATTN_MMA_F16_CASE(128, 128, 8, 2); +DECL_FATTN_MMA_F16_CASE(256, 256, 8, 2); 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 372641be9a..86d4ffae27 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 @@ -2,9 +2,9 @@ #include "../fattn-mma-f16.cuh" -DECL_FATTN_MMA_F16_CASE(64, 8, 4); -DECL_FATTN_MMA_F16_CASE(80, 8, 4); -DECL_FATTN_MMA_F16_CASE(96, 8, 4); -DECL_FATTN_MMA_F16_CASE(112, 8, 4); -DECL_FATTN_MMA_F16_CASE(128, 8, 4); -DECL_FATTN_MMA_F16_CASE(256, 8, 4); +DECL_FATTN_MMA_F16_CASE(64, 64, 8, 4); +DECL_FATTN_MMA_F16_CASE(80, 80, 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); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu index 9ff5968b6a..680a13ca6d 100644 --- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu @@ -2,9 +2,9 @@ #include "../fattn-mma-f16.cuh" -DECL_FATTN_MMA_F16_CASE(64, 8, 8); -DECL_FATTN_MMA_F16_CASE(80, 8, 8); -DECL_FATTN_MMA_F16_CASE(96, 8, 8); -DECL_FATTN_MMA_F16_CASE(112, 8, 8); -DECL_FATTN_MMA_F16_CASE(128, 8, 8); -DECL_FATTN_MMA_F16_CASE(256, 8, 8); +DECL_FATTN_MMA_F16_CASE(64, 64, 8, 8); +DECL_FATTN_MMA_F16_CASE(80, 80, 8, 8); +DECL_FATTN_MMA_F16_CASE(96, 96, 8, 8); +DECL_FATTN_MMA_F16_CASE(112, 112, 8, 8); +DECL_FATTN_MMA_F16_CASE(128, 128, 8, 8); +DECL_FATTN_MMA_F16_CASE(256, 256, 8, 8); 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 dd373a09d2..3428113dc8 100755 --- a/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +++ b/ggml/src/ggml-cuda/template-instances/generate_cu_files.py @@ -18,7 +18,7 @@ SOURCE_FATTN_MMA_START = """// This file has been autogenerated by generate_cu_f """ -SOURCE_FATTN_MMA_CASE = "DECL_FATTN_MMA_F16_CASE({head_size}, {ncols1}, {ncols2});\n" +SOURCE_FATTN_MMA_CASE = "DECL_FATTN_MMA_F16_CASE({head_size_kq}, {head_size_v}, {ncols1}, {ncols2});\n" TYPES_MMQ = [ "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0", @@ -57,18 +57,21 @@ for vkq_size in [16, 32]: with open(f"fattn-vec-f{vkq_size}-instance-hs{head_size}-{get_short_name(type_k)}-{get_short_name(type_v)}.cu", "w") as f: f.write(SOURCE_FATTN_VEC.format(vkq_size=vkq_size, head_size=head_size, type_k=type_k, type_v=type_v)) -for ncols in [8, 16, 32, 64, 128]: - for ncols2 in [1, 2, 4, 8]: +for ncols in [8, 16, 32, 64]: + for ncols2 in [1, 2, 4, 8, 16]: + if ncols2 > ncols: + continue ncols1 = ncols // ncols2 - if ncols == 128: - continue # Too much register pressure. with open(f"fattn-mma-f16-instance-ncols1_{ncols1}-ncols2_{ncols2}.cu", "w") as f: f.write(SOURCE_FATTN_MMA_START) - for head_size in [64, 80, 96, 112, 128, 256]: - if ncols == 128 and head_size == 256: - continue # Needs too much shared memory. - f.write(SOURCE_FATTN_MMA_CASE.format(ncols1=ncols1, ncols2=ncols2, head_size=head_size)) + for head_size_kq in [64, 80, 96, 112, 128, 256, 576]: + if head_size_kq != 576 and ncols2 == 16: + continue + if head_size_kq == 576 and ncols2 != 16: + 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)) for type in TYPES_MMQ: with open(f"mmq-instance-{get_short_name(type)}.cu", "w") as f: diff --git a/ggml/src/ggml-metal/ggml-metal-impl.h b/ggml/src/ggml-metal/ggml-metal-impl.h index 8721b272de..17eab976f3 100644 --- a/ggml/src/ggml-metal/ggml-metal-impl.h +++ b/ggml/src/ggml-metal/ggml-metal-impl.h @@ -207,6 +207,10 @@ typedef struct { float attn_factor; float beta_fast; float beta_slow; + int32_t sect_0; + int32_t sect_1; + int32_t sect_2; + int32_t sect_3; } ggml_metal_kargs_rope; typedef struct { @@ -299,21 +303,42 @@ typedef struct { } ggml_metal_kargs_mul_mv_ext; typedef struct { - int32_t nei0; - int32_t nei1; - uint64_t nbi1; + int32_t ne10; + int32_t ne11; // n_expert_used (bcast) + uint64_t nb11; + uint64_t nb12; + int32_t neh11; // n_tokens + uint64_t nbh11; + int32_t ne20; // n_expert_used + uint64_t nb21; +} ggml_metal_kargs_mul_mm_id_map0; + +typedef struct { + int32_t ne20; // n_expert_used + int32_t neh0; + int32_t neh1; + uint64_t nbh1; + uint64_t nbh2; + int32_t ne0; + uint64_t nb1; + uint64_t nb2; +} ggml_metal_kargs_mul_mm_id_map1; + +typedef struct { int32_t ne00; int32_t ne02; uint64_t nb01; uint64_t nb02; - int32_t ne11; - int32_t ne12; - int32_t ne13; - uint64_t nb10; - uint64_t nb11; - uint64_t nb12; - int32_t ne0; - int32_t ne1; + uint64_t nb03; + int32_t neh12; + uint64_t nbh10; + uint64_t nbh11; + uint64_t nbh12; + uint64_t nbh13; + int32_t neh0; + int32_t neh1; + int16_t r2; + int16_t r3; } ggml_metal_kargs_mul_mm_id; typedef struct { diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m index d92392edb7..576f9581bd 100644 --- a/ggml/src/ggml-metal/ggml-metal.m +++ b/ggml/src/ggml-metal/ggml-metal.m @@ -306,30 +306,36 @@ enum ggml_metal_kernel_type { GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32, GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32, GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32, - GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP1_F32, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F16, + GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F16, GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32, GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16, + GGML_METAL_KERNEL_TYPE_ROPE_MULTI_F32, + GGML_METAL_KERNEL_TYPE_ROPE_MULTI_F16, + GGML_METAL_KERNEL_TYPE_ROPE_VISION_F32, + GGML_METAL_KERNEL_TYPE_ROPE_VISION_F16, GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32, GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16, GGML_METAL_KERNEL_TYPE_IM2COL_F16, @@ -650,7 +656,8 @@ static void ggml_metal_mem_pool_reset(struct ggml_metal_mem_pool * mem_pool) { } if (mem_pool->heaps_to_remove.count > 0) { - for (NSUInteger i = 0; i < [mem_pool->heaps_to_remove count]; i++) { + // remove in reverse order + for (NSUInteger i = [mem_pool->heaps_to_remove count] - 1; ; --i) { NSUInteger index = [[mem_pool->heaps_to_remove objectAtIndex:i] intValue]; ggml_metal_heap_ptr * ptr = [mem_pool->heaps objectAtIndex:index]; @@ -659,6 +666,10 @@ static void ggml_metal_mem_pool_reset(struct ggml_metal_mem_pool * mem_pool) { [mem_pool->heaps removeObjectAtIndex:index]; [ptr release]; + + if (i == 0) { + break; + } } [mem_pool->heaps_to_remove removeAllObjects]; @@ -672,7 +683,7 @@ static void ggml_metal_mem_pool_clear(struct ggml_metal_mem_pool * mem_pool) { } static id ggml_metal_mem_pool_alloc(struct ggml_metal_mem_pool * mem_pool, size_t size) { - const size_t alignment = 32; + const size_t alignment = 256; const size_t size_aligned = GGML_PAD(size, alignment); @@ -1242,30 +1253,36 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32, mul_mm_iq1_m_f32, has_simdgroup_mm); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32, mul_mm_iq4_nl_f32, has_simdgroup_mm); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32, mul_mm_iq4_xs_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32, mul_mm_id_f32_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32, mul_mm_id_f16_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F32, mul_mm_id_bf16_f32, has_simdgroup_mm && use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32, mul_mm_id_q4_0_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32, mul_mm_id_q4_1_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32, mul_mm_id_q5_0_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32, mul_mm_id_q5_1_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32, mul_mm_id_q8_0_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32, mul_mm_id_q2_K_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32, mul_mm_id_q3_K_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32, mul_mm_id_q4_K_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32, mul_mm_id_q5_K_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32, mul_mm_id_q6_K_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32, mul_mm_id_iq2_xxs_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32, mul_mm_id_iq2_xs_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32, mul_mm_id_iq3_xxs_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32, mul_mm_id_iq3_s_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32, mul_mm_id_iq2_s_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32, mul_mm_id_iq1_s_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32, mul_mm_id_iq1_m_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32, mul_mm_id_iq4_nl_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32, mul_mm_id_iq4_xs_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16, mul_mm_id_map0_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP1_F32, mul_mm_id_map1_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F16, mul_mm_id_f32_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F16, mul_mm_id_f16_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F16, mul_mm_id_bf16_f16, has_simdgroup_mm && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F16, mul_mm_id_q4_0_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F16, mul_mm_id_q4_1_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F16, mul_mm_id_q5_0_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F16, mul_mm_id_q5_1_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F16, mul_mm_id_q8_0_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F16, mul_mm_id_q2_K_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F16, mul_mm_id_q3_K_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F16, mul_mm_id_q4_K_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F16, mul_mm_id_q5_K_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F16, mul_mm_id_q6_K_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F16, mul_mm_id_iq2_xxs_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F16, mul_mm_id_iq2_xs_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F16, mul_mm_id_iq3_xxs_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F16, mul_mm_id_iq3_s_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F16, mul_mm_id_iq2_s_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F16, mul_mm_id_iq1_s_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F16, mul_mm_id_iq1_m_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F16, mul_mm_id_iq4_nl_f16, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F16, mul_mm_id_iq4_xs_f16, has_simdgroup_mm); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32, rope_norm_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16, rope_norm_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_MULTI_F32, rope_multi_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_MULTI_F16, rope_multi_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_VISION_F32, rope_vision_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_VISION_F16, rope_vision_f16, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32, rope_neox_f32, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16, rope_neox_f16, true); GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16, im2col_f16, true); @@ -1628,16 +1645,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex case GGML_OP_NORM: return has_simdgroup_reduction && (op->ne[0] % 4 == 0 && ggml_is_contiguous_1(op->src[0])); case GGML_OP_ROPE: - { - const int mode = ((const int32_t *) op->op_params)[2]; - if (mode & GGML_ROPE_TYPE_MROPE) { - return false; - } - if (mode & GGML_ROPE_TYPE_VISION) { - return false; - } - return true; - } + return true; case GGML_OP_IM2COL: return op->src[0]->type == GGML_TYPE_F16; case GGML_OP_POOL_1D: @@ -2999,7 +3007,7 @@ static bool ggml_metal_encode_node( [encoder setBuffer:id_dst offset:offs_dst atIndex:3]; [encoder setThreadgroupMemoryLength:8192 atIndex:0]; - [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)]; + [encoder dispatchThreadgroups:MTLSizeMake((ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)]; } else { id pipeline = nil; @@ -3219,8 +3227,6 @@ static bool ggml_metal_encode_node( } break; case GGML_OP_MUL_MAT_ID: { - const int n_as = src0->ne[2]; - // src2 = ids const enum ggml_type src2t = src2->type; GGML_UNUSED(src2t); @@ -3234,24 +3240,21 @@ static bool ggml_metal_encode_node( GGML_ASSERT(ne03 == 1); GGML_ASSERT(ne13 == 1); + const uint32_t r2 = 1; + const uint32_t r3 = 1; + // find the break-even point where the matrix-matrix kernel becomes more efficient compared // to the matrix-vector kernel // ne20 = n_used_experts - // ne21 = n_rows - const int dst_rows = ne20*ne21; - const int dst_rows_min = n_as; - const int dst_rows_max = (device.maxThreadgroupMemoryLength/2 - 8192)/4; - - // max size of the rowids array in the kernel shared buffer - //GGML_ASSERT(dst_rows <= dst_rows_max); + // ne21 = n_rows (batch size) + const int ne21_mm_id_min = 32; // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel if ([device supportsFamily:MTLGPUFamilyApple7] && ne00 % 32 == 0 && ne00 >= 64 && - //ne01 / ne02 >= 512 && // NOTE: this is based on Mixtral shapes, might need adjustments - dst_rows > dst_rows_min && - dst_rows <= dst_rows_max) { + (ne21 >= ne21_mm_id_min)) { + GGML_ASSERT(ne00 % 4 == 0); // some Metal matrix data types require aligned pointers // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5) @@ -3262,62 +3265,169 @@ static bool ggml_metal_encode_node( default: break; } - id pipeline = nil; + const int64_t neh10 = ne10; // n_embd + const int64_t neh11 = ne21; // n_tokens + const int64_t neh12 = ne02; // n_expert - switch (src0->type) { - case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32 ].pipeline; break; - case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32 ].pipeline; break; - case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F32 ].pipeline; break; - case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32 ].pipeline; break; - case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32 ].pipeline; break; - case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32 ].pipeline; break; - case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32 ].pipeline; break; - case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32 ].pipeline; break; - case GGML_TYPE_Q2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32 ].pipeline; break; - case GGML_TYPE_Q3_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32 ].pipeline; break; - case GGML_TYPE_Q4_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32 ].pipeline; break; - case GGML_TYPE_Q5_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32 ].pipeline; break; - case GGML_TYPE_Q6_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32 ].pipeline; break; - case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32].pipeline; break; - case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32 ].pipeline; break; - case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32].pipeline; break; - case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32 ].pipeline; break; - case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32 ].pipeline; break; - case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32 ].pipeline; break; - case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32 ].pipeline; break; - case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32 ].pipeline; break; - case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32 ].pipeline; break; - default: GGML_ABORT("MUL_MAT_ID not implemented"); + const uint64_t nbh10 = ggml_type_size(GGML_TYPE_F16); + const uint64_t nbh11 = nbh10*neh10; + const uint64_t nbh12 = nbh11*neh11; + const uint64_t nbh13 = nbh12*neh12; + + const size_t s_src1 = ggml_type_size(GGML_TYPE_F16)*neh10*neh11*neh12; + id h_src1 = ggml_metal_mem_pool_alloc(mem_pool, s_src1); + if (!h_src1) { + GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, s_src1); + return false; } - ggml_metal_kargs_mul_mm_id args = { - /*.nei0 =*/ ne20, - /*.nei1 =*/ ne21, - /*.nbi1 =*/ nb21, - /*.ne00 =*/ ne00, - /*.ne02 =*/ ne02, - /*.nb01 =*/ nb01, - /*.nb02 =*/ nb02, - /*.ne11 =*/ ne11, - /*.ne12 =*/ ne12, - /*.ne13 =*/ ne13, - /*.nb10 =*/ nb10, - /*.nb11 =*/ nb11, - /*.nb12 =*/ nb12, - /*.ne0 =*/ ne0, - /*.ne1 =*/ ne1, - }; + const int64_t neh0 = ne0; + const int64_t neh1 = ne21; + const int64_t neh2 = ne02; - [encoder setComputePipelineState:pipeline]; - [encoder setBytes:&args length:sizeof(args) atIndex:0]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1]; - [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:3]; - [encoder setBuffer:id_src2 offset:offs_src2 atIndex:4]; + const uint64_t nbh0 = ggml_type_size(GGML_TYPE_F32); + const uint64_t nbh1 = nbh0*neh0; + const uint64_t nbh2 = nbh1*neh1; + //const uint64_t nbh3 = nbh2*neh2; - [encoder setThreadgroupMemoryLength:GGML_PAD(8192 + dst_rows*4/*sizeof(ushort2)*/, 16) atIndex:0]; + const size_t s_dst = ggml_type_size(GGML_TYPE_F32)*neh0*neh1*neh2; + id h_dst = ggml_metal_mem_pool_alloc(mem_pool, s_dst); + if (!h_dst) { + GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, s_dst); + return false; + } - [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 31)/32, (ne01 + 63)/64, n_as) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)]; + // tokens per expert + const size_t s_tpe = ggml_type_size(GGML_TYPE_I32)*ne02; + id h_tpe = ggml_metal_mem_pool_alloc(mem_pool, s_tpe); + if (!h_tpe) { + GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, s_tpe); + return false; + } + + // id map + // [n_expert_used, n_tokens] + const size_t s_ids = ggml_type_size(GGML_TYPE_I32)*ne20*ne21; + id h_ids = ggml_metal_mem_pool_alloc(mem_pool, s_ids); + if (!h_ids) { + GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, s_ids); + return false; + } + + { + const int nth = MIN(1024, ne10/4); + + ggml_metal_kargs_mul_mm_id_map0 args = { + ne10, + ne11, // n_expert_used (bcast) + nb11, + nb12, + neh11, // n_tokens + nbh11, + ne20, // n_expert_used + nb21, + }; + + id pipeline = nil; + + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP0_F16].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBytes:&args length:sizeof(args) atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2]; + [encoder setBuffer: h_src1 offset:0 atIndex:3]; + [encoder setBuffer: h_tpe offset:0 atIndex:4]; + [encoder setBuffer: h_ids offset:0 atIndex:5]; + + [encoder dispatchThreadgroups:MTLSizeMake(ne02, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } + + { + id pipeline = nil; + + switch (src0->type) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F16 ].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F16 ].pipeline; break; + case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F16 ].pipeline; break; + case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F16 ].pipeline; break; + case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F16 ].pipeline; break; + case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F16 ].pipeline; break; + case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F16 ].pipeline; break; + case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F16 ].pipeline; break; + case GGML_TYPE_Q2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F16 ].pipeline; break; + case GGML_TYPE_Q3_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F16 ].pipeline; break; + case GGML_TYPE_Q4_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F16 ].pipeline; break; + case GGML_TYPE_Q5_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F16 ].pipeline; break; + case GGML_TYPE_Q6_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F16 ].pipeline; break; + case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F16].pipeline; break; + case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F16 ].pipeline; break; + case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F16].pipeline; break; + case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F16 ].pipeline; break; + case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F16 ].pipeline; break; + case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F16 ].pipeline; break; + case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F16 ].pipeline; break; + case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F16 ].pipeline; break; + case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F16 ].pipeline; break; + default: GGML_ABORT("MUL_MAT_ID not implemented"); + } + + ggml_metal_kargs_mul_mm_id args = { + /*.ne00 =*/ ne00, + /*.ne02 =*/ ne02, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.nb03 =*/ nb03, + /*.neh12 =*/ neh12, + /*.nbh10 =*/ nbh10, + /*.nbh11 =*/ nbh11, + /*.nbh12 =*/ nbh12, + /*.nbh13 =*/ nbh13, + /*.neh0 =*/ neh0, + /*.neh1 =*/ neh1, + /*.r2 =*/ r2, + /*.r3 =*/ r3, + }; + + [encoder setComputePipelineState:pipeline]; + [encoder setBytes:&args length:sizeof(args) atIndex:0]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1]; + [encoder setBuffer: h_src1 offset:0 atIndex:2]; + [encoder setBuffer: h_tpe offset:0 atIndex:3]; + [encoder setBuffer: h_dst offset:0 atIndex:4]; + + [encoder setThreadgroupMemoryLength:8192 atIndex:0]; + [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 31)/32, (ne01 + 63)/64, ne02) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)]; + } + + { + GGML_ASSERT(ne0 % 4 == 0); + + const int nth = MIN(1024, ne0/4); + + ggml_metal_kargs_mul_mm_id_map1 args = { + ne20, // n_expert_used + neh0, + neh1, + nbh1, + nbh2, + ne0, + nb1, + nb2, + }; + + id pipeline = nil; + + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_MAP1_F32].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBytes:&args length:sizeof(args) atIndex:0]; + [encoder setBuffer: h_dst offset:0 atIndex:1]; + [encoder setBuffer: h_ids offset:0 atIndex:2]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:3]; + + [encoder dispatchThreadgroups:MTLSizeMake(ne20, ne21, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } } else { id pipeline = nil; @@ -3511,7 +3621,7 @@ static bool ggml_metal_encode_node( [encoder setBuffer:id_src2 offset:offs_src2 atIndex:4]; const int64_t _ne1 = 1; - const int64_t ne123 = dst_rows; + const int64_t ne123 = ne20*ne21; if (smem > 0) { [encoder setThreadgroupMemoryLength:smem atIndex:0]; @@ -3715,6 +3825,7 @@ static bool ggml_metal_encode_node( } break; case GGML_OP_ROPE: { + // make sure we have one or more position id(ne10) per token(ne02) GGML_ASSERT(ne10 % ne02 == 0); GGML_ASSERT(ne10 >= ne02); @@ -3741,20 +3852,42 @@ static bool ggml_metal_encode_node( memcpy(&beta_fast, (const int32_t *) dst->op_params + 9, sizeof(float)); memcpy(&beta_slow, (const int32_t *) dst->op_params + 10, sizeof(float)); - const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; + const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; + const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE; + const bool is_vision = mode == GGML_ROPE_TYPE_VISION; + + // mrope + const int sect_0 = ((const int32_t *) dst->op_params)[11]; + const int sect_1 = ((const int32_t *) dst->op_params)[12]; + const int sect_2 = ((const int32_t *) dst->op_params)[13]; + const int sect_3 = ((const int32_t *) dst->op_params)[14]; id pipeline = nil; - if (!is_neox) { + if (is_neox) { switch (src0->type) { - case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32].pipeline; break; - case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16].pipeline; break; + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16].pipeline; break; + default: GGML_ABORT("fatal error"); + }; + } else if (is_mrope && !is_vision) { + GGML_ASSERT(ne10*4 >= ne02); // need at least 4 pos per token + switch (src0->type) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_MULTI_F32].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_MULTI_F16].pipeline; break; + default: GGML_ABORT("fatal error"); + }; + } else if (is_vision) { + GGML_ASSERT(ne10*4 >= ne02); // need at least 4 pos per token + switch (src0->type) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_VISION_F32].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_VISION_F16].pipeline; break; default: GGML_ABORT("fatal error"); }; } else { switch (src0->type) { - case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32].pipeline; break; - case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16].pipeline; break; + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16].pipeline; break; default: GGML_ABORT("fatal error"); }; } @@ -3785,6 +3918,10 @@ static bool ggml_metal_encode_node( /*.attn_factor =*/ attn_factor, /*.beta_fast =*/ beta_fast, /*.beta_slow =*/ beta_slow, + /* sect_0 =*/ sect_0, + /* sect_1 =*/ sect_1, + /* sect_2 =*/ sect_2, + /* sect_3 =*/ sect_3, }; [encoder setComputePipelineState:pipeline]; diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal index 9f4147e939..9cfddf4503 100644 --- a/ggml/src/ggml-metal/ggml-metal.metal +++ b/ggml/src/ggml-metal/ggml-metal.metal @@ -2713,8 +2713,148 @@ kernel void kernel_rope_neox( } } +template +kernel void kernel_rope_multi( + constant ggml_metal_kargs_rope & args, + device const char * src0, + device const char * src1, + device const char * src2, + device char * dst, + ushort tiitg[[thread_index_in_threadgroup]], + ushort3 tptg [[threads_per_threadgroup]], + uint3 tgpig[[threadgroup_position_in_grid]]) { + const int i3 = tgpig[2]; + const int i2 = tgpig[1]; + const int i1 = tgpig[0]; + + float corr_dims[2]; + rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims); + + device const int32_t * pos = (device const int32_t *) src1; + + const float inv_ndims = -1.f/args.n_dims; + + float cos_theta; + float sin_theta; + + for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) { + if (i0 < args.n_dims) { + const int ic = i0/2; + + // mrope theta calculations + // note: the rest is the same as kernel_rope_neox + const int sect_dims = args.sect_0 + args.sect_1 + args.sect_2 + args.sect_3; + const int sec_w01 = args.sect_0 + args.sect_1; // end of section 1 + const int sec_w012 = args.sect_0 + args.sect_1 + args.sect_2; // end of section 2 + const int sector = ic % sect_dims; + + float theta_base; + if (sector < args.sect_0) { + theta_base = (float) pos[i2]; + } else if (sector < sec_w01) { + theta_base = (float) pos[i2 + args.ne02]; + } else if (sector < sec_w012) { + theta_base = (float) pos[i2 + args.ne02 * 2]; + } else { + theta_base = (float) pos[i2 + args.ne02 * 3]; + } + // end of mrope + + const float theta = theta_base * pow(args.freq_base, inv_ndims*i0); + + const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f; + + rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta); + + device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00); + device T * dst_data = (device T *)( dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + ic*args.nb0); + + const float x0 = src[0]; + const float x1 = src[args.n_dims/2]; + + dst_data[0] = x0*cos_theta - x1*sin_theta; + dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta; + } else { + device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00); + device T * dst_data = (device T *)( dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0); + + dst_data[0] = src[0]; + dst_data[1] = src[1]; + } + } +} + +template +kernel void kernel_rope_vision( + constant ggml_metal_kargs_rope & args, + device const char * src0, + device const char * src1, + device const char * src2, + device char * dst, + ushort tiitg[[thread_index_in_threadgroup]], + ushort3 tptg [[threads_per_threadgroup]], + uint3 tgpig[[threadgroup_position_in_grid]]) { + const int i3 = tgpig[2]; + const int i2 = tgpig[1]; + const int i1 = tgpig[0]; + + float corr_dims[2]; + rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims); + + device const int32_t * pos = (device const int32_t *) src1; + + const float inv_ndims = -1.f/args.n_dims; + + float cos_theta; + float sin_theta; + + for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) { + if (i0 < 2*args.n_dims) { // different from kernel_rope_multi + const int ic = i0/2; + + // mrope theta calculations (only support 2 dimensions) + const int sect_dims = args.sect_0 + args.sect_1; + const int sector = ic % sect_dims; + + float p; + float theta_base; + if (sector < args.sect_1) { + p = (float) sector; + theta_base = (float) pos[i2]; + } else { + p = (float) sector - args.sect_0; + theta_base = (float) pos[i2 + args.ne02]; + } + + const float theta = theta_base * pow(args.freq_base, 2.0f * inv_ndims * p); + // end of mrope + + const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f; + + rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta); + + device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00); + device T * dst_data = (device T *)( dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + ic*args.nb0); + + const float x0 = src[0]; + const float x1 = src[args.n_dims]; // different from kernel_rope_multi + + dst_data[0] = x0*cos_theta - x1*sin_theta; + dst_data[args.n_dims] = x0*sin_theta + x1*cos_theta; // different from kernel_rope_multi + } else { + device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00); + device T * dst_data = (device T *)( dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0); + + dst_data[0] = src[0]; + dst_data[1] = src[1]; + } + } +} + typedef decltype(kernel_rope_norm) kernel_rope_norm_t; typedef decltype(kernel_rope_neox) kernel_rope_neox_t; +typedef decltype(kernel_rope_multi) kernel_rope_multi_t; +typedef decltype(kernel_rope_vision) kernel_rope_vision_t; template [[host_name("kernel_rope_norm_f32")]] kernel kernel_rope_norm_t kernel_rope_norm; template [[host_name("kernel_rope_norm_f16")]] kernel kernel_rope_norm_t kernel_rope_norm; @@ -2722,6 +2862,12 @@ template [[host_name("kernel_rope_norm_f16")]] kernel kernel_rope_norm_t kernel_ template [[host_name("kernel_rope_neox_f32")]] kernel kernel_rope_neox_t kernel_rope_neox; template [[host_name("kernel_rope_neox_f16")]] kernel kernel_rope_neox_t kernel_rope_neox; +template [[host_name("kernel_rope_multi_f32")]] kernel kernel_rope_multi_t kernel_rope_multi; +template [[host_name("kernel_rope_multi_f16")]] kernel kernel_rope_multi_t kernel_rope_multi; + +template [[host_name("kernel_rope_vision_f32")]] kernel kernel_rope_vision_t kernel_rope_vision; +template [[host_name("kernel_rope_vision_f16")]] kernel kernel_rope_vision_t kernel_rope_vision; + typedef void (im2col_t)( device const float * x, device char * dst, @@ -6336,127 +6482,219 @@ kernel void kernel_mul_mm( } } -// same as kernel_mul_mm_impl, but src1 and dst are accessed via indices stored in rowids -// TODO: this kernel needs to be reimplemented from scratch for better performance -template -void kernel_mul_mm_id_impl( - int32_t ne00, - int32_t ne02, - uint64_t nb01, - uint64_t nb02, - int32_t ne11, - int32_t ne12, - uint64_t nb10, - uint64_t nb11, - uint64_t nb12, - int32_t ne0, - int32_t ne1, - int64_t ne0ne1, - device const char * src0, - device const char * src1, - threadgroup ushort2 * rowids, - device char * dst, - threadgroup char * shmem, +template +kernel void kernel_mul_mm_id_map0( + constant ggml_metal_kargs_mul_mm_id_map0 & args, + device const char * src1, + device const char * src2, + device char * hsrc1, + device char * htpe, + device char * hids, + uint3 tgpig[[threadgroup_position_in_grid]], + ushort3 tpitg[[thread_position_in_threadgroup]], + ushort3 ntg[[threads_per_threadgroup]]) { + const int ide = tgpig[0]; // expert id + + int n_all = 0; + + device int32_t * ids_i32 = (device int32_t *) (hids); + + for (int i21 = 0; i21 < args.neh11; i21++) { // n_tokens + device const int32_t * src2_i32 = (device const int32_t *) (src2 + i21*args.nb21); + + for (int i20 = 0; i20 < args.ne20; i20++) { // n_expert_used + if (src2_i32[i20] != ide) { + continue; + } + + device const float4 * src1_f32x4 = (device const float4 *) ( src1 + i21*args.nb12 + (i20%args.ne11)*args.nb11); + device T4 * hsrc1_f32x4 = (device T4 *) (hsrc1 + (ide*args.neh11 + n_all)*args.nbh11); + + for (int64_t i00 = tpitg.x; i00 < args.ne10/4; i00 += ntg.x) { + hsrc1_f32x4[i00] = (T4) (src1_f32x4[i00]); + } + + if (tpitg.x == 0) { + ids_i32[i21*args.ne20 + i20] = ide*args.neh11 + n_all; + } + + ++n_all; + } + } + + if (tpitg.x == 0) { + device int32_t * tpe_i32 = (device int32_t *) (htpe); + tpe_i32[ide] = n_all; + } +} + +typedef decltype(kernel_mul_mm_id_map0) kernel_mul_mm_id_map0_t; + +template [[host_name("kernel_mul_mm_id_map0_f16")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0; + +template +kernel void kernel_mul_mm_id_map1( + constant ggml_metal_kargs_mul_mm_id_map1 & args, + device const char * hdst, + device const char * hids, + device char * dst, + uint3 tgpig[[threadgroup_position_in_grid]], + ushort3 tpitg[[thread_position_in_threadgroup]], + ushort3 ntg[[threads_per_threadgroup]]) { + const int i20 = tgpig[0]; // used expert + const int i21 = tgpig[1]; // token + + device const int32_t * ids_i32 = (device const int32_t *) (hids); + device float4 * dst_f32x4 = (device float4 *) (dst + i20*args.nb1 + i21*args.nb2); + + const int id = ids_i32[i21*args.ne20 + i20]; + + const int ide = id / args.neh1; + const int idt = id % args.neh1; + + device const float4 * hdst_f32x4 = (device const float4 *) (hdst + idt*args.nbh1 + ide*args.nbh2); + + for (int64_t i0 = tpitg.x; i0 < args.neh0/4; i0 += ntg.x) { + dst_f32x4[i0] = hdst_f32x4[i0]; + } +} + +typedef decltype(kernel_mul_mm_id_map1) kernel_mul_mm_id_map1_t; + +template [[host_name("kernel_mul_mm_id_map1_f32")]] kernel kernel_mul_mm_id_map1_t kernel_mul_mm_id_map1; + +template +kernel void kernel_mul_mm_id( + constant ggml_metal_kargs_mul_mm_id & args, + device const char * src0, + device const char * src1, + device const char * tpe, + device char * dst, + threadgroup char * shmem [[threadgroup(0)]], uint3 tgpig[[threadgroup_position_in_grid]], ushort tiitg[[thread_index_in_threadgroup]], ushort sgitg[[simdgroup_index_in_threadgroup]]) { - threadgroup half * sa = (threadgroup half *)(shmem); - threadgroup float * sb = (threadgroup float *)(shmem + 4096); + threadgroup T * sa = (threadgroup T *)(shmem); + threadgroup half * sb = (threadgroup half *)(shmem + 4096); const int r0 = tgpig.y; const int r1 = tgpig.x; + const int im = tgpig.z; - if (r1*BLOCK_SIZE_N >= ne1) return; + device const int32_t * tpe_i32 = (device const int32_t *) (tpe); + + const int neh1 = tpe_i32[im]; + + if (r1*BLOCK_SIZE_N >= neh1) { + return; + } // if this block is of 64x32 shape or smaller - short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M; - short n_cols = (ne1 - r1 * BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1 * BLOCK_SIZE_N) : BLOCK_SIZE_N; + const short n_rows = (args.neh0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.neh0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M; + const short n_cols = ( neh1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? ( neh1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N; // a thread shouldn't load data outside of the matrix - short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1; - short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1; + const short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1; + const short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1; - simdgroup_half8x8 ma[4]; - simdgroup_float8x8 mb[2]; + simdgroup_T8x8 ma[4]; + simdgroup_half8x8 mb[2]; simdgroup_float8x8 mc[8]; - for (int i = 0; i < 8; i++){ + + for (short i = 0; i < 8; i++){ mc[i] = make_filled_simdgroup_matrix(0.f); } + short il = (tiitg % THREAD_PER_ROW); - ushort offset1 = il/nl; + const int i12 = im%args.neh12; + const int i13 = im/args.neh12; - threadgroup const auto & id = rowids[r1 * BLOCK_SIZE_N + thread_col]; + const uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03; + const short offset1 = il/nl; - device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01) + offset1; - device const float * y = (device const float *)(src1 - + nb12 * id[1] - + nb11 * (id[0] % ne11) - + nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL))); + device const block_q * x = (device const block_q *)(src0 + + args.nb01*(r0*BLOCK_SIZE_M + thread_row) + offset0) + offset1; - for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) { + device const half * y = (device const half *)(src1 + + args.nbh13*i13 + + args.nbh12*i12 + + args.nbh11*(r1*BLOCK_SIZE_N + thread_col) + + args.nbh10*(BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL))); + + for (int loop_k = 0; loop_k < args.ne00; loop_k += BLOCK_SIZE_K) { // load data and store to threadgroup memory - half4x4 temp_a; + T4x4 temp_a; dequantize_func(x, il, temp_a); + threadgroup_barrier(mem_flags::mem_threadgroup); - for (int i = 0; i < 16; i++) { - *(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \ - + (tiitg % THREAD_PER_ROW) * 16 + (i / 8) * 8) \ - + (tiitg / THREAD_PER_ROW) % 8 + (i & 7) * 8) = temp_a[i/4][i%4]; + #pragma unroll(16) + for (short i = 0; i < 16; i++) { + *(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \ + + (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \ + + (tiitg/THREAD_PER_ROW)%8 + (i&7)*8) = temp_a[i/4][i%4]; } - *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) = *((device float2x4 *)y); + *(threadgroup half2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = *((device half2x4 *) y); il = (il + 2 < nl) ? il + 2 : il % 2; - x = (il < 2) ? x + (2+nl-1)/nl : x; + x = (il < 2) ? x + (2 + nl - 1)/nl : x; y += BLOCK_SIZE_K; threadgroup_barrier(mem_flags::mem_threadgroup); // load matrices from threadgroup memory and conduct outer products - threadgroup half * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2)); - threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2)); + threadgroup const T * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2)); + threadgroup const half * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2)); - #pragma unroll(BLOCK_SIZE_K/8) - for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) { + #pragma unroll(4) + for (short ik = 0; ik < BLOCK_SIZE_K/8; ik++) { #pragma unroll(4) - for (int i = 0; i < 4; i++) { + for (short i = 0; i < 4; i++) { simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i); } + simdgroup_barrier(mem_flags::mem_none); + #pragma unroll(2) - for (int i = 0; i < 2; i++) { + for (short i = 0; i < 2; i++) { simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i); } - lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE; - lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE; - #pragma unroll(8) - for (int i = 0; i < 8; i++){ + for (short i = 0; i < 8; i++){ simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]); } + + lsma += (BLOCK_SIZE_M/SG_MAT_ROW)*SG_MAT_SIZE; + lsmb += (BLOCK_SIZE_N/SG_MAT_ROW)*SG_MAT_SIZE; } } - { + if ((r0 + 1) * BLOCK_SIZE_M <= args.neh0 && (r1 + 1) * BLOCK_SIZE_N <= neh1) { + device float * C = (device float *) dst + + (BLOCK_SIZE_M * r0 + 32*(sgitg & 1)) + \ + (BLOCK_SIZE_N * r1 + 16*(sgitg >> 1)) * args.neh0 + im*args.neh1*args.neh0; + + for (short i = 0; i < 8; i++) { + simdgroup_store(mc[i], C + 8 * (i%4) + 8 * args.neh0 * (i/4), args.neh0); + } + } else { + // block is smaller than 64x32, we should avoid writing data outside of the matrix threadgroup_barrier(mem_flags::mem_threadgroup); threadgroup float * temp_str = ((threadgroup float *) shmem) \ - + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M; - for (int i = 0; i < 8; i++) { - simdgroup_store(mc[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M); + + 32*(sgitg&1) + (16*(sgitg >> 1))*BLOCK_SIZE_M; + for (short i = 0; i < 8; i++) { + simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M); } threadgroup_barrier(mem_flags::mem_threadgroup); if (sgitg == 0) { for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) { - threadgroup const auto & jid = rowids[r1 * BLOCK_SIZE_N + j]; - int64_t joff = jid[0]*ne0 + jid[1]*ne0ne1; - - device float * D = (device float *) dst + (r0*BLOCK_SIZE_M) + joff; + device float * D = (device float *) dst + (r0*BLOCK_SIZE_M) + (r1*BLOCK_SIZE_N + j)*args.neh0 + im*args.neh1*args.neh0; device float4 * D4 = (device float4 *) D; threadgroup float * C = temp_str + (j*BLOCK_SIZE_M); @@ -6476,66 +6714,6 @@ void kernel_mul_mm_id_impl( } } -template -kernel void kernel_mul_mm_id( - constant ggml_metal_kargs_mul_mm_id & args, - device const char * src0s, - device const char * src1, - device char * dst, - device const char * ids, - threadgroup char * shmem [[threadgroup(0)]], - uint3 tgpig[[threadgroup_position_in_grid]], - ushort tiitg[[thread_index_in_threadgroup]], - ushort sgitg[[simdgroup_index_in_threadgroup]]) { - - const int32_t i02 = tgpig.z; - - tgpig.z = 0; - - device const char * src0 = src0s + i02*args.nb02; - - // row indices - threadgroup ushort2 * rowids = (threadgroup ushort2 *)(shmem + 8192); - - // TODO: parallelize this loop - int32_t _ne1 = 0; - for (ushort ii1 = 0; ii1 < args.nei1; ii1++) { - for (ushort ii0 = 0; ii0 < args.nei0; ii0++) { - int32_t id = ((device int32_t *) (ids + ii1*args.nbi1))[ii0]; - if (id == i02) { - if (tiitg == 0) { - rowids[_ne1] = ushort2(ii0, ii1); - } - _ne1++; - } - } - } - - threadgroup_barrier(mem_flags::mem_threadgroup); - - kernel_mul_mm_id_impl( - args.ne00, - args.ne02, - args.nb01, - args.nb02, - args.ne11, - args.ne12, - args.nb10, - args.nb11, - args.nb12, - args.ne0, - _ne1, - (int64_t)args.ne0*args.ne1, - src0, - src1, - rowids, - dst, - shmem, - tgpig, - tiitg, - sgitg); -} - #define QK_NL 16 // @@ -6576,63 +6754,64 @@ template [[host_name("kernel_get_rows_iq4_xs")]] kernel get_rows_q_t kernel_get // matrix-matrix multiplication // -typedef decltype(kernel_mul_mm) mat_mm_t; +typedef decltype(kernel_mul_mm) mul_mm_t; -template [[host_name("kernel_mul_mm_f32_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_f16_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_f32_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_f16_f32")]] kernel mul_mm_t kernel_mul_mm; #if defined(GGML_METAL_USE_BF16) -template [[host_name("kernel_mul_mm_bf16_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_bf16_f32")]] kernel mul_mm_t kernel_mul_mm; #endif -template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q5_0_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q5_1_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q5_K_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_iq2_xs_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_iq3_s_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_iq2_s_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_iq1_s_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_iq1_m_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_iq4_nl_f32")]] kernel mat_mm_t kernel_mul_mm; -template [[host_name("kernel_mul_mm_iq4_xs_f32")]] kernel mat_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q5_0_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q5_1_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q5_K_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_iq2_xs_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_iq3_s_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_iq2_s_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_iq1_s_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_iq1_m_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_iq4_nl_f32")]] kernel mul_mm_t kernel_mul_mm; +template [[host_name("kernel_mul_mm_iq4_xs_f32")]] kernel mul_mm_t kernel_mul_mm; // // indirect matrix-matrix multiplication // -typedef decltype(kernel_mul_mm_id) mat_mm_id_t; +typedef decltype(kernel_mul_mm_id) mul_mm_id; -template [[host_name("kernel_mul_mm_id_f32_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_f16_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_f32_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_f16_f16")]] kernel mul_mm_id kernel_mul_mm_id; #if defined(GGML_METAL_USE_BF16) -template [[host_name("kernel_mul_mm_id_bf16_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_bf16_f16")]] kernel mul_mm_id kernel_mul_mm_id; #endif -template [[host_name("kernel_mul_mm_id_q4_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q4_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q5_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q5_1_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q8_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q2_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q3_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q4_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q5_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_iq3_s_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_iq2_s_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_iq1_s_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_iq1_m_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; -template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q4_0_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q4_1_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q5_0_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q5_1_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q8_0_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q2_K_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q3_K_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q4_K_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q5_K_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_q6_K_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_iq2_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_iq2_xs_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_iq3_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_iq3_s_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_iq2_s_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_iq1_s_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_iq1_m_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_iq4_nl_f16")]] kernel mul_mm_id kernel_mul_mm_id; +template [[host_name("kernel_mul_mm_id_iq4_xs_f16")]] kernel mul_mm_id kernel_mul_mm_id; + // // matrix-vector multiplication diff --git a/ggml/src/ggml-opt.cpp b/ggml/src/ggml-opt.cpp index 7c3e24103a..58d77578f4 100644 --- a/ggml/src/ggml-opt.cpp +++ b/ggml/src/ggml-opt.cpp @@ -28,16 +28,19 @@ struct ggml_opt_dataset { }; struct ggml_opt_context { - ggml_backend_sched_t backend_sched = nullptr; - ggml_cgraph * allocated_graph = nullptr; - ggml_cgraph * allocated_graph_copy = nullptr; - struct ggml_context * ctx_static = nullptr; - struct ggml_context * ctx_static_cpu = nullptr; - struct ggml_context * ctx_compute = nullptr; - struct ggml_context * ctx_copy = nullptr; - ggml_backend_buffer_t buf_static = nullptr; - ggml_backend_buffer_t buf_static_cpu = nullptr; - std::mt19937 rng; + ggml_backend_sched_t backend_sched = nullptr; + ggml_cgraph * allocated_graph = nullptr; + ggml_cgraph * allocated_graph_copy = nullptr; + struct ggml_context * ctx_static = nullptr; + struct ggml_context * ctx_cpu = nullptr; + struct ggml_context * ctx_compute = nullptr; + struct ggml_context * ctx_copy = nullptr; + ggml_backend_buffer_t buf_static = nullptr; + ggml_backend_buffer_t buf_cpu = nullptr; + std::mt19937 rng; + enum ggml_opt_loss_type loss_type; + enum ggml_opt_build_type build_type; + enum ggml_opt_build_type build_type_alloc; struct ggml_tensor * inputs = nullptr; struct ggml_tensor * outputs = nullptr; @@ -50,6 +53,11 @@ struct ggml_opt_context { struct ggml_cgraph * gf = nullptr; struct ggml_cgraph * gb_grad = nullptr; struct ggml_cgraph * gb_opt = nullptr; + bool static_graphs = false; + bool eval_ready = false; + std::vector grad_accs; + std::vector grad_m; + std::vector grad_v; int64_t iter = 1; int32_t opt_period = 1; @@ -73,7 +81,13 @@ struct ggml_opt_result { // ====== Dataset ====== -ggml_opt_dataset_t ggml_opt_dataset_init(int64_t ne_datapoint, int64_t ne_label, int64_t ndata, int64_t ndata_shard) { +ggml_opt_dataset_t ggml_opt_dataset_init( + enum ggml_type type_data, + enum ggml_type type_label, + int64_t ne_datapoint, + int64_t ne_label, + int64_t ndata, + int64_t ndata_shard) { GGML_ASSERT(ne_datapoint > 0); GGML_ASSERT(ne_label >= 0); GGML_ASSERT(ndata > 0); @@ -92,11 +106,11 @@ ggml_opt_dataset_t ggml_opt_dataset_init(int64_t ne_datapoint, int64_t ne_label, result->ctx = ggml_init(params); } - result->data = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_datapoint, ndata); + result->data = ggml_new_tensor_2d(result->ctx, type_data, ne_datapoint, ndata); result->nbs_data = ggml_nbytes(result->data) * ndata_shard/ndata; if (ne_label > 0) { - result->labels = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_label, ndata); + result->labels = ggml_new_tensor_2d(result->ctx, type_label, ne_label, ndata); result->nbs_labels = ggml_nbytes(result->labels) * ndata_shard/ndata; } else { result->labels = nullptr; @@ -119,6 +133,10 @@ void ggml_opt_dataset_free(ggml_opt_dataset_t dataset) { delete dataset; } +int64_t ggml_opt_dataset_ndata(ggml_opt_dataset_t dataset) { + return dataset->ndata; +} + struct ggml_tensor * ggml_opt_dataset_data(ggml_opt_dataset_t dataset) { return dataset->data; } @@ -144,6 +162,8 @@ void ggml_opt_dataset_get_batch(ggml_opt_dataset_t dataset, struct ggml_tensor * GGML_ASSERT( data_batch && ggml_is_contiguous(data_batch)); GGML_ASSERT(!labels_batch || ggml_is_contiguous(labels_batch)); GGML_ASSERT((labels_batch == nullptr) == (dataset->labels == nullptr)); + GGML_ASSERT( data_batch->type == dataset->data->type); + GGML_ASSERT(!labels_batch || labels_batch->type == dataset->labels->type); const size_t nb_data_batch = ggml_nbytes(data_batch); GGML_ASSERT(nb_data_batch % dataset->nbs_data == 0); @@ -171,6 +191,31 @@ void ggml_opt_dataset_get_batch(ggml_opt_dataset_t dataset, struct ggml_tensor * } } +void ggml_opt_dataset_get_batch_host(ggml_opt_dataset_t dataset, void * data_batch, size_t nb_data_batch, void * labels_batch, int64_t ibatch) { + GGML_ASSERT((labels_batch == nullptr) == (dataset->labels == nullptr)); + GGML_ASSERT(nb_data_batch % dataset->nbs_data == 0); + + const int64_t shards_per_batch = nb_data_batch / dataset->nbs_data; + + GGML_ASSERT((ibatch + 1)*shards_per_batch <= int64_t(dataset->permutation.size())); + + for (int64_t ishard_batch = 0; ishard_batch < shards_per_batch; ++ishard_batch) { + const int64_t ishard = dataset->permutation[ibatch*shards_per_batch + ishard_batch]; + + const char * ptr_data = (const char *) dataset->data->data + ishard *dataset->nbs_data; + char * ptr_data_batch = (char *) data_batch + ishard_batch*dataset->nbs_data; + memcpy(ptr_data_batch, ptr_data, dataset->nbs_data); + + if (!labels_batch) { + continue; + } + + const char * ptr_labels = (const char *) dataset->labels->data + ishard *dataset->nbs_labels; + char * ptr_labels_batch = (char *) labels_batch + ishard_batch*dataset->nbs_labels; + memcpy(ptr_labels_batch, ptr_labels, dataset->nbs_labels); + } +} + // ====== Model / Context ====== struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata) { @@ -187,17 +232,18 @@ struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * us return result; } +struct ggml_opt_optimizer_params ggml_opt_get_constant_optimizer_params(void * userdata) { + return *((struct ggml_opt_optimizer_params *) userdata); +} + struct ggml_opt_params ggml_opt_default_params( ggml_backend_sched_t backend_sched, - struct ggml_context * ctx_compute, - struct ggml_tensor * inputs, - struct ggml_tensor * outputs, enum ggml_opt_loss_type loss_type) { return { /*backend_sched =*/ backend_sched, - /*ctx_compute =*/ ctx_compute, - /*inputs =*/ inputs, - /*logits =*/ outputs, + /*ctx_compute =*/ nullptr, + /*inputs =*/ nullptr, + /*logits =*/ nullptr, /*loss_type =*/ loss_type, /*build_type =*/ GGML_OPT_BUILD_TYPE_OPT, /*opt_period =*/ 1, @@ -266,195 +312,246 @@ static ggml_cgraph * dup_graph(ggml_context * ctx, ggml_cgraph * src) { return dst; } -static void ggml_opt_alloc_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph) { - GGML_ASSERT(graph); - if (opt_ctx->allocated_graph == graph) { - return; - } +static void ggml_opt_build(ggml_opt_context_t opt_ctx) { + GGML_ASSERT(opt_ctx->ctx_compute && "no compute context set, either use static graphs or set one with ggml_opt_prepare_alloc"); + GGML_ASSERT((!opt_ctx->static_graphs || opt_ctx->inputs->data) && "when using static graphs the inputs must be allocated statically"); - ggml_backend_sched_reset(opt_ctx->backend_sched); // clear allocation of previous graph + const bool accumulate = opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_GRAD && + !(opt_ctx->static_graphs && opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT && opt_ctx->opt_period == 1); - { - ggml_init_params params = { - /*.mem_size =*/ ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE, - /*.mem_buffer =*/ nullptr, - /*.no_alloc =*/ true, - }; - ggml_free(opt_ctx->ctx_copy); - opt_ctx->ctx_copy = ggml_init(params); - } - - opt_ctx->allocated_graph_copy = dup_graph(opt_ctx->ctx_copy, graph); - - ggml_backend_sched_alloc_graph(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy); - opt_ctx->allocated_graph = graph; -} - -ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) { - ggml_opt_context_t result = new struct ggml_opt_context; - result->backend_sched = params.backend_sched; - result->ctx_compute = params.ctx_compute; - result->inputs = params.inputs; - result->outputs = params.outputs; - result->opt_period = params.opt_period; - result->get_opt_pars = params.get_opt_pars; - result->get_opt_pars_ud = params.get_opt_pars_ud; - - GGML_ASSERT(result->inputs->data && "the inputs must be allocated statically"); - GGML_ASSERT(result->opt_period >= 1); - - const bool accumulate = params.build_type == GGML_OPT_BUILD_TYPE_GRAD || - (params.build_type == GGML_OPT_BUILD_TYPE_OPT && result->opt_period > 1); - - ggml_set_input(result->inputs); - ggml_set_output(result->outputs); - - result->gf = ggml_new_graph_custom(result->ctx_compute, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true); // Forward pass. - ggml_build_forward_expand(result->gf, result->outputs); + ggml_set_input(opt_ctx->inputs); + ggml_set_output(opt_ctx->outputs); int n_param = 0; - for (int i = 0; i < result->gf->n_nodes; ++i) { - if (result->gf->nodes[i]->flags & GGML_TENSOR_FLAG_PARAM) { + for (int i = 0; i < opt_ctx->gf->n_nodes; ++i) { + const struct ggml_tensor * node = opt_ctx->gf->nodes[i]; + if (node->flags & GGML_TENSOR_FLAG_PARAM) { n_param++; } + GGML_ASSERT(!(node->flags & GGML_TENSOR_FLAG_LOSS) && "support for extra loss terms not implemented"); } - { + if (!opt_ctx->ctx_static) { // The static context is used for: - // - gradients (1 tensor per param if using gradient accumulation) + // - gradients (1 per loss, 1 tensor per param if using gradient accumulation) // - optimizer momenta (2 tensors per param) - // - labels - // - loss + its gradient (up to 5 tensors) - // - pred - // - ncorrect (2 tensors). - const size_t tensors_per_param = (accumulate ? 1 : 0) + (params.build_type == GGML_OPT_BUILD_TYPE_OPT ? 2 : 0); - const size_t size_meta = (tensors_per_param*n_param + 9) * ggml_tensor_overhead(); + // - labels (if using static graphs) + // - loss (if using static graphs, up to 5 tensors) + // - pred (if using static graphs) + // - ncorrect (if using static graphs, 2 tensors). + constexpr size_t n_loss = 1; + const size_t tensors_per_param = (accumulate ? 1 : 0) + + (opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT ? 2 : 0); + const size_t tensors_const = opt_ctx->static_graphs ? 9 : 0; + const size_t size_meta = (n_loss + tensors_per_param*n_param + tensors_const) * ggml_tensor_overhead(); struct ggml_init_params params = { /*.mem_size =*/ size_meta, /*.mem_buffer =*/ nullptr, /*.no_alloc =*/ true, }; - result->ctx_static = ggml_init(params); + opt_ctx->ctx_static = ggml_init(params); } + GGML_ASSERT(opt_ctx->build_type <= opt_ctx->build_type_alloc); + { - // The static cpu context is used for: - // - optimizer parameters (1 for the entire context) + // The cpu context is allocated statically if using static graphs, dynamically otherwise. + // It is used for: + // - optimizer parameters (1 shared for all optimizer invocations) const size_t size_meta = 1 * ggml_tensor_overhead(); struct ggml_init_params params = { /*.mem_size =*/ size_meta, /*.mem_buffer =*/ nullptr, /*.no_alloc =*/ true, }; - result->ctx_static_cpu = ggml_init(params); + ggml_free(opt_ctx->ctx_cpu); + opt_ctx->ctx_cpu = ggml_init(params); + + ggml_backend_buffer_free(opt_ctx->buf_cpu); + opt_ctx->buf_cpu = nullptr; } + struct ggml_context * ctx_results = opt_ctx->static_graphs ? opt_ctx->ctx_static : opt_ctx->ctx_compute; - switch (params.loss_type) { + switch (opt_ctx->loss_type) { case GGML_OPT_LOSS_TYPE_MEAN: { - result->loss = ggml_sum(result->ctx_static, result->outputs); - ggml_set_name(result->loss, "loss_sum"); - const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs)); - result->loss = ggml_scale(result->ctx_static, result->loss, scale); - ggml_set_name(result->loss, "loss_mean"); - result->loss_per_datapoint = true; + opt_ctx->loss = ggml_sum(ctx_results, opt_ctx->outputs); + ggml_set_name(opt_ctx->loss, "loss_sum"); + const float scale = 1.0f / (opt_ctx->opt_period * ggml_nelements(opt_ctx->outputs)); + opt_ctx->loss = ggml_scale(ctx_results, opt_ctx->loss, scale); + ggml_set_name(opt_ctx->loss, "loss_mean"); + opt_ctx->loss_per_datapoint = true; break; } case GGML_OPT_LOSS_TYPE_SUM: { - result->loss = ggml_sum(result->ctx_static, result->outputs); - ggml_set_name(result->loss, "loss_sum"); - result->loss_per_datapoint = false; + opt_ctx->loss = ggml_sum(ctx_results, opt_ctx->outputs); + ggml_set_name(opt_ctx->loss, "loss_sum"); + opt_ctx->loss_per_datapoint = false; break; } case GGML_OPT_LOSS_TYPE_CROSS_ENTROPY: { - result->labels = ggml_dup_tensor(result->ctx_static, result->outputs); - ggml_set_input(result->labels); - ggml_set_name(result->labels, "labels"); - result->loss = ggml_cross_entropy_loss(result->ctx_static, result->outputs, result->labels); - ggml_set_name(result->loss, "loss_cross_entropy"); - if (result->opt_period > 1) { - result->loss = ggml_scale(result->ctx_static, result->loss, 1.0f / result->opt_period); - ggml_set_name(result->loss, "loss_cross_entropy_scaled"); + opt_ctx->labels = ggml_dup_tensor(ctx_results, opt_ctx->outputs); + ggml_set_input(opt_ctx->labels); + ggml_set_name(opt_ctx->labels, "labels"); + opt_ctx->loss = ggml_cross_entropy_loss(ctx_results, opt_ctx->outputs, opt_ctx->labels); + ggml_set_name(opt_ctx->loss, "loss_cross_entropy"); + if (opt_ctx->opt_period > 1) { + opt_ctx->loss = ggml_scale(ctx_results, opt_ctx->loss, 1.0f / opt_ctx->opt_period); + ggml_set_name(opt_ctx->loss, "loss_cross_entropy_scaled"); } - result->loss_per_datapoint = true; + opt_ctx->loss_per_datapoint = true; break; } case GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR: { - result->labels = ggml_dup_tensor(result->ctx_static, result->outputs); - ggml_set_input(result->labels); - ggml_set_name(result->labels, "labels"); - result->loss = ggml_sub(result->ctx_static, result->outputs, result->labels); - ggml_set_name(result->loss, "loss_error"); - result->loss = ggml_sqr(result->ctx_static, result->loss); - ggml_set_name(result->loss, "loss_squared_error"); - result->loss = ggml_sum(result->ctx_static, result->loss); - ggml_set_name(result->loss, "loss_sum_squared_error"); - const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs)); - result->loss = ggml_scale(result->ctx_static, result->loss, scale); - ggml_set_name(result->loss, "loss_mean_squared_error"); - result->loss_per_datapoint = true; + opt_ctx->labels = ggml_dup_tensor(ctx_results, opt_ctx->outputs); + ggml_set_input(opt_ctx->labels); + ggml_set_name(opt_ctx->labels, "labels"); + opt_ctx->loss = ggml_sub(ctx_results, opt_ctx->outputs, opt_ctx->labels); + ggml_set_name(opt_ctx->loss, "loss_error"); + opt_ctx->loss = ggml_sqr(ctx_results, opt_ctx->loss); + ggml_set_name(opt_ctx->loss, "loss_squared_error"); + opt_ctx->loss = ggml_sum(ctx_results, opt_ctx->loss); + ggml_set_name(opt_ctx->loss, "loss_sum_squared_error"); + const float scale = 1.0f / (opt_ctx->opt_period * ggml_nelements(opt_ctx->outputs)); + opt_ctx->loss = ggml_scale(ctx_results, opt_ctx->loss, scale); + ggml_set_name(opt_ctx->loss, "loss_mean_squared_error"); + opt_ctx->loss_per_datapoint = true; break; } } - ggml_set_output(result->loss); - ggml_set_loss(result->loss); - ggml_build_forward_expand(result->gf, result->loss); + ggml_set_output(opt_ctx->loss); + ggml_set_loss(opt_ctx->loss); + ggml_build_forward_expand(opt_ctx->gf, opt_ctx->loss); - result->pred = ggml_argmax(result->ctx_static, result->outputs); - ggml_set_name(result->pred, "pred"); - ggml_set_output(result->pred); - ggml_build_forward_expand(result->gf, result->pred); + if (opt_ctx->loss_type == GGML_OPT_LOSS_TYPE_CROSS_ENTROPY) { + opt_ctx->pred = ggml_argmax(ctx_results, opt_ctx->outputs); + ggml_set_name(opt_ctx->pred, "pred"); + ggml_set_output(opt_ctx->pred); + ggml_build_forward_expand(opt_ctx->gf, opt_ctx->pred); - if (result->labels) { - result->ncorrect = ggml_count_equal(result->ctx_static, result->pred, ggml_argmax(result->ctx_static, result->labels)); - ggml_set_name(result->ncorrect, "ncorrect"); - ggml_set_output(result->ncorrect); - ggml_build_forward_expand(result->gf, result->ncorrect); - } else { - result->ncorrect = nullptr; + opt_ctx->ncorrect = ggml_count_equal(ctx_results, opt_ctx->pred, ggml_argmax(ctx_results, opt_ctx->labels)); + ggml_set_name(opt_ctx->ncorrect, "ncorrect"); + ggml_set_output(opt_ctx->ncorrect); + ggml_build_forward_expand(opt_ctx->gf, opt_ctx->ncorrect); } - if (params.build_type == GGML_OPT_BUILD_TYPE_FORWARD) { - result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0)); - return result; + if (opt_ctx->buf_static) { + if (opt_ctx->build_type == GGML_OPT_BUILD_TYPE_FORWARD) { + return; + } + } else if (opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_FORWARD) { + opt_ctx->buf_static = ggml_backend_alloc_ctx_tensors( + opt_ctx->ctx_static, ggml_backend_sched_get_backend(opt_ctx->backend_sched, 0)); + return; } - // gb_grad == graph backward gradients, forward pass, then backward pass to calculate gradients. - result->gb_grad = ggml_graph_dup(result->ctx_compute, result->gf); - ggml_build_backward_expand(result->ctx_static, result->ctx_compute, result->gb_grad, accumulate); + if (opt_ctx->grad_accs.empty()) { + GGML_ASSERT(opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_GRAD); - if (params.build_type == GGML_OPT_BUILD_TYPE_GRAD) { - result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0)); - ggml_graph_reset(result->gb_grad); - return result; - } + const int n_nodes = opt_ctx->gf->n_nodes; + opt_ctx->grad_accs.resize(n_nodes); + for (int i = 0; i < n_nodes; ++i) { + ggml_tensor * node = opt_ctx->gf->nodes[i]; + if ((accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) || (node->flags & GGML_TENSOR_FLAG_LOSS)) { + opt_ctx->grad_accs[i] = ggml_new_tensor(opt_ctx->ctx_static, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne); + } else { + opt_ctx->grad_accs[i] = nullptr; + } + } - GGML_ASSERT(params.build_type == GGML_OPT_BUILD_TYPE_OPT); - - // gb_opt == graph backward optimize, forward pass, then backward pass to calculate gradients, then optimizer step. - result->gb_opt = ggml_graph_dup(result->ctx_compute, result->gb_grad); - - result->adamw_params = ggml_new_tensor_1d(result->ctx_static_cpu, GGML_TYPE_F32, 7); - ggml_set_input(result->adamw_params); - ggml_set_name(result->adamw_params, "adamw_params"); - - for (int i = result->gf->n_nodes-1; i >= 0; --i) { - struct ggml_tensor * node = result->gb_opt->nodes[i]; - struct ggml_tensor * grad = ggml_graph_get_grad(result->gb_opt, node); - - if (node->flags & GGML_TENSOR_FLAG_PARAM) { - struct ggml_tensor * m = ggml_dup_tensor(result->ctx_static, node); - struct ggml_tensor * v = ggml_dup_tensor(result->ctx_static, node); - struct ggml_tensor * opt_step = ggml_opt_step_adamw(result->ctx_compute, node, grad, m, v, result->adamw_params); - ggml_build_forward_expand(result->gb_opt, opt_step); + if (opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_OPT) { + opt_ctx->grad_m.resize(n_nodes); + opt_ctx->grad_v.resize(n_nodes); + for (int i = 0; i < n_nodes; ++i) { + ggml_tensor * node = opt_ctx->gf->nodes[i]; + if (node->flags & GGML_TENSOR_FLAG_PARAM) { + opt_ctx->grad_m[i] = ggml_new_tensor(opt_ctx->ctx_static, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne); + opt_ctx->grad_v[i] = ggml_new_tensor(opt_ctx->ctx_static, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne); + } else { + opt_ctx->grad_m[i] = nullptr; + opt_ctx->grad_v[i] = nullptr; + } + } } } - result->buf_static = ggml_backend_alloc_ctx_tensors( - result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0)); + // gb_grad == graph backward gradients, forward pass, then backward pass to calculate gradients. + opt_ctx->gb_grad = ggml_graph_dup(opt_ctx->ctx_compute, opt_ctx->gf, /*force_grads =*/ true); + ggml_build_backward_expand(opt_ctx->ctx_compute, opt_ctx->gb_grad, opt_ctx->grad_accs.data()); - result->buf_static_cpu = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx_static_cpu, ggml_backend_cpu_buffer_type()); + if (opt_ctx->buf_static) { + if (opt_ctx->build_type == GGML_OPT_BUILD_TYPE_GRAD) { + return; + } + } else if (opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_GRAD) { + opt_ctx->buf_static = ggml_backend_alloc_ctx_tensors(opt_ctx->ctx_static, ggml_backend_sched_get_backend(opt_ctx->backend_sched, 0)); + ggml_graph_reset(opt_ctx->gb_grad); + } - ggml_graph_reset(result->gb_opt); + GGML_ASSERT(opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT); + + // gb_opt == graph backward optimize, forward pass, then backward pass to calculate gradients, then optimizer step. + opt_ctx->gb_opt = ggml_graph_dup(opt_ctx->ctx_compute, opt_ctx->gb_grad, /*force_grads =*/ true); + + opt_ctx->adamw_params = ggml_new_tensor_1d(opt_ctx->ctx_cpu, GGML_TYPE_F32, 7); + ggml_set_input(opt_ctx->adamw_params); + ggml_set_name(opt_ctx->adamw_params, "adamw_params"); + + for (int i = opt_ctx->gf->n_nodes-1; i >= 0; --i) { + struct ggml_tensor * node = opt_ctx->gb_opt->nodes[i]; + struct ggml_tensor * grad = ggml_graph_get_grad(opt_ctx->gb_opt, node); + + if (grad && (node->flags & GGML_TENSOR_FLAG_PARAM)) { + struct ggml_tensor * m = opt_ctx->grad_m[i]; + struct ggml_tensor * v = opt_ctx->grad_v[i]; + struct ggml_tensor * opt_step = ggml_opt_step_adamw(opt_ctx->ctx_compute, node, grad, m, v, opt_ctx->adamw_params); + + ggml_set_name(m, (std::string("AdamW m for ") + std::string(node->name)).c_str()); + ggml_set_name(v, (std::string("AdamW v for ") + std::string(node->name)).c_str()); + ggml_set_name(opt_step, (std::string("AdamW step for ") + std::string(node->name)).c_str()); + + ggml_build_forward_expand(opt_ctx->gb_opt, opt_step); + } + } + + if (!opt_ctx->buf_static) { + opt_ctx->buf_static = ggml_backend_alloc_ctx_tensors( + opt_ctx->ctx_static, ggml_backend_sched_get_backend(opt_ctx->backend_sched, 0)); + ggml_graph_reset(opt_ctx->gb_opt); + } + + opt_ctx->buf_cpu = ggml_backend_alloc_ctx_tensors_from_buft(opt_ctx->ctx_cpu, ggml_backend_cpu_buffer_type()); +} + +ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) { + ggml_opt_context_t result = new struct ggml_opt_context; + result->backend_sched = params.backend_sched; + result->ctx_compute = params.ctx_compute; + result->loss_type = params.loss_type; + result->build_type = params.build_type; + result->build_type_alloc = params.build_type; + result->inputs = params.inputs; + result->outputs = params.outputs; + result->opt_period = params.opt_period; + result->get_opt_pars = params.get_opt_pars; + result->get_opt_pars_ud = params.get_opt_pars_ud; + + GGML_ASSERT(result->opt_period >= 1); + + result->static_graphs = result->ctx_compute; + + if (!result->static_graphs) { + GGML_ASSERT(!result->inputs); + GGML_ASSERT(!result->outputs); + return result; + } + + GGML_ASSERT(result->inputs); + GGML_ASSERT(result->outputs); + + result->gf = ggml_new_graph_custom(result->ctx_compute, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true); // Forward pass. + ggml_build_forward_expand(result->gf, result->outputs); + + ggml_opt_build(result); return result; } @@ -464,9 +561,9 @@ void ggml_opt_free(ggml_opt_context_t opt_ctx) { return; } ggml_backend_buffer_free(opt_ctx->buf_static); - ggml_backend_buffer_free(opt_ctx->buf_static_cpu); + ggml_backend_buffer_free(opt_ctx->buf_cpu); ggml_free(opt_ctx->ctx_static); - ggml_free(opt_ctx->ctx_static_cpu); + ggml_free(opt_ctx->ctx_cpu); delete opt_ctx; } @@ -582,8 +679,79 @@ void ggml_opt_result_accuracy(ggml_opt_result_t result, double * accuracy, doubl // ====== Computation ====== -static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph, ggml_opt_result * result) { - if (graph != opt_ctx->gf) { +void ggml_opt_prepare_alloc( + ggml_opt_context_t opt_ctx, + struct ggml_context * ctx_compute, + struct ggml_cgraph * gf, + struct ggml_tensor * inputs, + struct ggml_tensor * outputs) { + GGML_ASSERT(!opt_ctx->static_graphs); + opt_ctx->ctx_compute = ctx_compute; + opt_ctx->gf = gf; + opt_ctx->inputs = inputs; + opt_ctx->outputs = outputs; +} + +void ggml_opt_alloc(ggml_opt_context_t opt_ctx, bool backward) { + GGML_ASSERT(!opt_ctx->eval_ready); + if (opt_ctx->build_type == GGML_OPT_BUILD_TYPE_OPT && opt_ctx->opt_period > 1 && opt_ctx->opt_i == 0) { + ggml_graph_reset(opt_ctx->gb_grad); + } + if (backward) { + const int32_t opt_i_next = (opt_ctx->opt_i + 1) % opt_ctx->opt_period; + opt_ctx->build_type = opt_i_next == 0 ? GGML_OPT_BUILD_TYPE_OPT : GGML_OPT_BUILD_TYPE_GRAD; + } else { + opt_ctx->build_type = GGML_OPT_BUILD_TYPE_FORWARD; + } + + if (!opt_ctx->static_graphs) { + ggml_opt_build(opt_ctx); + } + + struct ggml_cgraph * graph = nullptr; + switch (opt_ctx->build_type) { + case GGML_OPT_BUILD_TYPE_FORWARD: { + graph = opt_ctx->gf; + } break; + case GGML_OPT_BUILD_TYPE_GRAD: { + graph = opt_ctx->gb_grad; + } break; + case GGML_OPT_BUILD_TYPE_OPT: { + graph = opt_ctx->gb_opt; + } break; + } + GGML_ASSERT(graph); + + if (opt_ctx->allocated_graph == graph) { + opt_ctx->eval_ready = true; + return; + } + + ggml_backend_sched_reset(opt_ctx->backend_sched); // clear allocation of previous graph + + if (opt_ctx->static_graphs) { + ggml_init_params params = { + /*.mem_size =*/ graph->size*ggml_tensor_overhead() + ggml_graph_overhead_custom(graph->size, graph->grads), + /*.mem_buffer =*/ nullptr, + /*.no_alloc =*/ true, + }; + ggml_free(opt_ctx->ctx_copy); + opt_ctx->ctx_copy = ggml_init(params); + + opt_ctx->allocated_graph_copy = dup_graph(opt_ctx->ctx_copy, graph); + } else { + opt_ctx->allocated_graph_copy = graph; + } + + ggml_backend_sched_alloc_graph(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy); + opt_ctx->allocated_graph = graph; + + opt_ctx->eval_ready = true; +} + +void ggml_opt_eval(ggml_opt_context_t opt_ctx, ggml_opt_result_t result) { + GGML_ASSERT(opt_ctx->eval_ready); + if (opt_ctx->allocated_graph == opt_ctx->gb_opt) { struct ggml_opt_optimizer_params opt_pars = opt_ctx->get_opt_pars(opt_ctx->get_opt_pars_ud); GGML_ASSERT(opt_pars.adamw.alpha > 0.0f); @@ -609,9 +777,19 @@ static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph, adamw_par_data[6] = beta2h; } - ggml_opt_alloc_graph(opt_ctx, graph); ggml_backend_sched_graph_compute(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy); opt_ctx->iter += opt_ctx->allocated_graph == opt_ctx->gb_opt; + opt_ctx->opt_i = (opt_ctx->opt_i + 1) % opt_ctx->opt_period; + + if (!opt_ctx->static_graphs) { + opt_ctx->gf = nullptr; + opt_ctx->gb_grad = nullptr; + opt_ctx->gb_opt = nullptr; + opt_ctx->allocated_graph = nullptr; + opt_ctx->allocated_graph_copy = nullptr; + } + + opt_ctx->eval_ready = false; if (!result) { return; @@ -635,12 +813,14 @@ static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph, ggml_backend_tensor_get(opt_ctx->loss, &loss, 0, ggml_nbytes(opt_ctx->loss)); result->loss.push_back(loss); - GGML_ASSERT(opt_ctx->pred->type == GGML_TYPE_I32); - std::vector pred(ndata); - ggml_backend_tensor_get(opt_ctx->pred, pred.data(), 0, ggml_nbytes(opt_ctx->pred)); - result->pred.insert(result->pred.end(), pred.begin(), pred.end()); + if (opt_ctx->pred) { + GGML_ASSERT(opt_ctx->pred->type == GGML_TYPE_I32); + std::vector pred(ndata); + ggml_backend_tensor_get(opt_ctx->pred, pred.data(), 0, ggml_nbytes(opt_ctx->pred)); + result->pred.insert(result->pred.end(), pred.begin(), pred.end()); + } - if (!opt_ctx->labels || result->ncorrect < 0) { + if (!opt_ctx->ncorrect || result->ncorrect < 0) { result->ncorrect = -1; return; } @@ -652,26 +832,6 @@ static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph, result->ncorrect += ncorrect; } -void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result * result) { - ggml_opt_eval_graph(opt_ctx, opt_ctx->gf, result); -} - -void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result * result) { - if (opt_ctx->opt_period == 1) { - ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_opt, result); - return; - } - - const int32_t opt_i_next = (opt_ctx->opt_i + 1) % opt_ctx->opt_period; - if (opt_i_next == 0) { - ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_opt, result); - ggml_opt_reset(opt_ctx, /*optimizer =*/ false); - } else { - ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_grad, result); - } - opt_ctx->opt_i = opt_i_next; -} - // ====== High-Level Functions ====== void ggml_opt_epoch( @@ -700,16 +860,18 @@ void ggml_opt_epoch( int64_t ibatch = 0; int64_t t_loop_start = ggml_time_us(); for (; ibatch < ibatch_split; ++ibatch) { + ggml_opt_alloc(opt_ctx, /*backward =*/ true); ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch); - ggml_opt_forward_backward(opt_ctx, result_train); + ggml_opt_eval(opt_ctx, result_train); if (callback_train) { callback_train(true, opt_ctx, dataset, result_train, ibatch+1, ibatch_split, t_loop_start); } } t_loop_start = ggml_time_us(); for (; ibatch < nbatches; ++ibatch) { + ggml_opt_alloc(opt_ctx, /*backward =*/ false); ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch); - ggml_opt_forward(opt_ctx, result_eval); + ggml_opt_eval(opt_ctx, result_eval); if (callback_eval) { callback_eval(false, opt_ctx, dataset, result_eval, ibatch+1-ibatch_split, nbatches-ibatch_split, t_loop_start); } @@ -726,13 +888,26 @@ void ggml_opt_epoch_callback_progress_bar( int64_t t_start_us) { fprintf(stderr, "%s[", train ? "train: " : "val: "); - constexpr int64_t bar_length = 25; + // The progress bar consists of partially filled blocks, unicode has 8 separate fill levels. + constexpr int64_t bar_length = 8; + const int64_t ibatch8 = 8 * ibatch; for (int64_t j = 0; j < bar_length; ++j) { - const int64_t ibatch_j = ibatch_max * j/bar_length; - if (ibatch_j < ibatch) { - fprintf(stderr, "="); - } else if (ibatch_max * (j - 1)/bar_length < ibatch) { - fprintf(stderr, ">"); + if (ibatch_max * (8*j + 8) / bar_length < ibatch8) { + fprintf(stderr, "\u2588"); // full block + } else if (ibatch_max * (8*j + 7) / bar_length < ibatch8) { + fprintf(stderr, "\u2589"); // 7/8 filled + } else if (ibatch_max * (8*j + 6) / bar_length < ibatch8) { + fprintf(stderr, "\u258A"); // 6/8 filled + } else if (ibatch_max * (8*j + 5) / bar_length < ibatch8) { + fprintf(stderr, "\u258B"); // 5/8 filled + } else if (ibatch_max * (8*j + 4) / bar_length < ibatch8) { + fprintf(stderr, "\u258C"); // 4/8 filled + } else if (ibatch_max * (8*j + 3) / bar_length < ibatch8) { + fprintf(stderr, "\u258D"); // 3/8 filled + } else if (ibatch_max * (8*j + 2) / bar_length < ibatch8) { + fprintf(stderr, "\u258E"); // 2/8 filled + } else if (ibatch_max * (8*j + 1) / bar_length < ibatch8) { + fprintf(stderr, "\u258F"); // 1/8 filled } else { fprintf(stderr, " "); } @@ -764,8 +939,8 @@ void ggml_opt_epoch_callback_progress_bar( const int64_t t_eta_m = t_eta_s / 60; t_eta_s -= t_eta_m * 60; - fprintf(stderr, "| data=%06" PRId64 "/%06" PRId64 ", loss=%.6lf+-%.6lf, accuracy=%.2lf+-%.2lf%%, " - "t=%02" PRId64 ":%02" PRId64 ":%02" PRId64 ", ETA=%02" PRId64 ":%02" PRId64 ":%02" PRId64 "]\r", + fprintf(stderr, "] data=%07" PRId64 "/%07" PRId64 " loss=%.5lf±%.5lf acc=%.2lf±%.2lf%% " + "t=%02" PRId64 ":%02" PRId64 ":%02" PRId64 " ETA=%02" PRId64 ":%02" PRId64 ":%02" PRId64 " \r", idata, idata_max, loss, loss_unc, 100.0*accuracy, 100.0*accuracy_unc, t_ibatch_h, t_ibatch_m, t_ibatch_s, t_eta_h, t_eta_m, t_eta_s); if (ibatch == ibatch_max) { @@ -806,7 +981,10 @@ void ggml_opt_fit( int64_t epoch = 1; - ggml_opt_params params = ggml_opt_default_params(backend_sched, ctx_compute, inputs, outputs, loss_type); + ggml_opt_params params = ggml_opt_default_params(backend_sched, loss_type); + params.ctx_compute = ctx_compute; + params.inputs = inputs; + params.outputs = outputs; params.opt_period = opt_period; params.get_opt_pars = get_opt_pars; params.get_opt_pars_ud = &epoch; diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c index ac918a60d9..84ec6dfe31 100644 --- a/ggml/src/ggml-quants.c +++ b/ggml/src/ggml-quants.c @@ -19,12 +19,6 @@ #define GROUP_MAX_EPS_IQ1_M 1e-7f #define GROUP_MAX_EPS_IQ1_S 1e-12f -#if defined(_MSC_VER) -// disable "possible loss of data" to avoid warnings for hundreds of casts -// we should just be careful :) -#pragma warning(disable: 4244 4267) -#endif - #define UNUSED GGML_UNUSED // reference implementation for deterministic creation of model files diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp index 9023eb0919..4f0abb5a60 100644 --- a/ggml/src/ggml-rpc/ggml-rpc.cpp +++ b/ggml/src/ggml-rpc/ggml-rpc.cpp @@ -151,6 +151,12 @@ struct rpc_msg_buffer_clear_req { uint8_t value; }; +struct rpc_msg_set_tensor_hash_req { + rpc_tensor tensor; + uint64_t offset; + uint64_t hash; +}; + struct rpc_msg_set_tensor_hash_rsp { uint8_t result; }; @@ -518,6 +524,11 @@ static rpc_tensor serialize_tensor(const ggml_tensor * tensor) { result.view_src = reinterpret_cast(tensor->view_src); result.view_offs = tensor->view_offs; result.data = reinterpret_cast(tensor->data); + + // Avoid sending uninitialized data over the wire + memset(result.name, 0, sizeof(result.name)); + memset(result.padding, 0, sizeof(result.padding)); + snprintf(result.name, GGML_MAX_NAME, "%s", tensor->name); return result; } @@ -543,15 +554,12 @@ static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggm ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context; rpc_tensor rpc_tensor = serialize_tensor(tensor); if (size > HASH_THRESHOLD) { - // input serialization format: | rpc_tensor | offset (8 bytes) | hash (8 bytes) - size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + sizeof(uint64_t); - std::vector input(input_size, 0); - uint64_t hash = fnv_hash((const uint8_t*)data, size); - memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor)); - memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset)); - memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), &hash, sizeof(hash)); + rpc_msg_set_tensor_hash_req request; + request.tensor = rpc_tensor; + request.offset = offset; + request.hash = fnv_hash((const uint8_t*)data, size); rpc_msg_set_tensor_hash_rsp response; - bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR_HASH, input.data(), input.size(), &response, sizeof(response)); + bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR_HASH, &request, sizeof(request), &response, sizeof(response)); GGML_ASSERT(status); if (response.result) { // the server has the same data, no need to send it @@ -859,7 +867,7 @@ public: bool free_buffer(const rpc_msg_free_buffer_req & request); bool buffer_clear(const rpc_msg_buffer_clear_req & request); bool set_tensor(const std::vector & input); - bool set_tensor_hash(const std::vector & input, rpc_msg_set_tensor_hash_rsp & response); + bool set_tensor_hash(const rpc_msg_set_tensor_hash_req & request, rpc_msg_set_tensor_hash_rsp & response); bool get_tensor(const rpc_msg_get_tensor_req & request, std::vector & response); bool copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_copy_tensor_rsp & response); bool graph_compute(const std::vector & input, rpc_msg_graph_compute_rsp & response); @@ -982,8 +990,21 @@ bool rpc_server::buffer_clear(const rpc_msg_buffer_clear_req & request) { } ggml_tensor * rpc_server::deserialize_tensor(struct ggml_context * ctx, const rpc_tensor * tensor) { + // Validate tensor type before using it + if (tensor->type >= GGML_TYPE_COUNT) { + GGML_LOG_ERROR("[%s] invalid tensor type received: %u\n", __func__, tensor->type); + return nullptr; + } + ggml_tensor * result = ggml_new_tensor_4d(ctx, (ggml_type) tensor->type, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]); + + // ggml_new_tensor_4d might fail if dimensions are invalid, although less likely to crash than invalid type + if (result == nullptr) { + GGML_LOG_ERROR("[%s] ggml_new_tensor_4d failed for type %u\\n", __func__, tensor->type); + return nullptr; + } + for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) { result->nb[i] = tensor->nb[i]; } @@ -1043,7 +1064,9 @@ bool rpc_server::set_tensor(const std::vector & input) { const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer); if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) { - GGML_ABORT("[%s] tensor->data out of bounds\n", __func__); + GGML_LOG_ERROR("[%s] tensor data region (data=0x%" PRIx64 ", offset=%" PRIu64 ", size=%zu) out of buffer bounds [0x%zx, 0x%zx)\n", + __func__, in_tensor->data, offset, size, p0, p1); + return false; } } @@ -1081,18 +1104,10 @@ bool rpc_server::get_cached_file(uint64_t hash, std::vector & data) { return true; } -bool rpc_server::set_tensor_hash(const std::vector & input, rpc_msg_set_tensor_hash_rsp & response) +bool rpc_server::set_tensor_hash(const rpc_msg_set_tensor_hash_req & request, rpc_msg_set_tensor_hash_rsp & response) { - // serialization format: | rpc_tensor | offset (8 bytes) | hash (8 bytes) | - if (input.size() != sizeof(rpc_tensor) + 16) { - return false; - } - const rpc_tensor * in_tensor = (const rpc_tensor *)input.data(); - uint64_t offset; - memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset)); - const uint64_t * hash = (const uint64_t *)(input.data() + sizeof(rpc_tensor) + sizeof(offset)); std::vector cached_file; - if (!get_cached_file(*hash, cached_file)) { + if (!get_cached_file(request.hash, cached_file)) { response.result = 0; return true; } @@ -1105,23 +1120,28 @@ bool rpc_server::set_tensor_hash(const std::vector & input, rpc_msg_set ggml_context_ptr ctx_ptr { ggml_init(params) }; GGML_ASSERT(ctx_ptr != nullptr); ggml_context * ctx = ctx_ptr.get(); - ggml_tensor * tensor = deserialize_tensor(ctx, in_tensor); + ggml_tensor * tensor = deserialize_tensor(ctx, &request.tensor); if (tensor == nullptr) { GGML_LOG_ERROR("[%s] error deserializing tensor\n", __func__); return false; } - GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu, hash: %" PRIx64 "\n", __func__, (void*)tensor->buffer, tensor->data, offset, size, *hash); + GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu, hash: %" PRIx64 "\n", + __func__, (void*)tensor->buffer, tensor->data, request.offset, size, request.hash); // sanitize tensor->data { const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer); const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer); - if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) { - GGML_ABORT("[%s] tensor->data out of bounds\n", __func__); + if (request.tensor.data + request.offset < p0 + || request.tensor.data + request.offset >= p1 + || size > (p1 - request.tensor.data - request.offset)) { + GGML_LOG_ERROR("[%s] tensor data region (data=0x%" PRIx64 ", offset=%" PRIu64 ", size=%zu, hash=0x%" PRIx64 ") out of buffer bounds [0x%zx, 0x%zx)\n", + __func__, request.tensor.data, request.offset, size, request.hash, p0, p1); + return false; } } - ggml_backend_tensor_set(tensor, cached_file.data(), offset, size); + ggml_backend_tensor_set(tensor, cached_file.data(), request.offset, size); response.result = 1; return true; } @@ -1183,7 +1203,9 @@ bool rpc_server::get_tensor(const rpc_msg_get_tensor_req & request, std::vector< if (request.tensor.data + request.offset < p0 || request.tensor.data + request.offset >= p1 || request.size > (p1 - request.tensor.data - request.offset)) { - GGML_ABORT("[%s] tensor->data out of bounds\n", __func__); + GGML_LOG_ERROR("[%s] requested tensor region (data=0x%" PRIx64 ", offset=%" PRIu64 ", size=%" PRIu64 ") out of buffer bounds [0x%zx, 0x%zx)\n", + __func__, request.tensor.data, request.offset, request.size, p0, p1); + return false; } } @@ -1237,22 +1259,50 @@ ggml_tensor * rpc_server::create_node(uint64_t id, struct ggml_context * ctx, const std::unordered_map & tensor_ptrs, std::unordered_map & tensor_map) { - if (id == 0) { - return nullptr; - } if (tensor_map.find(id) != tensor_map.end()) { return tensor_map[id]; } - const rpc_tensor * tensor = tensor_ptrs.at(id); + // Safely find the tensor pointer + auto it_ptr = tensor_ptrs.find(id); + if (it_ptr == tensor_ptrs.end()) { + return nullptr; + } + const rpc_tensor * tensor = it_ptr->second; + struct ggml_tensor * result = 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] = create_node(tensor->src[i], ctx, tensor_ptrs, tensor_map); + // Check if the source ID is 0 before calling create_node recursively + if (tensor->src[i] == 0) { + result->src[i] = nullptr; + } else { + result->src[i] = create_node(tensor->src[i], ctx, tensor_ptrs, tensor_map); + // If the recursive call failed for a non-zero ID, propagate the error + if (result->src[i] == nullptr) { + GGML_LOG_ERROR("[%s] failed to create source node %d (src_id=%" PRIu64 ") for node id %" PRIu64 "\n", + __func__, i, tensor->src[i], id); + // Must return nullptr to signal failure up the call stack + return nullptr; + } + } + } + + // Handle view_src similarly + if (tensor->view_src == 0) { + result->view_src = nullptr; + } else { + result->view_src = create_node(tensor->view_src, ctx, tensor_ptrs, tensor_map); + // If the recursive call failed for a non-zero ID, propagate the error + if (result->view_src == nullptr) { + GGML_LOG_ERROR("[%s] failed to create view_src node (view_src_id=%" PRIu64 ") for node id %" PRIu64 "\n", + __func__, tensor->view_src, id); + // Must return nullptr to signal failure up the call stack + return nullptr; + } } - result->view_src = create_node(tensor->view_src, ctx, tensor_ptrs, tensor_map); result->view_offs = tensor->view_offs; return result; } @@ -1278,6 +1328,7 @@ bool rpc_server::graph_compute(const std::vector & input, rpc_msg_graph GGML_PRINT_DEBUG("[%s] n_nodes: %u, n_tensors: %u\n", __func__, n_nodes, n_tensors); size_t buf_size = ggml_tensor_overhead()*(n_nodes + n_tensors) + ggml_graph_overhead_custom(n_nodes, false); + struct ggml_init_params params = { /*.mem_size =*/ buf_size, /*.mem_buffer =*/ NULL, @@ -1297,6 +1348,14 @@ bool rpc_server::graph_compute(const std::vector & input, rpc_msg_graph int64_t id; memcpy(&id, &nodes[i], sizeof(id)); graph->nodes[i] = create_node(id, ctx, tensor_ptrs, tensor_map); + + // Check if create_node failed for a *non-zero* ID. + // If id was 0, create_node returning nullptr is expected. + // If id was non-zero and create_node returned nullptr, it indicates a deserialization error. + if (graph->nodes[i] == nullptr && id != 0) { + GGML_LOG_ERROR("[%s] failed to create graph node %d (id=%" PRId64 ")\n", __func__, i, id); + return false; + } } ggml_status status = ggml_backend_graph_compute(backend, graph); response.result = status; @@ -1361,7 +1420,9 @@ static void rpc_serve_client(ggml_backend_t backend, const char * cache_dir, return; } rpc_msg_get_alloc_size_rsp response; - server.get_alloc_size(request, response); + if (!server.get_alloc_size(request, response)) { + return; + } if (!send_msg(sockfd, &response, sizeof(response))) { return; } @@ -1440,12 +1501,12 @@ static void rpc_serve_client(ggml_backend_t backend, const char * cache_dir, break; } case RPC_CMD_SET_TENSOR_HASH: { - std::vector input; - if (!recv_msg(sockfd, input)) { + rpc_msg_set_tensor_hash_req request; + if (!recv_msg(sockfd, &request, sizeof(request))) { return; } rpc_msg_set_tensor_hash_rsp response; - if (!server.set_tensor_hash(input, response)) { + if (!server.set_tensor_hash(request, response)) { return; } if (!send_msg(sockfd, &response, sizeof(response))) { @@ -1531,6 +1592,14 @@ static void rpc_serve_client(ggml_backend_t backend, const char * cache_dir, void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, const char * cache_dir, size_t free_mem, size_t total_mem) { + printf("Starting RPC server v%d.%d.%d\n", + RPC_PROTO_MAJOR_VERSION, + RPC_PROTO_MINOR_VERSION, + RPC_PROTO_PATCH_VERSION); + printf(" endpoint : %s\n", endpoint); + printf(" local cache : %s\n", cache_dir ? cache_dir : "n/a"); + printf(" backend memory : %zu MB\n", free_mem / (1024 * 1024)); + std::string host; int port; if (!parse_endpoint(endpoint, host, port)) { @@ -1690,6 +1759,9 @@ static void * ggml_backend_rpc_get_proc_address(ggml_backend_reg_t reg, const ch if (std::strcmp(name, "ggml_backend_rpc_add_device") == 0) { return (void *)ggml_backend_rpc_add_device; } + if (std::strcmp(name, "ggml_backend_rpc_start_server") == 0) { + return (void *)ggml_backend_rpc_start_server; + } return NULL; GGML_UNUSED(reg); diff --git a/ggml/src/ggml-sycl/CMakeLists.txt b/ggml/src/ggml-sycl/CMakeLists.txt index 6699b70bad..231fb71dab 100644 --- a/ggml/src/ggml-sycl/CMakeLists.txt +++ b/ggml/src/ggml-sycl/CMakeLists.txt @@ -52,9 +52,8 @@ target_compile_options(ggml-sycl PRIVATE "-Wno-narrowing") find_package(DNNL) set(GGML_SYCL_DNNL 0) if(DNNL_FOUND) - if (DEFINED ENV{ONEAPI_ROOT} AND NOT DEFINED DNNL_GPU_VENDOR) - # Assuming oneDNN packaged with oneapi release is used which - # supports only intel target + if (NOT DEFINED DNNL_GPU_VENDOR) + # default to intel target set(DNNL_GPU_VENDOR "INTEL") if(NOT "${GGML_SYCL_TARGET}" STREQUAL "INTEL") message(WARNING "oneDNN builds bundled with oneapi release only support INTEL target") @@ -108,6 +107,9 @@ endif() if (GGML_SYCL_TARGET STREQUAL "INTEL") # Intel devices use Intel oneMKL directly instead of oneMath to avoid the limitation of linking Intel oneMKL statically # See https://github.com/uxlfoundation/oneMath/issues/654 + if (CMAKE_CXX_COMPILER_ID STREQUAL "Clang") + set(SYCL_COMPILER ON) + endif() find_package(MKL REQUIRED) target_link_libraries(ggml-sycl PRIVATE MKL::MKL_SYCL::BLAS) target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_USE_INTEL_ONEMKL) diff --git a/ggml/src/ggml-sycl/backend.hpp b/ggml/src/ggml-sycl/backend.hpp index de814ef91a..f78a36ddf8 100644 --- a/ggml/src/ggml-sycl/backend.hpp +++ b/ggml/src/ggml-sycl/backend.hpp @@ -14,23 +14,24 @@ #define GGML_SYCL_BACKEND_HPP #include "binbcast.hpp" -#include "concat.hpp" #include "common.hpp" +#include "concat.hpp" #include "conv.hpp" #include "convert.hpp" +#include "cpy.hpp" #include "dequantize.hpp" #include "dmmv.hpp" +#include "element_wise.hpp" +#include "gla.hpp" +#include "im2col.hpp" #include "mmq.hpp" #include "mmvq.hpp" -#include "rope.hpp" #include "norm.hpp" +#include "outprod.hpp" +#include "quants.hpp" +#include "rope.hpp" #include "softmax.hpp" #include "tsembd.hpp" -#include "im2col.hpp" #include "wkv.hpp" -#include "outprod.hpp" -#include "element_wise.hpp" -#include "cpy.hpp" -#include "gla.hpp" -#endif // GGML_SYCL_BACKEND_HPP +#endif // GGML_SYCL_BACKEND_HPP diff --git a/ggml/src/ggml-sycl/common.hpp b/ggml/src/ggml-sycl/common.hpp index 0ab0fb0aa3..60909dde7d 100644 --- a/ggml/src/ggml-sycl/common.hpp +++ b/ggml/src/ggml-sycl/common.hpp @@ -42,6 +42,7 @@ void ggml_sycl_host_free(void* ptr); extern int g_ggml_sycl_debug; extern int g_ggml_sycl_disable_optimize; +extern int g_ggml_sycl_prioritize_dmmv; #define GGML_SYCL_DEBUG(...) \ do { \ @@ -80,10 +81,6 @@ extern int g_ggml_sycl_disable_optimize; // max batch size to use MMQ kernels when tensor cores are available #define MMQ_MAX_BATCH_SIZE 32 -#if defined(_MSC_VER) -#pragma warning(disable : 4244 4267) // possible loss of data -#endif - // dmmv = dequantize_mul_mat_vec #ifndef GGML_SYCL_DMMV_X #define GGML_SYCL_DMMV_X 32 @@ -118,17 +115,12 @@ static void crash() { GGML_ABORT("SYCL error"); } -#define SYCL_CHECK(err) \ - do { \ - auto err_ = (err); \ - if (err_ != 0) \ - ggml_sycl_error( \ - #err, \ - __func__, \ - __FILE__, \ - __LINE__, \ - "Meet error in this line code!"); \ - } while (0) +#define SYCL_CHECK(err) \ + do { \ + auto err_ = (err); \ + if (err_ != 0) \ + ggml_sycl_error(#err, __func__, __FILE__, __LINE__, "Exception caught in this line of code."); \ + } while (0) #if DPCT_COMPAT_RT_VERSION >= 11100 #define GGML_SYCL_ASSUME(x) __builtin_assume(x) @@ -493,5 +485,9 @@ static __dpct_inline__ Tp* get_pointer(sycl::local_accessor acc) { int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block_size); +constexpr size_t ceil_div(const size_t m, const size_t n) { + return (m + n - 1) / n; +} + bool gpu_has_xmx(sycl::device &dev); #endif // GGML_SYCL_COMMON_HPP diff --git a/ggml/src/ggml-sycl/convert.cpp b/ggml/src/ggml-sycl/convert.cpp index 76ac6a4dd1..b2f8a65693 100644 --- a/ggml/src/ggml-sycl/convert.cpp +++ b/ggml/src/ggml-sycl/convert.cpp @@ -437,41 +437,52 @@ static void dequantize_row_iq4_nl_sycl(const void *vx, dst_t *y, const int64_t k } template -static void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, - const sycl::nd_item<3> &item_ct1) { +static void convert_unary_nc(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01, + const int64_t ne02, const int64_t s01, const int64_t s02, const int64_t s03, + const sycl::nd_item<3> & item_ct1) { + const int64_t work_group_size = item_ct1.get_local_range(2); - const int64_t global_id = item_ct1.get_local_id(2) + work_group_size * item_ct1.get_group(2); + const int64_t global_id = item_ct1.get_local_id(2) + work_group_size * item_ct1.get_group(2); + + const int64_t i01 = item_ct1.get_group(1); + const int64_t i02 = item_ct1.get_group(0) % ne02; + const int64_t i03 = item_ct1.get_group(0) / ne02; // make each work-item deal with more elements since sycl global range can not exceed max int - const src_t * x = (const src_t *) vx; - for (int64_t i = global_id; i < k; i += work_group_size * item_ct1.get_group_range(2)) { - y[i] = x[i]; + const src_t * x = static_cast(vx); + const int64_t ix = i03 * s03 + i02 * s02 + i01 * s01; + const int64_t iy = ((i03 * ne02 + i02) * ne01 + i01) * ne00; + +#pragma unroll + for (int64_t i00 = global_id; i00 < ne00; i00 += work_group_size * item_ct1.get_group_range(2)) { + y[iy + i00] = static_cast(x[ix + i00]); } } template -static void convert_unary_sycl(const void *__restrict__ vx, - dst_t *__restrict__ y, const int64_t k, - dpct::queue_ptr stream) { - const int64_t num_blocks = (k + SYCL_DEQUANTIZE_BLOCK_SIZE - 1) / SYCL_DEQUANTIZE_BLOCK_SIZE; +static void convert_unary_nc_sycl(const void * __restrict__ vx, dst_t * __restrict__ y, + const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03, + const int64_t s01, const int64_t s02, const int64_t s03, dpct::queue_ptr queue) { + dpct::has_capability_or_fail(queue->get_device(), { sycl::aspect::fp16 }); + + sycl::range<3> global_size(ne02 * ne03, ne01, ceil_div(ne00, SYCL_DEQUANTIZE_BLOCK_SIZE)); // decrease global range when it exceeds the max int - int64_t local_size = downsample_sycl_global_range(num_blocks, SYCL_DEQUANTIZE_BLOCK_SIZE); - sycl::range<3> block_nums(1, 1, num_blocks); - sycl::range<3> local_range(1, 1, local_size); - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); + // TODO: Downsample logic is separated from the kernel, a rewrite is desirable + int64_t downsized_workgroup = downsample_sycl_global_range(global_size[0], SYCL_DEQUANTIZE_BLOCK_SIZE); + sycl::range<3> workgroup_size(1, 1, downsized_workgroup); - stream->parallel_for( - sycl::nd_range<3>(block_nums * local_range, local_range), - [=](sycl::nd_item<3> item_ct1) { - convert_unary(vx, y, k, item_ct1); - }); - } + queue->parallel_for(sycl::nd_range<3>(global_size * workgroup_size, workgroup_size), [=](sycl::nd_item<3> item_ct1) { + convert_unary_nc(vx, y, ne00, ne01, ne02, s01, s02, s03, item_ct1); + }); } -to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor *dst) { +template +static void convert_unary_sycl(const void * vx, dst_t * y, const int64_t k, dpct::queue_ptr queue) { + convert_unary_nc_sycl(vx, y, k, 1, 1, 1, k, k, k, queue); +} + +to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) { switch (type) { case GGML_TYPE_Q4_0: if (dst->src[0]->extra && @@ -574,3 +585,12 @@ to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) { return nullptr; } } + +to_fp16_nc_sycl_t get_to_fp16_nc_sycl(ggml_type type) { + switch (type) { + case GGML_TYPE_F32: + return convert_unary_nc_sycl; + default: + return nullptr; + } +} diff --git a/ggml/src/ggml-sycl/convert.hpp b/ggml/src/ggml-sycl/convert.hpp index 355dae22b4..f8cb573e36 100644 --- a/ggml/src/ggml-sycl/convert.hpp +++ b/ggml/src/ggml-sycl/convert.hpp @@ -1,6 +1,6 @@ // // MIT license -// Copyright (C) 2024 Intel Corporation +// Copyright (C) 2025 Intel Corporation // SPDX-License-Identifier: MIT // @@ -16,12 +16,19 @@ #include "common.hpp" template -using to_t_sycl_t = void (*)(const void *__restrict__ x, T *__restrict__ y, - int64_t k, dpct::queue_ptr stream); -typedef to_t_sycl_t to_fp32_sycl_t; +using to_t_sycl_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int64_t k, dpct::queue_ptr stream); +typedef to_t_sycl_t to_fp32_sycl_t; typedef to_t_sycl_t to_fp16_sycl_t; -to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor *dst); -to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst); +to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst); +to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor * dst); -#endif // GGML_SYCL_CONVERT_HPP +// Nc = Non-contiguous +template +using to_t_nc_sycl_t = void (*)(const void * x, T * y, int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne03, + int64_t s01, int64_t s02, int64_t s03, dpct::queue_ptr queue); + +typedef to_t_nc_sycl_t to_fp16_nc_sycl_t; +to_fp16_nc_sycl_t get_to_fp16_nc_sycl(ggml_type type); + +#endif // GGML_SYCL_CONVERT_HPP diff --git a/ggml/src/ggml-sycl/element_wise.cpp b/ggml/src/ggml-sycl/element_wise.cpp index fc25d98ddf..dcc6ec809a 100644 --- a/ggml/src/ggml-sycl/element_wise.cpp +++ b/ggml/src/ggml-sycl/element_wise.cpp @@ -21,6 +21,27 @@ static void acc_f32(const float * x, const float * y, float * dst, const int ne, } } +template +static void sgn(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { + for(auto i = item_ct1.get_global_id(2); i < (const size_t)k; i += item_ct1.get_global_range(2)) { + dst[i] = x[i] > static_cast(0.f) ? static_cast(1.f) : ((x[i] < static_cast(0.f) ? static_cast(-1.f) : static_cast(0.f))); + } +} + +template +static void abs_op(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { + for(auto i = item_ct1.get_global_id(2); i < (const size_t)k; i += item_ct1.get_global_range(2)) { + dst[i] = sycl::fabs(x[i]); + } +} + +template +static void elu_op(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { + for(auto i = item_ct1.get_global_id(2); i < (const size_t)k; i += item_ct1.get_global_range(2)) { + dst[i] = (x[i] > static_cast(0.f)) ? x[i] : sycl::expm1(x[i]); + } +} + template static void gelu(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { @@ -335,6 +356,37 @@ static void silu_sycl(const T *x, T *dst, const int k, }); } +template +static void sgn_sycl(const T * x, T * dst, const int k, queue_ptr stream) { + // hard code for now + const int num_blocks = ceil_div(k, 256); + stream->parallel_for( + sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range(1, 1, 256)), sycl::range(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) { + sgn(x, dst, k, item_ct1); + }); +} + +template +static void abs_sycl(const T * x, T * dst, const int k, queue_ptr stream) { + // hard code for now + const int num_blocks = ceil_div(k, 256); + stream->parallel_for( + sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) { + abs_op(x, dst, k, item_ct1); + }); +} + + +template +static void elu_sycl(const T * x, T * dst, const int k, queue_ptr stream) { + // hard code for now + const int num_blocks = ceil_div(k, 256); + stream->parallel_for( + sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) { + elu_op(x, dst, k, item_ct1); + }); +} + template static void gelu_quick_sycl(const T *x, T *dst, const int k, queue_ptr stream) { @@ -574,6 +626,106 @@ static void clamp_sycl(const T *x, T *dst, const float min, }); } +inline void ggml_sycl_op_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); + +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + sgn_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + sgn_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } +} + +inline void ggml_sycl_op_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); + +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + abs_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + abs_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } +} + + +inline void ggml_sycl_op_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); + +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + elu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + elu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } +} + inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { #if defined (GGML_SYCL_F16) GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); @@ -1388,3 +1540,20 @@ void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s done\n", __func__); } +void ggml_sycl_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_sgn(ctx, dst); + GGML_SYCL_DEBUG("call %s done\n", __func__); +} + +void ggml_sycl_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_abs(ctx, dst); + GGML_SYCL_DEBUG("call %s done\n", __func__); +} + +void ggml_sycl_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_elu(ctx, dst); + GGML_SYCL_DEBUG("call %s done\n", __func__); +} diff --git a/ggml/src/ggml-sycl/element_wise.hpp b/ggml/src/ggml-sycl/element_wise.hpp index e623cb56f7..f4199d69da 100644 --- a/ggml/src/ggml-sycl/element_wise.hpp +++ b/ggml/src/ggml-sycl/element_wise.hpp @@ -66,5 +66,10 @@ void ggml_sycl_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst); void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst); +void ggml_sycl_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst); + +void ggml_sycl_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst); + +void ggml_sycl_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst); #endif // GGML_SYCL_ELEMENTWISE_HPP diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp index 548f2d0a06..0ea729948e 100644 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp @@ -38,6 +38,7 @@ #include "ggml-sycl/backend.hpp" #include "ggml-sycl/common.hpp" +#include "ggml-sycl/element_wise.hpp" #include "ggml-sycl/presets.hpp" #include "ggml-sycl/gemm.hpp" #include "ggml-sycl/sycl_hw.hpp" @@ -48,6 +49,7 @@ static bool g_sycl_loaded = false; int g_ggml_sycl_debug = 0; int g_ggml_sycl_disable_optimize = 0; int g_ggml_sycl_disable_graph = 0; +int g_ggml_sycl_prioritize_dmmv = 0; static ggml_sycl_device_info ggml_sycl_init() { ggml_sycl_device_info info = {}; @@ -192,13 +194,15 @@ static void ggml_check_sycl() try { if (!initialized) { g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0); - g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 0); + g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 1); g_ggml_sycl_disable_graph = get_sycl_env("GGML_SYCL_DISABLE_GRAPH", 1); + g_ggml_sycl_prioritize_dmmv = get_sycl_env("GGML_SYCL_PRIORITIZE_DMMV", 0); GGML_SYCL_DEBUG("[SYCL] call ggml_check_sycl\n"); GGML_LOG_INFO("Running with Environment Variables:\n"); GGML_LOG_INFO(" GGML_SYCL_DEBUG: %d\n", g_ggml_sycl_debug); GGML_LOG_INFO(" GGML_SYCL_DISABLE_OPT: %d\n", g_ggml_sycl_disable_optimize); GGML_LOG_INFO(" GGML_SYCL_DISABLE_GRAPH: %d\n", g_ggml_sycl_disable_graph); + GGML_LOG_INFO(" GGML_SYCL_PRIORITIZE_DMMV: %d\n", g_ggml_sycl_prioritize_dmmv); GGML_LOG_INFO("Build with Macros:\n"); #if defined(GGML_SYCL_FORCE_MMQ) GGML_LOG_INFO(" GGML_SYCL_FORCE_MMQ: yes\n"); @@ -337,7 +341,7 @@ ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer, assert(tensor->view_src->buffer->buft == buffer->buft); return GGML_STATUS_SUCCESS; } - if (tensor->type == GGML_TYPE_Q4_0) { + if (tensor->type == GGML_TYPE_Q4_0 && !g_ggml_sycl_disable_optimize) { ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{}; tensor->extra = extra; ctx->tensor_extras.push_back(extra); //used to release it when destroy ctx. @@ -2693,35 +2697,31 @@ catch (sycl::exception const &exc) { std::exit(1); } -static void k_compute_batched_ptrs(const sycl::half *src0_as_f16, - const sycl::half *src1_as_f16, char *dst, - const void **ptrs_src, void **ptrs_dst, - int64_t ne12, int64_t ne13, int64_t ne23, - size_t nb02, size_t nb03, size_t nb12, - size_t nb13, size_t nbd2, size_t nbd3, - int64_t r2, int64_t r3, - const sycl::nd_item<3> &item_ct1) { - int64_t i13 = item_ct1.get_group(2) * item_ct1.get_local_range(2) + - item_ct1.get_local_id(2); - int64_t i12 = item_ct1.get_group(1) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); +static void k_compute_batched_ptrs(const sycl::half * src0_as_f16, const sycl::half * src1_as_f16, char * dst, + const void ** ptrs_src, void ** ptrs_dst, int64_t ne12, int64_t ne13, int64_t ne23, + size_t nb02, size_t nb03, size_t nb12, size_t nb13, size_t nbd2, size_t nbd3, + int64_t r2, int64_t r3, const sycl::nd_item<3> & item_ct1) { + const int64_t i13 = item_ct1.get_group(2) * item_ct1.get_local_range(2) + item_ct1.get_local_id(2); + const int64_t i12 = item_ct1.get_group(1) * item_ct1.get_local_range(1) + item_ct1.get_local_id(1); if (i13 >= ne13 || i12 >= ne12) { return; } - int64_t i03 = i13 / r3; - int64_t i02 = i12 / r2; + const int64_t i03 = i13 / r3; + const int64_t i02 = i12 / r2; - ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03; - ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12 + i13*nb13; - ptrs_dst[0*ne23 + i12 + i13*ne12] = ( char *) dst + i12*nbd2 + i13*nbd3; + const uint8_t * src0_bytes = reinterpret_cast(src0_as_f16); + const uint8_t * src1_bytes = reinterpret_cast(src1_as_f16); + uint8_t * dst_bytes = reinterpret_cast(dst); + + ptrs_src[0 * ne23 + i12 + i13 * ne12] = src0_bytes + i02 * nb02 + i03 * nb03; + ptrs_src[1 * ne23 + i12 + i13 * ne12] = src1_bytes + i12 * nb12 + i13 * nb13; + ptrs_dst[0 * ne23 + i12 + i13 * ne12] = dst_bytes + i12 * nbd2 + i13 * nbd3; } -static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, - const ggml_tensor *src0, - const ggml_tensor *src1, - ggml_tensor *dst) try { +static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, + const ggml_tensor * src1, ggml_tensor * dst) try { GGML_ASSERT(!ggml_is_transposed(src0)); GGML_ASSERT(!ggml_is_transposed(src1)); GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src0->buffer)); @@ -2729,103 +2729,107 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, GGML_TENSOR_BINARY_OP_LOCALS + // TODO: see https://github.com/ggml-org/llama.cpp/pull/13155 + // Batched mul_mat requires a rewrite to support both oneDNN and non-contiguous dst + GGML_ASSERT(ggml_is_contiguous(dst)); SYCL_CHECK(ggml_sycl_set_device(ctx.device)); - queue_ptr main_stream = ctx.stream();; + queue_ptr queue = ctx.stream(); - void * src0_ddq = src0->data; - sycl::half *src0_as_f16 = (sycl::half *)src0_ddq; - float * src1_ddf = (float *) src1->data; - float * dst_ddf = (float *) dst->data; + dpct::has_capability_or_fail(queue->get_device(), { sycl::aspect::fp16 }); + + const sycl::half * src0_f16 = static_cast(src0->data); + float * dst_ddf = static_cast(dst->data); + + const sycl::half * src1_f16 = static_cast(src1->data); + const size_t type_size_src1 = ggml_type_size(src1->type); + GGML_ASSERT(nb10 == type_size_src1); + + // SRC1 strides + int64_t s11 = nb11 / type_size_src1; + int64_t s12 = nb12 / type_size_src1; + int64_t s13 = nb13 / type_size_src1; + ggml_sycl_pool_alloc src1_f16_alloc(ctx.pool()); // convert src1 to fp16 - ggml_sycl_pool_alloc src1_f16_alloc(ctx.pool()); if (src1->type != GGML_TYPE_F16) { - const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type, dst); + const to_fp16_nc_sycl_t to_fp16_nc_sycl = get_to_fp16_nc_sycl(src1->type); + GGML_ASSERT(to_fp16_nc_sycl != nullptr); const int64_t ne_src1 = ggml_nelements(src1); src1_f16_alloc.alloc(ne_src1); - GGML_ASSERT(to_fp16_sycl != nullptr); - to_fp16_sycl(src1_ddf, src1_f16_alloc.get(), ne_src1, main_stream); + to_fp16_nc_sycl(src1_f16, src1_f16_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, queue); + + src1_f16 = src1_f16_alloc.get(); + s11 = ne10; + s12 = ne11 * s11; + s13 = ne12 * s12; } - sycl::half *src1_f16 = src1->type == GGML_TYPE_F16 ? (sycl::half *)src1_ddf - : src1_f16_alloc.get(); - char * dst_t; + ggml_sycl_pool_alloc dst_f16(ctx.pool()); + char * dst_t = reinterpret_cast(dst_ddf); - dpct::library_data_t cu_compute_type = dpct::library_data_t::real_float; - dpct::library_data_t cu_data_type = dpct::library_data_t::real_float; + dpct::library_data_t mkl_compute_type = dpct::library_data_t::real_float; + dpct::library_data_t mkl_data_type = dpct::library_data_t::real_float; // dst strides size_t nbd2 = dst->nb[2]; size_t nbd3 = dst->nb[3]; const float alpha_f32 = 1.0f; - const float beta_f32 = 0.0f; + const float beta_f32 = 0.0f; const void * alpha = &alpha_f32; const void * beta = &beta_f32; - dst_t = (char *) dst_ddf; - GGML_ASSERT(ne12 % ne02 == 0); GGML_ASSERT(ne13 % ne03 == 0); // broadcast factors - const int64_t r2 = ne12/ne02; - const int64_t r3 = ne13/ne03; + const int64_t r2 = ne12 / ne02; + const int64_t r3 = ne13 / ne03; if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) { // there is no broadcast and src0, src1 are contiguous across dims 2, 3 - SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch( - *main_stream, oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha, - (const char *) src0_as_f16, dpct::library_data_t::real_half, nb01 / nb00, nb02 / nb00, - (const char *) src1_f16, dpct::library_data_t::real_half, nb11 / nb10, nb12 / nb10, beta, (char *) dst_t, - cu_data_type, ne01, nb2 / nb0, ne12 * ne13, cu_compute_type))); + SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(*queue, oneapi::math::transpose::trans, + oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha, + src0_f16, dpct::library_data_t::real_half, nb01 / nb00, nb02 / nb00, + src1_f16, dpct::library_data_t::real_half, s11, s12, beta, dst_t, + mkl_data_type, ne0, ne1 * ne0, ne12 * ne13, mkl_compute_type))); } else { - const int ne23 = ne12*ne13; + const int ne23 = ne12 * ne13; - ggml_sycl_pool_alloc ptrs_src(ctx.pool(), 2*ne23); - ggml_sycl_pool_alloc< void *> ptrs_dst(ctx.pool(), 1*ne23); + ggml_sycl_pool_alloc ptrs_src(ctx.pool(), 2 * ne23); + ggml_sycl_pool_alloc ptrs_dst(ctx.pool(), 1 * ne23); ggml_sycl_pool_alloc> matrix_info(ctx.host_pool(), 1); sycl::range<3> block_dims(1, ne12, ne13); - /* - DPCT1049:47: The work-group size passed to the SYCL kernel may exceed - the limit. To get the device limit, query - info::device::max_work_group_size. Adjust the work-group size if needed. - */ - { - dpct::has_capability_or_fail(main_stream->get_device(), - {sycl::aspect::fp16}); - - main_stream->submit([&](sycl::handler &cgh) { - const void **ptrs_src_get = ptrs_src.get(); - void **ptrs_dst_get = ptrs_dst.get(); - size_t nb12_scaled = src1->type == GGML_TYPE_F16 ? nb12 : nb12 / 2; - size_t nb13_scaled = src1->type == GGML_TYPE_F16 ? nb13 : nb13 / 2; - cgh.parallel_for(sycl::nd_range<3>(block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - k_compute_batched_ptrs( - src0_as_f16, src1_f16, - dst_t, ptrs_src_get, - ptrs_dst_get, ne12, ne13, ne23, - nb02, nb03, nb12_scaled, nb13_scaled, - nbd2, nbd3, r2, r3, item_ct1); - }); + queue->submit([&](sycl::handler & cgh) { + const void ** ptrs_src_get = ptrs_src.get(); + void ** ptrs_dst_get = ptrs_dst.get(); + size_t nb12_scaled = src1->type == GGML_TYPE_F16 ? nb12 : s12 * sizeof(sycl::half); + size_t nb13_scaled = src1->type == GGML_TYPE_F16 ? nb13 : s13 * sizeof(sycl::half); + cgh.parallel_for(sycl::nd_range<3>(block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) { + k_compute_batched_ptrs(src0_f16, src1_f16, dst_t, ptrs_src_get, ptrs_dst_get, ne12, ne13, ne23, nb02, + nb03, nb12_scaled, nb13_scaled, nbd2, nbd3, r2, r3, item_ct1); }); - } + }); + SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch( - *main_stream, oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha, + *queue, oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha, (const void **) (ptrs_src.get() + 0 * ne23), dpct::library_data_t::real_half, nb01 / nb00, - (const void **) (ptrs_src.get() + 1 * ne23), dpct::library_data_t::real_half, nb11 / nb10, beta, - (void **) (ptrs_dst.get() + 0 * ne23), cu_data_type, ne01, ne23, cu_compute_type, matrix_info.get()))); + (const void **) (ptrs_src.get() + 1 * ne23), dpct::library_data_t::real_half, s11, beta, + (void **) (ptrs_dst.get() + 0 * ne23), mkl_data_type, ne0, ne23, mkl_compute_type, matrix_info.get()))); } +} catch (const sycl::exception & exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ << ", line:" << __LINE__ << std::endl; + std::exit(1); } -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); -} + +enum class mul_mat_algo { + DMMV = 0, + MMVQ = 1, + MUL_MAT_SYCL = 2, +}; inline bool ggml_sycl_supports_mmq(enum ggml_type type) { // TODO: accuracy issues in MMQ @@ -2833,6 +2837,33 @@ inline bool ggml_sycl_supports_mmq(enum ggml_type type) { return false; } +inline bool ggml_sycl_supports_reorder_mul_mat_sycl(enum ggml_type type) { + switch (type) { + case GGML_TYPE_Q4_0: + return true; + default: + return false; + } +} + +inline bool ggml_sycl_supports_reorder_dmmv(enum ggml_type type) { + switch (type) { + case GGML_TYPE_Q4_0: + return true; + default: + return false; + } +} + +inline bool ggml_sycl_supports_reorder_mmvq(enum ggml_type type) { + switch (type) { + case GGML_TYPE_Q4_0: + return true; + default: + return false; + } +} + static bool ggml_sycl_supports_dmmv(enum ggml_type type) { switch (type) { case GGML_TYPE_Q4_0: @@ -2861,7 +2892,7 @@ static void reorder_qw(char *data_device, const int ncols, const int nrows, GGML_ASSERT((size % sizeof(block_q4_0) == 0)); GGML_ASSERT((offset % sizeof(block_q4_0) == 0)); int offset_blks = offset / sizeof(block_q4_0); - auto qs_ptr = (uint8_t*)data_device + offset_blks * QK4_0 / 2;; + auto qs_ptr = (uint8_t*)data_device + offset_blks * QK4_0 / 2; auto d_ptr = (sycl::half*)(qs_ptr + ncols * nrows / 2) + offset_blks; stream->parallel_for( @@ -2889,25 +2920,44 @@ static void reorder_qw(const ggml_tensor * src0, dpct::queue_ptr stream) { reorder_qw(data_device, ncols, nrows, size, 0, stream); } -/* -* This function could be called when the OP (mul_mat) function support reorder optimizition. -*/ -static void opt_for_reorder(ggml_backend_sycl_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, - ggml_tensor * dst) { - if (!g_ggml_sycl_disable_optimize && //allow optimize, controlled by $GGML_SYCL_DISABLE_OPT - ctx->opt_feature.reorder && //allow this device due to good perf, skip the devices with bad perf. - dst->op == GGML_OP_MUL_MAT && //limit to some supported cases of Q4_0, to do for more cases. - src0->type == GGML_TYPE_Q4_0 && - src1->ne[2]==1 && src1->ne[3]==1) { +static bool should_reorder_tensor(ggml_backend_sycl_context& ctx, const ggml_tensor * dst) { + return !g_ggml_sycl_disable_optimize && //allow optimize, controlled by $GGML_SYCL_DISABLE_OPT + ctx.opt_feature.reorder && //allow this device due to good perf, skip the devices with bad perf. + dst->op == GGML_OP_MUL_MAT && //limit to some supported cases of Q4_0, to do for more cases. + dst->src[1]->ne[2]==1 && dst->src[1]->ne[3]==1; +} - ggml_tensor_extra_gpu* extra = (ggml_tensor_extra_gpu*)src0->extra; - if (!extra) return; //only happen in CI/UT permute case. - - if (extra->optimized_feature.reorder) return; //skip the tensor which is handled for reorder. - - reorder_qw(src0, ctx->stream()); - extra->optimized_feature.reorder = true; //used to decode/dequan in next steps. +static void opt_for_reorder(ggml_backend_sycl_context * ctx, const ggml_tensor * src0, const ggml_tensor * /* src1 */, + ggml_tensor * dst, mul_mat_algo mm_algorithm) { + if (!should_reorder_tensor(*ctx, dst)) { + return; } + + ggml_tensor_extra_gpu * extra = static_cast(src0->extra); + if (!extra || extra->optimized_feature.reorder) { + return; // Skip permutations and already reordered tensors + } + + switch (mm_algorithm) { + case mul_mat_algo::DMMV: + if (!ggml_sycl_supports_reorder_dmmv(src0->type)) { + return; + } + break; + case mul_mat_algo::MMVQ: + if (!ggml_sycl_supports_reorder_mmvq(src0->type)) { + return; + } + break; + case mul_mat_algo::MUL_MAT_SYCL: + if (!ggml_sycl_supports_reorder_mul_mat_sycl(src0->type)) { + return; + } + break; + } + + reorder_qw(src0, ctx->stream()); + extra->optimized_feature.reorder = true; // Used to decode/dequan in next steps and avoid re-reordering } static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { @@ -2916,7 +2966,8 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor int64_t min_compute_capability = INT_MAX; if (split) { - ggml_backend_sycl_split_buffer_type_context * buft_ctx = (ggml_backend_sycl_split_buffer_type_context *) src0->buffer->buft->context; + ggml_backend_sycl_split_buffer_type_context * buft_ctx = + (ggml_backend_sycl_split_buffer_type_context *) src0->buffer->buft->context; auto & tensor_split = buft_ctx->tensor_split; for (int id = 0; id < ggml_sycl_info().device_count; ++id) { // skip devices that are not going to do any work: @@ -2929,7 +2980,7 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor } } } else { - min_compute_capability = ggml_sycl_info().devices[ctx.device].cc; + min_compute_capability = ggml_sycl_info().devices[ctx.device].cc; } // check data types and tensor shapes for custom matrix multiplication kernels: @@ -2951,9 +3002,15 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor use_mul_mat_q = use_mul_mat_q && (src1->ne[1] <= MMQ_MAX_BATCH_SIZE); #endif // SYCL_USE_XMX + // mmvq path is faster in the CUDA backend. - if (ctx.stream()->get_backend() == sycl::backend::ext_oneapi_cuda) + if (!g_ggml_sycl_prioritize_dmmv && (ctx.stream()->get_backend() == sycl::backend::ext_oneapi_cuda + // Dispatch becomes obscure with the reorder, MMVQ when the reorder optimization + // is enabled takes precedence over DMMV, the current if-else implementation + // requires disabling DMMV if both conditions are met + || (should_reorder_tensor(ctx, dst) && ggml_sycl_supports_reorder_mmvq(src0->type)))) { use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q; + } if (!split && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) { // TODO: Refactor and cleanup of mul mat dispatching. @@ -2965,24 +3022,30 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor // The kernel from the if path is faster for that specific case, but does not support all mul mats. ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst); } - } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) { + } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && !ggml_is_transposed(src1) && src1->ne[1] == 1) { // KQV single-batch ggml_sycl_mul_mat_vec_nc(ctx, src0, src1, dst); } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) { // KQ + KQV multi-batch ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst); } else if (use_dequantize_mul_mat_vec) { - opt_for_reorder(&ctx, src0, src1, dst); //the OP function in this branch support reorder. - ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec, false); - // save_tensor_txt("1/dst_1.txt", (float*) dst->data, src0->ne[1], sizeof(float), ctx.stream()); + constexpr bool convert_src1_to_q8_1 = false; + opt_for_reorder(&ctx, src0, src1, dst, mul_mat_algo::DMMV); + ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec, convert_src1_to_q8_1); } else if (use_mul_mat_vec_q) { - ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_vec_q, true); + constexpr bool convert_src1_to_q8_1 = true; + opt_for_reorder(&ctx, src0, src1, dst, mul_mat_algo::MMVQ); + ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_vec_q, convert_src1_to_q8_1); } else if (use_mul_mat_q) { - ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_q, true); + constexpr bool convert_src1_to_q8_1 = true; + ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_q, convert_src1_to_q8_1); } else { - opt_for_reorder(&ctx, src0, src1, dst); //the OP function in this branch support reorder. - ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_sycl, false); + constexpr bool convert_src1_to_q8_1 = false; + // MUL_MAT_SYCL supports reorder + opt_for_reorder(&ctx, src0, src1, dst, mul_mat_algo::MUL_MAT_SYCL); + ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_sycl, convert_src1_to_q8_1); } + GGML_SYCL_DEBUG("call %s done\n", __func__); } @@ -3355,6 +3418,15 @@ 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_SGN: + ggml_sycl_sgn(ctx, dst); + break; + case GGML_UNARY_OP_ABS: + ggml_sycl_abs(ctx, dst); + break; + case GGML_UNARY_OP_ELU: + ggml_sycl_elu(ctx, dst); + break; default: return false; } @@ -3837,6 +3909,9 @@ 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_TANH: case GGML_UNARY_OP_EXP: + case GGML_UNARY_OP_SGN: + case GGML_UNARY_OP_ABS: + case GGML_UNARY_OP_ELU: #if defined (GGML_SYCL_F16) return ggml_is_contiguous(op->src[0]) && (op->type == op->src[0]->type); #else diff --git a/ggml/src/ggml-sycl/mmvq.cpp b/ggml/src/ggml-sycl/mmvq.cpp index 1b92ba2d60..3cade1a42a 100644 --- a/ggml/src/ggml-sycl/mmvq.cpp +++ b/ggml/src/ggml-sycl/mmvq.cpp @@ -1,6 +1,60 @@ #include "mmvq.hpp" + +#include "ggml.h" +#include "common.hpp" +#include "quants.hpp" #include "vecdotq.hpp" -#include + +template +static void mul_mat_vec_q_reorder(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols, const int nrows, const sycl::nd_item<3> & nd_item) { + using block_type = ggml_sycl_reordered::block_q_t; + using block_traits = typename block_type::traits; + + const auto sg = nd_item.get_sub_group(); + const int sg_range = sg.get_group_linear_range(); + const int workgroup_id = nd_item.get_group_linear_id(); + const int sg_id = sg.get_group_linear_id(); + const int row = workgroup_id * sg_range + sg_id; + + if (row >= nrows) { + return; + } + + const int blocks_per_row = ncols / block_traits::qk; + constexpr int blocks_per_subgroup = ceil_div(block_traits::vdr_mmvq * WARP_SIZE, block_traits::qi); + constexpr int block_elements_per_subgroup = block_traits::qi / block_traits::vdr_mmvq; + + static_assert(blocks_per_subgroup > 0); + static_assert(block_elements_per_subgroup > 0); + + const block_q8_1 * y = (const block_q8_1 *) vy; + + float partial_sum = 0.0f; + for (int i = sg.get_local_linear_id() / block_elements_per_subgroup; i < blocks_per_row; i += blocks_per_subgroup) { + const int ibx = row * blocks_per_row + i; // x block index + // TODO: Generalize offsets, right now only works for quantizations that don't split high and low bits + const int bx_offset = block_type::get_block_offset(ibx); + const int d_offset = block_type::get_d_offset(nrows, ncols, ibx); + + // Y block index that aligns with ibx + const int iby = i * block_type::block_to_q8_1_ratio(); + +#pragma unroll + for (int elem = 0; elem < block_elements_per_subgroup; elem += WARP_SIZE) { + // x block quant index when casting the quants to int + const int iqs = elem + block_traits::vdr_mmvq * (sg.get_local_linear_id() % block_elements_per_subgroup); + + partial_sum += reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, &y[iby], iqs); + } + } + + auto sum = sycl::reduce_over_group(nd_item.get_sub_group(), partial_sum, std::plus<>()); + + if (sg.leader()) { + dst[row] = sum; + } +} template static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, @@ -480,26 +534,39 @@ static void mul_mat_vec_q_iq4_xs_q8_1(const void *__restrict__ vx, } } -static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy, - float *dst, const int ncols, - const int nrows, +static void reorder_mul_mat_vec_q4_0_q8_1_sycl(const void * vx, const void * vy, float * dst, const int ncols, + const int nrows, dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK4_0 == 0); + const int block_num_y = ceil_div(nrows, GGML_SYCL_MMV_Y); + constexpr size_t num_subgroups = 16; + GGML_ASSERT(block_num_y % num_subgroups == 0); + + const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, (block_num_y * WARP_SIZE)); + const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE); + + stream->submit([&](sycl::handler & cgh) { + cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size), + [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q_reorder>(vx, vy, dst, ncols, nrows, + nd_item); + }); + }); +} + +static void mul_mat_vec_q4_0_q8_1_sycl(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, dpct::queue_ptr stream) { GGML_ASSERT(ncols % QK4_0 == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { - - stream->submit([&](sycl::handler &cgh) { - - cgh.parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) - [[sycl::reqd_sub_group_size(WARP_SIZE)]] { - mul_mat_vec_q( - vx, vy, dst, ncols, nrows, item_ct1); - }); + stream->submit([&](sycl::handler & cgh) { + cgh.parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + mul_mat_vec_q( + vx, vy, dst, ncols, nrows, item_ct1); + }); }); } } @@ -916,93 +983,95 @@ static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy, } } -void ggml_sycl_op_mul_mat_vec_q( - ggml_backend_sycl_context & ctx, - const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, - const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i, - float *dst_dd_i, const int64_t row_low, const int64_t row_high, - const int64_t src1_ncols, const int64_t src1_padded_col_size, - const dpct::queue_ptr &stream) { - +void ggml_sycl_op_mul_mat_vec_q(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, + ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i, + const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, + const int64_t row_high, const int64_t src1_ncols, const int64_t src1_padded_col_size, + const dpct::queue_ptr & stream) { const int64_t ne10 = src1->ne[0]; GGML_ASSERT(ne10 % QK8_1 == 0); - const int64_t ne00 = src0->ne[0]; + const int64_t ne00 = src0->ne[0]; const int64_t row_diff = row_high - row_low; int id; - SYCL_CHECK( - CHECK_TRY_ERROR(id = get_current_device_id())); + SYCL_CHECK(CHECK_TRY_ERROR(id = get_current_device_id())); const size_t q8_1_ts = sizeof(block_q8_1); const size_t q8_1_bs = QK8_1; // the main device has a larger memory buffer to hold the results from all GPUs // nrows_dst == nrows of the matrix that the kernel writes into - for (int i = 0; i < src1_ncols; i++) - { + for (int i = 0; i < src1_ncols; i++) { const size_t src1_ddq_i_offset = i * src1_padded_col_size * q8_1_ts / q8_1_bs; - const char* src1_ddq_i_bs = src1_ddq_i + src1_ddq_i_offset; - float* dst_dd_i_bs = dst_dd_i + i * dst->ne[0]; + const char * src1_ddq_i_bs = src1_ddq_i + src1_ddq_i_offset; + float * dst_dd_i_bs = dst_dd_i + i * dst->ne[0]; switch (src0->type) { - case GGML_TYPE_Q4_0: - mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_Q4_1: - mul_mat_vec_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_Q5_0: - mul_mat_vec_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_Q5_1: - mul_mat_vec_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_Q8_0: - mul_mat_vec_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_Q2_K: - mul_mat_vec_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_Q3_K: - mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_Q4_K: - mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_Q5_K: - mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_Q6_K: - mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ1_S: - mul_mat_vec_iq1_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ1_M: - mul_mat_vec_iq1_m_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ2_XXS: - mul_mat_vec_iq2_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ2_XS: - mul_mat_vec_iq2_xs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ2_S: - mul_mat_vec_iq2_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ3_XXS: - mul_mat_vec_iq3_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ3_S: - mul_mat_vec_iq3_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ4_NL: - mul_mat_vec_iq4_nl_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - case GGML_TYPE_IQ4_XS: - mul_mat_vec_iq4_xs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); - break; - default: - GGML_ABORT("fatal error"); + case GGML_TYPE_Q4_0: + if ((ggml_tensor_extra_gpu *) dst->src[0]->extra && + ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) { + GGML_SYCL_DEBUG("Calling reorder_mul_mat_vec_q4_0_q8_1_sycl\n"); + reorder_mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + } else { + GGML_SYCL_DEBUG("Calling mul_mat_vec_q4_0_q8_1_sycl\n"); + mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + } + break; + case GGML_TYPE_Q4_1: + mul_mat_vec_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_0: + mul_mat_vec_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_1: + mul_mat_vec_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + case GGML_TYPE_Q8_0: + mul_mat_vec_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + case GGML_TYPE_Q2_K: + mul_mat_vec_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + case GGML_TYPE_Q3_K: + mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + case GGML_TYPE_Q4_K: + mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_K: + mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + case GGML_TYPE_Q6_K: + mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ1_S: + mul_mat_vec_iq1_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ1_M: + mul_mat_vec_iq1_m_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ2_XXS: + mul_mat_vec_iq2_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ2_XS: + mul_mat_vec_iq2_xs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ2_S: + mul_mat_vec_iq2_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ3_XXS: + mul_mat_vec_iq3_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ3_S: + mul_mat_vec_iq3_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ4_NL: + mul_mat_vec_iq4_nl_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + case GGML_TYPE_IQ4_XS: + mul_mat_vec_iq4_xs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream); + break; + default: + GGML_ABORT("fatal error"); } } GGML_UNUSED(src1); diff --git a/ggml/src/ggml-sycl/quants.hpp b/ggml/src/ggml-sycl/quants.hpp new file mode 100644 index 0000000000..a74e30526c --- /dev/null +++ b/ggml/src/ggml-sycl/quants.hpp @@ -0,0 +1,61 @@ +// +// MIT license +// Copyright (C) 2025 Codeplay Software Ltd. +// Copyright (C) 2025 Intel Corporation +// SPDX-License-Identifier: MIT +// + +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// + +#ifndef GGML_SYCL_QUANTS_HPP +#define GGML_SYCL_QUANTS_HPP + +#include "ggml-common.h" +#include "ggml.h" + +namespace ggml_sycl_reordered { + + +// The reordered block moves quants (qs) and scales(d) to two +// uniform regions of memory that is contiguous in the same tensor. +// What this means is that instead of having: +// [d0, qs0] [d1, qs1] [d2, qs2] ... [dN, qsN] +// We have: +// [qs0, qs1, qs2, ..., qsN] [d0, d1, d2, ..., dN] +// +// Notes: out-of-bounds qs will run into d values +// Aligment relies on the allocated size of qs + +template struct block_q_t; + + +// qk number of weights / quants in a block +// qr number of weights in a byte (described as 'before dequantization') +// for quantization types that has low and high bits split, qr is calculated with +// using the lower bits, e.g for Q6 quants QR6 is 2 +// qi number of 32 bit integers needed to represent all the quants from a block (`qs` field) +// See ggml-common.h to see how these are calculated +template <> struct block_q_t { + struct traits { + static constexpr uint32_t qk = QK4_0; + static constexpr uint32_t qi = QI4_0; + static constexpr uint32_t qr = QR4_0; + static constexpr uint32_t vdr_mmvq = 2; + }; + + static constexpr int get_block_offset(const int block_index) { return block_index * (traits::qk / traits::qr); } + + static constexpr int get_d_offset(int nrows, int ncols, const int block_index) { + return (ncols / traits::qr * nrows) + block_index * sizeof(ggml_half); + } + + static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; } +}; + +} // namespace ggml_sycl_reordered + +#endif // GGML_SYCL_QUANTS_HPP diff --git a/ggml/src/ggml-sycl/vecdotq.hpp b/ggml/src/ggml-sycl/vecdotq.hpp index c5942008ad..cbf664fcf2 100644 --- a/ggml/src/ggml-sycl/vecdotq.hpp +++ b/ggml/src/ggml-sycl/vecdotq.hpp @@ -1,6 +1,6 @@ // // MIT license -// Copyright (C) 2024 Intel Corporation +// Copyright (C) 2025 Intel Corporation // SPDX-License-Identifier: MIT // @@ -14,8 +14,11 @@ #define GGML_SYCL_VECDOTQ_HPP #include "dpct/helper.hpp" +#include "ggml.h" +#include "quants.hpp" -typedef float (*vec_dot_q_sycl_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs); +typedef float (*vec_dot_q_sycl_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, + const int & iqs); static __dpct_inline__ int get_int_from_int8(const int8_t* x8, const int& i32) { const uint16_t* x16 = @@ -252,13 +255,60 @@ vec_dot_q6_K_q8_1_impl_mmvq(const int &vl, const int &vh, // VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called // MMVQ = mul_mat_vec_q, MMQ = mul_mat_q +template struct reorder_vec_dot_q_sycl { + static_assert(T != T, "ggml_type for reorder vecdot not implemented"); +}; + +template <> struct reorder_vec_dot_q_sycl { + static constexpr ggml_type gtype = GGML_TYPE_Q4_0; + + using q4_0_block = ggml_sycl_reordered::block_q_t; + using q4_0_traits = typename q4_0_block::traits; + + __dpct_inline__ float vec_dot_q4_0_q8_1_impl(const int * v, const int * u, const float & d4, const sycl::half2 & ds8) { + int sumi = 0; + +#pragma unroll + for (size_t i = 0; i < q4_0_traits::vdr_mmvq; ++i) { + const int vi0 = (v[i] >> 0) & 0x0F0F0F0F; + const int vi1 = (v[i] >> 4) & 0x0F0F0F0F; + + // SIMD dot product of quantized values + sumi = dpct::dp4a(vi0, u[2 * i + 0], sumi); + sumi = dpct::dp4a(vi1, u[2 * i + 1], sumi); + } + + const sycl::float2 ds8f = ds8.convert(); + + // second part effectively subtracts 8 from each quant value + return d4 * (sumi * ds8f.x() - (8 * q4_0_traits::vdr_mmvq / q4_0_traits::qi) * ds8f.y()); + } + + __dpct_inline__ float operator()(const void * __restrict__ vbq, const int ibx_offset, const int d_offset, + const block_q8_1 * __restrict__ bq8_1, const int & iqs) { + const uint8_t * bq4_0 = static_cast(vbq) + ibx_offset; + const ggml_half d = *(reinterpret_cast(static_cast(vbq) + d_offset)); + int v[q4_0_traits::vdr_mmvq]; + int u[2 * q4_0_traits::vdr_mmvq]; + +#pragma unroll + + for (size_t i = 0; i < q4_0_traits::vdr_mmvq; ++i) { + v[i] = get_int_from_uint8(bq4_0, iqs + i); + u[2 * i + 0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + u[2 * i + 1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + q4_0_traits::qi); + } + + return vec_dot_q4_0_q8_1_impl(v, u, d, bq8_1->ds); + }; +}; + #define VDR_Q4_0_Q8_1_MMVQ 2 #define VDR_Q4_0_Q8_1_MMQ 4 template -static __dpct_inline__ float vec_dot_q4_0_q8_1_impl(const int *v, const int *u, - const float &d4, - const sycl::half2 &ds8) { +static __dpct_inline__ float vec_dot_q4_0_q8_1_impl(const int * v, const int * u, const float & d4, + const sycl::half2 & ds8) { int sumi = 0; #pragma unroll for (int i = 0; i < vdr; ++i) { @@ -270,8 +320,7 @@ static __dpct_inline__ float vec_dot_q4_0_q8_1_impl(const int *v, const int *u, sumi = dpct::dp4a(vi1, u[2 * i + 1], sumi); } - const sycl::float2 ds8f = - ds8.convert(); + const sycl::float2 ds8f = ds8.convert(); // second part effectively subtracts 8 from each quant value return d4 * (sumi * ds8f.x() - (8 * vdr / QI4_0) * ds8f.y()); @@ -456,13 +505,13 @@ vec_dot_q4_0_q8_1(const void *__restrict__ vbq, const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq; int v[VDR_Q4_0_Q8_1_MMVQ]; - int u[2*VDR_Q4_0_Q8_1_MMVQ]; + int u[2 * VDR_Q4_0_Q8_1_MMVQ]; #pragma unroll for (int i = 0; i < VDR_Q4_0_Q8_1_MMVQ; ++i) { - v[i] = get_int_from_uint8(bq4_0->qs, iqs + i); - u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); - u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_0); + v[i] = get_int_from_uint8(bq4_0->qs, iqs + i); + u[2 * i + 0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + u[2 * i + 1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_0); } return vec_dot_q4_0_q8_1_impl(v, u, bq4_0->d, bq8_1->ds); diff --git a/ggml/src/ggml-vulkan/CMakeLists.txt b/ggml/src/ggml-vulkan/CMakeLists.txt index 9d028f718d..31816219c0 100644 --- a/ggml/src/ggml-vulkan/CMakeLists.txt +++ b/ggml/src/ggml-vulkan/CMakeLists.txt @@ -71,6 +71,22 @@ if (Vulkan_FOUND) add_compile_definitions(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) endif() + # Compile a test shader to determine whether GL_EXT_bfloat16 is supported. + # If it's not, there will be an error to stderr. + # If it's supported, set a define to indicate that we should compile those shaders + execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_bfloat16_support.comp" + OUTPUT_VARIABLE glslc_output + ERROR_VARIABLE glslc_error) + + if (${glslc_error} MATCHES ".*extension not supported: GL_EXT_bfloat16.*") + message(STATUS "GL_EXT_bfloat16 not supported by glslc") + set(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT OFF) + else() + message(STATUS "GL_EXT_bfloat16 supported by glslc") + set(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT ON) + add_compile_definitions(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + endif() + target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan) target_include_directories(ggml-vulkan PRIVATE ${CMAKE_CURRENT_BINARY_DIR}) @@ -142,6 +158,7 @@ if (Vulkan_FOUND) -DGGML_VULKAN_COOPMAT_GLSLC_SUPPORT=${GGML_VULKAN_COOPMAT_GLSLC_SUPPORT} -DGGML_VULKAN_COOPMAT2_GLSLC_SUPPORT=${GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT} -DGGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT=${GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT} + -DGGML_VULKAN_BFLOAT16_GLSLC_SUPPORT=${GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT} BUILD_COMMAND ${CMAKE_COMMAND} --build . INSTALL_COMMAND ${CMAKE_COMMAND} --install . INSTALL_DIR ${CMAKE_BINARY_DIR} diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index c0bdb9e17a..e2b357fdc1 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -51,6 +51,24 @@ #include "ggml-vulkan-shaders.hpp" +// remove this once it's more widely available in the SDK +#if !defined(VK_KHR_shader_bfloat16) + +#define VK_KHR_shader_bfloat16 1 +#define VK_KHR_SHADER_BFLOAT16_SPEC_VERSION 1 +#define VK_KHR_SHADER_BFLOAT16_EXTENSION_NAME "VK_KHR_shader_bfloat16" +#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_BFLOAT16_FEATURES_KHR ((VkStructureType)1000141000) +#define VK_COMPONENT_TYPE_BFLOAT16_KHR ((VkComponentTypeKHR)1000141000) + +typedef struct VkPhysicalDeviceShaderBfloat16FeaturesKHR { + VkStructureType sType; + void* pNext; + VkBool32 shaderBFloat16Type; + VkBool32 shaderBFloat16DotProduct; + VkBool32 shaderBFloat16CooperativeMatrix; +} VkPhysicalDeviceShaderBfloat16FeaturesKHR; +#endif + #define ROUNDUP_POW2(M, N) (((M) + (N) - 1) & ~((N) - 1)) #define CEIL_DIV(M, N) (((M) + (N)-1) / (N)) static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; } @@ -257,6 +275,7 @@ struct vk_device_struct { bool prefer_host_memory; bool float_controls_rte_fp16; bool subgroup_add; + bool subgroup_shuffle; bool integer_dot_product; @@ -266,8 +285,9 @@ struct vk_device_struct { bool subgroup_require_full_support; bool coopmat_support; - bool coopmat_acc_f32_support; - bool coopmat_acc_f16_support; + bool coopmat_acc_f32_support {}; + bool coopmat_acc_f16_support {}; + bool coopmat_bf16_support {}; uint32_t coopmat_m; uint32_t coopmat_n; uint32_t coopmat_k; @@ -293,6 +313,7 @@ struct vk_device_struct { vk_matmul_pipeline pipeline_matmul_f32 {}; vk_matmul_pipeline pipeline_matmul_f32_f16 {}; + vk_matmul_pipeline pipeline_matmul_bf16 {}; vk_matmul_pipeline2 pipeline_matmul_f16; vk_matmul_pipeline2 pipeline_matmul_f16_f32; @@ -301,6 +322,7 @@ struct vk_device_struct { vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_COUNT]; vk_matmul_pipeline pipeline_matmul_id_f32 {}; + vk_matmul_pipeline pipeline_matmul_id_bf16 {}; vk_matmul_pipeline2 pipeline_matmul_id_f16; vk_matmul_pipeline2 pipeline_matmul_id_f16_f32; @@ -319,11 +341,17 @@ struct vk_device_struct { vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT]; vk_pipeline pipeline_get_rows_f32[GGML_TYPE_COUNT]; vk_pipeline pipeline_acc_f32; - vk_pipeline pipeline_add_f32, pipeline_add_f32_norepeat; - vk_pipeline pipeline_add_f16_f32_f16, pipeline_add_f16_f32_f16_norepeat; - vk_pipeline pipeline_sub_f32, pipeline_sub_f32_norepeat; - vk_pipeline pipeline_mul_f32, pipeline_mul_f32_norepeat; - vk_pipeline pipeline_div_f32, pipeline_div_f32_norepeat; + + // [src0 0=fp32,1=fp16][src1 0=fp32,1=fp16][dst 0=fp32,1=fp16] + vk_pipeline pipeline_add[2][2][2]; + vk_pipeline pipeline_add_norepeat[2][2][2]; + vk_pipeline pipeline_sub[2][2][2]; + vk_pipeline pipeline_sub_norepeat[2][2][2]; + vk_pipeline pipeline_mul[2][2][2]; + vk_pipeline pipeline_mul_norepeat[2][2][2]; + vk_pipeline pipeline_div[2][2][2]; + vk_pipeline pipeline_div_norepeat[2][2][2]; + vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32; vk_pipeline pipeline_upscale_f32; vk_pipeline pipeline_scale_f32; @@ -333,8 +361,8 @@ struct vk_device_struct { vk_pipeline pipeline_clamp_f32; vk_pipeline pipeline_pad_f32; vk_pipeline pipeline_repeat_f32, pipeline_repeat_back_f32; - vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16; - vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16; + vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16, pipeline_cpy_f16_f32, pipeline_cpy_f32_bf16; + vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16, pipeline_contig_cpy_f16_f32, pipeline_contig_cpy_f32_bf16; vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT]; vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT]; vk_pipeline pipeline_norm_f32; @@ -342,14 +370,17 @@ struct vk_device_struct { vk_pipeline pipeline_rms_norm_f32; vk_pipeline pipeline_rms_norm_back_f32; vk_pipeline pipeline_l2_norm_f32; - vk_pipeline pipeline_gelu_f32; - vk_pipeline pipeline_gelu_quick_f32; - vk_pipeline pipeline_silu_f32; - vk_pipeline pipeline_silu_back_f32; - vk_pipeline pipeline_relu_f32; + + // [src/dst 0=fp32,1=fp16] + vk_pipeline pipeline_gelu[2]; + vk_pipeline pipeline_gelu_quick[2]; + vk_pipeline pipeline_silu[2]; + vk_pipeline pipeline_relu[2]; + vk_pipeline pipeline_tanh[2]; + vk_pipeline pipeline_sigmoid[2]; + vk_pipeline pipeline_leaky_relu_f32; - vk_pipeline pipeline_tanh_f32; - vk_pipeline pipeline_sigmoid_f32; + vk_pipeline pipeline_silu_back_f32; vk_pipeline pipeline_diag_mask_inf_f32; vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16; vk_pipeline pipeline_soft_max_f32_wg512, pipeline_soft_max_f32_f16_wg512; @@ -368,14 +399,24 @@ struct vk_device_struct { vk_pipeline pipeline_rwkv_wkv6_f32; vk_pipeline pipeline_rwkv_wkv7_f32; vk_pipeline pipeline_opt_step_adamw_f32; + vk_pipeline pipeline_conv2d_dw_whcn_f32; + vk_pipeline pipeline_conv2d_dw_cwhn_f32; // [2][2][2] is for {f16acc,f32acc}x{large,small_rows}x{unaligned, aligned} + vk_pipeline pipeline_flash_attn_f32_f16_D64_cm2[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D80_cm2[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D96_cm2[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D112_cm2[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D128_cm2[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D256_cm2[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_f32_f16_D64[GGML_TYPE_COUNT][2][2][2]; vk_pipeline pipeline_flash_attn_f32_f16_D80[GGML_TYPE_COUNT][2][2][2]; vk_pipeline pipeline_flash_attn_f32_f16_D96[GGML_TYPE_COUNT][2][2][2]; vk_pipeline pipeline_flash_attn_f32_f16_D112[GGML_TYPE_COUNT][2][2][2]; vk_pipeline pipeline_flash_attn_f32_f16_D128[GGML_TYPE_COUNT][2][2][2]; vk_pipeline pipeline_flash_attn_f32_f16_D256[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_split_k_reduce; std::unordered_map pipelines; @@ -680,6 +721,24 @@ struct vk_op_rwkv_wkv7_push_constants { uint32_t H; }; +struct vk_op_conv2d_dw_push_constants { + uint32_t ne; + uint32_t batches; + uint32_t channels; + uint32_t dst_w; + uint32_t dst_h; + uint32_t src_w; + uint32_t src_h; + uint32_t knl_w; + uint32_t knl_h; + int32_t stride_x; + int32_t stride_y; + int32_t pad_x; + int32_t pad_y; + int32_t dilation_x; + int32_t dilation_y; +}; + struct vk_op_upscale_push_constants { uint32_t ne; uint32_t a_offset; uint32_t d_offset; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03; @@ -1531,13 +1590,29 @@ static void ggml_vk_wait_events(vk_context& ctx, std::vector&& events // number of rows/cols for flash attention shader static constexpr uint32_t flash_attention_num_small_rows = 32; -static std::array fa_rows_cols(uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) { +static constexpr uint32_t scalar_flash_attention_num_small_rows = 1; +static constexpr uint32_t scalar_flash_attention_num_large_rows = 8; + +static uint32_t get_fa_num_small_rows(bool scalar) { + return scalar ? scalar_flash_attention_num_small_rows : flash_attention_num_small_rows; +} + +static std::array fa_rows_cols(bool scalar, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) { GGML_UNUSED(clamp); + if (scalar) { + if (small_rows) { + return {scalar_flash_attention_num_small_rows, 64}; + } else { + return {scalar_flash_attention_num_large_rows, 32}; + } + } + // small rows, large cols if (small_rows) { - return {flash_attention_num_small_rows, 64}; + return {get_fa_num_small_rows(scalar), 32}; } + // small cols to reduce register count if (ggml_is_quantized(type) || D == 256) { return {64, 32}; @@ -1582,7 +1657,7 @@ static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vec const uint32_t warps = warptile[0] / warptile[10]; const uint32_t load_bufs = (warptile[1] + warptile[2]) * (warptile[3] + bank_conflict_offset) * type_size; - const uint32_t mmid_row_ids = mul_mat_id ? 3072 * sizeof(uint32_t) : 0; + const uint32_t mmid_row_ids = mul_mat_id ? 4096 * sizeof(uint32_t) : 0; const uint32_t coopmat_stage = device->coopmat_support ? warptile[7] * warptile[8] / warps * sizeof(float) : 0; const uint32_t total_size = load_bufs + mmid_row_ids + coopmat_stage + lut_size; @@ -1791,6 +1866,12 @@ static void ggml_vk_load_shaders(vk_device& device) { if (!device->pipeline_matmul_id_f32) { device->pipeline_matmul_id_f32 = std::make_shared(); } + if (!device->pipeline_matmul_bf16) { + device->pipeline_matmul_bf16 = std::make_shared(); + } + if (!device->pipeline_matmul_id_bf16) { + device->pipeline_matmul_id_bf16 = std::make_shared(); + } std::vector> compiles; auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, @@ -1826,65 +1907,66 @@ static void ggml_vk_load_shaders(vk_device& device) { parameter_count, wg_denoms, specialization_constants, disable_robustness, require_full_subgroups, required_subgroup_size)); }; + auto const &fa_wg_denoms = [&](bool scalar, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::array { + return {fa_rows_cols(scalar, D, clamp, type, small_rows)[0], 1, 1}; + }; + + auto const &fa_spec_constants = [&](bool scalar, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::vector { + // For large number of rows, 128 invocations seems to work best. + // For small number of rows (e.g. N==1), 256 works better. But matrix granularity for 256 is 32, so we + // can't use 256 for D==80. + // For scalar, use 128 (arbitrary) + uint32_t wg_size = scalar ? 128 : ((small_rows && (D % 32) == 0) ? 256 : 128); + auto rows_cols = fa_rows_cols(scalar, D, clamp, type, small_rows); + + // 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); + + // mask dim1 is padded to 64, we rely on this to avoid clamping mask loads + GGML_ASSERT((GGML_KQ_MASK_PAD % rows_cols[0]) == 0); + return {wg_size, rows_cols[0], rows_cols[1], (D), clamp, D_split}; + }; + +#define CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, D) \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][0], "flash_attn_f32_f16_D" #D "_f16acc" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,false), fa_spec_constants(SCALAR, D,1,TYPE,false), 1, true); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,false), fa_spec_constants(SCALAR, D,0,TYPE,false), fa_rows_cols(SCALAR,D,0,TYPE,false)[1], true); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][0], "flash_attn_f32_f16_D" #D "_f32acc" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,false), fa_spec_constants(SCALAR, D,1,TYPE,false), 1, true); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,false), fa_spec_constants(SCALAR, D,0,TYPE,false), fa_rows_cols(SCALAR,D,0,TYPE,false)[1], true); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][0], "flash_attn_f32_f16_D" #D "_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,true), fa_spec_constants(SCALAR, D,1,TYPE,true), 1, true); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,true), fa_spec_constants(SCALAR, D,0,TYPE,true), fa_rows_cols(SCALAR,D,0,TYPE,true)[1], true); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][0], "flash_attn_f32_f16_D" #D "_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,true), fa_spec_constants(SCALAR, D,1,TYPE,true), 1, true); \ + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,true), fa_spec_constants(SCALAR, D,0,TYPE,true), fa_rows_cols(SCALAR,D,0,TYPE,true)[1], true); \ + +#define CREATE_FA(TYPE, NAMELC, SCALAR, SUFFIX) \ + CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 64) \ + CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 80) \ + CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 96) \ + CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 112) \ + CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 128) \ + CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 256) + + CREATE_FA(GGML_TYPE_F16, f16, true, ) + CREATE_FA(GGML_TYPE_Q4_0, q4_0, true, ) + CREATE_FA(GGML_TYPE_Q8_0, q8_0, true, ) #if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) if (device->coopmat2) { - - auto const &fa_wg_denoms = [&](uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::array { - return {fa_rows_cols(D, clamp, type, small_rows)[0], 1, 1}; - }; - - auto const &fa_spec_constants = [&](uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::vector { - // For large number of rows, 128 invocations seems to work best. - // For small number of rows (e.g. N==1), 256 works better. But matrix granularity for 256 is 32, so we - // can't use 256 for D==80. - uint32_t wg_size = (small_rows && (D % 32) == 0) ? 256 : 128; - auto rows_cols = fa_rows_cols(D, clamp, type, small_rows); - // mask dim1 is padded to 64, we rely on this to avoid clamping mask loads - GGML_ASSERT((GGML_KQ_MASK_PAD % rows_cols[0]) == 0); - return {wg_size, rows_cols[0], rows_cols[1], (D), clamp}; - }; - -#define CREATE_FA2(TYPE, NAMELC, D) \ - ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][0][0][0], "flash_attn_f32_f16_D" #D "_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_len, flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,1,TYPE,false), fa_spec_constants(D,1,TYPE,false), 1); \ - ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][0][0][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_len, flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,0,TYPE,false), fa_spec_constants(D,0,TYPE,false), fa_rows_cols(D,0,TYPE,false)[1]); \ - ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][1][0][0], "flash_attn_f32_f16_D" #D "_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _cm2_len, flash_attn_f32_f16_ ## NAMELC ## _cm2_data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,1,TYPE,false), fa_spec_constants(D,1,TYPE,false), 1); \ - ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][1][0][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _cm2_len, flash_attn_f32_f16_ ## NAMELC ## _cm2_data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,0,TYPE,false), fa_spec_constants(D,0,TYPE,false), fa_rows_cols(D,0,TYPE,false)[1]); \ - ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][0][1][0], "flash_attn_f32_f16_D" #D "_f16acc_smallrows" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_len, flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,1,TYPE,true), fa_spec_constants(D,1,TYPE,true), 1); \ - ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][0][1][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc_smallrows" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_len, flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,0,TYPE,true), fa_spec_constants(D,0,TYPE,true), fa_rows_cols(D,0,TYPE,true)[1]); \ - ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][1][1][0], "flash_attn_f32_f16_D" #D "_f32acc_smallrows" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _cm2_len, flash_attn_f32_f16_ ## NAMELC ## _cm2_data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,1,TYPE,true), fa_spec_constants(D,1,TYPE,true), 1); \ - ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][1][1][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc_smallrows" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _cm2_len, flash_attn_f32_f16_ ## NAMELC ## _cm2_data, "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,0,TYPE,true), fa_spec_constants(D,0,TYPE,true), fa_rows_cols(D,0,TYPE,true)[1]); \ - -#define CREATE_FA(TYPE, NAMELC) \ - CREATE_FA2(TYPE, NAMELC, 64) \ - CREATE_FA2(TYPE, NAMELC, 80) \ - CREATE_FA2(TYPE, NAMELC, 96) \ - CREATE_FA2(TYPE, NAMELC, 112) \ - CREATE_FA2(TYPE, NAMELC, 128) \ - CREATE_FA2(TYPE, NAMELC, 256) - - CREATE_FA(GGML_TYPE_F16, f16) - CREATE_FA(GGML_TYPE_Q4_0, q4_0) - CREATE_FA(GGML_TYPE_Q4_1, q4_1) - CREATE_FA(GGML_TYPE_Q5_0, q5_0) - CREATE_FA(GGML_TYPE_Q5_1, q5_1) - CREATE_FA(GGML_TYPE_Q8_0, q8_0) - // K dequants currently disabled because D dimension is rounded up to 256 and runs inefficiently - //CREATE_FA(GGML_TYPE_Q2_K, q2_k) - //CREATE_FA(GGML_TYPE_Q3_K, q3_k) - //CREATE_FA(GGML_TYPE_Q4_K, q4_k) - //CREATE_FA(GGML_TYPE_Q5_K, q5_k) - //CREATE_FA(GGML_TYPE_Q6_K, q6_k) - //CREATE_FA(GGML_TYPE_IQ1_S, iq1_s) - //CREATE_FA(GGML_TYPE_IQ1_M, iq1_m) - //CREATE_FA(GGML_TYPE_IQ2_XXS, iq2_xxs) - //CREATE_FA(GGML_TYPE_IQ2_XS, iq2_xs) - //CREATE_FA(GGML_TYPE_IQ2_S, iq2_s) - //CREATE_FA(GGML_TYPE_IQ3_XXS, iq3_xxs) - //CREATE_FA(GGML_TYPE_IQ3_S, iq3_s) - //CREATE_FA(GGML_TYPE_IQ4_XS, iq4_xs) - CREATE_FA(GGML_TYPE_IQ4_NL, iq4_nl) + CREATE_FA(GGML_TYPE_F16, f16, false, _cm2) + CREATE_FA(GGML_TYPE_Q4_0, q4_0, false, _cm2) + CREATE_FA(GGML_TYPE_Q4_1, q4_1, false, _cm2) + CREATE_FA(GGML_TYPE_Q5_0, q5_0, false, _cm2) + CREATE_FA(GGML_TYPE_Q5_1, q5_1, false, _cm2) + CREATE_FA(GGML_TYPE_Q8_0, q8_0, false, _cm2) + CREATE_FA(GGML_TYPE_IQ4_NL, iq4_nl, false, _cm2) + } +#endif +#undef CREATE_FA2 #undef CREATE_FA +#if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) + if (device->coopmat2) { + // Create 6 variants, {s,m,l}x{unaligned,aligned} #define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \ @@ -1900,6 +1982,11 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(PIPELINE_NAME . f32acc, NAMELC, , WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \ CREATE_MM2(pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3) +#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (device->coopmat_bf16_support) { + CREATE_MM(pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3) + } +#endif CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) @@ -1921,6 +2008,11 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, 4) +#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (device->coopmat_bf16_support) { + CREATE_MM(pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4) + } +#endif CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) @@ -1974,6 +2066,11 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_F32, pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); +#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (device->coopmat_bf16_support) { + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ) + } +#endif if (device->coopmat_acc_f16_support) { CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); @@ -2022,6 +2119,11 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); +#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (device->coopmat_bf16_support) { + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + } +#endif if (device->coopmat_acc_f16_support) { CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); @@ -2104,6 +2206,8 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); @@ -2139,6 +2243,8 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); @@ -2191,6 +2297,8 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_F16, pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_F16, pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); @@ -2226,6 +2334,8 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f32acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f32acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); @@ -2246,8 +2356,26 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc, matmul_id_iq3_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f32acc, matmul_id_iq4_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); -#undef CREATE_MM } + // reusing CREATE_MM from the fp32 path + if ((device->coopmat2 || device->coopmat_support) +#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) + && !device->coopmat_bf16_support +#endif + ) { + // use scalar tile sizes + l_warptile = { 128, 128, 128, 16, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 }; + m_warptile = { 128, 64, 64, 16, subgroup_size_8, 32, 2, 4, 2, 1, subgroup_size_8 }; + s_warptile = { subgroup_size_16, 32, 32, 16, 32, 32, 2, 2, 2, 1, subgroup_size_8 }; + + l_wg_denoms = {128, 128, 1 }; + m_wg_denoms = { 64, 64, 1 }; + s_wg_denoms = { 32, 32, 1 }; + + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id); + } +#undef CREATE_MM // mul mat vec @@ -2266,6 +2394,7 @@ static void ggml_vk_load_shaders(vk_device& device) { for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) { ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32_"+std::to_string(i+1), mul_mat_vec_f32_f32_f32_len, mul_mat_vec_f32_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32_"+std::to_string(i+1), mul_mat_vec_f16_f32_f32_len, mul_mat_vec_f16_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f32_f32_"+std::to_string(i+1), mul_mat_vec_bf16_f32_f32_len, mul_mat_vec_bf16_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); @@ -2288,6 +2417,7 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32_"+std::to_string(i+1), mul_mat_vec_f32_f16_f32_len, mul_mat_vec_f32_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32_"+std::to_string(i+1), mul_mat_vec_f16_f16_f32_len, mul_mat_vec_f16_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f16_f32_"+std::to_string(i+1), mul_mat_vec_bf16_f16_f32_len, mul_mat_vec_bf16_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true); @@ -2311,6 +2441,7 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32", mul_mat_vec_id_f32_f32_len, mul_mat_vec_id_f32_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32", mul_mat_vec_id_f16_f32_len, mul_mat_vec_id_f16_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_BF16], "mul_mat_vec_id_bf16_f32", mul_mat_vec_id_bf16_f32_len, mul_mat_vec_id_bf16_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true); @@ -2356,6 +2487,7 @@ static void ggml_vk_load_shaders(vk_device& device) { // get_rows ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_F32 ], "get_rows_f32", get_rows_f32_len, get_rows_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_F16 ], "get_rows_f16", get_rows_f16_len, get_rows_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_BF16], "get_rows_bf16", get_rows_bf16_len, get_rows_bf16_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q4_0], "get_rows_q4_0", get_rows_q4_0_len, get_rows_q4_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q4_1], "get_rows_q4_1", get_rows_q4_1_len, get_rows_q4_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q5_0], "get_rows_q5_0", get_rows_q5_0_len, get_rows_q5_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); @@ -2373,6 +2505,7 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F32 ], "get_rows_f32_f32", get_rows_f32_f32_len, get_rows_f32_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F16 ], "get_rows_f16_f32", get_rows_f16_f32_len, get_rows_f16_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_BF16], "get_rows_bf16_f32", get_rows_bf16_f32_len, get_rows_bf16_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q4_0], "get_rows_q4_0_f32", get_rows_q4_0_f32_len, get_rows_q4_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q4_1], "get_rows_q4_1_f32", get_rows_q4_1_f32_len, get_rows_q4_1_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q5_0], "get_rows_q5_0_f32", get_rows_q5_0_f32_len, get_rows_q5_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); @@ -2399,7 +2532,7 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true); } } - ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 7 * sizeof(uint32_t), {1, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 9 * sizeof(uint32_t), {1, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); @@ -2410,10 +2543,15 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f32, "cpy_f32_f32", cpy_f32_f32_len, cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f16, "cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_f16_f16, "cpy_f16_f16", cpy_f16_f16_len, cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f16_f32, "cpy_f16_f32", cpy_f16_f32_len, cpy_f16_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_bf16,"cpy_f32_bf16",cpy_f32_bf16_len,cpy_f32_bf16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f32, "contig_cpy_f32_f32", contig_cpy_f32_f32_len, contig_cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f16, "contig_cpy_f32_f16", contig_cpy_f32_f16_len, contig_cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f16_f16, "contig_cpy_f16_f16", contig_cpy_f16_f16_len, contig_cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f16_f32, "contig_cpy_f16_f32", contig_cpy_f16_f32_len, contig_cpy_f16_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_bf16,"contig_cpy_f32_bf16",contig_cpy_f32_bf16_len,contig_cpy_f32_bf16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + if (device->float_controls_rte_fp16) { ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_rte_len, cpy_f32_q4_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_rte_len, cpy_f32_q4_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1); @@ -2437,20 +2575,32 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q8_0], "cpy_q8_0_f32", cpy_q8_0_f32_len, cpy_q8_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_IQ4_NL], "cpy_iq4_nl_f32", cpy_iq4_nl_f32_len, cpy_iq4_nl_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_add_f32, "add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1); - ggml_vk_create_pipeline(device, device->pipeline_add_f32_norepeat, "add_f32_norepeat", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1); - ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16, "add_f16_f32_f16", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1); - ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16_norepeat, "add_f16_f32_f16_norepeat", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1); + auto get_suffix = [](bool src0_f16, bool src1_f16, bool dst_f16) { + std::string s; + s += std::string(src0_f16 ? "_f16" : "_f32"); + s += std::string(src1_f16 ? "_f16" : "_f32"); + s += std::string(dst_f16 ? "_f16" : "_f32"); + return s; + }; + +#define CREATE_BINARY(name, namemod, spec) \ + for (int s0 : {0,1}) for (int s1 : {0,1}) for (int d : {0,1}) \ + ggml_vk_create_pipeline(device, device->pipeline_ ## name ## namemod[s0][s1][d], \ + #name + get_suffix(s0, s1, d) + #namemod, name ## _len[s0][s1][d], name ## _data[s0][s1][d], \ + "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, spec, 1); + + CREATE_BINARY(add, , {0}) + CREATE_BINARY(add, _norepeat, {1}) + CREATE_BINARY(sub, , {0}) + CREATE_BINARY(sub, _norepeat, {1}) + CREATE_BINARY(mul, , {0}) + CREATE_BINARY(mul, _norepeat, {1}) + CREATE_BINARY(div, , {0}) + CREATE_BINARY(div, _norepeat, {1}) +#undef CREATE_BINARY ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_sub_f32, "sub_f32", sub_f32_len, sub_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1); - ggml_vk_create_pipeline(device, device->pipeline_sub_f32_norepeat, "sub_f32_norepeat", sub_f32_len, sub_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1); - ggml_vk_create_pipeline(device, device->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1); - ggml_vk_create_pipeline(device, device->pipeline_mul_f32_norepeat, "mul_f32_norepeat", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1); - ggml_vk_create_pipeline(device, device->pipeline_div_f32, "div_f32", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1); - ggml_vk_create_pipeline(device, device->pipeline_div_f32_norepeat, "div_f32_norepeat", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1); - ggml_vk_create_pipeline(device, device->pipeline_concat_f32, "concat_f32", concat_f32_len, concat_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_concat_f16, "concat_f16", concat_f16_len, concat_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_concat_i32, "concat_i32", concat_i32_len, concat_i32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); @@ -2470,14 +2620,20 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_repeat_f32, "repeat_f32", repeat_f32_len, repeat_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_repeat_back_f32, "repeat_back_f32", repeat_back_f32_len, repeat_back_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_gelu_f32, "gelu_f32", gelu_f32_len, gelu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_gelu_quick_f32, "gelu_quick_f32", gelu_quick_f32_len, gelu_quick_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_silu_f32, "silu_f32", silu_f32_len, silu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_silu_back_f32, "silu_back_f32", silu_back_f32_len, silu_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_relu_f32, "relu_f32", relu_f32_len, relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); +#define CREATE_UNARY(name) \ + ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32", name ## _f32_len, name ## _f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); \ + ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16", name ## _f16_len, name ## _f16_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + + CREATE_UNARY(gelu) + CREATE_UNARY(gelu_quick) + CREATE_UNARY(silu) + CREATE_UNARY(relu) + CREATE_UNARY(tanh) + CREATE_UNARY(sigmoid) +#undef CREATE_UNARY + ggml_vk_create_pipeline(device, device->pipeline_leaky_relu_f32, "leaky_relu_f32", leaky_relu_f32_len, leaky_relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_tanh_f32, "tanh_f32", tanh_f32_len, tanh_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_sigmoid_f32, "sigmoid_f32", sigmoid_f32_len, sigmoid_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_silu_back_f32, "silu_back_f32", silu_back_f32_len, silu_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {1, 512, 1}, {}, 1, true); @@ -2529,6 +2685,9 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_opt_step_adamw_f32, "opt_step_adamw_f32", opt_step_adamw_f32_len, opt_step_adamw_f32_data, "main", 5, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f32, "conv2d_dw_whcn_f32", conv2d_dw_whcn_f32_len, conv2d_dw_whcn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_cwhn_f32, "conv2d_dw_cwhn_f32", conv2d_dw_cwhn_f32_len, conv2d_dw_cwhn_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1); + for (auto &c : compiles) { c.wait(); } @@ -2578,6 +2737,7 @@ static vk_device ggml_vk_get_device(size_t idx) { bool coopmat2_support = false; device->coopmat_support = false; device->integer_dot_product = false; + bool bfloat16_support = false; for (const auto& properties : ext_props) { if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) { @@ -2608,6 +2768,9 @@ static vk_device ggml_vk_get_device(size_t idx) { !getenv("GGML_VK_DISABLE_INTEGER_DOT_PRODUCT")) { device->integer_dot_product = true; #endif + } else if (strcmp("VK_KHR_shader_bfloat16", properties.extensionName) == 0 && + !getenv("GGML_VK_DISABLE_BFLOAT16")) { + bfloat16_support = true; } } @@ -2700,6 +2863,9 @@ static vk_device ggml_vk_get_device(size_t idx) { device->subgroup_add = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) && (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eArithmetic); + device->subgroup_shuffle = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) && + (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eShuffle); + const bool force_disable_f16 = getenv("GGML_VK_DISABLE_F16") != nullptr; device->fp16 = !force_disable_f16 && fp16_storage && fp16_compute; @@ -2794,6 +2960,17 @@ static vk_device ggml_vk_get_device(size_t idx) { } #endif +#if defined(VK_KHR_shader_bfloat16) + VkPhysicalDeviceShaderBfloat16FeaturesKHR bfloat16_features {}; + bfloat16_features.pNext = nullptr; + bfloat16_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_BFLOAT16_FEATURES_KHR; + if (bfloat16_support) { + last_struct->pNext = (VkBaseOutStructure *)&bfloat16_features; + last_struct = (VkBaseOutStructure *)&bfloat16_features; + device_extensions.push_back("VK_KHR_shader_bfloat16"); + } +#endif + VkPhysicalDeviceMaintenance4Features maint4_features {}; maint4_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_4_FEATURES; if (maintenance4_support) { @@ -2991,6 +3168,25 @@ static vk_device ggml_vk_get_device(size_t idx) { device->coopmat_int_n = prop.NSize; device->coopmat_int_k = prop.KSize; } +#if defined(VK_KHR_shader_bfloat16) && defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (prop.AType == VK_COMPONENT_TYPE_BFLOAT16_KHR && + prop.BType == VK_COMPONENT_TYPE_BFLOAT16_KHR && + prop.CType == VK_COMPONENT_TYPE_FLOAT32_KHR && + prop.ResultType == VK_COMPONENT_TYPE_FLOAT32_KHR && + (vk::ScopeKHR)prop.scope == vk::ScopeKHR::eSubgroup + ) { + // coopmat sizes not set yet + if (device->coopmat_m == 0) { + device->coopmat_bf16_support = true; + device->coopmat_m = prop.MSize; + device->coopmat_n = prop.NSize; + device->coopmat_k = prop.KSize; + } else if (device->coopmat_m == prop.MSize && device->coopmat_n == prop.NSize && device->coopmat_k == prop.KSize) { + // Only enable if shape is identical + device->coopmat_bf16_support = true; + } + } +#endif } if (device->coopmat_m == 0 || !device->coopmat_acc_f32_support) { @@ -2998,11 +3194,19 @@ static vk_device ggml_vk_get_device(size_t idx) { GGML_LOG_DEBUG("ggml_vulkan: WARNING: No suitable matrix core mode found. Disabling matrix cores.\n"); device->coopmat_support = false; } + if (getenv("GGML_VK_DISABLE_BFLOAT16")) { + device->coopmat_bf16_support = false; + } } if (device->coopmat_support) { device_extensions.push_back("VK_KHR_cooperative_matrix"); } +#if defined(VK_KHR_shader_bfloat16) + if (device->coopmat_bf16_support) { + device_extensions.push_back("VK_KHR_shader_bfloat16"); + } +#endif #endif device->name = GGML_VK_NAME + std::to_string(idx); @@ -3459,6 +3663,9 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) { return ctx->device->pipeline_matmul_f32_f16; } + if (src0_type == GGML_TYPE_BF16 && src1_type == GGML_TYPE_BF16) { + return ctx->device->pipeline_matmul_bf16; + } if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) { if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { return ctx->device->pipeline_matmul_f16_f32.f16acc; @@ -3530,6 +3737,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * switch (a_type) { case GGML_TYPE_F32: case GGML_TYPE_F16: + case GGML_TYPE_BF16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: @@ -3562,6 +3770,9 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) { return ctx->device->pipeline_matmul_id_f32; } + if (src0_type == GGML_TYPE_BF16 && src1_type == GGML_TYPE_BF16) { + return ctx->device->pipeline_matmul_id_bf16; + } if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) { if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { return ctx->device->pipeline_matmul_id_f16_f32.f16acc; @@ -3615,6 +3826,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context switch (a_type) { case GGML_TYPE_F32: case GGML_TYPE_F16: + case GGML_TYPE_BF16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: @@ -4350,6 +4562,20 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_cpy_f16_f16; } } + if (src->type == GGML_TYPE_F16 && to == GGML_TYPE_F32) { + if (contig) { + return ctx->device->pipeline_contig_cpy_f16_f32; + } else { + return ctx->device->pipeline_cpy_f16_f32; + } + } + if (src->type == GGML_TYPE_F32 && to == GGML_TYPE_BF16) { + if (contig) { + return ctx->device->pipeline_contig_cpy_f32_bf16; + } else { + return ctx->device->pipeline_cpy_f32_bf16; + } + } if (src->type == GGML_TYPE_F32) { switch (to) { case GGML_TYPE_Q4_0: @@ -4477,8 +4703,12 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub const bool x_non_contig = (ctx->device->coopmat2 && src0->type == GGML_TYPE_F32) || !ggml_vk_dim01_contiguous(src0); const bool y_non_contig = (ctx->device->coopmat2 && src1->type == GGML_TYPE_F32) || + (src0->type == GGML_TYPE_BF16 && src1->type != GGML_TYPE_BF16) || !ggml_vk_dim01_contiguous(src1); + // If src0 is BF16, try to use a BF16 x BF16 multiply + ggml_type f16_type = src0->type == GGML_TYPE_BF16 ? GGML_TYPE_BF16 : GGML_TYPE_F16; + const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig; bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0; @@ -4488,25 +4718,25 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub if (mmp == nullptr) { // Fall back to f16 dequant mul mat - mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type, (ggml_prec)dst->op_params[0]); + mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? f16_type : src1->type, (ggml_prec)dst->op_params[0]); quantize_y = false; } const bool qx_needs_dequant = mmp == nullptr || x_non_contig; - const bool qy_needs_dequant = !quantize_y && ((src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig); + const bool qy_needs_dequant = !quantize_y && ((src1->type != f16_type && !y_f32_kernel) || y_non_contig); if (qx_needs_dequant) { // Fall back to dequant + f16 mulmat - mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, GGML_TYPE_F16, y_f32_kernel ? GGML_TYPE_F32 : GGML_TYPE_F16, (ggml_prec)dst->op_params[0]); + mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, f16_type, y_f32_kernel ? GGML_TYPE_F32 : f16_type, (ggml_prec)dst->op_params[0]); } // Not implemented GGML_ASSERT(y_non_contig || !qy_needs_dequant); // NOLINT - const uint32_t kpad = quantize_y ? 0 : ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11, qx_needs_dequant ? GGML_TYPE_F16 : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type))); + const uint32_t kpad = quantize_y ? 0 : ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11, qx_needs_dequant ? f16_type : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type))); const bool aligned = !quantize_y && ne10 == kpad && ne01 > 8 && ne11 > 8; - vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, qx_needs_dequant ? GGML_TYPE_F16 : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type)); + vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, qx_needs_dequant ? f16_type : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type)); // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) : ne11; @@ -4527,12 +4757,12 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub vk_pipeline to_q8_1 = nullptr; if (x_non_contig) { - to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, GGML_TYPE_F16); + to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, f16_type); } else { to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type); } if (y_non_contig) { - to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, GGML_TYPE_F16); + to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, f16_type); } else { to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type); } @@ -4949,6 +5179,8 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con const uint64_t nb01 = src0->nb[1]; const uint64_t nb02 = src0->nb[2]; + const uint64_t nb12 = src1->nb[2]; + // const uint64_t ne10 = src1->ne[0]; const uint64_t ne11 = src1->ne[1]; const uint64_t ne12 = src1->ne[2]; @@ -4974,6 +5206,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con const uint32_t row_stride_x = nb01 / sizeof(ggml_fp16_t); const uint32_t channel_stride_x = nb02 / sizeof(ggml_fp16_t); + const uint32_t channel_stride_y = nb12 / sizeof(float); const uint64_t qx_sz = ggml_nbytes(src0); const uint64_t qy_sz = ggml_nbytes(src1); @@ -5004,7 +5237,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con const uint64_t d_shader_offset = d_buf_offset - d_buffer_offset; // compute - const std::array pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, (uint32_t)(ne12 / ne02), (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) }; + const std::array pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, channel_stride_y, (uint32_t)(ne12 / ne02), (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) }; ggml_vk_sync_buffers(subctx); ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 }); @@ -5029,7 +5262,7 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, c // mul_mat_vec supports batching ne12*ne13 when ne11==1, or treating ne11 as the batch size (up to four) // when ne12 and ne13 are one. } else if ((dst->ne[1] == 1 || (dst->ne[1] <= mul_mat_vec_max_cols && src1->ne[2] * src1->ne[3] == 1)) && - (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) { + (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16 || ggml_is_quantized(src0->type))) { ggml_vk_mul_mat_vec_q_f16(ctx, subctx, src0, src1, dst, dryrun); } else { ggml_vk_mul_mat_q_f16(ctx, subctx, src0, src1, dst, dryrun); @@ -5056,7 +5289,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& const uint64_t nei0 = ids->ne[0]; const uint64_t nei1 = ids->ne[1]; - GGML_ASSERT(nei0 * nei1 <= 3072); + GGML_ASSERT(nei0 * nei1 <= 4096); const uint32_t nbi1 = ids->nb[1]; const uint32_t nbi2 = ids->nb[2]; @@ -5097,27 +5330,31 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& const bool x_non_contig = (ctx->device->coopmat2 && src0->type == GGML_TYPE_F32) || !ggml_vk_dim01_contiguous(src0); const bool y_non_contig = (ctx->device->coopmat2 && src1->type == GGML_TYPE_F32) || + (src0->type == GGML_TYPE_BF16 && src1->type != GGML_TYPE_BF16) || !ggml_vk_dim01_contiguous(src1); + // If src0 is BF16, try to use a BF16 x BF16 multiply + ggml_type f16_type = src0->type == GGML_TYPE_BF16 ? GGML_TYPE_BF16 : GGML_TYPE_F16; + const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig; - vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type, (ggml_prec)dst->op_params[0]); + vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? f16_type : src1->type, (ggml_prec)dst->op_params[0]); const bool qx_needs_dequant = mmp == nullptr || x_non_contig; - const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig; + const bool qy_needs_dequant = (src1->type != f16_type && !y_f32_kernel) || y_non_contig; if (qx_needs_dequant) { // Fall back to dequant + f16 mulmat - mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, GGML_TYPE_F16, y_f32_kernel ? GGML_TYPE_F32 : GGML_TYPE_F16, (ggml_prec)dst->op_params[0]); + mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, f16_type, y_f32_kernel ? GGML_TYPE_F32 : f16_type, (ggml_prec)dst->op_params[0]); } // Not implemented GGML_ASSERT(y_non_contig || !qy_needs_dequant); // NOLINT - const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1, qx_needs_dequant ? GGML_TYPE_F16 : src0->type)); + const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1, qx_needs_dequant ? f16_type : src0->type)); const bool aligned = ne10 == kpad && ne01 > 8 && nei1 > 8; - vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned, qx_needs_dequant ? GGML_TYPE_F16 : src0->type); + vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned, qx_needs_dequant ? f16_type : src0->type); // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) :ne11; @@ -5136,12 +5373,12 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& vk_pipeline to_fp16_vk_1 = nullptr; if (x_non_contig) { - to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, GGML_TYPE_F16); + to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, f16_type); } else { to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type); } if (y_non_contig) { - to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, GGML_TYPE_F16); + to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, f16_type); } else { to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type); } @@ -5501,20 +5738,57 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx assert(q->type == GGML_TYPE_F32); assert(k->type == v->type); + bool scalar = !ctx->device->coopmat2; + + uint32_t gqa_ratio = 1; + uint32_t qk_ratio = neq2 / nek2; + uint32_t workgroups_x = (uint32_t)neq1; + uint32_t workgroups_y = (uint32_t)neq2; + uint32_t workgroups_z = (uint32_t)neq3; + + // For scalar FA, we can use the "large" size to accommodate qga. + // For coopmat FA, we always use the small size (which is still pretty large for gqa). + const uint32_t max_gqa = scalar ? scalar_flash_attention_num_large_rows : get_fa_num_small_rows(false); + + if (N == 1 && qk_ratio > 1 && qk_ratio <= max_gqa && + qk_ratio * nek2 == neq2 && nek2 == nev2 && neq3 == 1 && nek3 == 1 && nev3 == 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; + } + vk_pipeline *pipelines; // XXX TODO other backends may be changing accumulator precision to default to f32 soon - bool f32acc = dst->op_params[3] == GGML_PREC_F32; - bool small_rows = N <= flash_attention_num_small_rows; - switch (D) { - case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64[k->type][f32acc][small_rows][0]; break; - case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80[k->type][f32acc][small_rows][0]; break; - case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96[k->type][f32acc][small_rows][0]; break; - case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112[k->type][f32acc][small_rows][0]; break; - case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128[k->type][f32acc][small_rows][0]; break; - case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256[k->type][f32acc][small_rows][0]; break; - default: - assert(!"unsupported D value"); - return; + bool f32acc = scalar || dst->op_params[3] == GGML_PREC_F32; + bool small_rows = N <= get_fa_num_small_rows(scalar); + + if (scalar) { + switch (D) { + case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64[k->type][f32acc][small_rows][0]; break; + case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80[k->type][f32acc][small_rows][0]; break; + case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96[k->type][f32acc][small_rows][0]; break; + case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112[k->type][f32acc][small_rows][0]; break; + case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128[k->type][f32acc][small_rows][0]; break; + case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256[k->type][f32acc][small_rows][0]; break; + default: + GGML_ASSERT(!"unsupported D value"); + return; + } + } else { + switch (D) { + case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64_cm2[k->type][f32acc][small_rows][0]; break; + case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80_cm2[k->type][f32acc][small_rows][0]; break; + case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96_cm2[k->type][f32acc][small_rows][0]; break; + case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112_cm2[k->type][f32acc][small_rows][0]; break; + case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128_cm2[k->type][f32acc][small_rows][0]; break; + case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256_cm2[k->type][f32acc][small_rows][0]; break; + default: + GGML_ASSERT(!"unsupported D value"); + return; + } } assert(pipelines); @@ -5532,27 +5806,14 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx vk_pipeline pipeline = pipelines[aligned]; assert(pipeline); - uint32_t gqa_ratio = 1; - uint32_t qk_ratio = neq2 / nek2; - uint32_t workgroups_x = (uint32_t)neq1; - uint32_t workgroups_y = (uint32_t)neq2; - uint32_t workgroups_z = (uint32_t)neq3; - - if (N == 1 && qk_ratio > 1 && gqa_ratio <= flash_attention_num_small_rows && - qk_ratio * nek2 == neq2 && nek2 == nev2 && neq3 == 1 && nek3 == 1 && nev3 == 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; - } - uint32_t split_kv = KV; uint32_t split_k = 1; + // 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 && ctx->device->shader_core_count > 0 && KV >= 512) { + if (workgroups_x == 1 && shader_core_count > 0 && KV >= 512) { // Try to run two workgroups per SM. split_k = ctx->device->shader_core_count * 2 / workgroups_y; if (split_k > 1) { @@ -5722,26 +5983,37 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const } return nullptr; case GGML_OP_ADD: - if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_f32_norepeat : ctx->device->pipeline_add_f32; - } - if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) { - return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_f16_f32_f16_norepeat : ctx->device->pipeline_add_f16_f32_f16; - } - return nullptr; case GGML_OP_SUB: - if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_sub_f32_norepeat : ctx->device->pipeline_sub_f32; - } - return nullptr; case GGML_OP_MUL: - if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_mul_f32_norepeat : ctx->device->pipeline_mul_f32; - } - return nullptr; case GGML_OP_DIV: - if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_div_f32_norepeat : ctx->device->pipeline_div_f32; + if ((src0->type != GGML_TYPE_F32 && src0->type != GGML_TYPE_F16) || + (src1->type != GGML_TYPE_F32 && src1->type != GGML_TYPE_F16) || + (dst->type != GGML_TYPE_F32 && dst->type != GGML_TYPE_F16)) { + return nullptr; + } + switch (op) { + case GGML_OP_ADD: + { + auto pipelines = ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_norepeat : ctx->device->pipeline_add; + return pipelines[src0->type == GGML_TYPE_F16][src1->type == GGML_TYPE_F16][dst->type == GGML_TYPE_F16]; + } + case GGML_OP_SUB: + { + auto pipelines = ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_sub_norepeat : ctx->device->pipeline_sub; + return pipelines[src0->type == GGML_TYPE_F16][src1->type == GGML_TYPE_F16][dst->type == GGML_TYPE_F16]; + } + case GGML_OP_MUL: + { + auto pipelines = ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_mul_norepeat : ctx->device->pipeline_mul; + return pipelines[src0->type == GGML_TYPE_F16][src1->type == GGML_TYPE_F16][dst->type == GGML_TYPE_F16]; + } + case GGML_OP_DIV: + { + auto pipelines = ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_div_norepeat : ctx->device->pipeline_div; + return pipelines[src0->type == GGML_TYPE_F16][src1->type == GGML_TYPE_F16][dst->type == GGML_TYPE_F16]; + } + default: + break; } return nullptr; case GGML_OP_CONCAT: @@ -5835,37 +6107,25 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const } return nullptr; case GGML_OP_UNARY: + if ((src0->type != GGML_TYPE_F32 && src0->type != GGML_TYPE_F16) || + (dst->type != GGML_TYPE_F32 && dst->type != GGML_TYPE_F16) || + (src0->type != dst->type)) { + return nullptr; + } + switch (ggml_get_unary_op(dst)) { case GGML_UNARY_OP_SILU: - if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ctx->device->pipeline_silu_f32; - } - break; + return ctx->device->pipeline_silu[dst->type == GGML_TYPE_F16]; case GGML_UNARY_OP_GELU: - if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ctx->device->pipeline_gelu_f32; - } - break; + return ctx->device->pipeline_gelu[dst->type == GGML_TYPE_F16]; case GGML_UNARY_OP_GELU_QUICK: - if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ctx->device->pipeline_gelu_quick_f32; - } - break; + return ctx->device->pipeline_gelu_quick[dst->type == GGML_TYPE_F16]; case GGML_UNARY_OP_RELU: - if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ctx->device->pipeline_relu_f32; - } - break; + return ctx->device->pipeline_relu[dst->type == GGML_TYPE_F16]; case GGML_UNARY_OP_TANH: - if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ctx->device->pipeline_tanh_f32; - } - break; + return ctx->device->pipeline_tanh[dst->type == GGML_TYPE_F16]; case GGML_UNARY_OP_SIGMOID: - if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - return ctx->device->pipeline_sigmoid_f32; - } - break; + return ctx->device->pipeline_sigmoid[dst->type == GGML_TYPE_F16]; default: break; } @@ -5988,6 +6248,15 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_leaky_relu_f32; } return nullptr; + case GGML_OP_CONV_2D_DW: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + if (ggml_is_contiguous(src1)) { + return ctx->device->pipeline_conv2d_dw_whcn_f32; + } else if (ggml_is_contiguous_channels(src1)) { + return ctx->device->pipeline_conv2d_dw_cwhn_f32; + } + } + return nullptr; default: return nullptr; } @@ -6014,6 +6283,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) { case GGML_OP_REPEAT_BACK: case GGML_OP_ROPE: case GGML_OP_RMS_NORM: + case GGML_OP_CONV_2D_DW: return true; default: return false; @@ -6310,6 +6580,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co case GGML_OP_CONCAT: case GGML_OP_UPSCALE: case GGML_OP_UNARY: + case GGML_OP_CONV_2D_DW: { const uint32_t ne = ggml_nelements(dst); if (ne > 262144) { @@ -7096,6 +7367,30 @@ static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, c }, dryrun); } +static void ggml_vk_conv_2d_dw(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { + vk_op_conv2d_dw_push_constants p{}; + p.ne = ggml_nelements(dst); + p.channels = dst->ne[2]; + p.batches = dst->ne[3]; + p.dst_w = dst->ne[0]; + p.dst_h = dst->ne[1]; + p.src_w = src1->ne[0]; + p.src_h = src1->ne[1]; + p.knl_w = src0->ne[0]; + p.knl_h = src0->ne[1]; + p.stride_x = dst->op_params[0]; + p.stride_y = dst->op_params[1]; + p.pad_x = dst->op_params[2]; + p.pad_y = dst->op_params[3]; + p.dilation_x = dst->op_params[4]; + p.dilation_y = dst->op_params[5]; + + GGML_ASSERT(src0->ne[3] == p.channels); + GGML_ASSERT(src1->ne[3] == p.batches); + + ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_2D_DW, std::move(p), dryrun); +} + static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { const float * op_params = (const float *)dst->op_params; ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f }, dryrun); @@ -8116,6 +8411,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_IM2COL: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_POOL_2D: + case GGML_OP_CONV_2D_DW: case GGML_OP_RWKV_WKV6: case GGML_OP_RWKV_WKV7: case GGML_OP_LEAKY_RELU: @@ -8179,6 +8475,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_IM2COL: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_POOL_2D: + case GGML_OP_CONV_2D_DW: case GGML_OP_LEAKY_RELU: { // These operations all go through ggml_vk_op_f32, so short-circuit and @@ -8352,6 +8649,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_POOL_2D: ggml_vk_pool_2d(ctx, compute_ctx, src0, node, dryrun); + break; + case GGML_OP_CONV_2D_DW: + ggml_vk_conv_2d_dw(ctx, compute_ctx, src0, src1, node, dryrun); + break; case GGML_OP_LEAKY_RELU: ggml_vk_leaky_relu(ctx, compute_ctx, src0, node, dryrun); @@ -8473,6 +8774,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * case GGML_OP_IM2COL: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_POOL_2D: + case GGML_OP_CONV_2D_DW: case GGML_OP_RWKV_WKV6: case GGML_OP_RWKV_WKV7: case GGML_OP_LEAKY_RELU: @@ -9209,7 +9511,10 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm case GGML_UNARY_OP_RELU: case GGML_UNARY_OP_TANH: case GGML_UNARY_OP_SIGMOID: - return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32; + return ggml_is_contiguous(op->src[0]) && + (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) && + (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) && + (op->src[0]->type == op->type); default: return false; } @@ -9227,6 +9532,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm switch (src0_type) { case GGML_TYPE_F32: case GGML_TYPE_F16: + case GGML_TYPE_BF16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: @@ -9262,19 +9568,23 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm if (a->ne[3] != b->ne[3]) { return false; } - if (!(ggml_vk_dim01_contiguous(op->src[0]) || op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) || + if (!(ggml_vk_dim01_contiguous(op->src[0]) || op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_BF16) || !(ggml_vk_dim01_contiguous(op->src[1]) || op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F16)) { return false; } + if (op->src[0]->type == GGML_TYPE_BF16 && op->src[1]->type == GGML_TYPE_F16) { + // We currently don't have a bf16 x f16 shader, or an fp16->bf16 copy shader. + // So don't support this combination for now. + return false; + } return true; } break; case GGML_OP_FLASH_ATTN_EXT: { ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context; - if (!ggml_vk_get_device(ctx->device)->coopmat2) { - return false; - } + auto device = ggml_vk_get_device(ctx->device); + bool coopmat2 = device->coopmat2; switch (op->src[0]->ne[0]) { case 64: case 80: @@ -9282,7 +9592,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm case 112: case 128: case 256: - case 575: // DeepSeek MLA break; default: return false; @@ -9308,10 +9617,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm switch (op->src[1]->type) { case GGML_TYPE_F16: case GGML_TYPE_Q4_0: + case GGML_TYPE_Q8_0: + // supported in scalar and coopmat2 paths + break; case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: case GGML_TYPE_Q5_1: - case GGML_TYPE_Q8_0: // K dequants currently disabled because D dimension is rounded up to 256 and runs inefficiently //case GGML_TYPE_Q2_K: //case GGML_TYPE_Q3_K: @@ -9327,10 +9638,18 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm //case GGML_TYPE_IQ3_S: //case GGML_TYPE_IQ4_XS: case GGML_TYPE_IQ4_NL: + // currently supported only in coopmat2 path + if (!coopmat2) { + return false; + } break; default: return false; } + if (!coopmat2 && !device->subgroup_shuffle) { + // scalar FA uses subgroupShuffle + return false; + } return true; } case GGML_OP_GET_ROWS: @@ -9338,6 +9657,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm switch (op->src[0]->type) { case GGML_TYPE_F32: case GGML_TYPE_F16: + case GGML_TYPE_BF16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: @@ -9368,6 +9688,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm switch (src1_type) { case GGML_TYPE_F32: case GGML_TYPE_F16: + case GGML_TYPE_BF16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: @@ -9381,6 +9702,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm } if (src1_type == GGML_TYPE_F32) { switch (src0_type) { + case GGML_TYPE_F16: case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: case GGML_TYPE_Q5_0: @@ -9419,6 +9741,9 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: + return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) && + (op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F16) && + (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16); case GGML_OP_SILU_BACK: case GGML_OP_RMS_NORM_BACK: case GGML_OP_SQR: @@ -9442,6 +9767,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm case GGML_OP_COUNT_EQUAL: case GGML_OP_IM2COL: case GGML_OP_TIMESTEP_EMBEDDING: + case GGML_OP_CONV_2D_DW: case GGML_OP_POOL_2D: case GGML_OP_RWKV_WKV6: case GGML_OP_RWKV_WKV7: diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt b/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt index d6e0b2a5a5..ad13f69b3f 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +++ b/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt @@ -12,6 +12,9 @@ endif() if (GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) add_compile_definitions(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) endif() +if (GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + add_compile_definitions(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) +endif() set(TARGET vulkan-shaders-gen) add_executable(${TARGET} vulkan-shaders-gen.cpp) install(TARGETS ${TARGET} RUNTIME) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp b/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp index dd828c2326..6567a8c54c 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp @@ -18,7 +18,11 @@ void main() { // fast path for when all four iterations are in-bounds if (idx + (num_iter-1)*num_threads < p.ne) { [[unroll]] for (uint i = 0; i < num_iter; ++i) { -#ifndef OPTIMIZATION_ERROR_WORKAROUND + +#if defined(DATA_D_BF16) + float f = float(data_a[get_aoffset() + idx]); + data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f)); +#elif !defined(OPTIMIZATION_ERROR_WORKAROUND) data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]); #else data_d[get_doffset() + idx] = data_a[get_aoffset() + idx]; @@ -31,7 +35,10 @@ void main() { continue; } -#ifndef OPTIMIZATION_ERROR_WORKAROUND +#if defined(DATA_D_BF16) + float f = float(data_a[get_aoffset() + idx]); + data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f)); +#elif !defined(OPTIMIZATION_ERROR_WORKAROUND) data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]); #else data_d[get_doffset() + idx] = data_a[get_aoffset() + idx]; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp b/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp new file mode 100644 index 0000000000..938c74da50 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp @@ -0,0 +1,105 @@ +#version 450 + +#include "types.comp" + +layout (push_constant) uniform parameter +{ + uint ne; + uint batches; + uint channels; + uint dst_w; + uint dst_h; + uint src_w; + uint src_h; + uint knl_w; + uint knl_h; + int stride_x; + int stride_y; + int pad_x; + int pad_y; + int dilation_x; + int dilation_y; +} p; + +layout (binding = 0) readonly buffer A {A_TYPE knl_data[];}; +layout (binding = 1) readonly buffer B {B_TYPE src_data[];}; +layout (binding = 2) writeonly buffer D {D_TYPE dst_data[];}; + +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE conv_2d_dw_whcn(uint idx) { + uint i0 = idx / p.dst_w; + uint dst_x = idx - i0 * p.dst_w; + uint i1 = i0 / p.dst_h; + uint dst_y = i0 - i1 * p.dst_h; + uint n = i1 / p.channels; + uint c = i1 - n * p.channels; + + uint src_i = n * p.channels * p.src_h * p.src_w + c * p.src_h * p.src_w; + uint knl_i = c * p.knl_h * p.knl_w; + + FLOAT_TYPE sum = 0.0; + for (uint knl_y = 0; knl_y < p.knl_h; ++knl_y) { + uint src_y = dst_y * p.stride_y + knl_y * p.dilation_y - p.pad_y; + if (src_y >= p.src_h) { // src_y < 0 will wrap to a large unsigned int + continue; + } + for (uint knl_x = 0; knl_x < p.knl_w; ++knl_x) { + uint src_x = dst_x * p.stride_x + knl_x * p.dilation_x - p.pad_x; + if (src_x >= p.src_w) { // src_x < 0 will wrap to a large unsigned int + continue; + } + FLOAT_TYPE v = FLOAT_TYPE(src_data[src_i + src_y * p.src_w + src_x]); + FLOAT_TYPE k = FLOAT_TYPE(knl_data[knl_i + knl_y * p.knl_w + knl_x]); + sum = fma(v, k, sum); + } + } + return sum; +} + +FLOAT_TYPE conv_2d_dw_cwhn(uint idx) { + uint i0 = idx / p.channels; + uint c = idx - i0 * p.channels; + uint i1 = i0 / p.dst_w; + uint dst_x = i0 - i1 * p.dst_w; + uint n = i1 / p.dst_h; + uint dst_y = i1 - n * p.dst_h; + + uint src_i = n * p.channels * p.src_h * p.src_w; + uint src_row = p.src_w * p.channels; + uint knl_row = p.knl_w * p.channels; + + FLOAT_TYPE sum = 0.0; + for (uint knl_y = 0; knl_y < p.knl_h; ++knl_y) { + uint src_y = dst_y * p.stride_y + knl_y * p.dilation_y - p.pad_y; + if (src_y >= p.src_h) { // src_y < 0 will wrap to a large unsigned int + continue; + } + for (uint knl_x = 0; knl_x < p.knl_w; ++knl_x) { + uint src_x = dst_x * p.stride_x + knl_x * p.dilation_x - p.pad_x; + if (src_x >= p.src_w) { // src_x < 0 will wrap to a large unsigned int + continue; + } + FLOAT_TYPE v = FLOAT_TYPE(src_data[src_i + src_y * src_row + src_x * p.channels + c]); + FLOAT_TYPE k = FLOAT_TYPE(knl_data[ knl_y * knl_row + knl_x * p.channels + c]); + sum = fma(v, k, sum); + } + } + return sum; +} + +void main() { + uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x; + if (idx >= p.ne) { + return; + } + + FLOAT_TYPE result = +#ifdef WHCN + conv_2d_dw_whcn(idx); +#else + conv_2d_dw_cwhn(idx); +#endif + dst_data[idx] = D_TYPE(result); +} + diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp b/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp index 29c9064942..f476a2e3dd 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp @@ -12,7 +12,10 @@ void main() { return; } -#ifndef OPTIMIZATION_ERROR_WORKAROUND +#if defined(DATA_D_BF16) + float f = float(data_a[get_aoffset() + src0_idx(idx)]); + data_d[get_doffset() + dst_idx(idx)] = D_TYPE(fp32_to_bf16(f)); +#elif !defined(OPTIMIZATION_ERROR_WORKAROUND) data_d[get_doffset() + dst_idx(idx)] = D_TYPE(data_a[get_aoffset() + src0_idx(idx)]); #else data_d[get_doffset() + dst_idx(idx)] = data_a[get_aoffset() + src0_idx(idx)]; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp index 2a162a2c81..0d9739d406 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp @@ -23,6 +23,12 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) { } #endif +#if defined(DATA_A_BF16) +vec2 dequantize(uint ib, uint iqs, uint a_offset) { + return vec2(bf16_to_fp32(data_a[a_offset + ib]), bf16_to_fp32(data_a[a_offset + ib + 1])); +} +#endif + #if defined(DATA_A_Q4_0) vec2 dequantize(uint ib, uint iqs, uint a_offset) { const uint vui = uint(data_a[a_offset + ib].qs[iqs]); @@ -428,7 +434,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) { } #endif -#if defined(DATA_A_F32) || defined(DATA_A_F16) +#if defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16) vec2 get_dm(uint ib, uint a_offset) { return vec2(0, 0); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp index 962d2353f8..9cb7da2daa 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp @@ -482,7 +482,7 @@ float16_t dequantFuncIQ2_XXS(const in decodeBufIQ2_XXS bl, const in uint blockCo const uint ib8 = (idx & 0x18) >> 3; // 0..3 const uint iqs = 8 * ib32 + ib8; - const uint8_t qs = bl.block.qs[iqs]; + const uint qs = bl.block.qs[iqs]; const uint signscale = pack32(u16vec2(bl16.block.qs[4*ib32+2], bl16.block.qs[4*ib32+3])); const float dscale = float(bl.block.d) * 0.25 * (0.5 + float(signscale >> 28)); diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp new file mode 100644 index 0000000000..e6545160d5 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp @@ -0,0 +1,483 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : enable +#extension GL_EXT_shader_16bit_storage : require + +#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + +#extension GL_KHR_shader_subgroup_shuffle : enable + +#include "types.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +layout (constant_id = 1) const uint32_t Br = 1; +layout (constant_id = 2) const uint32_t Bc = 32; +layout (constant_id = 3) const uint32_t D = 32; + +layout (constant_id = 5) const uint32_t D_split = 16; +const uint32_t D_per_thread = D / D_split; + +const uint32_t cols_per_iter = gl_WorkGroupSize.x / D_split; +const uint32_t cols_per_thread = Bc / cols_per_iter; + +layout (push_constant) uniform parameter { + uint32_t N; + uint32_t KV; + + uint32_t ne1; + uint32_t ne2; + uint32_t ne3; + + uint32_t neq2; + uint32_t neq3; + uint32_t nek2; + uint32_t nek3; + uint32_t nev2; + uint32_t nev3; + uint32_t nem1; + + uint32_t nb01; + uint32_t nb02; + uint32_t nb03; + uint32_t nb11; + uint32_t nb12; + uint32_t nb13; + uint32_t nb21; + uint32_t nb22; + uint32_t nb23; + uint32_t nb31; + + float scale; + float max_bias; + float logit_softcap; + + uint32_t mask; + uint32_t n_head_log2; + float m0; + float m1; + + uint32_t gqa_ratio; + uint32_t split_kv; + uint32_t k_num; +} p; + +layout (binding = 0) readonly buffer Q {float data_q[];}; +layout (binding = 0) readonly buffer QV4 {vec4 data_qv4[];}; +layout (binding = 1) readonly buffer K {float16_t data_k[];}; +layout (binding = 1) readonly buffer KV4 {f16vec4 data_kv4[];}; +layout (binding = 2) readonly buffer V {float16_t data_v[];}; +layout (binding = 2) readonly buffer VV4 {f16vec4 data_vv4[];}; +layout (binding = 3) readonly buffer M {float16_t data_m[];}; +layout (binding = 4) writeonly buffer O {D_TYPE data_o[];}; + +#if defined(A_TYPE_PACKED16) +#define BINDING_IDX_K 0 +#define BINDING_IDX_V 1 +layout (binding = 1) readonly buffer KV_PACKED16 {A_TYPE_PACKED16 data_packed16[];} kv_packed[2]; +#endif + +#if defined(DATA_A_Q4_0) +#define BLOCK_BYTE_SIZE 18 + +vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) { + uint vui_lo = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]); + uint vui_hi = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]); + uint shift = (iqs & 0x10) >> 2; + vui_lo >>= shift; + vui_hi >>= shift; + + return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f); +} +#endif + +#if defined(DATA_A_Q8_0) +#define BLOCK_BYTE_SIZE 34 +vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) { + const i8vec2 v0 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147 + const i8vec2 v1 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy; + + return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y); +} +#endif + +#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b)) + +// Store the output when doing grouped query attention. +// Rows index by Q's dimension 2, and the first N rows are valid. +D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N) +{ + uint32_t offset = (iq2 + r) * D + c; + data_o[o_offset + offset] = D_TYPE(elem); + return elem; +} + +// Store column zero. This is used to save per-row m and L values for split_k. +ACC_TYPE perElemOpStoreCol0(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N) +{ + if (r < N && c == 0) { + uint32_t offset = iq2 + r; + data_o[o_offset + offset] = D_TYPE(elem); + } + return elem; +} + +// Load the slope matrix, indexed by Q's dimension 2. +ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2) +{ + const uint32_t h = iq2 + (r % p.gqa_ratio); + + const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1); + const int exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1); + + return ACC_TYPE(pow(base, ACC_TYPE(exph))); +} + +shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x]; +shared vec4 tmpshv4[gl_WorkGroupSize.x]; + +shared float masksh[Bc][Br]; +shared vec4 Qf[Br][D / 4]; + +void main() { +#ifdef NEEDS_INIT_IQ_SHMEM + init_iq_shmem(gl_WorkGroupSize); +#endif + + const uint32_t tid = gl_LocalInvocationIndex; + const uint32_t N = p.N; + const uint32_t KV = p.KV; + + const uint32_t d_tid = gl_LocalInvocationIndex % D_split; + const uint32_t col_tid = gl_LocalInvocationIndex / D_split; + + uint32_t i = gl_WorkGroupID.x; + uint32_t split_k_index = 0; + + if (p.k_num > 1) { + i = 0; + split_k_index = gl_WorkGroupID.x; + } + + const uint32_t Tr = CEIL_DIV(N, Br); + + const uint32_t start_j = split_k_index * p.split_kv / Bc; + const uint32_t end_j = CEIL_DIV(min(KV, (split_k_index + 1) * p.split_kv), Bc); + + // When not using grouped query attention, all rows share the same iq2, equal to gl_WorkGroupID.y. + // When using grouped query attention, each workgroup does gqa_ratio consecutive values of iq2. + const uint32_t iq2 = gl_WorkGroupID.y * p.gqa_ratio; + const uint32_t iq3 = gl_WorkGroupID.z; + + // broadcast factors + const uint32_t rk2 = p.neq2/p.nek2; + const uint32_t rk3 = p.neq3/p.nek3; + + const uint32_t rv2 = p.neq2/p.nev2; + const uint32_t rv3 = p.neq3/p.nev3; + + // k indices + const uint32_t ik3 = iq3 / rk3; + const uint32_t ik2 = iq2 / rk2; + + // v indices + const uint32_t iv3 = iq3 / rv3; + const uint32_t iv2 = iq2 / rv2; + + // nb?1 are already divided by the type size and are in units of elements. + // When using grouped query attention, Q is indexed by iq2, so the stride + // should be nb02 (which is in bytes). + uint32_t q_stride = p.gqa_ratio > 1 ? (p.nb02 / 4) : p.nb01; + uint32_t k_stride = p.nb11; + uint32_t v_stride = p.nb21; + // When using grouped query attention, all rows use the same mask (stride 0). + // "p.gqa_ratio >> 16" is just a roundabout way of writing zero + // that prevents the compiler from folding the "&" through the select + // and breaking the alignment detection. + uint32_t m_stride = (p.gqa_ratio > 1) ? (p.gqa_ratio >> 16) : KV; + + uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4; + + [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) { + uint32_t d = (idx + tid) % (D / 4); + uint32_t r = (idx + tid) / (D / 4); + if (r < Br && d < D / 4 && + i * Br + r < N) { + Qf[r][d] = vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d]) * p.scale; + } + } + barrier(); + + vec4 Of[Br][D_per_thread / 4]; + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + Of[r][d] = vec4(0.0); + } + } + + float Lf[Br], Mf[Br]; + + // Use -FLT_MAX/2 rather than -inf to reduce the possibility of NaNs, e.g. when computing Mold-M. + const float NEG_FLT_MAX_OVER_2 = uintBitsToFloat(0xFEFFFFFF); + + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + Lf[r] = 0; + Mf[r] = NEG_FLT_MAX_OVER_2; + } + + float slope[Br]; + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + slope[r] = 1.0; + } + + // ALiBi + if (p.max_bias > 0.0f) { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + slope[r] = perElemOpComputeSlope(r, col_tid, ACC_TYPE(0), iq2); + } + } + +#if BLOCK_SIZE > 1 + uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / BLOCK_BYTE_SIZE; + uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / BLOCK_BYTE_SIZE; +#else + uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2; + uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2; +#endif + + [[dont_unroll]] + for (uint32_t j = start_j; j < end_j; ++j) { + + float Sf[Br][cols_per_thread]; + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + Sf[r][c] = 0.0; + } + } + + + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { +#if BLOCK_SIZE > 1 + uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid); + uint ib = coord / BLOCK_SIZE; + uint iqs = (coord % BLOCK_SIZE); + vec4 K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K); +#else + vec4 K_Tf = vec4(data_kv4[k_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * k_stride / 4 + d * D_split + d_tid]); +#endif + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + Sf[r][c] += dot(Qf[r][d * D_split + d_tid], K_Tf); + } + } + } + + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + // Compute sum across the D_split + [[unroll]] for (uint s = D_split / 2; s > 0; s >>= 1) { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + Sf[r][c] += subgroupShuffleXor(Sf[r][c], s); + } + } + } + + if (p.logit_softcap != 0.0f) { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + Sf[r][c] = p.logit_softcap * tanh(Sf[r][c]); + } + } + } + + if (p.mask != 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) { + masksh[c][r] = float(data_m[(i * Br + r) * m_stride + (j * Bc + c)]); + } + } + barrier(); + + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + float mvf = masksh[c * cols_per_iter + col_tid][r]; + + Sf[r][c] += slope[r]*mvf; + } + } + barrier(); + } + + float rowmaxf[Br], Pf[Br][cols_per_thread], rowsumf[Br], eMf[Br], Moldf[Br]; + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + rowmaxf[r] = Sf[r][0]; + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + rowmaxf[r] = max(rowmaxf[r], Sf[r][c]); + } + Moldf[r] = Mf[r]; + + // M = max(rowmax, Mold) + // P = e^(S - M) + // eM = e^(Mold - M) + Mf[r] = max(rowmaxf[r], Moldf[r]); + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + Pf[r][c] = exp(Sf[r][c] - Mf[r]); + } + eMf[r] = exp(Moldf[r] - Mf[r]); + + // Compute sum across row of P + rowsumf[r] = 0.0; + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + rowsumf[r] += Pf[r][c]; + } + + Lf[r] = eMf[r]*Lf[r] + rowsumf[r]; + } + + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + Of[r][d] = eMf[r] * Of[r][d]; + } + } + + [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) { + [[unroll]] for (uint32_t d = 0; d < D_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 < Br; ++r) { + Of[r][d] += Pf[r][c] * Vf; + } + } + } + + barrier(); + } + + // reduce across threads + + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + float rowmaxf, eMf; + + tmpsh[tid] = Mf[r]; + // Compute max across the row + barrier(); + [[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) { + if (tid < s) { + tmpsh[tid] = max(tmpsh[tid], tmpsh[tid + s]); + } + barrier(); + } + rowmaxf = tmpsh[d_tid]; + barrier(); + + float Moldf = Mf[r]; + + // M = max(rowmax, Mold) + // eM = e^(Mold - M) + Mf[r] = max(rowmaxf, Moldf); + eMf = exp(Moldf - Mf[r]); + + Lf[r] = eMf*Lf[r]; + + tmpsh[tid] = Lf[r]; + + // Compute sum across the row + barrier(); + [[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) { + if (tid < s) { + tmpsh[tid] = tmpsh[tid] + tmpsh[tid + s]; + } + barrier(); + } + Lf[r] = tmpsh[d_tid]; + barrier(); + + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + + Of[r][d] = eMf * Of[r][d]; + tmpshv4[tid] = Of[r][d]; + + barrier(); + [[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) { + if (tid < s) { + Of[r][d] += tmpshv4[tid + s]; + tmpshv4[tid] = Of[r][d]; + } + barrier(); + } + Of[r][d] = tmpshv4[d_tid]; + 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 = D * p.ne1 * split_k_index; + + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + if (r < N) { + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) { + perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N); + } + } + } + } + + o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2; + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + if (r < N) { + perElemOpStoreCol0(r, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N); + perElemOpStoreCol0(r, 0u, ACC_TYPE(Mf[r]), o_offset + p.ne1, iq2, N); + } + } + + return; + } + + float Lfrcp[Br]; + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + Lfrcp[r] = 1.0 / Lf[r]; + } + + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + Of[r][d] *= Lfrcp[r]; + } + } + + uint32_t o_offset = iq3*p.ne2*p.ne1; + + if (p.gqa_ratio > 1) { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + if (r < N) { + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) { + perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N); + } + } + } + } + } else { + [[unroll]] for (uint32_t r = 0; r < Br; ++r) { + if (i * Br + r < N) { + [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) { + [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) { + data_o[o_offset + iq2 * D + (i * Br + r) * p.ne1 * D + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]); + } + } + } + } + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp index e877ed7796..ee6b86a18d 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp @@ -20,9 +20,14 @@ void main() { const uint a_offset = get_aoffset() + i01*p.nb01 + i11*p.nb02 + i12*p.nb03; const uint d_offset = get_doffset() + i10*p.nb21 + i11*p.nb22 + i12*p.nb23; -#ifndef OPTIMIZATION_ERROR_WORKAROUND - data_d[d_offset + i00] = D_TYPE(data_a[a_offset + i00]); +#if defined(DATA_A_BF16) + FLOAT_TYPE v = FLOAT_TYPE(bf16_to_fp32(data_a[a_offset + i00])); #else - data_d[d_offset + i00] = data_a[a_offset + i00]; + FLOAT_TYPE v = FLOAT_TYPE(data_a[a_offset + i00]); +#endif +#ifndef OPTIMIZATION_ERROR_WORKAROUND + data_d[d_offset + i00] = D_TYPE(v); +#else + data_d[d_offset + i00] = D_TYPE(v); #endif } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp index 775b48cd05..bb429dd594 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp @@ -6,7 +6,7 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; -#if !defined(DATA_A_F32) && !defined(DATA_A_F16) +#if !defined(DATA_A_F32) && !defined(DATA_A_F16) && !defined(DATA_A_BF16) #define K_PER_ITER 8 #else #define K_PER_ITER 2 diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp index 48376637fb..bc633369f9 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp @@ -21,7 +21,9 @@ layout (push_constant) uniform parameter uint nrows_x; uint row_stride_x; uint channel_stride_x; + uint channel_stride_y; uint channel_x_divisor; + uint ne12; uint b_offset; uint d_offset; } p; @@ -33,6 +35,7 @@ void main() { const uint row_x = gl_GlobalInvocationID.y; const uint channel = gl_GlobalInvocationID.z; const uint channel_x = channel / p.channel_x_divisor; + const uint channel_y = channel % p.ne12; const uint nrows_y = p.ncols_x; const uint nrows_dst = p.nrows_x; @@ -56,7 +59,7 @@ void main() { const uint row_y = col_x; const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x; - const uint iy = channel*nrows_y + row_y; + const uint iy = channel_y*p.channel_stride_y + row_y; const vec4 av4 = vec4(data_a_v4[ix / 4]); const vec4 bv4 = vec4(data_b_v4[iy / 4]); @@ -72,7 +75,7 @@ void main() { const uint row_y = col_x; const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x; - const uint iy = channel*nrows_y + row_y; + const uint iy = channel_y*p.channel_stride_y + row_y; const vec4 av4 = vec4(data_a_v4[ix / 4]); const vec4 bv4 = vec4(data_b_v4[iy / 4]); @@ -89,7 +92,7 @@ void main() { const uint row_y = col_x; const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x; - const uint iy = channel*nrows_y + row_y; + const uint iy = channel_y*p.channel_stride_y + row_y; const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]); diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp index 23ce8ceec3..7859a1a60e 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp @@ -10,6 +10,10 @@ #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require #endif +#if defined(DATA_A_BF16) && defined(COOPMAT) +#extension GL_EXT_bfloat16 : enable +#endif + #ifdef COOPMAT #extension GL_KHR_cooperative_matrix : enable #extension GL_KHR_memory_scope_semantics : enable @@ -29,6 +33,10 @@ #define LOAD_VEC_B 1 #endif +#if !defined(TO_FLOAT_TYPE) +#define TO_FLOAT_TYPE FLOAT_TYPE +#endif + layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; @@ -95,7 +103,7 @@ shared FLOAT_TYPE buf_a[BM * SHMEM_STRIDE]; shared FLOAT_TYPE buf_b[BN * SHMEM_STRIDE]; #ifdef MUL_MAT_ID -shared u16vec2 row_ids[3072]; +shared u16vec2 row_ids[4096]; #endif // MUL_MAT_ID #define NUM_WARPS (BLOCK_SIZE / WARP) @@ -202,8 +210,8 @@ void main() { #endif #ifdef COOPMAT - coopmat cache_a; - coopmat cache_b; + coopmat cache_a; + coopmat cache_b; coopmat sums[cms_per_row * cms_per_col]; [[unroll]] for (uint i = 0; i < cms_per_row * cms_per_col; i++) { @@ -248,6 +256,21 @@ void main() { buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = FLOAT_TYPE(0.0f); } #endif +#elif defined(DATA_A_BF16) +#if LOAD_VEC_A == 4 + const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; + buf_a[buf_idx ] = TO_FLOAT_TYPE(data_a[idx].x); + buf_a[buf_idx + 1] = TO_FLOAT_TYPE(data_a[idx].y); + buf_a[buf_idx + 2] = TO_FLOAT_TYPE(data_a[idx].z); + buf_a[buf_idx + 3] = TO_FLOAT_TYPE(data_a[idx].w); +#else + if (ir * BM + loadc_a + l < p.M && block + loadr_a < end_k) { + buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = TO_FLOAT_TYPE(data_a[pos_a + (loadc_a + l) * p.stride_a + loadr_a]); + } else { + buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = TO_FLOAT_TYPE(uint16_t(0)); + } +#endif #elif defined(DATA_A_Q4_0) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + 4 * loadr_a; @@ -695,13 +718,13 @@ void main() { const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b; #endif const uint buf_idx = (loadc_b + l) * SHMEM_STRIDE + loadr_b * LOAD_VEC_B; - buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx].x); - buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx].y); - buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx].z); - buf_b[buf_idx + 3] = FLOAT_TYPE(data_b[idx].w); + buf_b[buf_idx + 0] = TO_FLOAT_TYPE(data_b[idx].x); + buf_b[buf_idx + 1] = TO_FLOAT_TYPE(data_b[idx].y); + buf_b[buf_idx + 2] = TO_FLOAT_TYPE(data_b[idx].z); + buf_b[buf_idx + 3] = TO_FLOAT_TYPE(data_b[idx].w); #elif !MUL_MAT_ID if (ic * BN + loadc_b + l < p.N && block + loadr_b < end_k) { - buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]); + buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = TO_FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]); } else { buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(0.0f); } @@ -709,7 +732,7 @@ void main() { const uint row_i = ic * BN + loadc_b + l; if (row_i < _ne1) { const u16vec2 row_idx = row_ids[row_i]; - buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]); + buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = TO_FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]); } else { buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(0.0f); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp index 06b7ab09ea..9184657573 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp @@ -14,6 +14,9 @@ #extension GL_EXT_buffer_reference : enable #extension GL_KHR_shader_subgroup_ballot : enable #extension GL_KHR_shader_subgroup_vote : enable +#ifdef DATA_A_BF16 +#extension GL_EXT_bfloat16 : enable +#endif #include "types.comp" @@ -80,10 +83,16 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; #define store_scales(a) #endif +#if defined(DATA_A_BF16) +#define MAT_TYPE bfloat16_t +#else +#define MAT_TYPE FLOAT_TYPE +#endif + #ifdef MUL_MAT_ID layout (binding = 3) readonly buffer IDS {int data_ids[];}; -shared u16vec4 row_ids[3072]; +shared u16vec4 row_ids[4096]; layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufB { B_TYPE b[]; @@ -271,8 +280,8 @@ void main() { // Manually partial unroll [[unroll]] for (uint j = 0; j < unroll_count; ++j) { - coopmat mat_a; - coopmat mat_b; + coopmat mat_a; + coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose); @@ -286,8 +295,8 @@ void main() { store_scales(tid); } while (block_k < end_k) { - coopmat mat_a; - coopmat mat_b; + coopmat mat_a; + coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose); @@ -310,8 +319,8 @@ void main() { // Manually partial unroll [[unroll]] for (uint j = 0; j < unroll_count; ++j) { - coopmat mat_a; - coopmat mat_b; + coopmat mat_a; + coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose); @@ -325,8 +334,8 @@ void main() { store_scales(tid); } while (block_k < end_k) { - coopmat mat_a; - coopmat mat_b; + coopmat mat_a; + coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose); @@ -350,8 +359,8 @@ void main() { // Manually partial unroll [[unroll]] for (uint j = 0; j < unroll_count; ++j) { - coopmat mat_a; - coopmat mat_b; + coopmat mat_a; + coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose); @@ -365,8 +374,8 @@ void main() { store_scales(tid); } while (block_k < end_k) { - coopmat mat_a; - coopmat mat_b; + coopmat mat_a; + coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose); @@ -405,8 +414,8 @@ void main() { fetch_scales(ir * BM, pos_a, stride_a, block_k + BK, tid, false); } - coopmat mat_a; - coopmat mat_b; + coopmat mat_a; + coopmat mat_b; coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA); #ifdef MUL_MAT_ID diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp index 284a35caa6..83de90eb7e 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp @@ -101,7 +101,7 @@ shared FLOAT_TYPE_VEC2 buf_b_ds[BN]; #define LOAD_VEC_B 4 #ifdef MUL_MAT_ID -shared u16vec2 row_ids[3072]; +shared u16vec2 row_ids[4096]; #endif // MUL_MAT_ID #define NUM_WARPS (BLOCK_SIZE / WARP) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/relu.comp b/ggml/src/ggml-vulkan/vulkan-shaders/relu.comp index 52a19b62a6..4f806270c7 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/relu.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/relu.comp @@ -17,5 +17,5 @@ void main() { return; } - data_d[i] = max(float(data_a[i]), 0); + data_d[i] = D_TYPE(max(float(data_a[i]), 0)); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/sigmoid.comp b/ggml/src/ggml-vulkan/vulkan-shaders/sigmoid.comp index 776581e2c4..5c9e5c3503 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/sigmoid.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/sigmoid.comp @@ -16,5 +16,5 @@ void main() { if (i >= p.KX) { return; } - data_d[i] = D_TYPE(1. / (1 + exp(-1. *data_a[i]))); + data_d[i] = D_TYPE(1. / (1 + exp(-1. * float(data_a[i])))); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp b/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp index 495f966bdc..8a6f868f58 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp @@ -16,5 +16,5 @@ void main() { if (i >= p.KX) { return; } - data_d[i] = D_TYPE(1. - 2. / (exp(2.*data_a[i]) + 1.)); + data_d[i] = D_TYPE(1. - 2. / (exp(2.*float(data_a[i])) + 1.)); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/test_bfloat16_support.comp b/ggml/src/ggml-vulkan/vulkan-shaders/test_bfloat16_support.comp new file mode 100644 index 0000000000..fd0ba401fe --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/test_bfloat16_support.comp @@ -0,0 +1,7 @@ +#version 460 + +#extension GL_EXT_bfloat16 : require + +void main() +{ +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp index f5b29bfb13..3bde717832 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp @@ -33,6 +33,19 @@ #endif #endif +#if defined(DATA_A_BF16) +#define QUANT_K 1 +#define QUANT_R 1 + +#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1 +#define A_TYPE uint16_t +#elif LOAD_VEC_A == 4 +#define A_TYPE u16vec4 +#elif LOAD_VEC_A == 8 +#error unsupported +#endif +#endif + #define QUANT_K_Q4_0 32 #define QUANT_R_Q4_0 2 @@ -1343,4 +1356,18 @@ void init_iq_shmem(uvec3 wgsize) } #endif +// returns the bfloat value in the low 16b. +// See ggml_compute_fp32_to_bf16 +uint32_t fp32_to_bf16(float f) +{ + uint32_t u = floatBitsToUint(f); + u = (u + (0x7fff + ((u >> 16) & 1))) >> 16; + return u; +} + +float bf16_to_fp32(uint32_t u) +{ + return uintBitsToFloat(u << 16); +} + #endif // !defined(GGML_TYPES_COMP) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp index cf74625cc5..d196137eb2 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp @@ -63,7 +63,8 @@ const std::vector type_names = { "iq3_xxs", "iq3_s", "iq4_xs", - "iq4_nl" + "iq4_nl", + "bf16", }; namespace { @@ -296,7 +297,6 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool std::string aligned_b_type_f16 = coopmat2 ? "float16_t" : fp16 ? "f16mat2x4" : "f16vec4"; std::map base_dict = { - {"FLOAT_TYPE", (coopmat2 || fp16) ? "float16_t" : "float"}, {"FLOAT_TYPE_VEC2", (coopmat2 || fp16) ? "f16vec2" : "vec2"}, }; std::string shader_name = "matmul"; @@ -318,12 +318,45 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool const std::string source_name = coopmat2 ? "mul_mm_cm2.comp" : "mul_mm.comp"; - // Shaders with f16 B_TYPE - string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc); - string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + auto const &FLOAT_TYPE = [&](const std::string &t) -> std::string { + if (t == "bf16") { + // scalar path promotes to float + if (!coopmat && !coopmat2) { + return "float"; + } + return "bfloat16_t"; + } + if (coopmat2 || fp16) { + return "float16_t"; + } + return "float"; + }; - string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); - string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); + // Shaders with f16 B_TYPE + string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + + string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); + + // bf16 + { + std::string load_vec_a_unaligned = "1"; + // For aligned matmul loads + std::string load_vec_a = coopmat2 ? "1" : "4"; + + // scalar path promotes to float + std::string to_float_type = (coopmat || coopmat2) ? "uintBitsToBFloat16EXT" : "bf16_to_fp32"; + + // If bfloat16 is not supported, then only compile the scalar (promote to fp32) shader +#if !defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT) + if (!(coopmat || coopmat2)) +#endif + { + string_to_spv(shader_name + "_bf16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("bf16")}, {"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_bf16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("bf16")}, {"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "uint16_t"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat, coopmat2, f16acc); + } + } for (const auto& tname : type_names) { std::string load_vec_quant = "2"; @@ -332,26 +365,30 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "iq4_nl")) load_vec_quant = "4"; + if (tname == "bf16") { + continue; + } + std::string data_a_key = "DATA_A_" + to_uppercase(tname); // For unaligned, load one at a time for f32/f16, or two at a time for quants - std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16") ? "1" : load_vec_quant; + std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? "1" : load_vec_quant; // For aligned matmul loads - std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16") ? load_vec : load_vec_quant; + std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? load_vec : load_vec_quant; // don't generate f32 variants for coopmat2 if (!coopmat2) { - string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); - string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); } if (tname != "f16" && tname != "f32") { - string_to_spv(shader_name + "_" + tname + "_f16", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); - string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); } #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) if (!coopmat && !coopmat2 && !matmul_id && (tname == "q4_0" || tname == "q4_1" || tname == "q5_0" || tname == "q5_1" || tname == "q8_0")) { - string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc); } #endif } @@ -384,7 +421,6 @@ void process_shaders() { #endif } -#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) // flash attention for (const auto& f16acc : {false, true}) { std::string acctype = f16acc ? "float16_t" : "float"; @@ -393,7 +429,9 @@ void process_shaders() { if (tname == "f32") { continue; } + if (tname == "bf16") continue; +#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) if (tname == "f16") { string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp", merge_maps(base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}}), true, false, true, f16acc); @@ -402,9 +440,17 @@ void process_shaders() { string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, true, f16acc); } +#endif + if (tname == "f16") { + string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp", + merge_maps(base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}}), true, false, false, f16acc); + } else if (tname == "q4_0" || tname == "q8_0") { + std::string data_a_key = "DATA_A_" + to_uppercase(tname); + string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp", + merge_maps(base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, false, f16acc); + } } } -#endif for (const auto& tname : type_names) { // mul mat vec @@ -417,12 +463,12 @@ void process_shaders() { string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}})); // Dequant shaders - if (tname != "f16") { + if (tname != "f16" && tname != "bf16") { string_to_spv("dequant_" + tname, "dequant_" + tname + ".comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float16_t"}})); } if (!string_ends_with(tname, "_k")) { - shader = (tname == "f32" || tname == "f16") ? "get_rows.comp" : "get_rows_quant.comp"; + shader = (tname == "f32" || tname == "f16" || tname == "bf16") ? "get_rows.comp" : "get_rows_quant.comp"; if (tname == "f16") { string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}})); @@ -447,9 +493,13 @@ void process_shaders() { string_to_spv("cpy_f32_f32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); string_to_spv("cpy_f32_f16", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}); string_to_spv("cpy_f16_f16", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}); + string_to_spv("cpy_f16_f32", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}); + string_to_spv("cpy_f32_bf16","copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}}); string_to_spv("contig_cpy_f32_f32", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); string_to_spv("contig_cpy_f32_f16", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}); string_to_spv("contig_cpy_f16_f16", "contig_copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}); + string_to_spv("contig_cpy_f16_f32", "contig_copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}); + string_to_spv("contig_cpy_f32_bf16","contig_copy.comp",{{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}}); for (std::string t : {"q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) { string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); @@ -457,8 +507,26 @@ void process_shaders() { string_to_spv("cpy_" + t + "_f32", "copy_from_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); } - string_to_spv("add_f32", "add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); - string_to_spv("add_f16_f32_f16", "add.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}}); + auto get_type_str = [](bool f16) { + return f16 ? "float16_t" : "float"; + }; + auto get_suffix = [](bool src0_f16, bool src1_f16, bool dst_f16) { + std::string s; + s += std::string(src0_f16 ? "_f16" : "_f32"); + s += std::string(src1_f16 ? "_f16" : "_f32"); + s += std::string(dst_f16 ? "_f16" : "_f32"); + return s; + }; + for (std::string op : {"add", "sub", "mul", "div"}) { + for (auto src0_f16 : {false, true}) { + for (auto src1_f16 : {false, true}) { + for (auto dst_f16 : {false, true}) { + auto name = op + get_suffix(src0_f16, src1_f16, dst_f16); + string_to_spv(name.c_str(), op + ".comp", {{"A_TYPE", get_type_str(src0_f16)}, {"B_TYPE", get_type_str(src1_f16)}, {"D_TYPE", get_type_str(dst_f16)}, {"FLOAT_TYPE", "float"}}); + } + } + } + } string_to_spv("sub_f32", "sub.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); @@ -493,14 +561,21 @@ void process_shaders() { string_to_spv("upscale_f32", "upscale.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}); - string_to_spv("gelu_f32", "gelu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); - string_to_spv("gelu_quick_f32", "gelu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); - string_to_spv("silu_f32", "silu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); - string_to_spv("silu_back_f32", "silu_back.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}); - string_to_spv("relu_f32", "relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); - string_to_spv("leaky_relu_f32", "leaky_relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); - string_to_spv("tanh_f32", "tanh.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); - string_to_spv("sigmoid_f32", "sigmoid.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("gelu_f16", "gelu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("gelu_f32", "gelu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("gelu_quick_f16", "gelu_quick.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("gelu_quick_f32", "gelu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("silu_f16", "silu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("silu_f32", "silu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("relu_f16", "relu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("relu_f32", "relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("tanh_f16", "tanh.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("tanh_f32", "tanh.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("sigmoid_f16", "sigmoid.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("sigmoid_f32", "sigmoid.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + + string_to_spv("leaky_relu_f32", "leaky_relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("silu_back_f32", "silu_back.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}); string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); @@ -544,6 +619,9 @@ void process_shaders() { string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}})); + string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}})); + string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}})); + for (auto &c : compiles) { c.wait(); } @@ -598,7 +676,12 @@ void write_output_files() { std::remove(path.c_str()); } } - + for (const char *op : {"add", "sub", "mul", "div"}) { + fprintf(hdr, "extern unsigned char *%s_data[2][2][2];\n", op); + fprintf(hdr, "extern uint64_t %s_len[2][2][2];\n", op); + fprintf(src, "unsigned char *%s_data[2][2][2] = {{{%s_f32_f32_f32_data, %s_f32_f32_f16_data}, {%s_f32_f16_f32_data, %s_f32_f16_f16_data}}, {{%s_f16_f32_f32_data, %s_f16_f32_f16_data}, {%s_f16_f16_f32_data, %s_f16_f16_f16_data}}};\n", op, op, op, op, op, op, op, op, op); + fprintf(src, "uint64_t %s_len[2][2][2] = {{{%s_f32_f32_f32_len, %s_f32_f32_f16_len}, {%s_f32_f16_f32_len, %s_f32_f16_f16_len}}, {{%s_f16_f32_f32_len, %s_f16_f32_f16_len}, {%s_f16_f16_f32_len, %s_f16_f16_f16_len}}};\n", op, op, op, op, op, op, op, op, op); + } fclose(hdr); fclose(src); } diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 2a39dc7bfd..8a6546240f 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -4,6 +4,7 @@ #include "ggml-backend.h" #include "ggml-impl.h" #include "ggml-threading.h" +#include "ggml-cpu.h" #include "ggml.h" // FIXME: required here for quantization functions @@ -382,58 +383,16 @@ void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int64_t n) { } } -// FIXME: these functions must detect the instruction set at runtime, since they are part of the core ggml library -// currently, the ggml_cpu_has_* functions are entirely compile-time void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int64_t n) { - int64_t i = 0; -#if defined(__F16C__) - //if (ggml_cpu_has_f16c()) { - for (; i + 7 < n; i += 8) { - __m256 x_vec = _mm256_loadu_ps(x + i); - __m128i y_vec = _mm256_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT); - _mm_storeu_si128((__m128i *)(y + i), y_vec); - } - for(; i + 3 < n; i += 4) { - __m128 x_vec = _mm_loadu_ps(x + i); - __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT); - _mm_storel_epi64((__m128i *)(y + i), y_vec); - } - //} -#endif - for (; i < n; i++) { + int i = 0; + for (; i < n; ++i) { y[i] = GGML_FP32_TO_FP16(x[i]); } } void ggml_bf16_to_fp32_row(const ggml_bf16_t * x, float * y, int64_t n) { - int64_t i = 0; -#if defined(__AVX512F__) - //if (ggml_cpu_has_avx512()) { - for (; i + 16 <= n; i += 16) { - _mm512_storeu_ps(y + i, - _mm512_castsi512_ps( - _mm512_slli_epi32( - _mm512_cvtepu16_epi32( - _mm256_loadu_si256( - (const __m256i *)(x + i))), - 16))); - } - //} -#endif -#if defined(__AVX2__) - //if (ggml_cpu_has_avx2()) { - for (; i + 8 <= n; i += 8) { - _mm256_storeu_ps(y + i, - _mm256_castsi256_ps( - _mm256_slli_epi32( - _mm256_cvtepu16_epi32( - _mm_loadu_si128( - (const __m128i *)(x + i))), - 16))); - } - //} -#endif - for (; i < n; i++) { + int i = 0; + for (; i < n; ++i) { y[i] = GGML_BF16_TO_FP32(x[i]); } } @@ -1340,6 +1299,10 @@ bool ggml_is_contiguous_2(const struct ggml_tensor * tensor) { return ggml_is_contiguous_n(tensor, 2); } +bool ggml_is_contiguously_allocated(const struct ggml_tensor * tensor) { + return ggml_nbytes(tensor) == ggml_nelements(tensor) * ggml_type_size(tensor->type)/ggml_blck_size(tensor->type); +} + bool ggml_is_permuted(const struct ggml_tensor * tensor) { static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); @@ -2769,11 +2732,11 @@ void ggml_mul_mat_set_prec( c = ggml_mul_mat_id(ctx, as, b, ids); as -> [cols, rows, n_expert] - ids -> [n_experts_used, n_tokens] (i32) b -> [cols, n_expert_used, n_tokens] + ids -> [n_expert_used, n_tokens] (i32) c -> [rows, n_expert_used, n_tokens] - in b, n_experts_used can be broadcasted to match the n_expert_used of ids + in b, n_expert_used can be broadcasted to match the n_expert_used of ids c ~= as[:,:,i] @ b[:,i%r,t], i = ids[e,t] for all e,t in ids */ @@ -5536,7 +5499,7 @@ static void ggml_compute_backward( // tensor = src0 * 1 + src1 * 0 if (src0_needs_grads) { // dsrc0 = dtensor * 1 - ggml_add_or_set(ctx, cgraph, isrc0, grad); + ggml_add_or_set(ctx, cgraph, isrc0, ggml_reshape(ctx, grad, src0)); } if (src1_needs_grads) { // dsrc1 = dtensor * 0 -> noop @@ -5817,10 +5780,9 @@ void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * } void ggml_build_backward_expand( - struct ggml_context * ctx_static, - struct ggml_context * ctx_compute, - struct ggml_cgraph * cgraph, - bool accumulate) { + struct ggml_context * ctx, + struct ggml_cgraph * cgraph, + struct ggml_tensor ** grad_accs) { GGML_ASSERT(cgraph->n_nodes > 0); GGML_ASSERT(cgraph->grads); GGML_ASSERT(cgraph->grad_accs); @@ -5893,21 +5855,24 @@ void ggml_build_backward_expand( GGML_ASSERT(!node->view_src || node->op == GGML_OP_CPY || node->op == GGML_OP_VIEW || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE); - const size_t igrad = ggml_hash_find(&cgraph->visited_hash_set, node); - GGML_ASSERT(igrad != GGML_HASHSET_FULL); - GGML_ASSERT(ggml_bitset_get(cgraph->visited_hash_set.used, igrad)); - if ((accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) || (node->flags & GGML_TENSOR_FLAG_LOSS)) { - cgraph->grad_accs[igrad] = ggml_dup_tensor(ctx_static, node); - cgraph->grads[igrad] = cgraph->grad_accs[igrad]; - ggml_format_name(cgraph->grad_accs[igrad], "grad acc for %s", node->name); + const size_t ihash = ggml_hash_find(&cgraph->visited_hash_set, node); + GGML_ASSERT(ihash != GGML_HASHSET_FULL); + GGML_ASSERT(ggml_bitset_get(cgraph->visited_hash_set.used, ihash)); + if (grad_accs && grad_accs[i]) { + cgraph->grad_accs[ihash] = grad_accs[i]; + cgraph->grads[ihash] = cgraph->grad_accs[ihash]; + } else if (node->flags & GGML_TENSOR_FLAG_LOSS) { + // loss tensors always need a gradient accumulator + cgraph->grad_accs[ihash] = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne); + cgraph->grads[ihash] = cgraph->grad_accs[ihash]; } - grads_needed[igrad] = true; + grads_needed[ihash] = true; } for (int i = n_nodes_f - 1; i >= 0; --i) { // inplace operations to add gradients are not created by ggml_compute_backward except for gradient accumulation // use allocator to automatically make inplace operations - ggml_compute_backward(ctx_compute, cgraph, i, grads_needed); + ggml_compute_backward(ctx, cgraph, i, grads_needed); } free(grads_needed); @@ -6053,8 +6018,8 @@ void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) { } } -struct ggml_cgraph * ggml_graph_dup(struct ggml_context * ctx, struct ggml_cgraph * cgraph) { - struct ggml_cgraph * result = ggml_new_graph_custom(ctx, cgraph->size, cgraph->grads != NULL); +struct ggml_cgraph * ggml_graph_dup(struct ggml_context * ctx, struct ggml_cgraph * cgraph, bool force_grads) { + struct ggml_cgraph * result = ggml_new_graph_custom(ctx, cgraph->size, cgraph->grads || force_grads); ggml_graph_cpy(cgraph, result); return result; } @@ -6073,6 +6038,9 @@ struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor) { } void ggml_graph_reset(struct ggml_cgraph * cgraph) { + if (!cgraph) { + return; + } GGML_ASSERT(cgraph->grads != NULL); for (int i = 0; i < cgraph->n_nodes; i++) { @@ -6382,8 +6350,8 @@ void ggml_set_output(struct ggml_tensor * tensor) { tensor->flags |= GGML_TENSOR_FLAG_OUTPUT; } -void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor) { - GGML_UNUSED(ctx); // TODO: remove this parameter +void ggml_set_param(struct ggml_tensor * tensor) { + GGML_ASSERT(tensor->op == GGML_OP_NONE); tensor->flags |= GGML_TENSOR_FLAG_PARAM; } diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index b81017b142..0e6226b900 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -104,6 +104,7 @@ class Keys: EXPERT_WEIGHTS_SCALE = "{arch}.expert_weights_scale" EXPERT_WEIGHTS_NORM = "{arch}.expert_weights_norm" EXPERT_GATING_FUNC = "{arch}.expert_gating_func" + MOE_EVERY_N_LAYERS = "{arch}.moe_every_n_layers" POOLING_TYPE = "{arch}.pooling_type" LOGIT_SCALE = "{arch}.logit_scale" DECODER_START_TOKEN_ID = "{arch}.decoder_start_token_id" @@ -230,8 +231,10 @@ class Keys: BLOCK_COUNT = "clip.vision.block_count" IMAGE_MEAN = "clip.vision.image_mean" IMAGE_STD = "clip.vision.image_std" + SPATIAL_MERGE_SIZE = "clip.vision.spatial_merge_size" USE_GELU = "clip.use_gelu" USE_SILU = "clip.use_silu" + N_WA_PATTERN = "clip.vision.n_wa_pattern" # used by qwen2.5vl class Attention: HEAD_COUNT = "clip.vision.attention.head_count" @@ -267,6 +270,7 @@ class MODEL_ARCH(IntEnum): REFACT = auto() BERT = auto() NOMIC_BERT = auto() + NOMIC_BERT_MOE = auto() JINA_BERT_V2 = auto() BLOOM = auto() STABLELM = auto() @@ -479,7 +483,9 @@ class MODEL_TENSOR(IntEnum): V_ENC_EMBD_PATCH = auto() V_ENC_EMBD_POS = auto() V_ENC_ATTN_Q = auto() + V_ENC_ATTN_Q_NORM = auto() V_ENC_ATTN_K = auto() + V_ENC_ATTN_K_NORM = auto() V_ENC_ATTN_V = auto() V_ENC_INPUT_NORM = auto() V_ENC_OUTPUT = auto() @@ -487,8 +493,11 @@ class MODEL_TENSOR(IntEnum): V_ENC_FFN_UP = auto() V_ENC_FFN_GATE = auto() V_ENC_FFN_DOWN = auto() + V_LAYER_SCALE_1 = auto() + V_LAYER_SCALE_2 = auto() V_PRE_NORM = auto() V_POST_NORM = auto() + V_MM_INP_NORM = auto() V_MM_INP_PROJ = auto() # gemma3 V_MM_SOFT_EMB_NORM = auto() # gemma3 V_RESMPL_POS_EMBD_K = auto() # minicpmv @@ -503,6 +512,7 @@ class MODEL_TENSOR(IntEnum): V_RESMPL_PROJ = auto() # minicpmv V_RESMPL_QUERY = auto() # minicpmv V_TOK_EMBD_IMG_BREAK = auto() # pixtral + V_MM_PATCH_MERGER = auto() # mistral small 3.1 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { @@ -521,6 +531,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.REFACT: "refact", MODEL_ARCH.BERT: "bert", MODEL_ARCH.NOMIC_BERT: "nomic-bert", + MODEL_ARCH.NOMIC_BERT_MOE: "nomic-bert-moe", MODEL_ARCH.JINA_BERT_V2: "jina-bert-v2", MODEL_ARCH.BLOOM: "bloom", MODEL_ARCH.STABLELM: "stablelm", @@ -733,7 +744,9 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = { MODEL_TENSOR.V_ENC_EMBD_PATCH: "v.patch_embd", MODEL_TENSOR.V_ENC_EMBD_POS: "v.position_embd", MODEL_TENSOR.V_ENC_ATTN_Q: "v.blk.{bid}.attn_q", + MODEL_TENSOR.V_ENC_ATTN_Q_NORM: "v.blk.{bid}.attn_q_norm", MODEL_TENSOR.V_ENC_ATTN_K: "v.blk.{bid}.attn_k", + MODEL_TENSOR.V_ENC_ATTN_K_NORM: "v.blk.{bid}.attn_k_norm", MODEL_TENSOR.V_ENC_ATTN_V: "v.blk.{bid}.attn_v", MODEL_TENSOR.V_ENC_INPUT_NORM: "v.blk.{bid}.ln1", MODEL_TENSOR.V_ENC_OUTPUT: "v.blk.{bid}.attn_out", @@ -741,9 +754,12 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = { MODEL_TENSOR.V_ENC_FFN_UP: "v.blk.{bid}.ffn_up", MODEL_TENSOR.V_ENC_FFN_GATE: "v.blk.{bid}.ffn_gate", MODEL_TENSOR.V_ENC_FFN_DOWN: "v.blk.{bid}.ffn_down", + MODEL_TENSOR.V_LAYER_SCALE_1: "v.blk.{bid}.ls1", + MODEL_TENSOR.V_LAYER_SCALE_2: "v.blk.{bid}.ls2", MODEL_TENSOR.V_PRE_NORM: "v.pre_ln", MODEL_TENSOR.V_POST_NORM: "v.post_ln", MODEL_TENSOR.V_MM_INP_PROJ: "mm.input_projection", + MODEL_TENSOR.V_MM_INP_NORM: "mm.input_norm", MODEL_TENSOR.V_MM_SOFT_EMB_NORM: "mm.soft_emb_norm", MODEL_TENSOR.V_RESMPL_POS_EMBD_K: "resampler.pos_embd_k", MODEL_TENSOR.V_RESMPL_ATTN_Q: "resampler.attn.q", @@ -757,6 +773,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = { MODEL_TENSOR.V_RESMPL_PROJ: "resampler.proj", MODEL_TENSOR.V_RESMPL_QUERY: "resampler.query", MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK: "v.token_embd.img_break", # pixtral + MODEL_TENSOR.V_MM_PATCH_MERGER: "mm.patch_merger", # mistral small 3.1 } MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { @@ -769,7 +786,9 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.V_ENC_EMBD_PATCH, MODEL_TENSOR.V_ENC_EMBD_POS, MODEL_TENSOR.V_ENC_ATTN_Q, + MODEL_TENSOR.V_ENC_ATTN_Q_NORM, MODEL_TENSOR.V_ENC_ATTN_K, + MODEL_TENSOR.V_ENC_ATTN_K_NORM, MODEL_TENSOR.V_ENC_ATTN_V, MODEL_TENSOR.V_ENC_INPUT_NORM, MODEL_TENSOR.V_ENC_OUTPUT, @@ -777,9 +796,12 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.V_ENC_FFN_UP, MODEL_TENSOR.V_ENC_FFN_GATE, MODEL_TENSOR.V_ENC_FFN_DOWN, + MODEL_TENSOR.V_LAYER_SCALE_1, + MODEL_TENSOR.V_LAYER_SCALE_2, MODEL_TENSOR.V_PRE_NORM, MODEL_TENSOR.V_POST_NORM, MODEL_TENSOR.V_MM_INP_PROJ, + MODEL_TENSOR.V_MM_INP_NORM, MODEL_TENSOR.V_MM_SOFT_EMB_NORM, MODEL_TENSOR.V_RESMPL_POS_EMBD_K, MODEL_TENSOR.V_RESMPL_ATTN_Q, @@ -793,6 +815,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.V_RESMPL_PROJ, MODEL_TENSOR.V_RESMPL_QUERY, MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK, + MODEL_TENSOR.V_MM_PATCH_MERGER, ], MODEL_ARCH.LLAMA: [ MODEL_TENSOR.TOKEN_EMBD, @@ -960,6 +983,22 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_UP, MODEL_TENSOR.LAYER_OUT_NORM, ], + MODEL_ARCH.NOMIC_BERT_MOE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.TOKEN_EMBD_NORM, + MODEL_TENSOR.TOKEN_TYPES, + MODEL_TENSOR.POS_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_OUT_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.LAYER_OUT_NORM, + ], MODEL_ARCH.JINA_BERT_V2: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.TOKEN_EMBD_NORM, @@ -2006,6 +2045,8 @@ class PoolingType(IntEnum): NONE = 0 MEAN = 1 CLS = 2 + LAST = 3 + RANK = 4 class GGMLQuantizationType(IntEnum): @@ -2136,6 +2177,9 @@ class VisionProjectorType: GEMMA3 = "gemma3" IDEFICS3 = "idefics3" PIXTRAL = "pixtral" + QWEN2VL = "qwen2vl_merger" + QWEN25VL = "qwen2.5vl_merger" + INTERNVL = "internvl" # Items here are (block size, type size) diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 48e9a470b7..ff50d3de31 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -728,6 +728,9 @@ class GGUFWriter: def add_expert_gating_func(self, value: ExpertGatingFuncType) -> None: self.add_uint32(Keys.LLM.EXPERT_GATING_FUNC.format(arch=self.arch), value.value) + def add_moe_every_n_layers(self, value: int) -> None: + self.add_uint32(Keys.LLM.MOE_EVERY_N_LAYERS.format(arch=self.arch), value) + def add_swin_norm(self, value: bool) -> None: self.add_bool(Keys.LLM.SWIN_NORM.format(arch=self.arch), value) @@ -969,6 +972,9 @@ class GGUFWriter: def add_vision_image_std(self, values: Sequence[float]) -> None: self.add_array(Keys.ClipVision.IMAGE_STD, values) + def add_vision_spatial_merge_size(self, value: int) -> None: + self.add_uint32(Keys.ClipVision.SPATIAL_MERGE_SIZE, value) + def add_vision_use_gelu(self, value: bool) -> None: self.add_bool(Keys.ClipVision.USE_GELU, value) @@ -978,6 +984,9 @@ class GGUFWriter: def add_vision_projector_scale_factor(self, value: int) -> None: self.add_uint32(Keys.ClipVision.Projector.SCALE_FACTOR, value) + def add_vision_n_wa_pattern(self, value: int) -> None: + self.add_uint32(Keys.ClipVision.N_WA_PATTERN, value) + def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes: pack_prefix = '' if not skip_pack_prefix: diff --git a/gguf-py/gguf/scripts/__init__.py b/gguf-py/gguf/scripts/__init__.py deleted file mode 100644 index 72cc73e700..0000000000 --- a/gguf-py/gguf/scripts/__init__.py +++ /dev/null @@ -1,7 +0,0 @@ -# pyright: reportUnusedImport=false - -from .gguf_convert_endian import main as gguf_convert_endian_entrypoint -from .gguf_dump import main as gguf_dump_entrypoint -from .gguf_set_metadata import main as gguf_set_metadata_entrypoint -from .gguf_new_metadata import main as gguf_new_metadata_entrypoint -from .gguf_editor_gui import main as gguf_editor_gui_entrypoint diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 1d70551973..ecf21b2b44 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -290,6 +290,7 @@ class TensorNameMap: "transformer.blocks.{bid}.ffn.router.layer", # dbrx "model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe "language_model.model.layers.{bid}.feed_forward.router", # llama4 + "encoder.layers.{bid}.mlp.router.layer", # nomic-bert-moe ), MODEL_TENSOR.FFN_GATE_INP_SHEXP: ( @@ -322,6 +323,7 @@ class TensorNameMap: "model.layers.layers.{bid}.mlp.up_proj", # plamo "model.layers.{bid}.feed_forward.w3", # internlm2 "encoder.layers.{bid}.mlp.fc11", # nomic-bert + "encoder.layers.{bid}.mlp.fc1", # nomic-bert-moe "model.layers.{bid}.mlp.c_fc", # starcoder2 "encoder.layer.{bid}.mlp.gated_layers_v", # jina-bert-v2 "model.layers.{bid}.residual_mlp.w3", # arctic @@ -337,6 +339,7 @@ class TensorNameMap: "model.layers.{bid}.mlp.experts.up_proj", # qwen2moe olmoe (merged) "model.layers.{bid}.block_sparse_moe.experts.w3", # phimoe (merged) "language_model.model.layers.{bid}.feed_forward.experts.up_proj", # llama4 + "encoder.layers.{bid}.mlp.experts.mlp.w1", # nomic-bert-moe ), MODEL_TENSOR.FFN_UP_SHEXP: ( @@ -418,6 +421,7 @@ class TensorNameMap: "model.layers.{bid}.block_sparse_moe.output_linear", # granitemoe "model.layers.{bid}.block_sparse_moe.experts.w2", # phimoe (merged) "language_model.model.layers.{bid}.feed_forward.experts.down_proj", # llama4 + "encoder.layers.{bid}.mlp.experts.mlp.w2", # nomic-bert-moe ), MODEL_TENSOR.FFN_DOWN_SHEXP: ( @@ -892,6 +896,7 @@ class TensorNameMap: MODEL_TENSOR.V_MMPROJ: ( "multi_modal_projector.linear_{bid}", + "visual.merger.mlp.{bid}", # qwen2vl ), MODEL_TENSOR.V_MMPROJ_FC: ( @@ -900,6 +905,7 @@ class TensorNameMap: MODEL_TENSOR.V_MMPROJ_MLP: ( "model.mm_projector.mlp.mlp.{bid}", + "mlp1.{bid}", # InternVL ), MODEL_TENSOR.V_MMPROJ_PEG: ( @@ -915,6 +921,7 @@ class TensorNameMap: "vpm.embeddings.patch_embedding", "model.vision_model.embeddings.patch_embedding", # SmolVLM "vision_tower.patch_conv", # pixtral + "visual.patch_embed.proj", # qwen2vl ), MODEL_TENSOR.V_ENC_EMBD_POS: ( @@ -928,6 +935,11 @@ class TensorNameMap: "vpm.encoder.layers.{bid}.self_attn.q_proj", "model.vision_model.encoder.layers.{bid}.self_attn.q_proj", # SmolVLM "vision_tower.transformer.layers.{bid}.attention.q_proj", # pixtral + "visual.blocks.{bid}.attn.q", # qwen2vl, generated + ), + + MODEL_TENSOR.V_ENC_ATTN_Q_NORM: ( + "vision_tower.vision_model.encoder.layers.{bid}.attn.q_norm", # InternVL ), MODEL_TENSOR.V_ENC_ATTN_K: ( @@ -935,6 +947,11 @@ class TensorNameMap: "vpm.encoder.layers.{bid}.self_attn.k_proj", "model.vision_model.encoder.layers.{bid}.self_attn.k_proj", # SmolVLM "vision_tower.transformer.layers.{bid}.attention.k_proj", # pixtral + "visual.blocks.{bid}.attn.k", # qwen2vl, generated + ), + + MODEL_TENSOR.V_ENC_ATTN_K_NORM: ( + "vision_tower.vision_model.encoder.layers.{bid}.attn.k_norm", # InternVL ), MODEL_TENSOR.V_ENC_ATTN_V: ( @@ -942,45 +959,65 @@ class TensorNameMap: "vpm.encoder.layers.{bid}.self_attn.v_proj", "model.vision_model.encoder.layers.{bid}.self_attn.v_proj", # SmolVLM "vision_tower.transformer.layers.{bid}.attention.v_proj", # pixtral + "visual.blocks.{bid}.attn.v", # qwen2vl, generated ), MODEL_TENSOR.V_ENC_INPUT_NORM: ( "vision_tower.vision_model.encoder.layers.{bid}.layer_norm1", + "vision_tower.vision_model.encoder.layers.{bid}.norm1", # InternVL "vpm.encoder.layers.{bid}.layer_norm1", "model.vision_model.encoder.layers.{bid}.layer_norm1", # SmolVLM "vision_tower.transformer.layers.{bid}.attention_norm", # pixtral + "visual.blocks.{bid}.norm1", # qwen2vl ), MODEL_TENSOR.V_ENC_OUTPUT: ( "vision_tower.vision_model.encoder.layers.{bid}.self_attn.out_proj", + "vision_tower.vision_model.encoder.layers.{bid}.attn.proj", # InternVL "vpm.encoder.layers.{bid}.self_attn.out_proj", "model.vision_model.encoder.layers.{bid}.self_attn.out_proj", # SmolVLM "vision_tower.transformer.layers.{bid}.attention.o_proj", # pixtral + "visual.blocks.{bid}.attn.proj", # qwen2vl ), MODEL_TENSOR.V_ENC_OUTPUT_NORM: ( "vision_tower.vision_model.encoder.layers.{bid}.layer_norm2", + "vision_tower.vision_model.encoder.layers.{bid}.norm2", # InternVL "vpm.encoder.layers.{bid}.layer_norm2", "model.vision_model.encoder.layers.{bid}.layer_norm2", # SmolVLM "vision_tower.transformer.layers.{bid}.ffn_norm", # pixtral + "visual.blocks.{bid}.norm2", # qwen2vl ), MODEL_TENSOR.V_ENC_FFN_UP: ( "vision_tower.vision_model.encoder.layers.{bid}.mlp.fc1", "vpm.encoder.layers.{bid}.mlp.fc1", - "model.vision_model.encoder.layers.{bid}.mlp.fc2", # SmolVLM, gemma3 (note: name is swapped) + "model.vision_model.encoder.layers.{bid}.mlp.fc1", # SmolVLM, gemma3 "vision_tower.transformer.layers.{bid}.feed_forward.up_proj", # pixtral + "visual.blocks.{bid}.mlp.fc1", # qwen2vl + "visual.blocks.{bid}.mlp.up_proj", # qwen2.5vl ), MODEL_TENSOR.V_ENC_FFN_GATE: ( "vision_tower.transformer.layers.{bid}.feed_forward.gate_proj", # pixtral + "visual.blocks.{bid}.mlp.gate_proj", # qwen2.5vl ), MODEL_TENSOR.V_ENC_FFN_DOWN: ( "vision_tower.vision_model.encoder.layers.{bid}.mlp.fc2", "vpm.encoder.layers.{bid}.mlp.fc2", - "model.vision_model.encoder.layers.{bid}.mlp.fc1", # SmolVLM, gemma3 (note: name is swapped) + "model.vision_model.encoder.layers.{bid}.mlp.fc2", # SmolVLM, gemma3 "vision_tower.transformer.layers.{bid}.feed_forward.down_proj", # pixtral + "visual.blocks.{bid}.mlp.fc2", # qwen2vl + "visual.blocks.{bid}.mlp.down_proj", # qwen2.5vl + ), + + MODEL_TENSOR.V_LAYER_SCALE_1: ( + "vision_tower.vision_model.encoder.layers.{bid}.ls1", # InternVL + ), + + MODEL_TENSOR.V_LAYER_SCALE_2: ( + "vision_tower.vision_model.encoder.layers.{bid}.ls2", # InternVL ), MODEL_TENSOR.V_PRE_NORM: ( @@ -991,12 +1028,17 @@ class TensorNameMap: MODEL_TENSOR.V_POST_NORM: ( "vision_tower.vision_model.post_layernorm", "model.vision_model.post_layernorm", # SmolVLM + "visual.merger.ln_q", # qwen2vl ), MODEL_TENSOR.V_MM_INP_PROJ: ( "multi_modal_projector.mm_input_projection", ), + MODEL_TENSOR.V_MM_INP_NORM: ( + "multi_modal_projector.norm", + ), + MODEL_TENSOR.V_MM_SOFT_EMB_NORM: ( "multi_modal_projector.mm_soft_emb_norm", ), @@ -1048,6 +1090,10 @@ class TensorNameMap: MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK: ( "v.token_embd.img_break", # for pixtral, this is a generated vector ), + + MODEL_TENSOR.V_MM_PATCH_MERGER: ( + "multi_modal_projector.patch_merger.merging_layer", # mistral small 3.1 + ), } # architecture-specific block mappings diff --git a/gguf-py/pyproject.toml b/gguf-py/pyproject.toml index 0c82725677..bb9b86ace7 100644 --- a/gguf-py/pyproject.toml +++ b/gguf-py/pyproject.toml @@ -1,6 +1,6 @@ [tool.poetry] name = "gguf" -version = "0.16.2" +version = "0.16.3" description = "Read and write ML models in GGUF for GGML" authors = ["GGML "] packages = [ @@ -36,8 +36,8 @@ requires = ["poetry-core>=1.0.0"] build-backend = "poetry.core.masonry.api" [tool.poetry.scripts] -gguf-convert-endian = "gguf.scripts:gguf_convert_endian_entrypoint" -gguf-dump = "gguf.scripts:gguf_dump_entrypoint" -gguf-set-metadata = "gguf.scripts:gguf_set_metadata_entrypoint" -gguf-new-metadata = "gguf.scripts:gguf_new_metadata_entrypoint" -gguf-editor-gui = "gguf.scripts:gguf_editor_gui_entrypoint" +gguf-convert-endian = "gguf.scripts.gguf_convert_endian:main" +gguf-dump = "gguf.scripts.gguf_dump:main" +gguf-set-metadata = "gguf.scripts.gguf_set_metadata:main" +gguf-new-metadata = "gguf.scripts.gguf_new_metadata:main" +gguf-editor-gui = "gguf.scripts.gguf_editor_gui:main" diff --git a/grammars/README.md b/grammars/README.md index 5aa12acc1b..a63198b5ae 100644 --- a/grammars/README.md +++ b/grammars/README.md @@ -1,6 +1,6 @@ # GBNF Guide -GBNF (GGML BNF) is a format for defining [formal grammars](https://en.wikipedia.org/wiki/Formal_grammar) to constrain model outputs in `llama.cpp`. For example, you can use it to force the model to generate valid JSON, or speak only in emojis. GBNF grammars are supported in various ways in `examples/main` and `examples/server`. +GBNF (GGML BNF) is a format for defining [formal grammars](https://en.wikipedia.org/wiki/Formal_grammar) to constrain model outputs in `llama.cpp`. For example, you can use it to force the model to generate valid JSON, or speak only in emojis. GBNF grammars are supported in various ways in `tools/main` and `tools/server`. ## Background @@ -110,21 +110,21 @@ While semantically correct, the syntax `x? x? x?.... x?` (with N repetitions) ma You can use GBNF grammars: -- In [llama-server](../examples/server)'s completion endpoints, passed as the `grammar` body field -- In [llama-cli](../examples/main), passed as the `--grammar` & `--grammar-file` flags +- In [llama-server](../tools/server)'s completion endpoints, passed as the `grammar` body field +- In [llama-cli](../tools/main), passed as the `--grammar` & `--grammar-file` flags - With [test-gbnf-validator](../tests/test-gbnf-validator.cpp), to test them against strings. ## JSON Schemas → GBNF `llama.cpp` supports converting a subset of https://json-schema.org/ to GBNF grammars: -- In [llama-server](../examples/server): +- In [llama-server](../tools/server): - For any completion endpoints, passed as the `json_schema` body field - For the `/chat/completions` endpoint, passed inside the `response_format` body field (e.g. `{"type", "json_object", "schema": {"items": {}}}` or `{ type: "json_schema", json_schema: {"schema": ...} }`) -- In [llama-cli](../examples/main), passed as the `--json` / `-j` flag +- In [llama-cli](../tools/main), passed as the `--json` / `-j` flag - To convert to a grammar ahead of time: - in CLI, with [examples/json_schema_to_grammar.py](../examples/json_schema_to_grammar.py) - - in JavaScript with [json-schema-to-grammar.mjs](../examples/server/public_legacy/json-schema-to-grammar.mjs) (this is used by the [server](../examples/server)'s Web UI) + - in JavaScript with [json-schema-to-grammar.mjs](../tools/server/public_legacy/json-schema-to-grammar.mjs) (this is used by the [server](../tools/server)'s Web UI) Take a look at [tests](../tests/test-json-schema-to-grammar.cpp) to see which features are likely supported (you'll also find usage examples in https://github.com/ggml-org/llama.cpp/pull/5978, https://github.com/ggml-org/llama.cpp/pull/6659 & https://github.com/ggml-org/llama.cpp/pull/6555). diff --git a/include/llama.h b/include/llama.h index a13350e15b..abedebdb78 100644 --- a/include/llama.h +++ b/include/llama.h @@ -4,6 +4,7 @@ #include "ggml.h" #include "ggml-cpu.h" #include "ggml-backend.h" +#include "ggml-opt.h" #include #include @@ -112,6 +113,7 @@ extern "C" { LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32, LLAMA_VOCAB_PRE_TYPE_LLAMA4 = 33, LLAMA_VOCAB_PRE_TYPE_PIXTRAL = 34, + LLAMA_VOCAB_PRE_TYPE_SEED_CODER = 35, }; enum llama_rope_type { @@ -351,19 +353,18 @@ extern "C" { enum ggml_type type_k; // data type for K cache [EXPERIMENTAL] enum ggml_type type_v; // data type for V cache [EXPERIMENTAL] - // Keep the booleans together and at the end of the struct to avoid misalignment during copy-by-value. - // TODO: move at the end of the struct - bool logits_all; // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead) - bool embeddings; // if true, extract embeddings (together with logits) - bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU - bool flash_attn; // whether to use flash attention [EXPERIMENTAL] - bool no_perf; // whether to measure performance timings - // Abort callback // if it returns true, execution of llama_decode() will be aborted // currently works only with CPU execution ggml_abort_callback abort_callback; void * abort_callback_data; + + // Keep the booleans together and at the end of the struct to avoid misalignment during copy-by-value. + bool embeddings; // if true, extract embeddings (together with logits) + bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU + bool flash_attn; // whether to use flash attention [EXPERIMENTAL] + bool no_perf; // whether to measure performance timings + bool op_offload; // whether to offload host tensor operations to device }; // model quantization parameters @@ -445,6 +446,10 @@ extern "C" { size_t n_paths, struct llama_model_params params); + LLAMA_API void llama_model_save_to_file( + const struct llama_model * model, + const char * path_model); + DEPRECATED(LLAMA_API void llama_free_model(struct llama_model * model), "use llama_model_free instead"); @@ -924,14 +929,19 @@ extern "C" { // Frees a batch of tokens allocated with llama_batch_init() LLAMA_API void llama_batch_free(struct llama_batch batch); - // Processes a batch of tokens with the ecoder part of the encoder-decoder model. - // Stores the encoder output internally for later use by the decoder cross-attention layers. + // Process a batch of tokens. + // In contrast to llama_decode() - this call does not use KV cache. + // For encode-decoder contexts, processes the batch using the encoder. + // Can store the encoder output internally for later use by the decoder's cross-attention layers. // 0 - success // < 0 - error. the KV cache state is restored to the state before this call LLAMA_API int32_t llama_encode( struct llama_context * ctx, struct llama_batch batch); + // Process a batch of tokens. + // Requires KV cache. + // For encode-decoder contexts, processes the batch using the decoder. // Positive return values does not mean a fatal error, but rather a warning. // 0 - success // 1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context) @@ -1232,6 +1242,7 @@ extern "C" { "will be removed in the future (see https://github.com/ggml-org/llama.cpp/pull/9896#discussion_r1800920915)"); /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 + /// Setting k <= 0 makes this a noop LLAMA_API struct llama_sampler * llama_sampler_init_top_k (int32_t k); /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 @@ -1427,6 +1438,37 @@ extern "C" { LLAMA_API void llama_perf_sampler_print(const struct llama_sampler * chain); LLAMA_API void llama_perf_sampler_reset( struct llama_sampler * chain); + // + // training + // + + // function that returns whether or not a given tensor contains trainable parameters + typedef bool (*llama_opt_param_filter)(const struct ggml_tensor * tensor, void * userdata); + + // always returns true + LLAMA_API bool llama_opt_param_filter_all(const struct ggml_tensor * tensor, void * userdata); + + struct llama_opt_params { + uint32_t n_ctx_train; // assumed context size post training, use context size specified in llama_context if 0 + + llama_opt_param_filter param_filter; // callback for determining which tensors contain trainable parameters + void * param_filter_ud; // userdata for determining which tensors contain trainable parameters + + ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters + void * get_opt_pars_ud; // userdata for calculating optimizer parameters + }; + + LLAMA_API void llama_opt_init(struct llama_context * lctx, struct llama_model * model, struct llama_opt_params lopt_params); + + LLAMA_API void llama_opt_epoch( + struct llama_context * lctx, + ggml_opt_dataset_t dataset, + ggml_opt_result_t result_train, + ggml_opt_result_t result_eval, + int64_t idata_split, + ggml_opt_epoch_callback callback_train, + ggml_opt_epoch_callback callback_eval); + #ifdef __cplusplus } #endif diff --git a/pyproject.toml b/pyproject.toml index ed62264ba6..3d71b055a8 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -40,5 +40,6 @@ build-backend = "poetry.core.masonry.api" [tool.poetry.scripts] llama-convert-hf-to-gguf = "convert_hf_to_gguf:main" +llama-convert-lora-to-gguf = "convert_lora_to_gguf:main" llama-convert-llama-ggml-to-gguf = "convert_llama_ggml_to_gguf:main" llama-ggml-vk-generate-shaders = "ggml_vk_generate_shaders:main" diff --git a/pyrightconfig.json b/pyrightconfig.json index 9acbbeb78a..5320fe5864 100644 --- a/pyrightconfig.json +++ b/pyrightconfig.json @@ -15,7 +15,7 @@ }, { // uses match expressions in steps.py - "root": "examples/server/tests", + "root": "tools/server/tests", "pythonVersion": "3.10", }, ], diff --git a/requirements/requirements-all.txt b/requirements/requirements-all.txt index eba0a59f62..9fa7d4d0ab 100644 --- a/requirements/requirements-all.txt +++ b/requirements/requirements-all.txt @@ -1,6 +1,6 @@ --r ../examples/llava/requirements.txt --r ../examples/server/bench/requirements.txt --r ../examples/server/tests/requirements.txt +-r ../tools/mtmd/requirements.txt +-r ../tools/server/bench/requirements.txt +-r ../tools/server/tests/requirements.txt -r ./requirements-compare-llama-bench.txt -r ./requirements-pydantic.txt diff --git a/scripts/compare-llama-bench.py b/scripts/compare-llama-bench.py index 6205fe88d7..c32b449f7d 100755 --- a/scripts/compare-llama-bench.py +++ b/scripts/compare-llama-bench.py @@ -19,9 +19,9 @@ logger = logging.getLogger("compare-llama-bench") # Properties by which to differentiate results per commit: KEY_PROPERTIES = [ - "cpu_info", "gpu_info", "backends", "n_gpu_layers", "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" + "cpu_info", "gpu_info", "backends", "n_gpu_layers", "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" ] # Properties that are boolean and are converted to Yes/No for the table: @@ -30,11 +30,11 @@ BOOL_PROPERTIES = ["embeddings", "cpu_strict", "use_mmap", "no_kv_offload", "fla # Header names for the table: PRETTY_NAMES = { "cpu_info": "CPU", "gpu_info": "GPU", "backends": "Backends", "n_gpu_layers": "GPU layers", - "model_filename": "File", "model_type": "Model", "model_size": "Model size [GiB]", - "model_n_params": "Num. of par.", "n_batch": "Batch size", "n_ubatch": "Microbatch size", - "embeddings": "Embeddings", "cpu_mask": "CPU mask", "cpu_strict": "CPU strict", "poll": "Poll", - "n_threads": "Threads", "type_k": "K type", "type_v": "V type", "split_mode": "Split mode", "main_gpu": "Main GPU", - "no_kv_offload": "NKVO", "flash_attn": "FlashAttention", "tensor_split": "Tensor split", "use_mmap": "Use mmap", + "tensor_buft_overrides": "Tensor overrides", "model_filename": "File", "model_type": "Model", "model_size": "Model size [GiB]", + "model_n_params": "Num. of par.", "n_batch": "Batch size", "n_ubatch": "Microbatch size", "embeddings": "Embeddings", + "cpu_mask": "CPU mask", "cpu_strict": "CPU strict", "poll": "Poll", "n_threads": "Threads", "type_k": "K type", "type_v": "V type", + "use_mmap": "Use mmap", "no_kv_offload": "NKVO", "split_mode": "Split mode", "main_gpu": "Main GPU", "tensor_split": "Tensor split", + "flash_attn": "FlashAttention", } DEFAULT_SHOW = ["model_type"] # Always show these properties by default. @@ -281,12 +281,12 @@ def get_rows(properties): The returned rows are unique in terms of property combinations. """ select_string = ", ".join( - [f"tb.{p}" for p in properties] + ["tb.n_prompt", "tb.n_gen", "AVG(tb.avg_ts)", "AVG(tc.avg_ts)"]) + [f"tb.{p}" for p in properties] + ["tb.n_prompt", "tb.n_gen", "tb.n_depth", "AVG(tb.avg_ts)", "AVG(tc.avg_ts)"]) equal_string = " AND ".join( [f"tb.{p} = tc.{p}" for p in KEY_PROPERTIES] + [ f"tb.build_commit = '{hexsha8_baseline}'", f"tc.build_commit = '{hexsha8_compare}'"] ) - group_order_string = ", ".join([f"tb.{p}" for p in properties] + ["tb.n_gen", "tb.n_prompt"]) + group_order_string = ", ".join([f"tb.{p}" for p in properties] + ["tb.n_gen", "tb.n_prompt", "tb.n_depth"]) query = (f"SELECT {select_string} FROM test tb JOIN test tc ON {equal_string} " f"GROUP BY {group_order_string} ORDER BY {group_order_string};") return cursor.execute(query).fetchall() @@ -309,7 +309,7 @@ else: rows_full = get_rows(KEY_PROPERTIES) properties_different = [] for i, kp_i in enumerate(KEY_PROPERTIES): - if kp_i in DEFAULT_SHOW or kp_i == "n_prompt" or kp_i == "n_gen": + if kp_i in DEFAULT_SHOW or kp_i in ["n_prompt", "n_gen", "n_depth"]: continue for row_full in rows_full: if row_full[i] != rows_full[0][i]: @@ -318,7 +318,7 @@ else: show = [] # Show CPU and/or GPU by default even if the hardware for all results is the same: - if "n_gpu_layers" not in properties_different: + if rows_full and "n_gpu_layers" not in properties_different: ngl = int(rows_full[0][KEY_PROPERTIES.index("n_gpu_layers")]) if ngl != 99 and "cpu_info" not in properties_different: @@ -338,19 +338,26 @@ else: pass rows_show = get_rows(show) +if not rows_show: + logger.error(f"No comparable data was found between {name_baseline} and {name_compare}.\n") + sys.exit(1) + table = [] for row in rows_show: - n_prompt = int(row[-4]) - n_gen = int(row[-3]) + n_prompt = int(row[-5]) + n_gen = int(row[-4]) + n_depth = int(row[-3]) if n_prompt != 0 and n_gen == 0: test_name = f"pp{n_prompt}" elif n_prompt == 0 and n_gen != 0: test_name = f"tg{n_gen}" else: test_name = f"pp{n_prompt}+tg{n_gen}" + if n_depth != 0: + test_name = f"{test_name}@d{n_depth}" # Regular columns test name avg t/s values Speedup # VVVVVVVVVVVVV VVVVVVVVV VVVVVVVVVVVVVV VVVVVVV - table.append(list(row[:-4]) + [test_name] + list(row[-2:]) + [float(row[-1]) / float(row[-2])]) + table.append(list(row[:-5]) + [test_name] + list(row[-2:]) + [float(row[-1]) / float(row[-2])]) # Some a-posteriori fixes to make the table contents prettier: for bool_property in BOOL_PROPERTIES: @@ -376,7 +383,7 @@ if "gpu_info" in show: for gns in GPU_NAME_STRIP: row_table[ip] = row_table[ip].replace(gns, "") - gpu_names = row_table[ip].split("/") + gpu_names = row_table[ip].split(", ") num_gpus = len(gpu_names) all_names_the_same = len(set(gpu_names)) == 1 if len(gpu_names) >= 2 and all_names_the_same: diff --git a/scripts/fetch_server_test_models.py b/scripts/fetch_server_test_models.py index e6775bfc58..ac483ef5d7 100755 --- a/scripts/fetch_server_test_models.py +++ b/scripts/fetch_server_test_models.py @@ -8,7 +8,7 @@ Example: python scripts/fetch_server_test_models.py - ( cd examples/server/tests && ./tests.sh -v -x -m slow ) + ( cd tools/server/tests && ./tests.sh -v -x -m slow ) ''' import ast import glob @@ -66,7 +66,7 @@ if __name__ == '__main__': models = sorted(list(set([ model - for test_file in glob.glob('examples/server/tests/unit/test_*.py') + for test_file in glob.glob('tools/server/tests/unit/test_*.py') for model in collect_hf_model_test_parameters(test_file) ])), key=lambda m: (m.hf_repo, m.hf_file)) diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last index 41feffca92..1f7c650c25 100644 --- a/scripts/sync-ggml.last +++ b/scripts/sync-ggml.last @@ -1 +1 @@ -13bcf9ce50651a8b4238ec6d136f46f2c1b23b6f +b59bddafe278877dfa22a80e53a637513862babb diff --git a/scripts/tool_bench.py b/scripts/tool_bench.py index 0f406bc42a..a2f2a2eb02 100755 --- a/scripts/tool_bench.py +++ b/scripts/tool_bench.py @@ -2,7 +2,7 @@ ''' Simplistic tool call benchmarks for llama-server and ollama. - Essentially runs the tests at server/examples/server/tests/unit/test_tool_call.py N times, at different temperatures and on different backends (current llama-server, baseline llama-server and ollama), + Essentially runs the tests at server/tools/server/tests/unit/test_tool_call.py N times, at different temperatures and on different backends (current llama-server, baseline llama-server and ollama), and plots the results of multiple runs (from same .jsonl file or multiple ones) as a success rate heatmap. Simple usage example: @@ -51,8 +51,8 @@ import typer sys.path.insert(0, Path(__file__).parent.parent.as_posix()) if True: - from examples.server.tests.utils import ServerProcess - from examples.server.tests.unit.test_tool_call import TIMEOUT_SERVER_START, do_test_calc_result, do_test_hello_world, do_test_weather + from tools.server.tests.utils import ServerProcess + from tools.server.tests.unit.test_tool_call import TIMEOUT_SERVER_START, do_test_calc_result, do_test_hello_world, do_test_weather @contextmanager diff --git a/scripts/xxd.cmake b/scripts/xxd.cmake index f5ad6ab9b1..14d2753808 100644 --- a/scripts/xxd.cmake +++ b/scripts/xxd.cmake @@ -1,5 +1,5 @@ # CMake equivalent of `xxd -i ${INPUT} ${OUTPUT}` -# Usage: cmake -DINPUT=examples/server/public/index.html -DOUTPUT=examples/server/index.html.hpp -P scripts/xxd.cmake +# Usage: cmake -DINPUT=tools/server/public/index.html -DOUTPUT=tools/server/index.html.hpp -P scripts/xxd.cmake SET(INPUT "" CACHE STRING "Input File") SET(OUTPUT "" CACHE STRING "Output File") diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt index 1cd316b03e..d4bf37b1cf 100644 --- a/src/CMakeLists.txt +++ b/src/CMakeLists.txt @@ -23,6 +23,7 @@ add_library(llama llama-memory.cpp llama-mmap.cpp llama-model-loader.cpp + llama-model-saver.cpp llama-model.cpp llama-quant.cpp llama-sampling.cpp diff --git a/src/llama-adapter.cpp b/src/llama-adapter.cpp index 7ac54d2391..8d94034aed 100644 --- a/src/llama-adapter.cpp +++ b/src/llama-adapter.cpp @@ -253,6 +253,9 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_ std::vector buft_extra; { auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); + if (!cpu_dev) { + throw std::runtime_error(format("%s: no CPU backend found", __func__)); + } auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev); auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t) @@ -291,6 +294,9 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_ LLAMA_LOG_WARN("%s: lora for '%s' cannot use buft '%s', fallback to CPU\n", __func__, model_tensor->name, ggml_backend_buft_name(buft)); auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); + if (!cpu_dev) { + throw std::runtime_error(format("%s: no CPU backend found", __func__)); + } buft = ggml_backend_dev_buffer_type(cpu_dev); break; diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index 62e1480bb5..f2bc8ca768 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -19,6 +19,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_REFACT, "refact" }, { LLM_ARCH_BERT, "bert" }, { LLM_ARCH_NOMIC_BERT, "nomic-bert" }, + { LLM_ARCH_NOMIC_BERT_MOE, "nomic-bert-moe" }, { LLM_ARCH_JINA_BERT_V2, "jina-bert-v2" }, { LLM_ARCH_BLOOM, "bloom" }, { LLM_ARCH_STABLELM, "stablelm" }, @@ -106,6 +107,7 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_EXPERT_WEIGHTS_SCALE, "%s.expert_weights_scale" }, { LLM_KV_EXPERT_WEIGHTS_NORM, "%s.expert_weights_norm" }, { LLM_KV_EXPERT_GATING_FUNC, "%s.expert_gating_func" }, + { LLM_KV_MOE_EVERY_N_LAYERS, "%s.moe_every_n_layers" }, { LLM_KV_POOLING_TYPE, "%s.pooling_type" }, { LLM_KV_LOGIT_SCALE, "%s.logit_scale" }, { LLM_KV_DECODER_START_TOKEN_ID, "%s.decoder_start_token_id" }, @@ -472,6 +474,24 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_NOMIC_BERT_MOE, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" }, + { LLM_TENSOR_TOKEN_TYPES, "token_types" }, + { LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_LAYER_OUT_NORM, "blk.%d.layer_output_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + }, + }, { LLM_ARCH_JINA_BERT_V2, { diff --git a/src/llama-arch.h b/src/llama-arch.h index 98ca00a1bd..41a023da3d 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -23,6 +23,7 @@ enum llm_arch { LLM_ARCH_REFACT, LLM_ARCH_BERT, LLM_ARCH_NOMIC_BERT, + LLM_ARCH_NOMIC_BERT_MOE, LLM_ARCH_JINA_BERT_V2, LLM_ARCH_BLOOM, LLM_ARCH_STABLELM, @@ -110,6 +111,7 @@ enum llm_kv { LLM_KV_EXPERT_WEIGHTS_SCALE, LLM_KV_EXPERT_WEIGHTS_NORM, LLM_KV_EXPERT_GATING_FUNC, + LLM_KV_MOE_EVERY_N_LAYERS, LLM_KV_POOLING_TYPE, LLM_KV_LOGIT_SCALE, LLM_KV_DECODER_START_TOKEN_ID, diff --git a/src/llama-batch.cpp b/src/llama-batch.cpp index 01d5ca57fd..a88b2fe308 100644 --- a/src/llama-batch.cpp +++ b/src/llama-batch.cpp @@ -189,7 +189,7 @@ llama_ubatch llama_sbatch::split_seq(size_t n_ubatch) { return ubatch; } -void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool simple_split, bool logits_all) { +llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split, bool logits_all) { GGML_ASSERT(batch.n_tokens >= 0); this->batch = &batch; this->n_embd = n_embd; @@ -203,6 +203,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim for (size_t i = 0; i < n_tokens; ++i) { ids[i] = i; } + if (simple_split) { seq.resize(1); llama_sbatch_seq & s = seq[0]; @@ -212,6 +213,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim s.length = n_tokens; return; } + std::sort(ids.begin(), ids.end(), [&batch](size_t a, size_t b) { int32_t n_seq_a = batch.n_seq_id ? batch.n_seq_id[a] : 1; @@ -239,6 +241,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim return n_seq_a > n_seq_b; } ); + // init seq llama_sbatch_seq * last_seq = nullptr; @@ -262,6 +265,7 @@ void llama_sbatch::from_batch(const llama_batch & batch, size_t n_embd, bool sim seq.push_back(new_seq); last_seq = &seq.back(); } + // keep shared prompts first at the end, then sort by length descending. std::sort(seq.begin(), seq.end(), [](llama_sbatch_seq & a, llama_sbatch_seq & b) { diff --git a/src/llama-batch.h b/src/llama-batch.h index f1df40d270..6305051b62 100644 --- a/src/llama-batch.h +++ b/src/llama-batch.h @@ -70,7 +70,8 @@ struct llama_sbatch { // sequence-wise split llama_ubatch split_seq(size_t n_ubatch); - void from_batch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false); + llama_sbatch() = default; + llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false); }; // temporary allocate memory for the input batch if needed diff --git a/src/llama-chat.cpp b/src/llama-chat.cpp index 41f89e3a9d..d12743e6b9 100644 --- a/src/llama-chat.cpp +++ b/src/llama-chat.cpp @@ -35,6 +35,7 @@ static const std::map LLM_CHAT_TEMPLATES = { { "mistral-v3", LLM_CHAT_TEMPLATE_MISTRAL_V3 }, { "mistral-v3-tekken", LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN }, { "mistral-v7", LLM_CHAT_TEMPLATE_MISTRAL_V7 }, + { "mistral-v7-tekken", LLM_CHAT_TEMPLATE_MISTRAL_V7_TEKKEN }, { "phi3", LLM_CHAT_TEMPLATE_PHI_3 }, { "phi4", LLM_CHAT_TEMPLATE_PHI_4 }, { "falcon3", LLM_CHAT_TEMPLATE_FALCON_3 }, @@ -50,8 +51,8 @@ static const std::map LLM_CHAT_TEMPLATES = { { "deepseek3", LLM_CHAT_TEMPLATE_DEEPSEEK_3 }, { "command-r", LLM_CHAT_TEMPLATE_COMMAND_R }, { "llama3", LLM_CHAT_TEMPLATE_LLAMA_3 }, - { "chatglm3", LLM_CHAT_TEMPLATE_CHATGML_3 }, - { "chatglm4", LLM_CHAT_TEMPLATE_CHATGML_4 }, + { "chatglm3", LLM_CHAT_TEMPLATE_CHATGLM_3 }, + { "chatglm4", LLM_CHAT_TEMPLATE_CHATGLM_4 }, { "glmedge", LLM_CHAT_TEMPLATE_GLMEDGE }, { "minicpm", LLM_CHAT_TEMPLATE_MINICPM }, { "exaone3", LLM_CHAT_TEMPLATE_EXAONE_3 }, @@ -122,6 +123,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) { } } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>")) { return LLM_CHAT_TEMPLATE_PHI_3; + } else if (tmpl_contains("[gMASK]")) { + return LLM_CHAT_TEMPLATE_CHATGLM_4; } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) { return tmpl_contains("") ? LLM_CHAT_TEMPLATE_FALCON_3 : LLM_CHAT_TEMPLATE_GLMEDGE; } else if (tmpl_contains("<|{{ item['role'] }}|>") && tmpl_contains("<|begin_of_image|>")) { @@ -154,9 +157,7 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) { return LLM_CHAT_TEMPLATE_LLAMA_3; } else if (tmpl_contains("[gMASK]sop")) { // chatglm3-6b - return LLM_CHAT_TEMPLATE_CHATGML_3; - } else if (tmpl_contains("[gMASK]")) { - return LLM_CHAT_TEMPLATE_CHATGML_4; + return LLM_CHAT_TEMPLATE_CHATGLM_3; } else if (tmpl_contains(LU8("<用户>"))) { // MiniCPM-3B-OpenHermes-2.5-v2-GGUF return LLM_CHAT_TEMPLATE_MINICPM; @@ -202,19 +203,20 @@ int32_t llm_chat_apply_template( if (add_ass) { ss << "<|im_start|>assistant\n"; } - } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7) { + } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7 || tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7_TEKKEN) { // Official mistral 'v7' template // See: https://huggingface.co/mistralai/Mistral-Large-Instruct-2411#basic-instruct-template-v7 + // https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503#basic-instruct-template-v7-tekken + const char * trailing_space = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7 ? " " : ""; for (auto message : chat) { std::string role(message->role); std::string content(message->content); if (role == "system") { - ss << "[SYSTEM_PROMPT] " << content << "[/SYSTEM_PROMPT]"; + ss << "[SYSTEM_PROMPT]" << trailing_space << content << "[/SYSTEM_PROMPT]"; } else if (role == "user") { - ss << "[INST] " << content << "[/INST]"; - } - else { - ss << " " << content << ""; + ss << "[INST]" << trailing_space << content << "[/INST]"; + } else { + ss << trailing_space << content << ""; } } } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V1 @@ -437,7 +439,7 @@ int32_t llm_chat_apply_template( if (add_ass) { ss << "<|start_header_id|>assistant<|end_header_id|>\n\n"; } - } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_3) { + } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_3) { // chatglm3-6b ss << "[gMASK]" << "sop"; for (auto message : chat) { @@ -447,14 +449,14 @@ int32_t llm_chat_apply_template( if (add_ass) { ss << "<|assistant|>"; } - } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_4) { + } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_4) { ss << "[gMASK]" << ""; for (auto message : chat) { std::string role(message->role); ss << "<|" << role << "|>" << "\n" << message->content; } if (add_ass) { - ss << "<|assistant|>"; + ss << "<|assistant|>\n"; } } else if (tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) { for (auto message : chat) { diff --git a/src/llama-chat.h b/src/llama-chat.h index dc30df711a..db24ade21e 100644 --- a/src/llama-chat.h +++ b/src/llama-chat.h @@ -14,6 +14,7 @@ enum llm_chat_template { LLM_CHAT_TEMPLATE_MISTRAL_V3, LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN, LLM_CHAT_TEMPLATE_MISTRAL_V7, + LLM_CHAT_TEMPLATE_MISTRAL_V7_TEKKEN, LLM_CHAT_TEMPLATE_PHI_3, LLM_CHAT_TEMPLATE_PHI_4, LLM_CHAT_TEMPLATE_FALCON_3, @@ -29,8 +30,8 @@ enum llm_chat_template { LLM_CHAT_TEMPLATE_DEEPSEEK_3, LLM_CHAT_TEMPLATE_COMMAND_R, LLM_CHAT_TEMPLATE_LLAMA_3, - LLM_CHAT_TEMPLATE_CHATGML_3, - LLM_CHAT_TEMPLATE_CHATGML_4, + LLM_CHAT_TEMPLATE_CHATGLM_3, + LLM_CHAT_TEMPLATE_CHATGLM_4, LLM_CHAT_TEMPLATE_GLMEDGE, LLM_CHAT_TEMPLATE_MINICPM, LLM_CHAT_TEMPLATE_EXAONE_3, diff --git a/src/llama-context.cpp b/src/llama-context.cpp index 983385f86d..62246c10da 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -6,11 +6,9 @@ #include "llama-model.h" #include "llama-kv-cache.h" -#include #include #include #include -#include // // llama_context @@ -95,6 +93,7 @@ llama_context::llama_context( } cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch); + cparams.op_offload = params.op_offload; const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max; @@ -114,12 +113,10 @@ llama_context::llama_context( } if (n_ctx_per_seq > hparams.n_ctx_train) { - LLAMA_LOG_WARN("%s: n_ctx_pre_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n", + LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n", __func__, n_ctx_per_seq, hparams.n_ctx_train); } - logits_all = params.logits_all; - if (!hparams.vocab_only) { // GPU backends for (auto * dev : model.devices) { @@ -177,44 +174,13 @@ llama_context::llama_context( } // init the memory module - // TODO: for now, always create a unified KV cache if (!hparams.vocab_only) { - kv_self.reset(static_cast(model.create_memory())); + llama_memory_params params_mem = { + /*.type_k =*/ params.type_k, + /*.type_v =*/ params.type_v, + }; - LLAMA_LOG_DEBUG("%s: n_ctx = %u\n", __func__, cparams.n_ctx); - - cparams.n_ctx = GGML_PAD(cparams.n_ctx, kv_self->get_padding(cparams)); - - LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx); - - uint32_t kv_size = cparams.n_ctx; - ggml_type type_k = params.type_k; - ggml_type type_v = params.type_v; - - if (llama_model_is_recurrent(&model)) { - // Mamba needs at least as many KV cells as there are sequences kept at any time - kv_size = std::max((uint32_t) 1, params.n_seq_max); - // it's probably best to keep as much precision as possible for the states - type_k = GGML_TYPE_F32; // required by ggml_ssm_conv for Mamba's conv_states - type_v = GGML_TYPE_F32; // required by ggml_ssm_scan for Mamba's ssm_states - } - - GGML_ASSERT(hparams.n_embd_head_k % ggml_blck_size(type_k) == 0); - GGML_ASSERT(hparams.n_embd_head_v % ggml_blck_size(type_v) == 0); - - if (!kv_self->init(model, cparams, type_k, type_v, kv_size, cparams.offload_kqv)) { - throw std::runtime_error("failed to initialize self-attention cache"); - } - - { - const size_t memory_size_k = kv_self->size_k_bytes(); - const size_t memory_size_v = kv_self->size_v_bytes(); - - LLAMA_LOG_INFO("%s: KV self size = %7.2f MiB, K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__, - (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f), - ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f), - ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f)); - } + memory.reset(model.create_memory(params_mem, cparams)); } // init backends @@ -278,7 +244,7 @@ llama_context::llama_context( } } - sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, pipeline_parallel)); + sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, pipeline_parallel, cparams.op_offload)); if (pipeline_parallel) { LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(sched.get())); @@ -286,7 +252,7 @@ llama_context::llama_context( } // reserve worst-case graph - if (!hparams.vocab_only) { + if (!hparams.vocab_only && memory) { const uint32_t n_seqs = 1; // TODO: worst-case number of sequences const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch); @@ -305,7 +271,9 @@ llama_context::llama_context( int n_nodes_tg = -1; // simulate full KV cache - kv_self->n = kv_self->size; + llama_kv_cache * kv_self = static_cast(memory.get()); + + kv_self->set_full(); cross.v_embd.clear(); @@ -391,7 +359,9 @@ llama_context::llama_context( } } -llama_context::~llama_context() = default; +llama_context::~llama_context() { + ggml_opt_free(opt_ctx); +} void llama_context::synchronize() { ggml_backend_sched_synchronize(sched.get()); @@ -427,6 +397,18 @@ const llama_model & llama_context::get_model() const { return model; } +const llama_cparams & llama_context::get_cparams() const { + return cparams; +} + +ggml_backend_sched_t llama_context::get_sched() const { + return sched.get(); +} + +ggml_context * llama_context::get_ctx_compute() const { + return ctx_compute.get(); +} + uint32_t llama_context::n_ctx() const { return cparams.n_ctx; } @@ -456,348 +438,21 @@ uint32_t llama_context::n_threads_batch() const { } llama_kv_cache * llama_context::get_kv_self() { - return kv_self.get(); + llama_kv_cache * kv_self = static_cast(memory.get()); + return kv_self; } const llama_kv_cache * llama_context::get_kv_self() const { - return kv_self.get(); -} - -ggml_tensor * llama_context::build_rope_shift( - ggml_context * ctx0, - ggml_tensor * cur, - ggml_tensor * shift, - ggml_tensor * factors, - float freq_base, - float freq_scale, - ggml_backend_buffer * bbuf) const { - const auto & n_ctx_orig = cparams.n_ctx_orig_yarn; - - const auto & yarn_ext_factor = cparams.yarn_ext_factor; - const auto & yarn_beta_fast = cparams.yarn_beta_fast; - const auto & yarn_beta_slow = cparams.yarn_beta_slow; - - const auto & hparams = model.hparams; - - const auto & n_rot = hparams.n_rot; - const auto & rope_type = hparams.rope_type; - - // See llm_build_deepseek2() for why attn_factor has to be scaled for YaRN RoPE to work correctly. - // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation. - const float yarn_attn_factor = model.arch == LLM_ARCH_DEEPSEEK2 ? 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale)) : cparams.yarn_attn_factor; - - ggml_tensor * tmp; - - if (ggml_is_quantized(cur->type)) { - // dequantize to f32 -> RoPE -> quantize back - tmp = ggml_cast(ctx0, cur, GGML_TYPE_F32); - - if (bbuf) { - for (const auto & backend : backends) { - // Figure out which backend KV cache belongs to - if (ggml_backend_supports_buft(backend.get(), ggml_backend_buffer_get_type(bbuf))) { - ggml_backend_sched_set_tensor_backend(sched.get(), tmp, backend.get()); - break; - } - } - } - - tmp = ggml_rope_ext_inplace(ctx0, tmp, - shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow); - - tmp = ggml_cpy(ctx0, tmp, cur); - } else { - // we rotate only the first n_rot dimensions - tmp = ggml_rope_ext_inplace(ctx0, cur, - shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow); - } - - return tmp; -} - -class llm_graph_input_k_shift : public llm_graph_input_i { -public: - llm_graph_input_k_shift(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {} - virtual ~llm_graph_input_k_shift() = default; - - void set_input(const llama_ubatch * ubatch) override; - - ggml_tensor * k_shift; // I32 [kv_size] - - const llama_kv_cache_unified * kv_self; -}; - -void llm_graph_input_k_shift::set_input(const llama_ubatch * ubatch) { - GGML_UNUSED(ubatch); - - if (k_shift) { - assert(ggml_backend_buffer_is_host(k_shift->buffer)); - - int32_t * data = (int32_t *) k_shift->data; - - for (uint32_t i = 0; i < kv_self->size; ++i) { - data[i] = kv_self->cells[i].delta; - } - } -} - -llm_graph_result_ptr llama_context::build_kv_self_shift( - ggml_context * ctx0, - ggml_cgraph * gf) const { - auto res = std::make_unique(); - - const auto & hparams = model.hparams; - - const auto & n_layer = hparams.n_layer; - - const auto & n_embd_head_k = hparams.n_embd_head_k; - //const auto & n_embd_head_v = hparams.n_embd_head_v; - - //GGML_ASSERT(kv_self->size == n_ctx); - - auto inp = std::make_unique(kv_self.get()); - - inp->k_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, cparams.n_ctx); - ggml_set_input(inp->k_shift); - - for (uint32_t il = 0; il < n_layer; ++il) { - const int64_t n_head_kv = hparams.n_head_kv(il); - const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); - - const bool is_swa = hparams.is_swa(il); - - // note: the swa rope params could become part of the cparams in the future - // if we decide to make them configurable, like the non-sliding ones - const float freq_base_l = is_swa ? hparams.rope_freq_base_train_swa : cparams.rope_freq_base; - const float freq_scale_l = is_swa ? hparams.rope_freq_scale_train_swa : cparams.rope_freq_scale; - - ggml_tensor * rope_factors = kv_self->cbs.get_rope_factors(n_ctx_per_seq(), il); - - ggml_tensor * k = - ggml_view_3d(ctx0, kv_self->k_l[il], - n_embd_head_k, n_head_kv, kv_self->size, - ggml_row_size(kv_self->k_l[il]->type, n_embd_head_k), - ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa), - 0); - - ggml_tensor * cur = build_rope_shift(ctx0, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l, kv_self->k_l[il]->buffer); - - ggml_build_forward_expand(gf, cur); - } - - res->add_input(std::move(inp)); - - return res; -} - -llm_graph_result_ptr llama_context::build_kv_self_defrag( - ggml_context * ctx0, - ggml_cgraph * gf) const { - auto res = std::make_unique(); - - const auto & hparams = model.hparams; - - const auto & ids = kv_self->defrag_info.ids; - -#if 0 - // CPU defrag - // - // TODO: optimizations are possible: - // - multiple threads - // - avoid copying to the host memory when already there - // - // likely not worth the effort, as we have ggml_graph based defrag - // - - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(); - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(); - - const uint32_t kv_size = size; - - std::vector buf_k; - std::vector buf_v; - - for (uint32_t il = 0; il < n_layer; ++il) { - const size_t k_size_row = ggml_row_size(k_l[il]->type, n_embd_k_gqa); - const size_t k_size = ggml_row_size(k_l[il]->type, n_embd_k_gqa*kv_size); - - const size_t v_size_el = ggml_type_size(v_l[il]->type); - const size_t v_size = ggml_row_size (v_l[il]->type, n_embd_v_gqa*kv_size); - - buf_k.resize(k_size); - buf_v.resize(v_size); - - ggml_backend_tensor_get(k_l[il], buf_k.data(), 0, buf_k.size()); - ggml_backend_tensor_get(v_l[il], buf_v.data(), 0, buf_v.size()); - - // batch move [i, i+nm) to [id, id+nm) - // note: cells can move only to a lower index - for (uint32_t i = 0; i < n_kv; ++i) { - const uint32_t id = ids[i]; - - if (i == id || id == n_kv) { - continue; - } - - uint32_t nm = 1; - - while (i + nm < n_kv && ids[i + nm] == id + nm) { - nm++; - } - - // move keys - { - const int64_t os = i*k_size_row; - const int64_t od = id*k_size_row; - - memcpy(buf_k.data() + od, buf_k.data() + os, nm*k_size_row); - } - - // move values (note: they are transposed) - { - const int64_t os = i; - const int64_t od = id; - - for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { - memcpy(buf_v.data() + (od + j*kv_size)*v_size_el, buf_v.data() + (os + j*kv_size)*v_size_el, nm*v_size_el); - } - } - - i += nm - 1; - } - - ggml_backend_tensor_set(k_l[il], buf_k.data(), 0, buf_k.size()); - ggml_backend_tensor_set(v_l[il], buf_v.data(), 0, buf_v.size()); - } -#else - for (uint32_t i = 0; i < ids.size(); ++i) { - const uint32_t id = ids[i]; - - if (i == id || id == ids.size()) { - continue; - } - - uint32_t nm = 1; - - while (i + nm < ids.size() && ids[i + nm] == id + nm) { - nm++; - } - - for (uint32_t il = 0; il < hparams.n_layer; ++il) { // NOLINT - const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); - const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); - - ggml_tensor * view_k_src = ggml_view_2d(ctx0, kv_self->k_l[il], - n_embd_k_gqa, nm, - ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa), - ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa*i)); - - ggml_tensor * view_k_dst = ggml_view_2d(ctx0, kv_self->k_l[il], - n_embd_k_gqa, nm, - ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa), - ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa*id)); - - ggml_tensor * view_v_src; - ggml_tensor * view_v_dst; - - if (cparams.flash_attn) { - // NOTE: the V cache is not transposed when using flash attention - view_v_src = ggml_view_2d(ctx0, kv_self->v_l[il], - n_embd_v_gqa, nm, - ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa), - ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa*i)); - - view_v_dst = ggml_view_2d(ctx0, kv_self->v_l[il], - n_embd_v_gqa, nm, - ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa), - ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa*id)); - } else { - view_v_src = ggml_view_2d(ctx0, kv_self->v_l[il], - nm, n_embd_v_gqa, - ggml_row_size(kv_self->v_l[il]->type, kv_self->size), - ggml_row_size(kv_self->v_l[il]->type, i)); - - view_v_dst = ggml_view_2d(ctx0, kv_self->v_l[il], - nm, n_embd_v_gqa, - ggml_row_size(kv_self->v_l[il]->type, kv_self->size), - ggml_row_size(kv_self->v_l[il]->type, id)); - } - - ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_k_src, view_k_dst)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_v_src, view_v_dst)); - } - - i += nm - 1; - } - - //LLAMA_LOG_INFO("gf->n_nodes = %d\n", gf->n_nodes); -#endif - - return res; + llama_kv_cache * kv_self = static_cast(memory.get()); + return kv_self; } void llama_context::kv_self_update() { - auto & kv = kv_self; - bool need_reserve = false; - if (kv->has_shift) { - if (!kv->get_can_shift()) { - GGML_ABORT("The current context does not support K-shift"); - } + llama_kv_cache * kv_self = static_cast(memory.get()); - LLAMA_LOG_DEBUG("%s: applying K-shift\n", __func__); - - // apply K-shift if needed - if (model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE) { - ggml_backend_sched_reset(sched.get()); - - auto * gf = graph_init(); - - auto res = build_kv_self_shift(ctx_compute.get(), gf); - - ggml_backend_sched_alloc_graph(sched.get(), gf); - - res->set_inputs(nullptr); - - graph_compute(gf, false); - - need_reserve = true; - } - - { - kv->has_shift = false; - - for (uint32_t i = 0; i < kv->size; ++i) { - kv->cells[i].delta = 0; - } - } - } - - // defragment the KV cache if needed - if (kv->do_defrag) { - LLAMA_LOG_DEBUG("%s: defragmenting KV cache\n", __func__); - - if (kv->defrag_prepare(graph_max_nodes())) { - ggml_backend_sched_reset(sched.get()); - - auto * gf = graph_init(); - - auto res = build_kv_self_defrag(ctx_compute.get(), gf); - - ggml_backend_sched_alloc_graph(sched.get(), gf); - - res->set_inputs(nullptr); - - graph_compute(gf, false); - - need_reserve = true; - } - - kv->do_defrag = false; - } + need_reserve = kv_self->update(*this); // reserve a worst case graph if needed if (need_reserve) { @@ -808,7 +463,7 @@ void llama_context::kv_self_update() { uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch); // simulate full KV cache - kv_self->n = kv_self->size; + kv_self->set_full(); llama_token token = model.vocab.token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph llama_ubatch ubatch = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr}; @@ -829,9 +484,6 @@ enum llama_pooling_type llama_context::pooling_type() const { } float * llama_context::get_logits() { - // reorder logits for backward compatibility - output_reorder(); - return logits; } @@ -874,9 +526,6 @@ float * llama_context::get_logits_ith(int32_t i) { } float * llama_context::get_embeddings() { - // reorder embeddings for backward compatibility - output_reorder(); - return embd; } @@ -1028,8 +677,8 @@ int llama_context::encode(llama_batch & inp_batch) { } // temporary allocate memory for the input batch if needed - // TODO: this is incorrect for multiple sequences because pos_max() is the maximum across all sequences - llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : kv_self->pos_max() + 1); + // note: during encode, we always pass the full sequence starting from pos = 0 + llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : 0); const llama_batch & batch = batch_allocr.batch; const int32_t n_tokens = batch.n_tokens; @@ -1054,11 +703,13 @@ int llama_context::encode(llama_batch & inp_batch) { t_compute_start_us = ggml_time_us(); } + embd_seq.clear(); + n_queued_tokens += n_tokens; const int64_t n_embd = hparams.n_embd; - sbatch.from_batch(batch, n_embd, /* simple_split */ true, /* logits_all */ true); + llama_sbatch sbatch = llama_sbatch(batch, n_embd, /* simple_split */ true, /* logits_all */ true); const llama_ubatch ubatch = sbatch.split_simple(n_tokens); @@ -1115,12 +766,12 @@ int llama_context::encode(llama_batch & inp_batch) { ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(sched.get(), t_embd); GGML_ASSERT(backend_embd != nullptr); - GGML_ASSERT(embd != nullptr); - switch (cparams.pooling_type) { case LLAMA_POOLING_TYPE_NONE: { // extract token embeddings + GGML_ASSERT(embd != nullptr); + GGML_ASSERT(n_tokens*n_embd <= (int64_t) embd_size); ggml_backend_tensor_get_async(backend_embd, t_embd, embd, 0, n_tokens*n_embd*sizeof(float)); } break; @@ -1145,11 +796,18 @@ int llama_context::encode(llama_batch & inp_batch) { } break; case LLAMA_POOLING_TYPE_RANK: { - // TODO: this likely should be the same logic as in llama_decoder_internal, but better to - // wait for an encoder model that requires this pooling type in order to test it - // https://github.com/ggerganov/llama.cpp/pull/9510 - GGML_ABORT("RANK pooling not implemented yet"); - } + // extract the rerank score - a single float per sequence + auto & embd_seq_out = embd_seq; + + for (uint32_t s = 0; s < ubatch.n_seqs; ++s) { + const llama_seq_id seq_id = ubatch.seq_id[s][0]; + if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { + continue; + } + embd_seq_out[seq_id].resize(1); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (seq_id)*sizeof(float), sizeof(float)); + } + } break; case LLAMA_POOLING_TYPE_UNSPECIFIED: { GGML_ABORT("unknown pooling type"); @@ -1187,14 +845,21 @@ int llama_context::encode(llama_batch & inp_batch) { } int llama_context::decode(llama_batch & inp_batch) { + if (!memory) { + LLAMA_LOG_WARN("%s: cannot decode batches with this context (use llama_encode() instead)\n", __func__); + return encode(inp_batch); + } + if (inp_batch.n_tokens == 0) { LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__); return -1; } + llama_kv_cache * kv_self = static_cast(memory.get()); + // temporary allocate memory for the input batch if needed - // TODO: this is incorrect for multiple sequences because pos_max() is the maximum across all sequences - llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : kv_self->pos_max() + 1); + // TODO: this is incorrect for multiple sequences because get_pos_max() is the maximum across all sequences + llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : kv_self->get_pos_max() + 1); const llama_batch & batch = batch_allocr.batch; @@ -1206,7 +871,7 @@ int llama_context::decode(llama_batch & inp_batch) { const int64_t n_tokens_all = batch.n_tokens; const int64_t n_embd = hparams.n_embd; - llama_kv_cache_guard kv_guard(kv_self.get()); + llama_kv_cache_guard kv_guard(kv_self); GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT @@ -1240,18 +905,14 @@ int llama_context::decode(llama_batch & inp_batch) { for (uint32_t i = 0; i < n_tokens_all; ++i) { n_outputs_all += batch.logits[i] != 0; } - } else if (logits_all || embd_pooled) { + } else if (embd_pooled) { n_outputs_all = n_tokens_all; } else { // keep last output only n_outputs_all = 1; } - const bool logits_all = n_outputs_all == n_tokens_all; - - sbatch.from_batch(batch, n_embd, - /* simple_split */ !kv_self->recurrent, - /* logits_all */ logits_all); + llama_sbatch sbatch = kv_self->sbatch_init(batch, /* logits_all */ n_outputs_all == n_tokens_all); // reserve output buffer if (output_reserve(n_outputs_all) < n_outputs_all) { @@ -1265,22 +926,7 @@ int llama_context::decode(llama_batch & inp_batch) { int64_t n_outputs_prev = 0; while (sbatch.n_tokens > 0) { - llama_ubatch ubatch = llama_ubatch(); - - const auto & n_ubatch = cparams.n_ubatch; - - if (kv_self->recurrent) { - if (embd_pooled) { - // Pooled embeddings cannot be split across ubatches (yet) - ubatch = sbatch.split_seq(cparams.n_ubatch); - } else { - // recurrent model architectures are easier to implement - // with equal-length sequences - ubatch = sbatch.split_equal(cparams.n_ubatch); - } - } else { - ubatch = sbatch.split_simple(n_ubatch); - } + llama_ubatch ubatch = kv_self->ubatch_next(sbatch, cparams.n_ubatch, embd_pooled); // count the outputs in this u_batch { @@ -1300,24 +946,12 @@ int llama_context::decode(llama_batch & inp_batch) { } // find KV slot - { - if (!kv_self->find_slot(ubatch)) { - LLAMA_LOG_WARN("%s: failed to find KV cache slot for ubatch of size %d\n", __func__, ubatch.n_tokens); + if (!kv_self->find_slot(ubatch)) { + LLAMA_LOG_WARN("%s: failed to find KV cache slot for ubatch of size %d\n", __func__, ubatch.n_tokens); - return 1; - } - - if (!kv_self->recurrent) { - // a heuristic, to avoid attending the full cache if it is not yet utilized - // after enough generations, the benefit from this heuristic disappears - // if we start defragmenting the cache, the benefit from this will be more important - const uint32_t pad = kv_self->get_padding(cparams); - kv_self->n = std::min(kv_self->size, std::max(pad, GGML_PAD(kv_self->cell_max(), pad))); - } + return 1; } - //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self->n, kv_self->used, kv_self->head); - ggml_backend_sched_reset(sched.get()); ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data); @@ -1431,43 +1065,68 @@ int llama_context::decode(llama_batch & inp_batch) { // finalize the batch processing kv_guard.commit(); + // set to total number of outputs in the batch, for use in llama_get_logits_ith + n_outputs = n_outputs_all; + // set output mappings { bool sorted_output = true; - GGML_ASSERT(sbatch.out_ids.size() == (size_t) n_outputs_all); + auto & out_ids = sbatch.out_ids; + + GGML_ASSERT(out_ids.size() == (size_t) n_outputs_all); for (int64_t i = 0; i < n_outputs_all; ++i) { - int64_t out_id = sbatch.out_ids[i]; + int64_t out_id = out_ids[i]; output_ids[out_id] = i; if (out_id != i) { sorted_output = false; } } - if (sorted_output) { - sbatch.out_ids.clear(); + // make the outputs have the same order they had in the user-provided batch + // note: this is mostly relevant for recurrent models atm + if (!sorted_output) { + const uint32_t n_vocab = model.vocab.n_tokens(); + const uint32_t n_embd = model.hparams.n_embd; + + GGML_ASSERT((size_t) n_outputs == out_ids.size()); + + // TODO: is there something more efficient which also minimizes swaps? + // selection sort, to minimize swaps (from https://en.wikipedia.org/wiki/Selection_sort) + for (int32_t i = 0; i < n_outputs - 1; ++i) { + int32_t j_min = i; + for (int32_t j = i + 1; j < n_outputs; ++j) { + if (out_ids[j] < out_ids[j_min]) { + j_min = j; + } + } + if (j_min == i) { continue; } + std::swap(out_ids[i], out_ids[j_min]); + if (logits_size > 0) { + for (uint32_t k = 0; k < n_vocab; k++) { + std::swap(logits[i*n_vocab + k], logits[j_min*n_vocab + k]); + } + } + if (embd_size > 0) { + for (uint32_t k = 0; k < n_embd; k++) { + std::swap(embd[i*n_embd + k], embd[j_min*n_embd + k]); + } + } + } + std::fill(output_ids.begin(), output_ids.end(), -1); + for (int32_t i = 0; i < n_outputs; ++i) { + output_ids[out_ids[i]] = i; + } } } - // set to total number of outputs in the batch, for use in llama_get_logits_ith - n_outputs = n_outputs_all; - // wait for the computation to finish (automatically done when obtaining the model output) //synchronize(); // decide if we need to defrag the kv cache - if (cparams.causal_attn && cparams.defrag_thold > 0.0f) { - // - do not defrag small contexts (i.e. < 2048 tokens) - // - count the padding towards the number of used tokens - const float fragmentation = kv_self->n >= 2048 ? std::max(0.0f, 1.0f - float(kv_self->used + kv_self->get_padding(cparams))/float(kv_self->n)) : 0.0f; - - // queue defragmentation for next llama_kv_cache_update - if (fragmentation > cparams.defrag_thold) { - LLAMA_LOG_DEBUG("%s: fragmentation: %.2f - requesting defrag\n", __func__, fragmentation); - - kv_self->defrag(); - } + if (cparams.defrag_thold > 0.0f) { + kv_self->defrag_sched(cparams.defrag_thold); } // Reset state for the next token before backend sync, to allow the CPU activities in the reset to @@ -1547,52 +1206,12 @@ int32_t llama_context::output_reserve(int32_t n_outputs) { // set all ids as invalid (negative) std::fill(output_ids.begin(), output_ids.end(), -1); - ggml_backend_buffer_clear(buf_output.get(), 0); - this->n_outputs = 0; this->n_outputs_max = n_outputs_max; return n_outputs_max; } -void llama_context::output_reorder() { - auto & out_ids = sbatch.out_ids; - if (!out_ids.empty()) { - const uint32_t n_vocab = model.vocab.n_tokens(); - const uint32_t n_embd = model.hparams.n_embd; - - GGML_ASSERT((size_t) n_outputs == out_ids.size()); - - // TODO: is there something more efficient which also minimizes swaps? - // selection sort, to minimize swaps (from https://en.wikipedia.org/wiki/Selection_sort) - for (int32_t i = 0; i < n_outputs - 1; ++i) { - int32_t j_min = i; - for (int32_t j = i + 1; j < n_outputs; ++j) { - if (out_ids[j] < out_ids[j_min]) { - j_min = j; - } - } - if (j_min == i) { continue; } - std::swap(out_ids[i], out_ids[j_min]); - if (logits_size > 0) { - for (uint32_t k = 0; k < n_vocab; k++) { - std::swap(logits[i*n_vocab + k], logits[j_min*n_vocab + k]); - } - } - if (embd_size > 0) { - for (uint32_t k = 0; k < n_embd; k++) { - std::swap(embd[i*n_embd + k], embd[j_min*n_embd + k]); - } - } - } - std::fill(output_ids.begin(), output_ids.end(), -1); - for (int32_t i = 0; i < n_outputs; ++i) { - output_ids[out_ids[i]] = i; - } - out_ids.clear(); - } -} - // // graph // @@ -1629,7 +1248,7 @@ llm_graph_result_ptr llama_context::graph_build( /*.backend_cpu =*/ backend_cpu, /*.cvec =*/ &cvec, /*.loras =*/ &loras, - /*.memory =*/ kv_self.get(), + /*.memory =*/ memory.get(), /*.cross =*/ &cross, /*.n_outputs =*/ n_outputs, /*.cb =*/ graph_get_cb(), @@ -2033,8 +1652,6 @@ size_t llama_context::state_write_data(llama_io_write_i & io) { { LLAMA_LOG_DEBUG("%s: - writing output ids\n", __func__); - output_reorder(); - const auto n_outputs = this->n_outputs; const auto & output_ids = this->output_ids; @@ -2088,6 +1705,8 @@ size_t llama_context::state_write_data(llama_io_write_i & io) { } LLAMA_LOG_DEBUG("%s: - writing KV self\n", __func__); + llama_kv_cache * kv_self = static_cast(memory.get()); + kv_self->state_write(io); return io.n_bytes(); @@ -2171,8 +1790,13 @@ size_t llama_context::state_read_data(llama_io_read_i & io) { } } - LLAMA_LOG_DEBUG("%s: - reading KV self\n", __func__); - kv_self->state_read(io); + if (memory) { + LLAMA_LOG_DEBUG("%s: - reading KV self\n", __func__); + + llama_kv_cache * kv_self = static_cast(memory.get()); + + kv_self->state_read(io); + } return io.n_bytes(); } @@ -2180,7 +1804,11 @@ size_t llama_context::state_read_data(llama_io_read_i & io) { size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id) { GGML_UNUSED(seq_id); - kv_self->state_write(io, seq_id); + if (memory) { + llama_kv_cache * kv_self = static_cast(memory.get()); + + kv_self->state_write(io, seq_id); + } return io.n_bytes(); } @@ -2188,7 +1816,11 @@ size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id s size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id) { GGML_UNUSED(seq_id); - kv_self->state_read(io, seq_id); + if (memory) { + llama_kv_cache * kv_self = static_cast(memory.get()); + + kv_self->state_read(io, seq_id); + } return io.n_bytes(); } @@ -2216,6 +1848,215 @@ void llama_context::perf_reset() { t_p_eval_us = n_p_eval = 0; } +// +// training +// + +static void llama_set_param(struct ggml_tensor * tensor, llama_opt_param_filter param_filter, void * userdata) { + if (!tensor || tensor->type != GGML_TYPE_F32) { + return; + } + if (!param_filter(tensor, userdata)) { + return; + } + if (strcmp(tensor->name, "token_embd.weight") == 0) { + return; // FIXME + } + if (strcmp(tensor->name, "rope_freqs.weight") == 0) { + return; // FIXME + } + ggml_set_param(tensor); +} + +void llama_context::opt_init(struct llama_model * model, struct llama_opt_params lopt_params) { + GGML_ASSERT(!opt_ctx); + model->hparams.n_ctx_train = lopt_params.n_ctx_train > 0 ? lopt_params.n_ctx_train : n_ctx(); + const uint32_t n_batch = std::min(this->n_batch(), model->hparams.n_ctx_train); + const uint32_t n_ubatch = std::min(this->n_ubatch(), n_batch); + GGML_ASSERT(model->hparams.n_ctx_train % n_batch == 0); + GGML_ASSERT(n_batch % n_ubatch == 0); + + ggml_opt_params opt_params = ggml_opt_default_params(sched.get(), GGML_OPT_LOSS_TYPE_CROSS_ENTROPY); + opt_params.opt_period = n_batch / n_ubatch; + opt_params.get_opt_pars = lopt_params.get_opt_pars; + opt_params.get_opt_pars_ud = lopt_params.get_opt_pars_ud; + + opt_ctx = ggml_opt_init(opt_params); + + llama_opt_param_filter param_filter = lopt_params.param_filter; + void * param_filter_ud = lopt_params.param_filter_ud; + + //llama_set_param(model->tok_embd, param_filter, param_filter_ud); // FIXME + llama_set_param(model->type_embd, param_filter, param_filter_ud); + llama_set_param(model->pos_embd, param_filter, param_filter_ud); + llama_set_param(model->tok_norm, param_filter, param_filter_ud); + llama_set_param(model->tok_norm_b, param_filter, param_filter_ud); + llama_set_param(model->output_norm, param_filter, param_filter_ud); + llama_set_param(model->output_norm_b, param_filter, param_filter_ud); + llama_set_param(model->output, param_filter, param_filter_ud); + llama_set_param(model->output_b, param_filter, param_filter_ud); + llama_set_param(model->output_norm_enc, param_filter, param_filter_ud); + llama_set_param(model->cls, param_filter, param_filter_ud); + llama_set_param(model->cls_b, param_filter, param_filter_ud); + llama_set_param(model->cls_out, param_filter, param_filter_ud); + llama_set_param(model->cls_out_b, param_filter, param_filter_ud); + + for (struct llama_layer & layer : model->layers) { + for (size_t i = 0; i < sizeof(layer)/sizeof(struct ggml_tensor *); ++i) { + llama_set_param(reinterpret_cast(&layer)[i], param_filter, param_filter_ud); + } + } +} + +void llama_context::opt_epoch_iter( + ggml_opt_dataset_t dataset, + ggml_opt_result_t result, + const std::vector & tokens, + const std::vector & labels_sparse, + llama_batch & batch, + ggml_opt_epoch_callback callback, + bool train, + int64_t idata_in_loop, + int64_t ndata_in_loop, + int64_t t_loop_start) { + GGML_ASSERT(opt_ctx); + const uint32_t n_ctx = llama_model_n_ctx_train(&model); + const uint32_t n_batch = std::min(this->n_batch(), n_ctx); + const uint32_t n_ubatch = std::min(this->n_ubatch(), n_batch); + + llama_kv_cache * kv_self = static_cast(memory.get()); + + kv_self->clear(); + llama_kv_cache_guard kv_guard(kv_self); + + for (uint32_t pos_ctx = 0; pos_ctx < n_ctx; pos_ctx += n_batch) { + batch.n_tokens = n_batch; + for (uint32_t pos_batch = 0; pos_batch < n_batch; ++pos_batch) { + batch.token [pos_batch] = tokens[pos_ctx + pos_batch]; + batch.pos [pos_batch] = pos_ctx + pos_batch; + batch.n_seq_id[pos_batch] = 1; + batch.seq_id [pos_batch][0] = 0; + batch.logits [pos_batch] = true; + } + + const auto n_tokens_all = batch.n_tokens; + + n_queued_tokens += n_tokens_all; + + // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens + const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE; + + embd_seq.clear(); + + int64_t n_outputs_all = n_tokens_all; + + llama_sbatch sbatch = kv_self->sbatch_init(batch, /*logits_all =*/ true); + + // reserve output buffer + if (output_reserve(n_outputs_all) < n_outputs_all) { + LLAMA_LOG_ERROR("%s: could not reserve space for batch with %" PRId64 " outputs\n", __func__, n_outputs_all); + GGML_ABORT("TODO: handle this error"); + }; + + for (uint32_t pos_batch = 0; pos_batch < n_batch; pos_batch += n_ubatch) { + llama_ubatch ubatch = kv_self->ubatch_next(sbatch, cparams.n_ubatch, embd_pooled); + + n_outputs = ubatch.n_tokens; + + // TODO: not sure if this is needed + if (!kv_self->find_slot(ubatch)) { + LLAMA_LOG_WARN("%s: failed to find KV cache slot for ubatch of size %d\n", __func__, ubatch.n_tokens); + + GGML_ABORT("TODO: handle this error"); + } + + auto * gf = graph_init(); + auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT); + + struct ggml_context * ctx_compute_opt; + { + const size_t size_gf = ggml_graph_size(gf); + const size_t size_meta = 4*size_gf*ggml_tensor_overhead() + 2*ggml_graph_overhead_custom(size_gf, /*grads = */ true); + struct ggml_init_params params = { + /*.mem_size =*/ size_meta, + /*.mem_buffer =*/ nullptr, + /*.no_alloc =*/ true, + }; + ctx_compute_opt = ggml_init(params); + } + ggml_opt_prepare_alloc(opt_ctx, ctx_compute_opt, gf, res->get_tokens(), res->get_logits()); + ggml_opt_alloc(opt_ctx, train); + res->set_inputs(&ubatch); + { + struct ggml_tensor * labels = ggml_opt_labels(opt_ctx); + GGML_ASSERT(labels->ne[1] == n_ubatch); + ggml_set_zero(labels); + const float onef = 1.0f; + for (uint32_t pos_ubatch = 0; pos_ubatch < n_ubatch; ++pos_ubatch) { + const uint32_t ilabel = pos_ctx + pos_batch + pos_ubatch; + GGML_ASSERT(labels_sparse[ilabel] < labels->ne[0]); + ggml_backend_tensor_set(labels, &onef, (pos_ubatch*labels->ne[0] + labels_sparse[ilabel])*sizeof(float), sizeof(float)); + } + } + ggml_opt_eval(opt_ctx, result); + if (callback) { + callback(train, opt_ctx, dataset, result, idata_in_loop + (pos_ctx + pos_batch)/n_ubatch + 1, ndata_in_loop, t_loop_start); + } + ggml_free(ctx_compute_opt); + } + } + + kv_guard.commit(); +} + +void llama_context::opt_epoch( + ggml_opt_dataset_t dataset, + ggml_opt_result_t result_train, + ggml_opt_result_t result_eval, + int64_t idata_split, + ggml_opt_epoch_callback callback_train, + ggml_opt_epoch_callback callback_eval) { + const uint32_t n_ctx = this->n_ctx(); + const uint32_t n_batch = std::min(cparams.n_batch, n_ctx); + const uint32_t n_ubatch = std::min(cparams.n_ubatch, n_batch); + const int64_t ndata = ggml_opt_dataset_ndata(dataset); + + GGML_ASSERT(idata_split >= 0); + GGML_ASSERT(idata_split <= ndata); + + const uint32_t ubatch_per_ctx = n_ctx / n_ubatch; + + struct llama_batch batch = llama_batch_init(n_batch, 0, 1); + std::vector tokens(n_ctx); + std::vector labels_sparse(n_ctx); + + int64_t idata = 0; + + int64_t t_loop_start = ggml_time_us(); + int64_t ndata_in_loop = idata_split*ubatch_per_ctx; + for (; idata < idata_split; ++idata) { + constexpr bool train = true; + const int64_t idata_in_loop = idata*ubatch_per_ctx; + + ggml_opt_dataset_get_batch_host(dataset, tokens.data(), n_ctx*sizeof(llama_token), labels_sparse.data(), idata); + opt_epoch_iter(dataset, result_train, tokens, labels_sparse, batch, + callback_train, train, idata_in_loop, ndata_in_loop, t_loop_start); + } + + t_loop_start = ggml_time_us(); + ndata_in_loop = (ndata - idata_split)*ubatch_per_ctx; + for (; idata < ndata; ++idata) { + constexpr bool train = false; + const int64_t idata_in_loop = (idata - idata_split)*ubatch_per_ctx; + + ggml_opt_dataset_get_batch_host(dataset, tokens.data(), n_ctx*sizeof(llama_token), labels_sparse.data(), idata); + opt_epoch_iter(dataset, result_eval, tokens, labels_sparse, batch, + callback_eval, train, idata_in_loop, ndata_in_loop, t_loop_start); + } + + llama_batch_free(batch); +} + // // interface implementation // @@ -2243,13 +2084,13 @@ llama_context_params llama_context_default_params() { /*.cb_eval_user_data =*/ nullptr, /*.type_k =*/ GGML_TYPE_F16, /*.type_v =*/ GGML_TYPE_F16, - /*.logits_all =*/ false, + /*.abort_callback =*/ nullptr, + /*.abort_callback_data =*/ nullptr, /*.embeddings =*/ false, /*.offload_kqv =*/ true, /*.flash_attn =*/ false, /*.no_perf =*/ true, - /*.abort_callback =*/ nullptr, - /*.abort_callback_data =*/ nullptr, + /*.op_offload =*/ true, }; return result; @@ -2543,7 +2384,7 @@ void llama_kv_cache_seq_cp( llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) { - return llama_kv_self_seq_cp(ctx, seq_id_src, seq_id_dst, p0, p1); + llama_kv_self_seq_cp(ctx, seq_id_src, seq_id_dst, p0, p1); } void llama_kv_self_seq_cp( @@ -2557,14 +2398,14 @@ void llama_kv_self_seq_cp( return; } - return kv->seq_cp(seq_id_src, seq_id_dst, p0, p1); + kv->seq_cp(seq_id_src, seq_id_dst, p0, p1); } // deprecated void llama_kv_cache_seq_keep( llama_context * ctx, llama_seq_id seq_id) { - return llama_kv_self_seq_keep(ctx, seq_id); + llama_kv_self_seq_keep(ctx, seq_id); } void llama_kv_self_seq_keep(llama_context * ctx, llama_seq_id seq_id) { @@ -2573,7 +2414,7 @@ void llama_kv_self_seq_keep(llama_context * ctx, llama_seq_id seq_id) { return; } - return kv->seq_keep(seq_id); + kv->seq_keep(seq_id); } // deprecated @@ -2583,7 +2424,7 @@ void llama_kv_cache_seq_add( llama_pos p0, llama_pos p1, llama_pos delta) { - return llama_kv_self_seq_add(ctx, seq_id, p0, p1, delta); + llama_kv_self_seq_add(ctx, seq_id, p0, p1, delta); } void llama_kv_self_seq_add( @@ -2597,7 +2438,7 @@ void llama_kv_self_seq_add( return; } - return kv->seq_add(seq_id, p0, p1, delta); + kv->seq_add(seq_id, p0, p1, delta); } // deprecated @@ -2607,7 +2448,7 @@ void llama_kv_cache_seq_div( llama_pos p0, llama_pos p1, int d) { - return llama_kv_self_seq_div(ctx, seq_id, p0, p1, d); + llama_kv_self_seq_div(ctx, seq_id, p0, p1, d); } void llama_kv_self_seq_div( @@ -2621,7 +2462,7 @@ void llama_kv_self_seq_div( return; } - return kv->seq_div(seq_id, p0, p1, d); + kv->seq_div(seq_id, p0, p1, d); } // deprecated @@ -2640,7 +2481,7 @@ llama_pos llama_kv_self_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) { // deprecated void llama_kv_cache_defrag(llama_context * ctx) { - return llama_kv_self_defrag(ctx); + llama_kv_self_defrag(ctx); } void llama_kv_self_defrag(llama_context * ctx) { @@ -2649,7 +2490,8 @@ void llama_kv_self_defrag(llama_context * ctx) { return; } - return kv->defrag(); + // force defrag + kv->defrag_sched(-1.0f); } // deprecated @@ -2833,3 +2675,34 @@ void llama_perf_context_print(const llama_context * ctx) { void llama_perf_context_reset(llama_context * ctx) { ctx->perf_reset(); } + +// +// training +// + +bool llama_opt_param_filter_all(const struct ggml_tensor * tensor, void * userdata) { + GGML_UNUSED(tensor); + GGML_UNUSED(userdata); + return true; +} + +void llama_opt_init(struct llama_context * ctx, struct llama_model * model, struct llama_opt_params lopt_params) { + ctx->opt_init(model, lopt_params); +} + +void llama_opt_epoch( + struct llama_context * ctx, + ggml_opt_dataset_t dataset, + ggml_opt_result_t result_train, + ggml_opt_result_t result_eval, + int64_t idata_split, + ggml_opt_epoch_callback callback_train, + ggml_opt_epoch_callback callback_eval) { + ctx->opt_epoch( + dataset, + result_train, + result_eval, + idata_split, + callback_train, + callback_eval); +} diff --git a/src/llama-context.h b/src/llama-context.h index 04facb544c..c0ceacb10c 100644 --- a/src/llama-context.h +++ b/src/llama-context.h @@ -7,6 +7,7 @@ #include "llama-adapter.h" #include "ggml-cpp.h" +#include "ggml-opt.h" #include #include @@ -27,7 +28,12 @@ struct llama_context { void synchronize(); - const llama_model & get_model() const; + const llama_model & get_model() const; + const llama_cparams & get_cparams() const; + + ggml_backend_sched_t get_sched() const; + + ggml_context * get_ctx_compute() const; uint32_t n_ctx() const; uint32_t n_ctx_per_seq() const; @@ -128,6 +134,32 @@ struct llama_context { llama_perf_context_data perf_get_data() const; void perf_reset(); + // + // training + // + + void opt_init(struct llama_model * model, struct llama_opt_params lopt_params); + + void opt_epoch( + ggml_opt_dataset_t dataset, + ggml_opt_result_t result_train, + ggml_opt_result_t result_eval, + int64_t idata_split, + ggml_opt_epoch_callback callback_train, + ggml_opt_epoch_callback callback_eval); + + void opt_epoch_iter( + ggml_opt_dataset_t dataset, + ggml_opt_result_t result, + const std::vector & tokens, + const std::vector & labels_sparse, + llama_batch & batch, + ggml_opt_epoch_callback callback, + bool train, + int64_t idata_in_loop, + int64_t ndata_in_loop, + int64_t t_loop_start); + private: // // output @@ -137,50 +169,30 @@ private: // Returns max number of outputs for which space was reserved. int32_t output_reserve(int32_t n_outputs); - // make the outputs have the same order they had in the user-provided batch - // TODO: maybe remove this - void output_reorder(); - // // graph // +public: int32_t graph_max_nodes() const; // zero-out inputs and create the ctx_compute for the compute graph ggml_cgraph * graph_init(); - llm_graph_result_ptr graph_build( - ggml_context * ctx, - ggml_cgraph * gf, - const llama_ubatch & ubatch, - llm_graph_type gtype); - // returns the result of ggml_backend_sched_graph_compute_async execution ggml_status graph_compute( ggml_cgraph * gf, bool batched); +private: + llm_graph_result_ptr graph_build( + ggml_context * ctx, + ggml_cgraph * gf, + const llama_ubatch & ubatch, + llm_graph_type gtype); + llm_graph_cb graph_get_cb() const; - // used by kv_self_update() - ggml_tensor * build_rope_shift( - ggml_context * ctx0, - ggml_tensor * cur, - ggml_tensor * shift, - ggml_tensor * factors, - float freq_base, - float freq_scale, - ggml_backend_buffer * bbuf) const; - - llm_graph_result_ptr build_kv_self_shift( - ggml_context * ctx0, - ggml_cgraph * gf) const; - - llm_graph_result_ptr build_kv_self_defrag( - ggml_context * ctx0, - ggml_cgraph * gf) const; - // TODO: read/write lora adapters and cvec size_t state_write_data(llama_io_write_i & io); size_t state_read_data (llama_io_read_i & io); @@ -197,14 +209,10 @@ private: llama_cparams cparams; llama_adapter_cvec cvec; llama_adapter_loras loras; - llama_sbatch sbatch; llama_cross cross; // TODO: tmp for handling cross-attention - need something better probably - std::unique_ptr kv_self; - - // TODO: remove - bool logits_all = false; + std::unique_ptr memory; // decode output (2-dimensional array: [n_outputs][n_vocab]) size_t logits_size = 0; // capacity (of floats) for logits @@ -231,6 +239,9 @@ private: ggml_context_ptr ctx_compute; + // training + ggml_opt_context_t opt_ctx = nullptr; + ggml_threadpool_t threadpool = nullptr; ggml_threadpool_t threadpool_batch = nullptr; diff --git a/src/llama-cparams.h b/src/llama-cparams.h index 30e550f023..246fa5777d 100644 --- a/src/llama-cparams.h +++ b/src/llama-cparams.h @@ -30,6 +30,7 @@ struct llama_cparams { bool flash_attn; bool no_perf; bool warmup; + bool op_offload; enum llama_pooling_type pooling_type; diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp index b52e3f6203..b0e3f63597 100644 --- a/src/llama-graph.cpp +++ b/src/llama-graph.cpp @@ -55,7 +55,21 @@ void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) { if (ubatch->pos && pos) { const int64_t n_tokens = ubatch->n_tokens; - ggml_backend_tensor_set(pos, ubatch->pos, 0, n_tokens*n_pos_per_token*ggml_element_size(pos)); + if (ubatch->token && n_pos_per_embd == 4) { + // in case we're using M-RoPE with text tokens, convert the 1D positions to 4D + // the 3 first dims are the same, and 4th dim is all 0 + std::vector pos_data(n_tokens*n_pos_per_embd); + // copy the first dimension + for (int i = 0; i < n_tokens; ++i) { + pos_data[ i] = ubatch->pos[i]; + pos_data[ n_tokens + i] = ubatch->pos[i]; + pos_data[2 * n_tokens + i] = ubatch->pos[i]; + pos_data[3 * n_tokens + i] = 0; // 4th dim is 0 + } + ggml_backend_tensor_set(pos, pos_data.data(), 0, pos_data.size()*ggml_element_size(pos)); + } else { + ggml_backend_tensor_set(pos, ubatch->pos, 0, n_tokens*n_pos_per_embd*ggml_element_size(pos)); + } } } @@ -71,7 +85,7 @@ void llm_graph_input_attn_temp::set_input(const llama_ubatch * ubatch) { ) * f_attn_temp_scale + 1.0; } - ggml_backend_tensor_set(attn_scale, attn_scale_data.data(), 0, n_tokens*n_pos_per_token*ggml_element_size(attn_scale)); + ggml_backend_tensor_set(attn_scale, attn_scale_data.data(), 0, n_tokens*ggml_element_size(attn_scale)); } } @@ -270,24 +284,7 @@ void llm_graph_input_s_copy::set_input(const llama_ubatch * ubatch) { // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n for (uint32_t i = 0; i < n_kv; ++i) { - const uint32_t cell_id = i + kv_self->head; - - ////////////////////////////////////////////// - // TODO: this should not mutate the KV cache ! - llama_kv_cell & kv_cell = const_cast(kv_self)->cells[i]; - - // prevent out-of-bound sources - if (kv_cell.src < 0 || (uint32_t) kv_cell.src >= kv_self->size) { - kv_cell.src = cell_id; - } - - data[i] = kv_cell.src; - - // TODO: do not mutate the KV cache - // ensure copy only happens once - if (kv_cell.src != (int32_t) cell_id) { - kv_cell.src = cell_id; - } + data[i] = kv_self->s_copy(i); } } } @@ -303,18 +300,7 @@ void llm_graph_input_s_mask::set_input(const llama_ubatch * ubatch) { // clear unused states for (int i = 0; i < n_kv; ++i) { - const uint32_t cell_id = i + kv_self->head; - - ////////////////////////////////////////////// - // TODO: this should not mutate the KV cache ! - llama_kv_cell & kv_cell = const_cast(kv_self)->cells[i]; - - data[i] = (float) (kv_cell.src >= 0); - - // only clear once - if (kv_cell.src < 0) { - kv_cell.src = cell_id; - } + data[i] = kv_self->s_mask(i); } } } @@ -592,7 +578,7 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) : res (std::make_unique()) { } -int64_t llm_graph_context::n_pos_per_token() const { +int64_t llm_graph_context::n_pos_per_embd() const { return arch == LLM_ARCH_QWEN2VL ? 4 : 1; } @@ -796,7 +782,7 @@ ggml_tensor * llm_graph_context::build_ffn( } break; } - if (type_gate == LLM_FFN_PAR) { + if (gate && type_gate == LLM_FFN_PAR) { cur = ggml_mul(ctx0, cur, tmp); cb(cur, "ffn_gate_par", il); } @@ -914,28 +900,35 @@ ggml_tensor * llm_graph_context::build_moe_ffn( ggml_tensor * up = build_lora_mm_id(up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens] cb(up, "ffn_moe_up", il); - ggml_tensor * gate = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens] - cb(gate, "ffn_moe_gate", il); + ggml_tensor * experts = nullptr; + if (gate_exps) { + cur = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens] + cb(cur, "ffn_moe_gate", il); + } else { + cur = up; + } switch (type_op) { case LLM_FFN_SILU: { - gate = ggml_silu(ctx0, gate); - cb(gate, "ffn_moe_silu", il); + cur = ggml_silu(ctx0, cur); + cb(cur, "ffn_moe_silu", il); } break; case LLM_FFN_GELU: { - gate = ggml_gelu(ctx0, gate); - cb(gate, "ffn_moe_gelu", il); + cur = ggml_gelu(ctx0, cur); + cb(cur, "ffn_moe_gelu", il); } break; default: GGML_ABORT("fatal error"); } - ggml_tensor * par = ggml_mul(ctx0, up, gate); // [n_ff, n_expert_used, n_tokens] - cb(par, "ffn_moe_gate_par", il); + if (gate_exps) { + cur = ggml_mul(ctx0, cur, up); // [n_ff, n_expert_used, n_tokens] + cb(cur, "ffn_moe_gate_par", il); + } - ggml_tensor * experts = build_lora_mm_id(down_exps, par, selected_experts); // [n_embd, n_expert_used, n_tokens] + experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens] cb(experts, "ffn_moe_down", il); if (!weight_before_ffn) { @@ -978,6 +971,7 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const { 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; cur = ggml_get_rows(ctx0, tok_embd, inp->tokens); @@ -1018,11 +1012,11 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const { } ggml_tensor * llm_graph_context::build_inp_pos() const { - auto inp = std::make_unique(n_pos_per_token()); + auto inp = std::make_unique(n_pos_per_embd()); auto & cur = inp->pos; - cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens*n_pos_per_token()); + cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens*n_pos_per_embd()); ggml_set_input(cur); res->add_input(std::move(inp)); @@ -1031,11 +1025,12 @@ ggml_tensor * llm_graph_context::build_inp_pos() const { } ggml_tensor * llm_graph_context::build_inp_attn_scale() const { - auto inp = std::make_unique(n_pos_per_token(), hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale); + auto inp = std::make_unique(hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale); auto & cur = inp->attn_scale; - cur = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 1, 1, n_tokens*n_pos_per_token()); + // this need to be 1x1xN for broadcasting + cur = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 1, 1, n_tokens); ggml_set_input(cur); res->add_input(std::move(inp)); @@ -1083,7 +1078,7 @@ ggml_tensor * llm_graph_context::build_inp_cls() const { } ggml_tensor * llm_graph_context::build_inp_s_copy() const { - const llama_kv_cache_unified * kv_self = static_cast(memory); + const llama_kv_cache_recurrent * kv_self = static_cast(memory); auto inp = std::make_unique(kv_self); @@ -1100,7 +1095,7 @@ ggml_tensor * llm_graph_context::build_inp_s_copy() const { } ggml_tensor * llm_graph_context::build_inp_s_mask() const { - const llama_kv_cache_unified * kv_self = static_cast(memory); + const llama_kv_cache_recurrent * kv_self = static_cast(memory); auto inp = std::make_unique(kv_self); @@ -1233,8 +1228,19 @@ ggml_tensor * llm_graph_context::build_attn_mha( ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32); if (v_mla) { +#if 0 + // v_mla can be applied as a matrix-vector multiplication with broadcasting across dimension 3 == n_tokens. + // However, the code is optimized for dimensions 0 and 1 being large, so this is ineffient. cur = ggml_reshape_4d(ctx0, cur, v_mla->ne[0], 1, n_head, n_tokens); cur = ggml_mul_mat(ctx0, v_mla, cur); +#else + // It's preferable to do the calculation as a matrix-matrix multiplication with n_tokens in dimension 1. + // The permutations are noops and only change how the tensor data is interpreted. + cur = ggml_permute(ctx0, cur, 0, 2, 1, 3); + cur = ggml_mul_mat(ctx0, v_mla, cur); + cur = ggml_permute(ctx0, cur, 0, 2, 1, 3); + cur = ggml_cont(ctx0, cur); // Needed because ggml_reshape_2d expects contiguous inputs. +#endif } cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens); @@ -1414,8 +1420,6 @@ ggml_tensor * llm_graph_context::build_attn( // store to KV cache { - GGML_ASSERT(!kv_self->recurrent); - const auto kv_head = kv_self->head; GGML_ASSERT(kv_self->size == n_ctx); @@ -1565,7 +1569,7 @@ ggml_tensor * llm_graph_context::build_copy_mask_state( ggml_tensor * state_mask, int32_t n_state, int32_t n_seqs) const { - const llama_kv_cache_unified * kv_self = static_cast(memory); + const llama_kv_cache_recurrent * kv_self = static_cast(memory); const auto n_kv = kv_self->n; const auto kv_head = kv_self->head; @@ -1597,7 +1601,7 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_load( ggml_tensor * state_mask, const llama_ubatch & ubatch, int il) const { - const llama_kv_cache_unified * kv_self = static_cast(memory); + const llama_kv_cache_recurrent * kv_self = static_cast(memory); const auto token_shift_count = hparams.token_shift_count; @@ -1618,7 +1622,7 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_store( ggml_tensor * token_shift, const llama_ubatch & ubatch, int il) const { - const llama_kv_cache_unified * kv_self = static_cast(memory); + const llama_kv_cache_recurrent * kv_self = static_cast(memory); const auto token_shift_count = hparams.token_shift_count; const auto n_embd = hparams.n_embd; diff --git a/src/llama-graph.h b/src/llama-graph.h index d192dc1495..832a8c09f2 100644 --- a/src/llama-graph.h +++ b/src/llama-graph.h @@ -19,6 +19,7 @@ struct llama_cparams; class llama_memory_i; class llama_kv_cache_unified; +class llama_kv_cache_recurrent; // certain models (typically multi-modal) can produce different types of graphs enum llm_graph_type { @@ -90,29 +91,27 @@ public: class llm_graph_input_pos : public llm_graph_input_i { public: - llm_graph_input_pos(int64_t n_pos_per_token) : n_pos_per_token(n_pos_per_token) {} + llm_graph_input_pos(int64_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {} virtual ~llm_graph_input_pos() = default; void set_input(const llama_ubatch * ubatch) override; ggml_tensor * pos = nullptr; // I32 [n_batch] - const int64_t n_pos_per_token = 1; + const int64_t n_pos_per_embd = 1; }; // temperature tuning, used by llama4 class llm_graph_input_attn_temp : public llm_graph_input_i { public: - llm_graph_input_attn_temp(int64_t n_pos_per_token, uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale) - : n_pos_per_token(n_pos_per_token), n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale) {} + llm_graph_input_attn_temp(uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale) + : n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale) {} virtual ~llm_graph_input_attn_temp() = default; void set_input(const llama_ubatch * ubatch) override; ggml_tensor * attn_scale = nullptr; // F32 [n_batch] - const int64_t n_pos_per_token = 1; - const uint32_t n_attn_temp_floor_scale; const float f_attn_temp_scale; }; @@ -188,26 +187,26 @@ public: class llm_graph_input_s_copy : public llm_graph_input_i { public: - llm_graph_input_s_copy(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {} + llm_graph_input_s_copy(const llama_kv_cache_recurrent * kv_self) : kv_self(kv_self) {} virtual ~llm_graph_input_s_copy() = default; void set_input(const llama_ubatch * ubatch) override; ggml_tensor * s_copy; // I32 [kv_size] - const llama_kv_cache_unified * kv_self; + const llama_kv_cache_recurrent * kv_self; }; class llm_graph_input_s_mask : public llm_graph_input_i { public: - llm_graph_input_s_mask(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {} + llm_graph_input_s_mask(const llama_kv_cache_recurrent * kv_self) : kv_self(kv_self) {} virtual ~llm_graph_input_s_mask() = default; void set_input(const llama_ubatch * ubatch) override; ggml_tensor * s_mask; // F32 [1, n_kv] - const llama_kv_cache_unified * kv_self; + const llama_kv_cache_recurrent * kv_self; }; class llm_graph_input_cross_embd : public llm_graph_input_i { @@ -299,6 +298,7 @@ class llm_graph_result_i { public: virtual ~llm_graph_result_i() = default; + virtual ggml_tensor * get_tokens() = 0; virtual ggml_tensor * get_logits() = 0; virtual ggml_tensor * get_embd() = 0; virtual ggml_tensor * get_embd_pooled() = 0; @@ -313,6 +313,7 @@ class llm_graph_result : public llm_graph_result_i { public: virtual ~llm_graph_result() = default; + ggml_tensor * get_tokens() override { return t_tokens; } ggml_tensor * get_logits() override { return t_logits; } ggml_tensor * get_embd() override { return t_embd; } ggml_tensor * get_embd_pooled() override { return t_embd_pooled; } @@ -329,6 +330,7 @@ public: } // important graph nodes + ggml_tensor * t_tokens = nullptr; ggml_tensor * t_logits = nullptr; ggml_tensor * t_embd = nullptr; ggml_tensor * t_embd_pooled = nullptr; @@ -352,8 +354,8 @@ struct llm_graph_params { const llama_cparams & cparams; const llama_ubatch & ubatch; - ggml_backend_sched * sched; - ggml_backend * backend_cpu; + ggml_backend_sched_t sched; + ggml_backend_t backend_cpu; const llama_adapter_cvec * cvec; const llama_adapter_loras * loras; @@ -404,9 +406,9 @@ struct llm_graph_context { ggml_context * ctx0 = nullptr; - ggml_backend_sched * sched; + ggml_backend_sched_t sched; - ggml_backend * backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove? + ggml_backend_t backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove? const llama_adapter_cvec * cvec; const llama_adapter_loras * loras; @@ -419,7 +421,7 @@ struct llm_graph_context { llm_graph_context(const llm_graph_params & params); - int64_t n_pos_per_token() const; + int64_t n_pos_per_embd() const; void cb(ggml_tensor * cur, const char * name, int il) const; diff --git a/src/llama-hparams.h b/src/llama-hparams.h index 80fcd65df0..7ee6a5b75a 100644 --- a/src/llama-hparams.h +++ b/src/llama-hparams.h @@ -66,6 +66,7 @@ struct llama_hparams { float expert_weights_scale = 0.0; bool expert_weights_norm = false; uint32_t expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_NONE; + uint32_t moe_every_n_layers = 0; float f_norm_eps; float f_norm_rms_eps; diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp index 7c9d46d811..3dcad65bb6 100644 --- a/src/llama-kv-cache.cpp +++ b/src/llama-kv-cache.cpp @@ -4,33 +4,41 @@ #include "llama-batch.h" #include "llama-cparams.h" #include "llama-model.h" +#include "llama-context.h" #include #include +#include #include #include #include -llama_kv_cache_unified::llama_kv_cache_unified(const llama_hparams & hparams, callbacks cbs) : hparams(hparams), cbs(std::move(cbs)) { +// +// llama_kv_cache_unified +// + +uint32_t llama_kv_cache_unified::get_padding(const llama_cparams & cparams) { + // the FA kernels require padding to avoid extra runtime boundary checks + return cparams.flash_attn ? 256u : 32u; } -bool llama_kv_cache_unified::init( +llama_kv_cache_unified::llama_kv_cache_unified( const llama_model & model, - const llama_cparams & cparams, ggml_type type_k, ggml_type type_v, + bool v_trans, + bool offload, uint32_t kv_size, - bool offload) { + uint32_t padding) : model(model), hparams(model.hparams), v_trans(v_trans), padding(padding) { const int32_t n_layer = hparams.n_layer; has_shift = false; + can_shift = true; - recurrent = llama_model_is_recurrent(&model); - v_trans = !recurrent && !cparams.flash_attn; - can_shift = !recurrent; + LLAMA_LOG_INFO("%s: kv_size = %d, type_k = '%s', type_v = '%s', n_layer = %d, can_shift = %d, padding = %d\n", + __func__, kv_size, ggml_type_name(type_k), ggml_type_name(type_v), n_layer, can_shift, padding); - LLAMA_LOG_INFO("%s: kv_size = %d, offload = %d, type_k = '%s', type_v = '%s', n_layer = %d, can_shift = %d\n", - __func__, kv_size, offload, ggml_type_name(type_k), ggml_type_name(type_v), n_layer, can_shift); + GGML_ASSERT(kv_size % padding == 0 && "kv_size must be a multiple of padding"); head = 0; size = kv_size; @@ -76,23 +84,20 @@ bool llama_kv_cache_unified::init( const char * dev_name = "CPU"; - ggml_backend_buffer_type_t buft; + ggml_backend_buffer_type_t buft = ggml_backend_cpu_buffer_type(); + if (offload) { auto * dev = model.dev_layer(i); buft = ggml_backend_dev_buffer_type(dev); dev_name = ggml_backend_dev_name(dev); - } else { - buft = ggml_backend_cpu_buffer_type(); } - LLAMA_LOG_DEBUG("%s: layer %3d: n_embd_k_gqa = %d, n_embd_v_gqa = %d, dev = %s\n", __func__, - i, n_embd_k_gqa, n_embd_v_gqa, dev_name); + LLAMA_LOG_DEBUG("%s: layer %3d: dev = %s\n", __func__, i, dev_name); ggml_context * ctx = ctx_for_buft(buft); if (!ctx) { - LLAMA_LOG_ERROR("%s: failed to create ggml context for kv cache\n", __func__); - return false; + throw std::runtime_error("failed to create ggml context for kv cache"); } ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size); @@ -110,55 +115,28 @@ bool llama_kv_cache_unified::init( ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); if (!buf) { - LLAMA_LOG_ERROR("%s: failed to allocate buffer for kv cache\n", __func__); - return false; + throw std::runtime_error("failed to allocate buffer for kv cache"); } ggml_backend_buffer_clear(buf, 0); LLAMA_LOG_INFO("%s: %10s KV buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0); bufs.emplace_back(buf); } - return true; -} + { + const size_t memory_size_k = size_k_bytes(); + const size_t memory_size_v = size_v_bytes(); -int32_t llama_kv_cache_unified::get_n_tokens() const { - int32_t result = 0; - - for (uint32_t i = 0; i < size; i++) { - result += cells[i].seq_id.size(); + LLAMA_LOG_INFO("%s: KV self size = %7.2f MiB, K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__, + (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f), + ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f), + ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f)); } - - return result; -} - -int32_t llama_kv_cache_unified::get_used_cells() const { - return used; -} - -size_t llama_kv_cache_unified::total_size() const { - size_t size = 0; - for (const auto & buf : bufs) { - size += ggml_backend_buffer_get_size(buf.get()); - } - - return size; -} - -llama_pos llama_kv_cache_unified::pos_max() const { - llama_pos pos_max = -1; - for (const auto & cell : cells) { - pos_max = std::max(pos_max, cell.pos); - } - - return pos_max; } void llama_kv_cache_unified::clear() { for (int32_t i = 0; i < (int32_t) size; ++i) { cells[i].pos = -1; cells[i].seq_id.clear(); - cells[i].src = -1; - cells[i].tail = -1; } head = 0; used = 0; @@ -179,35 +157,6 @@ bool llama_kv_cache_unified::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1 = std::numeric_limits::max(); } - // models like Mamba or RWKV can't have a state partially erased - if (recurrent) { - if (seq_id >= (int64_t) size) { - // could be fatal - return false; - } - if (0 <= seq_id) { - int32_t & tail_id = cells[seq_id].tail; - if (tail_id >= 0) { - const llama_kv_cell & cell = cells[tail_id]; - // partial intersection is invalid - if ((0 < p0 && p0 <= cell.pos) || (0 < p1 && p1 <= cell.pos)) { - return false; - } - // invalidate tails which will be cleared - if (p0 <= cell.pos && cell.pos < p1) { - tail_id = -1; - } - } - } else { - // seq_id is negative, then the range should include everything or nothing - if (p0 != p1 && (p0 != 0 || p1 != std::numeric_limits::max())) { - return false; - } - } - - return true; - } - for (uint32_t i = 0; i < size; ++i) { if (cells[i].pos >= p0 && cells[i].pos < p1) { if (seq_id < 0) { @@ -224,7 +173,6 @@ bool llama_kv_cache_unified::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos } cells[i].pos = -1; - cells[i].src = -1; if (new_head == size) { new_head = i; @@ -254,34 +202,6 @@ void llama_kv_cache_unified::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id p1 = std::numeric_limits::max(); } - if (recurrent) { - if ((uint32_t) seq_id_dst < size && (uint32_t) seq_id_src < size) { - llama_kv_cell & tail_src = cells[seq_id_src]; - llama_kv_cell & tail_dst = cells[seq_id_dst]; - if (tail_dst.tail >= 0) { - // clear destination seq_id if it wasn't empty - llama_kv_cell & cell_dst = cells[tail_dst.tail]; - - cell_dst.seq_id.erase(seq_id_dst); - tail_dst.tail = -1; - if (cell_dst.seq_id.empty()) { - cell_dst.pos = -1; - cell_dst.delta = -1; - cell_dst.src = -1; - used -= 1; - } - } - if (tail_src.tail >= 0) { - llama_kv_cell & cell_src = cells[tail_src.tail]; - - cell_src.seq_id.insert(seq_id_dst); - tail_dst.tail = tail_src.tail; - } - } - - return; - } - // otherwise, this is the KV of a Transformer-like model head = 0; @@ -296,17 +216,12 @@ void llama_kv_cache_unified::seq_keep(llama_seq_id seq_id) { uint32_t new_head = size; for (uint32_t i = 0; i < size; ++i) { - if (recurrent && (llama_seq_id) i != seq_id) { - cells[i].tail = -1; - } - if (!cells[i].has_seq_id(seq_id)) { if (cells[i].pos >= 0) { used--; } cells[i].pos = -1; - cells[i].src = -1; cells[i].seq_id.clear(); if (new_head == size){ @@ -344,20 +259,6 @@ void llama_kv_cache_unified::seq_add(llama_seq_id seq_id, llama_pos p0, llama_po return; } - if (recurrent) { - // for Mamba-like or RWKV models, only the pos needs to be shifted - if (0 <= seq_id && seq_id < (int64_t) size) { - const int32_t tail_id = cells[seq_id].tail; - if (tail_id >= 0) { - llama_kv_cell & cell = cells[tail_id]; - if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) { - cell.pos += delta; - } - } - } - return; - } - for (uint32_t i = 0; i < size; ++i) { if (cells[i].has_seq_id(seq_id) && cells[i].pos >= p0 && cells[i].pos < p1) { has_shift = true; @@ -400,21 +301,6 @@ void llama_kv_cache_unified::seq_div(llama_seq_id seq_id, llama_pos p0, llama_po return; } - if (recurrent) { - // for Mamba-like or RWKV models, only the pos needs to be changed - if (0 <= seq_id && seq_id < (int64_t) size) { - const int32_t tail_id = cells[seq_id].tail; - if (tail_id >= 0) { - llama_kv_cell & cell = cells[tail_id]; - if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) { - cell.pos /= d; - } - } - } - - return; - } - for (uint32_t i = 0; i < size; ++i) { if (cells[i].has_seq_id(seq_id) && cells[i].pos >= p0 && cells[i].pos < p1) { has_shift = true; @@ -440,23 +326,11 @@ llama_pos llama_kv_cache_unified::seq_pos_max(llama_seq_id seq_id) const { return result; } -void llama_kv_cache_unified::defrag() { - if (!recurrent) { - do_defrag = true; - } -} - void llama_kv_cache_unified::restore() { if (pending.ranges.empty()) { return; } - // TODO: tmp - move to llama_kv_cache_recurrent - if (recurrent) { - seq_rm(-1, -1, -1); - return; - } - uint32_t new_head = size; for (auto & range : pending.ranges) { @@ -469,7 +343,6 @@ void llama_kv_cache_unified::restore() { } cells[i].pos = -1; - cells[i].src = -1; } new_head = std::min(new_head, range.c0); @@ -481,11 +354,6 @@ void llama_kv_cache_unified::restore() { } void llama_kv_cache_unified::commit() { - // TODO: tmp - move to llama_kv_cache_recurrent - if (recurrent) { - return; - } - if (pending.ranges.empty()) { LLAMA_LOG_WARN("%s: no pending KV cache updates to commit - might indicate a bug (ref: %s)\n", __func__, "https://github.com/ggml-org/llama.cpp/pull/12695"); @@ -495,8 +363,98 @@ void llama_kv_cache_unified::commit() { pending.ranges.clear(); } -bool llama_kv_cache_unified::get_can_shift() const { - return can_shift; +bool llama_kv_cache_unified::update(llama_context & lctx) { + bool need_reserve = false; + + auto * sched = lctx.get_sched(); + + if (has_shift) { + if (!get_can_shift()) { + GGML_ABORT("The current KV cache / model configuration does not support K-shift"); + } + + LLAMA_LOG_DEBUG("%s: applying K-shift\n", __func__); + + // apply K-shift if needed + if (hparams.rope_type != LLAMA_ROPE_TYPE_NONE) { + ggml_backend_sched_reset(sched); + + auto * gf = lctx.graph_init(); + + auto res = build_graph_shift(lctx.get_cparams(), lctx.get_ctx_compute(), gf); + + ggml_backend_sched_alloc_graph(sched, gf); + + res->set_inputs(nullptr); + + lctx.graph_compute(gf, false); + + need_reserve = true; + } + + { + has_shift = false; + + for (uint32_t i = 0; i < size; ++i) { + cells[i].delta = 0; + } + } + } + + if (do_defrag) { + LLAMA_LOG_DEBUG("%s: defragmenting KV cache\n", __func__); + + if (defrag_prepare(lctx.graph_max_nodes())) { + ggml_backend_sched_reset(sched); + + auto * gf = lctx.graph_init(); + + auto res = build_graph_defrag(lctx.get_cparams(), lctx.get_ctx_compute(), gf); + + ggml_backend_sched_alloc_graph(sched, gf); + + res->set_inputs(nullptr); + + lctx.graph_compute(gf, false); + + need_reserve = true; + } + + do_defrag = false; + } + + return need_reserve; +} + +void llama_kv_cache_unified::defrag_sched(float thold) { + // - do not defrag small contexts (i.e. < 2048 tokens) + // - count the padding towards the number of used tokens + const float fragmentation = n >= 2048 ? std::max(0.0f, 1.0f - (float(used + padding)/n)) : 0.0f; + + // queue defragmentation for next llama_kv_cache_update + if (fragmentation > thold) { + LLAMA_LOG_DEBUG("%s: fragmentation: %.2f - requesting defrag\n", __func__, fragmentation); + + do_defrag = true; + } +} + +void llama_kv_cache_unified::set_full() { + n = size; +} + +llama_sbatch llama_kv_cache_unified::sbatch_init( + const llama_batch & batch, + bool logits_all) { + return llama_sbatch(batch, hparams.n_embd, true, logits_all); +} + +llama_ubatch llama_kv_cache_unified::ubatch_next( + llama_sbatch & sbatch, + uint32_t n_ubatch, + bool embd_pooled) const { + GGML_UNUSED(embd_pooled); + return sbatch.split_simple(n_ubatch); } bool llama_kv_cache_unified::find_slot( @@ -511,169 +469,6 @@ bool llama_kv_cache_unified::find_slot( head = 0; } - if (recurrent) { - // For recurrent state architectures (like Mamba or RWKV), - // each cache cell can store the state for a whole sequence. - // A slot should be always be contiguous. - - // can only process batches with an equal number of new tokens in each sequence - GGML_ASSERT(ubatch.equal_seqs); - - int32_t min = size - 1; - int32_t max = 0; - - // everything should fit if all seq_ids are smaller than the max - for (uint32_t s = 0; s < n_seqs; ++s) { - const uint32_t n_seq_id = ubatch.n_seq_id[s]; - for (uint32_t j = 0; j < n_seq_id; ++j) { - const llama_seq_id seq_id = ubatch.seq_id[s][j]; - - if (seq_id < 0 || (uint32_t) seq_id >= size) { - // too big seq_id - // TODO: would it be possible to resize the cache instead? - LLAMA_LOG_ERROR("%s: seq_id=%d >= n_seq_max=%d Try using a bigger --parallel value\n", __func__, seq_id, size); - return false; - } - if (j > 0) { - llama_kv_cell & seq = cells[seq_id]; - if (seq.tail >= 0) { - llama_kv_cell & cell = cells[seq.tail]; - // clear cells from seq_ids that become shared - // (should not normally happen, but let's handle it anyway) - cell.seq_id.erase(seq_id); - seq.tail = -1; - if (cell.seq_id.empty()) { - cell.pos = -1; - cell.src = -1; - used -= 1; - } - } - } - } - } - -#ifndef NDEBUG - { - std::vector tails_verif; - tails_verif.assign(size, -1); - for (uint32_t i = 0; i < size; ++i) { - llama_kv_cell & cell = cells[i]; - for (llama_seq_id seq_id : cell.seq_id) { - if (tails_verif[seq_id] != -1) { - LLAMA_LOG_ERROR("%s: duplicate tail for seq_id %d in cell %d and %d\n", __func__, seq_id, i, tails_verif[seq_id]); - } - tails_verif[seq_id] = i; - } - } - for (uint32_t i = 0; i < size; ++i) { - if (tails_verif[i] != cells[i].tail) { - LLAMA_LOG_ERROR("%s: wrong tail for seq_id %d, (%d instead of %d)\n", __func__, i, cells[i].tail, tails_verif[i]); - } - } - } -#endif - - // find next empty cell - uint32_t next_empty_cell = head; - - for (uint32_t i = 0; i < size; ++i) { - if (next_empty_cell >= size) { next_empty_cell -= size; } - llama_kv_cell & cell = cells[next_empty_cell]; - if (cell.is_empty()) { break; } - next_empty_cell += 1; - } - - // find usable cell range - for (uint32_t s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; - llama_kv_cell & seq_meta = cells[seq_id]; - bool has_cell = false; - if (seq_meta.tail >= 0) { - llama_kv_cell & cell = cells[seq_meta.tail]; - GGML_ASSERT(cell.has_seq_id(seq_id)); - // does this seq_id "own" the cell? - if (cell.seq_id.size() == 1) { has_cell = true; } - } - if (!has_cell) { - llama_kv_cell & empty_cell = cells[next_empty_cell]; - GGML_ASSERT(empty_cell.is_empty()); - // copy old tail into the empty cell - if (seq_meta.tail >= 0) { - llama_kv_cell & orig_cell = cells[seq_meta.tail]; - empty_cell.pos = orig_cell.pos; - empty_cell.src = orig_cell.src; - orig_cell.seq_id.erase(seq_id); - empty_cell.seq_id.insert(seq_id); // will be overwritten - } - seq_meta.tail = next_empty_cell; - // find next empty cell - if (s + 1 < n_seqs) { - next_empty_cell += 1; - for (uint32_t i = 0; i < size; ++i) { - if (next_empty_cell >= size) { next_empty_cell -= size; } - llama_kv_cell & cell = cells[next_empty_cell]; - if (cell.is_empty()) { break; } - next_empty_cell += 1; - } - } - } - if (min > seq_meta.tail) { min = seq_meta.tail; } - if (max < seq_meta.tail) { max = seq_meta.tail; } - } - - // gather and re-order - for (uint32_t s = 0; s < n_seqs; ++s) { - int32_t dst_id = s + min; - int32_t src_id = cells[ubatch.seq_id[s][0]].tail; - if (dst_id != src_id) { - llama_kv_cell & dst_cell = cells[dst_id]; - llama_kv_cell & src_cell = cells[src_id]; - - std::swap(dst_cell.pos, src_cell.pos); - std::swap(dst_cell.src, src_cell.src); - std::swap(dst_cell.seq_id, src_cell.seq_id); - - // swap tails (assuming they NEVER overlap) - for (const llama_seq_id seq_id : src_cell.seq_id) { - cells[seq_id].tail = src_id; - } - for (const llama_seq_id seq_id : dst_cell.seq_id) { - cells[seq_id].tail = dst_id; - } - } - } - - // update the pos of the used seqs - for (uint32_t s = 0; s < n_seqs; ++s) { - const llama_pos last_pos = ubatch.pos[n_seq_tokens * s + n_seq_tokens - 1]; - int32_t cell_id = s + min; - llama_kv_cell & cell = cells[cell_id]; - - if (cell.pos >= 0 && last_pos != cell.pos + (llama_pos) n_seq_tokens) { - // What should happen when the pos backtracks or skips a value? - // Clearing the state mid-batch would require special-casing which isn't done. - LLAMA_LOG_WARN("%s: non-consecutive token position %d after %d for sequence %d with %u new tokens\n", - __func__, last_pos, cell.pos, ubatch.seq_id[s][0], n_seq_tokens); - } - cell.pos = last_pos; - cell.seq_id.clear(); - for (int32_t j = 0; j < ubatch.n_seq_id[s]; ++j) { - const llama_seq_id seq_id = ubatch.seq_id[s][j]; - cell.seq_id.insert(seq_id); - cells[seq_id].tail = cell_id; - } - } - - // allow getting the range of used cells, from head to head + n - head = min; - n = max - min + 1; - used = std::count_if(cells.begin(), cells.end(), - [](const llama_kv_cell& cell){ return !cell.is_empty(); }); - - // sanity check - return n >= n_seqs; - } - // otherwise, one cell per token. if (n_tokens > size) { @@ -725,24 +520,50 @@ bool llama_kv_cache_unified::find_slot( pending.ranges.push_back({head, head + n_tokens}); + // a heuristic, to avoid attending the full cache if it is not yet utilized + // after enough generations, the benefit from this heuristic disappears + // if we start defragmenting the cache, the benefit from this will be more important + n = std::min(size, std::max(padding, GGML_PAD(cell_max(), padding))); + + //printf("n = %5d, used = %5d, head = %5d\n", n, used, head); + return true; } -uint32_t llama_kv_cache_unified::get_padding(const llama_cparams & cparams) const { - // the FA kernels require padding to avoid extra runtime boundary checks - return cparams.flash_attn ? 256u : 32u; -} +int32_t llama_kv_cache_unified::get_n_tokens() const { + int32_t result = 0; -uint32_t llama_kv_cache_unified::cell_max() const { - for (uint32_t i = size; i > 0; --i) { - const llama_kv_cell & cell = cells[i - 1]; - - if (cell.pos >= 0 && !cell.is_empty()) { - return i; - } + for (uint32_t i = 0; i < size; i++) { + result += cells[i].seq_id.size(); } - return 0; + return result; +} + +int32_t llama_kv_cache_unified::get_used_cells() const { + return used; +} + +bool llama_kv_cache_unified::get_can_shift() const { + return can_shift; +} + +llama_pos llama_kv_cache_unified::get_pos_max() const { + llama_pos pos_max = -1; + for (const auto & cell : cells) { + pos_max = std::max(pos_max, cell.pos); + } + + return pos_max; +} + +size_t llama_kv_cache_unified::total_size() const { + size_t size = 0; + for (const auto & buf : bufs) { + size += ggml_backend_buffer_get_size(buf.get()); + } + + return size; } size_t llama_kv_cache_unified::size_k_bytes() const { @@ -765,6 +586,269 @@ size_t llama_kv_cache_unified::size_v_bytes() const { return size_v_bytes; } +ggml_tensor * llama_kv_cache_unified::build_rope_shift( + const llama_cparams & cparams, + ggml_context * ctx, + ggml_tensor * cur, + ggml_tensor * shift, + ggml_tensor * factors, + float freq_base, + float freq_scale) const { + const auto & n_ctx_orig = cparams.n_ctx_orig_yarn; + + const auto & yarn_ext_factor = cparams.yarn_ext_factor; + const auto & yarn_beta_fast = cparams.yarn_beta_fast; + const auto & yarn_beta_slow = cparams.yarn_beta_slow; + + const auto & n_rot = hparams.n_rot; + const auto & rope_type = hparams.rope_type; + + // See llm_build_deepseek2() for why attn_factor has to be scaled for YaRN RoPE to work correctly. + // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation. + const float yarn_attn_factor = model.arch == LLM_ARCH_DEEPSEEK2 ? 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale)) : cparams.yarn_attn_factor; + + ggml_tensor * tmp; + + if (ggml_is_quantized(cur->type)) { + // dequantize to f32 -> RoPE -> quantize back + tmp = ggml_cast(ctx, cur, GGML_TYPE_F32); + + tmp = ggml_rope_ext(ctx, tmp, + shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow); + + tmp = ggml_cpy(ctx, tmp, cur); + } else { + // we rotate only the first n_rot dimensions + tmp = ggml_rope_ext_inplace(ctx, cur, + shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow); + } + + return tmp; +} + +class llm_graph_input_k_shift : public llm_graph_input_i { +public: + llm_graph_input_k_shift(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {} + virtual ~llm_graph_input_k_shift() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * k_shift; // I32 [kv_size] + + const llama_kv_cache_unified * kv_self; +}; + +void llm_graph_input_k_shift::set_input(const llama_ubatch * ubatch) { + GGML_UNUSED(ubatch); + + if (k_shift) { + assert(ggml_backend_buffer_is_host(k_shift->buffer)); + + int32_t * data = (int32_t *) k_shift->data; + + for (uint32_t i = 0; i < kv_self->size; ++i) { + data[i] = kv_self->cells[i].delta; + } + } +} + +llm_graph_result_ptr llama_kv_cache_unified::build_graph_shift( + const llama_cparams & cparams, + ggml_context * ctx, + ggml_cgraph * gf) const { + auto res = std::make_unique(); + + const auto & n_layer = hparams.n_layer; + + const auto & n_embd_head_k = hparams.n_embd_head_k; + //const auto & n_embd_head_v = hparams.n_embd_head_v; + + const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max; + + //GGML_ASSERT(kv_self->size == n_ctx); + + auto inp = std::make_unique(this); + + inp->k_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, cparams.n_ctx); + ggml_set_input(inp->k_shift); + + for (uint32_t il = 0; il < n_layer; ++il) { + const int64_t n_head_kv = hparams.n_head_kv(il); + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); + + const bool is_swa = hparams.is_swa(il); + + // note: the swa rope params could become part of the cparams in the future + // if we decide to make them configurable, like the non-sliding ones + const float freq_base_l = is_swa ? hparams.rope_freq_base_train_swa : cparams.rope_freq_base; + const float freq_scale_l = is_swa ? hparams.rope_freq_scale_train_swa : cparams.rope_freq_scale; + + ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il); + + ggml_tensor * k = + ggml_view_3d(ctx, k_l[il], + n_embd_head_k, n_head_kv, size, + ggml_row_size(k_l[il]->type, n_embd_head_k), + ggml_row_size(k_l[il]->type, n_embd_k_gqa), + 0); + + ggml_tensor * cur = build_rope_shift(cparams, ctx, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l); + + ggml_build_forward_expand(gf, cur); + } + + res->add_input(std::move(inp)); + + return res; +} + +llm_graph_result_ptr llama_kv_cache_unified::build_graph_defrag( + const llama_cparams & cparams, + ggml_context * ctx, + ggml_cgraph * gf) const { + auto res = std::make_unique(); + + const auto & ids = defrag_info.ids; + +#if 0 + // CPU defrag + // + // TODO: optimizations are possible: + // - multiple threads + // - avoid copying to the host memory when already there + // + // likely not worth the effort, as we have ggml_graph based defrag + // + + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(); + + const uint32_t kv_size = size; + + std::vector buf_k; + std::vector buf_v; + + for (uint32_t il = 0; il < n_layer; ++il) { + const size_t k_size_row = ggml_row_size(k_l[il]->type, n_embd_k_gqa); + const size_t k_size = ggml_row_size(k_l[il]->type, n_embd_k_gqa*kv_size); + + const size_t v_size_el = ggml_type_size(v_l[il]->type); + const size_t v_size = ggml_row_size (v_l[il]->type, n_embd_v_gqa*kv_size); + + buf_k.resize(k_size); + buf_v.resize(v_size); + + ggml_backend_tensor_get(k_l[il], buf_k.data(), 0, buf_k.size()); + ggml_backend_tensor_get(v_l[il], buf_v.data(), 0, buf_v.size()); + + // batch move [i, i+nm) to [id, id+nm) + // note: cells can move only to a lower index + for (uint32_t i = 0; i < n_kv; ++i) { + const uint32_t id = ids[i]; + + if (i == id || id == n_kv) { + continue; + } + + uint32_t nm = 1; + + while (i + nm < n_kv && ids[i + nm] == id + nm) { + nm++; + } + + // move keys + { + const int64_t os = i*k_size_row; + const int64_t od = id*k_size_row; + + memcpy(buf_k.data() + od, buf_k.data() + os, nm*k_size_row); + } + + // move values (note: they are transposed) + { + const int64_t os = i; + const int64_t od = id; + + for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { + memcpy(buf_v.data() + (od + j*kv_size)*v_size_el, buf_v.data() + (os + j*kv_size)*v_size_el, nm*v_size_el); + } + } + + i += nm - 1; + } + + ggml_backend_tensor_set(k_l[il], buf_k.data(), 0, buf_k.size()); + ggml_backend_tensor_set(v_l[il], buf_v.data(), 0, buf_v.size()); + } +#else + for (uint32_t i = 0; i < ids.size(); ++i) { + const uint32_t id = ids[i]; + + if (i == id || id == ids.size()) { + continue; + } + + uint32_t nm = 1; + + while (i + nm < ids.size() && ids[i + nm] == id + nm) { + nm++; + } + + for (uint32_t il = 0; il < hparams.n_layer; ++il) { // NOLINT + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); + const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); + + ggml_tensor * view_k_src = ggml_view_2d(ctx, k_l[il], + n_embd_k_gqa, nm, + ggml_row_size(k_l[il]->type, n_embd_k_gqa), + ggml_row_size(k_l[il]->type, n_embd_k_gqa*i)); + + ggml_tensor * view_k_dst = ggml_view_2d(ctx, k_l[il], + n_embd_k_gqa, nm, + ggml_row_size(k_l[il]->type, n_embd_k_gqa), + ggml_row_size(k_l[il]->type, n_embd_k_gqa*id)); + + ggml_tensor * view_v_src; + ggml_tensor * view_v_dst; + + if (cparams.flash_attn) { + // NOTE: the V cache is not transposed when using flash attention + view_v_src = ggml_view_2d(ctx, v_l[il], + n_embd_v_gqa, nm, + ggml_row_size(v_l[il]->type, n_embd_v_gqa), + ggml_row_size(v_l[il]->type, n_embd_v_gqa*i)); + + view_v_dst = ggml_view_2d(ctx, v_l[il], + n_embd_v_gqa, nm, + ggml_row_size(v_l[il]->type, n_embd_v_gqa), + ggml_row_size(v_l[il]->type, n_embd_v_gqa*id)); + } else { + view_v_src = ggml_view_2d(ctx, v_l[il], + nm, n_embd_v_gqa, + ggml_row_size(v_l[il]->type, size), + ggml_row_size(v_l[il]->type, i)); + + view_v_dst = ggml_view_2d(ctx, v_l[il], + nm, n_embd_v_gqa, + ggml_row_size(v_l[il]->type, size), + ggml_row_size(v_l[il]->type, id)); + } + + ggml_build_forward_expand(gf, ggml_cpy(ctx, view_k_src, view_k_dst)); + ggml_build_forward_expand(gf, ggml_cpy(ctx, view_v_src, view_v_dst)); + } + + i += nm - 1; + } + + //LLAMA_LOG_INFO("gf->n_nodes = %d\n", gf->n_nodes); +#endif + + return res; +} + bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) { const uint32_t n_layer = hparams.n_layer; @@ -867,7 +951,7 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) { cells[i0 + nf] = cell1; // clear the old cell and move the head there - cell1 = llama_kv_cell(); + cell1 = kv_cell(); head = n_used; if (!cont) { @@ -895,13 +979,25 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) { return false; } - LLAMA_LOG_DEBUG("(tmp log) KV defrag cell moves: %u\n", n_moves); + LLAMA_LOG_DEBUG("%s: (tmp log) KV defrag cell moves: %u\n", __func__, n_moves); - LLAMA_LOG_DEBUG("expected gf nodes: %u\n", 6*n_moves*n_layer); + LLAMA_LOG_DEBUG("%s: expected gf nodes: %u\n", __func__, 6*n_moves*n_layer); return true; } +uint32_t llama_kv_cache_unified::cell_max() const { + for (uint32_t i = size; i > 0; --i) { + const kv_cell & cell = cells[i - 1]; + + if (cell.pos >= 0 && !cell.is_empty()) { + return i; + } + } + + return 0; +} + void llama_kv_cache_unified::state_write(llama_io_write_i & io, llama_seq_id seq_id) const { std::vector> cell_ranges; // ranges, from inclusive, to exclusive uint32_t cell_count = 0; @@ -1110,7 +1206,7 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell clear(); for (uint32_t i = 0; i < cell_count; ++i) { - llama_kv_cell & cell = cells[i]; + kv_cell & cell = cells[i]; llama_pos pos; uint32_t n_seq_id; @@ -1133,15 +1229,6 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell } cell.seq_id.insert(seq_id); - - if (recurrent) { - int32_t & tail = cells[seq_id].tail; - if (tail != -1) { - LLAMA_LOG_ERROR("%s: duplicate tail for seq_id %d in cell %d and %d\n", __func__, seq_id, i, tail); - return false; - } - tail = i; - } } } @@ -1149,14 +1236,6 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell used = cell_count; } - if (recurrent) { - for (uint32_t i = 0; i < cell_count; ++i) { - uint32_t cell_id = head + i; - // make sure the recurrent states will keep their restored state - cells[cell_id].src = cell_id; - } - } - return true; } @@ -1174,7 +1253,1034 @@ bool llama_kv_cache_unified::state_read_data(llama_io_read_i & io, uint32_t cell LLAMA_LOG_ERROR("%s: not enough cells in kv cache to restore state (%u > %u)\n", __func__, cell_count, size); return false; } - if (v_trans != (bool) v_trans) { + if (this->v_trans != (bool) v_trans) { + LLAMA_LOG_ERROR("%s: incompatible V transposition\n", __func__); + return false; + } + + // For each layer, read the keys for each cell, one row is one cell, read as one contiguous block + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); + + // Read type of key + int32_t k_type_i_ref; + io.read_to(&k_type_i_ref, sizeof(k_type_i_ref)); + const int32_t k_type_i = (int32_t) k_l[il]->type; + if (k_type_i != k_type_i_ref) { + LLAMA_LOG_ERROR("%s: mismatched key type (%d != %d, layer %d)\n", __func__, k_type_i, k_type_i_ref, il); + return false; + } + + // Read row size of key + uint64_t k_size_row_ref; + io.read_to(&k_size_row_ref, sizeof(k_size_row_ref)); + const size_t k_size_row = ggml_row_size(k_l[il]->type, n_embd_k_gqa); + if (k_size_row != k_size_row_ref) { + LLAMA_LOG_ERROR("%s: mismatched key row size (%zu != %zu, layer %d)\n", __func__, k_size_row, (size_t) k_size_row_ref, il); + return false; + } + + if (cell_count) { + // Read and set the keys for the whole cell range + ggml_backend_tensor_set(k_l[il], io.read(cell_count * k_size_row), head * k_size_row, cell_count * k_size_row); + } + } + + if (!this->v_trans) { + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + + // Read type of value + int32_t v_type_i_ref; + io.read_to(&v_type_i_ref, sizeof(v_type_i_ref)); + const int32_t v_type_i = (int32_t)v_l[il]->type; + if (v_type_i != v_type_i_ref) { + LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il); + return false; + } + + // Read row size of value + uint64_t v_size_row_ref; + io.read_to(&v_size_row_ref, sizeof(v_size_row_ref)); + const size_t v_size_row = ggml_row_size(v_l[il]->type, n_embd_v_gqa); + if (v_size_row != v_size_row_ref) { + LLAMA_LOG_ERROR("%s: mismatched value row size (%zu != %zu, layer %d)\n", __func__, v_size_row, (size_t) v_size_row_ref, il); + return false; + } + + if (cell_count) { + // Read and set the values for the whole cell range + ggml_backend_tensor_set(v_l[il], io.read(cell_count * v_size_row), head * v_size_row, cell_count * v_size_row); + } + } + } else { + // For each layer, read the values for each cell (transposed) + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + + // Read type of value + int32_t v_type_i_ref; + io.read_to(&v_type_i_ref, sizeof(v_type_i_ref)); + const int32_t v_type_i = (int32_t)v_l[il]->type; + if (v_type_i != v_type_i_ref) { + LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il); + return false; + } + + // Read element size of value + uint32_t v_size_el_ref; + io.read_to(&v_size_el_ref, sizeof(v_size_el_ref)); + const size_t v_size_el = ggml_type_size(v_l[il]->type); + if (v_size_el != v_size_el_ref) { + LLAMA_LOG_ERROR("%s: mismatched value element size (%zu != %zu, layer %d)\n", __func__, v_size_el, (size_t) v_size_el_ref, il); + return false; + } + + // Read GQA embedding size + uint32_t n_embd_v_gqa_ref; + io.read_to(&n_embd_v_gqa_ref, sizeof(n_embd_v_gqa_ref)); + if (n_embd_v_gqa != n_embd_v_gqa_ref) { + LLAMA_LOG_ERROR("%s: mismatched GQA embedding size (%u != %u, layer %d)\n", __func__, n_embd_v_gqa, n_embd_v_gqa_ref, il); + return false; + } + + if (cell_count) { + // For each row in the transposed matrix, read the values for the whole cell range + for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { + const size_t dst_offset = (head + j * size) * v_size_el; + ggml_backend_tensor_set(v_l[il], io.read(cell_count * v_size_el), dst_offset, cell_count * v_size_el); + } + } + } + } + + return true; +} + +// +// llama_kv_cache_recurrent +// + +llama_kv_cache_recurrent::llama_kv_cache_recurrent( + const llama_model & model, + ggml_type type_k, + ggml_type type_v, + bool offload, + uint32_t kv_size) : hparams(model.hparams) { + const int32_t n_layer = hparams.n_layer; + + LLAMA_LOG_INFO("%s: kv_size = %d, type_k = '%s', type_v = '%s', n_layer = %d\n", + __func__, kv_size, ggml_type_name(type_k), ggml_type_name(type_v), n_layer); + + head = 0; + size = kv_size; + used = 0; + + this->type_k = type_k; + this->type_v = type_v; + + cells.clear(); + cells.resize(kv_size); + + // create a context for each buffer type + std::map ctx_map; + auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * { + auto it = ctx_map.find(buft); + if (it == ctx_map.end()) { + ggml_init_params params = { + /*.mem_size =*/ size_t(2u*n_layer*ggml_tensor_overhead()), + /*.mem_buffer =*/ NULL, + /*.no_alloc =*/ true, + }; + + ggml_context * ctx = ggml_init(params); + if (!ctx) { + return nullptr; + } + + ctx_map[buft] = ctx; + ctxs.emplace_back(ctx); + + return ctx; + } + + return it->second; + }; + + k_l.reserve(n_layer); + v_l.reserve(n_layer); + + for (int i = 0; i < n_layer; i++) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s(); + + const char * dev_name = "CPU"; + + ggml_backend_buffer_type_t buft = ggml_backend_cpu_buffer_type(); + + if (offload) { + auto * dev = model.dev_layer(i); + buft = ggml_backend_dev_buffer_type(dev); + + dev_name = ggml_backend_dev_name(dev); + } + + LLAMA_LOG_DEBUG("%s, layer %3d: dev = %s\n", __func__, i, dev_name); + + ggml_context * ctx = ctx_for_buft(buft); + if (!ctx) { + throw std::runtime_error("failed to create ggml context for kv cache"); + } + + ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size); + ggml_tensor * v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size); + ggml_format_name(k, "cache_k_l%d", i); + ggml_format_name(v, "cache_v_l%d", i); + k_l.push_back(k); + v_l.push_back(v); + } + + // allocate tensors and initialize the buffers to avoid NaNs in the padding + for (auto it : ctx_map) { + auto * buft = it.first; + auto * ctx = it.second; + + ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); + if (!buf) { + throw std::runtime_error("failed to allocate buffer for kv cache"); + } + ggml_backend_buffer_clear(buf, 0); + LLAMA_LOG_INFO("%s: %10s KV buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0); + bufs.emplace_back(buf); + } + + { + const size_t memory_size_k = size_k_bytes(); + const size_t memory_size_v = size_v_bytes(); + + LLAMA_LOG_INFO("%s: KV self size = %7.2f MiB, K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__, + (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f), + ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f), + ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f)); + } +} + +void llama_kv_cache_recurrent::clear() { + for (int32_t i = 0; i < (int32_t) size; ++i) { + cells[i].pos = -1; + cells[i].seq_id.clear(); + cells[i].src = -1; + cells[i].tail = -1; + } + head = 0; + used = 0; + + for (auto & buf : bufs) { + ggml_backend_buffer_clear(buf.get(), 0); + } +} + +bool llama_kv_cache_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) { + uint32_t new_head = size; + + if (p0 < 0) { + p0 = 0; + } + + if (p1 < 0) { + p1 = std::numeric_limits::max(); + } + + // models like Mamba or RWKV can't have a state partially erased + if (seq_id >= (int64_t) size) { + // could be fatal + return false; + } + if (0 <= seq_id) { + int32_t & tail_id = cells[seq_id].tail; + if (tail_id >= 0) { + const kv_cell & cell = cells[tail_id]; + // partial intersection is invalid + if ((0 < p0 && p0 <= cell.pos) || (0 < p1 && p1 <= cell.pos)) { + return false; + } + // invalidate tails which will be cleared + if (p0 <= cell.pos && cell.pos < p1) { + tail_id = -1; + } + } + } else { + // seq_id is negative, then the range should include everything or nothing + if (p0 != p1 && (p0 != 0 || p1 != std::numeric_limits::max())) { + return false; + } + } + + for (uint32_t i = 0; i < size; ++i) { + if (cells[i].pos >= p0 && cells[i].pos < p1) { + if (seq_id < 0) { + cells[i].seq_id.clear(); + } else if (cells[i].has_seq_id(seq_id)) { + cells[i].seq_id.erase(seq_id); + } else { + continue; + } + if (cells[i].is_empty()) { + // keep count of the number of used cells + if (cells[i].pos >= 0) { + used--; + } + cells[i].pos = -1; + cells[i].src = -1; + if (new_head == size) { + new_head = i; + } + } + } + } + + // If we freed up a slot, set head to it so searching can start there. + if (new_head != size && new_head < head) { + head = new_head; + } + + return true; +} + +void llama_kv_cache_recurrent::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) { + if (seq_id_src == seq_id_dst) { + return; + } + + if (p0 < 0) { + p0 = 0; + } + + if (p1 < 0) { + p1 = std::numeric_limits::max(); + } + + if ((uint32_t) seq_id_dst < size && (uint32_t) seq_id_src < size) { + kv_cell & tail_src = cells[seq_id_src]; + kv_cell & tail_dst = cells[seq_id_dst]; + if (tail_dst.tail >= 0) { + // clear destination seq_id if it wasn't empty + kv_cell & cell_dst = cells[tail_dst.tail]; + + cell_dst.seq_id.erase(seq_id_dst); + tail_dst.tail = -1; + if (cell_dst.seq_id.empty()) { + cell_dst.pos = -1; + cell_dst.src = -1; + used -= 1; + } + } + if (tail_src.tail >= 0) { + kv_cell & cell_src = cells[tail_src.tail]; + + cell_src.seq_id.insert(seq_id_dst); + tail_dst.tail = tail_src.tail; + } + } +} + +void llama_kv_cache_recurrent::seq_keep(llama_seq_id seq_id) { + uint32_t new_head = size; + + for (uint32_t i = 0; i < size; ++i) { + if ((llama_seq_id) i != seq_id) { + cells[i].tail = -1; + } + + if (!cells[i].has_seq_id(seq_id)) { + if (cells[i].pos >= 0) { + used--; + } + + cells[i].pos = -1; + cells[i].src = -1; + cells[i].seq_id.clear(); + + if (new_head == size){ + new_head = i; + } + } else { + cells[i].seq_id.clear(); + cells[i].seq_id.insert(seq_id); + } + } + + // If we freed up a slot, set head to it so searching can start there. + if (new_head != size && new_head < head) { + head = new_head; + } +} + +void llama_kv_cache_recurrent::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) { + if (delta == 0) { + return; + } + + if (p0 < 0) { + p0 = 0; + } + + if (p1 < 0) { + p1 = std::numeric_limits::max(); + } + + // If there is no range then return early to avoid looping over the + if (p0 == p1) { + return; + } + + // for Mamba-like or RWKV models, only the pos needs to be shifted + if (0 <= seq_id && seq_id < (int64_t) size) { + const int32_t tail_id = cells[seq_id].tail; + if (tail_id >= 0) { + kv_cell & cell = cells[tail_id]; + if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) { + cell.pos += delta; + } + } + } +} + +void llama_kv_cache_recurrent::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) { + if (d == 1) { + return; + } + + if (p0 < 0) { + p0 = 0; + } + + if (p1 < 0) { + p1 = std::numeric_limits::max(); + } + + // If there is no range then return early to avoid looping over the cache. + if (p0 == p1) { + return; + } + + // for Mamba-like or RWKV models, only the pos needs to be changed + if (0 <= seq_id && seq_id < (int64_t) size) { + const int32_t tail_id = cells[seq_id].tail; + if (tail_id >= 0) { + kv_cell & cell = cells[tail_id]; + if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) { + cell.pos /= d; + } + } + } +} + +llama_pos llama_kv_cache_recurrent::seq_pos_max(llama_seq_id seq_id) const { + llama_pos result = 0; + + for (uint32_t i = 0; i < size; ++i) { + if (cells[i].has_seq_id(seq_id)) { + result = std::max(result, cells[i].pos); + } + } + + return result; +} + +void llama_kv_cache_recurrent::restore() { + if (pending.ranges.empty()) { + return; + } + + seq_rm(-1, -1, -1); +} + +void llama_kv_cache_recurrent::commit() { + pending.ranges.clear(); +} + +bool llama_kv_cache_recurrent::update(llama_context & lctx) { + GGML_UNUSED(lctx); + return false; +} + +void llama_kv_cache_recurrent::defrag_sched(float thold) { + GGML_UNUSED(thold); + // noop +} + +void llama_kv_cache_recurrent::set_full() { + n = size; +} + +llama_sbatch llama_kv_cache_recurrent::sbatch_init( + const llama_batch & batch, + bool logits_all) { + return llama_sbatch(batch, hparams.n_embd, false, logits_all); +} + +llama_ubatch llama_kv_cache_recurrent::ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const { + if (embd_pooled) { + // Pooled embeddings cannot be split across ubatches (yet) + return sbatch.split_seq(n_ubatch); + } + + return sbatch.split_equal(n_ubatch); +} + +bool llama_kv_cache_recurrent::find_slot( + const llama_ubatch & ubatch) { + const uint32_t n_tokens = ubatch.n_tokens; + const uint32_t n_seqs = ubatch.n_seqs; + + const uint32_t n_seq_tokens = ubatch.n_seq_tokens; + + // if we have enough unused cells before the current head -> + // better to start searching from the beginning of the cache, hoping to fill it + if (head > used + 2*n_tokens) { + head = 0; + } + + // For recurrent state architectures (like Mamba or RWKV), + // each cache cell can store the state for a whole sequence. + // A slot should be always be contiguous. + + // can only process batches with an equal number of new tokens in each sequence + GGML_ASSERT(ubatch.equal_seqs); + + int32_t min = size - 1; + int32_t max = 0; + + // everything should fit if all seq_ids are smaller than the max + for (uint32_t s = 0; s < n_seqs; ++s) { + const uint32_t n_seq_id = ubatch.n_seq_id[s]; + for (uint32_t j = 0; j < n_seq_id; ++j) { + const llama_seq_id seq_id = ubatch.seq_id[s][j]; + + if (seq_id < 0 || (uint32_t) seq_id >= size) { + // too big seq_id + // TODO: would it be possible to resize the cache instead? + LLAMA_LOG_ERROR("%s: seq_id=%d >= n_seq_max=%d Try using a bigger --parallel value\n", __func__, seq_id, size); + return false; + } + if (j > 0) { + kv_cell & seq = cells[seq_id]; + if (seq.tail >= 0) { + kv_cell & cell = cells[seq.tail]; + // clear cells from seq_ids that become shared + // (should not normally happen, but let's handle it anyway) + cell.seq_id.erase(seq_id); + seq.tail = -1; + if (cell.seq_id.empty()) { + cell.pos = -1; + cell.src = -1; + used -= 1; + } + } + } + } + } + +#ifndef NDEBUG + { + std::vector tails_verif; + tails_verif.assign(size, -1); + for (uint32_t i = 0; i < size; ++i) { + kv_cell & cell = cells[i]; + for (llama_seq_id seq_id : cell.seq_id) { + if (tails_verif[seq_id] != -1) { + LLAMA_LOG_ERROR("%s: duplicate tail for seq_id %d in cell %d and %d\n", __func__, seq_id, i, tails_verif[seq_id]); + } + tails_verif[seq_id] = i; + } + } + for (uint32_t i = 0; i < size; ++i) { + if (tails_verif[i] != cells[i].tail) { + LLAMA_LOG_ERROR("%s: wrong tail for seq_id %d, (%d instead of %d)\n", __func__, i, cells[i].tail, tails_verif[i]); + } + } + } +#endif + + // find next empty cell + uint32_t next_empty_cell = head; + + for (uint32_t i = 0; i < size; ++i) { + if (next_empty_cell >= size) { next_empty_cell -= size; } + kv_cell & cell = cells[next_empty_cell]; + if (cell.is_empty()) { break; } + next_empty_cell += 1; + } + + // find usable cell range + for (uint32_t s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = ubatch.seq_id[s][0]; + kv_cell & seq_meta = cells[seq_id]; + bool has_cell = false; + if (seq_meta.tail >= 0) { + kv_cell & cell = cells[seq_meta.tail]; + GGML_ASSERT(cell.has_seq_id(seq_id)); + // does this seq_id "own" the cell? + if (cell.seq_id.size() == 1) { has_cell = true; } + } + if (!has_cell) { + kv_cell & empty_cell = cells[next_empty_cell]; + GGML_ASSERT(empty_cell.is_empty()); + // copy old tail into the empty cell + if (seq_meta.tail >= 0) { + kv_cell & orig_cell = cells[seq_meta.tail]; + empty_cell.pos = orig_cell.pos; + empty_cell.src = orig_cell.src; + orig_cell.seq_id.erase(seq_id); + empty_cell.seq_id.insert(seq_id); // will be overwritten + } + seq_meta.tail = next_empty_cell; + // find next empty cell + if (s + 1 < n_seqs) { + next_empty_cell += 1; + for (uint32_t i = 0; i < size; ++i) { + if (next_empty_cell >= size) { next_empty_cell -= size; } + kv_cell & cell = cells[next_empty_cell]; + if (cell.is_empty()) { break; } + next_empty_cell += 1; + } + } + } + if (min > seq_meta.tail) { min = seq_meta.tail; } + if (max < seq_meta.tail) { max = seq_meta.tail; } + } + + // gather and re-order + for (uint32_t s = 0; s < n_seqs; ++s) { + int32_t dst_id = s + min; + int32_t src_id = cells[ubatch.seq_id[s][0]].tail; + if (dst_id != src_id) { + kv_cell & dst_cell = cells[dst_id]; + kv_cell & src_cell = cells[src_id]; + + std::swap(dst_cell.pos, src_cell.pos); + std::swap(dst_cell.src, src_cell.src); + std::swap(dst_cell.seq_id, src_cell.seq_id); + + // swap tails (assuming they NEVER overlap) + for (const llama_seq_id seq_id : src_cell.seq_id) { + cells[seq_id].tail = src_id; + } + for (const llama_seq_id seq_id : dst_cell.seq_id) { + cells[seq_id].tail = dst_id; + } + } + } + + // update the pos of the used seqs + for (uint32_t s = 0; s < n_seqs; ++s) { + const llama_pos last_pos = ubatch.pos[n_seq_tokens * s + n_seq_tokens - 1]; + int32_t cell_id = s + min; + kv_cell & cell = cells[cell_id]; + + if (cell.pos >= 0 && last_pos != cell.pos + (llama_pos) n_seq_tokens) { + // What should happen when the pos backtracks or skips a value? + // Clearing the state mid-batch would require special-casing which isn't done. + LLAMA_LOG_WARN("%s: non-consecutive token position %d after %d for sequence %d with %u new tokens\n", + __func__, last_pos, cell.pos, ubatch.seq_id[s][0], n_seq_tokens); + } + cell.pos = last_pos; + cell.seq_id.clear(); + for (int32_t j = 0; j < ubatch.n_seq_id[s]; ++j) { + const llama_seq_id seq_id = ubatch.seq_id[s][j]; + cell.seq_id.insert(seq_id); + cells[seq_id].tail = cell_id; + } + } + + // allow getting the range of used cells, from head to head + n + head = min; + n = max - min + 1; + used = std::count_if(cells.begin(), cells.end(), + [](const kv_cell & cell){ return !cell.is_empty(); }); + + // sanity check + return n >= n_seqs; +} + +int32_t llama_kv_cache_recurrent::get_n_tokens() const { + int32_t result = 0; + + for (uint32_t i = 0; i < size; i++) { + result += cells[i].seq_id.size(); + } + + return result; +} + +int32_t llama_kv_cache_recurrent::get_used_cells() const { + return used; +} + +llama_pos llama_kv_cache_recurrent::get_pos_max() const { + llama_pos pos_max = -1; + for (const auto & cell : cells) { + pos_max = std::max(pos_max, cell.pos); + } + + return pos_max; +} + +bool llama_kv_cache_recurrent::get_can_shift() const { + return false; +} + +int32_t llama_kv_cache_recurrent::s_copy(int i) const { + const uint32_t cell_id = i + head; + + ////////////////////////////////////////////// + // TODO: this should not mutate the KV cache ! + kv_cell & cell = const_cast(cells[cell_id]); + + // prevent out-of-bound sources + if (cell.src < 0 || (uint32_t) cell.src >= size) { + cell.src = cell_id; + } + + int32_t res = cell.src; + + // TODO: do not mutate the KV cache + // ensure copy only happens once + if (cell.src != (int32_t) cell_id) { + cell.src = cell_id; + } + + return res; +} + +float llama_kv_cache_recurrent::s_mask(int i) const { + const uint32_t cell_id = i + head; + + ////////////////////////////////////////////// + // TODO: this should not mutate the KV cache ! + kv_cell & cell = const_cast(cells[cell_id]); + + float res = (float) (cell.src >= 0); + + // only clear once + if (cell.src < 0) { + cell.src = cell_id; + } + + return res; +} + +uint32_t llama_kv_cache_recurrent::cell_max() const { + for (uint32_t i = size; i > 0; --i) { + const kv_cell & cell = cells[i - 1]; + + if (cell.pos >= 0 && !cell.is_empty()) { + return i; + } + } + + return 0; +} + +size_t llama_kv_cache_recurrent::total_size() const { + size_t size = 0; + for (const auto & buf : bufs) { + size += ggml_backend_buffer_get_size(buf.get()); + } + + return size; +} + +size_t llama_kv_cache_recurrent::size_k_bytes() const { + size_t size_k_bytes = 0; + + for (const auto & k : k_l) { + size_k_bytes += ggml_nbytes(k); + } + + return size_k_bytes; +} + +size_t llama_kv_cache_recurrent::size_v_bytes() const { + size_t size_v_bytes = 0; + + for (const auto & v : v_l) { + size_v_bytes += ggml_nbytes(v); + } + + return size_v_bytes; +} + +void llama_kv_cache_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq_id) const { + std::vector> cell_ranges; // ranges, from inclusive, to exclusive + uint32_t cell_count = 0; + + // Count the number of cells with the specified seq_id + // Find all the ranges of cells with this seq id (or all, when -1) + uint32_t cell_range_begin = size; + for (uint32_t i = 0; i < size; ++i) { + const auto & cell = cells[i]; + if ((seq_id == -1 && !cell.is_empty()) || cell.has_seq_id(seq_id)) { + ++cell_count; + if (cell_range_begin == size) { + cell_range_begin = i; + } + } else { + if (cell_range_begin != size) { + cell_ranges.emplace_back(cell_range_begin, i); + cell_range_begin = size; + } + } + } + if (cell_range_begin != size) { + cell_ranges.emplace_back(cell_range_begin, size); + } + + // DEBUG CHECK: Sum of cell counts in ranges should equal the total cell count + uint32_t cell_count_check = 0; + for (const auto & range : cell_ranges) { + cell_count_check += range.second - range.first; + } + GGML_ASSERT(cell_count == cell_count_check); + + io.write(&cell_count, sizeof(cell_count)); + + state_write_meta(io, cell_ranges, seq_id); + state_write_data(io, cell_ranges); +} + +void llama_kv_cache_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_id) { + uint32_t cell_count; + io.read_to(&cell_count, sizeof(cell_count)); + + bool res = true; + res = res && state_read_meta(io, cell_count, seq_id); + res = res && state_read_data(io, cell_count); + + if (!res) { + if (seq_id == -1) { + clear(); + } else { + seq_rm(seq_id, -1, -1); + } + throw std::runtime_error("failed to restore kv cache"); + } +} + +void llama_kv_cache_recurrent::state_write_meta(llama_io_write_i & io, const std::vector> & cell_ranges, llama_seq_id seq_id) const { + for (const auto & range : cell_ranges) { + for (uint32_t i = range.first; i < range.second; ++i) { + const auto & cell = cells[i]; + const llama_pos pos = cell.pos; + const uint32_t n_seq_id = seq_id == -1 ? cell.seq_id.size() : 0; + + io.write(&pos, sizeof(pos)); + io.write(&n_seq_id, sizeof(n_seq_id)); + + if (n_seq_id) { + for (auto seq_id : cell.seq_id) { + io.write(&seq_id, sizeof(seq_id)); + } + } + } + } +} + +void llama_kv_cache_recurrent::state_write_data(llama_io_write_i & io, const std::vector> & cell_ranges) const { + const uint32_t v_trans = 0; + const uint32_t n_layer = hparams.n_layer; + + io.write(&v_trans, sizeof(v_trans)); + io.write(&n_layer, sizeof(n_layer)); + + std::vector tmp_buf; + + // Iterate and write all the keys first, each row is a cell + // Get whole range at a time + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); + + // Write key type + const int32_t k_type_i = (int32_t)k_l[il]->type; + io.write(&k_type_i, sizeof(k_type_i)); + + // Write row size of key + const uint64_t k_size_row = ggml_row_size(k_l[il]->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 + for (const auto & range : cell_ranges) { + const size_t range_size = range.second - range.first; + const size_t buf_size = range_size * k_size_row; + io.write_tensor(k_l[il], range.first * k_size_row, buf_size); + } + } + + if (!v_trans) { + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + + // Write value type + const int32_t v_type_i = (int32_t)v_l[il]->type; + io.write(&v_type_i, sizeof(v_type_i)); + + // Write row size of value + const uint64_t v_size_row = ggml_row_size(v_l[il]->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 + for (const auto & range : cell_ranges) { + const size_t range_size = range.second - range.first; + const size_t buf_size = range_size * v_size_row; + io.write_tensor(v_l[il], range.first * v_size_row, buf_size); + } + } + } else { + // When v is transposed, we also need the element size and get the element ranges from each row + const uint32_t kv_size = size; + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + + // Write value type + const int32_t v_type_i = (int32_t)v_l[il]->type; + io.write(&v_type_i, sizeof(v_type_i)); + + // Write element size + const uint32_t v_size_el = ggml_type_size(v_l[il]->type); + io.write(&v_size_el, sizeof(v_size_el)); + + // Write GQA embedding size + io.write(&n_embd_v_gqa, sizeof(n_embd_v_gqa)); + + // 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 + for (const auto & range : cell_ranges) { + const size_t range_size = range.second - range.first; + const size_t src_offset = (range.first + j * kv_size) * v_size_el; + const size_t buf_size = range_size * v_size_el; + io.write_tensor(v_l[il], src_offset, buf_size); + } + } + } + } +} + +bool llama_kv_cache_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id) { + if (dest_seq_id != -1) { + // single sequence + + seq_rm(dest_seq_id, -1, -1); + + llama_sbatch sbatch; + llama_ubatch batch = sbatch.reserve_ubatch(cell_count, /* has_embd */ false); + + batch.n_tokens = cell_count; + batch.n_seq_tokens = cell_count; + batch.n_seqs = 1; + + for (uint32_t i = 0; i < cell_count; ++i) { + llama_pos pos; + uint32_t n_seq_id; + + io.read_to(&pos, sizeof(pos)); + io.read_to(&n_seq_id, sizeof(n_seq_id)); + + if (n_seq_id != 0) { + LLAMA_LOG_ERROR("%s: invalid seq_id-agnostic kv cell\n", __func__); + return false; + } + + batch.pos[i] = pos; + } + batch.n_seq_id[0] = 1; + batch.seq_id[0] = &dest_seq_id; + if (!find_slot(batch)) { + LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__); + return false; + } + commit(); + + // DEBUG CHECK: kv.head should be our first cell, kv.head + cell_count - 1 should be our last cell (verify seq_id and pos values) + // Assume that this is one contiguous block of cells + GGML_ASSERT(head + cell_count <= size); + GGML_ASSERT(cells[head].pos == batch.pos[0]); + GGML_ASSERT(cells[head + cell_count - 1].pos == batch.pos[cell_count - 1]); + GGML_ASSERT(cells[head].has_seq_id(dest_seq_id)); + GGML_ASSERT(cells[head + cell_count - 1].has_seq_id(dest_seq_id)); + } else { + // whole KV cache restore + + if (cell_count > size) { + LLAMA_LOG_ERROR("%s: not enough cells in kv cache\n", __func__); + return false; + } + + clear(); + + for (uint32_t i = 0; i < cell_count; ++i) { + kv_cell & cell = cells[i]; + + llama_pos pos; + uint32_t n_seq_id; + + io.read_to(&pos, sizeof(pos)); + io.read_to(&n_seq_id, sizeof(n_seq_id)); + + cell.pos = pos; + + for (uint32_t j = 0; j < n_seq_id; ++j) { + llama_seq_id seq_id; + io.read_to(&seq_id, sizeof(seq_id)); + + // TODO: llama_kv_cache_recurrent should have a notion of max sequences + //if (seq_id < 0 || (uint32_t) seq_id >= llama_n_seq_max(ctx)) { + if (seq_id < 0) { + //LLAMA_LOG_ERROR("%s: invalid seq_id, %d is out of range [0, %u)\n", __func__, seq_id, llama_n_seq_max(ctx)); + LLAMA_LOG_ERROR("%s: invalid seq_id, %d is out of range [0, inf)\n", __func__, seq_id); + return false; + } + + cell.seq_id.insert(seq_id); + + int32_t & tail = cells[seq_id].tail; + if (tail != -1) { + LLAMA_LOG_ERROR("%s: duplicate tail for seq_id %d in cell %d and %d\n", __func__, seq_id, i, tail); + return false; + } + tail = i; + } + } + + head = 0; + used = cell_count; + } + + for (uint32_t i = 0; i < cell_count; ++i) { + uint32_t cell_id = head + i; + // make sure the recurrent states will keep their restored state + cells[cell_id].src = cell_id; + } + + return true; +} + +bool llama_kv_cache_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell_count) { + uint32_t v_trans; + uint32_t n_layer; + io.read_to(&v_trans, sizeof(v_trans)); + io.read_to(&n_layer, sizeof(n_layer)); + + if (n_layer != hparams.n_layer) { + LLAMA_LOG_ERROR("%s: mismatched layer count (%u instead of %u)\n", __func__, n_layer, hparams.n_layer); + return false; + } + if (cell_count > size) { + LLAMA_LOG_ERROR("%s: not enough cells in kv cache to restore state (%u > %u)\n", __func__, cell_count, size); + return false; + } + if (false != (bool) v_trans) { LLAMA_LOG_ERROR("%s: incompatible V transposition\n", __func__); return false; } @@ -1326,7 +2432,7 @@ void llama_kv_cache_view_update(llama_kv_cache_view * view, const llama_kv_cache view->cells_sequences = (llama_seq_id *)p; } - const std::vector & kv_cells = kvu->cells; + const std::vector & kv_cells = kvu->cells; llama_kv_cache_view_cell * c_curr = view->cells; llama_seq_id * cs_curr = view->cells_sequences; int32_t used_cells = 0; diff --git a/src/llama-kv-cache.h b/src/llama-kv-cache.h index 56c74035ae..bf3b4b6a44 100644 --- a/src/llama-kv-cache.h +++ b/src/llama-kv-cache.h @@ -2,32 +2,72 @@ #include "llama.h" #include "llama-io.h" +#include "llama-graph.h" #include "llama-memory.h" #include "ggml-cpp.h" -#include #include #include struct llama_cparams; struct llama_hparams; struct llama_ubatch; +struct llama_sbatch; +struct llama_model; +struct llama_context; struct llama_kv_cache : public llama_memory_i { - using llama_memory_i::llama_memory_i; + virtual ~llama_kv_cache() = default; - virtual void restore() = 0; // call if batch processing fails - restores the cache state - virtual void commit() = 0; // call after successful batch processing - clears any pending state + // call if batch processing fails - restores the cache state + virtual void restore() = 0; - virtual int32_t get_n_tokens() const = 0; - virtual int32_t get_used_cells() const = 0; // TODO: remove, this is too-specific to the unified cache + // call after successful batch processing - clears any pending state + virtual void commit() = 0; - virtual bool get_can_shift() const = 0; + // process any pending defrag/shift/etc. operations + // optionally call once before processing a new batch + virtual bool update(llama_context & lctx) = 0; + + // schedule a defrag if the fragmentation threshold is exceeded. otherwise, do nothing + virtual void defrag_sched(float thold) = 0; + + // simulate full cache, used for allocating worst-case compute buffers + virtual void set_full() = 0; + + // + // batch processing + // + + virtual llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) = 0; + + // different KV caches require different batch splitting strategies + virtual llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const = 0; + + // find an empty slot of size "n_tokens" in the cache + virtual bool find_slot(const llama_ubatch & batch) = 0; + + // getters + virtual int32_t get_n_tokens() const = 0; + virtual int32_t get_used_cells() const = 0; // TODO: remove, this is too-specific to the unified cache + virtual llama_pos get_pos_max() const = 0; + virtual bool get_can_shift() const = 0; bool get_can_edit() const override { return get_can_shift(); } + + // + // state write/read + // + + virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const = 0; + virtual void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1) = 0; }; +// +// llama_kv_cache_guard +// + struct llama_kv_cache_guard { llama_kv_cache_guard(llama_kv_cache * kv) : kv(kv) {} @@ -43,65 +83,50 @@ private: llama_kv_cache * kv; }; -struct llama_kv_cell { - llama_pos pos = -1; - llama_pos delta = 0; - int32_t src = -1; // used by recurrent state models to copy states - int32_t tail = -1; +// +// llama_kv_cache_unified +// - std::set seq_id; - - bool has_seq_id(const llama_seq_id & id) const { - return seq_id.find(id) != seq_id.end(); - } - - bool is_empty() const { - return seq_id.empty(); - } - - bool is_same_seq(const llama_kv_cell & other) const { - return seq_id == other.seq_id; - } -}; - -// ring-buffer of cached KV data -// TODO: pimpl // TODO: add notion of max sequences class llama_kv_cache_unified : public llama_kv_cache { public: - // can be used to query data from the model if needed - struct callbacks { - std::function get_rope_factors; + struct kv_cell { + llama_pos pos = -1; + llama_pos delta = 0; + + std::set seq_id; + + bool has_seq_id(const llama_seq_id & id) const { + return seq_id.find(id) != seq_id.end(); + } + + bool is_empty() const { + return seq_id.empty(); + } + + bool is_same_seq(const kv_cell & other) const { + return seq_id == other.seq_id; + } }; + static uint32_t get_padding(const llama_cparams & cparams); + llama_kv_cache_unified( - const llama_hparams & hparams, - callbacks cbs); - - virtual ~llama_kv_cache_unified() = default; - - // TODO: become constructor - bool init( - const llama_model & model, // TODO: do not reference the model - const llama_cparams & cparams, + const llama_model & model, ggml_type type_k, ggml_type type_v, + bool v_trans, + bool offload, uint32_t kv_size, - bool offload); + uint32_t padding); - int32_t get_n_tokens() const override; - int32_t get_used_cells() const override; + ~llama_kv_cache_unified() = default; - size_t total_size() const; - - // TODO: better data structures to reduce the cost of this operation - llama_pos pos_max() const; + // + // llama_memory_i + // void clear() override; - void defrag() override; - - virtual void restore() override; - virtual void commit() 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; @@ -111,63 +136,40 @@ public: llama_pos seq_pos_max(llama_seq_id seq_id) const override; - bool get_can_shift() const override; + // + // llama_kv_cache + // + + void restore() override; + void commit() override; + + bool update(llama_context & ctx) override; + + void defrag_sched(float thold) override; + + void set_full() override; + + llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override; + + llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override; - // find an empty slot of size "n_tokens" in the cache // updates the cache head // Note: On success, it's important that cache.head points // to the first cell of the slot. - bool find_slot(const llama_ubatch & batch); + bool find_slot(const llama_ubatch & batch) override; - // TODO: maybe not needed - uint32_t get_padding(const llama_cparams & cparams) const; + int32_t get_n_tokens() const override; + int32_t get_used_cells() const override; - // find how many cells are currently in use - uint32_t cell_max() const; + // TODO: better data structures to reduce the cost of this operation + llama_pos get_pos_max() const override; - size_t size_k_bytes() const; - size_t size_v_bytes() const; - - // defrag - - struct { - std::vector ids; - } defrag_info; - - // return true if cells have been moved - bool defrag_prepare(int32_t n_max_nodes); - - // commit/restore cache - - struct slot_range { - uint32_t c0 = 0; // note: these are cell indices, not sequence positions - uint32_t c1 = 0; - }; - - // pending cell updates that are not yet committed - struct { - std::vector ranges; - } pending; + bool get_can_shift() const override; // state write/load - void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const; - void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1); - - // members - - const llama_hparams & hparams; - - callbacks cbs; - - bool has_shift = false; - bool do_defrag = false; - - // TODO: remove this and implement llama_kv_cache_recurrent instead - bool recurrent = false; // with recurrent state models, a cell can hold the state for more than one past token - - bool v_trans = true; // the value tensor is transposed - bool can_shift = false; + void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override; + void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1) override; // Note: The value of head isn't only used to optimize searching // for a free KV slot. llama_decode_impl also uses it, so it @@ -179,18 +181,213 @@ public: // computed before each graph build uint32_t n = 0; - std::vector cells; + std::vector cells; std::vector k_l; // per layer std::vector v_l; private: + const llama_model & model; + const llama_hparams & hparams; + + bool has_shift = false; + bool do_defrag = false; + + bool v_trans = true; // the value tensor is transposed + bool can_shift = false; + + // required padding + uint32_t padding = 1; + ggml_type type_k = GGML_TYPE_F16; ggml_type type_v = GGML_TYPE_F16; std::vector ctxs; std::vector bufs; + // defrag + struct { + std::vector ids; + } defrag_info; + + // return true if cells have been moved + bool defrag_prepare(int32_t n_max_nodes); + + // commit/restore cache + struct slot_range { + uint32_t c0 = 0; // note: these are cell indices, not sequence positions + uint32_t c1 = 0; + }; + + // pending cell updates that are not yet committed + struct { + std::vector ranges; + } pending; + + // find how many cells are currently in use + uint32_t cell_max() const; + + size_t total_size() const; + + size_t size_k_bytes() const; + size_t size_v_bytes() const; + + ggml_tensor * build_rope_shift( + const llama_cparams & cparams, + ggml_context * ctx, + ggml_tensor * cur, + ggml_tensor * shift, + ggml_tensor * factors, + float freq_base, + float freq_scale) const; + + llm_graph_result_ptr build_graph_shift( + const llama_cparams & cparams, + ggml_context * ctx, + ggml_cgraph * gf) const; + + llm_graph_result_ptr build_graph_defrag( + const llama_cparams & cparams, + ggml_context * ctx, + ggml_cgraph * gf) const; + + void state_write_meta(llama_io_write_i & io, const std::vector> & cell_ranges, llama_seq_id seq_id = -1) const; + void state_write_data(llama_io_write_i & io, const std::vector> & cell_ranges) const; + + bool state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id = -1); + bool state_read_data(llama_io_read_i & io, uint32_t cell_count); +}; + +// +// llama_kv_cache_recurrent +// + +class llama_kv_cache_recurrent : public llama_kv_cache { +public: + struct kv_cell { + llama_pos pos = -1; + int32_t src = -1; // used to copy states + int32_t tail = -1; + + std::set seq_id; + + bool has_seq_id(const llama_seq_id & id) const { + return seq_id.find(id) != seq_id.end(); + } + + bool is_empty() const { + return seq_id.empty(); + } + + bool is_same_seq(const kv_cell & other) const { + return seq_id == other.seq_id; + } + }; + + llama_kv_cache_recurrent( + const llama_model & model, + ggml_type type_k, + ggml_type type_v, + bool offload, + uint32_t kv_size); + + ~llama_kv_cache_recurrent() = default; + + // + // llama_memory_i + // + + void clear() 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 delta) override; + void seq_div (llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) override; + + llama_pos seq_pos_max(llama_seq_id seq_id) const override; + + // + // llama_kv_cache + // + + void restore() override; + void commit() override; + + bool update(llama_context & lctx) override; + + void defrag_sched(float thold) override; + + void set_full() override; + + llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override; + + llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override; + + bool find_slot(const llama_ubatch & batch) override; + + int32_t get_n_tokens() const override; + int32_t get_used_cells() const override; + + // TODO: better data structures to reduce the cost of this operation + llama_pos get_pos_max() const override; + + bool get_can_shift() const override; + + // TODO: temporary methods - they are not really const as they do const_cast<>, fix this + int32_t s_copy(int i) const; + float s_mask(int i) const; + + // state write/load + + void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override; + void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1) override; + + // Note: The value of head isn't only used to optimize searching + // for a free KV slot. llama_decode_impl also uses it, so it + // cannot be freely changed after a slot has been allocated. + uint32_t head = 0; + uint32_t size = 0; + uint32_t used = 0; // used cells (i.e. at least one seq_id) + + // computed before each graph build + uint32_t n = 0; + + std::vector cells; + + std::vector k_l; // per layer + std::vector v_l; + +private: + //const llama_model & model; + const llama_hparams & hparams; + + // commit/restore cache + // TODO: rework for recurrent cache + struct slot_range { + uint32_t c0 = 0; // note: these are cell indices, not sequence positions + uint32_t c1 = 0; + }; + + // pending cell updates that are not yet committed + struct { + std::vector ranges; + } pending; + + ggml_type type_k = GGML_TYPE_F16; + ggml_type type_v = GGML_TYPE_F16; + + std::vector ctxs; + std::vector bufs; + + // find how many cells are currently in use + uint32_t cell_max() const; + + size_t total_size() const; + + size_t size_k_bytes() const; + size_t size_v_bytes() const; + void state_write_meta(llama_io_write_i & io, const std::vector> & cell_ranges, llama_seq_id seq_id = -1) const; void state_write_data(llama_io_write_i & io, const std::vector> & cell_ranges) const; @@ -198,11 +395,6 @@ private: bool state_read_data(llama_io_read_i & io, uint32_t cell_count); }; -// TODO: temporary reusing llama_kv_cache_unified -- implement recurrent cache and simplify llama_kv_cache_unified -//class llama_kv_cache_recurrent : public llama_kv_cache_unified { -//public: -// using llama_kv_cache_unified::llama_kv_cache_unified; -//}; // // kv cache view diff --git a/src/llama-memory.h b/src/llama-memory.h index dfa8c4e90f..c7412d5911 100644 --- a/src/llama-memory.h +++ b/src/llama-memory.h @@ -2,12 +2,22 @@ #include "llama.h" +struct llama_memory_params { + // kv cache + ggml_type type_k; + ggml_type type_v; + + // parameters for other types of memory + // ... +}; + // general concept of LLM memory // the KV cache is a type of LLM memory, but there can be other types class llama_memory_i { public: + virtual ~llama_memory_i() = default; + virtual void clear() = 0; - virtual void defrag() = 0; virtual bool seq_rm (llama_seq_id seq_id, llama_pos p0, llama_pos p1) = 0; virtual void seq_cp (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) = 0; diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp index ea73a8a7ba..4cce51668b 100644 --- a/src/llama-model-loader.cpp +++ b/src/llama-model-loader.cpp @@ -301,12 +301,12 @@ namespace GGUFMeta { GGUFMeta::GKV::get_kv(meta.get(), kid); switch (arr_info.gt) { - case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same::value)); break; - case GGUF_TYPE_INT32: GGML_ASSERT( - (std::is_same::value) || - (std::is_same::value)); break; + case GGUF_TYPE_UINT32: + case GGUF_TYPE_INT32: GGML_ASSERT((std::is_same::value) || + (std::is_same::value)); break; + case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same::value)); break; default: - throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str())); + throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str())); } result.resize(arr_info.length); @@ -330,12 +330,12 @@ namespace GGUFMeta { GGUFMeta::GKV::get_kv(meta.get(), kid); switch (arr_info.gt) { - case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same::value)); break; - case GGUF_TYPE_INT32: GGML_ASSERT( - (std::is_same::value) || - (std::is_same::value)); break; + case GGUF_TYPE_UINT32: + case GGUF_TYPE_INT32: GGML_ASSERT((std::is_same::value) || + (std::is_same::value)); break; + case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same::value)); break; default: - throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str())); + throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str())); } if (arr_info.length > N_MAX) { @@ -823,6 +823,10 @@ void llama_model_loader::init_mappings(bool prefetch, llama_mlocks * mlock_mmaps mmaps_used.reserve(files.size()); for (const auto & file : files) { auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU)); + if (!reg) { + throw std::runtime_error(format("%s: no CPU backend found", __func__)); + } + auto * is_numa_fn = (decltype(ggml_is_numa) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_is_numa"); std::unique_ptr mapping = std::make_unique(file.get(), prefetch ? -1 : 0, is_numa_fn()); mmaps_used.emplace_back(mapping->size(), 0); diff --git a/src/llama-model-saver.cpp b/src/llama-model-saver.cpp new file mode 100644 index 0000000000..a70b989234 --- /dev/null +++ b/src/llama-model-saver.cpp @@ -0,0 +1,281 @@ +#include "llama-model-saver.h" + +#include "gguf.h" + +#include "llama.h" +#include "llama-hparams.h" +#include "llama-model.h" +#include "llama-vocab.h" + +#include + +llama_model_saver::llama_model_saver(const struct llama_model & model) : model(model), llm_kv(model.arch) { + gguf_ctx = gguf_init_empty(); +} + +llama_model_saver::~llama_model_saver() { + gguf_free(gguf_ctx); +} + +void llama_model_saver::add_kv(const enum llm_kv key, const uint32_t value) { + gguf_set_val_u32(gguf_ctx, llm_kv(key).c_str(), value); +} + +void llama_model_saver::add_kv(const enum llm_kv key, const int32_t value) { + gguf_set_val_i32(gguf_ctx, llm_kv(key).c_str(), value); +} + +void llama_model_saver::add_kv(const enum llm_kv key, const float value) { + gguf_set_val_f32(gguf_ctx, llm_kv(key).c_str(), value); +} + +void llama_model_saver::add_kv(const enum llm_kv key, const bool value) { + gguf_set_val_bool(gguf_ctx, llm_kv(key).c_str(), value); +} + +void llama_model_saver::add_kv(const enum llm_kv key, const char * value) { + gguf_set_val_str(gguf_ctx, llm_kv(key).c_str(), value); +} + +[[noreturn]] +void llama_model_saver::add_kv(const enum llm_kv key, const char value) { + GGML_UNUSED(key); + GGML_UNUSED(value); + GGML_ABORT("fatal error"); // this should never be called, only needed to make the template below compile +} + +template +void llama_model_saver::add_kv(const enum llm_kv key, const Container & value, const bool per_layer) { + const size_t n_values = per_layer ? size_t(model.hparams.n_layer) : value.size(); + GGML_ASSERT(n_values <= value.size()); + + if (n_values == 0) { + return; + } + + if (per_layer) { + bool all_values_the_same = true; + for (size_t i = 1; i < n_values; ++i) { + if (value[i] != value[0]) { + all_values_the_same = false; + break; + } + } + if (all_values_the_same) { + add_kv(key, value[0]); + return; + } + } + + if (std::is_same::value) { + gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_UINT8, value.data(), n_values); + } else if (std::is_same::value) { + gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_INT8, value.data(), n_values); + } else if (std::is_same::value) { + gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_UINT32, value.data(), n_values); + } else if (std::is_same::value) { + gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_INT32, value.data(), n_values); + } else if (std::is_same::value) { + gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_FLOAT32, value.data(), n_values); + } else if (std::is_same::value) { + gguf_set_val_str(gguf_ctx, llm_kv(key).c_str(), reinterpret_cast(value.data())); + } else { + GGML_ABORT("fatal error"); + } +} + +void llama_model_saver::add_kv(const enum llm_kv key, const std::vector & value) { + std::vector tmp(value.size()); + for (size_t i = 0; i < value.size(); ++i) { + tmp[i] = value[i].c_str(); + } + gguf_set_arr_str(gguf_ctx, llm_kv(key).c_str(), tmp.data(), tmp.size()); +} + +void llama_model_saver::add_tensor(const struct ggml_tensor * tensor) { + if (!tensor) { + return; + } + if (gguf_find_tensor(gguf_ctx, tensor->name) >= 0) { + GGML_ASSERT(std::string(tensor->name) == "rope_freqs.weight"); // FIXME + return; + } + gguf_add_tensor(gguf_ctx, tensor); +} + +void llama_model_saver::add_kv_from_model() { + const llama_hparams & hparams = model.hparams; + const llama_vocab & vocab = model.vocab; + + const int32_t n_vocab = vocab.n_tokens(); + std::vector tokens(n_vocab); + std::vector scores(n_vocab); + std::vector token_types(n_vocab); + + for (int32_t id = 0; id < n_vocab; ++id) { + const llama_vocab::token_data & token_data = vocab.get_token_data(id); + + tokens[id] = token_data.text; + scores[id] = token_data.score; + + switch(token_data.attr) { + case LLAMA_TOKEN_ATTR_UNKNOWN: token_types[id] = LLAMA_TOKEN_TYPE_UNKNOWN; break; + case LLAMA_TOKEN_ATTR_UNUSED: token_types[id] = LLAMA_TOKEN_TYPE_UNUSED; break; + case LLAMA_TOKEN_ATTR_NORMAL: token_types[id] = LLAMA_TOKEN_TYPE_NORMAL; break; + case LLAMA_TOKEN_ATTR_CONTROL: token_types[id] = LLAMA_TOKEN_TYPE_CONTROL; break; + case LLAMA_TOKEN_ATTR_USER_DEFINED: token_types[id] = LLAMA_TOKEN_TYPE_USER_DEFINED; break; + case LLAMA_TOKEN_ATTR_BYTE: token_types[id] = LLAMA_TOKEN_TYPE_BYTE; break; + case LLAMA_TOKEN_ATTR_UNDEFINED: + default: token_types[id] = LLAMA_TOKEN_TYPE_UNDEFINED; break; + } + } + + // add_kv(LLM_KV_GENERAL_TYPE, ???); + add_kv(LLM_KV_GENERAL_ARCHITECTURE, model.arch_name()); + // add_kv(LLM_KV_GENERAL_QUANTIZATION_VERSION, ???); + // add_kv(LLM_KV_GENERAL_ALIGNMENT, ???); + add_kv(LLM_KV_GENERAL_NAME, model.name); + // add_kv(LLM_KV_GENERAL_AUTHOR, ???); + // add_kv(LLM_KV_GENERAL_VERSION, ???); + // add_kv(LLM_KV_GENERAL_URL, ???); + // add_kv(LLM_KV_GENERAL_DESCRIPTION, ???); + // add_kv(LLM_KV_GENERAL_LICENSE, ???); + // add_kv(LLM_KV_GENERAL_SOURCE_URL, ???); + // add_kv(LLM_KV_GENERAL_SOURCE_HF_REPO, ???); + + 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); + add_kv(LLM_KV_BLOCK_COUNT, hparams.n_layer); + add_kv(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead); + add_kv(LLM_KV_FEED_FORWARD_LENGTH, hparams.n_ff_arr, true); + add_kv(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); + add_kv(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_exp); + add_kv(LLM_KV_USE_PARALLEL_RESIDUAL, hparams.use_par_res); + // add_kv(LLM_KV_TENSOR_DATA_LAYOUT, ???); + add_kv(LLM_KV_EXPERT_COUNT, hparams.n_expert); + add_kv(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used); + add_kv(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared); + add_kv(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale); + add_kv(LLM_KV_POOLING_TYPE, uint32_t(hparams.pooling_type)); + add_kv(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale); + add_kv(LLM_KV_DECODER_START_TOKEN_ID, hparams.dec_start_token_id); + add_kv(LLM_KV_ATTN_LOGIT_SOFTCAPPING, hparams.f_attn_logit_softcapping); + add_kv(LLM_KV_FINAL_LOGIT_SOFTCAPPING, hparams.f_final_logit_softcapping); + add_kv(LLM_KV_SWIN_NORM, hparams.swin_norm); + add_kv(LLM_KV_RESCALE_EVERY_N_LAYERS, hparams.rescale_every_n_layers); + add_kv(LLM_KV_TIME_MIX_EXTRA_DIM, hparams.time_mix_extra_dim); + add_kv(LLM_KV_TIME_DECAY_EXTRA_DIM, hparams.time_decay_extra_dim); + add_kv(LLM_KV_RESIDUAL_SCALE, hparams.f_residual_scale); + add_kv(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale); + + add_kv(LLM_KV_ATTENTION_HEAD_COUNT, hparams.n_head_arr, true); + add_kv(LLM_KV_ATTENTION_HEAD_COUNT_KV, hparams.n_head_kv_arr, true); + add_kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS, hparams.f_max_alibi_bias); + add_kv(LLM_KV_ATTENTION_CLAMP_KQV, hparams.f_clamp_kqv); + add_kv(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k); + add_kv(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v); + add_kv(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + add_kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + add_kv(LLM_KV_ATTENTION_CAUSAL, hparams.causal_attn); + add_kv(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q); + add_kv(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv); + add_kv(LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, hparams.n_rel_attn_bkts); + add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa); + add_kv(LLM_KV_ATTENTION_SCALE, hparams.f_attention_scale); + + const float rope_scaling_factor = hparams.rope_freq_scale_train == 1.0f ? 0.0f : 1.0f/hparams.rope_freq_scale_train; + + add_kv(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot); + add_kv(LLM_KV_ROPE_FREQ_BASE, hparams.rope_freq_base_train); + // add_kv(LLM_KV_ROPE_SCALE_LINEAR, rope_scaling_factor); // old name + add_kv(LLM_KV_ROPE_SCALING_TYPE, llama_rope_scaling_type_name(hparams.rope_scaling_type_train)); + add_kv(LLM_KV_ROPE_SCALING_FACTOR, rope_scaling_factor); + add_kv(LLM_KV_ROPE_SCALING_ATTN_FACTOR, hparams.rope_attn_factor); + add_kv(LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, hparams.n_ctx_orig_yarn); + add_kv(LLM_KV_ROPE_SCALING_FINETUNED, hparams.rope_finetuned); + add_kv(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul); + + // TODO: implement split file support + // add_kv(LLM_KV_SPLIT_NO, ???); + // add_kv(LLM_KV_SPLIT_COUNT, ???); + // add_kv(LLM_KV_SPLIT_TENSORS_COUNT, ???); + + add_kv(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner); + add_kv(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv); + add_kv(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state); + add_kv(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank); + add_kv(LLM_KV_SSM_DT_B_C_RMS, hparams.ssm_dt_b_c_rms); + + add_kv(LLM_KV_WKV_HEAD_SIZE, hparams.wkv_head_size); + + add_kv(LLM_KV_TOKENIZER_MODEL, vocab.get_tokenizer_model()); + add_kv(LLM_KV_TOKENIZER_PRE, vocab.get_tokenizer_pre()); + add_kv(LLM_KV_TOKENIZER_LIST, tokens); + add_kv(LLM_KV_TOKENIZER_TOKEN_TYPE, token_types); + add_kv(LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, vocab.n_token_types()); + add_kv(LLM_KV_TOKENIZER_SCORES, scores); + add_kv(LLM_KV_TOKENIZER_MERGES, vocab.get_bpe_merges()); + // FIXME llama_token is type i32 but when reading in a GGUF file u32 is expected, not an issue for writing though + add_kv(LLM_KV_TOKENIZER_BOS_ID, uint32_t(vocab.token_bos())); + add_kv(LLM_KV_TOKENIZER_EOS_ID, uint32_t(vocab.token_eos())); + add_kv(LLM_KV_TOKENIZER_EOT_ID, uint32_t(vocab.token_eot())); + add_kv(LLM_KV_TOKENIZER_EOM_ID, uint32_t(vocab.token_eom())); + add_kv(LLM_KV_TOKENIZER_UNK_ID, uint32_t(vocab.token_unk())); + add_kv(LLM_KV_TOKENIZER_SEP_ID, uint32_t(vocab.token_sep())); + add_kv(LLM_KV_TOKENIZER_PAD_ID, uint32_t(vocab.token_pad())); + // add_kv(LLM_KV_TOKENIZER_CLS_ID, uint32_t(vocab.token_bos())); // deprecated + // add_kv(LLM_KV_TOKENIZER_MASK_ID, ???); + add_kv(LLM_KV_TOKENIZER_ADD_BOS, vocab.get_add_bos()); + add_kv(LLM_KV_TOKENIZER_ADD_EOS, vocab.get_add_eos()); + add_kv(LLM_KV_TOKENIZER_ADD_PREFIX, vocab.get_add_space_prefix()); + add_kv(LLM_KV_TOKENIZER_REMOVE_EXTRA_WS, vocab.get_remove_extra_whitespaces()); + add_kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, vocab.get_precompiled_charsmap()); + // add_kv(LLM_KV_TOKENIZER_HF_JSON, ???); + // add_kv(LLM_KV_TOKENIZER_RWKV, ???); + add_kv(LLM_KV_TOKENIZER_FIM_PRE_ID, uint32_t(vocab.token_fim_pre())); + add_kv(LLM_KV_TOKENIZER_FIM_SUF_ID, uint32_t(vocab.token_fim_suf())); + add_kv(LLM_KV_TOKENIZER_FIM_MID_ID, uint32_t(vocab.token_fim_mid())); + add_kv(LLM_KV_TOKENIZER_FIM_PAD_ID, uint32_t(vocab.token_fim_pad())); + add_kv(LLM_KV_TOKENIZER_FIM_REP_ID, uint32_t(vocab.token_fim_rep())); + add_kv(LLM_KV_TOKENIZER_FIM_SEP_ID, uint32_t(vocab.token_fim_sep())); + + // TODO: implement LoRA support + // add_kv(LLM_KV_ADAPTER_TYPE, ???); + // add_kv(LLM_KV_ADAPTER_LORA_ALPHA, ???); + + // deprecated + // add_kv(LLM_KV_TOKENIZER_PREFIX_ID, ???); + // add_kv(LLM_KV_TOKENIZER_SUFFIX_ID, ???); + // add_kv(LLM_KV_TOKENIZER_MIDDLE_ID, ???); +} + +void llama_model_saver::add_tensors_from_model() { + if (std::string(model.output->name) != std::string(model.tok_embd->name)) { + add_tensor(model.tok_embd); // some models use the same tensor for tok_embd and output + } + add_tensor(model.type_embd); + add_tensor(model.pos_embd); + add_tensor(model.tok_norm); + add_tensor(model.tok_norm_b); + add_tensor(model.output_norm); + add_tensor(model.output_norm_b); + add_tensor(model.output); + add_tensor(model.output_b); + add_tensor(model.output_norm_enc); + add_tensor(model.cls); + add_tensor(model.cls_b); + add_tensor(model.cls_out); + add_tensor(model.cls_out_b); + + for (const struct llama_layer & layer : model.layers) { + for (size_t i = 0; i < sizeof(layer)/sizeof(struct ggml_tensor *); ++i) { + add_tensor(reinterpret_cast(&layer)[i]); + } + } +} + +void llama_model_saver::save(const std::string & path_model) { + gguf_write_to_file(gguf_ctx, path_model.c_str(), false); +} + diff --git a/src/llama-model-saver.h b/src/llama-model-saver.h new file mode 100644 index 0000000000..a5a434c306 --- /dev/null +++ b/src/llama-model-saver.h @@ -0,0 +1,37 @@ +#pragma once + +#include "llama.h" +#include "llama-arch.h" + +#include + +struct llama_model_saver { + struct gguf_context * gguf_ctx = nullptr; + const struct llama_model & model; + const struct LLM_KV llm_kv; + + llama_model_saver(const struct llama_model & model); + ~llama_model_saver(); + + void add_kv(enum llm_kv key, uint32_t value); + void add_kv(enum llm_kv key, int32_t value); + void add_kv(enum llm_kv key, float value); + void add_kv(enum llm_kv key, bool value); + void add_kv(enum llm_kv key, const char * value); + + [[noreturn]] + void add_kv(enum llm_kv key, char value); // needed to make the template below compile + + template + void add_kv(enum llm_kv key, const Container & value, bool per_layer = false); + + void add_kv(enum llm_kv key, const std::vector & value); + + void add_tensor(const struct ggml_tensor * tensor); + + void add_kv_from_model(); + + void add_tensors_from_model(); + + void save(const std::string & path_model); +}; diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 6b7bfecf3a..3a4e72a36b 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -40,14 +40,17 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_335M: return "335M"; case LLM_TYPE_410M: return "410M"; case LLM_TYPE_450M: return "450M"; + case LLM_TYPE_475M: return "475M"; case LLM_TYPE_770M: return "770M"; case LLM_TYPE_780M: return "780M"; case LLM_TYPE_0_5B: return "0.5B"; + case LLM_TYPE_0_6B: return "0.6B"; case LLM_TYPE_1B: return "1B"; case LLM_TYPE_1_3B: return "1.3B"; case LLM_TYPE_1_4B: return "1.4B"; case LLM_TYPE_1_5B: return "1.5B"; case LLM_TYPE_1_6B: return "1.6B"; + case LLM_TYPE_1_7B: return "1.7B"; case LLM_TYPE_1_8B: return "1.8B"; case LLM_TYPE_2B: return "2B"; case LLM_TYPE_2_8B: return "2.8B"; @@ -66,6 +69,7 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_15B: return "15B"; case LLM_TYPE_16B: return "16B"; case LLM_TYPE_20B: return "20B"; + case LLM_TYPE_27B: return "27B"; case LLM_TYPE_30B: return "30B"; case LLM_TYPE_32B: return "32B"; case LLM_TYPE_34B: return "34B"; @@ -74,7 +78,9 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_65B: return "65B"; case LLM_TYPE_70B: return "70B"; case LLM_TYPE_236B: return "236B"; + case LLM_TYPE_290B: return "290B"; case LLM_TYPE_314B: return "314B"; + case LLM_TYPE_405B: return "405B"; case LLM_TYPE_671B: return "671B"; case LLM_TYPE_SMALL: return "0.1B"; case LLM_TYPE_MEDIUM: return "0.4B"; @@ -88,10 +94,10 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_16x3_8B: return "16x3.8B"; case LLM_TYPE_10B_128x3_66B: return "10B+128x3.66B"; case LLM_TYPE_57B_A14B: return "57B.A14B"; - case LLM_TYPE_27B: return "27B"; - case LLM_TYPE_290B: return "290B"; case LLM_TYPE_17B_16E: return "17Bx16E (Scout)"; case LLM_TYPE_17B_128E: return "17Bx128E (Maverick)"; + case LLM_TYPE_30B_A3B: return "30B.A3B"; + case LLM_TYPE_235B_A22B: return "235B.A22B"; default: return "?B"; } } @@ -111,6 +117,10 @@ static const std::map LLAMA_ROPE_SCALING_ { LLAMA_ROPE_SCALING_TYPE_LONGROPE, "longrope" }, }; +std::string llama_rope_scaling_type_name(llama_rope_scaling_type rope_scaling_type) { + return LLAMA_ROPE_SCALING_TYPES.at(rope_scaling_type); +} + static llama_rope_scaling_type llama_rope_scaling_type_from_string(const std::string & name) { for (const auto & kv : LLAMA_ROPE_SCALING_TYPES) { if (kv.second == name) { @@ -293,6 +303,10 @@ static buft_list_t make_cpu_buft_list(const std::vector & de // add extra buffer types, only if no GPU device is present // ref: https://github.com/ggml-org/llama.cpp/issues/12481#issuecomment-2743136094 auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); + if (cpu_dev == nullptr) { + throw std::runtime_error(format("%s: no CPU backend found", __func__)); + } + auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev); auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t) ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts"); @@ -577,6 +591,7 @@ void llama_model::load_hparams(llama_model_loader & ml) { switch (hparams.n_layer) { case 32: type = LLM_TYPE_7B; break; case 80: type = LLM_TYPE_70B; break; + case 162: type = LLM_TYPE_405B; break; default: type = LLM_TYPE_UNKNOWN; } } break; @@ -695,13 +710,19 @@ void llama_model::load_hparams(llama_model_loader & ml) { } } break; case LLM_ARCH_NOMIC_BERT: + case LLM_ARCH_NOMIC_BERT_MOE: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); ml.get_key(LLM_KV_ATTENTION_CAUSAL, hparams.causal_attn); ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type); + ml.get_key(LLM_KV_MOE_EVERY_N_LAYERS, hparams.moe_every_n_layers, 0); if (hparams.n_layer == 12 && hparams.n_embd == 768) { - type = LLM_TYPE_137M; + if (arch == LLM_ARCH_NOMIC_BERT) { + type = LLM_TYPE_137M; + } else if (arch == LLM_ARCH_NOMIC_BERT_MOE && hparams.moe_every_n_layers == 2) { + type = LLM_TYPE_475M; + } } } break; case LLM_ARCH_BLOOM: @@ -762,6 +783,7 @@ void llama_model::load_hparams(llama_model_loader & ml) { // fall through case LLM_ARCH_QWEN2: { + ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type, false); ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); switch (hparams.n_layer) { case 24: type = hparams.n_embd == 1024 ? LLM_TYPE_0_5B : LLM_TYPE_1B; break; @@ -791,6 +813,10 @@ void llama_model::load_hparams(llama_model_loader & ml) { { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); switch (hparams.n_layer) { + case 28: type = hparams.n_embd == 1024 ? LLM_TYPE_0_6B : LLM_TYPE_1_7B; break; + case 36: type = hparams.n_embd == 2560 ? LLM_TYPE_4B : LLM_TYPE_8B; break; + case 40: type = LLM_TYPE_14B; break; + case 64: type = LLM_TYPE_32B; break; default: type = LLM_TYPE_UNKNOWN; } } break; @@ -800,6 +826,8 @@ void llama_model::load_hparams(llama_model_loader & ml) { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); switch (hparams.n_layer) { + case 48: type = LLM_TYPE_30B_A3B; break; + case 94: type = LLM_TYPE_235B_A22B; break; default: type = LLM_TYPE_UNKNOWN; } } break; @@ -1464,6 +1492,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } ggml_backend_dev_t cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); + if (cpu_dev == nullptr) { + throw std::runtime_error(format("%s: no CPU backend found", __func__)); + } const int i_gpu_start = std::max((int) hparams.n_layer - n_gpu_layers, (int) 0); const int act_gpu_layers = devices.empty() ? 0 : std::min(n_gpu_layers, (int)n_layer + 1); auto get_layer_buft_list = [&](int il) -> llama_model::impl::layer_dev { @@ -1631,8 +1662,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) { for (const auto * overrides = ml.tensor_buft_overrides; overrides->pattern != nullptr; ++overrides) { std::regex pattern(overrides->pattern); if (std::regex_search(tensor_name, pattern)) { - LLAMA_LOG_DEBUG("tensor %s buffer type overriden to %s\n", tensor_name.c_str(), ggml_backend_buft_name(overrides->buft)); buft = overrides->buft; + LLAMA_LOG_DEBUG("tensor %s (%zu MiB %s) buffer type overridden to %s\n", + tensor_name.c_str(), + ggml_nbytes(t_meta) / 1024 / 1024, ggml_type_name(t_meta->type), + ggml_backend_buft_name(buft)); break; } } @@ -1649,6 +1683,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) { auto * buft_dev = ggml_backend_buft_get_device(buft); if (ml.use_mmap && buft_dev && buft == ggml_backend_dev_host_buffer_type(buft_dev)) { auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); + if (!cpu_dev) { + throw std::runtime_error("no CPU backend found"); + } buft = ggml_backend_dev_buffer_type(cpu_dev); } @@ -1830,7 +1867,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED); layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED); - layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + if (n_ff > 0) { + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + } if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) { layer.rope_long = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); @@ -1840,9 +1879,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); } - layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); - layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0); - layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + if (n_ff > 0) { + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + } // optional MLP bias layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED); @@ -2057,6 +2098,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } break; case LLM_ARCH_BERT: case LLM_ARCH_NOMIC_BERT: + case LLM_ARCH_NOMIC_BERT_MOE: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); type_embd = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, 0); @@ -2090,20 +2132,31 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0); } + if (arch == LLM_ARCH_NOMIC_BERT_MOE) { + layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0); + } + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0); layer.attn_out_norm = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0); layer.attn_out_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i), {n_embd}, 0); - layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); - layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0); - - if (arch == LLM_ARCH_BERT) { + if (hparams.moe_every_n_layers > 0 && i % hparams.moe_every_n_layers == 1) { layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0); - layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, 0); - layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, 0); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff, n_expert}, 0); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff, n_embd, n_expert}, 0); + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0); } else { - layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0); + + if (arch == LLM_ARCH_BERT || arch == LLM_ARCH_NOMIC_BERT_MOE) { + layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0); + layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, 0); + layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, 0); + } else { + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + } } layer.layer_out_norm = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0); @@ -3474,7 +3527,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) { // output output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); - output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } for (int i = 0; i < n_layer; ++i) { auto & layer = layers[i]; @@ -4079,6 +4136,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) { if (!dev) { // FIXME: workaround for CPU backend buft having a NULL device dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); + if (!dev) { + throw std::runtime_error(format("%s: no CPU backend found", __func__)); + } } ggml_backend_dev_props props; ggml_backend_dev_get_props(dev, &props); @@ -4208,7 +4268,7 @@ uint64_t llama_model::n_elements() const { } void llama_model::print_info() const { - const char * rope_scaling_type = LLAMA_ROPE_SCALING_TYPES.at(hparams.rope_scaling_type_train); + const std::string rope_scaling_type = llama_rope_scaling_type_name(hparams.rope_scaling_type_train); auto print_f = [](const std::function & f, uint32_t n) { bool is_var = false; @@ -4269,7 +4329,7 @@ void llama_model::print_info() const { LLAMA_LOG_INFO("%s: causal attn = %d\n", __func__, hparams.causal_attn); LLAMA_LOG_INFO("%s: pooling type = %d\n", __func__, hparams.pooling_type); LLAMA_LOG_INFO("%s: rope type = %d\n", __func__, hparams.rope_type); - LLAMA_LOG_INFO("%s: rope scaling = %s\n", __func__, rope_scaling_type); + LLAMA_LOG_INFO("%s: rope scaling = %s\n", __func__, rope_scaling_type.c_str()); LLAMA_LOG_INFO("%s: freq_base_train = %.1f\n", __func__, hparams.rope_freq_base_train); LLAMA_LOG_INFO("%s: freq_scale_train = %g\n", __func__, hparams.rope_freq_scale_train); LLAMA_LOG_INFO("%s: n_ctx_orig_yarn = %u\n", __func__, hparams.n_ctx_orig_yarn); @@ -4416,6 +4476,19 @@ const ggml_tensor * llama_model::get_tensor(const char * name) const { return it->second; } +ggml_tensor * llama_model::get_rope_factors(uint32_t n_ctx_per_seq, int il) const { + // choose long/short freq factors based on the context size + if (layers[il].rope_freqs != nullptr) { + return layers[il].rope_freqs; + } + + if (n_ctx_per_seq > hparams.n_ctx_orig_yarn) { + return layers[il].rope_long; + } + + return layers[il].rope_short; +} + struct llm_build_llama : public llm_graph_context { llm_build_llama(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { const int64_t n_embd_head = hparams.n_embd_head_v; @@ -4456,7 +4529,7 @@ struct llm_build_llama : public llm_graph_context { // self-attention { // rope freq factors for llama3; may return nullptr for llama2 and other models - ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il); // compute Q and K and RoPE them ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); @@ -4662,6 +4735,7 @@ struct llm_build_deci : public llm_graph_context { ggml_tensor * inpSA = inpL; const int64_t n_head_kv = hparams.n_head_kv(il); const int64_t n_head = hparams.n_head(il); + const int64_t n_ff = hparams.n_ff(il); if (n_head == 0) { // attention-free layer of Llama-3_1-Nemotron-51B @@ -4681,7 +4755,7 @@ struct llm_build_deci : public llm_graph_context { } else if (n_head > 0) { // self-attention // rope freq factors for llama3; may return nullptr for llama2 and other models - ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il); // compute Q and K and RoPE them ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); @@ -4737,6 +4811,11 @@ struct llm_build_deci : public llm_graph_context { inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } + // FFN-free layer of Llama-3_1-Nemotron-Ultra-253B + if (n_ff == 0) { + continue; + } + // For Granite architecture if (hparams.f_residual_scale) { cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); @@ -5730,6 +5809,11 @@ struct llm_build_bert : public llm_graph_context { cur = build_lora_mm(model.layers[il].wqkv, cur); cb(cur, "wqkv", il); + if (model.arch == LLM_ARCH_NOMIC_BERT_MOE) { + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + } + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); @@ -5782,13 +5866,29 @@ struct llm_build_bert : public llm_graph_context { cb(ffn_inp, "ffn_inp", il); // feed-forward network - if (model.arch == LLM_ARCH_BERT) { + if (hparams.moe_every_n_layers > 0 && il % hparams.moe_every_n_layers == 1) { + // MoE branch + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + nullptr, + model.layers[il].ffn_down_exps, + nullptr, + hparams.n_expert, + hparams.n_expert_used, + LLM_FFN_GELU, + false, false, + 0.0f, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il); + cb(cur, "ffn_moe_out", il); + } else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE) { cur = build_ffn(cur, model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, NULL, NULL, NULL, model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL, LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); } else if (model.arch == LLM_ARCH_JINA_BERT_V2) { cur = build_ffn(cur, model.layers[il].ffn_up, NULL, NULL, @@ -5796,6 +5896,7 @@ struct llm_build_bert : public llm_graph_context { model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL, LLM_FFN_GELU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); } else { cur = build_ffn(cur, model.layers[il].ffn_up, NULL, NULL, @@ -5803,8 +5904,8 @@ struct llm_build_bert : public llm_graph_context { model.layers[il].ffn_down, NULL, NULL, NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); } - cb(cur, "ffn_out", il); // attentions bypass the intermediate layer cur = ggml_add(ctx0, cur, ffn_inp); @@ -7141,7 +7242,7 @@ struct llm_build_phi3 : public llm_graph_context { // self-attention { // rope freq factors for 128k context - ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il); ggml_tensor* attn_norm_output = build_norm(inpL, model.layers[il].attn_norm, @@ -7893,7 +7994,7 @@ struct llm_build_minicpm3 : public llm_graph_context { for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; - ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il); // norm cur = build_norm(inpL, @@ -8660,7 +8761,7 @@ struct llm_build_mamba : public llm_graph_context { ggml_tensor * state_mask, const llama_ubatch & ubatch, int il) const { - const llama_kv_cache_unified * kv_self = static_cast(memory); + const llama_kv_cache_recurrent * kv_self = static_cast(memory); const auto kv_head = kv_self->head; @@ -8961,7 +9062,7 @@ struct llm_build_cohere2 : public llm_graph_context { // self-attention { // rope freq factors for 128k context - ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il); // compute Q and K and RoPE them ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); @@ -9899,7 +10000,7 @@ struct llm_build_deepseek : public llm_graph_context { // self-attention { // rope freq factors for llama3; may return nullptr for llama2 and other models - ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il); // compute Q and K and RoPE them ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); @@ -11263,7 +11364,7 @@ struct llm_build_exaone : public llm_graph_context { // self-attention { // rope freq factors for llama3; may return nullptr for llama2 and other models - ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il); // compute Q and K and RoPE them ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); @@ -11408,7 +11509,7 @@ struct llm_build_rwkv6_base : public llm_graph_context { ggml_tensor * state_mask, const llama_ubatch & ubatch, int il) const { - const llama_kv_cache_unified * kv_self = static_cast(memory); + const llama_kv_cache_recurrent * kv_self = static_cast(memory); const auto n_tokens = ubatch.n_tokens; const auto n_seqs = ubatch.n_seqs; @@ -11804,7 +11905,7 @@ struct llm_build_rwkv7_base : public llm_graph_context { ggml_tensor *& first_layer_value, const llama_ubatch & ubatch, int il) const { - const llama_kv_cache_unified * kv_self = static_cast(memory); + const llama_kv_cache_recurrent * kv_self = static_cast(memory); const auto n_tokens = ubatch.n_tokens; const auto n_seqs = ubatch.n_seqs; @@ -12644,7 +12745,7 @@ struct llm_build_bailingmoe : public llm_graph_context { // self-attention { // rope freq factors for llama3; may return nullptr for llama2 and other models - ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il); // compute Q and K and RoPE them ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); @@ -12764,36 +12865,46 @@ struct llm_build_bailingmoe : public llm_graph_context { } }; -llama_memory_i * llama_model::create_memory() const { +llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const { llama_memory_i * res; switch (arch) { + case LLM_ARCH_BERT: + case LLM_ARCH_JINA_BERT_V2: + case LLM_ARCH_NOMIC_BERT: + case LLM_ARCH_NOMIC_BERT_MOE: + { + res = nullptr; + } break; case LLM_ARCH_MAMBA: case LLM_ARCH_RWKV6: case LLM_ARCH_RWKV6QWEN2: case LLM_ARCH_RWKV7: case LLM_ARCH_ARWKV7: { - res = new llama_kv_cache_unified(hparams, { - /*.get_rope_factors =*/ nullptr - }); + res = new llama_kv_cache_recurrent( + *this, + GGML_TYPE_F32, + GGML_TYPE_F32, + cparams.offload_kqv, + std::max((uint32_t) 1, cparams.n_seq_max)); } break; default: { - res = new llama_kv_cache_unified(hparams, { - /*.get_rope_factors =*/ [this](uint32_t n_ctx_per_seq, int il) { - // choose long/short freq factors based on the context size - if (layers[il].rope_freqs != nullptr) { - return layers[il].rope_freqs; - } + const auto padding = llama_kv_cache_unified::get_padding(cparams); - if (n_ctx_per_seq > hparams.n_ctx_orig_yarn) { - return layers[il].rope_long; - } + cparams.n_ctx = GGML_PAD(cparams.n_ctx, padding); - return layers[il].rope_short; - } - }); + LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx); + + res = new llama_kv_cache_unified( + *this, + params.type_k, + params.type_v, + !cparams.flash_attn, + cparams.offload_kqv, + cparams.n_ctx, + padding); } } @@ -12842,6 +12953,7 @@ llm_graph_result_ptr llama_model::build_graph( case LLM_ARCH_BERT: case LLM_ARCH_JINA_BERT_V2: case LLM_ARCH_NOMIC_BERT: + case LLM_ARCH_NOMIC_BERT_MOE: { llm = std::make_unique(*this, params, gf); } break; @@ -13174,8 +13286,6 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_DECI: case LLM_ARCH_BAICHUAN: case LLM_ARCH_STARCODER: - case LLM_ARCH_PLAMO: - case LLM_ARCH_ORION: case LLM_ARCH_INTERNLM2: case LLM_ARCH_MINICPM: case LLM_ARCH_XVERSE: @@ -13200,6 +13310,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_DBRX: case LLM_ARCH_BERT: case LLM_ARCH_NOMIC_BERT: + case LLM_ARCH_NOMIC_BERT_MOE: case LLM_ARCH_STABLELM: case LLM_ARCH_BITNET: case LLM_ARCH_QWEN: @@ -13212,6 +13323,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_PHI2: case LLM_ARCH_PHI3: case LLM_ARCH_PHIMOE: + case LLM_ARCH_PLAMO: case LLM_ARCH_GEMMA: case LLM_ARCH_GEMMA2: case LLM_ARCH_GEMMA3: @@ -13219,6 +13331,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_OPENELM: case LLM_ARCH_GPTNEOX: case LLM_ARCH_CODESHELL: + case LLM_ARCH_ORION: case LLM_ARCH_NEMOTRON: case LLM_ARCH_EXAONE: case LLM_ARCH_MINICPM3: @@ -13291,6 +13404,14 @@ const char * llama_model_chat_template(const llama_model * model, const char * n : LLM_KV(model->arch)(LLM_KV_TOKENIZER_CHAT_TEMPLATE); const auto & it = model->gguf_kv.find(key); if (it == model->gguf_kv.end()) { + // one-off fix for very popular models (so we are not flooded with issues) + // do not extend this list unless absolutely necessary + // Mistral-Small-2503 does not have built-in chat template + llama_vocab_pre_type pre_type = model->vocab.get_pre_type(); + if (pre_type == LLAMA_VOCAB_PRE_TYPE_TEKKEN && model->layers.size() == 40) { + return "mistral-v7-tekken"; + } + return nullptr; } diff --git a/src/llama-model.h b/src/llama-model.h index fd82d106cc..6bdec263b7 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -36,14 +36,17 @@ enum llm_type { LLM_TYPE_335M, LLM_TYPE_410M, LLM_TYPE_450M, + LLM_TYPE_475M, LLM_TYPE_770M, LLM_TYPE_780M, LLM_TYPE_0_5B, + LLM_TYPE_0_6B, LLM_TYPE_1B, LLM_TYPE_1_3B, LLM_TYPE_1_4B, LLM_TYPE_1_5B, LLM_TYPE_1_6B, + LLM_TYPE_1_7B, LLM_TYPE_1_8B, LLM_TYPE_2B, LLM_TYPE_2_8B, @@ -62,6 +65,7 @@ enum llm_type { LLM_TYPE_15B, LLM_TYPE_16B, LLM_TYPE_20B, + LLM_TYPE_27B, LLM_TYPE_30B, LLM_TYPE_32B, LLM_TYPE_34B, @@ -70,7 +74,9 @@ enum llm_type { LLM_TYPE_65B, LLM_TYPE_70B, LLM_TYPE_236B, + LLM_TYPE_290B, LLM_TYPE_314B, + LLM_TYPE_405B, LLM_TYPE_671B, LLM_TYPE_SMALL, LLM_TYPE_MEDIUM, @@ -84,12 +90,14 @@ enum llm_type { LLM_TYPE_16x3_8B, LLM_TYPE_10B_128x3_66B, LLM_TYPE_57B_A14B, - LLM_TYPE_27B, - LLM_TYPE_290B, LLM_TYPE_17B_16E, // llama4 Scout LLM_TYPE_17B_128E, // llama4 Maverick + LLM_TYPE_30B_A3B, + LLM_TYPE_235B_A22B, }; +std::string llama_rope_scaling_type_name(llama_rope_scaling_type rope_scaling_type); + struct llama_layer_posnet { // resnet struct ggml_tensor * norm1 = nullptr; @@ -390,8 +398,11 @@ struct llama_model { const struct ggml_tensor * get_tensor(const char * name) const; + ggml_tensor * get_rope_factors(uint32_t n_ctx_per_seq, int il) const; + + // note: can mutate `cparams` // TODO: move this to new llm_arch_model_i interface - llama_memory_i * create_memory() const; // TODO: params + llama_memory_i * create_memory(const llama_memory_params & params, llama_cparams & cparams) const; // TODO: move this to new llm_arch_model_i interface llm_graph_result_ptr build_graph( diff --git a/src/llama-quant.cpp b/src/llama-quant.cpp index 7dc5422763..820d5128e2 100644 --- a/src/llama-quant.cpp +++ b/src/llama-quant.cpp @@ -519,7 +519,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std:: nthread = std::thread::hardware_concurrency(); } - // mmap consistently increases speed Linux, and also increases speed on Windows with + // mmap consistently increases speed on Linux, and also increases speed on Windows with // hot cache. It may cause a slowdown on macOS, possibly related to free memory. #if defined(__linux__) || defined(_WIN32) constexpr bool use_mmap = true; @@ -529,7 +529,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std:: llama_model_kv_override * kv_overrides = nullptr; if (params->kv_overrides) { - auto v = (std::vector*)params->kv_overrides; + auto * v = (std::vector*)params->kv_overrides; kv_overrides = v->data(); } diff --git a/src/llama-sampling.cpp b/src/llama-sampling.cpp index d149798502..804b11e0a9 100644 --- a/src/llama-sampling.cpp +++ b/src/llama-sampling.cpp @@ -232,7 +232,7 @@ static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k) // } if (k <= 0) { - k = cur_p->size; + return; } k = std::min(k, (int) cur_p->size); @@ -298,6 +298,7 @@ static void llama_sampler_top_k_impl(llama_token_data_array * cur_p, int32_t k) } cur_p->sorted = true; } + cur_p->size = k; } @@ -1749,23 +1750,35 @@ static const char * llama_sampler_top_n_sigma_name(const struct llama_sampler * static void llama_sampler_top_n_sigma_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { const auto * ctx = (llama_sampler_top_n_sigma *) smpl->ctx; + if (ctx->n <= 0.0f || cur_p->size <= 1) { + return; + } + // find max logit and calculate mean float max = cur_p->data[0].logit; float logits_sum = 0; + size_t valid_count = 0; for (size_t i = 0; i < cur_p->size; ++i) { - if (cur_p->data[i].logit > max) { - max = cur_p->data[i].logit; + // Only count non-negative infinity values + if (cur_p->data[i].logit != -INFINITY) { + if (cur_p->data[i].logit > max) { + max = cur_p->data[i].logit; + } + logits_sum += cur_p->data[i].logit; + valid_count++; } - logits_sum += cur_p->data[i].logit; } - float mean = logits_sum/cur_p->size; + float mean = valid_count > 0 ? logits_sum/valid_count : 0; // calculate standard deviation float acc = 0; for (size_t i = 0; i < cur_p->size; ++i) { - acc += pow(cur_p->data[i].logit - mean, 2); + // Skip -infinity in std calculation + if (cur_p->data[i].logit != -INFINITY) { + acc += pow(cur_p->data[i].logit - mean, 2); + } } - float std = sqrt(acc/cur_p->size); + float std = valid_count > 0 ? sqrt(acc/valid_count) : 0; //apply mask for (size_t i = 0; i < cur_p->size; ++i) { diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp index 50ded286f3..9389ca805a 100644 --- a/src/llama-vocab.cpp +++ b/src/llama-vocab.cpp @@ -1,5 +1,7 @@ #include "llama-vocab.h" +#include "ggml.h" +#include "gguf.h" #include "llama-impl.h" #include "llama-model-loader.h" @@ -415,6 +417,13 @@ struct llm_tokenizer_bpe : llm_tokenizer { "'(?:[sSdDmMtT]|[lL][lL]|[vV][eE]|[rR][eE])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+", }; break; + case LLAMA_VOCAB_PRE_TYPE_SEED_CODER: + regex_exprs = { + // original regex from tokenizer.json + // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1}| ?[^\\s\\p{L}\\p{N}\r\n]+|\\s*[\r\n]+|\\s+(?!\\S)|\\s+" + "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1}| ?[^\\s\\p{L}\\p{N}\\r\\n]+|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }; + break; default: // default regex for BPE tokenization pre-processing regex_exprs = { @@ -1227,6 +1236,9 @@ struct fragment_buffer_variant { struct llama_vocab::impl { uint32_t n_token_types = 0; // for BERT-style token types + std::string tokenizer_model; + std::string tokenizer_pre; + enum llama_vocab_type type = LLAMA_VOCAB_TYPE_SPM; enum llama_vocab_pre_type pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT; @@ -1362,9 +1374,6 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { // determine vocab type { - std::string tokenizer_model; - std::string tokenizer_pre; - ml.get_key(LLM_KV_TOKENIZER_MODEL, tokenizer_model); ml.get_key(LLM_KV_TOKENIZER_PRE, tokenizer_pre, false); @@ -1459,7 +1468,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { const int precompiled_charsmap_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP).c_str()); if (precompiled_charsmap_keyidx != -1) { - size_t n_precompiled_charsmap = gguf_get_arr_n(ctx, precompiled_charsmap_keyidx); + const gguf_type pc_type = gguf_get_arr_type(ctx, precompiled_charsmap_keyidx); + GGML_ASSERT(pc_type == GGUF_TYPE_INT8 || pc_type == GGUF_TYPE_UINT8); + + const size_t n_precompiled_charsmap = gguf_get_arr_n(ctx, precompiled_charsmap_keyidx); const char * pc = (const char *) gguf_get_arr_data(ctx, precompiled_charsmap_keyidx); precompiled_charsmap.assign(pc, pc + n_precompiled_charsmap); #ifdef IS_BIG_ENDIAN @@ -1634,6 +1646,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { tokenizer_pre == "bailingmoe") { pre_type = LLAMA_VOCAB_PRE_TYPE_BAILINGMOE; clean_spaces = false; + } else if ( + tokenizer_pre == "seed-coder") { + pre_type = LLAMA_VOCAB_PRE_TYPE_SEED_CODER; + clean_spaces = false; } else { throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str())); } @@ -2778,6 +2794,14 @@ void llama_vocab::load(llama_model_loader & ml, const LLM_KV & kv) { pimpl->load(ml, kv); } +std::string llama_vocab::get_tokenizer_model() const { + return pimpl->tokenizer_model; +} + +std::string llama_vocab::get_tokenizer_pre() const { + return pimpl->tokenizer_pre; +} + enum llama_vocab_type llama_vocab::get_type() const { return pimpl->type; } @@ -3000,6 +3024,20 @@ int llama_vocab::find_bpe_rank(const std::string & token_left, const std::string return it->second; } +std::vector llama_vocab::get_bpe_merges() const { + std::vector result(pimpl->bpe_ranks.size()); + + for (const auto & pair : pimpl->bpe_ranks) { + result[pair.second] = pair.first.first + " " + pair.first.second; + } + + return result; +} + +std::vector llama_vocab::get_precompiled_charsmap() const { + return pimpl->precompiled_charsmap; +} + int32_t llama_vocab::tokenize( const char * text, int32_t text_len, diff --git a/src/llama-vocab.h b/src/llama-vocab.h index 5ce3552143..daa6cf3082 100644 --- a/src/llama-vocab.h +++ b/src/llama-vocab.h @@ -21,6 +21,9 @@ struct llama_vocab { void load(llama_model_loader & ml, const LLM_KV & kv); + std::string get_tokenizer_model() const; + std::string get_tokenizer_pre() const; + enum llama_vocab_type get_type() const; enum llama_vocab_pre_type get_pre_type() const; @@ -80,6 +83,9 @@ struct llama_vocab { int max_token_len() const; int find_bpe_rank(const std::string & token_left, const std::string & token_right) const; + std::vector get_bpe_merges() const; + + std::vector get_precompiled_charsmap() const; int32_t tokenize( const char * text, diff --git a/src/llama.cpp b/src/llama.cpp index d5164720b2..9fdddf7b07 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -4,6 +4,7 @@ #include "llama-mmap.h" #include "llama-vocab.h" #include "llama-model-loader.h" +#include "llama-model-saver.h" #include "llama-model.h" #include "ggml.h" @@ -253,6 +254,13 @@ struct llama_model * llama_model_load_from_splits( return llama_model_load_from_file_impl(splits.front(), splits, params); } +void llama_model_save_to_file(const struct llama_model * model, const char * path_model) { + llama_model_saver ms(*model); + ms.add_kv_from_model(); + ms.add_tensors_from_model(); + ms.save(path_model); +} + // // chat templates // @@ -338,3 +346,4 @@ const char * llama_print_system_info(void) { return s.c_str(); } + diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index ae68275251..709d5ad96a 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -111,10 +111,13 @@ if (NOT WIN32) # TODO: disabled on loongarch64 because the ggml-ci node lacks Python 3.8 if (NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64") llama_build_and_test(test-json-schema-to-grammar.cpp WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/..) - target_include_directories(test-json-schema-to-grammar PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../examples/server) + target_include_directories(test-json-schema-to-grammar PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../tools/server) + endif() + + if (NOT GGML_BACKEND_DL) + llama_build(test-quantize-stats.cpp) endif() - llama_build(test-quantize-stats.cpp) llama_build(test-gbnf-validator.cpp) # build test-tokenizer-1-bpe target once and add many tests @@ -162,6 +165,10 @@ if (NOT GGML_BACKEND_DL) llama_build_and_test(test-rope.cpp) endif() +# libmtmd +set(LLAMA_TEST_NAME test-mtmd-c-api) +llama_build_and_test(test-mtmd-c-api.c) +target_link_libraries(${LLAMA_TEST_NAME} PRIVATE mtmd) # dummy executable - not installed get_filename_component(TEST_TARGET test-c.c NAME_WE) diff --git a/tests/run-json-schema-to-grammar.mjs b/tests/run-json-schema-to-grammar.mjs index b20ac1d6b5..450c3dde0a 100644 --- a/tests/run-json-schema-to-grammar.mjs +++ b/tests/run-json-schema-to-grammar.mjs @@ -1,5 +1,5 @@ import { readFileSync } from "fs" -import { SchemaConverter } from "../examples/server/public_legacy/json-schema-to-grammar.mjs" +import { SchemaConverter } from "../tools/server/public_legacy/json-schema-to-grammar.mjs" const [, , file] = process.argv const url = `file://${file}` diff --git a/tests/test-arg-parser.cpp b/tests/test-arg-parser.cpp index 537fc63a4c..21dbd54042 100644 --- a/tests/test-arg-parser.cpp +++ b/tests/test-arg-parser.cpp @@ -126,6 +126,53 @@ int main(void) { assert(params.cpuparams.n_threads == 1010); #endif // _WIN32 + if (common_has_curl()) { + printf("test-arg-parser: test curl-related functions\n\n"); + const char * GOOD_URL = "https://raw.githubusercontent.com/ggml-org/llama.cpp/refs/heads/master/README.md"; + const char * BAD_URL = "https://www.google.com/404"; + const char * BIG_FILE = "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v1.bin"; + + { + printf("test-arg-parser: test good URL\n\n"); + auto res = common_remote_get_content(GOOD_URL, {}); + assert(res.first == 200); + assert(res.second.size() > 0); + std::string str(res.second.data(), res.second.size()); + assert(str.find("llama.cpp") != std::string::npos); + } + + { + printf("test-arg-parser: test bad URL\n\n"); + auto res = common_remote_get_content(BAD_URL, {}); + assert(res.first == 404); + } + + { + printf("test-arg-parser: test max size error\n"); + common_remote_params params; + params.max_size = 1; + try { + common_remote_get_content(GOOD_URL, params); + assert(false && "it should throw an error"); + } catch (std::exception & e) { + printf(" expected error: %s\n\n", e.what()); + } + } + + { + printf("test-arg-parser: test timeout error\n"); + common_remote_params params; + params.timeout = 1; + try { + common_remote_get_content(BIG_FILE, params); + assert(false && "it should throw an error"); + } catch (std::exception & e) { + printf(" expected error: %s\n\n", e.what()); + } + } + } else { + printf("test-arg-parser: no curl, skipping curl-related functions\n"); + } printf("test-arg-parser: all tests OK\n\n"); } diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index d70acb7719..543db93402 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -823,7 +823,7 @@ struct test_case { ggml_build_forward_expand(gf, out); ggml_graph_cpy(gf, gb); - ggml_build_backward_expand(ctx.get(), ctx.get(), gb, false); + ggml_build_backward_expand(ctx.get(), gb, nullptr); if (expect.size() != 1 || expect[0] != 0.0f) { GGML_ASSERT(ggml_graph_n_nodes(gb) > ggml_graph_n_nodes(gf)); for (ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != NULL; t = ggml_get_next_tensor(ctx.get(), t)) { @@ -1026,7 +1026,7 @@ struct test_example : public test_case { // Step 3: return the output tensor. return out; } - // In order to also check the gradients for your op, add calls like ggml_set_param(ctx, a) + // In order to also check the gradients for your op, add calls like ggml_set_param(a) // immediately after you create the tensors. // This is optional and only makes sense if a backward pass has actually been implemented for the new op. }; @@ -1058,7 +1058,7 @@ struct test_unary : public test_case { auto ne = ne_a; ne[0] *= 3; a = ggml_new_tensor(ctx, type, 4, ne.data()); if (grad_supported) { - ggml_set_param(ctx, a); + ggml_set_param(a); } ggml_set_name(a, "a"); @@ -1067,7 +1067,7 @@ struct test_unary : public test_case { } else { a = ggml_new_tensor(ctx, type, 4, ne_a.data()); if (grad_supported) { - ggml_set_param(ctx, a); + ggml_set_param(a); } ggml_set_name(a, "a"); } @@ -1133,7 +1133,7 @@ struct test_get_rows : public test_case { const bool grad_supported = ggml_is_matrix(in) && ggml_is_vector(rows); if (grad_supported) { - ggml_set_param(ctx, in); + ggml_set_param(in); // rows is a constant input -> no gradients } @@ -1322,7 +1322,7 @@ struct test_repeat : public test_case { ggml_set_name(target, "target"); ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_set_param(ctx, src); + ggml_set_param(src); ggml_set_name(src, "src"); ggml_tensor * out = ggml_repeat(ctx, src, target); @@ -1406,7 +1406,7 @@ struct test_dup : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_set_param(ctx, src); + ggml_set_param(src); ggml_set_name(src, "src"); if (_use_permute) { @@ -1442,7 +1442,7 @@ struct test_set : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data()); - ggml_set_param(ctx, src); + ggml_set_param(src); ggml_set_name(src, "src"); auto ne_dst = ne; @@ -1450,7 +1450,7 @@ struct test_set : public test_case { ne_dst[i] *= 2; } ggml_tensor* dst = ggml_new_tensor(ctx, type_dst, 4, ne_dst.data()); - ggml_set_param(ctx, dst); + ggml_set_param(dst); ggml_set_name(dst, "dst"); size_t offset = 0; @@ -1498,7 +1498,7 @@ struct test_cpy : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data()); - ggml_set_param(ctx, src); + ggml_set_param(src); ggml_set_name(src, "src"); if (_src_use_permute) { @@ -1536,7 +1536,7 @@ struct test_cont : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_set_param(ctx, src); + ggml_set_param(src); ggml_set_name(src, "src"); src = ggml_transpose(ctx, src); @@ -1583,8 +1583,8 @@ struct test_bin_bcast : public test_case { // The backward pass supports broadcasting only for GGML_ADD: const bool grad_supported = op == ggml_add || ggml_are_same_shape(a, b); if (grad_supported) { - ggml_set_param(ctx, a); - ggml_set_param(ctx, b); + ggml_set_param(a); + ggml_set_param(b); } ggml_tensor * out = op(ctx, a, b); @@ -1632,11 +1632,11 @@ struct test_add1 : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_set_param(ctx, a); + ggml_set_param(a); ggml_set_name(a, "a"); ggml_tensor * b = ggml_new_tensor_1d(ctx, type, 1); - // ggml_set_param(ctx, b); // TODO: implement + // ggml_set_param(b); // TODO: implement ggml_set_name(b, "b"); ggml_tensor * out = ggml_add1(ctx, a, b); @@ -1667,7 +1667,7 @@ struct test_scale : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_set_param(ctx, a); + ggml_set_param(a); ggml_set_name(a, "a"); ggml_tensor * out = ggml_scale(ctx, a, scale); @@ -1762,7 +1762,7 @@ struct test_rms_norm : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_set_param(ctx, a); + ggml_set_param(a); ggml_set_name(a, "a"); if (v) { @@ -1981,7 +1981,7 @@ struct test_mul_mat : public test_case { const std::array bs; // dims 3 and 4 const std::array nr; // repeat in dims 3 and 4 const std::array per; // permutation of dimensions - const bool v; // whether a is a non-contiguous view + const bool v; // whether a and b are non-contiguous views std::string vars() override { return VARS_TO_STR9(type_a, type_b, m, n, k, bs, nr, per, v); @@ -2028,9 +2028,9 @@ struct test_mul_mat : public test_case { b = ggml_new_tensor_4d(ctx, type_b, ne_b[per[0]], ne_b[per[1]], ne_b[per[2]], ne_b[per[3]]); if (!ggml_is_quantized(type_a)) { if (bs[1] == 1 && nr[1] == 1) { - ggml_set_param(ctx, a); + ggml_set_param(a); } - ggml_set_param(ctx, b); + ggml_set_param(b); } ggml_set_name(a, "a"); ggml_set_name(b, "b"); @@ -2040,19 +2040,29 @@ struct test_mul_mat : public test_case { ggml_set_name(a, "a_permuted"); ggml_set_name(b, "b_permuted"); } else { - if (v) { - a = ggml_new_tensor_4d(ctx, type_a, k*2, m, bs[0], bs[1]); - a = ggml_view_4d(ctx, a, k, m, bs[0], bs[1], a->nb[1], a->nb[2], a->nb[3], 0); + a = ggml_new_tensor_4d(ctx, type_a, k*2, m, bs[0], bs[1]); + b = ggml_new_tensor_4d(ctx, type_b, k*2, n, bs[0]*nr[0], bs[1]*nr[1]); + + if (!ggml_is_quantized(type_a)) { + if (bs[1] == 1 && nr[1] == 1) { + ggml_set_param(a); + } + ggml_set_param(b); + } + + a = ggml_view_4d(ctx, a, k, m, bs[0], bs[1], a->nb[1], a->nb[2], a->nb[3], 0); + b = ggml_view_4d(ctx, b, k, n, bs[0]*nr[0], bs[1]*nr[1], b->nb[1], b->nb[2], b->nb[3], 0); } else { a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0], bs[1]); - } - b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]); - if (!ggml_is_quantized(type_a)) { - if (bs[1] == 1 && nr[1] == 1) { - ggml_set_param(ctx, a); + b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]); + + if (!ggml_is_quantized(type_a)) { + if (bs[1] == 1 && nr[1] == 1) { + ggml_set_param(a); + } + ggml_set_param(b); } - ggml_set_param(ctx, b); } ggml_set_name(a, "a"); ggml_set_name(b, "b"); @@ -2201,7 +2211,7 @@ struct test_sqr : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_set_param(ctx, a); + ggml_set_param(a); ggml_set_name(a, "a"); ggml_tensor * out = ggml_sqr(ctx, a); @@ -2230,7 +2240,7 @@ struct test_sqrt : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_set_param(ctx, a); + ggml_set_param(a); ggml_set_name(a, "a"); ggml_tensor * out = ggml_sqrt(ctx, a); @@ -2270,7 +2280,7 @@ struct test_log : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_set_param(ctx, a); + ggml_set_param(a); ggml_set_name(a, "a"); ggml_tensor * out = ggml_log(ctx, a); @@ -2306,7 +2316,7 @@ struct test_sin : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_set_param(ctx, a); + ggml_set_param(a); ggml_set_name(a, "a"); ggml_tensor * out = ggml_sin(ctx, a); @@ -2349,7 +2359,7 @@ struct test_cos : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_set_param(ctx, a); + ggml_set_param(a); ggml_set_name(a, "a"); ggml_tensor * out = ggml_cos(ctx, a); @@ -2429,7 +2439,7 @@ struct test_diag_mask_inf : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_set_param(ctx, a); + ggml_set_param(a); ggml_set_name(a, "a"); ggml_tensor * out = ggml_diag_mask_inf(ctx, a, n_past); @@ -2468,7 +2478,7 @@ struct test_soft_max : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_set_param(ctx, a); + ggml_set_param(a); ggml_set_name(a, "a"); ggml_tensor * mask = nullptr; @@ -2550,7 +2560,7 @@ struct test_rope : public test_case { auto ne = ne_a; ne[0] *= 2; ne[1] *= 4; ne[2] *= 3; a = ggml_new_tensor(ctx, type, 4, ne.data()); if (forward) { - ggml_set_param(ctx, a); + ggml_set_param(a); } ggml_set_name(a, "a"); @@ -2559,7 +2569,7 @@ struct test_rope : public test_case { } else { a = ggml_new_tensor(ctx, type, 4, ne_a.data()); if (forward) { - ggml_set_param(ctx, a); + ggml_set_param(a); } ggml_set_name(a, "a"); } @@ -2673,7 +2683,7 @@ struct test_pool2d : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * input = ggml_new_tensor(ctx, type_input, 4, ne_input.data()); - ggml_set_param(ctx, input); + ggml_set_param(input); ggml_set_name(input, "input"); ggml_tensor * out = ggml_pool_2d(ctx, input, pool_type, k0, k1, s0, s1, p0, p1); @@ -2749,7 +2759,7 @@ struct test_im2col : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * input = ggml_new_tensor(ctx, type_input, 4, ne_input.data()); - ggml_set_param(ctx, input); + ggml_set_param(input); ggml_set_name(input, "input"); ggml_tensor * kernel = ggml_new_tensor(ctx, type_kernel, 4, ne_kernel.data()); @@ -2762,6 +2772,48 @@ struct test_im2col : public test_case { } }; +// GGML_OP_CONV_2D_DW +struct test_conv_2d_dw : public test_case { + const std::array ne_input; + const std::array ne_kernel; + const int stride; + const int padding; + const int dilation; + const bool cwhn; + + std::string vars() override { + return VARS_TO_STR6(ne_input, ne_kernel, stride, padding, dilation, cwhn); + } + + test_conv_2d_dw(std::array ne_input = {64, 64, 16, 1}, + std::array ne_kernel = {3, 3, 1, 16}, + int stride = 1, int padding = 0, int dilation = 1, bool cwhn = false) + : ne_input(ne_input), ne_kernel(ne_kernel), stride(stride), padding(padding), dilation(dilation), cwhn(cwhn) {} + + ggml_tensor * build_graph(ggml_context * ctx) override { + ggml_tensor * input = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_input.data()); + ggml_set_name(input, "input"); + + ggml_tensor * kernel = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_kernel.data()); + ggml_set_name(kernel, "kernel"); + + if (cwhn) { + // change memory layout to channel-most-contiguous (CWHN), + // then permute it back so NE matches the original input + input = ggml_cont(ctx, ggml_permute(ctx, input, 1, 2, 0, 3)); + input = ggml_permute(ctx, input, 2, 0, 1, 3); + kernel = ggml_cont(ctx, ggml_permute(ctx, kernel, 2, 3, 1, 0)); + kernel = ggml_permute(ctx, kernel, 3, 2, 0, 1); + } + + ggml_tensor * out = ggml_conv_2d_dw_direct( + ctx, kernel, input, + stride, stride, padding, padding, dilation, dilation); + ggml_set_name(out, "out"); + return out; + } +}; + // GGML_OP_CONCAT struct test_concat : public test_case { const ggml_type type; @@ -2884,7 +2936,7 @@ struct test_sum : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_set_param(ctx, a); + ggml_set_param(a); ggml_set_name(a, "a"); ggml_tensor * out = ggml_sum(ctx, a); @@ -2913,7 +2965,7 @@ struct test_sum_rows : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_set_param(ctx, a); + ggml_set_param(a); ggml_set_name(a, "a"); ggml_tensor * out = ggml_sum_rows(ctx, a); @@ -2938,7 +2990,7 @@ struct test_mean : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_set_param(ctx, a); + ggml_set_param(a); ggml_set_name(a, "a"); ggml_tensor * out = ggml_mean(ctx, a); @@ -3084,11 +3136,11 @@ struct test_acc : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data()); - ggml_set_param(ctx, a); + ggml_set_param(a); ggml_set_name(a, "a"); ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne_b.data()); - ggml_set_param(ctx, b); + ggml_set_param(b); ggml_set_name(b, "b"); ggml_tensor * out = ggml_acc(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], b->nb[1]); @@ -3325,7 +3377,7 @@ struct test_cross_entropy_loss : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * logits = ggml_new_tensor(ctx, type, 4, ne.data()); - ggml_set_param(ctx, logits); + ggml_set_param(logits); ggml_set_name(logits, "logits"); ggml_tensor * labels = ggml_new_tensor(ctx, type, 4, ne.data()); @@ -3407,7 +3459,7 @@ struct test_opt_step_adamw : public test_case { ggml_tensor * build_graph(ggml_context * ctx) override { ggml_tensor * a = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]); - ggml_set_param(ctx, a); // Despite tensor a having gradients the output tensor will not. + ggml_set_param(a); // Despite tensor a having gradients the output tensor will not. ggml_set_name(a, "a"); ggml_tensor * grad = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]); @@ -3972,6 +4024,11 @@ static std::vector> make_test_cases_eval() { // test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F16, {1024, 1024, 256, 1}, {3, 3, 256, 1}, 1, 1, 1, 1, 1, 1, true)); // test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32, {1024, 1024, 256, 1}, {3, 3, 256, 1}, 1, 1, 1, 1, 1, 1, true)); + test_cases.emplace_back(new test_conv_2d_dw({17, 34, 9, 1}, {3, 3, 1, 9}, 1, 0, 1, false)); + test_cases.emplace_back(new test_conv_2d_dw({17, 34, 9, 1}, {3, 3, 1, 9}, 1, 0, 1, true)); + test_cases.emplace_back(new test_conv_2d_dw({32, 8, 64, 1}, {3, 3, 1, 64}, 2, 1, 1, false)); + test_cases.emplace_back(new test_conv_2d_dw({32, 8, 64, 1}, {3, 3, 1, 64}, 2, 1, 1, true)); + test_cases.emplace_back(new test_conv_transpose_1d()); test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 3, 0, 1)); test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 2, 0, 1)); @@ -4184,6 +4241,11 @@ static std::vector> make_test_cases_eval() { test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 2, 1, 3})); test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 1, 3, 2})); test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 3, 2, 1})); + + // test cases with large ne00/ne10 to cover stream-k fixup + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 1, 1024, {3, 2}, {1, 1})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 8, 1024, {3, 2}, {1, 1})); + test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 1024, {3, 2}, {1, 1})); } } for (ggml_type type_a : other_types) { @@ -4541,6 +4603,9 @@ static std::vector> make_test_cases_perf() { } } + test_cases.emplace_back(new test_conv_2d_dw({512, 512, 256, 1}, {3, 3, 1, 256}, 1, 1, 1, false)); + test_cases.emplace_back(new test_conv_2d_dw({512, 512, 256, 1}, {3, 3, 1, 256}, 1, 1, 1, true)); + return test_cases; } diff --git a/tests/test-chat-template.cpp b/tests/test-chat-template.cpp index be1a640068..a0a50f9881 100644 --- a/tests/test-chat-template.cpp +++ b/tests/test-chat-template.cpp @@ -181,8 +181,8 @@ int main(void) { }, { /* .name= */ "ChatGLM4", - /* .template_str= */ U8C("[gMASK]{% for item in messages %}{% if item['tools'] is defined %}<|system|>\n你是一个名为 ChatGLM 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的，你的任务是针对用户的问题和要求提供适当的答复和支持。\n\n# 可用工具{% set tools = item['tools'] %}{% for tool in tools %}{% if tool['type'] == 'function' %}\n\n## {{ tool['function']['name'] }}\n\n{{ tool['function'] | tojson(indent=4) }}\n......{% endif %}{% endfor %}{% endif %}{% if item['content'] %}<|{{ item['role'] }}|>{{ item['metadata'] }}\n{{ item['content'] }}{% endif %}{% endfor %}{% if add_generation_prompt %}<|assistant|>{% endif %}"), - /* .expected_output= */ "[gMASK]<|system|>\nYou are a helpful assistant<|user|>\nHello<|assistant|>\nHi there<|user|>\nWho are you<|assistant|>\n I am an assistant <|user|>\nAnother question<|assistant|>", + /* .template_str= */ U8C("[gMASK]{% for item in messages %}{% if item['tools'] is defined %}<|system|>\n你是一个名为 ChatGLM 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的，你的任务是针对用户的问题和要求提供适当的答复和支持。\n\n# 可用工具{% set tools = item['tools'] %}{% for tool in tools %}{% if tool['type'] == 'function' %}\n\n## {{ tool['function']['name'] }}\n\n{{ tool['function'] | tojson(indent=4) }}\n......{% endif %}{% endfor %}{% endif %}{% if item['content'] %}<|{{ item['role'] }}|>{{ item['metadata'] }}\n{{ item['content'] }}{% endif %}{% endfor %}{% if add_generation_prompt %}<|assistant|>\n{% endif %}"), + /* .expected_output= */ "[gMASK]<|system|>\nYou are a helpful assistant<|user|>\nHello<|assistant|>\nHi there<|user|>\nWho are you<|assistant|>\n I am an assistant <|user|>\nAnother question<|assistant|>\n", /* .expected_output_jinja= */ "", /* .bos_token= */ "", /* .eos_token= */ "", diff --git a/tests/test-json-schema-to-grammar.cpp b/tests/test-json-schema-to-grammar.cpp index e35134f3cb..38cf01d6d8 100755 --- a/tests/test-json-schema-to-grammar.cpp +++ b/tests/test-json-schema-to-grammar.cpp @@ -597,6 +597,22 @@ static void test_all(const std::string & lang, std::function +#include + +#include "mtmd.h" + +int main(void) { + printf("\n\nTesting libmtmd C API...\n"); + printf("--------\n\n"); + + struct mtmd_context_params params = mtmd_context_params_default(); + printf("Default image marker: %s\n", params.image_marker); + + mtmd_input_chunks * chunks = mtmd_test_create_input_chunks(); + + if (!chunks) { + fprintf(stderr, "Failed to create input chunks\n"); + return 1; + } + + size_t n_chunks = mtmd_input_chunks_size(chunks); + printf("Number of chunks: %zu\n", n_chunks); + assert(n_chunks > 0); + + for (size_t i = 0; i < n_chunks; i++) { + const mtmd_input_chunk * chunk = mtmd_input_chunks_get(chunks, i); + assert(chunk != NULL); + enum mtmd_input_chunk_type type = mtmd_input_chunk_get_type(chunk); + printf("Chunk %zu type: %d\n", i, type); + + if (type == MTMD_INPUT_CHUNK_TYPE_TEXT) { + size_t n_tokens; + const llama_token * tokens = mtmd_input_chunk_get_tokens_text(chunk, &n_tokens); + printf(" Text chunk with %zu tokens\n", n_tokens); + assert(tokens != NULL); + assert(n_tokens > 0); + for (size_t j = 0; j < n_tokens; j++) { + assert(tokens[j] >= 0); + printf(" > Token %zu: %d\n", j, tokens[j]); + } + + } else if (type == MTMD_INPUT_CHUNK_TYPE_IMAGE) { + const mtmd_image_tokens * image_tokens = mtmd_input_chunk_get_tokens_image(chunk); + size_t n_tokens = mtmd_image_tokens_get_n_tokens(image_tokens); + size_t nx = mtmd_image_tokens_get_nx(image_tokens); + size_t ny = mtmd_image_tokens_get_ny(image_tokens); + const char * id = mtmd_image_tokens_get_id(image_tokens); + assert(n_tokens > 0); + assert(nx > 0); + assert(ny > 0); + assert(id != NULL); + printf(" Image chunk with %zu tokens\n", n_tokens); + printf(" Image size: %zu x %zu\n", nx, ny); + printf(" Image ID: %s\n", id); + } + } + + // Free the chunks + mtmd_input_chunks_free(chunks); + + printf("\n\nDONE: test libmtmd C API...\n"); + + return 0; +} diff --git a/tests/test-opt.cpp b/tests/test-opt.cpp index f90c92b4b8..558f877210 100644 --- a/tests/test-opt.cpp +++ b/tests/test-opt.cpp @@ -57,7 +57,8 @@ static helper_ctx_data helper_get_ctx_data( enum ggml_opt_loss_type loss_type = GGML_OPT_LOSS_TYPE_SUM) { std::vector datasets(ndata); for (int64_t ndata_shard = 1; ndata_shard <= ndata; ++ndata_shard) { - ggml_opt_dataset_t dataset = ggml_opt_dataset_init(ne_datapoint, ne_label, ndata, ndata_shard); + ggml_opt_dataset_t dataset = ggml_opt_dataset_init( + GGML_TYPE_F32, GGML_TYPE_F32, ne_datapoint, ne_label, ndata, ndata_shard); float * data = ggml_get_data_f32(ggml_opt_dataset_data( dataset)); float * labels = ggml_get_data_f32(ggml_opt_dataset_labels(dataset)); @@ -74,7 +75,8 @@ static helper_ctx_data helper_get_ctx_data( datasets[ndata_shard-1] = dataset; } - ggml_opt_dataset_t dataset_unsupervised = ggml_opt_dataset_init(1, 0, ndata, /*ndata_shard =*/ 1); + ggml_opt_dataset_t dataset_unsupervised = ggml_opt_dataset_init( + GGML_TYPE_F32, GGML_TYPE_F32, 1, 0, ndata, /*ndata_shard =*/ 1); float * data = ggml_get_data_f32(ggml_opt_dataset_data(dataset_unsupervised)); @@ -113,7 +115,7 @@ static helper_ctx_data helper_get_ctx_data( struct ggml_tensor * weights = ggml_new_tensor_1d(ctx_static, GGML_TYPE_F32, 1); ggml_set_name(weights, "weights"); - ggml_set_param(ctx_static, weights); + ggml_set_param(weights); struct ggml_tensor * intermediary = ggml_add(ctx_compute, inputs, weights); @@ -127,8 +129,11 @@ static helper_ctx_data helper_get_ctx_data( GGML_ASSERT(nbatch_logical % nbatch_physical == 0); const int32_t opt_period = nbatch_logical / nbatch_physical; - struct ggml_opt_params opt_params = ggml_opt_default_params(backend_sched, ctx_compute, inputs, outputs, loss_type); - opt_params.opt_period = opt_period; + struct ggml_opt_params opt_params = ggml_opt_default_params(backend_sched, loss_type); + opt_params.ctx_compute = ctx_compute; + opt_params.inputs = inputs; + opt_params.outputs = outputs; + opt_params.opt_period = opt_period; if (!optimizer_defaults) { opt_params.get_opt_pars = helper_get_test_opt_pars; } @@ -264,8 +269,9 @@ static std::pair test_grad(ggml_backend_sched_t backend_sched, ggml_ba for (int idata = 0; idata < ndata; ++idata) { const float idataf = idata; + ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true); ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs)); - ggml_opt_forward_backward(cd.opt_ctx, cd.result); + ggml_opt_eval(cd.opt_ctx, cd.result); ggml_backend_tensor_get(ggml_opt_grad_acc(cd.opt_ctx, cd.weights), grad_history.data() + idata, 0, sizeof(float)); } @@ -334,8 +340,9 @@ static std::pair test_forward_backward( } else { for (int idata = 0; idata < ndata; ++idata) { const float idataf = idata; + ggml_opt_alloc(cd.opt_ctx, /*backward =*/ false); ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs)); - ggml_opt_forward(cd.opt_ctx, cd.result); + ggml_opt_eval(cd.opt_ctx, cd.result); ggml_backend_tensor_get(loss, loss_history.data() + idata, 0, sizeof(float)); } } @@ -367,7 +374,8 @@ static std::pair test_forward_backward( float w0; ggml_backend_tensor_get(cd.weights, &w0, 0, sizeof(float)); for (int i = 0; i < 10; ++i) { - ggml_opt_forward_backward(cd.opt_ctx, nullptr); + ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true); + ggml_opt_eval(cd.opt_ctx, cd.result); } ggml_backend_tensor_set(cd.weights, &w0, 0, sizeof(float)); @@ -387,8 +395,9 @@ static std::pair test_forward_backward( } else { for (int idata = 0; idata < ndata; ++idata) { const float idataf = idata; + ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true); ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs)); - ggml_opt_forward_backward(cd.opt_ctx, cd.result); + ggml_opt_eval(cd.opt_ctx, cd.result); ggml_backend_tensor_get(loss, loss_history.data() + idata, 0, sizeof(float)); } } @@ -492,14 +501,16 @@ static std::pair test_idata_split(ggml_backend_sched_t backend_sched, int idata = 0; for (; idata < idata_split; ++idata) { const float idataf = idata; + ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true); ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs)); - ggml_opt_forward_backward(cd.opt_ctx, cd.result); + ggml_opt_eval(cd.opt_ctx, cd.result); ggml_backend_tensor_get(loss, loss_history.data() + idata, 0, sizeof(float)); } for (; idata < ndata; ++idata) { const float idataf = idata; + ggml_opt_alloc(cd.opt_ctx, /*backward =*/ false); ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs)); - ggml_opt_forward(cd.opt_ctx, cd.result2); + ggml_opt_eval(cd.opt_ctx, cd.result2); ggml_backend_tensor_get(loss, loss_history.data() + idata, 0, sizeof(float)); } } @@ -573,7 +584,6 @@ static std::pair test_gradient_accumulation( struct helper_ctx_data cd = helper_get_ctx_data( backend_sched, backend, /*init_opt_ctx =*/ true, /*optimizer_defaults =*/ false, /*nbatch_logical =*/ 6, nbatch_physical, loss_type); - struct ggml_tensor * loss = ggml_opt_loss(cd.opt_ctx); std::vector grad_history(ndata); for (int64_t idata = 0; idata < ndata; ++idata) { @@ -584,15 +594,17 @@ static std::pair test_gradient_accumulation( if (nbatch_physical == 1) { for (int idata = 0; idata < ndata; ++idata) { const float idataf = idata; + ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true); ggml_backend_tensor_set(cd.inputs, &idataf, 0, 1*sizeof(float)); - ggml_opt_forward_backward(cd.opt_ctx, cd.result); + ggml_opt_eval(cd.opt_ctx, cd.result); ggml_backend_tensor_get(ggml_opt_grad_acc(cd.opt_ctx, cd.weights), grad_history.data() + idata, 0, 1*sizeof(float)); } } else if (nbatch_physical == 2) { for (int idata = 0; idata < ndata; idata += 2) { const float idataf[2] = {float(idata + 0), float(idata + 1)}; + ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true); ggml_backend_tensor_set(cd.inputs, idataf, 0, 2*sizeof(float)); - ggml_opt_forward_backward(cd.opt_ctx, cd.result); + ggml_opt_eval(cd.opt_ctx, cd.result); grad_history[idata + 0] = 0.0f; ggml_backend_tensor_get(ggml_opt_grad_acc(cd.opt_ctx, cd.weights), grad_history.data() + idata + 1, 0, 1*sizeof(float)); @@ -617,7 +629,7 @@ static std::pair test_gradient_accumulation( } subtest_ok = subtest_ok && almost_equal(grad_history[1], 2.0, atol); subtest_ok = subtest_ok && almost_equal(grad_history[3], 4.0, atol); - subtest_ok = subtest_ok && almost_equal(grad_history[5], 0.0, atol); + subtest_ok = subtest_ok && almost_equal(grad_history[5], 6.0, atol); } else if (loss_type == GGML_OPT_LOSS_TYPE_MEAN) { if (nbatch_physical == 1) { subtest_ok = subtest_ok && almost_equal(grad_history[0], 1.0/ndata, atol); @@ -630,7 +642,7 @@ static std::pair test_gradient_accumulation( } subtest_ok = subtest_ok && almost_equal(grad_history[1], 2.0/ndata, atol); subtest_ok = subtest_ok && almost_equal(grad_history[3], 4.0/ndata, atol); - subtest_ok = subtest_ok && almost_equal(grad_history[5], 0.0/ndata, atol); + subtest_ok = subtest_ok && almost_equal(grad_history[5], 6.0/ndata, atol); } else { GGML_ASSERT(false); } @@ -692,7 +704,8 @@ static std::pair test_regression(ggml_backend_sched_t backend_sched, g std::mt19937 gen(12345); std::normal_distribution nd{0.0f, 0.1f}; - ggml_opt_dataset_t dataset = ggml_opt_dataset_init(1, 1, ndata_regression, ndata_regression); + ggml_opt_dataset_t dataset = ggml_opt_dataset_init( + GGML_TYPE_F32, GGML_TYPE_F32, 1, 1, ndata_regression, ndata_regression); float * data = ggml_get_data_f32(ggml_opt_dataset_data( dataset)); float * labels = ggml_get_data_f32(ggml_opt_dataset_labels(dataset)); @@ -733,15 +746,14 @@ static std::pair test_regression(ggml_backend_sched_t backend_sched, g struct ggml_tensor * a = ggml_new_tensor_1d(ctx_static, GGML_TYPE_F32, 1); ggml_set_name(a, "a"); - ggml_set_param(ctx_static, a); + ggml_set_param(a); struct ggml_tensor * b = ggml_new_tensor_1d(ctx_static, GGML_TYPE_F32, 1); ggml_set_name(b, "b"); - ggml_set_param(ctx_static, b); + ggml_set_param(b); struct ggml_tensor * f = ggml_add(ctx_compute, ggml_mul(ctx_compute, x, a), b); ggml_set_name(f, "f"); - ggml_set_param(ctx_static, f); ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx_static, backend); const float a0 = 1.0f; @@ -853,7 +865,7 @@ int main(void) { backends_modded.insert(backends_modded.end(), backends.begin(), backends.end()); ggml_backend_sched_t backend_sched = ggml_backend_sched_new( - backends_modded.data(), nullptr, backends_modded.size(), GGML_DEFAULT_GRAPH_SIZE, false); + backends_modded.data(), nullptr, backends_modded.size(), GGML_DEFAULT_GRAPH_SIZE, false, true); printf("Backend %zu/%zu: %s\n", i + 1, dev_count, ggml_backend_dev_name(devs[i])); printf(" Device description: %s\n", ggml_backend_dev_description(devs[i])); diff --git a/tests/test-quantize-stats.cpp b/tests/test-quantize-stats.cpp index db01059119..a284a1f0c5 100644 --- a/tests/test-quantize-stats.cpp +++ b/tests/test-quantize-stats.cpp @@ -1,4 +1,5 @@ #include "ggml.h" +#include "ggml-cpu.h" #include "llama.h" #include "common.h" diff --git a/tests/test-sampling.cpp b/tests/test-sampling.cpp index f1f87d454d..60ac62b385 100644 --- a/tests/test-sampling.cpp +++ b/tests/test-sampling.cpp @@ -360,7 +360,7 @@ int main(void) { test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 3, 4, 0, 1}, {0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, 1.0f, 1.1f, 4, 7, {}); test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.571429f, 0.428571f, 0.0f, 0.0f}, 1.00f); - test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f, 0.0f, 0.0f, 0.0f}, 0.00f); + test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 0.00f); // top_n_sigma == 0 now represents a no-op rather than greedy decoding as of PR#13345 test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 3.00f); test_sampler_queue(10000, "k", 10000, 1.0f, 1.0f); diff --git a/tools/CMakeLists.txt b/tools/CMakeLists.txt new file mode 100644 index 0000000000..d64956b843 --- /dev/null +++ b/tools/CMakeLists.txt @@ -0,0 +1,39 @@ +# dependencies + +find_package(Threads REQUIRED) + +# third-party + +# ... + +# flags + +llama_add_compile_flags() + +# tools + +if (EMSCRIPTEN) +else() + add_subdirectory(batched-bench) + add_subdirectory(gguf-split) + add_subdirectory(imatrix) + add_subdirectory(llama-bench) + add_subdirectory(main) + add_subdirectory(perplexity) + add_subdirectory(quantize) + if (LLAMA_BUILD_SERVER) + add_subdirectory(server) + endif() + add_subdirectory(run) + add_subdirectory(tokenize) + add_subdirectory(tts) + add_subdirectory(mtmd) + if (GGML_RPC) + add_subdirectory(rpc) + endif() + if (NOT GGML_BACKEND_DL) + # these examples use the backends directly and cannot be built with dynamic loading + add_subdirectory(cvector-generator) + add_subdirectory(export-lora) + endif() +endif() diff --git a/examples/batched-bench/CMakeLists.txt b/tools/batched-bench/CMakeLists.txt similarity index 100% rename from examples/batched-bench/CMakeLists.txt rename to tools/batched-bench/CMakeLists.txt diff --git a/examples/batched-bench/README.md b/tools/batched-bench/README.md similarity index 100% rename from examples/batched-bench/README.md rename to tools/batched-bench/README.md diff --git a/examples/batched-bench/batched-bench.cpp b/tools/batched-bench/batched-bench.cpp similarity index 100% rename from examples/batched-bench/batched-bench.cpp rename to tools/batched-bench/batched-bench.cpp diff --git a/examples/cvector-generator/CMakeLists.txt b/tools/cvector-generator/CMakeLists.txt similarity index 100% rename from examples/cvector-generator/CMakeLists.txt rename to tools/cvector-generator/CMakeLists.txt diff --git a/examples/cvector-generator/README.md b/tools/cvector-generator/README.md similarity index 100% rename from examples/cvector-generator/README.md rename to tools/cvector-generator/README.md diff --git a/examples/cvector-generator/completions.txt b/tools/cvector-generator/completions.txt similarity index 100% rename from examples/cvector-generator/completions.txt rename to tools/cvector-generator/completions.txt diff --git a/examples/cvector-generator/cvector-generator.cpp b/tools/cvector-generator/cvector-generator.cpp similarity index 100% rename from examples/cvector-generator/cvector-generator.cpp rename to tools/cvector-generator/cvector-generator.cpp diff --git a/examples/cvector-generator/mean.hpp b/tools/cvector-generator/mean.hpp similarity index 100% rename from examples/cvector-generator/mean.hpp rename to tools/cvector-generator/mean.hpp diff --git a/examples/cvector-generator/negative.txt b/tools/cvector-generator/negative.txt similarity index 100% rename from examples/cvector-generator/negative.txt rename to tools/cvector-generator/negative.txt diff --git a/examples/cvector-generator/pca.hpp b/tools/cvector-generator/pca.hpp similarity index 100% rename from examples/cvector-generator/pca.hpp rename to tools/cvector-generator/pca.hpp diff --git a/examples/cvector-generator/positive.txt b/tools/cvector-generator/positive.txt similarity index 100% rename from examples/cvector-generator/positive.txt rename to tools/cvector-generator/positive.txt diff --git a/examples/export-lora/CMakeLists.txt b/tools/export-lora/CMakeLists.txt similarity index 100% rename from examples/export-lora/CMakeLists.txt rename to tools/export-lora/CMakeLists.txt diff --git a/examples/export-lora/README.md b/tools/export-lora/README.md similarity index 100% rename from examples/export-lora/README.md rename to tools/export-lora/README.md diff --git a/examples/export-lora/export-lora.cpp b/tools/export-lora/export-lora.cpp similarity index 100% rename from examples/export-lora/export-lora.cpp rename to tools/export-lora/export-lora.cpp diff --git a/examples/gguf-split/CMakeLists.txt b/tools/gguf-split/CMakeLists.txt similarity index 100% rename from examples/gguf-split/CMakeLists.txt rename to tools/gguf-split/CMakeLists.txt diff --git a/examples/gguf-split/README.md b/tools/gguf-split/README.md similarity index 100% rename from examples/gguf-split/README.md rename to tools/gguf-split/README.md diff --git a/examples/gguf-split/gguf-split.cpp b/tools/gguf-split/gguf-split.cpp similarity index 100% rename from examples/gguf-split/gguf-split.cpp rename to tools/gguf-split/gguf-split.cpp diff --git a/examples/gguf-split/tests.sh b/tools/gguf-split/tests.sh similarity index 100% rename from examples/gguf-split/tests.sh rename to tools/gguf-split/tests.sh diff --git a/examples/imatrix/CMakeLists.txt b/tools/imatrix/CMakeLists.txt similarity index 100% rename from examples/imatrix/CMakeLists.txt rename to tools/imatrix/CMakeLists.txt diff --git a/examples/imatrix/README.md b/tools/imatrix/README.md similarity index 98% rename from examples/imatrix/README.md rename to tools/imatrix/README.md index 9aa2b20347..6d8897d98b 100644 --- a/examples/imatrix/README.md +++ b/tools/imatrix/README.md @@ -1,4 +1,4 @@ -# llama.cpp/examples/imatrix +# llama.cpp/tools/imatrix Compute an importance matrix for a model and given text dataset. Can be used during quantization to enhance the quality of the quantized models. More information is available here: https://github.com/ggml-org/llama.cpp/pull/4861 diff --git a/examples/imatrix/imatrix.cpp b/tools/imatrix/imatrix.cpp similarity index 97% rename from examples/imatrix/imatrix.cpp rename to tools/imatrix/imatrix.cpp index 31b675e8f9..81d0404d68 100644 --- a/examples/imatrix/imatrix.cpp +++ b/tools/imatrix/imatrix.cpp @@ -24,7 +24,8 @@ static void print_usage(int, char ** argv) { LOG("\n %s \\\n" " -m model.gguf -f some-text.txt [-o imatrix.dat] [--process-output] \\\n" " [--no-ppl] [--chunk 123] [--output-frequency 10] [--save-frequency 0] \\\n" - " [--in-file imatrix-prev-0.dat --in-file imatrix-prev-1.dat ...]\n" , argv[0]); + " [--in-file imatrix-prev-0.dat --in-file imatrix-prev-1.dat ...] \\\n" + " [--parse-special]\n" , argv[0]); LOG("\n"); } @@ -46,7 +47,7 @@ private: common_params m_params; std::mutex m_mutex; int m_last_call = 0; - std::vector m_src1_data; + std::vector m_src1_data; std::vector m_ids; // the expert ids from ggml_mul_mat_id }; @@ -93,11 +94,13 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * const bool is_host = ggml_backend_buffer_is_host(src1->buffer); if (!is_host) { - m_src1_data.resize(ggml_nelements(src1)); - ggml_backend_tensor_get(src1, m_src1_data.data(), 0, ggml_nbytes(src1)); + const size_t src1_nbytes = ggml_nbytes(src1); + m_src1_data.resize(src1_nbytes); + ggml_backend_tensor_get(src1, m_src1_data.data(), 0, src1_nbytes); } - const float * data = is_host ? (const float *) src1->data : m_src1_data.data(); + const char * data = is_host ? (const char *) src1->data : m_src1_data.data(); + GGML_ASSERT(src1->nb[0] == ggml_element_size(src1)); // this has been adapted to the new format of storing merged experts in a single 3d tensor // ref: https://github.com/ggml-org/llama.cpp/pull/6387 @@ -144,7 +147,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * const int64_t i11 = idx % src1->ne[1]; const int64_t i12 = row; - const float * x = (const float *)((const char *)data + i11*src1->nb[1] + i12*src1->nb[2]); + const float * x = (const float *)(data + i11*src1->nb[1] + i12*src1->nb[2]); for (int j = 0; j < (int)src1->ne[0]; ++j) { e.values[e_start + j] += x[j]*x[j]; @@ -180,7 +183,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * ++e.ncall; LOG_DBGV(2, "%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type); for (int row = 0; row < (int)src1->ne[1]; ++row) { - const float * x = data + row * src1->ne[0]; + const float * x = (const float *) (data + row * src1->nb[1]); for (int j = 0; j < (int)src1->ne[0]; ++j) { e.values[j] += x[j]*x[j]; e.counts[j]++; @@ -437,7 +440,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) { auto tim1 = std::chrono::high_resolution_clock::now(); LOG_INF("%s: tokenizing the input ..\n", __func__); - std::vector tokens = common_tokenize(ctx, params.prompt, true); + std::vector tokens = common_tokenize(ctx, params.prompt, true, params.parse_special); auto tim2 = std::chrono::high_resolution_clock::now(); LOG_INF("%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast(tim2-tim1).count()); @@ -583,7 +586,6 @@ int main(int argc, char ** argv) { params.out_file = "imatrix.dat" ; params.n_ctx = 512; - params.logits_all = true; params.escape = false; if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_IMATRIX, print_usage)) { diff --git a/examples/llama-bench/CMakeLists.txt b/tools/llama-bench/CMakeLists.txt similarity index 100% rename from examples/llama-bench/CMakeLists.txt rename to tools/llama-bench/CMakeLists.txt diff --git a/examples/llama-bench/README.md b/tools/llama-bench/README.md similarity index 55% rename from examples/llama-bench/README.md rename to tools/llama-bench/README.md index 6bbe4bb75f..4fb2a24e19 100644 --- a/examples/llama-bench/README.md +++ b/tools/llama-bench/README.md @@ -1,4 +1,4 @@ -# llama.cpp/examples/llama-bench +# llama.cpp/tools/llama-bench Performance testing tool for llama.cpp. @@ -20,19 +20,30 @@ Performance testing tool for llama.cpp. ## Syntax ``` -usage: ./llama-bench [options] +usage: llama-bench [options] options: -h, --help + --numa numa mode (default: disabled) + -r, --repetitions number of times to repeat each test (default: 5) + --prio <0|1|2|3> process/thread priority (default: 0) + --delay <0...N> (seconds) delay between each test (default: 0) + -o, --output output format printed to stdout (default: md) + -oe, --output-err output format printed to stderr (default: none) + -v, --verbose verbose output + --progress print test progress indicators + +test parameters: -m, --model (default: models/7B/ggml-model-q4_0.gguf) -p, --n-prompt (default: 512) -n, --n-gen (default: 128) -pg (default: ) + -d, --n-depth (default: 0) -b, --batch-size (default: 2048) -ub, --ubatch-size (default: 512) -ctk, --cache-type-k (default: f16) -ctv, --cache-type-v (default: f16) - -t, --threads (default: 8) + -t, --threads (default: 16) -C, --cpu-mask (default: 0x0) --cpu-strict <0|1> (default: 0) --poll <0...100> (default: 50) @@ -43,17 +54,15 @@ options: -nkvo, --no-kv-offload <0|1> (default: 0) -fa, --flash-attn <0|1> (default: 0) -mmp, --mmap <0|1> (default: 1) - --numa (default: disabled) -embd, --embeddings <0|1> (default: 0) -ts, --tensor-split (default: 0) - -r, --repetitions (default: 5) - --prio <0|1|2|3> (default: 0) - --delay <0...N> (seconds) (default: 0) - -o, --output (default: md) - -oe, --output-err (default: none) - -v, --verbose (default: 0) + -ot --override-tensors =;... + (default: disabled) + -nopo, --no-op-offload <0|1> (default: 0) -Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times. +Multiple values can be given for each parameter by separating them with ',' +or by specifying the parameter multiple times. Ranges can be given as +'start-end' or 'start-end+step' or 'start-end*mult'. ``` llama-bench can perform three types of tests: @@ -66,6 +75,8 @@ With the exception of `-r`, `-o` and `-v`, all options can be specified multiple Each test is repeated the number of times given by `-r`, and the results are averaged. The results are given in average tokens per second (t/s) and standard deviation. Some output formats (e.g. json) also include the individual results of each repetition. +Using the `-d ` option, each test can be run at a specified context depth, prefilling the KV cache with `` tokens. + For a description of the other options, see the [main example](../main/README.md). Note: @@ -148,6 +159,19 @@ $ ./llama-bench -ngl 10,20,30,31,32,33,34,35 | llama 7B mostly Q4_0 | 3.56 GiB | 6.74 B | CUDA | 35 | pp 512 | 2400.01 ± 7.72 | | llama 7B mostly Q4_0 | 3.56 GiB | 6.74 B | CUDA | 35 | tg 128 | 131.66 ± 0.49 | +### Different prefilled context + +``` +$ ./llama-bench -d 0,512 +``` + +| model | size | params | backend | ngl | test | t/s | +| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: | +| qwen2 7B Q4_K - Medium | 4.36 GiB | 7.62 B | CUDA | 99 | pp512 | 7340.20 ± 23.45 | +| qwen2 7B Q4_K - Medium | 4.36 GiB | 7.62 B | CUDA | 99 | tg128 | 120.60 ± 0.59 | +| qwen2 7B Q4_K - Medium | 4.36 GiB | 7.62 B | CUDA | 99 | pp512 @ d512 | 6425.91 ± 18.88 | +| qwen2 7B Q4_K - Medium | 4.36 GiB | 7.62 B | CUDA | 99 | tg128 @ d512 | 116.71 ± 0.60 | + ## Output formats By default, llama-bench outputs the results in markdown format. The results can be output in other formats by using the `-o` option. @@ -170,9 +194,9 @@ $ ./llama-bench -o csv ``` ```csv -build_commit,build_number,cuda,metal,gpu_blas,blas,cpu_info,gpu_info,model_filename,model_type,model_size,model_n_params,n_batch,n_threads,f16_kv,n_gpu_layers,main_gpu,mul_mat_q,tensor_split,n_prompt,n_gen,test_time,avg_ns,stddev_ns,avg_ts,stddev_ts -"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","512","0","2023-09-23T12:09:01Z","212155977","732372","2413.341687","8.305961" -"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","0","128","2023-09-23T12:09:02Z","969320879","2728399","132.052051","0.371342" +build_commit,build_number,cpu_info,gpu_info,backends,model_filename,model_type,model_size,model_n_params,n_batch,n_ubatch,n_threads,cpu_mask,cpu_strict,poll,type_k,type_v,n_gpu_layers,split_mode,main_gpu,no_kv_offload,flash_attn,tensor_split,use_mmap,embeddings,n_prompt,n_gen,n_depth,test_time,avg_ns,stddev_ns,avg_ts,stddev_ts +"8cf427ff","5163","AMD Ryzen 7 7800X3D 8-Core Processor","NVIDIA GeForce RTX 4080","CUDA","models/Qwen2.5-7B-Instruct-Q4_K_M.gguf","qwen2 7B Q4_K - Medium","4677120000","7615616512","2048","512","8","0x0","0","50","f16","f16","99","layer","0","0","0","0.00","1","0","512","0","0","2025-04-24T11:57:09Z","70285660","982040","7285.676949","100.064434" +"8cf427ff","5163","AMD Ryzen 7 7800X3D 8-Core Processor","NVIDIA GeForce RTX 4080","CUDA","models/Qwen2.5-7B-Instruct-Q4_K_M.gguf","qwen2 7B Q4_K - Medium","4677120000","7615616512","2048","512","8","0x0","0","50","f16","f16","99","layer","0","0","0","0.00","1","0","0","128","0","2025-04-24T11:57:10Z","1067431600","3834831","119.915244","0.430617" ``` ### JSON @@ -184,64 +208,78 @@ $ ./llama-bench -o json ```json [ { - "build_commit": "3469684", - "build_number": 1275, - "cuda": true, - "metal": false, - "gpu_blas": true, - "blas": true, - "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K", - "gpu_info": "NVIDIA GeForce RTX 3090 Ti", - "model_filename": "models/7B/ggml-model-q4_0.gguf", - "model_type": "llama 7B mostly Q4_0", - "model_size": 3825065984, - "model_n_params": 6738415616, - "n_batch": 512, - "n_threads": 16, - "f16_kv": true, + "build_commit": "8cf427ff", + "build_number": 5163, + "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor", + "gpu_info": "NVIDIA GeForce RTX 4080", + "backends": "CUDA", + "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf", + "model_type": "qwen2 7B Q4_K - Medium", + "model_size": 4677120000, + "model_n_params": 7615616512, + "n_batch": 2048, + "n_ubatch": 512, + "n_threads": 8, + "cpu_mask": "0x0", + "cpu_strict": false, + "poll": 50, + "type_k": "f16", + "type_v": "f16", "n_gpu_layers": 99, + "split_mode": "layer", "main_gpu": 0, - "mul_mat_q": true, + "no_kv_offload": false, + "flash_attn": false, "tensor_split": "0.00", + "use_mmap": true, + "embeddings": false, "n_prompt": 512, "n_gen": 0, - "test_time": "2023-09-23T12:09:57Z", - "avg_ns": 212365953, - "stddev_ns": 985423, - "avg_ts": 2410.974041, - "stddev_ts": 11.163766, - "samples_ns": [ 213837238, 211635853, 212328053, 211329715, 212698907 ], - "samples_ts": [ 2394.34, 2419.25, 2411.36, 2422.75, 2407.16 ] + "n_depth": 0, + "test_time": "2025-04-24T11:58:50Z", + "avg_ns": 72135640, + "stddev_ns": 1453752, + "avg_ts": 7100.002165, + "stddev_ts": 140.341520, + "samples_ns": [ 74601900, 71632900, 71745200, 71952700, 70745500 ], + "samples_ts": [ 6863.1, 7147.55, 7136.37, 7115.79, 7237.21 ] }, { - "build_commit": "3469684", - "build_number": 1275, - "cuda": true, - "metal": false, - "gpu_blas": true, - "blas": true, - "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K", - "gpu_info": "NVIDIA GeForce RTX 3090 Ti", - "model_filename": "models/7B/ggml-model-q4_0.gguf", - "model_type": "llama 7B mostly Q4_0", - "model_size": 3825065984, - "model_n_params": 6738415616, - "n_batch": 512, - "n_threads": 16, - "f16_kv": true, + "build_commit": "8cf427ff", + "build_number": 5163, + "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor", + "gpu_info": "NVIDIA GeForce RTX 4080", + "backends": "CUDA", + "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf", + "model_type": "qwen2 7B Q4_K - Medium", + "model_size": 4677120000, + "model_n_params": 7615616512, + "n_batch": 2048, + "n_ubatch": 512, + "n_threads": 8, + "cpu_mask": "0x0", + "cpu_strict": false, + "poll": 50, + "type_k": "f16", + "type_v": "f16", "n_gpu_layers": 99, + "split_mode": "layer", "main_gpu": 0, - "mul_mat_q": true, + "no_kv_offload": false, + "flash_attn": false, "tensor_split": "0.00", + "use_mmap": true, + "embeddings": false, "n_prompt": 0, "n_gen": 128, - "test_time": "2023-09-23T12:09:59Z", - "avg_ns": 977425219, - "stddev_ns": 9268593, - "avg_ts": 130.965708, - "stddev_ts": 1.238924, - "samples_ns": [ 984472709, 974901233, 989474741, 970729355, 967548060 ], - "samples_ts": [ 130.019, 131.295, 129.362, 131.86, 132.293 ] + "n_depth": 0, + "test_time": "2025-04-24T11:58:51Z", + "avg_ns": 1076767880, + "stddev_ns": 9449585, + "avg_ts": 118.881588, + "stddev_ts": 1.041811, + "samples_ns": [ 1075361300, 1065089400, 1071761200, 1081934900, 1089692600 ], + "samples_ts": [ 119.03, 120.178, 119.43, 118.307, 117.464 ] } ] ``` @@ -254,8 +292,8 @@ $ ./llama-bench -o jsonl ``` ```json lines -{"build_commit":"3469684","build_number":1275,"cuda":true,"metal":false,"gpu_blas":true,"blas":true,"cpu_info":"13th Gen Intel(R) Core(TM) i9-13900K","gpu_info":"NVIDIA GeForce RTX 3090 Ti","model_filename":"models/7B/ggml-model-q4_0.gguf","model_type":"llama 7B mostly Q4_0","model_size":3825065984,"model_n_params":6738415616,"n_batch":512,"n_threads":16,"f16_kv":true,"n_gpu_layers":99,"main_gpu":0,"mul_mat_q":true,"tensor_split":"0.00","n_prompt":512,"n_gen":0,"test_time":"2023-09-23T12:09:57Z","avg_ns":212365953,"stddev_ns":985423,"avg_ts":2410.974041,"stddev_ts":11.163766,"samples_ns":[213837238,211635853,212328053,211329715,212698907],"samples_ts":[2394.34,2419.25,2411.36,2422.75,2407.16]} -{"build_commit":"3469684","build_number":1275,"cuda":true,"metal":false,"gpu_blas":true,"blas":true,"cpu_info":"13th Gen Intel(R) Core(TM) i9-13900K","gpu_info":"NVIDIA GeForce RTX 3090 Ti","model_filename":"models/7B/ggml-model-q4_0.gguf","model_type":"llama 7B mostly Q4_0","model_size":3825065984,"model_n_params":6738415616,"n_batch":512,"n_threads":16,"f16_kv":true,"n_gpu_layers":99,"main_gpu":0,"mul_mat_q":true,"tensor_split":"0.00","n_prompt":0,"n_gen":128,"test_time":"2023-09-23T12:09:59Z","avg_ns":977425219,"stddev_ns":9268593,"avg_ts":130.965708,"stddev_ts":1.238924,"samples_ns":[984472709,974901233,989474741,970729355,967548060],"samples_ts":[130.019,131.295,129.362,131.86,132.293]} +{"build_commit": "8cf427ff", "build_number": 5163, "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor", "gpu_info": "NVIDIA GeForce RTX 4080", "backends": "CUDA", "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf", "model_type": "qwen2 7B Q4_K - Medium", "model_size": 4677120000, "model_n_params": 7615616512, "n_batch": 2048, "n_ubatch": 512, "n_threads": 8, "cpu_mask": "0x0", "cpu_strict": false, "poll": 50, "type_k": "f16", "type_v": "f16", "n_gpu_layers": 99, "split_mode": "layer", "main_gpu": 0, "no_kv_offload": false, "flash_attn": false, "tensor_split": "0.00", "use_mmap": true, "embeddings": false, "n_prompt": 512, "n_gen": 0, "n_depth": 0, "test_time": "2025-04-24T11:59:33Z", "avg_ns": 70497220, "stddev_ns": 883196, "avg_ts": 7263.609157, "stddev_ts": 90.940578, "samples_ns": [ 71551000, 71222800, 70364100, 69439100, 69909100 ],"samples_ts": [ 7155.74, 7188.71, 7276.44, 7373.37, 7323.8 ]} +{"build_commit": "8cf427ff", "build_number": 5163, "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor", "gpu_info": "NVIDIA GeForce RTX 4080", "backends": "CUDA", "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf", "model_type": "qwen2 7B Q4_K - Medium", "model_size": 4677120000, "model_n_params": 7615616512, "n_batch": 2048, "n_ubatch": 512, "n_threads": 8, "cpu_mask": "0x0", "cpu_strict": false, "poll": 50, "type_k": "f16", "type_v": "f16", "n_gpu_layers": 99, "split_mode": "layer", "main_gpu": 0, "no_kv_offload": false, "flash_attn": false, "tensor_split": "0.00", "use_mmap": true, "embeddings": false, "n_prompt": 0, "n_gen": 128, "n_depth": 0, "test_time": "2025-04-24T11:59:33Z", "avg_ns": 1068078400, "stddev_ns": 6279455, "avg_ts": 119.844681, "stddev_ts": 0.699739, "samples_ns": [ 1066331700, 1064864900, 1079042600, 1063328400, 1066824400 ],"samples_ts": [ 120.038, 120.203, 118.624, 120.377, 119.982 ]} ``` @@ -271,25 +309,32 @@ $ ./llama-bench -o sql CREATE TABLE IF NOT EXISTS test ( build_commit TEXT, build_number INTEGER, - cuda INTEGER, - metal INTEGER, - gpu_blas INTEGER, - blas INTEGER, cpu_info TEXT, gpu_info TEXT, + backends TEXT, model_filename TEXT, model_type TEXT, model_size INTEGER, model_n_params INTEGER, n_batch INTEGER, + n_ubatch INTEGER, n_threads INTEGER, - f16_kv INTEGER, + cpu_mask TEXT, + cpu_strict INTEGER, + poll INTEGER, + type_k TEXT, + type_v TEXT, n_gpu_layers INTEGER, + split_mode TEXT, main_gpu INTEGER, - mul_mat_q INTEGER, + no_kv_offload INTEGER, + flash_attn INTEGER, tensor_split TEXT, + use_mmap INTEGER, + embeddings INTEGER, n_prompt INTEGER, n_gen INTEGER, + n_depth INTEGER, test_time TEXT, avg_ns INTEGER, stddev_ns INTEGER, @@ -297,6 +342,6 @@ CREATE TABLE IF NOT EXISTS test ( stddev_ts REAL ); -INSERT INTO test (build_commit, build_number, cuda, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '512', '0', '2023-09-23T12:10:30Z', '212693772', '743623', '2407.240204', '8.409634'); -INSERT INTO test (build_commit, build_number, cuda, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '0', '128', '2023-09-23T12:10:31Z', '977925003', '4037361', '130.891159', '0.537692'); +INSERT INTO test (build_commit, build_number, cpu_info, gpu_info, backends, model_filename, model_type, model_size, model_n_params, n_batch, n_ubatch, n_threads, cpu_mask, cpu_strict, poll, type_k, type_v, n_gpu_layers, split_mode, main_gpu, no_kv_offload, flash_attn, tensor_split, use_mmap, embeddings, n_prompt, n_gen, n_depth, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('8cf427ff', '5163', 'AMD Ryzen 7 7800X3D 8-Core Processor', 'NVIDIA GeForce RTX 4080', 'CUDA', 'models/Qwen2.5-7B-Instruct-Q4_K_M.gguf', 'qwen2 7B Q4_K - Medium', '4677120000', '7615616512', '2048', '512', '8', '0x0', '0', '50', 'f16', 'f16', '99', 'layer', '0', '0', '0', '0.00', '1', '0', '512', '0', '0', '2025-04-24T12:00:08Z', '69905000', '519516', '7324.546977', '54.032613'); +INSERT INTO test (build_commit, build_number, cpu_info, gpu_info, backends, model_filename, model_type, model_size, model_n_params, n_batch, n_ubatch, n_threads, cpu_mask, cpu_strict, poll, type_k, type_v, n_gpu_layers, split_mode, main_gpu, no_kv_offload, flash_attn, tensor_split, use_mmap, embeddings, n_prompt, n_gen, n_depth, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('8cf427ff', '5163', 'AMD Ryzen 7 7800X3D 8-Core Processor', 'NVIDIA GeForce RTX 4080', 'CUDA', 'models/Qwen2.5-7B-Instruct-Q4_K_M.gguf', 'qwen2 7B Q4_K - Medium', '4677120000', '7615616512', '2048', '512', '8', '0x0', '0', '50', 'f16', 'f16', '99', 'layer', '0', '0', '0', '0.00', '1', '0', '0', '128', '0', '2025-04-24T12:00:09Z', '1063608780', '4464130', '120.346696', '0.504647'); ``` diff --git a/examples/llama-bench/llama-bench.cpp b/tools/llama-bench/llama-bench.cpp similarity index 67% rename from examples/llama-bench/llama-bench.cpp rename to tools/llama-bench/llama-bench.cpp index cbcbfcee86..ca0d0aed5e 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/tools/llama-bench/llama-bench.cpp @@ -36,6 +36,46 @@ static uint64_t get_time_ns() { return std::chrono::nanoseconds(clock::now().time_since_epoch()).count(); } +static bool tensor_buft_override_equal(const llama_model_tensor_buft_override& a, const llama_model_tensor_buft_override& b) { + if (a.pattern != b.pattern) { + // cString comparison that may be null + if (a.pattern == nullptr || b.pattern == nullptr) { + return false; + } + if (strcmp(a.pattern, b.pattern) != 0) { + return false; + } + } + if (a.buft != b.buft) { + return false; + } + return true; +} + +static bool vec_tensor_buft_override_equal(const std::vector& a, const std::vector& b) { + if (a.size() != b.size()) { + return false; + } + for (size_t i = 0; i < a.size(); i++) { + if (!tensor_buft_override_equal(a[i], b[i])) { + return false; + } + } + return true; +} + +static bool vec_vec_tensor_buft_override_equal(const std::vector>& a, const std::vector>& b) { + if (a.size() != b.size()) { + return false; + } + for (size_t i = 0; i < a.size(); i++) { + if (!vec_tensor_buft_override_equal(a[i], b[i])) { + return false; + } + } + return true; +} + template static std::string join(const std::vector & values, const std::string & delim) { std::ostringstream str; for (size_t i = 0; i < values.size(); i++) { @@ -155,11 +195,46 @@ static std::string pair_str(const std::pair & p) { return buf; } +static std::vector parse_int_range(const std::string & s) { + // first[-last[(+|*)step]] + std::regex range_regex(R"(^(\d+)(?:-(\d+)(?:([\+|\*])(\d+))?)?(?:,|$))"); + + std::smatch match; + std::string::const_iterator search_start(s.cbegin()); + std::vector result; + while (std::regex_search(search_start, s.cend(), match, range_regex)) { + int first = std::stoi(match[1]); + int last = match[2].matched ? std::stoi(match[2]) : first; + char op = match[3].matched ? match[3].str()[0] : '+'; + int step = match[4].matched ? std::stoi(match[4]) : 1; + + for (int i = first; i <= last;) { + result.push_back(i); + + if (op == '+') { + i += step; + } else if (op == '*') { + i *= step; + } else { + throw std::invalid_argument("invalid range format"); + } + } + search_start = match.suffix().first; + } + + if (search_start != s.cend()) { + throw std::invalid_argument("invalid range format"); + } + + return result; +} + struct cmd_params { std::vector model; std::vector n_prompt; std::vector n_gen; std::vector> n_pg; + std::vector n_depth; std::vector n_batch; std::vector n_ubatch; std::vector type_k; @@ -175,8 +250,10 @@ struct cmd_params { std::vector no_kv_offload; std::vector flash_attn; std::vector> tensor_split; + std::vector> tensor_buft_overrides; std::vector use_mmap; std::vector embeddings; + std::vector no_op_offload; ggml_numa_strategy numa; int reps; ggml_sched_priority prio; @@ -192,6 +269,7 @@ static const cmd_params cmd_params_defaults = { /* n_prompt */ { 512 }, /* n_gen */ { 128 }, /* n_pg */ {}, + /* n_depth */ { 0 }, /* n_batch */ { 2048 }, /* n_ubatch */ { 512 }, /* type_k */ { GGML_TYPE_F16 }, @@ -207,8 +285,10 @@ static const cmd_params cmd_params_defaults = { /* no_kv_offload */ { false }, /* flash_attn */ { false }, /* tensor_split */ { std::vector(llama_max_devices(), 0.0f) }, + /* tensor_buft_overrides*/ { std::vector{ { nullptr, nullptr } } }, /* use_mmap */ { true }, /* embeddings */ { false }, + /* no_op_offload */ { false }, /* numa */ GGML_NUMA_STRATEGY_DISABLED, /* reps */ 5, /* prio */ GGML_SCHED_PRIO_NORMAL, @@ -224,12 +304,29 @@ static void print_usage(int /* argc */, char ** argv) { printf("\n"); printf("options:\n"); printf(" -h, --help\n"); + printf(" --numa numa mode (default: disabled)\n"); + printf(" -r, --repetitions number of times to repeat each test (default: %d)\n", + cmd_params_defaults.reps); + printf(" --prio <0|1|2|3> process/thread priority (default: %d)\n", + cmd_params_defaults.prio); + printf(" --delay <0...N> (seconds) delay between each test (default: %d)\n", + cmd_params_defaults.delay); + printf(" -o, --output output format printed to stdout (default: %s)\n", + output_format_str(cmd_params_defaults.output_format)); + printf(" -oe, --output-err output format printed to stderr (default: %s)\n", + output_format_str(cmd_params_defaults.output_format_stderr)); + printf(" -v, --verbose verbose output\n"); + printf(" --progress print test progress indicators\n"); + printf("\n"); + printf("test parameters:\n"); printf(" -m, --model (default: %s)\n", join(cmd_params_defaults.model, ",").c_str()); printf(" -p, --n-prompt (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str()); printf(" -n, --n-gen (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str()); printf(" -pg (default: %s)\n", join(transform_to_str(cmd_params_defaults.n_pg, pair_str), ",").c_str()); + printf(" -d, --n-depth (default: %s)\n", + join(cmd_params_defaults.n_depth, ",").c_str()); printf(" -b, --batch-size (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str()); printf(" -ub, --ubatch-size (default: %s)\n", @@ -261,23 +358,17 @@ static void print_usage(int /* argc */, char ** argv) { join(cmd_params_defaults.flash_attn, ",").c_str()); printf(" -mmp, --mmap <0|1> (default: %s)\n", join(cmd_params_defaults.use_mmap, ",").c_str()); - printf(" --numa (default: disabled)\n"); printf(" -embd, --embeddings <0|1> (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str()); printf(" -ts, --tensor-split (default: 0)\n"); - printf(" -r, --repetitions (default: %d)\n", cmd_params_defaults.reps); - printf(" --prio <0|1|2|3> (default: %d)\n", cmd_params_defaults.prio); - printf(" --delay <0...N> (seconds) (default: %d)\n", cmd_params_defaults.delay); - printf(" -o, --output (default: %s)\n", - output_format_str(cmd_params_defaults.output_format)); - printf(" -oe, --output-err (default: %s)\n", - output_format_str(cmd_params_defaults.output_format_stderr)); - printf(" -v, --verbose (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0"); - printf(" --progress (default: %s)\n", cmd_params_defaults.progress ? "1" : "0"); + printf(" -ot --override-tensors =;...\n"); + printf(" (default: disabled)\n"); + printf(" -nopo, --no-op-offload <0|1> (default: 0)\n"); printf("\n"); printf( - "Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter " - "multiple times.\n"); + "Multiple values can be given for each parameter by separating them with ','\n" + "or by specifying the parameter multiple times. Ranges can be given as\n" + "'start-end' or 'start-end+step' or 'start-end*mult'.\n"); } static ggml_type ggml_type_from_name(const std::string & s) { @@ -331,179 +422,190 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { std::replace(arg.begin(), arg.end(), '_', '-'); } - if (arg == "-h" || arg == "--help") { - print_usage(argc, argv); - exit(0); - } else if (arg == "-m" || arg == "--model") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], split_delim); - params.model.insert(params.model.end(), p.begin(), p.end()); - } else if (arg == "-p" || arg == "--n-prompt") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], split_delim); - params.n_prompt.insert(params.n_prompt.end(), p.begin(), p.end()); - } else if (arg == "-n" || arg == "--n-gen") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], split_delim); - params.n_gen.insert(params.n_gen.end(), p.begin(), p.end()); - } else if (arg == "-pg") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], ','); - if (p.size() != 2) { - invalid_param = true; - break; - } - params.n_pg.push_back({ std::stoi(p[0]), std::stoi(p[1]) }); - } else if (arg == "-b" || arg == "--batch-size") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], split_delim); - params.n_batch.insert(params.n_batch.end(), p.begin(), p.end()); - } else if (arg == "-ub" || arg == "--ubatch-size") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], split_delim); - params.n_ubatch.insert(params.n_ubatch.end(), p.begin(), p.end()); - } else if (arg == "-ctk" || arg == "--cache-type-k") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], split_delim); - std::vector types; - for (const auto & t : p) { - ggml_type gt = ggml_type_from_name(t); - if (gt == GGML_TYPE_COUNT) { + try { + if (arg == "-h" || arg == "--help") { + print_usage(argc, argv); + exit(0); + } else if (arg == "-m" || arg == "--model") { + if (++i >= argc) { invalid_param = true; break; } - types.push_back(gt); - } - if (invalid_param) { - break; - } - params.type_k.insert(params.type_k.end(), types.begin(), types.end()); - } else if (arg == "-ctv" || arg == "--cache-type-v") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], split_delim); - std::vector types; - for (const auto & t : p) { - ggml_type gt = ggml_type_from_name(t); - if (gt == GGML_TYPE_COUNT) { + auto p = string_split(argv[i], split_delim); + params.model.insert(params.model.end(), p.begin(), p.end()); + } else if (arg == "-p" || arg == "--n-prompt") { + if (++i >= argc) { invalid_param = true; break; } - types.push_back(gt); - } - if (invalid_param) { - break; - } - params.type_v.insert(params.type_v.end(), types.begin(), types.end()); - } else if (arg == "-t" || arg == "--threads") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], split_delim); - params.n_threads.insert(params.n_threads.end(), p.begin(), p.end()); - } else if (arg == "-C" || arg == "--cpu-mask") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], split_delim); - params.cpu_mask.insert(params.cpu_mask.end(), p.begin(), p.end()); - } else if (arg == "--cpu-strict") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], split_delim); - params.cpu_strict.insert(params.cpu_strict.end(), p.begin(), p.end()); - } else if (arg == "--poll") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], split_delim); - params.poll.insert(params.poll.end(), p.begin(), p.end()); - } else if (arg == "-ngl" || arg == "--n-gpu-layers") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], split_delim); - params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end()); - } else if (llama_supports_rpc() && (arg == "-rpc" || arg == "--rpc")) { - if (++i >= argc) { - invalid_param = true; - break; - } - params.rpc_servers.push_back(argv[i]); - } else if (arg == "-sm" || arg == "--split-mode") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], split_delim); - std::vector modes; - for (const auto & m : p) { - llama_split_mode mode; - if (m == "none") { - mode = LLAMA_SPLIT_MODE_NONE; - } else if (m == "layer") { - mode = LLAMA_SPLIT_MODE_LAYER; - } else if (m == "row") { - mode = LLAMA_SPLIT_MODE_ROW; - } else { + auto p = parse_int_range(argv[i]); + params.n_prompt.insert(params.n_prompt.end(), p.begin(), p.end()); + } else if (arg == "-n" || arg == "--n-gen") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = parse_int_range(argv[i]); + params.n_gen.insert(params.n_gen.end(), p.begin(), p.end()); + } else if (arg == "-pg") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = string_split(argv[i], ','); + if (p.size() != 2) { + invalid_param = true; + break; + } + params.n_pg.push_back({ std::stoi(p[0]), std::stoi(p[1]) }); + } else if (arg == "-d" || arg == "--n-depth") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = parse_int_range(argv[i]); + params.n_depth.insert(params.n_depth.end(), p.begin(), p.end()); + } else if (arg == "-b" || arg == "--batch-size") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = parse_int_range(argv[i]); + params.n_batch.insert(params.n_batch.end(), p.begin(), p.end()); + } else if (arg == "-ub" || arg == "--ubatch-size") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = parse_int_range(argv[i]); + params.n_ubatch.insert(params.n_ubatch.end(), p.begin(), p.end()); + } else if (arg == "-ctk" || arg == "--cache-type-k") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = string_split(argv[i], split_delim); + + std::vector types; + for (const auto & t : p) { + ggml_type gt = ggml_type_from_name(t); + if (gt == GGML_TYPE_COUNT) { + invalid_param = true; + break; + } + types.push_back(gt); + } + if (invalid_param) { + break; + } + params.type_k.insert(params.type_k.end(), types.begin(), types.end()); + } else if (arg == "-ctv" || arg == "--cache-type-v") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = string_split(argv[i], split_delim); + + std::vector types; + for (const auto & t : p) { + ggml_type gt = ggml_type_from_name(t); + if (gt == GGML_TYPE_COUNT) { + invalid_param = true; + break; + } + types.push_back(gt); + } + if (invalid_param) { + break; + } + params.type_v.insert(params.type_v.end(), types.begin(), types.end()); + } else if (arg == "-t" || arg == "--threads") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = parse_int_range(argv[i]); + params.n_threads.insert(params.n_threads.end(), p.begin(), p.end()); + } else if (arg == "-C" || arg == "--cpu-mask") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = string_split(argv[i], split_delim); + params.cpu_mask.insert(params.cpu_mask.end(), p.begin(), p.end()); + } else if (arg == "--cpu-strict") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = string_split(argv[i], split_delim); + params.cpu_strict.insert(params.cpu_strict.end(), p.begin(), p.end()); + } else if (arg == "--poll") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = parse_int_range(argv[i]); + params.poll.insert(params.poll.end(), p.begin(), p.end()); + } else if (arg == "-ngl" || arg == "--n-gpu-layers") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = parse_int_range(argv[i]); + params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end()); + } else if (llama_supports_rpc() && (arg == "-rpc" || arg == "--rpc")) { + if (++i >= argc) { + invalid_param = true; + break; + } + params.rpc_servers.push_back(argv[i]); + } else if (arg == "-sm" || arg == "--split-mode") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = string_split(argv[i], split_delim); + + std::vector modes; + for (const auto & m : p) { + llama_split_mode mode; + if (m == "none") { + mode = LLAMA_SPLIT_MODE_NONE; + } else if (m == "layer") { + mode = LLAMA_SPLIT_MODE_LAYER; + } else if (m == "row") { + mode = LLAMA_SPLIT_MODE_ROW; + } else { + invalid_param = true; + break; + } + modes.push_back(mode); + } + if (invalid_param) { + break; + } + params.split_mode.insert(params.split_mode.end(), modes.begin(), modes.end()); + } else if (arg == "-mg" || arg == "--main-gpu") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.main_gpu = parse_int_range(argv[i]); + } else if (arg == "-nkvo" || arg == "--no-kv-offload") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = string_split(argv[i], split_delim); + params.no_kv_offload.insert(params.no_kv_offload.end(), p.begin(), p.end()); + } else if (arg == "--numa") { + if (++i >= argc) { invalid_param = true; break; } - modes.push_back(mode); - } - if (invalid_param) { - break; - } - params.split_mode.insert(params.split_mode.end(), modes.begin(), modes.end()); - } else if (arg == "-mg" || arg == "--main-gpu") { - if (++i >= argc) { - invalid_param = true; - break; - } - params.main_gpu = string_split(argv[i], split_delim); - } else if (arg == "-nkvo" || arg == "--no-kv-offload") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], split_delim); - params.no_kv_offload.insert(params.no_kv_offload.end(), p.begin(), p.end()); - } else if (arg == "--numa") { - if (++i >= argc) { - invalid_param = true; - break; - } else { std::string value(argv[i]); - /**/ if (value == "distribute" || value == "") { + if (value == "distribute" || value == "") { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; } else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; @@ -513,89 +615,182 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { invalid_param = true; break; } - } - } else if (arg == "-fa" || arg == "--flash-attn") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], split_delim); - params.flash_attn.insert(params.flash_attn.end(), p.begin(), p.end()); - } else if (arg == "-mmp" || arg == "--mmap") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], split_delim); - params.use_mmap.insert(params.use_mmap.end(), p.begin(), p.end()); - } else if (arg == "-embd" || arg == "--embeddings") { - if (++i >= argc) { - invalid_param = true; - break; - } - auto p = string_split(argv[i], split_delim); - params.embeddings.insert(params.embeddings.end(), p.begin(), p.end()); - } else if (arg == "-ts" || arg == "--tensor-split") { - if (++i >= argc) { - invalid_param = true; - break; - } - for (auto ts : string_split(argv[i], split_delim)) { - // split string by ; and / - const std::regex regex{ R"([;/]+)" }; - std::sregex_token_iterator it{ ts.begin(), ts.end(), regex, -1 }; - std::vector split_arg{ it, {} }; - GGML_ASSERT(split_arg.size() <= llama_max_devices()); + } else if (arg == "-fa" || arg == "--flash-attn") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = string_split(argv[i], split_delim); + params.flash_attn.insert(params.flash_attn.end(), p.begin(), p.end()); + } else if (arg == "-mmp" || arg == "--mmap") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = string_split(argv[i], split_delim); + params.use_mmap.insert(params.use_mmap.end(), p.begin(), p.end()); + } else if (arg == "-embd" || arg == "--embeddings") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = string_split(argv[i], split_delim); + params.embeddings.insert(params.embeddings.end(), p.begin(), p.end()); + } else if (arg == "-nopo" || arg == "--no-op-offload") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = string_split(argv[i], split_delim); + params.no_op_offload.insert(params.no_op_offload.end(), p.begin(), p.end()); + } else if (arg == "-ts" || arg == "--tensor-split") { + if (++i >= argc) { + invalid_param = true; + break; + } + for (auto ts : string_split(argv[i], split_delim)) { + // split string by ; and / + const std::regex regex{ R"([;/]+)" }; + std::sregex_token_iterator it{ ts.begin(), ts.end(), regex, -1 }; + std::vector split_arg{ it, {} }; + GGML_ASSERT(split_arg.size() <= llama_max_devices()); - std::vector tensor_split(llama_max_devices()); - for (size_t i = 0; i < llama_max_devices(); ++i) { - if (i < split_arg.size()) { - tensor_split[i] = std::stof(split_arg[i]); - } else { - tensor_split[i] = 0.0f; + std::vector tensor_split(llama_max_devices()); + for (size_t i = 0; i < llama_max_devices(); ++i) { + if (i < split_arg.size()) { + tensor_split[i] = std::stof(split_arg[i]); + } else { + tensor_split[i] = 0.0f; + } + } + params.tensor_split.push_back(tensor_split); + } + } else if (arg == "-ot" || arg == "--override-tensor") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto value = argv[i]; + /* static */ std::map buft_list; + if (buft_list.empty()) { + // enumerate all the devices and add their buffer types to the list + for (size_t i = 0; i < ggml_backend_dev_count(); ++i) { + auto * dev = ggml_backend_dev_get(i); + auto * buft = ggml_backend_dev_buffer_type(dev); + if (buft) { + buft_list[ggml_backend_buft_name(buft)] = buft; + } } } - params.tensor_split.push_back(tensor_split); - } - } else if (arg == "-r" || arg == "--repetitions") { - if (++i >= argc) { + auto override_group_span_len = std::strcspn(value, ","); + bool last_group = false; + do { + if (override_group_span_len == 0) { + // Adds an empty override-tensors for an empty span + params.tensor_buft_overrides.push_back({{}}); + if (value[override_group_span_len] == '\0') { + value = &value[override_group_span_len]; + last_group = true; + } else { + value = &value[override_group_span_len + 1]; + override_group_span_len = std::strcspn(value, ","); + } + continue; + } + // Stamps null terminators into the argv + // value for this option to avoid the + // memory leak present in the implementation + // over in arg.cpp. Acceptable because we + // only parse these args once in this program. + auto override_group = value; + if (value[override_group_span_len] == '\0') { + value = &value[override_group_span_len]; + last_group = true; + } else { + value[override_group_span_len] = '\0'; + value = &value[override_group_span_len + 1]; + } + std::vector group_tensor_buft_overrides{}; + auto override_span_len = std::strcspn(override_group, ";"); + while (override_span_len > 0) { + auto override = override_group; + if (override_group[override_span_len] != '\0') { + override_group[override_span_len] = '\0'; + override_group = &override_group[override_span_len + 1]; + } else { + override_group = &override_group[override_span_len]; + } + auto tensor_name_span_len = std::strcspn(override, "="); + if (tensor_name_span_len >= override_span_len) { + invalid_param = true; + break; + } + override[tensor_name_span_len] = '\0'; + auto tensor_name = override; + auto buffer_type = &override[tensor_name_span_len + 1]; + if (buft_list.find(buffer_type) == buft_list.end()) { + printf("Available buffer types:\n"); + for (const auto & it : buft_list) { + printf(" %s\n", ggml_backend_buft_name(it.second)); + } + invalid_param = true; + break; + } + group_tensor_buft_overrides.push_back({tensor_name, buft_list.at(buffer_type)}); + override_span_len = std::strcspn(override_group, ";"); + } + if (invalid_param) { + break; + } + group_tensor_buft_overrides.push_back({nullptr,nullptr}); + params.tensor_buft_overrides.push_back(group_tensor_buft_overrides); + override_group_span_len = std::strcspn(value, ","); + } while (!last_group); + } else if (arg == "-r" || arg == "--repetitions") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.reps = std::stoi(argv[i]); + } else if (arg == "--prio") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.prio = (enum ggml_sched_priority) std::stoi(argv[i]); + } else if (arg == "--delay") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.delay = std::stoi(argv[i]); + } else if (arg == "-o" || arg == "--output") { + if (++i >= argc) { + invalid_param = true; + break; + } + invalid_param = !output_format_from_str(argv[i], params.output_format); + } else if (arg == "-oe" || arg == "--output-err") { + if (++i >= argc) { + invalid_param = true; + break; + } + invalid_param = !output_format_from_str(argv[i], params.output_format_stderr); + } else if (arg == "-v" || arg == "--verbose") { + params.verbose = true; + } else if (arg == "--progress") { + params.progress = true; + } else { invalid_param = true; break; } - params.reps = std::stoi(argv[i]); - } else if (arg == "--prio") { - if (++i >= argc) { - invalid_param = true; - break; - } - params.prio = (enum ggml_sched_priority) std::stoi(argv[i]); - } else if (arg == "--delay") { - if (++i >= argc) { - invalid_param = true; - break; - } - params.delay = std::stoi(argv[i]); - } else if (arg == "-o" || arg == "--output") { - if (++i >= argc) { - invalid_param = true; - break; - } - invalid_param = !output_format_from_str(argv[i], params.output_format); - } else if (arg == "-oe" || arg == "--output-err") { - if (++i >= argc) { - invalid_param = true; - break; - } - invalid_param = !output_format_from_str(argv[i], params.output_format_stderr); - } else if (arg == "-v" || arg == "--verbose") { - params.verbose = true; - } else if (arg == "--progress") { - params.progress = true; - } else { + } catch (const std::exception & e) { + fprintf(stderr, "error: %s\n", e.what()); invalid_param = true; break; } } + if (invalid_param) { fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str()); print_usage(argc, argv); @@ -615,6 +810,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { if (params.n_pg.empty()) { params.n_pg = cmd_params_defaults.n_pg; } + if (params.n_depth.empty()) { + params.n_depth = cmd_params_defaults.n_depth; + } if (params.n_batch.empty()) { params.n_batch = cmd_params_defaults.n_batch; } @@ -648,12 +846,18 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { if (params.tensor_split.empty()) { params.tensor_split = cmd_params_defaults.tensor_split; } + if (params.tensor_buft_overrides.empty()) { + params.tensor_buft_overrides = cmd_params_defaults.tensor_buft_overrides; + } if (params.use_mmap.empty()) { params.use_mmap = cmd_params_defaults.use_mmap; } if (params.embeddings.empty()) { params.embeddings = cmd_params_defaults.embeddings; } + if (params.no_op_offload.empty()) { + params.no_op_offload = cmd_params_defaults.no_op_offload; + } if (params.n_threads.empty()) { params.n_threads = cmd_params_defaults.n_threads; } @@ -674,6 +878,7 @@ struct cmd_params_instance { std::string model; int n_prompt; int n_gen; + int n_depth; int n_batch; int n_ubatch; ggml_type type_k; @@ -689,8 +894,10 @@ struct cmd_params_instance { bool no_kv_offload; bool flash_attn; std::vector tensor_split; + std::vector tensor_buft_overrides; bool use_mmap; bool embeddings; + bool no_op_offload; llama_model_params to_llama_mparams() const { llama_model_params mparams = llama_model_default_params(); @@ -733,19 +940,26 @@ struct cmd_params_instance { mparams.tensor_split = tensor_split.data(); mparams.use_mmap = use_mmap; + if (tensor_buft_overrides.empty()) { + mparams.tensor_buft_overrides = nullptr; + } else { + GGML_ASSERT(tensor_buft_overrides.back().pattern == nullptr && "Tensor buffer overrides not terminated with empty pattern"); + mparams.tensor_buft_overrides = tensor_buft_overrides.data(); + } + return mparams; } bool equal_mparams(const cmd_params_instance & other) const { return model == other.model && n_gpu_layers == other.n_gpu_layers && rpc_servers_str == other.rpc_servers_str && split_mode == other.split_mode && main_gpu == other.main_gpu && use_mmap == other.use_mmap && - tensor_split == other.tensor_split; + tensor_split == other.tensor_split && vec_tensor_buft_override_equal(tensor_buft_overrides, other.tensor_buft_overrides); } llama_context_params to_llama_cparams() const { llama_context_params cparams = llama_context_default_params(); - cparams.n_ctx = n_prompt + n_gen; + cparams.n_ctx = n_prompt + n_gen + n_depth; cparams.n_batch = n_batch; cparams.n_ubatch = n_ubatch; cparams.type_k = type_k; @@ -753,6 +967,7 @@ struct cmd_params_instance { cparams.offload_kqv = !no_kv_offload; cparams.flash_attn = flash_attn; cparams.embeddings = embeddings; + cparams.op_offload = !no_op_offload; return cparams; } @@ -769,8 +984,10 @@ static std::vector get_cmd_params_instances(const cmd_param for (const auto & sm : params.split_mode) for (const auto & mg : params.main_gpu) for (const auto & ts : params.tensor_split) + for (const auto & ot : params.tensor_buft_overrides) for (const auto & mmp : params.use_mmap) for (const auto & embd : params.embeddings) + for (const auto & nopo : params.no_op_offload) for (const auto & nb : params.n_batch) for (const auto & nub : params.n_ubatch) for (const auto & tk : params.type_k) @@ -780,6 +997,7 @@ static std::vector get_cmd_params_instances(const cmd_param for (const auto & nt : params.n_threads) for (const auto & cm : params.cpu_mask) for (const auto & cs : params.cpu_strict) + for (const auto & nd : params.n_depth) for (const auto & pl : params.poll) { for (const auto & n_prompt : params.n_prompt) { if (n_prompt == 0) { @@ -789,6 +1007,7 @@ static std::vector get_cmd_params_instances(const cmd_param /* .model = */ m, /* .n_prompt = */ n_prompt, /* .n_gen = */ 0, + /* .n_depth = */ nd, /* .n_batch = */ nb, /* .n_ubatch = */ nub, /* .type_k = */ tk, @@ -804,8 +1023,10 @@ static std::vector get_cmd_params_instances(const cmd_param /* .no_kv_offload= */ nkvo, /* .flash_attn = */ fa, /* .tensor_split = */ ts, + /* .tensor_buft_overrides = */ ot, /* .use_mmap = */ mmp, /* .embeddings = */ embd, + /* .no_op_offload= */ nopo, }; instances.push_back(instance); } @@ -818,6 +1039,7 @@ static std::vector get_cmd_params_instances(const cmd_param /* .model = */ m, /* .n_prompt = */ 0, /* .n_gen = */ n_gen, + /* .n_depth = */ nd, /* .n_batch = */ nb, /* .n_ubatch = */ nub, /* .type_k = */ tk, @@ -833,8 +1055,10 @@ static std::vector get_cmd_params_instances(const cmd_param /* .no_kv_offload= */ nkvo, /* .flash_attn = */ fa, /* .tensor_split = */ ts, + /* .tensor_buft_overrides = */ ot, /* .use_mmap = */ mmp, /* .embeddings = */ embd, + /* .no_op_offload= */ nopo, }; instances.push_back(instance); } @@ -847,6 +1071,7 @@ static std::vector get_cmd_params_instances(const cmd_param /* .model = */ m, /* .n_prompt = */ n_pg.first, /* .n_gen = */ n_pg.second, + /* .n_depth = */ nd, /* .n_batch = */ nb, /* .n_ubatch = */ nub, /* .type_k = */ tk, @@ -862,8 +1087,10 @@ static std::vector get_cmd_params_instances(const cmd_param /* .no_kv_offload= */ nkvo, /* .flash_attn = */ fa, /* .tensor_split = */ ts, + /* .tensor_buft_overrides = */ ot, /* .use_mmap = */ mmp, /* .embeddings = */ embd, + /* .no_op_offload= */ nopo, }; instances.push_back(instance); } @@ -896,10 +1123,13 @@ struct test { bool no_kv_offload; bool flash_attn; std::vector tensor_split; + std::vector tensor_buft_overrides; bool use_mmap; bool embeddings; + bool no_op_offload; int n_prompt; int n_gen; + int n_depth; std::string test_time; std::vector samples_ns; @@ -927,10 +1157,13 @@ struct test { no_kv_offload = inst.no_kv_offload; flash_attn = inst.flash_attn; tensor_split = inst.tensor_split; + tensor_buft_overrides = inst.tensor_buft_overrides; use_mmap = inst.use_mmap; embeddings = inst.embeddings; + no_op_offload = inst.no_op_offload; n_prompt = inst.n_prompt; n_gen = inst.n_gen; + n_depth = inst.n_depth; // RFC 3339 date-time format time_t t = time(NULL); std::strftime(buf, sizeof(buf), "%FT%TZ", gmtime(&t)); @@ -972,9 +1205,9 @@ struct test { "build_commit", "build_number", "cpu_info", "gpu_info", "backends", "model_filename", "model_type", "model_size", "model_n_params", "n_batch", "n_ubatch", "n_threads", "cpu_mask", "cpu_strict", "poll", "type_k", "type_v", "n_gpu_layers", - "split_mode", "main_gpu", "no_kv_offload", "flash_attn", "tensor_split", "use_mmap", - "embeddings", "n_prompt", "n_gen", "test_time", "avg_ns", "stddev_ns", - "avg_ts", "stddev_ts", + "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", }; return fields; } @@ -984,8 +1217,8 @@ struct test { static field_type get_field_type(const std::string & field) { if (field == "build_number" || field == "n_batch" || field == "n_ubatch" || field == "n_threads" || field == "poll" || field == "model_size" || field == "model_n_params" || field == "n_gpu_layers" || - field == "main_gpu" || field == "n_prompt" || field == "n_gen" || field == "avg_ns" || - field == "stddev_ns") { + field == "main_gpu" || field == "n_prompt" || field == "n_gen" || field == "n_depth" || + field == "avg_ns" || field == "stddev_ns" || field == "no_op_offload") { return INT; } if (field == "f16_kv" || field == "no_kv_offload" || field == "cpu_strict" || field == "flash_attn" || @@ -1000,6 +1233,7 @@ struct test { std::vector get_values() const { std::string tensor_split_str; + std::string tensor_buft_overrides_str; int max_nonzero = 0; for (size_t i = 0; i < llama_max_devices(); i++) { if (tensor_split[i] > 0) { @@ -1014,6 +1248,26 @@ struct test { tensor_split_str += "/"; } } + if (tensor_buft_overrides.size() == 1) { + // Last element of tensor_buft_overrides is always a null pattern + // so if it is only one element long, it must be a null pattern. + GGML_ASSERT(tensor_buft_overrides[0].pattern == nullptr); + tensor_buft_overrides_str += "none"; + } else { + for (size_t i = 0; i < tensor_buft_overrides.size()-1; i++) { + // Last element of tensor_buft_overrides is always a null pattern + if (tensor_buft_overrides[i].pattern == nullptr) { + tensor_buft_overrides_str += "none"; + } else { + tensor_buft_overrides_str += tensor_buft_overrides[i].pattern; + tensor_buft_overrides_str += "="; + tensor_buft_overrides_str += ggml_backend_buft_name(tensor_buft_overrides[i].buft); + } + if (i + 2 < tensor_buft_overrides.size()) { + tensor_buft_overrides_str += ";"; + } + } + } std::vector values = { build_commit, std::to_string(build_number), cpu_info, @@ -1037,10 +1291,13 @@ struct test { std::to_string(no_kv_offload), std::to_string(flash_attn), tensor_split_str, + tensor_buft_overrides_str, std::to_string(use_mmap), std::to_string(embeddings), + std::to_string(no_op_offload), std::to_string(n_prompt), std::to_string(n_gen), + std::to_string(n_depth), test_time, std::to_string(avg_ns()), std::to_string(stdev_ns()), @@ -1218,7 +1475,10 @@ struct markdown_printer : public printer { return 4; } if (field == "test") { - return 13; + return 15; + } + if (field == "no_op_offload") { + return 4; } int width = std::max((int) field.length(), 10); @@ -1251,9 +1511,15 @@ struct markdown_printer : public printer { if (field == "embeddings") { return "embd"; } + if (field == "no_op_offload") { + return "nopo"; + } if (field == "tensor_split") { return "ts"; } + if (field == "tensor_buft_overrides") { + return "ot"; + } return field; } @@ -1307,12 +1573,18 @@ struct markdown_printer : public printer { if (params.tensor_split.size() > 1 || params.tensor_split != cmd_params_defaults.tensor_split) { fields.emplace_back("tensor_split"); } + if (params.tensor_buft_overrides.size() > 1 || !vec_vec_tensor_buft_override_equal(params.tensor_buft_overrides, cmd_params_defaults.tensor_buft_overrides)) { + fields.emplace_back("tensor_buft_overrides"); + } if (params.use_mmap.size() > 1 || params.use_mmap != cmd_params_defaults.use_mmap) { fields.emplace_back("use_mmap"); } if (params.embeddings.size() > 1 || params.embeddings != cmd_params_defaults.embeddings) { fields.emplace_back("embeddings"); } + if (params.no_op_offload.size() > 1 || params.no_op_offload != cmd_params_defaults.no_op_offload) { + fields.emplace_back("no_op_offload"); + } fields.emplace_back("test"); fields.emplace_back("t/s"); @@ -1362,6 +1634,10 @@ struct markdown_printer : public printer { } else { snprintf(buf, sizeof(buf), "pp%d+tg%d", t.n_prompt, t.n_gen); } + if (t.n_depth > 0) { + int len = strlen(buf); + snprintf(buf + len, sizeof(buf) - len, " @ d%d", t.n_depth); + } value = buf; } else if (field == "t/s") { snprintf(buf, sizeof(buf), "%.2f ± %.2f", t.avg_ts(), t.stdev_ts()); @@ -1620,6 +1896,14 @@ int main(int argc, char ** argv) { for (int i = 0; i < params.reps; i++) { llama_kv_self_clear(ctx); + if (t.n_depth > 0) { + if (params.progress) { + fprintf(stderr, "llama-bench: benchmark %d/%zu: depth run %d/%d\n", params_idx, params_count, + i + 1, params.reps); + } + test_prompt(ctx, t.n_depth, t.n_batch, t.n_threads); + } + uint64_t t_start = get_time_ns(); if (t.n_prompt > 0) { diff --git a/examples/main/CMakeLists.txt b/tools/main/CMakeLists.txt similarity index 100% rename from examples/main/CMakeLists.txt rename to tools/main/CMakeLists.txt diff --git a/examples/main/README.md b/tools/main/README.md similarity index 99% rename from examples/main/README.md rename to tools/main/README.md index e4b3590b5d..4f16ad6b2b 100644 --- a/examples/main/README.md +++ b/tools/main/README.md @@ -1,4 +1,4 @@ -# llama.cpp/examples/main +# llama.cpp/tools/main This example program allows you to use various LLaMA language models easily and efficiently. It is specifically designed to work with the [llama.cpp](https://github.com/ggml-org/llama.cpp) project, which provides a plain C/C++ implementation with optional 4-bit quantization support for faster, lower memory inference, and is optimized for desktop CPUs. This program can be used to perform various inference tasks with LLaMA models, including generating text based on user-provided prompts and chat-like interactions with reverse prompts. diff --git a/examples/main/main.cpp b/tools/main/main.cpp similarity index 99% rename from examples/main/main.cpp rename to tools/main/main.cpp index c59b941bf5..1bd2be2d94 100644 --- a/examples/main/main.cpp +++ b/tools/main/main.cpp @@ -99,14 +99,6 @@ int main(int argc, char ** argv) { console::init(params.simple_io, params.use_color); atexit([]() { console::cleanup(); }); - if (params.logits_all) { - LOG_ERR("************\n"); - LOG_ERR("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__); - LOG_ERR("************\n\n"); - - return 0; - } - if (params.embedding) { LOG_ERR("************\n"); LOG_ERR("%s: please use the 'embedding' tool for embedding calculations\n", __func__); @@ -160,7 +152,12 @@ int main(int argc, char ** argv) { LOG_INF("%s: llama threadpool init, n_threads = %d\n", __func__, (int) params.cpuparams.n_threads); - auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU)); + auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); + if (!cpu_dev) { + LOG_ERR("%s: no CPU backend found\n", __func__); + return 1; + } + auto * reg = ggml_backend_dev_backend_reg(cpu_dev); auto * ggml_threadpool_new_fn = (decltype(ggml_threadpool_new) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_new"); auto * ggml_threadpool_free_fn = (decltype(ggml_threadpool_free) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_free"); diff --git a/examples/llava/CMakeLists.txt b/tools/mtmd/CMakeLists.txt similarity index 89% rename from examples/llava/CMakeLists.txt rename to tools/mtmd/CMakeLists.txt index 6409b4f5e6..dfafa9cf81 100644 --- a/examples/llava/CMakeLists.txt +++ b/tools/mtmd/CMakeLists.txt @@ -28,6 +28,7 @@ endif() add_library(mtmd OBJECT mtmd.cpp + mtmd-helper.cpp mtmd.h clip.cpp clip.h @@ -64,13 +65,7 @@ endif() add_executable(llama-llava-cli deprecation-warning.cpp) add_executable(llama-gemma3-cli deprecation-warning.cpp) add_executable(llama-minicpmv-cli deprecation-warning.cpp) - -set(TARGET llama-qwen2vl-cli) -add_executable(${TARGET} qwen2vl-cli.cpp) -set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-qwen2vl-cli) -install(TARGETS ${TARGET} RUNTIME) -target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT}) -target_compile_features(${TARGET} PRIVATE cxx_std_17) +add_executable(llama-qwen2vl-cli deprecation-warning.cpp) set(TARGET llama-mtmd-cli) add_executable(${TARGET} mtmd-cli.cpp) diff --git a/examples/llava/README-quantize.md b/tools/mtmd/README-quantize.md similarity index 100% rename from examples/llava/README-quantize.md rename to tools/mtmd/README-quantize.md diff --git a/examples/llava/README.md b/tools/mtmd/README.md similarity index 85% rename from examples/llava/README.md rename to tools/mtmd/README.md index f58d9de710..ab258ea174 100644 --- a/examples/llava/README.md +++ b/tools/mtmd/README.md @@ -16,25 +16,7 @@ The naming and structure related to multimodal support have evolved, which might ## Pre-quantized models -These are ready-to-use models, most of them come with `Q4_K_M` quantization by default: - -```sh -# Gemma 3 -llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF -llama-mtmd-cli -hf ggml-org/gemma-3-12b-it-GGUF -llama-mtmd-cli -hf ggml-org/gemma-3-27b-it-GGUF - -# SmolVLM -llama-mtmd-cli -hf ggml-org/SmolVLM-Instruct-GGUF -llama-mtmd-cli -hf ggml-org/SmolVLM-256M-Instruct-GGUF -llama-mtmd-cli -hf ggml-org/SmolVLM-500M-Instruct-GGUF -llama-mtmd-cli -hf ggml-org/SmolVLM2-2.2B-Instruct-GGUF -llama-mtmd-cli -hf ggml-org/SmolVLM2-256M-Video-Instruct-GGUF -llama-mtmd-cli -hf ggml-org/SmolVLM2-500M-Video-Instruct-GGUF - -# Pixtral 12B -llama-mtmd-cli -hf ggml-org/pixtral-12b-GGUF -``` +See the list of pre-quantized model [here](../../docs/multimodal.md) ## How it works and what is `mmproj`? @@ -57,7 +39,18 @@ Built upon `clip.cpp` (similar to `llava.cpp`), `libmtmd` offers several advanta ## How to obtain `mmproj` -Multimodal projector (`mmproj`) files are specific to each model architecture. Please refer to the relevant guide for instructions on how to obtain or create them: +Multimodal projector (`mmproj`) files are specific to each model architecture. + +For the following models, you can use `convert_hf_to_gguf.py`with `--mmproj` flag to get the `mmproj` file: +- [Gemma 3](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d) - Note: 1B variant does not have vision support +- SmolVLM (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB)) +- SmolVLM2 (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB)) +- [Pixtral 12B](https://huggingface.co/mistral-community/pixtral-12b) - only works with `transformers`-compatible checkpoint +- Qwen 2 VL and Qwen 2.5 VL (from [Qwen](https://huggingface.co/Qwen)) +- [Mistral Small 3.1 24B](https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503) +- InternVL 2.5 and InternVL 3 from [OpenGVLab](https://huggingface.co/OpenGVLab) (note: we don't support conversion of `InternVL3-*-hf` model, only non-HF version is supported ; `InternLM2Model` **text** model is not supported) + +For older models, please refer to the relevant guide for instructions on how to obtain or create them: - [LLaVA](../../docs/multimodal/llava.md) - [MobileVLM](../../docs/multimodal/MobileVLM.md) @@ -67,9 +60,3 @@ Multimodal projector (`mmproj`) files are specific to each model architecture. P - [MiniCPM-o 2.6](../../docs/multimodal/minicpmo2.6.md) - [IBM Granite Vision](../../docs/multimodal/granitevision.md) - [Google Gemma 3](../../docs/multimodal/gemma3.md) - -For the following models, you can use `convert_hf_to_gguf.py`with `--mmproj` flag to get the `mmproj` file: -- [Gemma 3](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d) - Note: 1B variant does not have vision support -- SmolVLM (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB)) -- SmolVLM2 (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB)) -- [Pixtral 12B](https://huggingface.co/mistral-community/pixtral-12b) - only works with `transformers`-compatible checkpoint diff --git a/examples/llava/android/adb_run.sh b/tools/mtmd/android/adb_run.sh similarity index 100% rename from examples/llava/android/adb_run.sh rename to tools/mtmd/android/adb_run.sh diff --git a/examples/llava/android/build_64.sh b/tools/mtmd/android/build_64.sh similarity index 100% rename from examples/llava/android/build_64.sh rename to tools/mtmd/android/build_64.sh diff --git a/examples/llava/clip-impl.h b/tools/mtmd/clip-impl.h similarity index 84% rename from examples/llava/clip-impl.h rename to tools/mtmd/clip-impl.h index 53ac381304..23036ba72f 100644 --- a/examples/llava/clip-impl.h +++ b/tools/mtmd/clip-impl.h @@ -2,8 +2,6 @@ #include "gguf.h" #include "clip.h" -#include "clip.h" - #include #include #include @@ -17,22 +15,15 @@ #define KEY_FTYPE "general.file_type" #define KEY_NAME "general.name" #define KEY_DESCRIPTION "general.description" -#define KEY_HAS_TEXT_ENC "clip.has_text_encoder" -#define KEY_HAS_VIS_ENC "clip.has_vision_encoder" -#define KEY_HAS_LLAVA_PROJ "clip.has_llava_projector" -#define KEY_HAS_MINICPMV_PROJ "clip.has_minicpmv_projector" -#define KEY_HAS_GLM_PROJ "clip.has_glm_projector" #define KEY_MINICPMV_VERSION "clip.minicpmv_version" -#define KEY_HAS_QWEN2VL_MERGER "clip.has_qwen2vl_merger" #define KEY_USE_GELU "clip.use_gelu" #define KEY_USE_SILU "clip.use_silu" -#define KEY_N_EMBD "clip.%s.embedding_length" -#define KEY_N_FF "clip.%s.feed_forward_length" -#define KEY_N_BLOCK "clip.%s.block_count" -#define KEY_N_HEAD "clip.%s.attention.head_count" -#define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon" -#define KEY_PROJ_DIM "clip.%s.projection_dim" -#define KEY_TOKENS "tokenizer.ggml.tokens" +#define KEY_N_EMBD "clip.vision.embedding_length" +#define KEY_N_FF "clip.vision.feed_forward_length" +#define KEY_N_BLOCK "clip.vision.block_count" +#define KEY_N_HEAD "clip.vision.attention.head_count" +#define KEY_LAYER_NORM_EPS "clip.vision.attention.layer_norm_epsilon" +#define KEY_PROJ_DIM "clip.vision.projection_dim" #define KEY_IMAGE_SIZE "clip.vision.image_size" #define KEY_PATCH_SIZE "clip.vision.patch_size" #define KEY_IMAGE_MEAN "clip.vision.image_mean" @@ -40,10 +31,13 @@ #define KEY_FEATURE_LAYER "clip.vision.feature_layer" #define KEY_PROJ_SCALE_FACTOR "clip.vision.projector.scale_factor" #define KEY_PROJ_TYPE "clip.projector_type" +#define KEY_SPATIAL_MERGE_SIZE "clip.vision.spatial_merge_size" #define KEY_MM_PATCH_MERGE_TYPE "clip.vision.mm_patch_merge_type" #define KEY_IMAGE_GRID_PINPOINTS "clip.vision.image_grid_pinpoints" #define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution" +#define KEY_WIN_ATTN_PATTERN "clip.vision.n_wa_pattern" +#define KEY_ATTN_WINDOW_SIZE "clip.vision.window_size" // @@ -59,11 +53,16 @@ #define TN_ATTN_Q "%s.blk.%d.attn_q.%s" #define TN_ATTN_V "%s.blk.%d.attn_v.%s" #define TN_ATTN_OUTPUT "%s.blk.%d.attn_out.%s" +#define TN_ATTN_K_NORM "%s.blk.%d.attn_k_norm.%s" +#define TN_ATTN_Q_NORM "%s.blk.%d.attn_q_norm.%s" #define TN_FFN_DOWN "%s.blk.%d.ffn_down.%s" #define TN_FFN_GATE "%s.blk.%d.ffn_gate.%s" #define TN_FFN_UP "%s.blk.%d.ffn_up.%s" -#define TN_LN_1 "%s.blk.%d.ln1.%s" -#define TN_LN_2 "%s.blk.%d.ln2.%s" +#define TN_FFN_GATE "%s.blk.%d.ffn_gate.%s" +#define TN_LN_1 "%s.blk.%d.ln1.%s" // layer norm +#define TN_LN_2 "%s.blk.%d.ln2.%s" // layer norm +#define TN_LS_1 "%s.blk.%d.ls1.%s" // layer scale +#define TN_LS_2 "%s.blk.%d.ls2.%s" // layer scale #define TN_LN_PRE "%s.pre_ln.%s" #define TN_LN_POST "%s.post_ln.%s" #define TN_LLAVA_PROJ "mm.%d.%s" @@ -71,10 +70,14 @@ #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s" #define TN_MVLM_PROJ_PEG "mm.model.peg.%d.%s" #define TN_IMAGE_NEWLINE "model.image_newline" +#define TN_MM_INP_NORM "mm.input_norm.weight" #define TN_MM_INP_PROJ "mm.input_projection.weight" // gemma3 #define TN_MM_SOFT_EMB_N "mm.soft_emb_norm.weight" // gemma3 #define TN_MM_PROJECTOR "mm.model.fc.weight" // idefics3 +#define TN_MM_PATCH_MERGER "mm.patch_merger.weight" // mistral small 3.1 #define TN_TOK_IMG_BREAK "v.token_embd.img_break" // pixtral +#define TN_TOK_GLM_BOI "adapter.boi" // glm-edge (these embeddings are not in text model) +#define TN_TOK_GLM_EOI "adapter.eoi" // glm-edge (these embeddings are not in text model) // mimicpmv #define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k" @@ -91,17 +94,22 @@ #define TN_GLM_ADAPTER_GATE "adapter.linear.gate.%s" #define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s" +// align x to upper multiple of n +#define CLIP_ALIGN(x, n) ((((x) + (n) - 1) / (n)) * (n)) + enum projector_type { PROJECTOR_TYPE_MLP, PROJECTOR_TYPE_MLP_NORM, PROJECTOR_TYPE_LDP, PROJECTOR_TYPE_LDPV2, - PROJECTOR_TYPE_RESAMPLER, + PROJECTOR_TYPE_MINICPMV, PROJECTOR_TYPE_GLM_EDGE, - PROJECTOR_TYPE_MERGER, + PROJECTOR_TYPE_QWEN2VL, PROJECTOR_TYPE_GEMMA3, PROJECTOR_TYPE_IDEFICS3, PROJECTOR_TYPE_PIXTRAL, + PROJECTOR_TYPE_QWEN25VL, + PROJECTOR_TYPE_INTERNVL, PROJECTOR_TYPE_UNKNOWN, }; @@ -109,12 +117,14 @@ static std::map PROJECTOR_TYPE_NAMES = { { PROJECTOR_TYPE_MLP, "mlp" }, { PROJECTOR_TYPE_LDP, "ldp" }, { PROJECTOR_TYPE_LDPV2, "ldpv2"}, - { PROJECTOR_TYPE_RESAMPLER, "resampler"}, + { PROJECTOR_TYPE_MINICPMV, "resampler"}, { PROJECTOR_TYPE_GLM_EDGE, "adapter"}, - { PROJECTOR_TYPE_MERGER, "qwen2vl_merger"}, + { PROJECTOR_TYPE_QWEN2VL, "qwen2vl_merger"}, + { PROJECTOR_TYPE_QWEN25VL, "qwen2.5vl_merger"}, { PROJECTOR_TYPE_GEMMA3, "gemma3"}, { PROJECTOR_TYPE_IDEFICS3, "idefics3"}, { PROJECTOR_TYPE_PIXTRAL, "pixtral"}, + { PROJECTOR_TYPE_INTERNVL, "internvl"}, }; static projector_type clip_projector_type_from_string(const std::string & str) { @@ -229,6 +239,15 @@ struct clip_image_u8_batch { struct clip_image_f32_batch { std::vector entries; + + clip_image_f32_batch clone() const { + clip_image_f32_batch new_batch; + new_batch.entries.reserve(entries.size()); + for (const auto & entry : entries) { + new_batch.entries.emplace_back(new clip_image_f32(*entry)); + } + return new_batch; + } }; // diff --git a/examples/llava/clip-quantize-cli.cpp b/tools/mtmd/clip-quantize-cli.cpp similarity index 100% rename from examples/llava/clip-quantize-cli.cpp rename to tools/mtmd/clip-quantize-cli.cpp diff --git a/examples/llava/clip.cpp b/tools/mtmd/clip.cpp similarity index 50% rename from examples/llava/clip.cpp rename to tools/mtmd/clip.cpp index da8a590f0e..41ba45a79b 100644 --- a/examples/llava/clip.cpp +++ b/tools/mtmd/clip.cpp @@ -28,9 +28,22 @@ #include #include #include +#include +#include struct clip_logger_state g_logger_state = {GGML_LOG_LEVEL_CONT, clip_log_callback_default, NULL}; +enum ffn_op_type { + FFN_GELU, + FFN_SILU, + FFN_GELU_QUICK, +}; + +enum norm_type { + NORM_TYPE_NORMAL, + NORM_TYPE_RMS, +}; + //#define CLIP_DEBUG_FUNCTIONS #ifdef CLIP_DEBUG_FUNCTIONS @@ -154,13 +167,19 @@ enum patch_merge_type { struct clip_hparams { int32_t image_size; int32_t patch_size; - int32_t hidden_size; - int32_t n_intermediate; + int32_t n_embd; + int32_t n_ff; int32_t projection_dim; int32_t n_head; int32_t n_layer; int32_t proj_scale_factor = 0; // idefics3 + // for models using dynamic image size, we need to have a smaller image size to warmup + // otherwise, user will get OOM everytime they load the model + int32_t warmup_image_size = 0; + + ffn_op_type ffn_op = FFN_GELU; + patch_merge_type mm_patch_merge_type = PATCH_MERGE_FLAT; float eps = 1e-6; @@ -169,161 +188,161 @@ struct clip_hparams { std::vector image_grid_pinpoints; int32_t image_crop_resolution; std::unordered_set vision_feature_layer; + int32_t attn_window_size = 0; + int32_t n_wa_pattern = 0; + int32_t spatial_merge_size = 0; }; struct clip_layer { // attention - struct ggml_tensor * k_w = nullptr; - struct ggml_tensor * k_b = nullptr; - struct ggml_tensor * q_w = nullptr; - struct ggml_tensor * q_b = nullptr; - struct ggml_tensor * v_w = nullptr; - struct ggml_tensor * v_b = nullptr; + ggml_tensor * k_w = nullptr; + ggml_tensor * k_b = nullptr; + ggml_tensor * q_w = nullptr; + ggml_tensor * q_b = nullptr; + ggml_tensor * v_w = nullptr; + ggml_tensor * v_b = nullptr; - struct ggml_tensor * o_w = nullptr; - struct ggml_tensor * o_b = nullptr; + ggml_tensor * o_w = nullptr; + ggml_tensor * o_b = nullptr; + + ggml_tensor * k_norm = nullptr; + ggml_tensor * q_norm = nullptr; // layernorm 1 - struct ggml_tensor * ln_1_w = nullptr; - struct ggml_tensor * ln_1_b = nullptr; + ggml_tensor * ln_1_w = nullptr; + ggml_tensor * ln_1_b = nullptr; - // ff - struct ggml_tensor * ff_i_w = nullptr; // legacy naming - struct ggml_tensor * ff_i_b = nullptr; // legacy naming - struct ggml_tensor * ff_o_w = nullptr; // legacy naming - struct ggml_tensor * ff_o_b = nullptr; // legacy naming - - struct ggml_tensor * ff_up_w = nullptr; - struct ggml_tensor * ff_up_b = nullptr; - struct ggml_tensor * ff_gate_w = nullptr; - struct ggml_tensor * ff_gate_b = nullptr; - struct ggml_tensor * ff_down_w = nullptr; - struct ggml_tensor * ff_down_b = nullptr; + ggml_tensor * ff_up_w = nullptr; + ggml_tensor * ff_up_b = nullptr; + ggml_tensor * ff_gate_w = nullptr; + ggml_tensor * ff_gate_b = nullptr; + ggml_tensor * ff_down_w = nullptr; + ggml_tensor * ff_down_b = nullptr; // layernorm 2 - struct ggml_tensor * ln_2_w = nullptr; - struct ggml_tensor * ln_2_b = nullptr; + ggml_tensor * ln_2_w = nullptr; + ggml_tensor * ln_2_b = nullptr; + + // layer scale (no bias) + ggml_tensor * ls_1_w = nullptr; + ggml_tensor * ls_2_w = nullptr; }; struct clip_vision_model { struct clip_hparams hparams; // embeddings - struct ggml_tensor * class_embedding = nullptr; - struct ggml_tensor * patch_embeddings_0 = nullptr; - struct ggml_tensor * patch_embeddings_1 = nullptr; // second Conv2D kernel when we decouple Conv3D along temproal dimension (Qwen2VL) - struct ggml_tensor * patch_bias = nullptr; - struct ggml_tensor * position_embeddings = nullptr; + ggml_tensor * class_embedding = nullptr; + ggml_tensor * patch_embeddings_0 = nullptr; + ggml_tensor * patch_embeddings_1 = nullptr; // second Conv2D kernel when we decouple Conv3D along temproal dimension (Qwen2VL) + ggml_tensor * patch_bias = nullptr; + ggml_tensor * position_embeddings = nullptr; - struct ggml_tensor * pre_ln_w = nullptr; - struct ggml_tensor * pre_ln_b = nullptr; + ggml_tensor * pre_ln_w = nullptr; + ggml_tensor * pre_ln_b = nullptr; std::vector layers; - struct ggml_tensor * post_ln_w; - struct ggml_tensor * post_ln_b; + ggml_tensor * post_ln_w; + ggml_tensor * post_ln_b; - struct ggml_tensor * projection; + ggml_tensor * projection; // LLaVA projection - struct ggml_tensor * mm_0_w = nullptr; - struct ggml_tensor * mm_0_b = nullptr; - struct ggml_tensor * mm_2_w = nullptr; - struct ggml_tensor * mm_2_b = nullptr; + ggml_tensor * mm_input_norm_w = nullptr; + ggml_tensor * mm_0_w = nullptr; + ggml_tensor * mm_0_b = nullptr; + ggml_tensor * mm_2_w = nullptr; + ggml_tensor * mm_2_b = nullptr; - struct ggml_tensor * image_newline = nullptr; + ggml_tensor * image_newline = nullptr; // Yi type models with mlp+normalization projection - struct ggml_tensor * mm_1_w = nullptr; // Yi type models have 0, 1, 3, 4 - struct ggml_tensor * mm_1_b = nullptr; - struct ggml_tensor * mm_3_w = nullptr; - struct ggml_tensor * mm_3_b = nullptr; - struct ggml_tensor * mm_4_w = nullptr; - struct ggml_tensor * mm_4_b = nullptr; + ggml_tensor * mm_1_w = nullptr; // Yi type models have 0, 1, 3, 4 + ggml_tensor * mm_1_b = nullptr; + ggml_tensor * mm_3_w = nullptr; + ggml_tensor * mm_3_b = nullptr; + ggml_tensor * mm_4_w = nullptr; + ggml_tensor * mm_4_b = nullptr; - //GLMV-Edge projection - struct ggml_tensor * mm_model_adapter_conv_w = nullptr; - struct ggml_tensor * mm_model_adapter_conv_b = nullptr; + // GLMV-Edge projection + ggml_tensor * mm_model_adapter_conv_w = nullptr; + ggml_tensor * mm_model_adapter_conv_b = nullptr; + ggml_tensor * mm_glm_tok_boi = nullptr; + ggml_tensor * mm_glm_tok_eoi = nullptr; // MobileVLM projection - struct ggml_tensor * mm_model_mlp_1_w = nullptr; - struct ggml_tensor * mm_model_mlp_1_b = nullptr; - struct ggml_tensor * mm_model_mlp_3_w = nullptr; - struct ggml_tensor * mm_model_mlp_3_b = nullptr; - struct ggml_tensor * mm_model_block_1_block_0_0_w = nullptr; - struct ggml_tensor * mm_model_block_1_block_0_1_w = nullptr; - struct ggml_tensor * mm_model_block_1_block_0_1_b = nullptr; - struct ggml_tensor * mm_model_block_1_block_1_fc1_w = nullptr; - struct ggml_tensor * mm_model_block_1_block_1_fc1_b = nullptr; - struct ggml_tensor * mm_model_block_1_block_1_fc2_w = nullptr; - struct ggml_tensor * mm_model_block_1_block_1_fc2_b = nullptr; - struct ggml_tensor * mm_model_block_1_block_2_0_w = nullptr; - struct ggml_tensor * mm_model_block_1_block_2_1_w = nullptr; - struct ggml_tensor * mm_model_block_1_block_2_1_b = nullptr; - struct ggml_tensor * mm_model_block_2_block_0_0_w = nullptr; - struct ggml_tensor * mm_model_block_2_block_0_1_w = nullptr; - struct ggml_tensor * mm_model_block_2_block_0_1_b = nullptr; - struct ggml_tensor * mm_model_block_2_block_1_fc1_w = nullptr; - struct ggml_tensor * mm_model_block_2_block_1_fc1_b = nullptr; - struct ggml_tensor * mm_model_block_2_block_1_fc2_w = nullptr; - struct ggml_tensor * mm_model_block_2_block_1_fc2_b = nullptr; - struct ggml_tensor * mm_model_block_2_block_2_0_w = nullptr; - struct ggml_tensor * mm_model_block_2_block_2_1_w = nullptr; - struct ggml_tensor * mm_model_block_2_block_2_1_b = nullptr; + ggml_tensor * mm_model_mlp_1_w = nullptr; + ggml_tensor * mm_model_mlp_1_b = nullptr; + ggml_tensor * mm_model_mlp_3_w = nullptr; + ggml_tensor * mm_model_mlp_3_b = nullptr; + ggml_tensor * mm_model_block_1_block_0_0_w = nullptr; + ggml_tensor * mm_model_block_1_block_0_1_w = nullptr; + ggml_tensor * mm_model_block_1_block_0_1_b = nullptr; + ggml_tensor * mm_model_block_1_block_1_fc1_w = nullptr; + ggml_tensor * mm_model_block_1_block_1_fc1_b = nullptr; + ggml_tensor * mm_model_block_1_block_1_fc2_w = nullptr; + ggml_tensor * mm_model_block_1_block_1_fc2_b = nullptr; + ggml_tensor * mm_model_block_1_block_2_0_w = nullptr; + ggml_tensor * mm_model_block_1_block_2_1_w = nullptr; + ggml_tensor * mm_model_block_1_block_2_1_b = nullptr; + ggml_tensor * mm_model_block_2_block_0_0_w = nullptr; + ggml_tensor * mm_model_block_2_block_0_1_w = nullptr; + ggml_tensor * mm_model_block_2_block_0_1_b = nullptr; + ggml_tensor * mm_model_block_2_block_1_fc1_w = nullptr; + ggml_tensor * mm_model_block_2_block_1_fc1_b = nullptr; + ggml_tensor * mm_model_block_2_block_1_fc2_w = nullptr; + ggml_tensor * mm_model_block_2_block_1_fc2_b = nullptr; + ggml_tensor * mm_model_block_2_block_2_0_w = nullptr; + ggml_tensor * mm_model_block_2_block_2_1_w = nullptr; + ggml_tensor * mm_model_block_2_block_2_1_b = nullptr; // MobileVLM_V2 projection - struct ggml_tensor * mm_model_mlp_0_w = nullptr; - struct ggml_tensor * mm_model_mlp_0_b = nullptr; - struct ggml_tensor * mm_model_mlp_2_w = nullptr; - struct ggml_tensor * mm_model_mlp_2_b = nullptr; - struct ggml_tensor * mm_model_peg_0_w = nullptr; - struct ggml_tensor * mm_model_peg_0_b = nullptr; + ggml_tensor * mm_model_mlp_0_w = nullptr; + ggml_tensor * mm_model_mlp_0_b = nullptr; + ggml_tensor * mm_model_mlp_2_w = nullptr; + ggml_tensor * mm_model_mlp_2_b = nullptr; + ggml_tensor * mm_model_peg_0_w = nullptr; + ggml_tensor * mm_model_peg_0_b = nullptr; // MINICPMV projection - struct ggml_tensor * mm_model_pos_embed_k = nullptr; - struct ggml_tensor * mm_model_query = nullptr; - struct ggml_tensor * mm_model_proj = nullptr; - struct ggml_tensor * mm_model_kv_proj = nullptr; - struct ggml_tensor * mm_model_attn_q_w = nullptr; - struct ggml_tensor * mm_model_attn_q_b = nullptr; - struct ggml_tensor * mm_model_attn_k_w = nullptr; - struct ggml_tensor * mm_model_attn_k_b = nullptr; - struct ggml_tensor * mm_model_attn_v_w = nullptr; - struct ggml_tensor * mm_model_attn_v_b = nullptr; - struct ggml_tensor * mm_model_attn_o_w = nullptr; - struct ggml_tensor * mm_model_attn_o_b = nullptr; - struct ggml_tensor * mm_model_ln_q_w = nullptr; - struct ggml_tensor * mm_model_ln_q_b = nullptr; - struct ggml_tensor * mm_model_ln_kv_w = nullptr; - struct ggml_tensor * mm_model_ln_kv_b = nullptr; - struct ggml_tensor * mm_model_ln_post_w = nullptr; - struct ggml_tensor * mm_model_ln_post_b = nullptr; + ggml_tensor * mm_model_pos_embed_k = nullptr; + ggml_tensor * mm_model_query = nullptr; + ggml_tensor * mm_model_proj = nullptr; + ggml_tensor * mm_model_kv_proj = nullptr; + ggml_tensor * mm_model_attn_q_w = nullptr; + ggml_tensor * mm_model_attn_q_b = nullptr; + ggml_tensor * mm_model_attn_k_w = nullptr; + ggml_tensor * mm_model_attn_k_b = nullptr; + ggml_tensor * mm_model_attn_v_w = nullptr; + ggml_tensor * mm_model_attn_v_b = nullptr; + ggml_tensor * mm_model_attn_o_w = nullptr; + ggml_tensor * mm_model_attn_o_b = nullptr; + ggml_tensor * mm_model_ln_q_w = nullptr; + ggml_tensor * mm_model_ln_q_b = nullptr; + ggml_tensor * mm_model_ln_kv_w = nullptr; + ggml_tensor * mm_model_ln_kv_b = nullptr; + ggml_tensor * mm_model_ln_post_w = nullptr; + ggml_tensor * mm_model_ln_post_b = nullptr; // gemma3 - struct ggml_tensor * mm_input_proj_w = nullptr; - struct ggml_tensor * mm_soft_emb_norm_w = nullptr; + ggml_tensor * mm_input_proj_w = nullptr; + ggml_tensor * mm_soft_emb_norm_w = nullptr; // pixtral - struct ggml_tensor * token_embd_img_break = nullptr; + ggml_tensor * token_embd_img_break = nullptr; + ggml_tensor * mm_patch_merger_w = nullptr; }; struct clip_ctx { - bool has_text_encoder = false; - bool has_vision_encoder = false; bool has_llava_projector = false; - bool has_minicpmv_projector = false; - bool has_glm_projector = false; - bool has_qwen2vl_merger = false; - int minicpmv_version = 2; + int minicpmv_version = 0; struct clip_vision_model vision_model; projector_type proj_type = PROJECTOR_TYPE_MLP; - int32_t max_feature_layer; // unused in newer models like gemma3 float image_mean[3]; float image_std[3]; - bool use_gelu = false; - bool use_silu = false; gguf_context_ptr ctx_gguf; ggml_context_ptr ctx_data; @@ -344,9 +363,12 @@ struct clip_ctx { clip_ctx(clip_context_params & ctx_params) { backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr); - backend = ctx_params.use_gpu - ? ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr) - : nullptr; + if (!backend_cpu) { + throw std::runtime_error("failed to initialize CPU backend"); + } + backend = ctx_params.use_gpu + ? ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr) + : nullptr; if (backend) { LOG_INF("%s: CLIP using %s backend\n", __func__, ggml_backend_name(backend)); @@ -361,7 +383,7 @@ struct clip_ctx { backend_buft.push_back(ggml_backend_get_default_buffer_type(backend_cpu)); sched.reset( - ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), 8192, false) + ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), 8192, false, true) ); } @@ -373,952 +395,362 @@ struct clip_ctx { } }; -static ggml_cgraph * clip_image_build_graph_siglip(clip_ctx * ctx, const clip_image_f32_batch & imgs) { - const auto & model = ctx->vision_model; - const auto & hparams = model.hparams; +struct clip_graph { + clip_ctx * ctx; + const clip_vision_model & model; + const clip_hparams & hparams; - const int image_size = hparams.image_size; - int image_size_width = image_size; - int image_size_height = image_size; + // we only support single image per batch + const clip_image_f32 & img; - const int patch_size = hparams.patch_size; - const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size)); - const int hidden_size = hparams.hidden_size; - const int n_head = hparams.n_head; - const int d_head = hidden_size / n_head; - const int n_layer = hparams.n_layer; - const float eps = hparams.eps; + const int patch_size; + const int n_patches_x; + const int n_patches_y; + const int n_patches; + const int n_embd; + const int n_head; + const int d_head; + const int n_layer; + const float eps; + const float kq_scale; - GGML_ASSERT(imgs.entries.size() == 1); // batch_size == 1 + ggml_context_ptr ctx0_ptr; + ggml_context * ctx0; + ggml_cgraph * gf; - struct ggml_init_params params = { - /*.mem_size =*/ ctx->buf_compute_meta.size(), - /*.mem_buffer =*/ ctx->buf_compute_meta.data(), - /*.no_alloc =*/ true, - }; - - ggml_context_ptr ctx0_ptr(ggml_init(params)); - auto ctx0 = ctx0_ptr.get(); - - struct ggml_cgraph * gf = ggml_new_graph(ctx0); - - // input raw - struct ggml_tensor * inp_raw = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, 3); - ggml_set_name(inp_raw, "inp_raw"); - ggml_set_input(inp_raw); - - struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1); - inp = ggml_reshape_2d(ctx0, inp, num_patches, hidden_size); - inp = ggml_cont(ctx0, ggml_transpose(ctx0, inp)); - inp = ggml_add(ctx0, inp, model.patch_bias); - - // position embeddings - struct ggml_tensor * embeddings = ggml_add(ctx0, inp, model.position_embeddings); - - // loop over layers - for (int il = 0; il < n_layer; il++) { - struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states - - // layernorm1 - { - cur = ggml_norm(ctx0, cur, eps); - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_1_w), model.layers[il].ln_1_b); - } - - // self-attention - { - - struct ggml_tensor * Q = - ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].q_w, cur), model.layers[il].q_b); - - Q = ggml_reshape_3d(ctx0, Q, d_head, n_head, num_patches); - Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3)); - - struct ggml_tensor * K = - ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].k_w, cur), model.layers[il].k_b); - - K = ggml_reshape_3d(ctx0, K, d_head, n_head, num_patches); - K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3)); - - struct ggml_tensor * V = - ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].v_w, cur), model.layers[il].v_b); - - V = ggml_reshape_3d(ctx0, V, d_head, n_head, num_patches); - V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3)); - - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - KQ = ggml_soft_max_ext(ctx0, KQ, nullptr, 1.0f / sqrtf((float)d_head), 0.0f); - - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ); - KQV = ggml_reshape_3d(ctx0, KQV, d_head, num_patches, n_head); - KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - - cur = ggml_cont_2d(ctx0, KQV, hidden_size, num_patches); - } - - // attention output - cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].o_w, cur), model.layers[il].o_b); - - // re-add the layer input, e.g., residual - cur = ggml_add(ctx0, cur, embeddings); - - embeddings = cur; // embeddings = residual, cur = hidden_states - - // layernorm2 - { - cur = ggml_norm(ctx0, cur, eps); - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_2_w), model.layers[il].ln_2_b); - } - - cur = ggml_mul_mat(ctx0, model.layers[il].ff_i_w, cur); - cur = ggml_add(ctx0, cur, model.layers[il].ff_i_b); - - // siglip uses gelu - cur = ggml_gelu(ctx0, cur); - - cur = ggml_mul_mat(ctx0, model.layers[il].ff_o_w, cur); - cur = ggml_add(ctx0, cur, model.layers[il].ff_o_b); - - // residual 2 - cur = ggml_add(ctx0, embeddings, cur); - - embeddings = cur; + clip_graph(clip_ctx * ctx, const clip_image_f32 & img) : + ctx(ctx), + model(ctx->vision_model), + hparams(model.hparams), + img(img), + patch_size(hparams.patch_size), + n_patches_x(img.nx / patch_size), + n_patches_y(img.ny / patch_size), + n_patches(n_patches_x * n_patches_y), + n_embd(hparams.n_embd), + n_head(hparams.n_head), + d_head(n_embd / n_head), + n_layer(hparams.n_layer), + eps(hparams.eps), + kq_scale(1.0f / sqrtf((float)d_head)) { + struct ggml_init_params params = { + /*.mem_size =*/ ctx->buf_compute_meta.size(), + /*.mem_buffer =*/ ctx->buf_compute_meta.data(), + /*.no_alloc =*/ true, + }; + ctx0_ptr.reset(ggml_init(params)); + ctx0 = ctx0_ptr.get(); + gf = ggml_new_graph(ctx0); } - // post-layernorm - if (model.post_ln_w) { - embeddings = ggml_norm(ctx0, embeddings, eps); - ggml_set_name(embeddings, "post_ln"); - - embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.post_ln_w), model.post_ln_b); - } - - if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) { - const int batch_size = 1; - const int mm_tokens_per_image = 256; // default value for gemma3 - const int tokens_per_side = sqrt(mm_tokens_per_image); - const int patches_per_image = sqrt(num_patches); - const int kernel_size = patches_per_image / tokens_per_side; - - embeddings = ggml_cont(ctx0, ggml_transpose(ctx0, embeddings)); - embeddings = ggml_reshape_4d(ctx0, embeddings, patches_per_image, patches_per_image, hidden_size, batch_size); - - // doing a pool2d to reduce the number of output tokens to 256 - embeddings = ggml_pool_2d(ctx0, embeddings, GGML_OP_POOL_AVG, kernel_size, kernel_size, kernel_size, kernel_size, 0, 0); - embeddings = ggml_reshape_3d(ctx0, embeddings, embeddings->ne[0] * embeddings->ne[0], hidden_size, batch_size); - embeddings = ggml_cont(ctx0, ggml_transpose(ctx0, embeddings)); - - // apply norm before projection - embeddings = ggml_rms_norm(ctx0, embeddings, eps); - embeddings = ggml_mul(ctx0, embeddings, model.mm_soft_emb_norm_w); - - // apply projection - embeddings = ggml_mul_mat(ctx0, - ggml_cont(ctx0, ggml_transpose(ctx0, model.mm_input_proj_w)), - embeddings); - - } else if (ctx->proj_type == PROJECTOR_TYPE_IDEFICS3) { - // https://github.com/huggingface/transformers/blob/0a950e0bbe1ed58d5401a6b547af19f15f0c195e/src/transformers/models/idefics3/modeling_idefics3.py#L578 - - ggml_tensor * cur = embeddings; - const int scale_factor = model.hparams.proj_scale_factor; - const int n_embd = cur->ne[0]; - const int seq = cur->ne[1]; - const int bsz = 1; // batch size, always 1 for now since we don't support batching - const int height = std::sqrt(seq); - const int width = std::sqrt(seq); - GGML_ASSERT(scale_factor != 0); - cur = ggml_reshape_4d(ctx0, cur, n_embd * scale_factor, width / scale_factor, height, bsz); - cur = ggml_permute(ctx0, cur, 0, 2, 1, 3); - cur = ggml_reshape_4d(ctx0, ggml_cont(ctx0, cur), - n_embd * scale_factor * scale_factor, - height / scale_factor, - width / scale_factor, - bsz); - cur = ggml_permute(ctx0, cur, 0, 2, 1, 3); - cur = ggml_reshape_3d(ctx0, ggml_cont(ctx0, cur), - n_embd * scale_factor * scale_factor, - seq / (scale_factor * scale_factor), - bsz); - - cur = ggml_mul_mat(ctx0, model.projection, cur); - embeddings = cur; - } else { - GGML_ABORT("SigLIP: Unsupported projector type"); - } - - // build the graph - ggml_build_forward_expand(gf, embeddings); - - return gf; -} - -// implementation of the 2D RoPE without adding a new op in ggml -// this is not efficient (use double the memory), but works on all backends -// TODO: there was a more efficient which relies on ggml_view and ggml_rope_ext_inplace, but the rope inplace does not work well with non-contiguous tensors ; we should fix that and revert back to the original implementation in https://github.com/ggml-org/llama.cpp/pull/13065 -static ggml_tensor * build_rope_2d( - ggml_context * ctx0, - ggml_tensor * cur, - ggml_tensor * pos_h, - ggml_tensor * pos_w, - const float freq_base -) { - const int64_t n_dim = cur->ne[0]; - const int64_t n_head = cur->ne[1]; - const int64_t n_pos = cur->ne[2]; - - // for example, if we have cur tensor of shape (n_dim=8, n_head, n_pos) - // we will have a list of 4 inv_freq: 1e-0, 1e-1, 1e-2, 1e-3 - // first half of cur will use 1e-0, 1e-2 (even) - // second half of cur will use 1e-1, 1e-3 (odd) - // the trick here is to rotate just half of n_dim, so inv_freq will automatically be even - // ^ don't ask me why, it's math! -2(2i) / n_dim == -2i / (n_dim/2) - // then for the second half, we use freq_scale to shift the inv_freq - // ^ why? replace (2i) with (2i+1) in the above equation - const float freq_scale_odd = std::pow(freq_base, (float)-2/n_dim); - - // first half - ggml_tensor * first; - { - first = ggml_view_3d(ctx0, cur, - n_dim/2, n_head, n_pos, - ggml_row_size(cur->type, n_dim), - ggml_row_size(cur->type, n_dim*n_head), - 0); - first = ggml_rope_ext( - ctx0, - first, - pos_h, // positions - nullptr, // freq factors - n_dim/2, // n_dims - 0, 0, freq_base, - 1.0f, 0.0f, 1.0f, 0.0f, 0.0f - ); - } - - // second half - ggml_tensor * second; - { - second = ggml_view_3d(ctx0, cur, - n_dim/2, n_head, n_pos, - ggml_row_size(cur->type, n_dim), - ggml_row_size(cur->type, n_dim*n_head), - n_dim/2 * ggml_element_size(cur)); - second = ggml_cont(ctx0, second); // copy, because ggml_rope don't play well with non-contiguous tensors - second = ggml_rope_ext( - ctx0, - second, - pos_w, // positions - nullptr, // freq factors - n_dim/2, // n_dims - 0, 0, freq_base, - freq_scale_odd, - 0.0f, 1.0f, 0.0f, 0.0f - ); - } - - cur = ggml_concat(ctx0, first, second, 0); - return cur; -} - -static ggml_cgraph * clip_image_build_graph_pixtral(clip_ctx * ctx, const clip_image_f32_batch & imgs) { - const auto & model = ctx->vision_model; - const auto & hparams = model.hparams; - - GGML_ASSERT(ctx->proj_type == PROJECTOR_TYPE_PIXTRAL); - GGML_ASSERT(imgs.entries.size() == 1); // batch_size == 1 - - int image_size_width = imgs.entries[0]->nx; - int image_size_height = imgs.entries[0]->ny; - - const int patch_size = hparams.patch_size; - const int n_patches_x = image_size_width / patch_size; - const int n_patches_y = image_size_height / patch_size; - const int num_patches = n_patches_x * n_patches_y; - const int hidden_size = hparams.hidden_size; - const int n_head = hparams.n_head; - const int d_head = hidden_size / n_head; - const int n_layer = hparams.n_layer; - const float eps = hparams.eps; - - struct ggml_init_params params = { - /*.mem_size =*/ ctx->buf_compute_meta.size(), - /*.mem_buffer =*/ ctx->buf_compute_meta.data(), - /*.no_alloc =*/ true, - }; - - ggml_context_ptr ctx0_ptr(ggml_init(params)); - auto ctx0 = ctx0_ptr.get(); - - struct ggml_cgraph * gf = ggml_new_graph(ctx0); - - // input raw - struct ggml_tensor * inp_raw = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, 3); - ggml_set_name(inp_raw, "inp_raw"); - ggml_set_input(inp_raw); - - // 2D input positions - struct ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_patches); - ggml_set_name(pos_h, "pos_h"); - ggml_set_input(pos_h); - struct ggml_tensor * pos_w = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_patches); - ggml_set_name(pos_w, "pos_w"); - ggml_set_input(pos_w); - - struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1); - inp = ggml_reshape_2d(ctx0, inp, num_patches, hidden_size); - inp = ggml_cont(ctx0, ggml_transpose(ctx0, inp)); - - struct ggml_tensor * embeddings = inp; - - // pre-layer norm - embeddings = ggml_mul(ctx0, ggml_rms_norm(ctx0, embeddings, eps), model.pre_ln_w); - - // loop over layers - for (int il = 0; il < n_layer; il++) { - struct ggml_tensor * cur = embeddings; - - // pre-attention norm - cur = ggml_mul(ctx0, ggml_rms_norm(ctx0, cur, eps), model.layers[il].ln_1_w); - - // self-attention - { - struct ggml_tensor * Q = ggml_mul_mat(ctx0, model.layers[il].q_w, cur); - - Q = ggml_reshape_3d(ctx0, Q, d_head, n_head, num_patches); - Q = build_rope_2d(ctx0, Q, pos_h, pos_w, hparams.rope_theta); - Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3)); - - struct ggml_tensor * K = ggml_mul_mat(ctx0, model.layers[il].k_w, cur); - - K = ggml_reshape_3d(ctx0, K, d_head, n_head, num_patches); - K = build_rope_2d(ctx0, K, pos_h, pos_w, hparams.rope_theta); - K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3)); - - struct ggml_tensor * V = ggml_mul_mat(ctx0, model.layers[il].v_w, cur); - - V = ggml_reshape_3d(ctx0, V, d_head, n_head, num_patches); - V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3)); - - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - KQ = ggml_soft_max_ext(ctx0, KQ, nullptr, 1.0f / sqrtf((float)d_head), 0.0f); - - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ); - KQV = ggml_reshape_3d(ctx0, KQV, d_head, num_patches, n_head); - KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - - cur = ggml_cont_2d(ctx0, KQV, hidden_size, num_patches); - - cur = ggml_mul_mat(ctx0, model.layers[il].o_w, cur); - } - - // re-add the layer input, e.g., residual - cur = ggml_add(ctx0, cur, embeddings); - - embeddings = cur; // embeddings = residual, cur = hidden_states - - // pre-ffn norm - cur = ggml_mul(ctx0, ggml_rms_norm(ctx0, cur, eps), model.layers[il].ln_2_w); - - // feed-forward - { - ggml_tensor * gate_proj = ggml_mul_mat(ctx0, model.layers[il].ff_gate_w, cur); - ggml_tensor * up_proj = ggml_mul_mat(ctx0, model.layers[il].ff_up_w, cur); - gate_proj = ggml_silu(ctx0, gate_proj); // pixtral uses silu - cur = ggml_mul(ctx0, up_proj, gate_proj); - cur = ggml_mul_mat(ctx0, model.layers[il].ff_down_w, cur); - } - - // residual 2 - cur = ggml_add(ctx0, embeddings, cur); - - embeddings = cur; - } - - // LlavaMultiModalProjector (with GELU activation) - { - embeddings = ggml_mul_mat(ctx0, model.mm_1_w, embeddings); - embeddings = ggml_add(ctx0, embeddings, model.mm_1_b); - - embeddings = ggml_gelu(ctx0, embeddings); - embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings); - embeddings = ggml_add(ctx0, embeddings, model.mm_2_b); - } - - // arrangement of the [IMG_BREAK] token - { - // not efficient, but works - // the trick is to view the embeddings as a 3D tensor with shape [hidden_size, n_patches_per_row, n_rows] - // and then concatenate the [IMG_BREAK] token to the end of each row, aka n_patches_per_row dimension - // after the concatenation, we have a tensor with shape [hidden_size, n_patches_per_row + 1, n_rows] - - const int n_embd_text = embeddings->ne[0]; - const int n_tokens_output = num_patches + n_patches_y - 1; // one [IMG_BREAK] per row, except the last row - - ggml_tensor * cur = ggml_reshape_3d(ctx0, embeddings, n_embd_text, n_patches_x, n_patches_y); - ggml_tensor * tok = ggml_new_tensor_3d(ctx0, embeddings->type, n_embd_text, 1, n_patches_y); - tok = ggml_scale(ctx0, tok, 0.0); // clear the tensor - tok = ggml_add(ctx0, tok, model.token_embd_img_break); - cur = ggml_concat(ctx0, cur, tok, 1); - embeddings = ggml_view_2d(ctx0, cur, - n_embd_text, n_tokens_output, - ggml_row_size(cur->type, n_embd_text), 0); - } - - // build the graph - ggml_build_forward_expand(gf, embeddings); - - return gf; -} - -static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_image_f32_batch & imgs, struct clip_image_size load_image_size, bool is_inf = false) { - if (!ctx->has_vision_encoder) { - LOG_ERR("This gguf file seems to have no vision encoder\n"); - return nullptr; - } - - const auto & model = ctx->vision_model; - const auto & hparams = model.hparams; - - const int image_size = hparams.image_size; - int image_size_width = image_size; - int image_size_height = image_size; - if (ctx->has_minicpmv_projector) { - LOG_DBG("%s: %d %d\n", __func__, load_image_size.width, load_image_size.height); - image_size_width = load_image_size.width; - image_size_height = load_image_size.height; - if (is_inf) { - image_size_width = imgs.entries[0]->nx; - image_size_height = imgs.entries[0]->ny; - } - } - else if (ctx->has_qwen2vl_merger) { - // use the image's native resolution when image is avaible - if (is_inf) { - // if (imgs->data->nx && imgs->data->ny) { - image_size_width = imgs.entries[0]->nx; - image_size_height = imgs.entries[0]->ny; - } - } - const int patch_size = hparams.patch_size; - const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size)); - const int patches_w = image_size_width / patch_size; - const int patches_h = image_size_height / patch_size; - const int num_positions = num_patches + (model.class_embedding ? 1 : 0); - const int num_position_ids = ctx->has_qwen2vl_merger ? num_positions * 4 : num_positions; - const int hidden_size = hparams.hidden_size; - const int n_head = hparams.n_head; - const int d_head = hidden_size / n_head; - const float eps = hparams.eps; - int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4}; - - const int batch_size = imgs.entries.size(); - - if (ctx->has_llava_projector || ctx->has_minicpmv_projector || ctx->has_glm_projector) { - GGML_ASSERT(batch_size == 1); - } - - struct ggml_init_params params = { - /*.mem_size =*/ ctx->buf_compute_meta.size(), - /*.mem_buffer =*/ ctx->buf_compute_meta.data(), - /*.no_alloc =*/ true, - }; - - ggml_context_ptr ctx0_ptr(ggml_init(params)); - auto ctx0 = ctx0_ptr.get(); - - struct ggml_cgraph * gf = ggml_new_graph(ctx0); - - struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, 3, batch_size); - ggml_set_name(inp_raw, "inp_raw"); - ggml_set_input(inp_raw); - - struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1); - - if (ctx->has_qwen2vl_merger) { - GGML_ASSERT(image_size_width % (patch_size * 2) == 0); - GGML_ASSERT(image_size_height % (patch_size * 2) == 0); - - auto inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1); - inp = ggml_add(ctx0, inp, inp_1); - inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 2, 0, 3)); // [w, h, c, b] -> [c, w, h, b] - inp = ggml_reshape_4d( - ctx0, inp, - hidden_size * 2, patches_w / 2, patches_h, batch_size); - inp = ggml_reshape_4d( - ctx0, inp, - hidden_size * 2, patches_w / 2, 2, batch_size * (patches_h / 2)); - inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 0, 2, 1, 3)); - inp = ggml_reshape_3d( - ctx0, inp, - hidden_size, patches_w * patches_h, batch_size); - } - else { - inp = ggml_reshape_3d(ctx0, inp, num_patches, hidden_size, batch_size); - inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3)); - } - - if (model.patch_bias) { - // inp = ggml_add(ctx0, inp, ggml_repeat(ctx0, model.patch_bias, inp)); - inp = ggml_add(ctx0, inp, model.patch_bias); - } - struct ggml_tensor * embeddings = inp; - struct ggml_tensor * pos_embed = nullptr; - - if (ctx->has_llava_projector) { - // concat class_embeddings and patch_embeddings - if (model.class_embedding) { - embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size); - ggml_set_name(embeddings, "embeddings"); - ggml_set_input(embeddings); - embeddings = ggml_acc(ctx0, embeddings, model.class_embedding, - embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0); - embeddings = ggml_acc(ctx0, embeddings, inp, - embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]); - } - } - - struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_position_ids); - ggml_set_name(positions, "positions"); - ggml_set_input(positions); - - if (!ctx->has_qwen2vl_merger) { // qwen2vl use rope position embedding - embeddings = - ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions)); - } - - if (ctx->has_minicpmv_projector) { - int pos_w = image_size_width/patch_size; - int pos_h = image_size_height/patch_size; - if (ctx->minicpmv_version == 2) { - pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 4096, pos_w * pos_h, 1); - } - else if (ctx->minicpmv_version == 3) { - pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 3584, pos_w * pos_h, 1); - } - else if (ctx->minicpmv_version == 4) { - pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 3584, pos_w * pos_h, 1); - } - ggml_set_name(pos_embed, "pos_embed"); - ggml_set_input(pos_embed); - } - - // pre-layernorm - if (model.pre_ln_w) { - embeddings = ggml_norm(ctx0, embeddings, eps); - ggml_set_name(embeddings, "pre_ln"); - - embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.pre_ln_w), model.pre_ln_b); - } - - std::vector embedding_stack; - const auto & vision_feature_layer = hparams.vision_feature_layer; - - // loop over layers - for (int il = 0; il < ctx->max_feature_layer; il++) { - struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states - - // If this is an embedding feature layer, save the output. - // NOTE: 0 index here refers to the input to the encoder. - if (vision_feature_layer.find(il) != vision_feature_layer.end()) { - embedding_stack.push_back(embeddings); - } - - //const size_t nb_q_w = model.layers[il].q_w->nb[0]; - - // layernorm1 - { - cur = ggml_norm(ctx0, cur, eps); - - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_1_w), - model.layers[il].ln_1_b); - } - - // self-attention - { - - struct ggml_tensor * Q = - ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].q_w, cur), model.layers[il].q_b); - - Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_positions, batch_size); - if (ctx->has_qwen2vl_merger) { - Q = ggml_rope_multi( - ctx0, Q, positions, nullptr, - d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1); - } - Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3)); - Q = ggml_reshape_3d(ctx0, Q, d_head, num_positions, n_head * batch_size); - - struct ggml_tensor * K = - ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].k_w, cur), model.layers[il].k_b); - - K = ggml_reshape_4d(ctx0, K, d_head, n_head, num_positions, batch_size); - if (ctx->has_qwen2vl_merger) { - K = ggml_rope_multi( - ctx0, K, positions, nullptr, - d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1); - } - K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3)); - K = ggml_reshape_3d(ctx0, K, d_head, num_positions, n_head * batch_size); - - struct ggml_tensor * V = - ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].v_w, cur), model.layers[il].v_b); - - V = ggml_reshape_4d(ctx0, V, d_head, n_head, num_positions, batch_size); - V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3)); - V = ggml_reshape_3d(ctx0, V, num_positions, d_head, n_head * batch_size); - - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - KQ = ggml_soft_max_ext(ctx0, KQ, nullptr, 1.0f / sqrtf((float)d_head), 0.0f); - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ); - KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_positions, n_head, batch_size); - KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - - cur = ggml_cont_3d(ctx0, KQV, hidden_size, num_positions, batch_size); - } - - // attention output - cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].o_w, cur), model.layers[il].o_b); - - // re-add the layer input, e.g., residual - cur = ggml_add(ctx0, cur, embeddings); - - embeddings = cur; // embeddings = residual, cur = hidden_states - - // layernorm2 - { - cur = ggml_norm(ctx0, cur, eps); - - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_2_w), model.layers[il].ln_2_b); - } - - cur = ggml_mul_mat(ctx0, model.layers[il].ff_i_w, cur); - cur = ggml_add(ctx0, cur, model.layers[il].ff_i_b); - - if (ctx->use_gelu) { - cur = ggml_gelu_inplace(ctx0, cur); - } else if (ctx->use_silu) { - cur = ggml_silu_inplace(ctx0, cur); + ggml_cgraph * build_siglip() { + ggml_tensor * inp = build_inp(); + ggml_tensor * cur = build_vit( + inp, n_patches, + NORM_TYPE_NORMAL, + hparams.ffn_op, + model.position_embeddings, + nullptr); + + if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) { + const int batch_size = 1; + GGML_ASSERT(n_patches_x == n_patches_y); + const int patches_per_image = n_patches_x; + const int kernel_size = hparams.proj_scale_factor; + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); + cur = ggml_reshape_4d(ctx0, cur, patches_per_image, patches_per_image, n_embd, batch_size); + + // doing a pool2d to reduce the number of output tokens + cur = ggml_pool_2d(ctx0, cur, GGML_OP_POOL_AVG, kernel_size, kernel_size, kernel_size, kernel_size, 0, 0); + cur = ggml_reshape_3d(ctx0, cur, cur->ne[0] * cur->ne[0], n_embd, batch_size); + cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); + + // apply norm before projection + cur = ggml_rms_norm(ctx0, cur, eps); + cur = ggml_mul(ctx0, cur, model.mm_soft_emb_norm_w); + + // apply projection + cur = ggml_mul_mat(ctx0, + ggml_cont(ctx0, ggml_transpose(ctx0, model.mm_input_proj_w)), + cur); + + } else if (ctx->proj_type == PROJECTOR_TYPE_IDEFICS3) { + // https://github.com/huggingface/transformers/blob/0a950e0bbe1ed58d5401a6b547af19f15f0c195e/src/transformers/models/idefics3/modeling_idefics3.py#L578 + + const int scale_factor = model.hparams.proj_scale_factor; + const int n_embd = cur->ne[0]; + const int seq = cur->ne[1]; + const int bsz = 1; // batch size, always 1 for now since we don't support batching + const int height = std::sqrt(seq); + const int width = std::sqrt(seq); + GGML_ASSERT(scale_factor != 0); + cur = ggml_reshape_4d(ctx0, cur, n_embd * scale_factor, width / scale_factor, height, bsz); + cur = ggml_permute(ctx0, cur, 0, 2, 1, 3); + cur = ggml_reshape_4d(ctx0, ggml_cont(ctx0, cur), + n_embd * scale_factor * scale_factor, + height / scale_factor, + width / scale_factor, + bsz); + cur = ggml_permute(ctx0, cur, 0, 2, 1, 3); + cur = ggml_reshape_3d(ctx0, ggml_cont(ctx0, cur), + n_embd * scale_factor * scale_factor, + seq / (scale_factor * scale_factor), + bsz); + + cur = ggml_mul_mat(ctx0, model.projection, cur); } else { - cur = ggml_gelu_quick_inplace(ctx0, cur); + GGML_ABORT("SigLIP: Unsupported projector type"); } - cur = ggml_mul_mat(ctx0, model.layers[il].ff_o_w, cur); - cur = ggml_add(ctx0, cur, model.layers[il].ff_o_b); + // build the graph + ggml_build_forward_expand(gf, cur); - // residual 2 - cur = ggml_add(ctx0, embeddings, cur); - - embeddings = cur; + return gf; } - // post-layernorm - if (model.post_ln_w) { - embeddings = ggml_norm(ctx0, embeddings, eps); - ggml_set_name(embeddings, "post_ln"); + ggml_cgraph * build_pixtral() { + const int n_merge = hparams.spatial_merge_size; - embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.post_ln_w), model.post_ln_b); - } + // 2D input positions + ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches); + ggml_set_name(pos_h, "pos_h"); + ggml_set_input(pos_h); - // final layer is a vision feature layer - if (vision_feature_layer.find(ctx->max_feature_layer) != vision_feature_layer.end()) { - embedding_stack.push_back(embeddings); - } + ggml_tensor * pos_w = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches); + ggml_set_name(pos_w, "pos_w"); + ggml_set_input(pos_w); - // If feature layers are explicitly set, stack them (if we have multiple) - if (!embedding_stack.empty()) { - embeddings = embedding_stack[0]; - for (size_t i = 1; i < embedding_stack.size(); i++) { - embeddings = ggml_concat(ctx0, embeddings, embedding_stack[i], 0); + auto add_pos = [&](ggml_tensor * cur, const clip_layer &) { + return build_rope_2d(ctx0, cur, pos_h, pos_w, hparams.rope_theta); + }; + + ggml_tensor * inp = build_inp(); + ggml_tensor * cur = build_vit( + inp, n_patches, + NORM_TYPE_RMS, + hparams.ffn_op, + nullptr, // no learned pos embd + add_pos); + + // mistral small 3.1 patch merger + // ref: https://github.com/huggingface/transformers/blob/7a3e208892c06a5e278144eaf38c8599a42f53e7/src/transformers/models/mistral3/modeling_mistral3.py#L67 + if (model.mm_patch_merger_w) { + GGML_ASSERT(hparams.spatial_merge_size > 0); + + cur = ggml_mul(ctx0, ggml_rms_norm(ctx0, cur, eps), model.mm_input_norm_w); + + // reshape image tokens to 2D grid + cur = ggml_reshape_3d(ctx0, cur, n_embd, n_patches_x, n_patches_y); + cur = ggml_permute(ctx0, cur, 2, 0, 1, 3); // [x, y, n_embd] + cur = ggml_cont(ctx0, cur); + + // torch.nn.functional.unfold is just an im2col under the hood + // we just need a dummy kernel to make it work + ggml_tensor * kernel = ggml_view_3d(ctx0, cur, n_merge, n_merge, cur->ne[2], 0, 0, 0); + cur = ggml_im2col(ctx0, kernel, cur, n_merge, n_merge, 0, 0, 1, 1, true, inp->type); + + // project to n_embd + cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], cur->ne[1] * cur->ne[2]); + cur = ggml_mul_mat(ctx0, model.mm_patch_merger_w, cur); } - } - // llava projector - if (ctx->has_llava_projector) { - embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]); - - struct ggml_tensor * patches = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_patches); - ggml_set_name(patches, "patches"); - ggml_set_input(patches); - - // shape [1, 576, 1024] - // ne is whcn, ne = [1024, 576, 1, 1] - embeddings = ggml_get_rows(ctx0, embeddings, patches); - - // print_tensor_info(embeddings, "embeddings"); - - // llava projector - if (ctx->proj_type == PROJECTOR_TYPE_MLP) { - embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); - embeddings = ggml_add(ctx0, embeddings, model.mm_0_b); - - embeddings = ggml_gelu(ctx0, embeddings); - if (model.mm_2_w) { - embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings); - embeddings = ggml_add(ctx0, embeddings, model.mm_2_b); - } - } - else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) { - embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); - embeddings = ggml_add(ctx0, embeddings, model.mm_0_b); - // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false); - // First LayerNorm - embeddings = ggml_norm(ctx0, embeddings, eps); - embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_1_w), - model.mm_1_b); - - // GELU activation - embeddings = ggml_gelu(ctx0, embeddings); - - // Second linear layer - embeddings = ggml_mul_mat(ctx0, model.mm_3_w, embeddings); - embeddings = ggml_add(ctx0, embeddings, model.mm_3_b); - - // Second LayerNorm - embeddings = ggml_norm(ctx0, embeddings, eps); - embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_4_w), - model.mm_4_b); - } - else if (ctx->proj_type == PROJECTOR_TYPE_LDP) { - // MobileVLM projector - int n_patch = 24; - struct ggml_tensor * mlp_1 = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w, embeddings); - mlp_1 = ggml_add(ctx0, mlp_1, model.mm_model_mlp_1_b); - mlp_1 = ggml_gelu(ctx0, mlp_1); - struct ggml_tensor * mlp_3 = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, mlp_1); - mlp_3 = ggml_add(ctx0, mlp_3, model.mm_model_mlp_3_b); - // mlp_3 shape = [1, 576, 2048], ne = [2048, 576, 1, 1] - - // block 1 - struct ggml_tensor * block_1 = nullptr; - { - // transpose from [1, 576, 2048] --> [1, 2048, 576] --> [1, 2048, 24, 24] - mlp_3 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_3, 1, 0, 2, 3)); - mlp_3 = ggml_reshape_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]); - // stride = 1, padding = 1, bias is nullptr - block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1); - - // layer norm - // // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1] - block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3)); - // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1] - block_1 = ggml_norm(ctx0, block_1, eps); - block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_0_1_w), model.mm_model_block_1_block_0_1_b); - block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3)); - - // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1] - // hardswish - struct ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1); - - block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0); - // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1] - // pointwise conv - block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]); - block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc1_w, block_1); - block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc1_b); - block_1 = ggml_relu(ctx0, block_1); - block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc2_w, block_1); - block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc2_b); - block_1 = ggml_hardsigmoid(ctx0, block_1); - // block_1_hw shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1], block_1 shape = [1, 2048], ne = [2048, 1, 1, 1] - block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]); - block_1 = ggml_mul(ctx0, block_1_hw, block_1); - - int w = block_1->ne[0], h = block_1->ne[1]; - block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]); - block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3)); - - // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1] - block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_2_0_w, block_1); - block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]); - - // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1] - block_1 = ggml_norm(ctx0, block_1, eps); - block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_2_1_w), model.mm_model_block_1_block_2_1_b); - block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3)); - // block1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1] - // residual - block_1 = ggml_add(ctx0, mlp_3, block_1); - } - - // block_2 - { - // stride = 2 - block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_2_block_0_0_w, block_1, 2, 2, 1, 1, 1, 1); - - // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1] - // layer norm - block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3)); - // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1] - block_1 = ggml_norm(ctx0, block_1, eps); - block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_0_1_w), model.mm_model_block_2_block_0_1_b); - block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3)); - // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1] - // hardswish - struct ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1); - - // not sure the parameters is right for globalAvgPooling - block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0); - // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1] - // pointwise conv - block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]); - block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc1_w, block_1); - block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc1_b); - block_1 = ggml_relu(ctx0, block_1); - block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc2_w, block_1); - block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc2_b); - block_1 = ggml_hardsigmoid(ctx0, block_1); - - // block_1_hw shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1], block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1] - block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]); - block_1 = ggml_mul(ctx0, block_1_hw, block_1); - - int w = block_1->ne[0], h = block_1->ne[1]; - block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]); - block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3)); - // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1] - block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_2_0_w, block_1); - block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]); - - - // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1] - block_1 = ggml_norm(ctx0, block_1, eps); - block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_2_1_w), model.mm_model_block_2_block_2_1_b); - block_1 = ggml_reshape_3d(ctx0, block_1, block_1->ne[0], block_1->ne[1] * block_1->ne[2], block_1->ne[3]); - // block_1 shape = [1, 144, 2048], ne = [2048, 144, 1] - } - embeddings = block_1; - } - else if (ctx->proj_type == PROJECTOR_TYPE_LDPV2) + // LlavaMultiModalProjector (always using GELU activation) { - int n_patch = 24; - struct ggml_tensor * mlp_0 = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings); - mlp_0 = ggml_add(ctx0, mlp_0, model.mm_model_mlp_0_b); - mlp_0 = ggml_gelu(ctx0, mlp_0); - struct ggml_tensor * mlp_2 = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, mlp_0); - mlp_2 = ggml_add(ctx0, mlp_2, model.mm_model_mlp_2_b); - // mlp_2 ne = [2048, 576, 1, 1] - // // AVG Pool Layer 2*2, strides = 2 - mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 0, 2, 3)); - // mlp_2 ne = [576, 2048, 1, 1] - mlp_2 = ggml_reshape_4d(ctx0, mlp_2, n_patch, n_patch, mlp_2->ne[1], mlp_2->ne[2]); - // mlp_2 ne [24, 24, 2048, 1] - mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0); - // weight ne = [3, 3, 2048, 1] - struct ggml_tensor * peg_0 = ggml_conv_2d_dw(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 1); - peg_0 = ggml_cont(ctx0, ggml_permute(ctx0, peg_0, 1, 2, 0, 3)); - peg_0 = ggml_add(ctx0, peg_0, model.mm_model_peg_0_b); - mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 2, 0, 3)); - peg_0 = ggml_add(ctx0, peg_0, mlp_2); - peg_0 = ggml_reshape_3d(ctx0, peg_0, peg_0->ne[0], peg_0->ne[1] * peg_0->ne[2], peg_0->ne[3]); - embeddings = peg_0; + cur = ggml_mul_mat(ctx0, model.mm_1_w, cur); + if (model.mm_1_b) { + cur = ggml_add(ctx0, cur, model.mm_1_b); + } + + cur = ggml_gelu(ctx0, cur); + cur = ggml_mul_mat(ctx0, model.mm_2_w, cur); + if (model.mm_2_b) { + cur = ggml_add(ctx0, cur, model.mm_2_b); + } } - else { - GGML_ABORT("fatal error"); + + // arrangement of the [IMG_BREAK] token + { + // not efficient, but works + // the trick is to view the embeddings as a 3D tensor with shape [n_embd, n_patches_per_row, n_rows] + // and then concatenate the [IMG_BREAK] token to the end of each row, aka n_patches_per_row dimension + // after the concatenation, we have a tensor with shape [n_embd, n_patches_per_row + 1, n_rows] + + const int p_y = n_merge > 0 ? n_patches_y / n_merge : n_patches_y; + const int p_x = n_merge > 0 ? n_patches_x / n_merge : n_patches_x; + const int p_total = p_x * p_y; + const int n_embd_text = cur->ne[0]; + const int n_tokens_output = p_total + p_y - 1; // one [IMG_BREAK] per row, except the last row + + ggml_tensor * tmp = ggml_reshape_3d(ctx0, cur, n_embd_text, p_x, p_y); + ggml_tensor * tok = ggml_new_tensor_3d(ctx0, tmp->type, n_embd_text, 1, p_y); + tok = ggml_scale(ctx0, tok, 0.0); // clear the tensor + tok = ggml_add(ctx0, tok, model.token_embd_img_break); + tmp = ggml_concat(ctx0, tmp, tok, 1); + cur = ggml_view_2d(ctx0, tmp, + n_embd_text, n_tokens_output, + ggml_row_size(tmp->type, n_embd_text), 0); } + + // build the graph + ggml_build_forward_expand(gf, cur); + + return gf; } - // minicpmv projector - else if (ctx->has_minicpmv_projector) - { - if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) { - struct ggml_tensor * q = model.mm_model_query; - { // layernorm - q = ggml_norm(ctx0, q, eps); - q = ggml_add(ctx0, ggml_mul(ctx0, q, model.mm_model_ln_q_w), model.mm_model_ln_q_b); - } - struct ggml_tensor * v = ggml_mul_mat(ctx0, model.mm_model_kv_proj, embeddings); - { // layernorm - v = ggml_norm(ctx0, v, eps); - v = ggml_add(ctx0, ggml_mul(ctx0, v, model.mm_model_ln_kv_w), model.mm_model_ln_kv_b); - } - struct ggml_tensor * k; - { // position - // q = ggml_add(ctx0, q, model.mm_model_pos_embed); - k = ggml_add(ctx0, v, pos_embed); - } - { // attention - int hidden_size = 4096; - const int d_head = 128; - int n_head = hidden_size/d_head; - int num_query = 96; - if (ctx->minicpmv_version == 2) { - hidden_size = 4096; - n_head = hidden_size/d_head; - num_query = 96; - } - else if (ctx->minicpmv_version == 3) { - hidden_size = 3584; - n_head = hidden_size/d_head; - num_query = 64; - } - else if (ctx->minicpmv_version == 4) { - hidden_size = 3584; - n_head = hidden_size/d_head; - num_query = 64; - } + // Qwen2VL and Qwen2.5VL use M-RoPE + ggml_cgraph * build_qwen2vl() { + GGML_ASSERT(model.patch_bias == nullptr); + GGML_ASSERT(model.class_embedding == nullptr); - struct ggml_tensor * Q = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q), model.mm_model_attn_q_b); - struct ggml_tensor * K = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_k_w, k), model.mm_model_attn_k_b); - struct ggml_tensor * V = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_v_w, v), model.mm_model_attn_v_b); - // permute - Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_query, batch_size); - Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3)); - Q = ggml_reshape_3d(ctx0, Q, d_head, num_query, n_head * batch_size); - K = ggml_reshape_4d(ctx0, K, d_head, n_head, num_positions, batch_size); - K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3)); - K = ggml_reshape_3d(ctx0, K, d_head, num_positions, n_head * batch_size); - V = ggml_reshape_4d(ctx0, V, d_head, n_head, num_positions, batch_size); - V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3)); - V = ggml_reshape_3d(ctx0, V, num_positions, d_head, n_head * batch_size); - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - KQ = ggml_soft_max_ext(ctx0, KQ, nullptr, 1.0f / sqrtf((float)d_head), 0.0f); - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ); - KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_query, n_head, batch_size); - KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - KQV = ggml_cont_3d(ctx0, KQV, hidden_size, num_query, batch_size); + const int batch_size = 1; + const bool use_window_attn = hparams.n_wa_pattern > 0; + const int n_wa_pattern = hparams.n_wa_pattern; + const int n_pos = n_patches; + const int num_position_ids = n_pos * 4; // m-rope requires 4 dim per position - embeddings = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_o_w, KQV), model.mm_model_attn_o_b); - } - { // layernorm - embeddings = ggml_norm(ctx0, embeddings, eps); - embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_post_w), model.mm_model_ln_post_b); - } - embeddings = ggml_mul_mat(ctx0, model.mm_model_proj, embeddings); + norm_type norm_t = ctx->proj_type == PROJECTOR_TYPE_QWEN25VL + ? NORM_TYPE_RMS // qwen 2.5 vl + : NORM_TYPE_NORMAL; // qwen 2 vl + + int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4}; + + ggml_tensor * inp_raw = build_inp_raw(); + ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1); + + GGML_ASSERT(img.nx % (patch_size * 2) == 0); + GGML_ASSERT(img.ny % (patch_size * 2) == 0); + + // second conv dimension + { + auto inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1); + inp = ggml_add(ctx0, inp, inp_1); + + inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 2, 0, 3)); // [w, h, c, b] -> [c, w, h, b] + inp = ggml_reshape_4d( + ctx0, inp, + n_embd * 2, n_patches_x / 2, n_patches_y, batch_size); + inp = ggml_reshape_4d( + ctx0, inp, + n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2)); + inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 0, 2, 1, 3)); + inp = ggml_reshape_3d( + ctx0, inp, + n_embd, n_patches_x * n_patches_y, batch_size); } - else { - GGML_ASSERT(false); + + ggml_tensor * inpL = inp; + ggml_tensor * window_mask = nullptr; + ggml_tensor * window_idx = nullptr; + ggml_tensor * inv_window_idx = nullptr; + + ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_position_ids); + ggml_set_name(positions, "positions"); + ggml_set_input(positions); + + // pre-layernorm + if (model.pre_ln_w) { + inpL = build_norm(inpL, model.pre_ln_w, model.pre_ln_b, norm_t, eps, -1); } - } - // glm projector - else if (ctx->has_glm_projector) { - if (ctx->proj_type == PROJECTOR_TYPE_GLM_EDGE) { - size_t gridsz = (size_t)sqrt(embeddings->ne[1]); - embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings,1,0,2,3)); - embeddings = ggml_reshape_3d(ctx0, embeddings, gridsz, gridsz, embeddings->ne[1]); - embeddings = ggml_conv_2d(ctx0, model.mm_model_adapter_conv_w, embeddings, 2, 2, 0, 0, 1, 1); - embeddings = ggml_reshape_3d(ctx0, embeddings,embeddings->ne[0]*embeddings->ne[1] , embeddings->ne[2], batch_size); - embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings, 1, 0, 2, 3)); - embeddings = ggml_add(ctx0, embeddings, model.mm_model_adapter_conv_b); - //GLU + + if (use_window_attn) { + // handle window attention inputs + inv_window_idx = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos / 4); + ggml_set_name(inv_window_idx, "inv_window_idx"); + ggml_set_input(inv_window_idx); + // mask for window attention + window_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_pos, n_pos); + ggml_set_name(window_mask, "window_mask"); + ggml_set_input(window_mask); + + // inpL shape: [n_embd, n_patches_x * n_patches_y, batch_size] + GGML_ASSERT(batch_size == 1); + inpL = ggml_reshape_2d(ctx0, inpL, n_embd * 4, n_patches_x * n_patches_y * batch_size / 4); + inpL = ggml_get_rows(ctx0, inpL, inv_window_idx); + inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_patches_x * n_patches_y, batch_size); + } + + // loop over layers + for (int il = 0; il < n_layer; il++) { + auto & layer = model.layers[il]; + const bool full_attn = use_window_attn ? (il + 1) % n_wa_pattern == 0 : true; + + ggml_tensor * cur = inpL; // inpL = residual, cur = hidden_states + + // layernorm1 + cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, norm_t, eps, il); + cb(cur, "ln1", il); + + // self-attention { - embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings); - embeddings = ggml_norm(ctx0, embeddings, eps); - embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_q_w), model.mm_model_ln_q_b); - embeddings = ggml_gelu_inplace(ctx0, embeddings); - struct ggml_tensor * x = embeddings; - embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, embeddings); - x = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w,x); - embeddings = ggml_silu_inplace(ctx0, embeddings); - embeddings = ggml_mul(ctx0, embeddings,x); - embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, embeddings); + ggml_tensor * Qcur = ggml_add(ctx0, + ggml_mul_mat(ctx0, layer.q_w, cur), layer.q_b); + ggml_tensor * Kcur = ggml_add(ctx0, + ggml_mul_mat(ctx0, layer.k_w, cur), layer.k_b); + ggml_tensor * Vcur = ggml_add(ctx0, + ggml_mul_mat(ctx0, layer.v_w, cur), layer.v_b); + + Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_patches); + Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_patches); + Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_patches); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + // apply M-RoPE + Qcur = ggml_rope_multi( + ctx0, Qcur, positions, nullptr, + d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1); + Kcur = ggml_rope_multi( + ctx0, Kcur, positions, nullptr, + d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1); + + cb(Qcur, "Qcur_rope", il); + cb(Kcur, "Kcur_rope", il); + + ggml_tensor * attn_mask = full_attn ? nullptr : window_mask; + + cur = build_attn(layer.o_w, layer.o_b, + Qcur, Kcur, Vcur, attn_mask, kq_scale, il); + cb(cur, "attn_out", il); } - } else { - GGML_ABORT("fatal error"); + + // re-add the layer input, e.g., residual + cur = ggml_add(ctx0, cur, inpL); + + inpL = cur; // inpL = residual, cur = hidden_states + + cb(cur, "ffn_inp", il); + + // layernorm2 + cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, norm_t, eps, il); + cb(cur, "ffn_inp_normed", il); + + // ffn + cur = build_ffn(cur, + layer.ff_up_w, layer.ff_up_b, + layer.ff_gate_w, layer.ff_gate_b, + layer.ff_down_w, layer.ff_down_b, + hparams.ffn_op, il); + + cb(cur, "ffn_out", il); + + // residual 2 + cur = ggml_add(ctx0, inpL, cur); + cb(cur, "layer_out", il); + + inpL = cur; } - } - else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) { - embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size * 4, num_positions / 4, batch_size); + + // post-layernorm + if (model.post_ln_w) { + inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, norm_t, eps, n_layer); + } + + // multimodal projection + ggml_tensor * embeddings = inpL; + embeddings = ggml_reshape_3d(ctx0, embeddings, n_embd * 4, n_pos / 4, batch_size); embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); embeddings = ggml_add(ctx0, embeddings, model.mm_0_b); @@ -1329,30 +761,971 @@ static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_im // Second linear layer embeddings = ggml_mul_mat(ctx0, model.mm_1_w, embeddings); embeddings = ggml_add(ctx0, embeddings, model.mm_1_b); + + if (use_window_attn) { + window_idx = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos / 4); + ggml_set_name(window_idx, "window_idx"); + ggml_set_input(window_idx); + + // embeddings shape: [n_embd, n_patches_x * n_patches_y, batch_size] + GGML_ASSERT(batch_size == 1); + embeddings = ggml_reshape_2d(ctx0, embeddings, hparams.projection_dim, n_patches_x * n_patches_y / 4); + embeddings = ggml_get_rows(ctx0, embeddings, window_idx); + embeddings = ggml_reshape_3d(ctx0, embeddings, hparams.projection_dim, n_patches_x * n_patches_y / 4, batch_size); + } + + // build the graph + ggml_build_forward_expand(gf, embeddings); + + return gf; } - // build the graph - ggml_build_forward_expand(gf, embeddings); + ggml_cgraph * build_minicpmv() { + const int batch_size = 1; - return gf; -} + GGML_ASSERT(model.class_embedding == nullptr); + const int n_pos = n_patches; + + // position embeddings for the projector (not for ViT) + int n_output_dim = clip_n_mmproj_embd(ctx); + ggml_tensor * pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_output_dim, n_pos, batch_size); + ggml_set_name(pos_embed, "pos_embed"); + ggml_set_input(pos_embed); + + // for selecting learned pos embd, used by ViT + struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos); + ggml_set_name(positions, "positions"); + ggml_set_input(positions); + + ggml_tensor * learned_pos_embd = ggml_get_rows(ctx0, model.position_embeddings, positions); + + ggml_tensor * inp = build_inp(); + ggml_tensor * embeddings = build_vit( + inp, n_patches, + NORM_TYPE_NORMAL, + hparams.ffn_op, + learned_pos_embd, + nullptr); + + // resampler projector (it is just another transformer) + + ggml_tensor * q = model.mm_model_query; + ggml_tensor * v = ggml_mul_mat(ctx0, model.mm_model_kv_proj, embeddings); + + // norm + q = build_norm(q, model.mm_model_ln_q_w, model.mm_model_ln_q_b, NORM_TYPE_NORMAL, eps, -1); + v = build_norm(v, model.mm_model_ln_kv_w, model.mm_model_ln_kv_b, NORM_TYPE_NORMAL, eps, -1); + + // k = v + pos_embed + ggml_tensor * k = ggml_add(ctx0, v, pos_embed); + + // attention + { + int n_embd = clip_n_mmproj_embd(ctx); + const int d_head = 128; + int n_head = n_embd/d_head; + int num_query = 96; + if (ctx->minicpmv_version == 2) { + num_query = 96; + } else if (ctx->minicpmv_version == 3) { + num_query = 64; + } else if (ctx->minicpmv_version == 4) { + num_query = 64; + } + + ggml_tensor * Q = ggml_add(ctx0, + ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q), + model.mm_model_attn_q_b); + ggml_tensor * K = ggml_add(ctx0, + ggml_mul_mat(ctx0, model.mm_model_attn_k_w, k), + model.mm_model_attn_k_b); + ggml_tensor * V = ggml_add(ctx0, + ggml_mul_mat(ctx0, model.mm_model_attn_v_w, v), + model.mm_model_attn_v_b); + + Q = ggml_reshape_3d(ctx0, Q, d_head, n_head, num_query); + K = ggml_reshape_3d(ctx0, K, d_head, n_head, n_pos); + V = ggml_reshape_3d(ctx0, V, d_head, n_head, n_pos); + + cb(Q, "resampler_Q", -1); + cb(K, "resampler_K", -1); + cb(V, "resampler_V", -1); + + embeddings = build_attn( + model.mm_model_attn_o_w, + model.mm_model_attn_o_b, + Q, K, V, nullptr, kq_scale, -1); + cb(embeddings, "resampler_attn_out", -1); + } + // layernorm + embeddings = build_norm(embeddings, model.mm_model_ln_post_w, model.mm_model_ln_post_b, NORM_TYPE_NORMAL, eps, -1); + + // projection + embeddings = ggml_mul_mat(ctx0, model.mm_model_proj, embeddings); + + // build the graph + ggml_build_forward_expand(gf, embeddings); + + return gf; + } + + ggml_cgraph * build_internvl() { + GGML_ASSERT(model.class_embedding != nullptr); + GGML_ASSERT(model.position_embeddings != nullptr); + + const int n_pos = n_patches + 1; + ggml_tensor * inp = build_inp(); + + // add CLS token + inp = ggml_concat(ctx0, inp, model.class_embedding, 1); + + // The larger models use a different ViT, which uses RMS norm instead of layer norm + // ref: https://github.com/ggml-org/llama.cpp/pull/13443#issuecomment-2869786188 + norm_type norm_t = (hparams.n_embd == 3200 && hparams.n_layer == 45) + ? NORM_TYPE_RMS // 6B ViT (Used by InternVL 2.5/3 - 26B, 38B, 78B) + : NORM_TYPE_NORMAL; // 300M ViT (Used by all smaller InternVL models) + + ggml_tensor * cur = build_vit( + inp, n_pos, + norm_t, + hparams.ffn_op, + model.position_embeddings, + nullptr); + + // remove CLS token + cur = ggml_view_2d(ctx0, cur, + n_embd, n_patches, + ggml_row_size(cur->type, n_embd), 0); + + // pixel shuffle + { + const int scale_factor = model.hparams.proj_scale_factor; + const int bsz = 1; // batch size, always 1 for now since we don't support batching + const int height = n_patches_y; + const int width = n_patches_x; + GGML_ASSERT(scale_factor > 0); + cur = ggml_reshape_4d(ctx0, cur, n_embd * scale_factor, height / scale_factor, width, bsz); + cur = ggml_permute(ctx0, cur, 0, 2, 1, 3); + cur = ggml_reshape_4d(ctx0, ggml_cont(ctx0, cur), + n_embd * scale_factor * scale_factor, + height / scale_factor, + width / scale_factor, + bsz); + cur = ggml_permute(ctx0, cur, 0, 2, 1, 3); + // flatten to 2D + cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, cur), + n_embd * scale_factor * scale_factor, + cur->ne[1] * cur->ne[2]); + } + + // projector (always using GELU activation) + { + // projector LayerNorm uses pytorch's default eps = 1e-5 + // ref: https://huggingface.co/OpenGVLab/InternVL3-8B-Instruct/blob/a34d3e4e129a5856abfd6aa6de79776484caa14e/modeling_internvl_chat.py#L79 + cur = build_norm(cur, model.mm_0_w, model.mm_0_b, NORM_TYPE_NORMAL, 1e-5, -1); + cur = ggml_mul_mat(ctx0, model.mm_1_w, cur); + cur = ggml_add(ctx0, cur, model.mm_1_b); + cur = ggml_gelu(ctx0, cur); + cur = ggml_mul_mat(ctx0, model.mm_3_w, cur); + cur = ggml_add(ctx0, cur, model.mm_3_b); + } + + // build the graph + ggml_build_forward_expand(gf, cur); + + return gf; + } + + // this graph is used by llava, granite and glm + // due to having embedding_stack (used by granite), we cannot reuse build_vit + ggml_cgraph * build_llava() { + const int batch_size = 1; + const int n_pos = n_patches + (model.class_embedding ? 1 : 0); + + GGML_ASSERT(n_patches_x == n_patches_y && "only square images supported"); + + // Calculate the deepest feature layer based on hparams and projector type + int max_feature_layer = n_layer; + { + // Get the index of the second to last layer; this is the default for models that have a llava projector + int il_last = hparams.n_layer - 1; + int deepest_feature_layer = -1; + + if (ctx->proj_type == PROJECTOR_TYPE_MINICPMV || ctx->proj_type == PROJECTOR_TYPE_GLM_EDGE) { + il_last += 1; + } + + // If we set explicit vision feature layers, only go up to the deepest one + // NOTE: only used by granite-vision models for now + for (const auto & feature_layer : hparams.vision_feature_layer) { + if (feature_layer > deepest_feature_layer) { + deepest_feature_layer = feature_layer; + } + } + max_feature_layer = deepest_feature_layer < 0 ? il_last : deepest_feature_layer; + } + + ggml_tensor * inp = build_inp(); + + // concat class_embeddings and patch_embeddings + if (model.class_embedding) { + inp = ggml_concat(ctx0, inp, model.class_embedding, 1); + } + + ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos); + ggml_set_name(positions, "positions"); + ggml_set_input(positions); + + inp = ggml_add(ctx0, inp, ggml_get_rows(ctx0, model.position_embeddings, positions)); + + ggml_tensor * inpL = inp; + + // pre-layernorm + if (model.pre_ln_w) { + inpL = build_norm(inpL, model.pre_ln_w, model.pre_ln_b, NORM_TYPE_NORMAL, eps, -1); + cb(inpL, "pre_ln", -1); + } + + std::vector embedding_stack; + const auto & vision_feature_layer = hparams.vision_feature_layer; + + // loop over layers + for (int il = 0; il < max_feature_layer; il++) { + auto & layer = model.layers[il]; + ggml_tensor * cur = inpL; // inpL = residual, cur = hidden_states + + // If this is an embedding feature layer, save the output. + // NOTE: 0 index here refers to the input to the encoder. + if (vision_feature_layer.find(il) != vision_feature_layer.end()) { + embedding_stack.push_back(cur); + } + + // layernorm1 + cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, NORM_TYPE_NORMAL, eps, il); + cb(cur, "layer_inp_normed", il); + + // self-attention + { + ggml_tensor * Qcur = ggml_mul_mat(ctx0, layer.q_w, cur); + if (layer.q_b) { + Qcur = ggml_add(ctx0, Qcur, layer.q_b); + } + + ggml_tensor * Kcur = ggml_mul_mat(ctx0, layer.k_w, cur); + if (layer.k_b) { + Kcur = ggml_add(ctx0, Kcur, layer.k_b); + } + + ggml_tensor * Vcur = ggml_mul_mat(ctx0, layer.v_w, cur); + if (layer.v_b) { + Vcur = ggml_add(ctx0, Vcur, layer.v_b); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_pos); + Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_pos); + Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_pos); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(layer.o_w, layer.o_b, + Qcur, Kcur, Vcur, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + } + + // re-add the layer input, e.g., residual + cur = ggml_add(ctx0, cur, inpL); + + inpL = cur; // inpL = residual, cur = hidden_states + + cb(cur, "ffn_inp", il); + + // layernorm2 + cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, NORM_TYPE_NORMAL, eps, il); + cb(cur, "ffn_inp_normed", il); + + // ffn + cur = build_ffn(cur, + layer.ff_up_w, layer.ff_up_b, + layer.ff_gate_w, layer.ff_gate_b, + layer.ff_down_w, layer.ff_down_b, + hparams.ffn_op, il); + + cb(cur, "ffn_out", il); + + // residual 2 + cur = ggml_add(ctx0, inpL, cur); + cb(cur, "layer_out", il); + + inpL = cur; + } + + // post-layernorm + if (model.post_ln_w) { + inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, NORM_TYPE_NORMAL, eps, -1); + } + + ggml_tensor * embeddings = inpL; + + // process vision feature layers (used by granite) + { + // final layer is a vision feature layer + if (vision_feature_layer.find(max_feature_layer) != vision_feature_layer.end()) { + embedding_stack.push_back(inpL); + } + + // If feature layers are explicitly set, stack them (if we have multiple) + if (!embedding_stack.empty()) { + embeddings = embedding_stack[0]; + for (size_t i = 1; i < embedding_stack.size(); i++) { + embeddings = ggml_concat(ctx0, embeddings, embedding_stack[i], 0); + } + } + } + + // llava projector (also used by granite) + if (ctx->has_llava_projector) { + embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]); + + ggml_tensor * patches = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches); + ggml_set_name(patches, "patches"); + ggml_set_input(patches); + + // shape [1, 576, 1024] + // ne is whcn, ne = [1024, 576, 1, 1] + embeddings = ggml_get_rows(ctx0, embeddings, patches); + + // print_tensor_info(embeddings, "embeddings"); + + // llava projector + if (ctx->proj_type == PROJECTOR_TYPE_MLP) { + embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); + embeddings = ggml_add(ctx0, embeddings, model.mm_0_b); + + embeddings = ggml_gelu(ctx0, embeddings); + if (model.mm_2_w) { + embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings); + embeddings = ggml_add(ctx0, embeddings, model.mm_2_b); + } + } + else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) { + embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); + embeddings = ggml_add(ctx0, embeddings, model.mm_0_b); + // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false); + // First LayerNorm + embeddings = ggml_norm(ctx0, embeddings, eps); + embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_1_w), + model.mm_1_b); + + // GELU activation + embeddings = ggml_gelu(ctx0, embeddings); + + // Second linear layer + embeddings = ggml_mul_mat(ctx0, model.mm_3_w, embeddings); + embeddings = ggml_add(ctx0, embeddings, model.mm_3_b); + + // Second LayerNorm + embeddings = ggml_norm(ctx0, embeddings, eps); + embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_4_w), + model.mm_4_b); + } + else if (ctx->proj_type == PROJECTOR_TYPE_LDP) { + // MobileVLM projector + int n_patch = 24; + ggml_tensor * mlp_1 = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w, embeddings); + mlp_1 = ggml_add(ctx0, mlp_1, model.mm_model_mlp_1_b); + mlp_1 = ggml_gelu(ctx0, mlp_1); + ggml_tensor * mlp_3 = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, mlp_1); + mlp_3 = ggml_add(ctx0, mlp_3, model.mm_model_mlp_3_b); + // mlp_3 shape = [1, 576, 2048], ne = [2048, 576, 1, 1] + + // block 1 + ggml_tensor * block_1 = nullptr; + { + // transpose from [1, 576, 2048] --> [1, 2048, 576] --> [1, 2048, 24, 24] + mlp_3 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_3, 1, 0, 2, 3)); + mlp_3 = ggml_reshape_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]); + // stride = 1, padding = 1, bias is nullptr + block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1); + + // layer norm + // // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1] + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3)); + // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1] + block_1 = ggml_norm(ctx0, block_1, eps); + block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_0_1_w), model.mm_model_block_1_block_0_1_b); + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3)); + + // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1] + // hardswish + ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1); + + block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0); + // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1] + // pointwise conv + block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]); + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc1_w, block_1); + block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc1_b); + block_1 = ggml_relu(ctx0, block_1); + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc2_w, block_1); + block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc2_b); + block_1 = ggml_hardsigmoid(ctx0, block_1); + // block_1_hw shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1], block_1 shape = [1, 2048], ne = [2048, 1, 1, 1] + block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]); + block_1 = ggml_mul(ctx0, block_1_hw, block_1); + + int w = block_1->ne[0], h = block_1->ne[1]; + block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]); + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3)); + + // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1] + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_2_0_w, block_1); + block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]); + + // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1] + block_1 = ggml_norm(ctx0, block_1, eps); + block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_2_1_w), model.mm_model_block_1_block_2_1_b); + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3)); + // block1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1] + // residual + block_1 = ggml_add(ctx0, mlp_3, block_1); + } + + // block_2 + { + // stride = 2 + block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_2_block_0_0_w, block_1, 2, 2, 1, 1, 1, 1); + + // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1] + // layer norm + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3)); + // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1] + block_1 = ggml_norm(ctx0, block_1, eps); + block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_0_1_w), model.mm_model_block_2_block_0_1_b); + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3)); + // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1] + // hardswish + ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1); + + // not sure the parameters is right for globalAvgPooling + block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0); + // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1] + // pointwise conv + block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]); + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc1_w, block_1); + block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc1_b); + block_1 = ggml_relu(ctx0, block_1); + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc2_w, block_1); + block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc2_b); + block_1 = ggml_hardsigmoid(ctx0, block_1); + + // block_1_hw shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1], block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1] + block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]); + block_1 = ggml_mul(ctx0, block_1_hw, block_1); + + int w = block_1->ne[0], h = block_1->ne[1]; + block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]); + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3)); + // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1] + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_2_0_w, block_1); + block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]); + + + // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1] + block_1 = ggml_norm(ctx0, block_1, eps); + block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_2_1_w), model.mm_model_block_2_block_2_1_b); + block_1 = ggml_reshape_3d(ctx0, block_1, block_1->ne[0], block_1->ne[1] * block_1->ne[2], block_1->ne[3]); + // block_1 shape = [1, 144, 2048], ne = [2048, 144, 1] + } + embeddings = block_1; + } + else if (ctx->proj_type == PROJECTOR_TYPE_LDPV2) + { + int n_patch = 24; + ggml_tensor * mlp_0 = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings); + mlp_0 = ggml_add(ctx0, mlp_0, model.mm_model_mlp_0_b); + mlp_0 = ggml_gelu(ctx0, mlp_0); + ggml_tensor * mlp_2 = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, mlp_0); + mlp_2 = ggml_add(ctx0, mlp_2, model.mm_model_mlp_2_b); + // mlp_2 ne = [2048, 576, 1, 1] + // // AVG Pool Layer 2*2, strides = 2 + mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 0, 2, 3)); + // mlp_2 ne = [576, 2048, 1, 1] + mlp_2 = ggml_reshape_4d(ctx0, mlp_2, n_patch, n_patch, mlp_2->ne[1], mlp_2->ne[2]); + // mlp_2 ne [24, 24, 2048, 1] + mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0); + // weight ne = [3, 3, 2048, 1] + ggml_tensor * peg_0 = ggml_conv_2d_dw(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 1); + peg_0 = ggml_cont(ctx0, ggml_permute(ctx0, peg_0, 1, 2, 0, 3)); + peg_0 = ggml_add(ctx0, peg_0, model.mm_model_peg_0_b); + mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 2, 0, 3)); + peg_0 = ggml_add(ctx0, peg_0, mlp_2); + peg_0 = ggml_reshape_3d(ctx0, peg_0, peg_0->ne[0], peg_0->ne[1] * peg_0->ne[2], peg_0->ne[3]); + embeddings = peg_0; + } + else { + GGML_ABORT("fatal error"); + } + } + + // glm projector + else if (ctx->proj_type == PROJECTOR_TYPE_GLM_EDGE) { + size_t gridsz = (size_t)sqrt(embeddings->ne[1]); + embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings,1,0,2,3)); + embeddings = ggml_reshape_3d(ctx0, embeddings, gridsz, gridsz, embeddings->ne[1]); + embeddings = ggml_conv_2d(ctx0, model.mm_model_adapter_conv_w, embeddings, 2, 2, 0, 0, 1, 1); + embeddings = ggml_reshape_3d(ctx0, embeddings,embeddings->ne[0]*embeddings->ne[1] , embeddings->ne[2], batch_size); + embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings, 1, 0, 2, 3)); + embeddings = ggml_add(ctx0, embeddings, model.mm_model_adapter_conv_b); + // GLU + { + embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings); + embeddings = ggml_norm(ctx0, embeddings, eps); + embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_q_w), model.mm_model_ln_q_b); + embeddings = ggml_gelu_inplace(ctx0, embeddings); + ggml_tensor * x = embeddings; + embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, embeddings); + x = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w,x); + embeddings = ggml_silu_inplace(ctx0, embeddings); + embeddings = ggml_mul(ctx0, embeddings,x); + embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, embeddings); + } + // arrangement of BOI/EOI token embeddings + // note: these embeddings are not present in text model, hence we cannot process them as text tokens + // see: https://huggingface.co/THUDM/glm-edge-v-2b/blob/main/siglip.py#L53 + { + embeddings = ggml_concat(ctx0, model.mm_glm_tok_boi, embeddings, 1); // BOI + embeddings = ggml_concat(ctx0, embeddings, model.mm_glm_tok_eoi, 1); // EOI + } + } + + else { + GGML_ABORT("llava: unknown projector type"); + } + + // build the graph + ggml_build_forward_expand(gf, embeddings); + + return gf; + } + +private: + // + // utility functions + // + + void cb(ggml_tensor * cur, const char * name, int il) const { + // TODO: implement this + GGML_UNUSED(cur); + GGML_UNUSED(name); + GGML_UNUSED(il); + } + + // build vision transformer (ViT) cgraph + // this function should cover most of the models + // if your model has specific features, you should probably duplicate this function + ggml_tensor * build_vit( + ggml_tensor * inp, + int64_t n_pos, + norm_type norm_t, + ffn_op_type ffn_t, + ggml_tensor * learned_pos_embd, + std::function add_pos + ) { + if (learned_pos_embd) { + inp = ggml_add(ctx0, inp, learned_pos_embd); + cb(inp, "pos_embed", -1); + } + + ggml_tensor * inpL = inp; + + // pre-layernorm + if (model.pre_ln_w) { + inpL = build_norm(inpL, model.pre_ln_w, model.pre_ln_b, norm_t, eps, -1); + cb(inpL, "pre_ln", -1); + } + + // loop over layers + for (int il = 0; il < n_layer; il++) { + auto & layer = model.layers[il]; + ggml_tensor * cur = inpL; // inpL = residual, cur = hidden_states + + // layernorm1 + cur = build_norm(cur, layer.ln_1_w, layer.ln_1_b, norm_t, eps, il); + cb(cur, "layer_inp_normed", il); + + // self-attention + { + ggml_tensor * Qcur = ggml_mul_mat(ctx0, layer.q_w, cur); + if (layer.q_b) { + Qcur = ggml_add(ctx0, Qcur, layer.q_b); + } + + ggml_tensor * Kcur = ggml_mul_mat(ctx0, layer.k_w, cur); + if (layer.k_b) { + Kcur = ggml_add(ctx0, Kcur, layer.k_b); + } + + ggml_tensor * Vcur = ggml_mul_mat(ctx0, layer.v_w, cur); + if (layer.v_b) { + Vcur = ggml_add(ctx0, Vcur, layer.v_b); + } + + if (layer.q_norm) { + Qcur = build_norm(Qcur, layer.q_norm, NULL, norm_t, eps, il); + cb(Qcur, "Qcur_norm", il); + } + + if (layer.k_norm) { + Kcur = build_norm(Kcur, layer.k_norm, NULL, norm_t, eps, il); + cb(Kcur, "Kcur_norm", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, d_head, n_head, n_pos); + Kcur = ggml_reshape_3d(ctx0, Kcur, d_head, n_head, n_pos); + Vcur = ggml_reshape_3d(ctx0, Vcur, d_head, n_head, n_pos); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + if (add_pos) { + Qcur = add_pos(Qcur, layer); + Kcur = add_pos(Kcur, layer); + cb(Qcur, "Qcur_pos", il); + cb(Kcur, "Kcur_pos", il); + } + + cur = build_attn(layer.o_w, layer.o_b, + Qcur, Kcur, Vcur, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + } + + if (layer.ls_1_w) { + cur = ggml_mul(ctx0, cur, layer.ls_1_w); + cb(cur, "attn_out_scaled", il); + } + + // re-add the layer input, e.g., residual + cur = ggml_add(ctx0, cur, inpL); + + inpL = cur; // inpL = residual, cur = hidden_states + + cb(cur, "ffn_inp", il); + + // layernorm2 + cur = build_norm(cur, layer.ln_2_w, layer.ln_2_b, norm_t, eps, il); + cb(cur, "ffn_inp_normed", il); + + // ffn + cur = build_ffn(cur, + layer.ff_up_w, layer.ff_up_b, + layer.ff_gate_w, layer.ff_gate_b, + layer.ff_down_w, layer.ff_down_b, + ffn_t, il); + + cb(cur, "ffn_out", il); + + if (layer.ls_2_w) { + cur = ggml_mul(ctx0, cur, layer.ls_2_w); + cb(cur, "ffn_out_scaled", il); + } + + // residual 2 + cur = ggml_add(ctx0, inpL, cur); + cb(cur, "layer_out", il); + + inpL = cur; + } + + // post-layernorm + if (model.post_ln_w) { + inpL = build_norm(inpL, model.post_ln_w, model.post_ln_b, norm_t, eps, -1); + } + return inpL; + } + + // build the input after conv2d (inp_raw --> patches) + // returns tensor with shape [n_embd, n_patches] + ggml_tensor * build_inp() { + ggml_tensor * inp_raw = build_inp_raw(); + ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1); + inp = ggml_reshape_2d(ctx0, inp, n_patches, n_embd); + inp = ggml_cont(ctx0, ggml_transpose(ctx0, inp)); + if (model.patch_bias) { + inp = ggml_add(ctx0, inp, model.patch_bias); + cb(inp, "patch_bias", -1); + } + return inp; + } + + ggml_tensor * build_inp_raw() { + ggml_tensor * inp_raw = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, img.nx, img.ny, 3); + ggml_set_name(inp_raw, "inp_raw"); + ggml_set_input(inp_raw); + return inp_raw; + } + + ggml_tensor * build_norm( + ggml_tensor * cur, + ggml_tensor * mw, + ggml_tensor * mb, + norm_type type, + float norm_eps, + int il) const { + + cur = type == NORM_TYPE_RMS + ? ggml_rms_norm(ctx0, cur, norm_eps) + : ggml_norm(ctx0, cur, norm_eps); + + if (mw || mb) { + cb(cur, "norm", il); + } + + if (mw) { + cur = ggml_mul(ctx0, cur, mw); + if (mb) { + cb(cur, "norm_w", il); + } + } + + if (mb) { + cur = ggml_add(ctx0, cur, mb); + } + + return cur; + } + + ggml_tensor * build_ffn( + ggml_tensor * cur, + ggml_tensor * up, + ggml_tensor * up_b, + ggml_tensor * gate, + ggml_tensor * gate_b, + ggml_tensor * down, + ggml_tensor * down_b, + ffn_op_type type_op, + int il) const { + + ggml_tensor * tmp = up ? ggml_mul_mat(ctx0, up, cur) : cur; + cb(tmp, "ffn_up", il); + + if (up_b) { + tmp = ggml_add(ctx0, tmp, up_b); + cb(tmp, "ffn_up_b", il); + } + + if (gate) { + cur = ggml_mul_mat(ctx0, gate, cur); + cb(cur, "ffn_gate", il); + + if (gate_b) { + cur = ggml_add(ctx0, cur, gate_b); + cb(cur, "ffn_gate_b", il); + } + } else { + cur = tmp; + } + + switch (type_op) { + case FFN_SILU: + { + cur = ggml_silu(ctx0, cur); + cb(cur, "ffn_silu", il); + } break; + case FFN_GELU: + { + cur = ggml_gelu(ctx0, cur); + cb(cur, "ffn_gelu", il); + } break; + case FFN_GELU_QUICK: + { + cur = ggml_gelu_quick(ctx0, cur); + cb(cur, "ffn_relu", il); + } break; + } + + // we only support parallel ffn for now + if (gate) { + cur = ggml_mul(ctx0, cur, tmp); + cb(cur, "ffn_gate_par", il); + } + + if (down) { + cur = ggml_mul_mat(ctx0, down, cur); + } + + if (down_b) { + cb(cur, "ffn_down", il); + } + + if (down_b) { + cur = ggml_add(ctx0, cur, down_b); + } + + return cur; + } + + ggml_tensor * build_attn( + ggml_tensor * wo, + ggml_tensor * wo_b, + ggml_tensor * q_cur, + ggml_tensor * k_cur, + ggml_tensor * v_cur, + ggml_tensor * kq_mask, + 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 + ggml_build_forward_expand(gf, q_cur); + ggml_build_forward_expand(gf, k_cur); + ggml_build_forward_expand(gf, v_cur); + + ggml_tensor * q = ggml_permute(ctx0, q_cur, 0, 2, 1, 3); + //cb(q, "q", il); + + ggml_tensor * k = ggml_permute(ctx0, k_cur, 0, 2, 1, 3); + //cb(k, "k", il); + + ggml_tensor * v = ggml_permute(ctx0, v_cur, 1, 2, 0, 3); + v = ggml_cont(ctx0, v); + //cb(k, "v", il); + + ggml_tensor * cur; + + // TODO @ngxson : support flash attention + { + const auto n_tokens = q->ne[1]; + const auto n_head = q->ne[2]; + // const auto n_kv = k->ne[1]; // for flash attention + + ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + // F32 may not needed for vision encoders? + // ggml_mul_mat_set_prec(kq, GGML_PREC_F32); + + kq = ggml_soft_max_ext(ctx0, kq, kq_mask, kq_scale, 0.0f); + + ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq); + cur = ggml_permute(ctx0, kqv, 0, 2, 1, 3); + cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens); + } + + cb(cur, "kqv_out", il); + + if (wo) { + cur = ggml_mul_mat(ctx0, wo, cur); + } + + if (wo_b) { + cur = ggml_add(ctx0, cur, wo_b); + } + + return cur; + } + + // implementation of the 2D RoPE without adding a new op in ggml + // this is not efficient (use double the memory), but works on all backends + // TODO: there was a more efficient which relies on ggml_view and ggml_rope_ext_inplace, but the rope inplace does not work well with non-contiguous tensors ; we should fix that and revert back to the original implementation in https://github.com/ggml-org/llama.cpp/pull/13065 + static ggml_tensor * build_rope_2d( + ggml_context * ctx0, + ggml_tensor * cur, + ggml_tensor * pos_h, + ggml_tensor * pos_w, + const float freq_base + ) { + const int64_t n_dim = cur->ne[0]; + const int64_t n_head = cur->ne[1]; + const int64_t n_pos = cur->ne[2]; + + // for example, if we have cur tensor of shape (n_dim=8, n_head, n_pos) + // we will have a list of 4 inv_freq: 1e-0, 1e-1, 1e-2, 1e-3 + // first half of cur will use 1e-0, 1e-2 (even) + // second half of cur will use 1e-1, 1e-3 (odd) + // the trick here is to rotate just half of n_dim, so inv_freq will automatically be even + // ^ don't ask me why, it's math! -2(2i) / n_dim == -2i / (n_dim/2) + // then for the second half, we use freq_scale to shift the inv_freq + // ^ why? replace (2i) with (2i+1) in the above equation + const float freq_scale_odd = std::pow(freq_base, (float)-2/n_dim); + + // first half + ggml_tensor * first; + { + first = ggml_view_3d(ctx0, cur, + n_dim/2, n_head, n_pos, + ggml_row_size(cur->type, n_dim), + ggml_row_size(cur->type, n_dim*n_head), + 0); + first = ggml_rope_ext( + ctx0, + first, + pos_h, // positions + nullptr, // freq factors + n_dim/2, // n_dims + 0, 0, freq_base, + 1.0f, 0.0f, 1.0f, 0.0f, 0.0f + ); + } + + // second half + ggml_tensor * second; + { + second = ggml_view_3d(ctx0, cur, + n_dim/2, n_head, n_pos, + ggml_row_size(cur->type, n_dim), + ggml_row_size(cur->type, n_dim*n_head), + n_dim/2 * ggml_element_size(cur)); + second = ggml_cont(ctx0, second); // copy, because ggml_rope don't play well with non-contiguous tensors + second = ggml_rope_ext( + ctx0, + second, + pos_w, // positions + nullptr, // freq factors + n_dim/2, // n_dims + 0, 0, freq_base, + freq_scale_odd, + 0.0f, 1.0f, 0.0f, 0.0f + ); + } + + cur = ggml_concat(ctx0, first, second, 0); + return cur; + } + +}; + +static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch & imgs) { + GGML_ASSERT(imgs.entries.size() == 1 && "n_batch > 1 is not supported"); + clip_graph graph(ctx, *imgs.entries[0]); -static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch & imgs, struct clip_image_size load_image_size, bool is_inf = false) { ggml_cgraph * res; + switch (ctx->proj_type) { case PROJECTOR_TYPE_GEMMA3: case PROJECTOR_TYPE_IDEFICS3: { - res = clip_image_build_graph_siglip(ctx, imgs); + res = graph.build_siglip(); } break; case PROJECTOR_TYPE_PIXTRAL: { - res = clip_image_build_graph_pixtral(ctx, imgs); + res = graph.build_pixtral(); + } break; + case PROJECTOR_TYPE_QWEN2VL: + case PROJECTOR_TYPE_QWEN25VL: + { + res = graph.build_qwen2vl(); + } break; + case PROJECTOR_TYPE_MINICPMV: + { + res = graph.build_minicpmv(); + } break; + case PROJECTOR_TYPE_INTERNVL: + { + res = graph.build_internvl(); } break; default: { - // TODO: we should have one build_* function per model - res = clip_image_build_graph_legacy(ctx, imgs, load_image_size, is_inf); + res = graph.build_llava(); } break; } return res; @@ -1365,7 +1738,7 @@ struct clip_model_loader { clip_ctx & ctx_clip; std::string fname; - size_t model_size; // in bytes + size_t model_size = 0; // in bytes // TODO @ngxson : we should not pass clip_ctx here, it should be clip_vision_model clip_model_loader(const char * fname, clip_ctx & ctx_clip) : ctx_clip(ctx_clip), fname(fname) { @@ -1406,7 +1779,7 @@ struct clip_model_loader { const char * name = gguf_get_tensor_name(ctx_gguf.get(), i); const size_t offset = gguf_get_tensor_offset(ctx_gguf.get(), i); enum ggml_type type = gguf_get_tensor_type(ctx_gguf.get(), i); - struct ggml_tensor * cur = ggml_get_tensor(meta, name); + ggml_tensor * cur = ggml_get_tensor(meta, name); size_t tensor_size = ggml_nbytes(cur); model_size += tensor_size; LOG_DBG("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n", @@ -1417,10 +1790,11 @@ struct clip_model_loader { void load_hparams() { auto & hparams = ctx_clip.vision_model.hparams; + std::string log_ffn_op; // for logging // projector type + std::string proj_type; { - std::string proj_type; get_string(KEY_PROJ_TYPE, proj_type, false); if (!proj_type.empty()) { ctx_clip.proj_type = clip_projector_type_from_string(proj_type); @@ -1432,33 +1806,47 @@ struct clip_model_loader { // other hparams { - get_bool(KEY_HAS_TEXT_ENC, ctx_clip.has_text_encoder, false); - get_bool(KEY_HAS_VIS_ENC, ctx_clip.has_vision_encoder, false); - GGML_ASSERT(ctx_clip.has_vision_encoder); - GGML_ASSERT(!ctx_clip.has_text_encoder); + get_i32(KEY_MINICPMV_VERSION, ctx_clip.minicpmv_version, false); // legacy - // legacy keys, use KEY_PROJ_TYPE instead - get_bool(KEY_HAS_LLAVA_PROJ, ctx_clip.has_llava_projector, false); - get_bool(KEY_HAS_MINICPMV_PROJ, ctx_clip.has_minicpmv_projector, false); - get_i32(KEY_MINICPMV_VERSION, ctx_clip.minicpmv_version, false); - get_bool(KEY_HAS_GLM_PROJ, ctx_clip.has_glm_projector, false); - get_bool(KEY_HAS_QWEN2VL_MERGER, ctx_clip.has_qwen2vl_merger, false); - // !!! do NOT extend the list above, use KEY_PROJ_TYPE instead - - get_bool(KEY_USE_GELU, ctx_clip.use_gelu, false); - get_bool(KEY_USE_SILU, ctx_clip.use_silu, false); - - get_u32(string_format(KEY_N_EMBD, "vision"), hparams.hidden_size); - get_u32(string_format(KEY_N_HEAD, "vision"), hparams.n_head); - get_u32(string_format(KEY_N_FF, "vision"), hparams.n_intermediate); - get_u32(string_format(KEY_N_BLOCK, "vision"), hparams.n_layer); - get_u32(string_format(KEY_PROJ_DIM, "vision"), hparams.projection_dim); - get_f32(string_format(KEY_LAYER_NORM_EPS, "vision"), hparams.eps); - get_u32(KEY_IMAGE_SIZE, hparams.image_size); - get_u32(KEY_PATCH_SIZE, hparams.patch_size); - get_u32(KEY_IMAGE_CROP_RESOLUTION, hparams.image_crop_resolution, false); + get_u32(KEY_N_EMBD, hparams.n_embd); + get_u32(KEY_N_HEAD, hparams.n_head); + get_u32(KEY_N_FF, hparams.n_ff); + get_u32(KEY_N_BLOCK, hparams.n_layer); + get_u32(KEY_PROJ_DIM, hparams.projection_dim); + get_f32(KEY_LAYER_NORM_EPS, hparams.eps); + get_u32(KEY_IMAGE_SIZE, hparams.image_size); + get_u32(KEY_PATCH_SIZE, hparams.patch_size); + get_u32(KEY_IMAGE_CROP_RESOLUTION, hparams.image_crop_resolution, false); get_arr_int(KEY_IMAGE_GRID_PINPOINTS, hparams.image_grid_pinpoints, false); + // default warmup value + hparams.warmup_image_size = hparams.image_size; + + ctx_clip.has_llava_projector = ctx_clip.proj_type == PROJECTOR_TYPE_MLP + || ctx_clip.proj_type == PROJECTOR_TYPE_MLP_NORM + || ctx_clip.proj_type == PROJECTOR_TYPE_LDP + || ctx_clip.proj_type == PROJECTOR_TYPE_LDPV2; + + { + bool use_gelu = false; + bool use_silu = false; + get_bool(KEY_USE_GELU, use_gelu, false); + get_bool(KEY_USE_SILU, use_silu, false); + if (use_gelu && use_silu) { + throw std::runtime_error(string_format("%s: both use_gelu and use_silu are set to true\n", __func__)); + } + if (use_gelu) { + hparams.ffn_op = FFN_GELU; + log_ffn_op = "gelu"; + } else if (use_silu) { + hparams.ffn_op = FFN_SILU; + log_ffn_op = "silu"; + } else { + hparams.ffn_op = FFN_GELU_QUICK; + log_ffn_op = "gelu_quick"; + } + } + { std::string mm_patch_merge_type; get_string(KEY_MM_PATCH_MERGE_TYPE, mm_patch_merge_type, false); @@ -1491,35 +1879,78 @@ struct clip_model_loader { for (auto & layer : vision_feature_layer) { hparams.vision_feature_layer.insert(layer); } - // Calculate the deepest feature layer based on hparams and projector type - ctx_clip.max_feature_layer = get_deepest_feature_layer(&ctx_clip); - LOG_INF("%s: text_encoder: %d\n", __func__, ctx_clip.has_text_encoder); - LOG_INF("%s: vision_encoder: %d\n", __func__, ctx_clip.has_vision_encoder); - LOG_INF("%s: llava_projector: %d\n", __func__, ctx_clip.has_llava_projector); - LOG_INF("%s: minicpmv_projector: %d\n", __func__, ctx_clip.has_minicpmv_projector); + // model-specific params + switch (ctx_clip.proj_type) { + case PROJECTOR_TYPE_MINICPMV: + { + if (ctx_clip.minicpmv_version == 0) { + ctx_clip.minicpmv_version = 2; // default to 2 if not set + } + } break; + case PROJECTOR_TYPE_IDEFICS3: + case PROJECTOR_TYPE_INTERNVL: + { + get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor, false); + } break; + case PROJECTOR_TYPE_PIXTRAL: + { + hparams.rope_theta = 10000.0f; + hparams.warmup_image_size = hparams.patch_size * 8; + get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.spatial_merge_size, false); + } break; + case PROJECTOR_TYPE_GEMMA3: + { + // default value (used by all model sizes in gemma 3 family) + // number of patches for each **side** is reduced by a factor of 4 + hparams.proj_scale_factor = 4; + // test model (tinygemma3) has a different value, we optionally read it + get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor, false); + } break; + case PROJECTOR_TYPE_QWEN2VL: + { + // max image size = sqrt(max_pixels) = 3584 + // ref: https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct/blob/main/preprocessor_config.json + // however, the model use unreasonable memory past 1024 size, we force it to 1024 otherwise it's unusable + // ref: https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct/discussions/10 + hparams.image_size = 1024; + hparams.warmup_image_size = hparams.patch_size * 8; + } break; + case PROJECTOR_TYPE_QWEN25VL: + { + // max image size = sqrt(max_pixels) + // https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/blob/main/preprocessor_config.json + // however, the model use unreasonable memory past 1024 size, we force it to 1024 otherwise it's unusable + // ref: https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct/discussions/10 + hparams.image_size = 1024; + hparams.warmup_image_size = hparams.patch_size * 8; + get_u32(KEY_WIN_ATTN_PATTERN, hparams.n_wa_pattern); + } break; + default: + break; + } + + LOG_INF("%s: projector: %s\n", __func__, proj_type.c_str()); + LOG_INF("%s: n_embd: %d\n", __func__, hparams.n_embd); + LOG_INF("%s: n_head: %d\n", __func__, hparams.n_head); + LOG_INF("%s: n_ff: %d\n", __func__, hparams.n_ff); + LOG_INF("%s: n_layer: %d\n", __func__, hparams.n_layer); + LOG_INF("%s: projection_dim: %d\n", __func__, hparams.projection_dim); + LOG_INF("%s: image_size: %d\n", __func__, hparams.image_size); + LOG_INF("%s: patch_size: %d\n", __func__, hparams.patch_size); + LOG_INF("\n"); + LOG_INF("%s: has_llava_proj: %d\n", __func__, ctx_clip.has_llava_projector); LOG_INF("%s: minicpmv_version: %d\n", __func__, ctx_clip.minicpmv_version); - LOG_INF("%s: glm_projector: %d\n", __func__, ctx_clip.has_glm_projector); + LOG_INF("%s: proj_scale_factor: %d\n", __func__, hparams.proj_scale_factor); + LOG_INF("%s: n_wa_pattern: %d\n", __func__, hparams.n_wa_pattern); + LOG_INF("%s: ffn_op: %s\n", __func__, log_ffn_op.c_str()); LOG_INF("%s: model size: %.2f MiB\n", __func__, model_size / 1024.0 / 1024.0); LOG_INF("%s: metadata size: %.2f MiB\n", __func__, ggml_get_mem_size(ctx_meta.get()) / 1024.0 / 1024.0); } - - // model-specific params - switch (ctx_clip.proj_type) { - case PROJECTOR_TYPE_IDEFICS3: - { - get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor, false); - } break; - case PROJECTOR_TYPE_PIXTRAL: - { - hparams.rope_theta = 10000.0f; - } break; - default: - break; - } } void load_tensors() { + auto & hparams = ctx_clip.vision_model.hparams; std::map tensor_offset; std::vector tensors_to_load; @@ -1542,14 +1973,14 @@ struct clip_model_loader { // helper function auto get_tensor = [&](const std::string & name, bool required = true) { - struct ggml_tensor * cur = ggml_get_tensor(ctx_meta.get(), name.c_str()); + ggml_tensor * cur = ggml_get_tensor(ctx_meta.get(), name.c_str()); if (!cur && required) { throw std::runtime_error(string_format("%s: unable to find tensor %s\n", __func__, name.c_str())); } if (cur) { tensors_to_load.push_back(cur); // add tensors to context - struct ggml_tensor * data_tensor = ggml_dup_tensor(ctx_clip.ctx_data.get(), cur); + ggml_tensor * data_tensor = ggml_dup_tensor(ctx_clip.ctx_data.get(), cur); ggml_set_name(data_tensor, cur->name); cur = data_tensor; } @@ -1569,22 +2000,24 @@ struct clip_model_loader { vision_model.patch_bias = get_tensor(TN_PATCH_BIAS, false); vision_model.patch_embeddings_0 = get_tensor(TN_PATCH_EMBD, false); vision_model.patch_embeddings_1 = get_tensor(TN_PATCH_EMBD_1, false); - if (vision_model.patch_embeddings_1 == nullptr) { - ctx_clip.has_qwen2vl_merger = false; - } vision_model.position_embeddings = get_tensor(string_format(TN_POS_EMBD, "v"), false); // layers - vision_model.layers.resize(vision_model.hparams.n_layer); - for (int il = 0; il < vision_model.hparams.n_layer; ++il) { + vision_model.layers.resize(hparams.n_layer); + for (int il = 0; il < hparams.n_layer; ++il) { auto & layer = vision_model.layers[il]; layer.k_w = get_tensor(string_format(TN_ATTN_K, "v", il, "weight")); layer.q_w = get_tensor(string_format(TN_ATTN_Q, "v", il, "weight")); layer.v_w = get_tensor(string_format(TN_ATTN_V, "v", il, "weight")); layer.o_w = get_tensor(string_format(TN_ATTN_OUTPUT, "v", il, "weight")); + layer.k_norm = get_tensor(string_format(TN_ATTN_K_NORM, "v", il, "weight"), false); + layer.q_norm = get_tensor(string_format(TN_ATTN_Q_NORM, "v", il, "weight"), false); layer.ln_1_w = get_tensor(string_format(TN_LN_1, "v", il, "weight"), false); layer.ln_2_w = get_tensor(string_format(TN_LN_2, "v", il, "weight"), false); + layer.ls_1_w = get_tensor(string_format(TN_LS_1, "v", il, "weight"), false); // no bias + layer.ls_2_w = get_tensor(string_format(TN_LS_2, "v", il, "weight"), false); // no bias + layer.k_b = get_tensor(string_format(TN_ATTN_K, "v", il, "bias"), false); layer.q_b = get_tensor(string_format(TN_ATTN_Q, "v", il, "bias"), false); layer.v_b = get_tensor(string_format(TN_ATTN_V, "v", il, "bias"), false); @@ -1592,7 +2025,7 @@ struct clip_model_loader { layer.ln_1_b = get_tensor(string_format(TN_LN_1, "v", il, "bias"), false); layer.ln_2_b = get_tensor(string_format(TN_LN_2, "v", il, "bias"), false); - // new naming + // ffn layer.ff_up_w = get_tensor(string_format(TN_FFN_UP, "v", il, "weight")); layer.ff_up_b = get_tensor(string_format(TN_FFN_UP, "v", il, "bias"), false); layer.ff_gate_w = get_tensor(string_format(TN_FFN_GATE, "v", il, "weight"), false); @@ -1600,11 +2033,18 @@ struct clip_model_loader { layer.ff_down_w = get_tensor(string_format(TN_FFN_DOWN, "v", il, "weight")); layer.ff_down_b = get_tensor(string_format(TN_FFN_DOWN, "v", il, "bias"), false); - // legacy naming (the in and out is reversed! don't ask me why) - layer.ff_i_w = layer.ff_down_w; - layer.ff_o_w = layer.ff_up_w; - layer.ff_i_b = layer.ff_down_b; - layer.ff_o_b = layer.ff_up_b; + // some models already exported with legacy (incorrect) naming which is quite messy, let's fix it here + // note: Qwen model converted from the old surgery script has n_ff = 0, so we cannot use n_ff to check! + if (layer.ff_up_w && layer.ff_down_w && layer.ff_down_w->ne[0] == hparams.n_embd) { + // swap up and down weights + ggml_tensor * tmp = layer.ff_up_w; + layer.ff_up_w = layer.ff_down_w; + layer.ff_down_w = tmp; + // swap up and down biases + tmp = layer.ff_up_b; + layer.ff_up_b = layer.ff_down_b; + layer.ff_down_b = tmp; + } } switch (ctx_clip.proj_type) { @@ -1669,7 +2109,7 @@ struct clip_model_loader { vision_model.mm_model_peg_0_w = get_tensor(string_format(TN_MVLM_PROJ_PEG, 0, "weight")); vision_model.mm_model_peg_0_b = get_tensor(string_format(TN_MVLM_PROJ_PEG, 0, "bias")); } break; - case PROJECTOR_TYPE_RESAMPLER: + case PROJECTOR_TYPE_MINICPMV: { // vision_model.mm_model_pos_embed = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD); vision_model.mm_model_pos_embed_k = get_tensor(TN_MINICPMV_POS_EMBD_K); @@ -1695,14 +2135,17 @@ struct clip_model_loader { { vision_model.mm_model_adapter_conv_w = get_tensor(string_format(TN_GLM_ADAPER_CONV, "weight")); vision_model.mm_model_adapter_conv_b = get_tensor(string_format(TN_GLM_ADAPER_CONV, "bias")); - vision_model.mm_model_mlp_0_w = get_tensor(string_format(TN_GLM_ADAPTER_LINEAR,"weight")); - vision_model.mm_model_ln_q_w = get_tensor(string_format(TN_GLM_ADAPTER_NORM_1,"weight")); - vision_model.mm_model_ln_q_b = get_tensor(string_format(TN_GLM_ADAPTER_NORM_1,"bias")); - vision_model.mm_model_mlp_1_w = get_tensor(string_format(TN_GLM_ADAPTER_D_H_2_4H,"weight")); - vision_model.mm_model_mlp_2_w = get_tensor(string_format(TN_GLM_ADAPTER_GATE,"weight")); - vision_model.mm_model_mlp_3_w = get_tensor(string_format(TN_GLM_ADAPTER_D_4H_2_H,"weight")); + vision_model.mm_model_mlp_0_w = get_tensor(string_format(TN_GLM_ADAPTER_LINEAR, "weight")); + vision_model.mm_model_ln_q_w = get_tensor(string_format(TN_GLM_ADAPTER_NORM_1, "weight")); + vision_model.mm_model_ln_q_b = get_tensor(string_format(TN_GLM_ADAPTER_NORM_1, "bias")); + vision_model.mm_model_mlp_1_w = get_tensor(string_format(TN_GLM_ADAPTER_D_H_2_4H, "weight")); + vision_model.mm_model_mlp_2_w = get_tensor(string_format(TN_GLM_ADAPTER_GATE, "weight")); + vision_model.mm_model_mlp_3_w = get_tensor(string_format(TN_GLM_ADAPTER_D_4H_2_H, "weight")); + vision_model.mm_glm_tok_boi = get_tensor(string_format(TN_TOK_GLM_BOI, "weight")); + vision_model.mm_glm_tok_eoi = get_tensor(string_format(TN_TOK_GLM_EOI, "weight")); } break; - case PROJECTOR_TYPE_MERGER: + case PROJECTOR_TYPE_QWEN2VL: + case PROJECTOR_TYPE_QWEN25VL: { vision_model.mm_0_w = get_tensor(string_format(TN_LLAVA_PROJ, 0, "weight")); vision_model.mm_0_b = get_tensor(string_format(TN_LLAVA_PROJ, 0, "bias")); @@ -1721,11 +2164,23 @@ struct clip_model_loader { case PROJECTOR_TYPE_PIXTRAL: { vision_model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight")); - vision_model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias")); + vision_model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"), false); vision_model.mm_2_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight")); - vision_model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias")); + vision_model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"), false); // [IMG_BREAK] token embedding vision_model.token_embd_img_break = get_tensor(TN_TOK_IMG_BREAK); + // for mistral small 3.1 + vision_model.mm_input_norm_w = get_tensor(TN_MM_INP_NORM, false); + vision_model.mm_patch_merger_w = get_tensor(TN_MM_PATCH_MERGER, false); + } break; + case PROJECTOR_TYPE_INTERNVL: + { + vision_model.mm_0_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "weight")); + vision_model.mm_0_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "bias")); + vision_model.mm_1_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "weight")); + vision_model.mm_1_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "bias")); + vision_model.mm_3_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "weight")); + vision_model.mm_3_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "bias")); } break; default: GGML_ASSERT(false && "unknown projector type"); @@ -1745,7 +2200,7 @@ struct clip_model_loader { ctx_clip.buf.reset(ggml_backend_alloc_ctx_tensors_from_buft(ctx_clip.ctx_data.get(), buft)); ggml_backend_buffer_set_usage(ctx_clip.buf.get(), GGML_BACKEND_BUFFER_USAGE_WEIGHTS); for (auto & t : tensors_to_load) { - struct ggml_tensor * cur = ggml_get_tensor(ctx_clip.ctx_data.get(), t->name); + ggml_tensor * cur = ggml_get_tensor(ctx_clip.ctx_data.get(), t->name); const size_t offset = tensor_offset[t->name]; fin.seekg(offset, std::ios::beg); if (!fin) { @@ -1774,16 +2229,14 @@ struct clip_model_loader { // create a fake batch clip_image_f32_batch batch; clip_image_f32_ptr img(clip_image_f32_init()); - clip_image_size image_size; - image_size.width = ctx_clip.vision_model.hparams.image_size; - image_size.height = ctx_clip.vision_model.hparams.image_size; - img->nx = image_size.width; - img->ny = image_size.height; - img->buf.resize(image_size.width * image_size.height * 3); + img->nx = ctx_clip.vision_model.hparams.warmup_image_size; + img->ny = ctx_clip.vision_model.hparams.warmup_image_size; + img->buf.resize(img->nx * img->ny * 3); batch.entries.push_back(std::move(img)); - ggml_cgraph * gf = clip_image_build_graph(&ctx_clip, batch, image_size, false); + ggml_cgraph * gf = clip_image_build_graph(&ctx_clip, batch); ggml_backend_sched_reserve(ctx_clip.sched.get(), gf); + for (size_t i = 0; i < ctx_clip.backend_ptrs.size(); ++i) { ggml_backend_t backend = ctx_clip.backend_ptrs[i]; ggml_backend_buffer_type_t buft = ctx_clip.backend_buft[i]; @@ -1866,9 +2319,10 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity) { struct clip_ctx * clip_init(const char * fname, struct clip_context_params ctx_params) { g_logger_state.verbosity_thold = ctx_params.verbosity; - clip_ctx * ctx_clip = new clip_ctx(ctx_params); + clip_ctx * ctx_clip = nullptr; try { + ctx_clip = new clip_ctx(ctx_params); clip_model_loader loader(fname, *ctx_clip); loader.load_hparams(); loader.load_tensors(); @@ -2181,8 +2635,8 @@ struct image_manipulation { float target_width_f = static_cast(inp_size.width) * scale; float target_height_f = static_cast(inp_size.height) * scale; - int aligned_width = GGML_PAD((int)target_width_f, align_size); - int aligned_height = GGML_PAD((int)target_height_f, align_size); + int aligned_width = CLIP_ALIGN((int)target_width_f, align_size); + int aligned_height = CLIP_ALIGN((int)target_height_f, align_size); return {aligned_width, aligned_height}; } @@ -2280,7 +2734,7 @@ struct llava_uhd { // no pinpoints, dynamically calculate the grid size (e.g. minicpmv) - auto best_size = get_best_resize(original_size, slice_size, patch_size, has_slices); + auto best_size = get_best_resize(original_size, slice_size, patch_size, !has_slices); res.overview_size = best_size; if (!has_slices) { @@ -2479,11 +2933,6 @@ int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip) { // returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector // res_imgs memory is being allocated here, previous allocations will be freed if found bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, struct clip_image_f32_batch * res_imgs) { - if (!ctx->has_vision_encoder) { - LOG_ERR("%s: This gguf file seems to have no vision encoder\n", __func__); - return false; - } - clip_image_size original_size{img->nx, img->ny}; bool pad_to_square = true; auto & params = ctx->vision_model.hparams; @@ -2504,12 +2953,11 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str } return true; } - else if (ctx->has_qwen2vl_merger) { + else if (ctx->proj_type == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type == PROJECTOR_TYPE_QWEN25VL) { clip_image_u8 resized; - auto patch_size = clip_get_patch_size(ctx) * 2; - int nx = ceil((float)img->nx / patch_size) * patch_size; - int ny = ceil((float)img->ny / patch_size) * patch_size; - image_manipulation::bicubic_resize(*img, resized, nx, ny); + auto patch_size = params.patch_size * 2; + auto new_size = image_manipulation::calc_size_preserved_ratio(original_size, patch_size, params.image_size); + image_manipulation::bicubic_resize(*img, resized, new_size.width, new_size.height); clip_image_f32_ptr img_f32(clip_image_f32_init()); // clip_image_f32_ptr res(clip_image_f32_init()); @@ -2518,9 +2966,11 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str res_imgs->entries.push_back(std::move(img_f32)); return true; } - else if (ctx->has_glm_projector + else if (ctx->proj_type == PROJECTOR_TYPE_GLM_EDGE || ctx->proj_type == PROJECTOR_TYPE_GEMMA3 - || ctx->proj_type == PROJECTOR_TYPE_IDEFICS3) { + || ctx->proj_type == PROJECTOR_TYPE_IDEFICS3 + || ctx->proj_type == PROJECTOR_TYPE_INTERNVL // TODO @ngxson : support dynamic resolution + ) { clip_image_u8 resized_image; int sz = params.image_size; image_manipulation::resize_and_pad_image(*img, resized_image, {sz, sz}); @@ -2594,15 +3044,18 @@ void clip_free(clip_ctx * ctx) { delete ctx; } +// deprecated size_t clip_embd_nbytes(const struct clip_ctx * ctx) { - return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float); + const int32_t nx = ctx->vision_model.hparams.image_size; + const int32_t ny = ctx->vision_model.hparams.image_size; + return clip_embd_nbytes_by_img(ctx, nx, ny); } -size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w) { +size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_w, int img_h) { clip_image_f32 img; img.nx = img_w; img.ny = img_h; - return clip_n_patches_by_img(ctx, &img) * clip_n_mmproj_embd(ctx) * sizeof(float); + return clip_n_output_tokens(ctx, &img) * clip_n_mmproj_embd(ctx) * sizeof(float); } int32_t clip_get_image_size(const struct clip_ctx * ctx) { @@ -2614,7 +3067,7 @@ int32_t clip_get_patch_size(const struct clip_ctx * ctx) { } int32_t clip_get_hidden_size(const struct clip_ctx * ctx) { - return ctx->vision_model.hparams.hidden_size; + return ctx->vision_model.hparams.n_embd; } const char * clip_patch_merge_type(const struct clip_ctx * ctx) { @@ -2632,21 +3085,49 @@ size_t get_clip_image_grid_size(const struct clip_ctx * ctx) { return ctx->vision_model.hparams.image_grid_pinpoints.size(); } +// deprecated int clip_n_patches(const struct clip_ctx * ctx) { clip_image_f32 img; img.nx = ctx->vision_model.hparams.image_size; img.ny = ctx->vision_model.hparams.image_size; - return clip_n_patches_by_img(ctx, &img); + return clip_n_output_tokens(ctx, &img); } +// deprecated int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * img) { + return clip_n_output_tokens(ctx, img); +} + +int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img) { + const auto & params = ctx->vision_model.hparams; + const int n_total = clip_n_output_tokens(ctx, img); + if (ctx->proj_type == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type == PROJECTOR_TYPE_QWEN25VL) { + return img->nx / (params.patch_size * 2) + (int)(img->nx % params.patch_size > 0); + } + return n_total; +} + +int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 * img) { + const auto & params = ctx->vision_model.hparams; + if (ctx->proj_type == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type == PROJECTOR_TYPE_QWEN25VL) { + return img->ny / (params.patch_size * 2) + (int)(img->ny % params.patch_size > 0); + } + return 1; +} + +int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img) { const auto & params = ctx->vision_model.hparams; int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size); - if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2 || ctx->proj_type == PROJECTOR_TYPE_GLM_EDGE) { + if (ctx->proj_type == PROJECTOR_TYPE_LDP + || ctx->proj_type == PROJECTOR_TYPE_LDPV2 + || ctx->proj_type == PROJECTOR_TYPE_GLM_EDGE) { n_patches /= 4; - } else if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) { + if (ctx->vision_model.mm_glm_tok_boi) { + n_patches += 2; // for BOI and EOI token embeddings + } + } else if (ctx->proj_type == PROJECTOR_TYPE_MINICPMV) { if (ctx->minicpmv_version == 2) { n_patches = 96; } @@ -2656,18 +3137,25 @@ int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * i else if (ctx->minicpmv_version == 4) { n_patches = 64; } - } else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) { + else { + GGML_ABORT("Unknown minicpmv version"); + } + } else if (ctx->proj_type == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type == PROJECTOR_TYPE_QWEN25VL) { int patch_size = params.patch_size * 2; int x_patch = img->nx / patch_size + (int)(img->nx % patch_size > 0); int y_patch = img->ny / patch_size + (int)(img->ny % patch_size > 0); n_patches = x_patch * y_patch; } else if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) { - n_patches = 256; - } else if (ctx->proj_type == PROJECTOR_TYPE_IDEFICS3) { - n_patches /= ctx->vision_model.hparams.proj_scale_factor; + int n_per_side = params.image_size / params.patch_size; + int n_per_side_2d_pool = n_per_side / params.proj_scale_factor; + n_patches = n_per_side_2d_pool * n_per_side_2d_pool; + } else if (ctx->proj_type == PROJECTOR_TYPE_IDEFICS3 || ctx->proj_type == PROJECTOR_TYPE_INTERNVL) { + // both W and H are divided by proj_scale_factor + n_patches /= (params.proj_scale_factor * params.proj_scale_factor); } else if (ctx->proj_type == PROJECTOR_TYPE_PIXTRAL) { - int n_patches_x = img->nx / params.patch_size; - int n_patches_y = img->ny / params.patch_size; + int n_merge = params.spatial_merge_size; + int n_patches_x = img->nx / params.patch_size / (n_merge > 0 ? n_merge : 1); + int n_patches_y = img->ny / params.patch_size / (n_merge > 0 ? n_merge : 1); n_patches = n_patches_y*n_patches_x + n_patches_y - 1; // + one [IMG_BREAK] per row, except the last row } @@ -2761,11 +3249,6 @@ static std::vector> get_2d_sincos_pos_embed(int embed_dim, co } bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) { - if (!ctx->has_vision_encoder) { - LOG_ERR("%s: This gguf file seems to have no vision encoder\n", __func__); - return false; - } - clip_image_f32_batch imgs; clip_image_f32_ptr img_copy(clip_image_f32_init()); *img_copy = *img; @@ -2776,54 +3259,66 @@ bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f3 bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs_c_ptr, float * vec) { const clip_image_f32_batch & imgs = *imgs_c_ptr; - - if (!ctx->has_vision_encoder) { - LOG_ERR("%s: This gguf file seems to have no vision encoder\n", __func__); - return false; - } - int batch_size = imgs.entries.size(); - if (ctx->has_llava_projector) { - GGML_ASSERT(batch_size == 1); // TODO: support multiple images - } - if (ctx->has_minicpmv_projector) { - GGML_ASSERT(batch_size == 1); - } - if (ctx->has_glm_projector) { - GGML_ASSERT(batch_size == 1); + + // TODO @ngxson : implement batch size > 1 as a loop + // we don't need true batching support because the cgraph will gonna be big anyway + if (batch_size != 1) { + return false; // only support batch size of 1 } // build the inference graph ggml_backend_sched_reset(ctx->sched.get()); - ggml_cgraph * gf = clip_image_build_graph(ctx, imgs, ctx->load_image_size, true); + ggml_cgraph * gf = clip_image_build_graph(ctx, imgs); ggml_backend_sched_alloc_graph(ctx->sched.get(), gf); // set inputs - const auto & model = ctx->vision_model; + const auto & model = ctx->vision_model; const auto & hparams = model.hparams; - // TODO @ngxson : this is ugly, need to refactor later - bool support_dynamic_size = ctx->has_minicpmv_projector - || ctx->has_qwen2vl_merger - || ctx->proj_type == PROJECTOR_TYPE_PIXTRAL; + const int image_size_width = imgs.entries[0]->nx; + const int image_size_height = imgs.entries[0]->ny; - const int image_size = hparams.image_size; - int image_size_width = image_size; - int image_size_height = image_size; - if (support_dynamic_size) { - image_size_width = imgs.entries[0]->nx; - image_size_height = imgs.entries[0]->ny; - } const int patch_size = hparams.patch_size; const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size)); - const int num_positions = num_patches + (model.class_embedding ? 1 : 0); - const int pos_w = ctx->load_image_size.width / patch_size; + const int n_pos = num_patches + (model.class_embedding ? 1 : 0); + const int pos_w = ctx->load_image_size.width / patch_size; const int pos_h = ctx->load_image_size.height / patch_size; + const bool use_window_attn = hparams.n_wa_pattern > 0; // for qwen2.5vl + + auto get_inp_tensor = [&gf](const char * name) { + ggml_tensor * inp = ggml_graph_get_tensor(gf, name); + if (inp == nullptr) { + GGML_ABORT("Failed to get tensor %s", name); + } + if (!(inp->flags & GGML_TENSOR_FLAG_INPUT)) { + GGML_ABORT("Tensor %s is not an input tensor", name); + } + return inp; + }; + + auto set_input_f32 = [&get_inp_tensor](const char * name, std::vector & values) { + ggml_tensor * cur = get_inp_tensor(name); + GGML_ASSERT(cur->type == GGML_TYPE_F32); + GGML_ASSERT(ggml_nelements(cur) == (int64_t)values.size()); + ggml_backend_tensor_set(cur, values.data(), 0, ggml_nbytes(cur)); + }; + + auto set_input_i32 = [&get_inp_tensor](const char * name, std::vector & values) { + ggml_tensor * cur = get_inp_tensor(name); + GGML_ASSERT(cur->type == GGML_TYPE_I32); + GGML_ASSERT(ggml_nelements(cur) == (int64_t)values.size()); + ggml_backend_tensor_set(cur, values.data(), 0, ggml_nbytes(cur)); + }; + + // set input pixel values { - struct ggml_tensor * inp_raw = ggml_graph_get_tensor(gf, "inp_raw"); - std::vector inp_data(ggml_nelements(inp_raw)); - float * data = inp_data.data(); + size_t nelem = 0; + for (const auto & img : imgs.entries) { + nelem += img->nx * img->ny * 3; + } + std::vector inp_raw(nelem); // layout of data (note: the channel dim is unrolled to better visualize the layout): // @@ -2839,18 +3334,10 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima for (size_t i = 0; i < imgs.entries.size(); i++) { const int nx = imgs.entries[i]->nx; const int ny = imgs.entries[i]->ny; - - if (ctx->has_glm_projector - || ctx->has_llava_projector - || ctx->proj_type == PROJECTOR_TYPE_GEMMA3 - || ctx->proj_type == PROJECTOR_TYPE_IDEFICS3) { - GGML_ASSERT(nx == image_size && ny == image_size); - } - const int n = nx * ny; for (int b = 0; b < batch_size; b++) { - float * batch_entry = data + b * (3*n); + float * batch_entry = inp_raw.data() + b * (3*n); for (int y = 0; y < ny; y++) { for (int x = 0; x < nx; x++) { size_t base_src = 3*(y * nx + x); // idx of the first channel @@ -2862,148 +3349,219 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima } } } - ggml_backend_tensor_set(inp_raw, data, 0, ggml_nbytes(inp_raw)); + set_input_f32("inp_raw", inp_raw); } - if (ctx->has_minicpmv_projector) { - { - // inspired from siglip: - // -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit - // -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit/blob/d66538faeba44480d0bfaa42145eef26f9423199/modeling_siglip.py#L316 - struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions"); - int* positions_data = (int*)malloc(ggml_nbytes(positions)); - int bucket_coords_h[1024]; - int bucket_coords_w[1024]; - for (int i = 0; i < pos_h; i++){ - bucket_coords_h[i] = std::floor(70.0*i/pos_h); - } - for (int i = 0; i < pos_w; i++){ - bucket_coords_w[i] = std::floor(70.0*i/pos_w); - } - for (int i = 0, id = 0; i < pos_h; i++){ - for (int j = 0; j < pos_w; j++){ - positions_data[id++] = bucket_coords_h[i]*70 + bucket_coords_w[j]; - } - } - ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions)); - free(positions_data); - } - { - // inspired from resampler of Qwen-VL: - // -> https://huggingface.co/Qwen/Qwen-VL/tree/main - // -> https://huggingface.co/Qwen/Qwen-VL/blob/0547ed36a86561e2e42fecec8fd0c4f6953e33c4/visual.py#L23 - struct ggml_tensor * pos_embed = ggml_graph_get_tensor(gf, "pos_embed"); - int embed_dim = 4096; - if (ctx->minicpmv_version == 2) { - embed_dim = 4096; - } - else if (ctx->minicpmv_version == 3) { - embed_dim = 3584; - } - else if (ctx->minicpmv_version == 4) { - embed_dim = 3584; - } - auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h)); - - float * pos_embed_data = (float *)malloc(ggml_nbytes(pos_embed)); - for(int i=0;i < pos_w * pos_h; ++i){ - for(int j=0; j < embed_dim; ++j){ - pos_embed_data[i * embed_dim + j] = pos_embed_t[i][j]; - } - } - - ggml_backend_tensor_set(pos_embed, pos_embed_data, 0, ggml_nbytes(pos_embed)); - free(pos_embed_data); - } - } - else { - if (model.class_embedding) { - struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings"); - - void* zero_mem = malloc(ggml_nbytes(embeddings)); - memset(zero_mem, 0, ggml_nbytes(embeddings)); - ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings)); - free(zero_mem); - } - - if (ctx->has_qwen2vl_merger) { - struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions"); - - const int pw = image_size_width / patch_size; - const int ph = image_size_height / patch_size; - int* positions_data = (int*)malloc(ggml_nbytes(positions)); - - int ptr = 0; - for (int y = 0; y < ph; y+=2) + // set input per projector + switch (ctx->proj_type) { + case PROJECTOR_TYPE_MINICPMV: { - for (int x = 0; x < pw; x+=2) - { - for (int dy = 0; dy < 2; dy++) { - for (int dx = 0; dx < 2; dx++) { - positions_data[ptr] = y + dy; - positions_data[num_patches + ptr] = x + dx; - positions_data[num_patches * 2 + ptr] = y + dy; - positions_data[num_patches * 3 + ptr] = x + dx; - ptr++; + // inspired from siglip: + // -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit + // -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit/blob/d66538faeba44480d0bfaa42145eef26f9423199/modeling_siglip.py#L316 + std::vector positions(pos_h * pos_w); + int bucket_coords_h[1024]; + int bucket_coords_w[1024]; + for (int i = 0; i < pos_h; i++){ + bucket_coords_h[i] = std::floor(70.0*i/pos_h); + } + for (int i = 0; i < pos_w; i++){ + bucket_coords_w[i] = std::floor(70.0*i/pos_w); + } + for (int i = 0, id = 0; i < pos_h; i++){ + for (int j = 0; j < pos_w; j++){ + positions[id++] = bucket_coords_h[i]*70 + bucket_coords_w[j]; + } + } + set_input_i32("positions", positions); + + // inspired from resampler of Qwen-VL: + // -> https://huggingface.co/Qwen/Qwen-VL/tree/main + // -> https://huggingface.co/Qwen/Qwen-VL/blob/0547ed36a86561e2e42fecec8fd0c4f6953e33c4/visual.py#L23 + int embed_dim = clip_n_mmproj_embd(ctx); + + // TODO @ngxson : this is very inefficient, can we do this using ggml_sin and ggml_cos? + auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h)); + + std::vector pos_embed(embed_dim * pos_w * pos_h); + for(int i = 0; i < pos_w * pos_h; ++i){ + for(int j = 0; j < embed_dim; ++j){ + pos_embed[i * embed_dim + j] = pos_embed_t[i][j]; + } + } + + set_input_f32("pos_embed", pos_embed); + } break; + case PROJECTOR_TYPE_QWEN2VL: + { + const int merge_ratio = 2; + const int pw = image_size_width / patch_size; + const int ph = image_size_height / patch_size; + std::vector positions(n_pos * 4); + int ptr = 0; + for (int y = 0; y < ph; y += merge_ratio) { + for (int x = 0; x < pw; x += merge_ratio) { + for (int dy = 0; dy < 2; dy++) { + for (int dx = 0; dx < 2; dx++) { + positions[ ptr] = y + dy; + positions[ num_patches + ptr] = x + dx; + positions[2 * num_patches + ptr] = y + dy; + positions[3 * num_patches + ptr] = x + dx; + ptr++; + } } } } - } - ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions)); - free(positions_data); - } - else if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) { - // do nothing - } - else if (ctx->proj_type == PROJECTOR_TYPE_IDEFICS3) { - // do nothing - } - else if (ctx->proj_type == PROJECTOR_TYPE_PIXTRAL) { - // set the 2D positions - int n_patches_per_col = image_size_width / patch_size; - std::vector pos_data(num_positions); - struct ggml_tensor * pos; - // dimension H - pos = ggml_graph_get_tensor(gf, "pos_h"); - for (int i = 0; i < num_positions; i++) { - pos_data[i] = i / n_patches_per_col; - } - ggml_backend_tensor_set(pos, pos_data.data(), 0, ggml_nbytes(pos)); - // dimension W - pos = ggml_graph_get_tensor(gf, "pos_w"); - for (int i = 0; i < num_positions; i++) { - pos_data[i] = i % n_patches_per_col; - } - ggml_backend_tensor_set(pos, pos_data.data(), 0, ggml_nbytes(pos)); - } - else { - struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions"); + set_input_i32("positions", positions); + } break; + case PROJECTOR_TYPE_QWEN25VL: + { + // pw * ph = number of tokens output by ViT after apply patch merger + // ipw * ipw = number of vision token been processed inside ViT + const int merge_ratio = 2; + const int pw = image_size_width / patch_size / merge_ratio; + const int ph = image_size_height / patch_size / merge_ratio; + const int ipw = image_size_width / patch_size; + const int iph = image_size_height / patch_size; - int* positions_data = (int*)malloc(ggml_nbytes(positions)); - for (int i = 0; i < num_positions; i++) { - positions_data[i] = i; - } - ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions)); - free(positions_data); + std::vector idx (ph * pw); + std::vector inv_idx(ph * pw); + + if (use_window_attn) { + const int attn_window_size = 112; + const int grid_window = attn_window_size / patch_size / merge_ratio; + int dst = 0; + // [num_vision_tokens, num_vision_tokens] attention mask tensor + std::vector mask(pow(ipw * iph, 2), std::numeric_limits::lowest()); + int mask_row = 0; + + for (int y = 0; y < ph; y += grid_window) { + for (int x = 0; x < pw; x += grid_window) { + const int win_h = std::min(grid_window, ph - y); + const int win_w = std::min(grid_window, pw - x); + const int dst_0 = dst; + // group all tokens belong to the same window togather (to a continue range) + for (int dy = 0; dy < win_h; dy++) { + for (int dx = 0; dx < win_w; dx++) { + const int src = (y + dy) * pw + (x + dx); + GGML_ASSERT(src < (int)idx.size()); + GGML_ASSERT(dst < (int)inv_idx.size()); + idx [src] = dst; + inv_idx[dst] = src; + dst++; + } + } + + for (int r=0; r < win_h * win_w * merge_ratio * merge_ratio; r++) { + int row_offset = mask_row * (ipw * iph); + std::fill( + mask.begin() + row_offset + (dst_0 * merge_ratio * merge_ratio), + mask.begin() + row_offset + (dst * merge_ratio * merge_ratio), + 0.0); + mask_row++; + } + } + } + + set_input_i32("window_idx", idx); + set_input_i32("inv_window_idx", inv_idx); + set_input_f32("window_mask", mask); + } else { + for (int i = 0; i < ph * pw; i++) { + idx[i] = i; + } + } + + const int mpow = merge_ratio * merge_ratio; + std::vector positions(n_pos * 4); + + int ptr = 0; + for (int y = 0; y < iph; y += merge_ratio) { + for (int x = 0; x < ipw; x += merge_ratio) { + for (int dy = 0; dy < 2; dy++) { + for (int dx = 0; dx < 2; dx++) { + auto remap = idx[ptr / mpow]; + remap = (remap * mpow) + (ptr % mpow); + + positions[ remap] = y + dy; + positions[ num_patches + remap] = x + dx; + positions[2 * num_patches + remap] = y + dy; + positions[3 * num_patches + remap] = x + dx; + ptr++; + } + } + } + } + + set_input_i32("positions", positions); + } break; + case PROJECTOR_TYPE_PIXTRAL: + { + // set the 2D positions + int n_patches_per_col = image_size_width / patch_size; + std::vector pos_data(n_pos); + // dimension H + for (int i = 0; i < n_pos; i++) { + pos_data[i] = i / n_patches_per_col; + } + set_input_i32("pos_h", pos_data); + // dimension W + for (int i = 0; i < n_pos; i++) { + pos_data[i] = i % n_patches_per_col; + } + set_input_i32("pos_w", pos_data); + } break; + case PROJECTOR_TYPE_GLM_EDGE: + { + // llava and other models + std::vector positions(n_pos); + for (int i = 0; i < n_pos; i++) { + positions[i] = i; + } + set_input_i32("positions", positions); + } break; + case PROJECTOR_TYPE_MLP: + case PROJECTOR_TYPE_MLP_NORM: + case PROJECTOR_TYPE_LDP: + case PROJECTOR_TYPE_LDPV2: + { + // llava and other models + std::vector positions(n_pos); + for (int i = 0; i < n_pos; i++) { + positions[i] = i; + } + set_input_i32("positions", positions); - if (!ctx->has_glm_projector) { - struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches"); // The patches vector is used to get rows to index into the embeds with; // we should skip dim 0 only if we have CLS to avoid going out of bounds // when retrieving the rows. int patch_offset = model.class_embedding ? 1 : 0; - int* patches_data = (int*)malloc(ggml_nbytes(patches)); + std::vector patches(num_patches); for (int i = 0; i < num_patches; i++) { - patches_data[i] = i + patch_offset; + patches[i] = i + patch_offset; } - ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches)); - free(patches_data); - } - } + set_input_i32("patches", patches); + } break; + case PROJECTOR_TYPE_GEMMA3: + case PROJECTOR_TYPE_IDEFICS3: + case PROJECTOR_TYPE_INTERNVL: + { + // do nothing + } break; + default: + GGML_ABORT("Unknown projector type"); } - ggml_backend_cpu_set_n_threads(ctx->backend_cpu, n_threads); + // ggml_backend_cpu_set_n_threads(ctx->backend_cpu, n_threads); + ggml_backend_dev_t dev = ggml_backend_get_device(ctx->backend_cpu); + ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr; + if (reg) { + auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads"); + if (ggml_backend_set_n_threads_fn) { + ggml_backend_set_n_threads_fn(ctx->backend_cpu, n_threads); + } + } auto status = ggml_backend_sched_graph_compute(ctx->sched.get(), gf); if (status != GGML_STATUS_SUCCESS) { @@ -3012,7 +3570,15 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima } // the last node is the embedding tensor - struct ggml_tensor * embeddings = ggml_graph_node(gf, -1); + ggml_tensor * embeddings = ggml_graph_node(gf, -1); + + // sanity check (only support batch size of 1 for now) + const int n_tokens_out = embeddings->ne[1]; + const int expected_n_tokens_out = clip_n_output_tokens(ctx, imgs.entries[0].get()); + if (n_tokens_out != expected_n_tokens_out) { + LOG_ERR("%s: expected %d tokens, got %d\n", __func__, expected_n_tokens_out, n_tokens_out); + GGML_ABORT("Invalid number of output tokens"); + } // copy the embeddings to the location passed by the user ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings)); @@ -3043,7 +3609,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i for (int i = 0; i < n_tensors; ++i) { const char * name = gguf_get_tensor_name(ctx_src, i); - struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name); + ggml_tensor * cur = ggml_get_tensor(ctx_data, name); gguf_add_tensor(ctx_out, cur); } @@ -3064,7 +3630,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i for (int i = 0; i < n_tensors; ++i) { const std::string name = gguf_get_tensor_name(ctx_src, i); - struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name.c_str()); + ggml_tensor * cur = ggml_get_tensor(ctx_data, name.c_str()); enum ggml_type new_type; void * new_data; @@ -3163,10 +3729,10 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) { return ctx->vision_model.mm_model_peg_0_b->ne[0]; case PROJECTOR_TYPE_MLP: case PROJECTOR_TYPE_PIXTRAL: - return ctx->vision_model.mm_2_b->ne[0]; + return ctx->vision_model.mm_2_w->ne[1]; case PROJECTOR_TYPE_MLP_NORM: return ctx->vision_model.mm_3_b->ne[0]; - case PROJECTOR_TYPE_RESAMPLER: + case PROJECTOR_TYPE_MINICPMV: if (ctx->minicpmv_version == 2) { return 4096; } else if (ctx->minicpmv_version == 3) { @@ -3174,36 +3740,36 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) { } else if (ctx->minicpmv_version == 4) { return 3584; } - break; // Should not happen if version is valid + GGML_ABORT("Unknown minicpmv version"); case PROJECTOR_TYPE_GLM_EDGE: return ctx->vision_model.mm_model_mlp_3_w->ne[1]; - case PROJECTOR_TYPE_MERGER: + case PROJECTOR_TYPE_QWEN2VL: + case PROJECTOR_TYPE_QWEN25VL: return ctx->vision_model.mm_1_b->ne[0]; case PROJECTOR_TYPE_GEMMA3: return ctx->vision_model.mm_input_proj_w->ne[0]; case PROJECTOR_TYPE_IDEFICS3: return ctx->vision_model.projection->ne[1]; + case PROJECTOR_TYPE_INTERNVL: + return ctx->vision_model.mm_3_w->ne[1]; default: - break; // Fall through to throw + GGML_ABORT("Unknown projector type"); } - - std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type]; - throw std::runtime_error(string_format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); } int clip_is_minicpmv(const struct clip_ctx * ctx) { - if (ctx->has_minicpmv_projector) { + if (ctx->proj_type == PROJECTOR_TYPE_MINICPMV) { return ctx->minicpmv_version; } return 0; } bool clip_is_glm(const struct clip_ctx * ctx) { - return ctx->has_glm_projector; + return ctx->proj_type == PROJECTOR_TYPE_GLM_EDGE; } bool clip_is_qwen2vl(const struct clip_ctx * ctx) { - return ctx->has_qwen2vl_merger; + return ctx->proj_type == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type == PROJECTOR_TYPE_QWEN25VL; } bool clip_is_llava(const struct clip_ctx * ctx) { @@ -3214,29 +3780,6 @@ bool clip_is_gemma3(const struct clip_ctx * ctx) { return ctx->proj_type == PROJECTOR_TYPE_GEMMA3; } -// Determine the number of encoder layers to iterate over -int get_deepest_feature_layer(const struct clip_ctx * ctx) { - // Get the index of the second to last layer; this is the - // default for models that have a llava projector - const auto & hparams = ctx->vision_model.hparams; - int n_layer = hparams.n_layer - 1; - int deepest_feature_layer = -1; - - // Handle other projectors; incrementing here indicates that we - // should use the last encoder layer for the vision features. - if (ctx->has_minicpmv_projector || ctx->has_glm_projector || ctx->has_qwen2vl_merger) { - n_layer += 1; - } - - // If we set explicit vision feature layers, only go up to the deepest one - for (const auto & feature_layer : hparams.vision_feature_layer) { - if (feature_layer > deepest_feature_layer) { - deepest_feature_layer = feature_layer; - } - } - return deepest_feature_layer < 0 ? n_layer : deepest_feature_layer; -} - bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec) { clip_image_f32 clip_img; clip_img.buf.resize(h * w * 3); diff --git a/examples/llava/clip.h b/tools/mtmd/clip.h similarity index 81% rename from examples/llava/clip.h rename to tools/mtmd/clip.h index 5fc45d3e23..0b0eb02956 100644 --- a/examples/llava/clip.h +++ b/tools/mtmd/clip.h @@ -47,7 +47,7 @@ CLIP_API struct clip_ctx * clip_init(const char * fname, struct clip_context_par CLIP_API void clip_free(struct clip_ctx * ctx); CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx); -CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w); +CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_w, int img_h); CLIP_API int32_t clip_get_image_size (const struct clip_ctx * ctx); CLIP_API int32_t clip_get_patch_size (const struct clip_ctx * ctx); @@ -59,18 +59,29 @@ CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx); CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx); CLIP_API size_t get_clip_image_grid_size(const struct clip_ctx * ctx); -CLIP_API int clip_n_patches (const struct clip_ctx * ctx); -CLIP_API int clip_n_patches_by_img (const struct clip_ctx * ctx, struct clip_image_f32 * img); -CLIP_API int clip_n_mmproj_embd (const struct clip_ctx * ctx); +GGML_DEPRECATED(CLIP_API int clip_n_patches(const struct clip_ctx * ctx), + "use clip_n_output_tokens instead"); +GGML_DEPRECATED(CLIP_API int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * img), + "use clip_n_output_tokens instead"); + +CLIP_API int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img); + +// for M-RoPE, this will be the number of token positions in X and Y directions +// for other models, X will be the total number of tokens and Y will be 1 +CLIP_API int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img); +CLIP_API int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 * img); + +// this should be equal to the embedding dimension of the text model +CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx); CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip); CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size); CLIP_API struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip); -CLIP_API struct clip_image_size * clip_image_size_init(); -CLIP_API struct clip_image_u8 * clip_image_u8_init (); -CLIP_API struct clip_image_f32 * clip_image_f32_init(); -CLIP_API struct clip_image_f32_batch * clip_image_f32_batch_init(); // only used by libllava +CLIP_API struct clip_image_size * clip_image_size_init(void); +CLIP_API struct clip_image_u8 * clip_image_u8_init (void); +CLIP_API struct clip_image_f32 * clip_image_f32_init(void); +CLIP_API struct clip_image_f32_batch * clip_image_f32_batch_init(void); // only used by libllava // nx, ny are the output image dimensions CLIP_API unsigned char * clip_image_u8_get_data(struct clip_image_u8 * img, uint32_t * nx, uint32_t * ny); @@ -114,8 +125,6 @@ CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx); CLIP_API bool clip_is_llava(const struct clip_ctx * ctx); CLIP_API bool clip_is_gemma3(const struct clip_ctx * ctx); -CLIP_API int get_deepest_feature_layer(const struct clip_ctx * ctx); - CLIP_API bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec); diff --git a/examples/llava/convert_image_encoder_to_gguf.py b/tools/mtmd/convert_image_encoder_to_gguf.py similarity index 100% rename from examples/llava/convert_image_encoder_to_gguf.py rename to tools/mtmd/convert_image_encoder_to_gguf.py diff --git a/examples/llava/deprecation-warning.cpp b/tools/mtmd/deprecation-warning.cpp similarity index 100% rename from examples/llava/deprecation-warning.cpp rename to tools/mtmd/deprecation-warning.cpp diff --git a/examples/llava/glmedge-convert-image-encoder-to-gguf.py b/tools/mtmd/glmedge-convert-image-encoder-to-gguf.py similarity index 100% rename from examples/llava/glmedge-convert-image-encoder-to-gguf.py rename to tools/mtmd/glmedge-convert-image-encoder-to-gguf.py diff --git a/examples/llava/glmedge-surgery.py b/tools/mtmd/glmedge-surgery.py similarity index 100% rename from examples/llava/glmedge-surgery.py rename to tools/mtmd/glmedge-surgery.py diff --git a/examples/llava/llava.cpp b/tools/mtmd/llava.cpp similarity index 96% rename from examples/llava/llava.cpp rename to tools/mtmd/llava.cpp index 03a22cbb4c..ebef8b3c1e 100644 --- a/examples/llava/llava.cpp +++ b/tools/mtmd/llava.cpp @@ -2,6 +2,7 @@ #include "llava.h" #include "llama.h" +#include "ggml-cpp.h" #include #include @@ -112,7 +113,7 @@ static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair< } // Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out) -static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out) { +static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out, clip_image_f32 * img_input) { struct { struct ggml_context * ctx; } model; @@ -175,7 +176,7 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector model.ctx = ggml_init(params); - struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_patches(ctx_clip), num_images - 1); // example: 4096 x 576 x 4 + struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_output_tokens(ctx_clip, img_input), num_images - 1); // example: 4096 x 576 x 4 // ggml_tensor_printf(image_features,"image_features",__LINE__,false,false); // fill it with the image embeddings, ignoring the base for (size_t i = 1; i < num_images; i++) { @@ -209,13 +210,17 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector struct ggml_tensor *flatten = ggml_view_2d(model.ctx, permuted_cont, clip_n_mmproj_embd(ctx_clip), num_patches_height * num_patches_width * num_patches_per_side * num_patches_per_side, size_ele * clip_n_mmproj_embd(ctx_clip), 0); // ggml_tensor_printf(flatten,"flatten",__LINE__,false,false); ggml_build_forward_expand(gf, flatten); - ggml_graph_compute_with_ctx(model.ctx, gf, 1); + + ggml_backend_ptr backend { ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr) }; + GGML_ASSERT(backend != nullptr && "failed to initialize CPU backend"); + ggml_backend_graph_compute(backend.get(), gf); + struct ggml_tensor* result = ggml_graph_node(gf, -1); memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context // append without newline tokens (default behavior in llava_arch when not using unpad ): - memcpy(image_embd_out + clip_n_patches(ctx_clip) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches - *n_img_pos_out = static_cast(result->ne[1]+clip_n_patches(ctx_clip)); + memcpy(image_embd_out + clip_n_output_tokens(ctx_clip, img_input) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches + *n_img_pos_out = static_cast(result->ne[1]+clip_n_output_tokens(ctx_clip, img_input)); // Debug: Test single segments // Current findings: sending base image, sending a segment embedding all works similar to python @@ -313,7 +318,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip), image_embd_v[i], clip_embd_nbytes_by_img(ctx_clip, nx, ny)); - n_img_pos_out += clip_n_patches_by_img(ctx_clip, img_res); + n_img_pos_out += clip_n_output_tokens(ctx_clip, img_res); } *n_img_pos = n_img_pos_out; for (size_t i = 0; i < image_embd_v.size(); i++) { @@ -342,8 +347,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli } else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) { // flat / default llava-1.5 type embedding - *n_img_pos = clip_n_patches(ctx_clip); clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0); + *n_img_pos = clip_n_output_tokens(ctx_clip, img_res); bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd); // image_embd shape is 576 x 4096 if (!encoded) { LOG_ERR("Unable to encode image\n"); @@ -381,7 +386,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size); int n_img_pos_out; - clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out); + clip_image_f32 * img_input = clip_image_f32_get_img(img_res_v.get(), 0); + clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out, img_input); *n_img_pos = n_img_pos_out; for (size_t i = 0; i < image_embd_v.size(); i++) { diff --git a/examples/llava/llava.h b/tools/mtmd/llava.h similarity index 100% rename from examples/llava/llava.h rename to tools/mtmd/llava.h diff --git a/examples/llava/llava_surgery.py b/tools/mtmd/llava_surgery.py similarity index 100% rename from examples/llava/llava_surgery.py rename to tools/mtmd/llava_surgery.py diff --git a/examples/llava/llava_surgery_v2.py b/tools/mtmd/llava_surgery_v2.py similarity index 100% rename from examples/llava/llava_surgery_v2.py rename to tools/mtmd/llava_surgery_v2.py diff --git a/examples/llava/minicpmv-convert-image-encoder-to-gguf.py b/tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py similarity index 100% rename from examples/llava/minicpmv-convert-image-encoder-to-gguf.py rename to tools/mtmd/minicpmv-convert-image-encoder-to-gguf.py diff --git a/examples/llava/minicpmv-surgery.py b/tools/mtmd/minicpmv-surgery.py similarity index 100% rename from examples/llava/minicpmv-surgery.py rename to tools/mtmd/minicpmv-surgery.py diff --git a/examples/llava/mtmd-cli.cpp b/tools/mtmd/mtmd-cli.cpp similarity index 76% rename from examples/llava/mtmd-cli.cpp rename to tools/mtmd/mtmd-cli.cpp index 250e8c9a9e..4977d5480b 100644 --- a/examples/llava/mtmd-cli.cpp +++ b/tools/mtmd/mtmd-cli.cpp @@ -63,7 +63,7 @@ static void sigint_handler(int signo) { #endif struct mtmd_cli_context { - mtmd_context_ptr ctx_vision; + mtmd::context_ptr ctx_vision; common_init_result llama_init; llama_model * model; @@ -72,6 +72,8 @@ struct mtmd_cli_context { llama_batch batch; int n_batch; + mtmd::bitmaps bitmaps; + // note: we know that gemma3 template is "linear", meaning each turn is completely separated to another // so here we don't need to keep track of chat history common_chat_templates_ptr tmpls; @@ -90,10 +92,15 @@ struct mtmd_cli_context { batch = llama_batch_init(params.n_batch, 0, 1); n_batch = params.n_batch; + if (!model || !lctx) { + exit(1); + } + if (!llama_model_chat_template(model, nullptr) && params.chat_template.empty()) { LOG_ERR("Model does not have chat template.\n"); LOG_ERR(" For old llava models, you may need to use '--chat-template vicuna'\n"); LOG_ERR(" For MobileVLM models, use '--chat-template deepseek'\n"); + LOG_ERR(" For Mistral Small 3.1, use '--chat-template mistral-v7'\n"); exit(1); } @@ -112,12 +119,12 @@ struct mtmd_cli_context { void init_vision_context(common_params & params) { const char * clip_path = params.mmproj.path.c_str(); - ctx_vision.reset(mtmd_init_from_file(clip_path, model, mtmd_context_params{ - /* use_gpu */ params.mmproj_use_gpu, - /* timings */ true, - /* n_threads */ params.cpuparams.n_threads, - /* verbosity */ params.verbosity > 0 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_INFO, - })); + mtmd_context_params mparams = mtmd_context_params_default(); + mparams.use_gpu = params.mmproj_use_gpu; + mparams.print_timings = true; + mparams.n_threads = params.cpuparams.n_threads; + mparams.verbosity = params.verbosity > 0 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_INFO; + ctx_vision.reset(mtmd_init_from_file(clip_path, model, mparams)); if (!ctx_vision.get()) { LOG_ERR("Failed to load vision model from %s\n", clip_path); exit(1); @@ -134,38 +141,14 @@ struct mtmd_cli_context { antiprompt_tokens.begin() ); } -}; -struct decode_embd_batch { - std::vector pos; - std::vector n_seq_id; - std::vector seq_id_0; - std::vector seq_ids; - std::vector logits; - llama_batch batch; - decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) { - pos .resize(n_tokens); - n_seq_id.resize(n_tokens); - seq_ids .resize(n_tokens + 1); - logits .resize(n_tokens); - seq_id_0.resize(1); - seq_id_0[0] = seq_id; - seq_ids [n_tokens] = nullptr; - batch = { - /*n_tokens =*/ n_tokens, - /*tokens =*/ nullptr, - /*embd =*/ embd, - /*pos =*/ pos.data(), - /*n_seq_id =*/ n_seq_id.data(), - /*seq_id =*/ seq_ids.data(), - /*logits =*/ logits.data(), - }; - for (int i = 0; i < n_tokens; i++) { - batch.pos [i] = pos_0 + i; - batch.n_seq_id[i] = 1; - batch.seq_id [i] = seq_id_0.data(); - batch.logits [i] = false; + bool load_image(const std::string & fname) { + mtmd::bitmap bmp(mtmd_helper_bitmap_init_from_file(fname.c_str())); + if (!bmp.ptr) { + return false; } + bitmaps.entries.push_back(std::move(bmp)); + return true; } }; @@ -173,7 +156,7 @@ static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int llama_tokens generated_tokens; for (int i = 0; i < n_predict; i++) { if (i > n_predict || !g_is_generating || g_is_interrupted) { - printf("\n"); + LOG("\n"); break; } @@ -182,15 +165,15 @@ static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int common_sampler_accept(smpl, token_id, true); if (llama_vocab_is_eog(ctx.vocab, token_id) || ctx.check_antiprompt(generated_tokens)) { - printf("\n"); + LOG("\n"); break; // end of generation } - printf("%s", common_token_to_piece(ctx.lctx, token_id).c_str()); + LOG("%s", common_token_to_piece(ctx.lctx, token_id).c_str()); fflush(stdout); if (g_is_interrupted) { - printf("\n"); + LOG("\n"); break; } @@ -205,9 +188,7 @@ static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int return 0; } -static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, std::vector & images_fname, bool add_bos = false) { - std::vector bitmaps; - +static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, bool add_bos = false) { common_chat_templates_inputs tmpl_inputs; tmpl_inputs.messages = {msg}; tmpl_inputs.add_generation_prompt = true; @@ -215,35 +196,43 @@ static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, std::vect auto formatted_chat = common_chat_templates_apply(ctx.tmpls.get(), tmpl_inputs); LOG_DBG("formatted_chat.prompt: %s\n", formatted_chat.prompt.c_str()); - for (auto & fname : images_fname) { - mtmd_bitmap bitmap; - if (mtmd_helper_bitmap_init_from_file(fname.c_str(), bitmap)) { - LOG_ERR("Unable to load image %s\n", fname.c_str()); - return 2; // image not found - } - bitmaps.push_back(std::move(bitmap)); - } - mtmd_input_text text; - text.text = formatted_chat.prompt; + text.text = formatted_chat.prompt.c_str(); text.add_special = add_bos; text.parse_special = true; - mtmd_input_chunks chunks; if (g_is_interrupted) return 0; - int32_t res = mtmd_tokenize(ctx.ctx_vision.get(), chunks, text, bitmaps); + mtmd::input_chunks chunks(mtmd_input_chunks_init()); + auto bitmaps_c_ptr = ctx.bitmaps.c_ptr(); + int32_t res = mtmd_tokenize(ctx.ctx_vision.get(), + chunks.ptr.get(), // output + &text, // text + bitmaps_c_ptr.data(), + bitmaps_c_ptr.size()); if (res != 0) { LOG_ERR("Unable to tokenize prompt, res = %d\n", res); return 1; } - if (mtmd_helper_eval(ctx.ctx_vision.get(), ctx.lctx, chunks, ctx.n_past, 0, ctx.n_batch)) { + ctx.bitmaps.entries.clear(); + + llama_pos new_n_past; + if (mtmd_helper_eval_chunks(ctx.ctx_vision.get(), + ctx.lctx, // lctx + chunks.ptr.get(), // chunks + ctx.n_past, // n_past + 0, // seq_id + ctx.n_batch, // n_batch + true, // logits_last + &new_n_past)) { LOG_ERR("Unable to eval prompt\n"); return 1; } - ctx.n_past += mtmd_helper_get_n_tokens(chunks); + ctx.n_past = new_n_past; + + LOG("\n"); return 0; } @@ -267,14 +256,14 @@ int main(int argc, char ** argv) { } mtmd_cli_context ctx(params); - printf("%s: %s\n", __func__, params.model.path.c_str()); + LOG("%s: loading model: %s\n", __func__, params.model.path.c_str()); bool is_single_turn = !params.prompt.empty() && !params.image.empty(); struct common_sampler * smpl = common_sampler_init(ctx.model, params.sampling); int n_predict = params.n_predict < 0 ? INT_MAX : params.n_predict; - // ctrl+C handling + // Ctrl+C handling { #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) struct sigaction sigint_action; @@ -300,7 +289,12 @@ int main(int argc, char ** argv) { common_chat_msg msg; msg.role = "user"; msg.content = params.prompt; - if (eval_message(ctx, msg, params.image, true)) { + for (const auto & image : params.image) { + if (!ctx.load_image(image)) { + return 1; // error is already printed by libmtmd + } + } + if (eval_message(ctx, msg, true)) { return 1; } if (!g_is_interrupted && generate_response(ctx, smpl, n_predict)) { @@ -315,7 +309,6 @@ int main(int argc, char ** argv) { LOG("\n"); bool is_first_msg = true; - std::vector images_fname; std::string content; while (!g_is_interrupted) { @@ -340,10 +333,17 @@ int main(int argc, char ** argv) { continue; } g_is_generating = true; - if (line.find("/image") == 0) { + if (line == "/image" || line.find("/image ") == 0) { + if (line.size() < 8) { + LOG_ERR("ERR: Missing image filename\n"); + continue; + } std::string image = line.substr(7); - images_fname.push_back(string_strip(image)); - content += "<__image__>"; + if (ctx.load_image(image)) { + LOG("Image %s loaded\n", image.c_str()); + content += "<__image__>"; + } + // else, error is already printed by libmtmd continue; } else { content += line; @@ -351,26 +351,20 @@ int main(int argc, char ** argv) { common_chat_msg msg; msg.role = "user"; msg.content = content; - int ret = eval_message(ctx, msg, images_fname, is_first_msg); - if (g_is_interrupted) break; - if (ret == 2) { - // non-fatal error - images_fname.clear(); - content.clear(); - continue; - } + int ret = eval_message(ctx, msg, is_first_msg); if (ret) { return 1; } + if (g_is_interrupted) break; if (generate_response(ctx, smpl, n_predict)) { return 1; } - images_fname.clear(); content.clear(); is_first_msg = false; } } if (g_is_interrupted) LOG("\nInterrupted by user\n"); + LOG("\n\n"); llama_perf_context_print(ctx.lctx); return g_is_interrupted ? 130 : 0; } diff --git a/tools/mtmd/mtmd-helper.cpp b/tools/mtmd/mtmd-helper.cpp new file mode 100644 index 0000000000..7a3288672d --- /dev/null +++ b/tools/mtmd/mtmd-helper.cpp @@ -0,0 +1,310 @@ +#include "mtmd.h" +#include "llama.h" + +#include +#include +#include + +#define LOG_INF(...) fprintf(stdout, __VA_ARGS__) +#define LOG_ERR(...) fprintf(stderr, __VA_ARGS__) + +size_t mtmd_helper_get_n_tokens(const mtmd_input_chunks * chunks) { + size_t n_tokens = 0; + for (size_t i = 0; i < mtmd_input_chunks_size(chunks); i++) { + auto chunk = mtmd_input_chunks_get(chunks, i); + auto chunk_type = mtmd_input_chunk_get_type(chunk); + if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) { + size_t n_tokens_text; + mtmd_input_chunk_get_tokens_text(chunk, &n_tokens_text); + n_tokens += n_tokens_text; + } else if (chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE) { + auto tokens_image = mtmd_input_chunk_get_tokens_image(chunk); + n_tokens += mtmd_image_tokens_get_n_tokens(tokens_image); + } else { + GGML_ASSERT(false && "chunk type not supported"); + } + } + return n_tokens; +} + +llama_pos mtmd_helper_get_n_pos(const mtmd_input_chunks * chunks) { + llama_pos n_pos = 0; + for (size_t i = 0; i < mtmd_input_chunks_size(chunks); i++) { + auto chunk = mtmd_input_chunks_get(chunks, i); + auto chunk_type = mtmd_input_chunk_get_type(chunk); + if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) { + size_t n_tokens_text; + mtmd_input_chunk_get_tokens_text(chunk, &n_tokens_text); + n_pos += n_tokens_text; + } else if (chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE) { + auto tokens_image = mtmd_input_chunk_get_tokens_image(chunk); + n_pos += mtmd_image_tokens_get_n_pos(tokens_image); + } else { + GGML_ASSERT(false && "chunk type not supported"); + } + } + return n_pos; +} + +// helper struct to make working with embd batch easier +// note: this will be removed after llama_batch_ext refactoring +struct decode_embd_batch { + int n_pos_per_embd; + int n_mmproj_embd; + std::vector pos; + std::vector pos_view; // used by mrope + std::vector n_seq_id; + std::vector seq_id_0; + std::vector seq_ids; + std::vector logits; + llama_batch batch; + decode_embd_batch(float * embd, int32_t n_tokens, int n_pos_per_embd, int n_mmproj_embd) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) { + pos .resize(n_tokens * n_pos_per_embd); + n_seq_id.resize(n_tokens); + seq_ids .resize(n_tokens + 1); + logits .resize(n_tokens); + seq_id_0.resize(1); + seq_ids [n_tokens] = nullptr; + batch = { + /*n_tokens =*/ n_tokens, + /*tokens =*/ nullptr, + /*embd =*/ embd, + /*pos =*/ pos.data(), + /*n_seq_id =*/ n_seq_id.data(), + /*seq_id =*/ seq_ids.data(), + /*logits =*/ logits.data(), + }; + } + + void set_position_normal(llama_pos pos_0, llama_seq_id seq_id) { + seq_id_0[0] = seq_id; + for (int i = 0; i < batch.n_tokens; i++) { + batch.pos [i] = pos_0 + i; + batch.n_seq_id[i] = 1; + batch.seq_id [i] = seq_id_0.data(); + batch.logits [i] = false; + } + } + + void set_position_mrope(llama_pos pos_0, int nx, int ny, llama_seq_id seq_id) { + GGML_ASSERT(n_pos_per_embd == 4); + seq_id_0[0] = seq_id; + for (int y = 0; y < ny; y++) { + for (int x = 0; x < nx; x++) { + int i = y * nx + x; + pos[i ] = pos_0; + pos[i + batch.n_tokens ] = pos_0 + y; + pos[i + batch.n_tokens * 2] = pos_0 + x; + pos[i + batch.n_tokens * 3] = 0; // last pos dim is unused + } + } + for (int i = 0; i < batch.n_tokens; i++) { + batch.n_seq_id[i] = 1; + batch.seq_id [i] = seq_id_0.data(); + batch.logits [i] = false; + } + } + + llama_batch get_view(int offset, int n_tokens) { + llama_pos * pos_ptr; + pos_view.clear(); + pos_view.reserve(n_tokens * n_pos_per_embd); + if (n_pos_per_embd > 1) { + // mrope + // for example, with layout of src: 1234...1234...1234...1234... + // offset 2 will give us dst: 34...34...34...34... + for (int i = 0; i < n_pos_per_embd; i++) { + // assume n_tokens is less than or equal to batch.n_tokens + // batch.n_tokens is number of **total** tokens + // n_tokens is number of viewed token + size_t src_idx = i * batch.n_tokens + offset; + pos_view.insert(pos_view.end(), + pos.data() + src_idx, + pos.data() + src_idx + n_tokens); + } + pos_ptr = pos_view.data(); + } else { + // normal + pos_ptr = pos.data() + offset; + } + return { + /*n_tokens =*/ n_tokens, + /*tokens =*/ nullptr, + /*embd =*/ batch.embd + offset * n_mmproj_embd, + /*pos =*/ pos_ptr, + /*n_seq_id =*/ batch.n_seq_id + offset, + /*seq_id =*/ batch.seq_id + offset, + /*logits =*/ batch.logits + offset, + }; + } +}; + +// Helper function for decoding an image whose embeddings have already been calculated +int32_t mtmd_helper_decode_image_chunk( + mtmd_context * ctx, + struct llama_context * lctx, + const mtmd_input_chunk * chunk, + float * encoded_embd, + llama_pos n_past, + llama_seq_id seq_id, + int32_t n_batch, + llama_pos * new_n_past) { + if (mtmd_input_chunk_get_type(chunk) != MTMD_INPUT_CHUNK_TYPE_IMAGE) { + LOG_ERR("failed to decode image chunk: input chunk not of image type\n"); + return -1; + } + const auto image_tokens = mtmd_input_chunk_get_tokens_image(chunk); + if (!image_tokens) { + LOG_ERR("failed to decode image chunk: image tokens are null\n"); + return -1; + } + + const llama_model * model = llama_get_model(lctx); + int n_mmproj_embd = llama_model_n_embd(model); + int n_pos_per_embd = mtmd_decode_use_mrope(ctx) ? 4 : 1; + + int32_t n_tokens = mtmd_image_tokens_get_n_tokens(image_tokens); + int32_t i_batch = 0; + int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch; + decode_embd_batch batch_embd(encoded_embd, n_tokens, n_pos_per_embd, n_mmproj_embd); + + const int nx = mtmd_image_tokens_get_nx(image_tokens); + const int ny = mtmd_image_tokens_get_ny(image_tokens); + + if (mtmd_decode_use_mrope(ctx)) { + batch_embd.set_position_mrope(n_past, nx, ny, seq_id); + } else { + batch_embd.set_position_normal(n_past, seq_id); + } + + if (mtmd_decode_use_non_causal(ctx)) { + llama_set_causal_attn(lctx, false); + // TODO @ngxson : need to make sure only one image is processed at a time, and n_ubatch must be enough to hold the image + } + + while (i_batch < n_img_batches) { // split into batches + int pos_offset = i_batch*n_batch; + int n_tokens_batch = std::min(n_batch, n_tokens - pos_offset); + llama_batch batch_embd_view = batch_embd.get_view(pos_offset, n_tokens_batch); + + LOG_INF("decoding image batch %d/%d, n_tokens_batch = %d\n", i_batch+1, n_img_batches, n_tokens_batch); + + int64_t t1 = ggml_time_ms(); + int32_t ret = llama_decode(lctx, batch_embd_view); + if (ret != 0) { + LOG_ERR("failed to decode image\n"); + llama_set_causal_attn(lctx, true); // restore causal attn + return ret; + } + + LOG_INF("image decoded (batch %d/%d) in %" PRId64 " ms\n", i_batch+1, n_img_batches, ggml_time_ms() - t1); + + i_batch++; + } + + n_past += mtmd_image_tokens_get_n_pos(image_tokens); + *new_n_past = n_past; + + if (mtmd_decode_use_non_causal(ctx)) { + llama_set_causal_attn(lctx, true); + } + return 0; +} + +int32_t mtmd_helper_eval_chunk_single(mtmd_context * ctx, + struct llama_context * lctx, + const mtmd_input_chunk * chunk, + llama_pos n_past, + llama_seq_id seq_id, + int32_t n_batch, + bool logits_last, + llama_pos * new_n_past) { + int32_t ret; + llama_batch text_batch = llama_batch_init(n_batch, 0, 1); + auto chunk_type = mtmd_input_chunk_get_type(chunk); + + if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) { + size_t n_tokens; + const auto tokens = mtmd_input_chunk_get_tokens_text(chunk, &n_tokens); + // LOG_INF("decoding text chunk, n_tokens = %zu\n", n_tokens); + size_t i = 0; + while (i < n_tokens) { // split into batches + text_batch.n_tokens = 0; // clear the batch + for (; i < n_tokens && text_batch.n_tokens < n_batch; i++) { + text_batch.n_tokens++; + text_batch.token [i] = tokens[i]; + text_batch.pos [i] = n_past++; + text_batch.n_seq_id[i] = 1; + text_batch.seq_id [i][0] = seq_id; + text_batch.logits [i] = false; + } + bool is_last_token = (i == n_tokens); + if (logits_last && is_last_token) { + text_batch.logits[text_batch.n_tokens - 1] = true; + } + ret = llama_decode(lctx, text_batch); + if (ret != 0) { + LOG_ERR("failed to decode text\n"); + llama_batch_free(text_batch); + return ret; + } + *new_n_past += text_batch.n_tokens; + } + + } else if (chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE) { + const auto image_tokens = mtmd_input_chunk_get_tokens_image(chunk); + int64_t t0 = ggml_time_ms(); + + LOG_INF("encoding image or slice...\n"); + + ret = mtmd_encode(ctx, image_tokens); + if (ret != 0) { + LOG_ERR("failed to encode image\n"); + llama_batch_free(text_batch); + return ret; + } + + LOG_INF("image/slice encoded in %" PRId64 " ms\n", ggml_time_ms() - t0); + + float * embd = mtmd_get_output_embd(ctx); + ret = mtmd_helper_decode_image_chunk(ctx, lctx, chunk, embd, n_past, seq_id, n_batch, new_n_past); + if (ret != 0) { + LOG_ERR("failed to decode image\n"); + llama_batch_free(text_batch); + return ret; + } + } else { + GGML_ABORT("chunk type not supported"); + } + + return 0; +} + +int32_t mtmd_helper_eval_chunks(mtmd_context * ctx, + struct llama_context * lctx, + const mtmd_input_chunks * chunks, + llama_pos n_past, + llama_seq_id seq_id, + int32_t n_batch, + bool logits_last, + llama_pos * new_n_past) { + size_t n_chunks = mtmd_input_chunks_size(chunks); + if (n_chunks == 0) { + LOG_ERR("no chunks to eval\n"); + return 0; + } + + for (size_t i = 0; i < n_chunks; i++) { + bool chunk_logits_last = (i == n_chunks - 1) && logits_last; + auto chunk = mtmd_input_chunks_get(chunks, i); + + int32_t res = mtmd_helper_eval_chunk_single(ctx, lctx, chunk, n_past, seq_id, n_batch, chunk_logits_last, &n_past); + if (res != 0) { + LOG_ERR("failed to eval chunk %zu\n", i); + return res; + } + *new_n_past = n_past; + } + + return 0; +} diff --git a/examples/llava/mtmd.cpp b/tools/mtmd/mtmd.cpp similarity index 63% rename from examples/llava/mtmd.cpp rename to tools/mtmd/mtmd.cpp index a994ef0166..2a852d9c19 100644 --- a/examples/llava/mtmd.cpp +++ b/tools/mtmd/mtmd.cpp @@ -12,6 +12,30 @@ #include #include +// represents raw image data, layout is RGBRGBRGB... +// length of data must be nx * ny * 3 +struct mtmd_bitmap { + uint32_t nx; + uint32_t ny; + std::vector data; + std::string id; // optional user-defined id, for ex: can be set to image hash, useful for KV cache tracking +}; + +struct mtmd_image_tokens_deleter { + void operator()(mtmd_image_tokens * val); // forward declaration +}; +using mtmd_image_tokens_ptr = std::unique_ptr; + +struct mtmd_input_chunk { + mtmd_input_chunk_type type; + std::vector tokens_text; + mtmd_image_tokens_ptr tokens_image; +}; + +struct mtmd_input_chunks { + std::vector entries; +}; + // slice template, used by some llava-uhd models to correctly place the special tokens around image embeddings // models not having it (llava-1.6) will process embeddings without any special tokens in-between enum mtmd_slice_tmpl { @@ -21,6 +45,16 @@ enum mtmd_slice_tmpl { // TODO @ngxson : add support for idefics (SmolVLM) }; +mtmd_context_params mtmd_context_params_default() { + mtmd_context_params params; + params.use_gpu = true; + params.print_timings = true; + params.n_threads = 4; + params.verbosity = GGML_LOG_LEVEL_INFO; + params.image_marker = MTMD_DEFAULT_IMAGE_MARKER; + return params; +} + struct mtmd_context { struct clip_ctx * ctx_clip; const struct llama_model * text_model; @@ -40,11 +74,14 @@ struct mtmd_context { llama_token tok_sli_img_end = LLAMA_TOKEN_NULL; // single slice llama_token tok_row_end = LLAMA_TOKEN_NULL; // end of row + bool use_mrope = false; // for Qwen2VL, we need to use M-RoPE + // TODO @ngxson : add timings mtmd_context(const char * mmproj_fname, const llama_model * text_model, const mtmd_context_params & ctx_params) : + text_model (text_model), print_timings(ctx_params.print_timings), n_threads (ctx_params.n_threads), image_marker (ctx_params.image_marker) @@ -56,9 +93,8 @@ struct mtmd_context { if (!ctx_clip) { throw std::runtime_error(string_format("Failed to load CLIP model from %s\n", mmproj_fname)); } - this->text_model = text_model; - GGML_ASSERT(!clip_is_qwen2vl(ctx_clip) && "Qwen2VL model is not supported yet, use llama-qwen2vl-cli instead"); + use_mrope = clip_is_qwen2vl(ctx_clip); int minicpmv_version = clip_is_minicpmv(ctx_clip); if (minicpmv_version == 2) { @@ -126,9 +162,20 @@ struct mtmd_image_tokens_data { struct mtmd_image_tokens { uint32_t nx; // number of tokens in x direction uint32_t ny; // number of tokens in y direction + bool use_mrope_pos = false; // use M-RoPE position counting (the whole image is 1 temporal position) uint32_t n_tokens() const { return nx * ny; } clip_image_f32_batch batch_f32; // preprocessed image patches std::string id; // optional user-defined ID, useful for KV cache tracking + + mtmd_image_tokens clone() { + return mtmd_image_tokens{ + nx, + ny, + use_mrope_pos, + batch_f32.clone(), + id + }; + } }; mtmd_context * mtmd_init_from_file(const char * mmproj_fname, @@ -169,12 +216,13 @@ static std::vector mtmd_tokenize_text_internal( } int32_t mtmd_tokenize(mtmd_context * ctx, - std::vector & output, - const mtmd_input_text & text, - const std::vector & bitmaps) { + mtmd_input_chunks * output, + const mtmd_input_text * text, + const mtmd_bitmap ** bitmaps, + size_t n_bitmaps) { auto vocab = llama_model_get_vocab(ctx->text_model); - std::string prompt_modified(text.text); + std::string prompt_modified(text->text); std::string marker_modified(ctx->image_marker); projector_type proj_type = clip_get_projector_type(ctx->ctx_clip); @@ -186,11 +234,6 @@ int32_t mtmd_tokenize(mtmd_context * ctx, marker_modified = "" + ctx->image_marker + ""; string_replace_all(prompt_modified, ctx->image_marker, marker_modified); - } else if (proj_type == PROJECTOR_TYPE_GLM_EDGE) { - // <|begin_of_image|> ... (image embeddings) ... <|end_of_image|> - marker_modified = "<|begin_of_image|>" + ctx->image_marker + "<|end_of_image|>"; - string_replace_all(prompt_modified, ctx->image_marker, marker_modified); - } else if (proj_type == PROJECTOR_TYPE_IDEFICS3) { // https://github.com/huggingface/transformers/blob/a42ba80fa520c784c8f11a973ca9034e5f859b79/src/transformers/models/idefics3/processing_idefics3.py#L192-L215 marker_modified = "" + ctx->image_marker + ""; @@ -202,14 +245,26 @@ int32_t mtmd_tokenize(mtmd_context * ctx, string_replace_all(prompt_modified, ctx->image_marker, marker_modified); } - // llava-1.5, llava-1.6, Yi-VL, Yi-34B, granite: don't need to add prefix and suffix - // for glm-edge, we don't need to add because the tokens are already in the returned embeddings + else if (proj_type == PROJECTOR_TYPE_QWEN2VL || proj_type == PROJECTOR_TYPE_QWEN25VL) { + // <|vision_start|> ... (image embeddings) ... <|vision_end|> + marker_modified = "<|vision_start|>" + ctx->image_marker + "<|vision_end|>"; + string_replace_all(prompt_modified, ctx->image_marker, marker_modified); - // TODO @ngxson : glm-edge : remove BOI / EOI tokens embeddings, decode them as normal tokens + } + + else if (proj_type == PROJECTOR_TYPE_INTERNVL) { + // ... (image embeddings) ... + marker_modified = "" + ctx->image_marker + ""; + string_replace_all(prompt_modified, ctx->image_marker, marker_modified); + + } + + // llava-1.5, llava-1.6, Yi-VL, Yi-34B, granite: don't need to add prefix and suffix + // for glm-edge, BOI and EOI token's embeddings are not present in the text model std::vector parts = string_split_str(prompt_modified, ctx->image_marker); - output.clear(); - output.reserve(parts.size()); + output->entries.clear(); + output->entries.reserve(parts.size()); size_t i_img = 0; @@ -220,7 +275,7 @@ int32_t mtmd_tokenize(mtmd_context * ctx, std::move(tokens), {}, }; - output.emplace_back(std::move(chunk)); + output->entries.emplace_back(std::move(chunk)); }; // utility for splitting batch of multiple images into chunks of batch having single images @@ -229,7 +284,7 @@ int32_t mtmd_tokenize(mtmd_context * ctx, for (auto & entry : batch_f32.entries) { mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens); - image_tokens->nx = clip_n_patches_by_img(ctx->ctx_clip, entry.get()); + image_tokens->nx = clip_n_output_tokens(ctx->ctx_clip, entry.get()); image_tokens->ny = 1; image_tokens->batch_f32.entries.push_back(std::move(entry)); image_tokens->id = id; @@ -246,9 +301,9 @@ int32_t mtmd_tokenize(mtmd_context * ctx, }; for (const auto & part : parts) { - //printf("tokenizing part: %s\n", part.c_str()); + // printf("tokenizing part: %s\n", part.c_str()); bool add_bos = &parts.front() == ∂ - auto tokens = mtmd_tokenize_text_internal(vocab, part, text.add_special && add_bos, text.parse_special); + auto tokens = mtmd_tokenize_text_internal(vocab, part, text->add_special && add_bos, text->parse_special); if (tokens.empty()) { continue; } @@ -257,22 +312,22 @@ int32_t mtmd_tokenize(mtmd_context * ctx, std::move(tokens), {}, }; - output.emplace_back(std::move(chunk)); + output->entries.emplace_back(std::move(chunk)); if (&parts.back() != &part) { // add image token to middle of 2 parts - if (i_img >= bitmaps.size()) { + if (i_img >= n_bitmaps) { LOG_ERR("%s: error: not enough images for %d parts\n", __func__, (int)parts.size()); return 1; } // convert mtmd_bitmap to clip_image_u8 clip_image_u8_ptr img_u8(clip_image_u8_init()); - img_u8->nx = bitmaps[i_img].nx; - img_u8->ny = bitmaps[i_img].ny; - img_u8->buf.resize(bitmaps[i_img].data.size()); - std::memcpy(img_u8->buf.data(), bitmaps[i_img].data.data(), img_u8->nx * img_u8->ny * 3); + img_u8->nx = bitmaps[i_img]->nx; + img_u8->ny = bitmaps[i_img]->ny; + img_u8->buf.resize(bitmaps[i_img]->data.size()); + std::memcpy(img_u8->buf.data(), bitmaps[i_img]->data.data(), img_u8->nx * img_u8->ny * 3); clip_image_size img_u8_size{img_u8->nx, img_u8->ny}; // preprocess image @@ -285,12 +340,12 @@ int32_t mtmd_tokenize(mtmd_context * ctx, if (ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_5 || ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_6) { // split batch into chunks of single images - auto chunks = split_batch_to_chunk(std::move(batch_f32), bitmaps[i_img].id); + auto chunks = split_batch_to_chunk(std::move(batch_f32), bitmaps[i_img]->id); GGML_ASSERT(chunks.size() > 0); // add overview image add_text_chunk({ctx->tok_ov_img_start}); - output.emplace_back(std::move(chunks.front())); + output->entries.emplace_back(std::move(chunks.front())); chunks.erase(chunks.begin()); add_text_chunk({ctx->tok_ov_img_end}); @@ -308,7 +363,7 @@ int32_t mtmd_tokenize(mtmd_context * ctx, if (ctx->tok_sli_img_start != LLAMA_TOKEN_NULL) { add_text_chunk({ctx->tok_sli_img_start}); } - output.emplace_back(std::move(chunks[y * n_col + x])); + output->entries.emplace_back(std::move(chunks[y * n_col + x])); if (ctx->tok_sli_img_end != LLAMA_TOKEN_NULL) { add_text_chunk({ctx->tok_sli_img_end}); } @@ -325,30 +380,33 @@ int32_t mtmd_tokenize(mtmd_context * ctx, } else { size_t n_tokens = 0; for (const auto & entry : batch_f32.entries) { - n_tokens += clip_n_patches_by_img(ctx->ctx_clip, entry.get()); + n_tokens += clip_n_output_tokens(ctx->ctx_clip, entry.get()); } mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens); - image_tokens->nx = n_tokens; - image_tokens->ny = 1; // TODO + if (ctx->use_mrope) { + // for Qwen2VL, we need this information for M-RoPE decoding positions + image_tokens->nx = clip_n_output_tokens_x(ctx->ctx_clip, batch_f32.entries[0].get()); + image_tokens->ny = clip_n_output_tokens_y(ctx->ctx_clip, batch_f32.entries[0].get()); + image_tokens->use_mrope_pos = true; + } else { + // other models, we only need the total number of tokens + image_tokens->nx = n_tokens; + image_tokens->ny = 1; + } image_tokens->batch_f32 = std::move(batch_f32); - image_tokens->id = bitmaps[i_img].id; // optional + image_tokens->id = bitmaps[i_img]->id; // optional LOG_DBG("image_tokens->nx = %d\n", image_tokens->nx); LOG_DBG("image_tokens->ny = %d\n", image_tokens->ny); LOG_DBG("batch_f32 size = %d\n", (int)image_tokens->batch_f32.entries.size()); - if (clip_is_glm(ctx->ctx_clip)) { - // glm-edge - image_tokens->nx += 2; // add 2 for the begin_of_image and end_of_image token embeddings - } - mtmd_input_chunk chunk{ MTMD_INPUT_CHUNK_TYPE_IMAGE, {}, std::move(image_tokens), }; - output.emplace_back(std::move(chunk)); + output->entries.emplace_back(std::move(chunk)); } i_img++; // move to next image @@ -358,28 +416,12 @@ int32_t mtmd_tokenize(mtmd_context * ctx, return 0; } -void mtmd_image_tokens_free(mtmd_image_tokens * image_tokens) { +static void mtmd_image_tokens_free(mtmd_image_tokens * image_tokens) { if (image_tokens) { delete image_tokens; } } -size_t mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens) { - return image_tokens->n_tokens(); -} - -size_t mtmd_image_tokens_get_nx(const mtmd_image_tokens * image_tokens) { - return image_tokens->nx; -} - -size_t mtmd_image_tokens_get_ny(const mtmd_image_tokens * image_tokens) { - return image_tokens->ny; -} - -std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens) { - return image_tokens->id; -} - int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) { int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip); ctx->image_embd_v.resize(image_tokens->n_tokens() * n_mmproj_embd); @@ -397,7 +439,7 @@ int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) // TODO @ngxson : llava does not support batched encoding ; this should be fixed inside clip_image_batch_encode() const auto & entries = image_tokens->batch_f32.entries; for (size_t i = 0; i < entries.size(); i++) { - int n_tokens_per_image = clip_n_patches_by_img(ctx->ctx_clip, entries[i].get()); + int n_tokens_per_image = clip_n_output_tokens(ctx->ctx_clip, entries[i].get()); ok = clip_image_encode( ctx->ctx_clip, ctx->n_threads, @@ -419,182 +461,6 @@ float * mtmd_get_output_embd(mtmd_context * ctx) { return ctx->image_embd_v.data(); } -size_t mtmd_helper_get_n_tokens(mtmd_input_chunks & chunks) { - size_t n_tokens = 0; - for (auto & chunk : chunks) { - if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) { - n_tokens += chunk.tokens_text.size(); - } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) { - n_tokens += chunk.tokens_image->n_tokens(); - } else { - GGML_ASSERT(false && "chunk type not supported"); - } - } - return n_tokens; -} - -// helper struct to make working with embd batch easier -// note: this will be removed after llama_batch_ext refactoring -struct decode_embd_batch { - std::vector pos; - std::vector n_seq_id; - std::vector seq_id_0; - std::vector seq_ids; - std::vector logits; - llama_batch batch; - decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) { - pos .resize(n_tokens); - n_seq_id.resize(n_tokens); - seq_ids .resize(n_tokens + 1); - logits .resize(n_tokens); - seq_id_0.resize(1); - seq_id_0[0] = seq_id; - seq_ids [n_tokens] = nullptr; - batch = { - /*n_tokens =*/ n_tokens, - /*tokens =*/ nullptr, - /*embd =*/ embd, - /*pos =*/ pos.data(), - /*n_seq_id =*/ n_seq_id.data(), - /*seq_id =*/ seq_ids.data(), - /*logits =*/ logits.data(), - }; - for (int i = 0; i < n_tokens; i++) { - batch.pos [i] = pos_0 + i; - batch.n_seq_id[i] = 1; - batch.seq_id [i] = seq_id_0.data(); - batch.logits [i] = false; - } - } -}; - -int32_t mtmd_helper_eval(mtmd_context * ctx, - llama_context * lctx, - mtmd_input_chunks & chunks, - llama_pos pos0, - llama_seq_id seq_id, - int32_t n_batch) { - int32_t ret; - llama_pos n_past = pos0; - llama_batch text_batch = llama_batch_init(n_batch, 0, 1); - int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip); - - for (auto & chunk : chunks) { - bool is_last = &chunk == &chunks.back(); - if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) { - text_batch.n_tokens = chunk.tokens_text.size(); - size_t i = 0; - while (i < chunk.tokens_text.size()) { // split into batches - for (; i < chunk.tokens_text.size() && text_batch.n_tokens < n_batch; i++) { - text_batch.token [i] = chunk.tokens_text[i]; - text_batch.pos [i] = n_past++; - text_batch.n_seq_id[i] = 1; - text_batch.seq_id [i][0] = seq_id; - text_batch.logits [i] = false; - } - if (is_last) { - // always get logits for last input chunk - text_batch.logits[text_batch.n_tokens - 1] = true; - } - ret = llama_decode(lctx, text_batch); - if (ret != 0) { - LOG_ERR("failed to decode text\n"); - llama_batch_free(text_batch); - return ret; - } - } - - } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) { - GGML_ASSERT(!is_last && "logits for last image chunk is not yet support"); - GGML_ASSERT(chunk.tokens_image != nullptr); - int64_t t0 = ggml_time_ms(); - if (ctx->print_timings) { - LOG_INF("encoding image or slice...\n"); - } - ret = mtmd_encode(ctx, chunk.tokens_image.get()); - if (ret != 0) { - LOG_ERR("failed to encode image\n"); - llama_batch_free(text_batch); - return ret; - } - if (ctx->print_timings) { - LOG_INF("image/slice encoded in %" PRId64 " ms\n", ggml_time_ms() - t0); - } - - int32_t n_tokens = mtmd_image_tokens_get_n_tokens(chunk.tokens_image.get()); - int32_t i_batch = 0; - int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch; - float * embd = mtmd_get_output_embd(ctx); - - if (mtmd_decode_use_non_causal(ctx)) { - llama_set_causal_attn(lctx, false); - // TODO @ngxson : need to make sure only one image is processed at a time, and n_ubatch must be enough to hold the image - } - - while (i_batch < n_img_batches) { // split into batches - int32_t pos_offset = i_batch*n_batch; - int32_t n_tokens_batch = std::min(n_batch, n_tokens - pos_offset); - float * embd_batch = embd + pos_offset*n_mmproj_embd; - decode_embd_batch batch_img(embd_batch, n_tokens_batch, n_past, 0); - - printf("decoding image batch %d/%d, n_tokens_batch = %d\n", i_batch+1, n_img_batches, n_tokens_batch); - - int64_t t1 = ggml_time_ms(); - ret = llama_decode(lctx, batch_img.batch); - if (ret != 0) { - LOG_ERR("failed to decode image\n"); - llama_set_causal_attn(lctx, true); // restore causal attn - llama_batch_free(text_batch); - return ret; - } - - if (ctx->print_timings) { - LOG_INF("image decoded (batch %d/%d) in %" PRId64 " ms\n", i_batch+1, n_img_batches, ggml_time_ms() - t1); - } - - i_batch++; - n_past += n_tokens_batch; - } - - if (mtmd_decode_use_non_causal(ctx)) { - llama_set_causal_attn(lctx, true); - } - - } else { - GGML_ASSERT(false && "chunk type not supported"); - } - } - - llama_batch_free(text_batch); - return 0; -} - -int32_t mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len, mtmd_bitmap & output) { - clip_image_u8_ptr img_u8(clip_image_u8_init()); - bool ok = clip_image_load_from_bytes(buf, len, img_u8.get()); - if (!ok) { - LOG_ERR("Unable to load image from buffer\n"); - return 1; - } - unsigned char * data = clip_image_u8_get_data(img_u8.get(), &output.nx, &output.ny); - output.data.resize(output.nx * output.ny * 3); - std::memcpy(output.data.data(), data, output.nx * output.ny * 3); - return 0; -} - -int32_t mtmd_helper_bitmap_init_from_file(const char * fname, mtmd_bitmap & output) { - clip_image_u8_ptr img_u8(clip_image_u8_init()); - bool ok = clip_image_load_from_file(fname, img_u8.get()); - if (!ok) { - LOG_ERR("Unable to load image %s\n", fname); - return 1; - } - unsigned char * data = clip_image_u8_get_data(img_u8.get(), &output.nx, &output.ny); - output.data.resize(output.nx * output.ny * 3); - std::memcpy(output.data.data(), data, output.nx * output.ny * 3); - return 0; -} - bool mtmd_decode_use_non_causal(mtmd_context * ctx) { projector_type proj_type = clip_get_projector_type(ctx->ctx_clip); if (proj_type == PROJECTOR_TYPE_GEMMA3) { @@ -603,6 +469,210 @@ bool mtmd_decode_use_non_causal(mtmd_context * ctx) { return false; } +bool mtmd_decode_use_mrope(mtmd_context * ctx) { + return ctx->use_mrope; +} + void mtmd_image_tokens_deleter::operator()(mtmd_image_tokens * val) { mtmd_image_tokens_free(val); } + +// these 2 helpers below use internal clip_image_u8_ptr, +// so unfortunately they cannot moved to mtmd-helper.h +// however, in theory, user can decode image file to bitmap using +// whichever library they want, and then use mtmd_bitmap_init() to create bitmap + +mtmd_bitmap * mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len) { + clip_image_u8_ptr img_u8(clip_image_u8_init()); + bool ok = clip_image_load_from_bytes(buf, len, img_u8.get()); + if (!ok) { + LOG_ERR("Unable to load image from buffer\n"); + return nullptr; + } + uint32_t nx, ny; + unsigned char * data = clip_image_u8_get_data(img_u8.get(), &nx, &ny); + return mtmd_bitmap_init(nx, ny, data); +} + +mtmd_bitmap * mtmd_helper_bitmap_init_from_file(const char * fname) { + clip_image_u8_ptr img_u8(clip_image_u8_init()); + bool ok = clip_image_load_from_file(fname, img_u8.get()); + if (!ok) { + LOG_ERR("Unable to load image %s\n", fname); + return nullptr; + } + uint32_t nx, ny; + unsigned char * data = clip_image_u8_get_data(img_u8.get(), &nx, &ny); + return mtmd_bitmap_init(nx, ny, data); +} + +// +// public API functions +// + +// mtmd_bitmap + +mtmd_bitmap * mtmd_bitmap_init(uint32_t nx, + uint32_t ny, + const unsigned char * data) { + mtmd_bitmap * bitmap = new mtmd_bitmap; + bitmap->nx = nx; + bitmap->ny = ny; + size_t data_size = (size_t)nx * ny * 3; + bitmap->data.resize(data_size); + std::memcpy(bitmap->data.data(), data, data_size); + return bitmap; +} + +uint32_t mtmd_bitmap_get_nx(const mtmd_bitmap * bitmap) { + return bitmap->nx; +} + +uint32_t mtmd_bitmap_get_ny(const mtmd_bitmap * bitmap) { + return bitmap->ny; +} + +const unsigned char * mtmd_bitmap_get_data(const mtmd_bitmap * bitmap) { + return bitmap->data.data(); +} + +const char * mtmd_bitmap_get_id(const mtmd_bitmap * bitmap) { + return bitmap->id.c_str(); +} + +void mtmd_bitmap_set_id(mtmd_bitmap * bitmap, const char * id) { + if (id) { + bitmap->id = std::string(id); + } else { + bitmap->id.clear(); + } +} + +void mtmd_bitmap_free(mtmd_bitmap * bitmap) { + if (bitmap) { + delete bitmap; + } +} + +// mtmd_input_chunks + +mtmd_input_chunks * mtmd_input_chunks_init() { + return new mtmd_input_chunks; +} + +size_t mtmd_input_chunks_size(const mtmd_input_chunks * chunks) { + return chunks->entries.size(); +} + +const mtmd_input_chunk * mtmd_input_chunks_get(const mtmd_input_chunks * chunks, size_t idx) { + if (idx >= chunks->entries.size()) { + return nullptr; + } + return &chunks->entries[idx]; +} + +void mtmd_input_chunks_free(mtmd_input_chunks * chunks) { + if (chunks) { + delete chunks; + } +} + +// mtmd_input_chunk + +enum mtmd_input_chunk_type mtmd_input_chunk_get_type(const mtmd_input_chunk * chunk) { + return chunk->type; +} + +const llama_token * mtmd_input_chunk_get_tokens_text(const mtmd_input_chunk * chunk, size_t * n_tokens_output) { + if (chunk->type == MTMD_INPUT_CHUNK_TYPE_TEXT) { + *n_tokens_output = chunk->tokens_text.size(); + return chunk->tokens_text.data(); + } + *n_tokens_output = 0; + return nullptr; +} + +const mtmd_image_tokens * mtmd_input_chunk_get_tokens_image(const mtmd_input_chunk * chunk) { + if (chunk->type == MTMD_INPUT_CHUNK_TYPE_IMAGE) { + return chunk->tokens_image.get(); + } + return nullptr; +} + +mtmd_input_chunk * mtmd_input_chunk_copy(const mtmd_input_chunk * chunk) { + mtmd_input_chunk * copy = new mtmd_input_chunk{ + chunk->type, + chunk->tokens_text, + mtmd_image_tokens_ptr(), + }; + if (chunk->tokens_image) { + // copy the image tokens + copy->tokens_image = mtmd_image_tokens_ptr(new mtmd_image_tokens()); + *copy->tokens_image = chunk->tokens_image->clone(); + } + return copy; +} + +void mtmd_input_chunk_free(mtmd_input_chunk * chunk) { + if (chunk) { + delete chunk; + } +} + +// mtmd_image_tokens + +size_t mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens) { + return image_tokens->n_tokens(); +} + +size_t mtmd_image_tokens_get_nx(const mtmd_image_tokens * image_tokens) { + return image_tokens->nx; +} + +size_t mtmd_image_tokens_get_ny(const mtmd_image_tokens * image_tokens) { + return image_tokens->ny; +} + +const char * mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens) { + return image_tokens->id.c_str(); +} + +llama_pos mtmd_image_tokens_get_n_pos(const mtmd_image_tokens * image_tokens) { + if (image_tokens->use_mrope_pos) { + return 1; // for M-RoPE, the whole image is 1 in temporal dimension + } + return image_tokens->n_tokens(); +} + +// test function + +mtmd_input_chunks * mtmd_test_create_input_chunks() { + mtmd_input_chunks * chunks = mtmd_input_chunks_init(); + if (!chunks) { + return nullptr; + } + + // create a text chunk + std::vector tokens_text = { 1, 2, 3, 4, 5 }; + mtmd_input_chunk chunk_text{ + MTMD_INPUT_CHUNK_TYPE_TEXT, + std::move(tokens_text), + {}, + }; + chunks->entries.emplace_back(std::move(chunk_text)); + + // create an image chunk + mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens); + image_tokens->nx = 4; + image_tokens->ny = 4; + image_tokens->batch_f32.entries.resize(16); + image_tokens->id = "image_1"; + mtmd_input_chunk chunk_image{ + MTMD_INPUT_CHUNK_TYPE_IMAGE, + {}, + std::move(image_tokens), + }; + chunks->entries.emplace_back(std::move(chunk_image)); + + return chunks; +} diff --git a/tools/mtmd/mtmd.h b/tools/mtmd/mtmd.h new file mode 100644 index 0000000000..0ada78c90f --- /dev/null +++ b/tools/mtmd/mtmd.h @@ -0,0 +1,331 @@ +#ifndef MTMD_H +#define MTMD_H + +#include "ggml.h" +#include "llama.h" +#include "clip.h" + +#include +#include +#include + +#ifdef __cplusplus +#include +#include +#include +#include +#endif + +/** + * libmtmd: A library for multimodal support in llama.cpp. + * + * WARNING: This API is experimental and subject to many BREAKING CHANGES. + * Issues related to API usage may receive lower priority support. + * + * For the usage, see an example in mtmd-cli.cpp + */ + +#ifdef LLAMA_SHARED +# if defined(_WIN32) && !defined(__MINGW32__) +# ifdef LLAMA_BUILD +# define MTMD_API __declspec(dllexport) +# else +# define MTMD_API __declspec(dllimport) +# endif +# else +# define MTMD_API __attribute__ ((visibility ("default"))) +# endif +#else +# define MTMD_API +#endif + +#define MTMD_DEFAULT_IMAGE_MARKER "<__image__>" + +#ifdef __cplusplus +extern "C" { +#endif + +enum mtmd_input_chunk_type { + MTMD_INPUT_CHUNK_TYPE_TEXT, + MTMD_INPUT_CHUNK_TYPE_IMAGE, +}; + +// opaque types +struct mtmd_context; +struct mtmd_bitmap; +struct mtmd_image_tokens; +struct mtmd_input_chunk; +struct mtmd_input_chunks; + +struct mtmd_input_text { + const char * text; + bool add_special; + bool parse_special; +}; + +// +// C API +// + +typedef struct mtmd_context mtmd_context; +typedef struct mtmd_bitmap mtmd_bitmap; +typedef struct mtmd_image_tokens mtmd_image_tokens; +typedef struct mtmd_input_chunk mtmd_input_chunk; +typedef struct mtmd_input_chunks mtmd_input_chunks; +typedef struct mtmd_input_text mtmd_input_text; + +struct mtmd_context_params { + bool use_gpu; + bool print_timings; + int n_threads; + enum ggml_log_level verbosity; + const char * image_marker; +}; + +MTMD_API struct mtmd_context_params mtmd_context_params_default(void); + +// initialize the mtmd context +// return nullptr on failure +MTMD_API mtmd_context * mtmd_init_from_file(const char * mmproj_fname, + const struct llama_model * text_model, + const struct mtmd_context_params ctx_params); + +MTMD_API void mtmd_free(mtmd_context * ctx); + +// whether we need to set non-causal mask before llama_decode +MTMD_API bool mtmd_decode_use_non_causal(mtmd_context * ctx); + +// whether the current model use M-RoPE for llama_decode +MTMD_API bool mtmd_decode_use_mrope(mtmd_context * ctx); + + +// mtmd_bitmap +// +// length of data must be nx * ny * 3 +// the data is in RGBRGBRGB... format +MTMD_API mtmd_bitmap * mtmd_bitmap_init (uint32_t nx, + uint32_t ny, + const unsigned char * data); +MTMD_API uint32_t mtmd_bitmap_get_nx (const mtmd_bitmap * bitmap); +MTMD_API uint32_t mtmd_bitmap_get_ny (const mtmd_bitmap * bitmap); +MTMD_API const unsigned char * mtmd_bitmap_get_data(const mtmd_bitmap * bitmap); +MTMD_API void mtmd_bitmap_free (mtmd_bitmap * bitmap); +// bitmap ID is optional, but useful for KV cache tracking +// these getters/setters are dedicated functions, so you can for example calculate the hash of the image based on mtmd_bitmap_get_data() +MTMD_API const char * mtmd_bitmap_get_id(const mtmd_bitmap * bitmap); +MTMD_API void mtmd_bitmap_set_id(mtmd_bitmap * bitmap, const char * id); + + +// mtmd_input_chunks +// +// this is simply a list of mtmd_input_chunk +// the elements can only be populated via mtmd_tokenize() +MTMD_API mtmd_input_chunks * mtmd_input_chunks_init(void); +MTMD_API size_t mtmd_input_chunks_size(const mtmd_input_chunks * chunks); +MTMD_API const mtmd_input_chunk * mtmd_input_chunks_get (const mtmd_input_chunks * chunks, size_t idx); +MTMD_API void mtmd_input_chunks_free(mtmd_input_chunks * chunks); + +// mtmd_input_chunk +// +// the instance will be constructed via mtmd_tokenize() +// it will be freed along with mtmd_input_chunks +MTMD_API enum mtmd_input_chunk_type mtmd_input_chunk_get_type (const mtmd_input_chunk * chunk); +MTMD_API const llama_token * mtmd_input_chunk_get_tokens_text (const mtmd_input_chunk * chunk, size_t * n_tokens_output); +MTMD_API const mtmd_image_tokens * mtmd_input_chunk_get_tokens_image(const mtmd_input_chunk * chunk); + +// in case you want to use custom logic to handle the chunk (i.e. KV cache management) +// you can move the chunk ownership to your own code by copying it +// remember to free the chunk when you are done with it +MTMD_API mtmd_input_chunk * mtmd_input_chunk_copy(const mtmd_input_chunk * chunk); +MTMD_API void mtmd_input_chunk_free(mtmd_input_chunk * chunk); + + +// mtmd_image_tokens +// +// the instance will be constructed via mtmd_tokenize() +// it will be freed along with mtmd_input_chunk +MTMD_API size_t mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens); +MTMD_API size_t mtmd_image_tokens_get_nx (const mtmd_image_tokens * image_tokens); +MTMD_API size_t mtmd_image_tokens_get_ny (const mtmd_image_tokens * image_tokens); +MTMD_API const char * mtmd_image_tokens_get_id (const mtmd_image_tokens * image_tokens); +// number of temporal positions (always 1 for M-RoPE, n_tokens otherwise) +MTMD_API llama_pos mtmd_image_tokens_get_n_pos (const mtmd_image_tokens * image_tokens); + +// tokenize an input text prompt and an image +// the prompt must have the input image marker (default: "<__image__>") in it +// the marker will be replaced with the image tokens +// for example: +// "here is an image: <__image__>\ndescribe it in detail." +// this will gives 3 chunks: +// 1. "here is an image: " +// 2. (image tokens) +// 3. "\ndescribe it in detail." +// number of bitmaps must be equal to the number of image markers in the prompt +// this function is thread-safe (shared ctx) +// return values: +// 0 on success +// 1 on number of images not matching the number of markers +// 2 on image preprocessing error +MTMD_API int32_t mtmd_tokenize(mtmd_context * ctx, + mtmd_input_chunks * output, + const mtmd_input_text * text, + const mtmd_bitmap ** bitmaps, + size_t n_bitmaps); + +// returns 0 on success +MTMD_API int32_t mtmd_encode(mtmd_context * ctx, + const mtmd_image_tokens * image_tokens); + +// get output embeddings from the last encode pass +MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx); + +///////////////////////////////////////// + +// +// Helper functions (can be implemented based on other functions) +// +// Please note that these helpers are not guaranteed to be stable. +// BREAKING CHANGES are expected. +// + +// helper function to construct a mtmd_bitmap from a file +// returns nullptr on failure +// this function is thread-safe +MTMD_API mtmd_bitmap * mtmd_helper_bitmap_init_from_file(const char * fname); + +// helper function to construct a mtmd_bitmap from a buffer containing a file +// the file content must be an image in format supported by stb_image (jpg, png, bmp, gif, etc.) +// returns nullptr on failure +// this function is thread-safe +MTMD_API mtmd_bitmap * mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len); + +// helper to count the total number of tokens from a list of chunks, useful to keep track of KV cache +MTMD_API size_t mtmd_helper_get_n_tokens(const mtmd_input_chunks * chunks); + +// helper to count the total position of tokens from a list of chunks, useful to keep track of n_past +// normally, n_pos is equal to n_tokens, but for M-RoPE it is different +MTMD_API llama_pos mtmd_helper_get_n_pos(const mtmd_input_chunks * chunks); + +// helper function that automatically: +// 1. run llama_decode() on text chunks +// 2. run mtmd_encode() on image chunks, then mtmd_get_output_embd() and then llama_decode() +// if any of the mtmd_encode() or llama_decode() calls return non-zero, stop and forward the error +// otherwise, returns 0 on success +// this function is NOT thread-safe +MTMD_API int32_t mtmd_helper_eval_chunks(mtmd_context * ctx, + struct llama_context * lctx, + const mtmd_input_chunks * chunks, + llama_pos n_past, + llama_seq_id seq_id, + int32_t n_batch, + bool logits_last, + llama_pos * new_n_past); + +// works like mtmd_helper_eval_chunks(), but only for a single chunk +// this function is NOT thread-safe +MTMD_API int32_t mtmd_helper_eval_chunk_single(mtmd_context * ctx, + struct llama_context * lctx, + const mtmd_input_chunk * chunk, + llama_pos n_past, + llama_seq_id seq_id, + int32_t n_batch, + bool logits_last, + llama_pos * new_n_past); + +// helper function to decode an image whose embeddings have already been calculated +// this helper will handle batching and pre/post decoding setup (for ex. gemma 3 requires non-causal attention) +// ret 0 on success, -1 on chunk not being a valid image chunk, 1 on decode failure +MTMD_API int32_t mtmd_helper_decode_image_chunk(mtmd_context * ctx, + struct llama_context * lctx, + const mtmd_input_chunk * chunk, + float * encoded_embd, + llama_pos n_past, + llama_seq_id seq_id, + int32_t n_batch, + llama_pos * new_n_past); + +///////////////////////////////////////// + +// test function, to be used in test-mtmd-c-api.c +MTMD_API mtmd_input_chunks * mtmd_test_create_input_chunks(void); + +#ifdef __cplusplus +} // extern "C" +#endif + +// +// C++ wrappers +// + +#ifdef __cplusplus + +namespace mtmd { + +struct mtmd_context_deleter { + void operator()(mtmd_context * val) { mtmd_free(val); } +}; +using context_ptr = std::unique_ptr; + +struct mtmd_bitmap_deleter { + void operator()(mtmd_bitmap * val) { mtmd_bitmap_free(val); } +}; +using bitmap_ptr = std::unique_ptr; + +struct mtmd_input_chunks_deleter { + void operator()(mtmd_input_chunks * val) { mtmd_input_chunks_free(val); } +}; +using input_chunks_ptr = std::unique_ptr; + +struct mtmd_input_chunk_deleter { + void operator()(mtmd_input_chunk * val) { mtmd_input_chunk_free(val); } +}; +using input_chunk_ptr = std::unique_ptr; + +struct bitmap { + bitmap_ptr ptr; + bitmap() : ptr(nullptr) {} + bitmap(mtmd_bitmap * bitmap) : ptr(bitmap) {} + bitmap(bitmap && other) noexcept : ptr(std::move(other.ptr)) {} + bitmap(uint32_t nx, uint32_t ny, const unsigned char * data) { + ptr.reset(mtmd_bitmap_init(nx, ny, data)); + } + ~bitmap() = default; + uint32_t nx() { return mtmd_bitmap_get_nx(ptr.get()); } + uint32_t ny() { return mtmd_bitmap_get_ny(ptr.get()); } + const unsigned char * data() { return mtmd_bitmap_get_data(ptr.get()); } + std::string id() { return mtmd_bitmap_get_id(ptr.get()); } + void set_id(const char * id) { mtmd_bitmap_set_id(ptr.get(), id); } +}; + +struct bitmaps { + std::vector entries; + ~bitmaps() = default; + // return list of pointers to mtmd_bitmap + // example: + // auto bitmaps_c_ptr = bitmaps.c_ptr(); + // int32_t res = mtmd_tokenize(... bitmaps_c_ptr.data(), bitmaps_c_ptr.size()); + std::vector c_ptr() { + std::vector res(entries.size()); + for (size_t i = 0; i < entries.size(); i++) { + res[i] = entries[i].ptr.get(); + } + return res; + } +}; + +struct input_chunks { + input_chunks_ptr ptr; + input_chunks() = default; + input_chunks(mtmd_input_chunks * chunks) : ptr(chunks) {} + ~input_chunks() = default; + size_t size() { return mtmd_input_chunks_size(ptr.get()); } + const mtmd_input_chunk * operator[](size_t idx) { + return mtmd_input_chunks_get(ptr.get(), idx); + } +}; + +} // namespace mtmd + +#endif + +#endif diff --git a/examples/llava/qwen2vl-cli.cpp b/tools/mtmd/qwen2vl-test.cpp similarity index 89% rename from examples/llava/qwen2vl-cli.cpp rename to tools/mtmd/qwen2vl-test.cpp index eca7b7f10b..7f9e3dca88 100644 --- a/examples/llava/qwen2vl-cli.cpp +++ b/tools/mtmd/qwen2vl-test.cpp @@ -23,7 +23,12 @@ #include #include #include +#include +#include +#include +// THIS FILE IS ONLY USED FOR TESTING THE QWEN2VL MODEL +// IT IS NOT A PRODUCTION CODE static bool qwen2vl_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past, int * st_pos_id, struct clip_image_size * image_size) { @@ -89,20 +94,12 @@ static bool qwen2vl_eval_image_embed(llama_context * ctx_llama, const struct lla static bool eval_tokens(struct llama_context * ctx_llama, std::vector tokens, int n_batch, int * n_past, int * st_pos_id) { int N = (int) tokens.size(); - std::vector pos; for (int i = 0; i < N; i += n_batch) { int n_eval = (int) tokens.size() - i; if (n_eval > n_batch) { n_eval = n_batch; } auto batch = llama_batch_get_one(&tokens[i], n_eval); - // TODO: add mrope pos ids somewhere else - pos.resize(batch.n_tokens * 4); - std::fill(pos.begin(), pos.end(), 0); - for (int j = 0; j < batch.n_tokens * 3; j ++) { - pos[j] = *st_pos_id + (j % batch.n_tokens); - } - batch.pos = pos.data(); if (llama_decode(ctx_llama, batch)) { LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past); @@ -367,14 +364,14 @@ static void debug_test_mrope_2d() { // 1. Initialize backend ggml_backend_t backend = NULL; std::string backend_name = ""; -#ifdef GGML_USE_CUDA - fprintf(stderr, "%s: using CUDA backend\n", __func__); - backend = ggml_backend_cuda_init(0); // init device 0 - backend_name = "cuda"; - if (!backend) { - fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__); - } -#endif +// #ifdef GGML_USE_CUDA +// fprintf(stderr, "%s: using CUDA backend\n", __func__); +// backend = ggml_backend_cuda_init(0); // init device 0 +// backend_name = "cuda"; +// if (!backend) { +// fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__); +// } +// #endif // if there aren't GPU Backends fallback to CPU backend if (!backend) { backend = ggml_backend_cpu_init(); @@ -483,28 +480,82 @@ static void debug_test_mrope_2d() { ggml_backend_free(backend); } -static void debug_dump_img_embed(struct llava_context * ctx_llava) { - int n_embd = llama_model_n_embd(llama_get_model(ctx_llava->ctx_llama)); - int ne = n_embd * 4; - float vals[56 * 56 * 3]; +enum model_output_type { + conv3d, + patch_embed, + patch_win_attn_scatter, + first_attn_layer, + last_attn_layer, + attn_softmax, + final_layer, +}; + +static void debug_dump_img_embed(struct llava_context * ctx_llava, model_output_type output_type) { + constexpr int ih = 140; + constexpr int iw = 196; + // constexpr int ih = 56; + // constexpr int iw = 56; + // int n_embd = llama_model_n_embd(llama_get_model(ctx_llava->ctx_llama)); + int n_embd = 1280; + int merge = 1; + if (output_type == model_output_type::final_layer) { + n_embd = 2048; + merge = 2; + } + else if (output_type == model_output_type::attn_softmax) { + merge = 1; + n_embd = (ih/14/merge) * (iw/14/merge) * 16; + } + + int ne = (ih/14/merge) * (iw/14/merge) * n_embd; + float vals[iw * ih * 3]; // float embd[ne]; std::vector embd; embd.resize(ne); - for (int i = 0; i < 56*56; i++) + for (int i = 0; i < iw*ih; i++) { for (int c = 0; c < 3; c++) - vals[i * 3 + c] = (float)(i % (56 * 56)) / (56*56); + vals[i * 3 + c] = (float)i / (iw*ih); } - clip_encode_float_image(ctx_llava->ctx_clip, 16, vals, 56, 56, embd.data()); + clip_encode_float_image(ctx_llava->ctx_clip, 8, vals, ih, iw, embd.data()); - std::ofstream outFile("img_embed.bin", std::ios::binary); + std::string file_postfix = ""; + switch (output_type) + { + case model_output_type::conv3d: + file_postfix = "conv3d"; + break; + case model_output_type::patch_embed: + file_postfix = "patch_embed"; + break; + case model_output_type::patch_win_attn_scatter: + file_postfix = "scatter"; + break; + case model_output_type::first_attn_layer: + file_postfix = "first_attn"; + break; + case model_output_type::last_attn_layer: + file_postfix = "last_attn"; + break; + case model_output_type::attn_softmax: + file_postfix = "attn_softmax"; + break; + case model_output_type::final_layer: + file_postfix = "final"; + break; + default: + break; + } + auto output_path = "img_embed_" + file_postfix + ".bin"; + + std::ofstream outFile(output_path, std::ios::binary); if (outFile.is_open()) { outFile.write(reinterpret_cast(embd.data()), ne * sizeof(float)); outFile.close(); - std::cout << "Data successfully written to mrope.bin" << std::endl; + std::cout << "Data successfully written to ::[ " << output_path << std::endl; } else { std::cerr << "Error opening file!" << std::endl; } @@ -551,8 +602,9 @@ int main(int argc, char ** argv) { } else if (params.image[0].empty()) { auto ctx_llava = llava_init_context(¶ms, model); - debug_test_mrope_2d(); - debug_dump_img_embed(ctx_llava); + // debug_test_mrope_2d(); + debug_dump_img_embed(ctx_llava, model_output_type::final_layer); + // debug_dump_img_embed(ctx_llava, model_output_type::last_attn_layer); llama_perf_context_print(ctx_llava->ctx_llama); ctx_llava->model = NULL; diff --git a/examples/llava/requirements.txt b/tools/mtmd/requirements.txt similarity index 100% rename from examples/llava/requirements.txt rename to tools/mtmd/requirements.txt diff --git a/examples/llava/test-1.jpeg b/tools/mtmd/test-1.jpeg similarity index 100% rename from examples/llava/test-1.jpeg rename to tools/mtmd/test-1.jpeg diff --git a/examples/llava/tests.sh b/tools/mtmd/tests.sh similarity index 72% rename from examples/llava/tests.sh rename to tools/mtmd/tests.sh index e612857edc..05ac7a04d8 100755 --- a/examples/llava/tests.sh +++ b/tools/mtmd/tests.sh @@ -36,17 +36,10 @@ add_test() { arr_tmpl+=("$tmpl") } -add_test_big() { - if [ "$RUN_BIG_TESTS" = true ]; then - add_test "$@" - fi -} - add_test "llama-mtmd-cli" "ggml-org/SmolVLM-500M-Instruct-GGUF:Q8_0" add_test "llama-mtmd-cli" "ggml-org/SmolVLM2-2.2B-Instruct-GGUF:Q4_K_M" add_test "llama-mtmd-cli" "ggml-org/SmolVLM2-500M-Video-Instruct-GGUF:Q8_0" add_test "llama-mtmd-cli" "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M" -add_test "llama-mtmd-cli" "guinmoon/MobileVLM-3B-GGUF:Q4_K_M" "deepseek" add_test "llama-mtmd-cli" "THUDM/glm-edge-v-5b-gguf:Q4_K_M" add_test "llama-mtmd-cli" "second-state/Llava-v1.5-7B-GGUF:Q2_K" "vicuna" add_test "llama-mtmd-cli" "cjpais/llava-1.6-mistral-7b-gguf:Q3_K" "vicuna" @@ -54,10 +47,24 @@ add_test "llama-mtmd-cli" "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M" add_test "llama-mtmd-cli" "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K" # model from openbmb is corrupted add_test "llama-mtmd-cli" "openbmb/MiniCPM-V-2_6-gguf:Q2_K" add_test "llama-mtmd-cli" "openbmb/MiniCPM-o-2_6-gguf:Q4_0" -add_test "llama-qwen2vl-cli" "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M" +add_test "llama-mtmd-cli" "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M" +add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M" +add_test "llama-mtmd-cli" "ggml-org/InternVL2_5-1B-GGUF:Q8_0" +add_test "llama-mtmd-cli" "ggml-org/InternVL3-1B-Instruct-GGUF:Q8_0" # to test the big models, run: ./tests.sh big -add_test_big "llama-mtmd-cli" "ggml-org/pixtral-12b-GGUF:Q4_K_M" +if [ "$RUN_BIG_TESTS" = true ]; then + add_test "llama-mtmd-cli" "ggml-org/pixtral-12b-GGUF:Q4_K_M" + add_test "llama-mtmd-cli" "ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF" "mistral-v7" + add_test "llama-mtmd-cli" "ggml-org/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M" + add_test "llama-mtmd-cli" "ggml-org/Qwen2-VL-7B-Instruct-GGUF:Q4_K_M" + add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M" + add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-7B-Instruct-GGUF:Q4_K_M" + add_test "llama-mtmd-cli" "ggml-org/InternVL3-8B-Instruct-GGUF:Q4_K_M" + add_test "llama-mtmd-cli" "ggml-org/InternVL3-14B-Instruct-GGUF:Q4_K_M" + # add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-32B-Instruct-GGUF:Q4_K_M" # does not work on my mac M3 Ultra + # add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-72B-Instruct-GGUF:Q4_K_M" # too big +fi # these models always give the wrong answer, not sure why # add_test "llama-mtmd-cli" "ggml-org/SmolVLM-Instruct-GGUF:Q4_K_M" @@ -66,6 +73,7 @@ add_test_big "llama-mtmd-cli" "ggml-org/pixtral-12b-GGUF:Q4_K_M" # this model has broken chat template, not usable # add_test "llama-mtmd-cli" "cmp-nct/Yi-VL-6B-GGUF:Q5_K" +# add_test "llama-mtmd-cli" "guinmoon/MobileVLM-3B-GGUF:Q4_K_M" "deepseek" ############### diff --git a/examples/perplexity/CMakeLists.txt b/tools/perplexity/CMakeLists.txt similarity index 100% rename from examples/perplexity/CMakeLists.txt rename to tools/perplexity/CMakeLists.txt diff --git a/examples/perplexity/README.md b/tools/perplexity/README.md similarity index 100% rename from examples/perplexity/README.md rename to tools/perplexity/README.md diff --git a/examples/perplexity/perplexity.cpp b/tools/perplexity/perplexity.cpp similarity index 99% rename from examples/perplexity/perplexity.cpp rename to tools/perplexity/perplexity.cpp index 175f2804b5..b5cdf5beb1 100644 --- a/examples/perplexity/perplexity.cpp +++ b/tools/perplexity/perplexity.cpp @@ -1554,7 +1554,10 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par if (int(batch_indeces.size()) != num_answers) { batch_indeces.resize(num_answers); } - for (int s = 0; s < num_answers; ++s) batch_indeces[s] = s0 + s; + + for (int s = 0; s < num_answers; ++s) { + batch_indeces[s] = s0 + s; + } for (size_t i = 0; i < cur_task.common_prefix; ++i) { //llama_batch_add(batch, cur_task.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3}, false); @@ -1970,7 +1973,6 @@ int main(int argc, char ** argv) { common_params params; params.n_ctx = 512; - params.logits_all = true; params.escape = false; if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) { diff --git a/examples/quantize/CMakeLists.txt b/tools/quantize/CMakeLists.txt similarity index 100% rename from examples/quantize/CMakeLists.txt rename to tools/quantize/CMakeLists.txt diff --git a/examples/quantize/README.md b/tools/quantize/README.md similarity index 100% rename from examples/quantize/README.md rename to tools/quantize/README.md diff --git a/examples/quantize/quantize.cpp b/tools/quantize/quantize.cpp similarity index 100% rename from examples/quantize/quantize.cpp rename to tools/quantize/quantize.cpp diff --git a/examples/quantize/tests.sh b/tools/quantize/tests.sh similarity index 100% rename from examples/quantize/tests.sh rename to tools/quantize/tests.sh diff --git a/examples/rpc/CMakeLists.txt b/tools/rpc/CMakeLists.txt similarity index 100% rename from examples/rpc/CMakeLists.txt rename to tools/rpc/CMakeLists.txt diff --git a/examples/rpc/README.md b/tools/rpc/README.md similarity index 100% rename from examples/rpc/README.md rename to tools/rpc/README.md diff --git a/examples/rpc/rpc-server.cpp b/tools/rpc/rpc-server.cpp similarity index 72% rename from examples/rpc/rpc-server.cpp rename to tools/rpc/rpc-server.cpp index 0277e25cb5..581c74018c 100644 --- a/examples/rpc/rpc-server.cpp +++ b/tools/rpc/rpc-server.cpp @@ -2,24 +2,6 @@ #define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING #endif -#include "ggml-cpu.h" - -#ifdef GGML_USE_CUDA -#include "ggml-cuda.h" -#endif - -#ifdef GGML_USE_METAL -#include "ggml-metal.h" -#endif - -#ifdef GGML_USE_VULKAN -#include "ggml-vulkan.h" -#endif - -#ifdef GGML_USE_SYCL -#include "ggml-sycl.h" -#endif - #include "ggml-rpc.h" #ifdef _WIN32 # define NOMINMAX @@ -154,6 +136,7 @@ struct rpc_server_params { size_t backend_mem = 0; bool use_cache = false; int n_threads = std::max(1U, std::thread::hardware_concurrency()/2); + std::string device; }; static void print_usage(int /*argc*/, char ** argv, rpc_server_params params) { @@ -161,6 +144,7 @@ static void print_usage(int /*argc*/, char ** argv, rpc_server_params params) { fprintf(stderr, "options:\n"); fprintf(stderr, " -h, --help show this help message and exit\n"); fprintf(stderr, " -t, --threads number of threads for the CPU backend (default: %d)\n", params.n_threads); + fprintf(stderr, " -d DEV, --device device to use\n"); fprintf(stderr, " -H HOST, --host HOST host to bind to (default: %s)\n", params.host.c_str()); fprintf(stderr, " -p PORT, --port PORT port to bind to (default: %d)\n", params.port); fprintf(stderr, " -m MEM, --mem MEM backend memory size (in MB)\n"); @@ -186,6 +170,22 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params & fprintf(stderr, "error: invalid number of threads: %d\n", params.n_threads); return false; } + } else if (arg == "-d" || arg == "--device") { + if (++i >= argc) { + return false; + } + params.device = argv[i]; + if (ggml_backend_dev_by_name(params.device.c_str()) == nullptr) { + fprintf(stderr, "error: unknown device: %s\n", params.device.c_str()); + fprintf(stderr, "available devices:\n"); + for (size_t i = 0; i < ggml_backend_dev_count(); i++) { + auto * dev = ggml_backend_dev_get(i); + size_t free, total; + ggml_backend_dev_memory(dev, &free, &total); + printf(" %s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / 1024 / 1024, free / 1024 / 1024); + } + return false; + } } else if (arg == "-p" || arg == "--port") { if (++i >= argc) { return false; @@ -214,66 +214,55 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params & } static ggml_backend_t create_backend(const rpc_server_params & params) { - ggml_backend_t backend = NULL; -#ifdef GGML_USE_CUDA - fprintf(stderr, "%s: using CUDA backend\n", __func__); - backend = ggml_backend_cuda_init(0); // init device 0 - if (!backend) { - fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__); - } -#elif GGML_USE_METAL - fprintf(stderr, "%s: using Metal backend\n", __func__); - backend = ggml_backend_metal_init(); - if (!backend) { - fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__); - } -#elif GGML_USE_VULKAN - fprintf(stderr, "%s: using Vulkan backend\n", __func__); - backend = ggml_backend_vk_init(0); // init device 0 - if (!backend) { - fprintf(stderr, "%s: ggml_backend_vulkan_init() failed\n", __func__); - } -#elif GGML_USE_SYCL - fprintf(stderr, "%s: using SYCL backend\n", __func__); - backend = ggml_backend_sycl_init(0); // init device 0 - if (!backend) { - fprintf(stderr, "%s: ggml_backend_sycl_init() failed\n", __func__); - } -#endif + ggml_backend_t backend = nullptr; - // if there aren't GPU Backends fallback to CPU backend - if (!backend) { - fprintf(stderr, "%s: using CPU backend\n", __func__); - backend = ggml_backend_cpu_init(); - ggml_backend_cpu_set_n_threads(backend, params.n_threads); + if (!params.device.empty()) { + ggml_backend_dev_t dev = ggml_backend_dev_by_name(params.device.c_str()); + if (dev) { + backend = ggml_backend_dev_init(dev, nullptr); + if (!backend) { + fprintf(stderr, "Failed to create backend for device %s\n", params.device.c_str()); + return nullptr; + } + } } + + // try to initialize a GPU backend first + if (!backend) { + backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr); + } + + // if there aren't GPU backends fallback to CPU backend + if (!backend) { + backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr); + } + + if (backend) { + fprintf(stderr, "%s: using %s backend\n", __func__, ggml_backend_name(backend)); + + // set the number of threads + ggml_backend_dev_t dev = ggml_backend_get_device(backend); + ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr; + if (reg) { + auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads"); + if (ggml_backend_set_n_threads_fn) { + ggml_backend_set_n_threads_fn(backend, params.n_threads); + } + } + } + return backend; } -static void get_backend_memory(size_t * free_mem, size_t * total_mem) { -#ifdef GGML_USE_CUDA - ggml_backend_cuda_get_device_memory(0, free_mem, total_mem); -#elif GGML_USE_VULKAN - ggml_backend_vk_get_device_memory(0, free_mem, total_mem); -#elif GGML_USE_SYCL - ggml_backend_sycl_get_device_memory(0, free_mem, total_mem); -#else - #ifdef _WIN32 - MEMORYSTATUSEX status; - status.dwLength = sizeof(status); - GlobalMemoryStatusEx(&status); - *total_mem = status.ullTotalPhys; - *free_mem = status.ullAvailPhys; - #else - long pages = sysconf(_SC_PHYS_PAGES); - long page_size = sysconf(_SC_PAGE_SIZE); - *total_mem = pages * page_size; - *free_mem = *total_mem; - #endif -#endif +static void get_backend_memory(ggml_backend_t backend, size_t * free_mem, size_t * total_mem) { + ggml_backend_dev_t dev = ggml_backend_get_device(backend); + GGML_ASSERT(dev != nullptr); + ggml_backend_dev_memory(dev, free_mem, total_mem); } int main(int argc, char * argv[]) { + ggml_backend_load_all(); + rpc_server_params params; if (!rpc_server_params_parse(argc, argv, params)) { fprintf(stderr, "Invalid parameters\n"); @@ -301,25 +290,33 @@ int main(int argc, char * argv[]) { free_mem = params.backend_mem; total_mem = params.backend_mem; } else { - get_backend_memory(&free_mem, &total_mem); + get_backend_memory(backend, &free_mem, &total_mem); } const char * cache_dir = nullptr; - std::string cache_dir_str = fs_get_cache_directory() + "rpc/"; + std::string cache_dir_str; if (params.use_cache) { + cache_dir_str = fs_get_cache_directory() + "rpc/"; if (!fs_create_directory_with_parents(cache_dir_str)) { fprintf(stderr, "Failed to create cache directory: %s\n", cache_dir_str.c_str()); return 1; } cache_dir = cache_dir_str.c_str(); } - printf("Starting RPC server v%d.%d.%d\n", - RPC_PROTO_MAJOR_VERSION, - RPC_PROTO_MINOR_VERSION, - RPC_PROTO_PATCH_VERSION); - printf(" endpoint : %s\n", endpoint.c_str()); - printf(" local cache : %s\n", cache_dir ? cache_dir : "n/a"); - printf(" backend memory : %zu MB\n", free_mem / (1024 * 1024)); - ggml_backend_rpc_start_server(backend, endpoint.c_str(), cache_dir, free_mem, total_mem); + + ggml_backend_reg_t reg = ggml_backend_reg_by_name("RPC"); + if (!reg) { + fprintf(stderr, "Failed to find RPC backend\n"); + return 1; + } + + auto start_server_fn = (decltype(ggml_backend_rpc_start_server)*) ggml_backend_reg_get_proc_address(reg, "ggml_backend_rpc_start_server"); + if (!start_server_fn) { + fprintf(stderr, "Failed to obtain RPC backend start server function\n"); + return 1; + } + + start_server_fn(backend, endpoint.c_str(), cache_dir, free_mem, total_mem); + ggml_backend_free(backend); return 0; } diff --git a/examples/run/CMakeLists.txt b/tools/run/CMakeLists.txt similarity index 100% rename from examples/run/CMakeLists.txt rename to tools/run/CMakeLists.txt diff --git a/examples/run/README.md b/tools/run/README.md similarity index 90% rename from examples/run/README.md rename to tools/run/README.md index 89a5520798..5fd769b44c 100644 --- a/examples/run/README.md +++ b/tools/run/README.md @@ -42,6 +42,8 @@ Examples: llama-run ollama://smollm:135m llama-run hf://QuantFactory/SmolLM-135M-GGUF/SmolLM-135M.Q2_K.gguf llama-run huggingface://bartowski/SmolLM-1.7B-Instruct-v0.2-GGUF/SmolLM-1.7B-Instruct-v0.2-IQ3_M.gguf + llama-run ms://QuantFactory/SmolLM-135M-GGUF/SmolLM-135M.Q2_K.gguf + llama-run modelscope://bartowski/SmolLM-1.7B-Instruct-v0.2-GGUF/SmolLM-1.7B-Instruct-v0.2-IQ3_M.gguf llama-run https://example.com/some-file1.gguf llama-run some-file2.gguf llama-run file://some-file3.gguf diff --git a/examples/run/linenoise.cpp/linenoise.cpp b/tools/run/linenoise.cpp/linenoise.cpp similarity index 100% rename from examples/run/linenoise.cpp/linenoise.cpp rename to tools/run/linenoise.cpp/linenoise.cpp diff --git a/examples/run/linenoise.cpp/linenoise.h b/tools/run/linenoise.cpp/linenoise.h similarity index 100% rename from examples/run/linenoise.cpp/linenoise.h rename to tools/run/linenoise.cpp/linenoise.h diff --git a/examples/run/run.cpp b/tools/run/run.cpp similarity index 97% rename from examples/run/run.cpp rename to tools/run/run.cpp index e63c2aac33..a189ae7faf 100644 --- a/examples/run/run.cpp +++ b/tools/run/run.cpp @@ -267,7 +267,7 @@ class Opt { "Commands:\n" " model\n" " Model is a string with an optional prefix of \n" - " huggingface:// (hf://), ollama://, https:// or file://.\n" + " huggingface:// (hf://), modelscope:// (ms://), ollama://, https:// or file://.\n" " If no protocol is specified and a file exists in the specified\n" " path, file:// is assumed, otherwise if a file does not exist in\n" " the specified path, ollama:// is assumed. Models that are being\n" @@ -282,6 +282,9 @@ class Opt { " llama-run hf://QuantFactory/SmolLM-135M-GGUF/SmolLM-135M.Q2_K.gguf\n" " llama-run " "huggingface://bartowski/SmolLM-1.7B-Instruct-v0.2-GGUF/SmolLM-1.7B-Instruct-v0.2-IQ3_M.gguf\n" + " llama-run ms://QuantFactory/SmolLM-135M-GGUF/SmolLM-135M.Q2_K.gguf\n" + " llama-run " + "modelscope://bartowski/SmolLM-1.7B-Instruct-v0.2-GGUF/SmolLM-1.7B-Instruct-v0.2-IQ3_M.gguf\n" " llama-run https://example.com/some-file1.gguf\n" " llama-run some-file2.gguf\n" " llama-run file://some-file3.gguf\n" @@ -689,7 +692,7 @@ class LlamaData { return 0; } - int huggingface_dl(std::string & model, const std::string & bn) { + int dl_from_endpoint(std::string & model_endpoint, std::string & model, const std::string & bn) { // Find the second occurrence of '/' after protocol string size_t pos = model.find('/'); pos = model.find('/', pos + 1); @@ -697,8 +700,6 @@ class LlamaData { std::vector headers = { "User-Agent: llama-cpp", "Accept: application/json" }; std::string url; - std::string model_endpoint = get_model_endpoint(); - if (pos == std::string::npos) { auto [model_name, manifest_url] = extract_model_and_tag(model, model_endpoint + "v2/"); hfr = model_name; @@ -720,6 +721,16 @@ class LlamaData { return download(url, bn, true, headers); } + int modelscope_dl(std::string & model, const std::string & bn) { + std::string model_endpoint = "https://modelscope.cn/models/"; + return dl_from_endpoint(model_endpoint, model, bn); + } + + int huggingface_dl(std::string & model, const std::string & bn) { + std::string model_endpoint = get_model_endpoint(); + return dl_from_endpoint(model_endpoint, model, bn); + } + int ollama_dl(std::string & model, const std::string & bn) { const std::vector headers = { "Accept: application/vnd.docker.distribution.manifest.v2+json" }; if (model.find('/') == std::string::npos) { @@ -837,6 +848,9 @@ class LlamaData { rm_until_substring(model_, "hf.co/"); rm_until_substring(model_, "://"); ret = huggingface_dl(model_, bn); + } else if (string_starts_with(model_, "ms://") || string_starts_with(model_, "modelscope://")) { + rm_until_substring(model_, "://"); + ret = modelscope_dl(model_, bn); } else if ((string_starts_with(model_, "https://") || string_starts_with(model_, "http://")) && !string_starts_with(model_, "https://ollama.com/library/")) { ret = download(model_, bn, true); diff --git a/examples/server/CMakeLists.txt b/tools/server/CMakeLists.txt similarity index 91% rename from examples/server/CMakeLists.txt rename to tools/server/CMakeLists.txt index aee90388e4..17109fddbd 100644 --- a/examples/server/CMakeLists.txt +++ b/tools/server/CMakeLists.txt @@ -34,8 +34,9 @@ endforeach() add_executable(${TARGET} ${TARGET_SRCS}) install(TARGETS ${TARGET} RUNTIME) +target_include_directories(${TARGET} PRIVATE ../llava) target_include_directories(${TARGET} PRIVATE ${CMAKE_SOURCE_DIR}) -target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT}) +target_link_libraries(${TARGET} PRIVATE common mtmd ${CMAKE_THREAD_LIBS_INIT}) if (LLAMA_SERVER_SSL) find_package(OpenSSL REQUIRED) diff --git a/examples/server/README.md b/tools/server/README.md similarity index 88% rename from examples/server/README.md rename to tools/server/README.md index a2a0903261..7b944c35ba 100644 --- a/examples/server/README.md +++ b/tools/server/README.md @@ -7,13 +7,15 @@ Set of LLM REST APIs and a simple web front end 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 - * Reranking endoint (WIP: https://github.com/ggml-org/llama.cpp/pull/9510) + * Reranking endoint (https://github.com/ggml-org/llama.cpp/pull/9510) * Parallel decoding with multi-user support * Continuous batching - * Multimodal (wip) + * Multimodal ([documentation](../../docs/multimodal.md)) / with OpenAI-compatible API support * Monitoring endpoints * Schema-constrained JSON response format * [Function calling](../../docs/function-calling.md) / tool use for ~any model + * Speculative decoding + * Easy-to-use web UI The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggml-org/llama.cpp/issues/4216). @@ -27,6 +29,7 @@ The project is under active development, and we are [looking for feedback and co | -------- | ----------- | | `-h, --help, --usage` | print usage and exit | | `--version` | show version and build info | +| `--completion-bash` | print source-able bash completion script for llama.cpp | | `--verbose-prompt` | print a verbose prompt before generation (default: false) | | `-t, --threads N` | number of threads to use during generation (default: -1)
(env: LLAMA_ARG_THREADS) | | `-tb, --threads-batch N` | number of threads to use during batch and prompt processing (default: same as --threads) | @@ -41,7 +44,7 @@ The project is under active development, and we are [looking for feedback and co | `--prio-batch N` | set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: 0)
| | `--poll-batch <0\|1>` | use polling to wait for work (default: same as --poll) | | `-c, --ctx-size N` | size of the prompt context (default: 4096, 0 = loaded from model)
(env: LLAMA_ARG_CTX_SIZE) | -| `-n, --predict, --n-predict N` | number of tokens to predict (default: -1, -1 = infinity, -2 = until context filled)
(env: LLAMA_ARG_N_PREDICT) | +| `-n, --predict, --n-predict N` | number of tokens to predict (default: -1, -1 = infinity)
(env: LLAMA_ARG_N_PREDICT) | | `-b, --batch-size N` | logical maximum batch size (default: 2048)
(env: LLAMA_ARG_BATCH) | | `-ub, --ubatch-size N` | physical maximum batch size (default: 512)
(env: LLAMA_ARG_UBATCH) | | `--keep N` | number of tokens to keep from the initial prompt (default: 0, -1 = all) | @@ -69,6 +72,7 @@ The project is under active development, and we are [looking for feedback and co | `--numa TYPE` | attempt optimizations that help on some NUMA systems
- distribute: spread execution evenly over all nodes
- isolate: only spawn threads on CPUs on the node that execution started on
- numactl: use the CPU map provided by numactl
if run without this previously, it is recommended to drop the system page cache before using this
see https://github.com/ggml-org/llama.cpp/issues/1437
(env: LLAMA_ARG_NUMA) | | `-dev, --device ` | comma-separated list of devices to use for offloading (none = don't offload)
use --list-devices to see a list of available devices
(env: LLAMA_ARG_DEVICE) | | `--list-devices` | print list of available devices and exit | +| `--override-tensor, -ot =,...` | override tensor buffer type | | `-ngl, --gpu-layers, --n-gpu-layers N` | number of layers to store in VRAM
(env: LLAMA_ARG_N_GPU_LAYERS) | | `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:
- none: use one GPU only
- layer (default): split layers and KV across GPUs
- row: split rows across GPUs
(env: LLAMA_ARG_SPLIT_MODE) | | `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1
(env: LLAMA_ARG_TENSOR_SPLIT) | @@ -82,15 +86,18 @@ The project is under active development, and we are [looking for feedback and co | `--control-vector-layer-range START END` | layer range to apply the control vector(s) to, start and end inclusive | | `-m, --model FNAME` | model path (default: `models/$filename` with filename from `--hf-file` or `--model-url` if set, otherwise models/7B/ggml-model-f16.gguf)
(env: LLAMA_ARG_MODEL) | | `-mu, --model-url MODEL_URL` | model download url (default: unused)
(env: LLAMA_ARG_MODEL_URL) | -| `-hfr, --hf-repo REPO` | Hugging Face model repository (default: unused)
(env: LLAMA_ARG_HF_REPO) | -| `-hff, --hf-file FILE` | Hugging Face model file (default: unused)
(env: LLAMA_ARG_HF_FILE) | +| `-hf, -hfr, --hf-repo /[:quant]` | Hugging Face model repository; quant is optional, case-insensitive, default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.
mmproj is also downloaded automatically if available. to disable, add --no-mmproj
example: unsloth/phi-4-GGUF:q4_k_m
(default: unused)
(env: LLAMA_ARG_HF_REPO) | +| `-hfd, -hfrd, --hf-repo-draft /[:quant]` | Same as --hf-repo, but for the draft model (default: unused)
(env: LLAMA_ARG_HFD_REPO) | +| `-hff, --hf-file FILE` | Hugging Face model file. If specified, it will override the quant in --hf-repo (default: unused)
(env: LLAMA_ARG_HF_FILE) | +| `-hfv, -hfrv, --hf-repo-v /[:quant]` | Hugging Face model repository for the vocoder model (default: unused)
(env: LLAMA_ARG_HF_REPO_V) | +| `-hffv, --hf-file-v FILE` | Hugging Face model file for the vocoder model (default: unused)
(env: LLAMA_ARG_HF_FILE_V) | | `-hft, --hf-token TOKEN` | Hugging Face access token (default: value from HF_TOKEN environment variable)
(env: HF_TOKEN) | | `--log-disable` | Log disable | | `--log-file FNAME` | Log to file | | `--log-colors` | Enable colored logging
(env: LLAMA_LOG_COLORS) | | `-v, --verbose, --log-verbose` | Set verbosity level to infinity (i.e. log all messages, useful for debugging) | | `-lv, --verbosity, --log-verbosity N` | Set the verbosity threshold. Messages with a higher verbosity will be ignored.
(env: LLAMA_LOG_VERBOSITY) | -| `--log-prefix` | Enable prefx in log messages
(env: LLAMA_LOG_PREFIX) | +| `--log-prefix` | Enable prefix in log messages
(env: LLAMA_LOG_PREFIX) | | `--log-timestamps` | Enable timestamps in log messages
(env: LLAMA_LOG_TIMESTAMPS) | @@ -98,9 +105,9 @@ The project is under active development, and we are [looking for feedback and co | Argument | Explanation | | -------- | ----------- | -| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'
(default: dry;top_k;typ_p;top_p;min_p;xtc;temperature) | +| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'
(default: penalties;dry;top_n_sigma;top_k;typ_p;top_p;min_p;xtc;temperature) | | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) | -| `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: dkypmxt) | +| `--sampling-seq, --sampler-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) | | `--top-k N` | top-k sampling (default: 40, 0 = disabled) | @@ -127,22 +134,26 @@ The project is under active development, and we are [looking for feedback and co | `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') | | `--grammar-file FNAME` | file to read grammar from | | `-j, --json-schema SCHEMA` | JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object
For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead | -| `--jinja` | Enable experimental Jinja templating engine (required for tool use) | -| `--reasoning-format FORMAT` | Controls extraction of model thinking traces and the format / field in which they are returned (default: `deepseek`; allowed values: `deepseek`, `none`; requires `--jinja`). `none` will leave thinking traces inline in `message.content` in a model-specific format, while `deepseek` will return them separately under `message.reasoning_content` | +| `-jf, --json-schema-file FILE` | File containing a JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object
For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead | + **Example-specific params** | Argument | Explanation | | -------- | ----------- | -| `--no-context-shift` | disables context shift on inifinite text generation (default: disabled)
(env: LLAMA_ARG_NO_CONTEXT_SHIFT) | +| `--no-context-shift` | disables context shift on infinite text generation (default: disabled)
(env: LLAMA_ARG_NO_CONTEXT_SHIFT) | | `-sp, --special` | special tokens output enabled (default: false) | | `--no-warmup` | skip warming up the model with an empty run | | `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) | | `--pooling {none,mean,cls,last,rank}` | pooling type for embeddings, use model default if unspecified
(env: LLAMA_ARG_POOLING) | | `-cb, --cont-batching` | enable continuous batching (a.k.a dynamic batching) (default: enabled)
(env: LLAMA_ARG_CONT_BATCHING) | | `-nocb, --no-cont-batching` | disable continuous batching
(env: LLAMA_ARG_NO_CONT_BATCHING) | +| `--mmproj FILE` | path to a multimodal projector file. see tools/mtmd/README.md
note: if -hf is used, this argument can be omitted
(env: LLAMA_ARG_MMPROJ) | +| `--mmproj-url URL` | URL to a multimodal projector file. see tools/mtmd/README.md
(env: LLAMA_ARG_MMPROJ_URL) | +| `--no-mmproj` | explicitly disable multimodal projector, useful when using -hf
(env: LLAMA_ARG_NO_MMPROJ) | +| `--no-mmproj-offload` | do not offload multimodal projector to GPU
(env: LLAMA_ARG_NO_MMPROJ_OFFLOAD) | | `-a, --alias STRING` | set alias for model name (to be used by REST API)
(env: LLAMA_ARG_ALIAS) | -| `--host HOST` | ip address to listen (default: 127.0.0.1)
(env: LLAMA_ARG_HOST) | +| `--host HOST` | ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: 127.0.0.1)
(env: LLAMA_ARG_HOST) | | `--port PORT` | port to listen (default: 8080)
(env: LLAMA_ARG_PORT) | | `--path PATH` | path to serve static files from (default: )
(env: LLAMA_ARG_STATIC_PATH) | | `--no-webui` | Disable the Web UI (default: enabled)
(env: LLAMA_ARG_NO_WEBUI) | @@ -154,22 +165,35 @@ The project is under active development, and we are [looking for feedback and co | `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate
(env: LLAMA_ARG_SSL_CERT_FILE) | | `-to, --timeout N` | server read/write timeout in seconds (default: 600)
(env: LLAMA_ARG_TIMEOUT) | | `--threads-http N` | number of threads used to process HTTP requests (default: -1)
(env: LLAMA_ARG_THREADS_HTTP) | -| `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting (default: 0)
(env: LLAMA_ARG_CACHE_REUSE) | +| `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting (default: 0)
[(card)](https://ggml.ai/f0.png)
(env: LLAMA_ARG_CACHE_REUSE) | | `--metrics` | enable prometheus compatible metrics endpoint (default: disabled)
(env: LLAMA_ARG_ENDPOINT_METRICS) | | `--slots` | enable slots monitoring endpoint (default: disabled)
(env: LLAMA_ARG_ENDPOINT_SLOTS) | | `--props` | enable changing global properties via POST /props (default: disabled)
(env: LLAMA_ARG_ENDPOINT_PROPS) | | `--no-slots` | disables slots monitoring endpoint
(env: LLAMA_ARG_NO_ENDPOINT_SLOTS) | | `--slot-save-path PATH` | path to save slot kv cache (default: disabled) | -| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)
if suffix/prefix are specified, template will be disabled
list of built-in templates:
chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, exaone3, gemma, granite, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, monarch, openchat, orion, phi3, rwkv-world, vicuna, vicuna-orca, zephyr
(env: LLAMA_ARG_CHAT_TEMPLATE) | +| `--jinja` | use jinja template for chat (default: disabled)
(env: LLAMA_ARG_JINJA) | +| `--reasoning-format FORMAT` | reasoning format (default: deepseek; allowed values: deepseek, none)
controls whether thought tags are extracted from the response, and in which format they're returned. 'none' leaves thoughts unparsed in `message.content`, 'deepseek' puts them in `message.reasoning_content` (for DeepSeek R1 & Command R7B only).
only supported for non-streamed responses
(env: LLAMA_ARG_THINK) | +| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)
if suffix/prefix are specified, template will be disabled
only commonly used templates are accepted (unless --jinja is set before this flag):
list of built-in templates:
bailing, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, falcon3, gemma, gigachat, glmedge, granite, 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, phi3, phi4, rwkv-world, smolvlm, vicuna, vicuna-orca, yandex, zephyr
(env: LLAMA_ARG_CHAT_TEMPLATE) | +| `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)
if suffix/prefix are specified, template will be disabled
only commonly used templates are accepted (unless --jinja is set before this flag):
list of built-in templates:
bailing, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, falcon3, gemma, gigachat, glmedge, granite, 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, phi3, phi4, rwkv-world, smolvlm, vicuna, vicuna-orca, yandex, zephyr
(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) | | `-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: 0.50, 0.0 = disabled)
| | `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) | | `--draft-max, --draft, --draft-n N` | number of tokens to draft for speculative decoding (default: 16)
(env: LLAMA_ARG_DRAFT_MAX) | -| `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 5)
(env: LLAMA_ARG_DRAFT_MIN) | -| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.9)
(env: LLAMA_ARG_DRAFT_P_MIN) | +| `--draft-min, --draft-n-min N` | minimum number of draft tokens to use for speculative decoding (default: 0)
(env: LLAMA_ARG_DRAFT_MIN) | +| `--draft-p-min P` | minimum speculative decoding probability (greedy) (default: 0.8)
(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)
(env: LLAMA_ARG_CTX_SIZE_DRAFT) | | `-devd, --device-draft ` | comma-separated list of devices to use for offloading the draft model (none = don't offload)
use --list-devices to see a list of available devices | | `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | number of layers to store in VRAM for the draft model
(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) | | `-md, --model-draft FNAME` | draft model for speculative decoding (default: unused)
(env: LLAMA_ARG_MODEL_DRAFT) | +| `-mv, --model-vocoder FNAME` | vocoder model for audio generation (default: unused) | +| `--tts-use-guide-tokens` | Use guide tokens to improve TTS word recall | +| `--embd-bge-small-en-default` | use default bge-small-en-v1.5 model (note: can download weights from the internet) | +| `--embd-e5-small-en-default` | use default e5-small-v2 model (note: can download weights from the internet) | +| `--embd-gte-small-default` | use default gte-small model (note: can download weights from the internet) | +| `--fim-qwen-1.5b-default` | use default Qwen 2.5 Coder 1.5B (note: can download weights from the internet) | +| `--fim-qwen-3b-default` | use default Qwen 2.5 Coder 3B (note: can download weights from the internet) | +| `--fim-qwen-7b-default` | use default Qwen 2.5 Coder 7B (note: can download weights from the internet) | +| `--fim-qwen-7b-spec` | use Qwen 2.5 Coder 7B + 0.5B draft for speculative decoding (note: can download weights from the internet) | +| `--fim-qwen-14b-spec` | use Qwen 2.5 Coder 14B + 0.5B draft for speculative decoding (note: can download weights from the internet) | Note: If both command line argument and environment variable are both set for the same param, the argument will take precedence over env var. @@ -193,6 +217,12 @@ services: LLAMA_ARG_PORT: 8080 ``` +### Multimodal support + +Multimodal support was added in [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) and is currently an experimental feature. + +For more details, please refer to [multimodal documentation](../../docs/multimodal.md) + ## Build `llama-server` is built alongside everything else from the root of the project @@ -232,7 +262,7 @@ To build or to run the dev server (with hot reload): ```sh # make sure you have nodejs installed -cd examples/server/webui +cd tools/server/webui npm i # to run the dev server @@ -242,7 +272,7 @@ npm run dev npm run build ``` After `public/index.html.gz` has been generated we need to generate the c++ -headers (like build/examples/server/index.html.gz.hpp) that will be included +headers (like build/tools/server/index.html.gz.hpp) that will be included by server.cpp. This is done by building `llama-server` as described in the [build](#build) section above. @@ -749,6 +779,9 @@ This endpoint is public (no API key check). By default, it is read-only. To make "total_slots": 1, "model_path": "../models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf", "chat_template": "...", + "modalities": { + "vision": false + }, "build_info": "b(build number)-(build commit hash)" } ``` @@ -757,6 +790,7 @@ This endpoint is public (no API key check). By default, it is read-only. To make - `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 +- `modalities` - the list of supported modalities ### POST `/props`: Change server global properties. @@ -1069,6 +1103,8 @@ print(completion.choices[0].text) Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used. +If model supports multimodal, you can input the media file via `image_url` content part. We support both base64 and remote URL as input. See OAI documentation for more. + *Options:* See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs/api-reference/chat). llama.cpp `/completion`-specific features such as `mirostat` are also supported. @@ -1228,12 +1264,12 @@ Apart from error types supported by OAI, we also have custom types that are spec ### Legacy completion web UI -A new chat-based UI has replaced the old completion-based since [this PR](https://github.com/ggml-org/llama.cpp/pull/10175). If you want to use the old completion, start the server with `--path ./examples/server/public_legacy` +A new chat-based UI has replaced the old completion-based since [this PR](https://github.com/ggml-org/llama.cpp/pull/10175). If you want to use the old completion, start the server with `--path ./tools/server/public_legacy` For example: ```sh -./llama-server -m my_model.gguf -c 8192 --path ./examples/server/public_legacy +./llama-server -m my_model.gguf -c 8192 --path ./tools/server/public_legacy ``` ### Extending or building alternative Web Front End diff --git a/examples/server/bench/README.md b/tools/server/bench/README.md similarity index 100% rename from examples/server/bench/README.md rename to tools/server/bench/README.md diff --git a/examples/server/bench/bench.py b/tools/server/bench/bench.py similarity index 100% rename from examples/server/bench/bench.py rename to tools/server/bench/bench.py diff --git a/examples/server/bench/prometheus.yml b/tools/server/bench/prometheus.yml similarity index 100% rename from examples/server/bench/prometheus.yml rename to tools/server/bench/prometheus.yml diff --git a/examples/server/bench/requirements.txt b/tools/server/bench/requirements.txt similarity index 100% rename from examples/server/bench/requirements.txt rename to tools/server/bench/requirements.txt diff --git a/examples/server/bench/script.js b/tools/server/bench/script.js similarity index 100% rename from examples/server/bench/script.js rename to tools/server/bench/script.js diff --git a/examples/server/chat-llama2.sh b/tools/server/chat-llama2.sh similarity index 100% rename from examples/server/chat-llama2.sh rename to tools/server/chat-llama2.sh diff --git a/examples/server/chat.mjs b/tools/server/chat.mjs similarity index 100% rename from examples/server/chat.mjs rename to tools/server/chat.mjs diff --git a/examples/server/chat.sh b/tools/server/chat.sh similarity index 100% rename from examples/server/chat.sh rename to tools/server/chat.sh diff --git a/examples/server/httplib.h b/tools/server/httplib.h similarity index 100% rename from examples/server/httplib.h rename to tools/server/httplib.h diff --git a/tools/server/public/index.html.gz b/tools/server/public/index.html.gz new file mode 100644 index 0000000000..d7363e13eb Binary files /dev/null and b/tools/server/public/index.html.gz differ diff --git a/examples/server/public/loading.html b/tools/server/public/loading.html similarity index 100% rename from examples/server/public/loading.html rename to tools/server/public/loading.html diff --git a/examples/server/public_legacy/colorthemes.css b/tools/server/public_legacy/colorthemes.css similarity index 100% rename from examples/server/public_legacy/colorthemes.css rename to tools/server/public_legacy/colorthemes.css diff --git a/examples/server/public_legacy/completion.js b/tools/server/public_legacy/completion.js similarity index 100% rename from examples/server/public_legacy/completion.js rename to tools/server/public_legacy/completion.js diff --git a/examples/server/public_legacy/favicon.ico b/tools/server/public_legacy/favicon.ico similarity index 100% rename from examples/server/public_legacy/favicon.ico rename to tools/server/public_legacy/favicon.ico diff --git a/examples/server/public_legacy/index-new.html b/tools/server/public_legacy/index-new.html similarity index 100% rename from examples/server/public_legacy/index-new.html rename to tools/server/public_legacy/index-new.html diff --git a/examples/server/public_legacy/index.html b/tools/server/public_legacy/index.html similarity index 100% rename from examples/server/public_legacy/index.html rename to tools/server/public_legacy/index.html diff --git a/examples/server/public_legacy/index.js b/tools/server/public_legacy/index.js similarity index 100% rename from examples/server/public_legacy/index.js rename to tools/server/public_legacy/index.js diff --git a/examples/server/public_legacy/json-schema-to-grammar.mjs b/tools/server/public_legacy/json-schema-to-grammar.mjs similarity index 99% rename from examples/server/public_legacy/json-schema-to-grammar.mjs rename to tools/server/public_legacy/json-schema-to-grammar.mjs index f767ce7b72..b12bf2ab09 100644 --- a/examples/server/public_legacy/json-schema-to-grammar.mjs +++ b/tools/server/public_legacy/json-schema-to-grammar.mjs @@ -2,6 +2,9 @@ const SPACE_RULE = '| " " | "\\n"{1,2} [ \\t]{0,20}'; function _buildRepetition(itemRule, minItems, maxItems, opts={}) { + if (maxItems == 0) { + return ''; + } if (minItems === 0 && maxItems === 1) { return `${itemRule}?`; } diff --git a/examples/server/public_legacy/loading.html b/tools/server/public_legacy/loading.html similarity index 100% rename from examples/server/public_legacy/loading.html rename to tools/server/public_legacy/loading.html diff --git a/examples/server/public_legacy/prompt-formats.js b/tools/server/public_legacy/prompt-formats.js similarity index 100% rename from examples/server/public_legacy/prompt-formats.js rename to tools/server/public_legacy/prompt-formats.js diff --git a/examples/server/public_legacy/style.css b/tools/server/public_legacy/style.css similarity index 100% rename from examples/server/public_legacy/style.css rename to tools/server/public_legacy/style.css diff --git a/examples/server/public_legacy/system-prompts.js b/tools/server/public_legacy/system-prompts.js similarity index 100% rename from examples/server/public_legacy/system-prompts.js rename to tools/server/public_legacy/system-prompts.js diff --git a/examples/server/public_legacy/theme-beeninorder.css b/tools/server/public_legacy/theme-beeninorder.css similarity index 100% rename from examples/server/public_legacy/theme-beeninorder.css rename to tools/server/public_legacy/theme-beeninorder.css diff --git a/examples/server/public_legacy/theme-ketivah.css b/tools/server/public_legacy/theme-ketivah.css similarity index 100% rename from examples/server/public_legacy/theme-ketivah.css rename to tools/server/public_legacy/theme-ketivah.css diff --git a/examples/server/public_legacy/theme-mangotango.css b/tools/server/public_legacy/theme-mangotango.css similarity index 100% rename from examples/server/public_legacy/theme-mangotango.css rename to tools/server/public_legacy/theme-mangotango.css diff --git a/examples/server/public_legacy/theme-playground.css b/tools/server/public_legacy/theme-playground.css similarity index 100% rename from examples/server/public_legacy/theme-playground.css rename to tools/server/public_legacy/theme-playground.css diff --git a/examples/server/public_legacy/theme-polarnight.css b/tools/server/public_legacy/theme-polarnight.css similarity index 100% rename from examples/server/public_legacy/theme-polarnight.css rename to tools/server/public_legacy/theme-polarnight.css diff --git a/examples/server/public_legacy/theme-snowstorm.css b/tools/server/public_legacy/theme-snowstorm.css similarity index 100% rename from examples/server/public_legacy/theme-snowstorm.css rename to tools/server/public_legacy/theme-snowstorm.css diff --git a/examples/server/public_simplechat/datautils.mjs b/tools/server/public_simplechat/datautils.mjs similarity index 100% rename from examples/server/public_simplechat/datautils.mjs rename to tools/server/public_simplechat/datautils.mjs diff --git a/examples/server/public_simplechat/index.html b/tools/server/public_simplechat/index.html similarity index 100% rename from examples/server/public_simplechat/index.html rename to tools/server/public_simplechat/index.html diff --git a/examples/server/public_simplechat/readme.md b/tools/server/public_simplechat/readme.md similarity index 97% rename from examples/server/public_simplechat/readme.md rename to tools/server/public_simplechat/readme.md index 21410199f6..24e026d455 100644 --- a/examples/server/public_simplechat/readme.md +++ b/tools/server/public_simplechat/readme.md @@ -7,7 +7,7 @@ by Humans for All. To run from the build dir -bin/llama-server -m path/model.gguf --path ../examples/server/public_simplechat +bin/llama-server -m path/model.gguf --path ../tools/server/public_simplechat Continue reading for the details. @@ -51,17 +51,17 @@ One could run this web frontend directly using server itself or if anyone is thi frontend to configure the server over http(s) or so, then run this web frontend using something like python's http module. -### running using examples/server +### running using tools/server -./llama-server -m path/model.gguf --path examples/server/public_simplechat [--port PORT] +./llama-server -m path/model.gguf --path tools/server/public_simplechat [--port PORT] ### running using python3's server module -first run examples/server +first run tools/server * ./llama-server -m path/model.gguf -next run this web front end in examples/server/public_simplechat -* cd ../examples/server/public_simplechat +next run this web front end in tools/server/public_simplechat +* cd ../tools/server/public_simplechat * python3 -m http.server PORT ### using the front end @@ -248,7 +248,7 @@ Set max_tokens to 1024, so that a relatively large previous reponse doesnt eat u available wrt next query-response. However dont forget that the server when started should also be started with a model context size of 1k or more, to be on safe side. - The /completions endpoint of examples/server doesnt take max_tokens, instead it takes the + The /completions endpoint of tools/server doesnt take max_tokens, instead it takes the internal n_predict, for now add the same here on the client side, maybe later add max_tokens to /completions endpoint handling code on server side. diff --git a/examples/server/public_simplechat/simplechat.css b/tools/server/public_simplechat/simplechat.css similarity index 100% rename from examples/server/public_simplechat/simplechat.css rename to tools/server/public_simplechat/simplechat.css diff --git a/examples/server/public_simplechat/simplechat.js b/tools/server/public_simplechat/simplechat.js similarity index 100% rename from examples/server/public_simplechat/simplechat.js rename to tools/server/public_simplechat/simplechat.js diff --git a/examples/server/public_simplechat/simplechat_screens.webp b/tools/server/public_simplechat/simplechat_screens.webp similarity index 100% rename from examples/server/public_simplechat/simplechat_screens.webp rename to tools/server/public_simplechat/simplechat_screens.webp diff --git a/examples/server/public_simplechat/ui.mjs b/tools/server/public_simplechat/ui.mjs similarity index 100% rename from examples/server/public_simplechat/ui.mjs rename to tools/server/public_simplechat/ui.mjs diff --git a/examples/server/server.cpp b/tools/server/server.cpp similarity index 93% rename from examples/server/server.cpp rename to tools/server/server.cpp index c580ec1232..7169ffdcee 100644 --- a/examples/server/server.cpp +++ b/tools/server/server.cpp @@ -7,6 +7,7 @@ #include "log.h" #include "sampling.h" #include "speculative.h" +#include "mtmd.h" // Change JSON_ASSERT from assert() to GGML_ASSERT: #define JSON_ASSERT GGML_ASSERT @@ -146,6 +147,7 @@ struct slot_params { {"top_k", sampling.top_k}, {"top_p", sampling.top_p}, {"min_p", sampling.min_p}, + {"top_n_sigma", sampling.top_n_sigma}, {"xtc_probability", sampling.xtc_probability}, {"xtc_threshold", sampling.xtc_threshold}, {"typical_p", sampling.typ_p}, @@ -196,8 +198,8 @@ struct server_task { int id_target = -1; // used by SERVER_TASK_TYPE_INFERENCE - slot_params params; - llama_tokens prompt_tokens; + slot_params params; + server_tokens prompt_tokens; int id_selected_slot = -1; // used by SERVER_TASK_TYPE_SLOT_SAVE, SERVER_TASK_TYPE_SLOT_RESTORE, SERVER_TASK_TYPE_SLOT_ERASE @@ -248,6 +250,7 @@ struct server_task { params.sampling.top_k = json_value(data, "top_k", defaults.sampling.top_k); params.sampling.top_p = json_value(data, "top_p", defaults.sampling.top_p); params.sampling.min_p = json_value(data, "min_p", defaults.sampling.min_p); + params.sampling.top_n_sigma = json_value(data, "top_n_sigma", defaults.sampling.top_n_sigma); params.sampling.xtc_probability = json_value(data, "xtc_probability", defaults.sampling.xtc_probability); params.sampling.xtc_threshold = json_value(data, "xtc_threshold", defaults.sampling.xtc_threshold); params.sampling.typ_p = json_value(data, "typical_p", defaults.sampling.typ_p); @@ -1246,6 +1249,9 @@ struct server_slot { llama_context * ctx = nullptr; llama_context * ctx_dft = nullptr; + // multimodal + mtmd_context * mctx = nullptr; + common_speculative * spec = nullptr; std::vector lora; @@ -1273,14 +1279,14 @@ struct server_slot { int32_t n_prompt_tokens_processed = 0; // input prompt tokens - llama_tokens prompt_tokens; + server_tokens prompt_tokens; size_t last_nl_pos = 0; std::string generated_text; llama_tokens generated_tokens; - llama_tokens cache_tokens; + server_tokens cache_tokens; std::vector generated_token_probs; @@ -1474,7 +1480,7 @@ struct server_slot { {"is_processing", is_processing()}, {"non_causal", is_non_causal()}, {"params", params.to_json()}, - {"prompt", common_detokenize(ctx, prompt_tokens)}, + {"prompt", prompt_tokens.detokenize(ctx, true)}, {"next_token", { {"has_next_token", has_next_token}, @@ -1847,13 +1853,16 @@ struct server_context { llama_model * model = nullptr; llama_context * ctx = nullptr; + // multimodal + mtmd_context * mctx = nullptr; + const llama_vocab * vocab = nullptr; llama_model * model_dft = nullptr; llama_context_params cparams_dft; - llama_batch batch = {}; + llama_batch batch {}; bool clean_kv_cache = true; bool add_bos_token = true; @@ -1876,6 +1885,8 @@ struct server_context { common_chat_templates_ptr chat_templates; ~server_context() { + mtmd_free(mctx); + // Clear any sampling context for (server_slot & slot : slots) { common_sampler_free(slot.smpl); @@ -1963,6 +1974,36 @@ struct server_context { chat_templates = common_chat_templates_init(model, "chatml"); } + std::string & mmproj_path = params_base.mmproj.path; + if (!mmproj_path.empty()) { + 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; + mparams.verbosity = params_base.verbosity > 0 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_INFO; + 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()); + return false; + } + SRV_INF("loaded multimodal model, '%s'\n", mmproj_path.c_str()); + + if (params_base.ctx_shift) { + params_base.ctx_shift = false; + SRV_WRN("%s\n", "ctx_shift is not supported by multimodal, it will be disabled"); + } + + if (params_base.n_cache_reuse) { + params_base.n_cache_reuse = 0; + SRV_WRN("%s\n", "cache_reuse is not supported by multimodal, it will be disabled"); + } + + if (!params_base.speculative.model.path.empty()) { + SRV_ERR("%s\n", "err: speculative decode is not supported by multimodal"); + return false; + } + } + return true; } @@ -1978,6 +2019,8 @@ struct server_context { slot.ctx = ctx; slot.n_ctx = n_ctx_slot; slot.n_predict = params_base.n_predict; + slot.mctx = mctx; + slot.cache_tokens.has_mtmd = mctx != nullptr; if (model_dft) { slot.batch_spec = llama_batch_init(params_base.speculative.n_max + 1, 0, 1); @@ -2014,8 +2057,6 @@ struct server_context { // note that n_batch can be > n_ctx (e.g. for non-causal attention models such as BERT where the KV cache is not used) { const int32_t n_batch = llama_n_batch(ctx); - - // only a single seq_id per token is needed batch = llama_batch_init(std::max(n_batch, params_base.n_parallel), 0, 1); } @@ -2052,7 +2093,7 @@ struct server_context { } // length of the Longest Common Subsequence between the current slot's prompt and the input prompt - int cur_lcs_len = common_lcs(slot.cache_tokens, task.prompt_tokens); + int cur_lcs_len = slot.cache_tokens.get_common_prefix(task.prompt_tokens); // fraction of the common subsequence length compared to the current slot's prompt length float cur_similarity = static_cast(cur_lcs_len) / static_cast(slot.cache_tokens.size()); @@ -2094,18 +2135,6 @@ struct server_context { return ret; } - bool can_be_detokenized(const struct llama_context * ctx, const std::vector & tokens) { - const llama_model * model = llama_get_model(ctx); - const llama_vocab * vocab = llama_model_get_vocab(model); - const int32_t n_vocab = llama_vocab_n_tokens(vocab); - for (const auto & token : tokens) { - if (token < 0 || token >= n_vocab) { - return false; - } - } - return true; - } - bool launch_slot_with_task(server_slot & slot, server_task && task) { slot.reset(); slot.id_task = task.id; @@ -2120,8 +2149,7 @@ struct server_context { slot.lora = slot.params.lora; } - bool can_detokenize = can_be_detokenized(ctx, slot.prompt_tokens); - if (!can_detokenize) { + if (!slot.prompt_tokens.validate(ctx)) { send_error(task, "Prompt contains invalid tokens", ERROR_TYPE_INVALID_REQUEST); return false; } @@ -2383,6 +2411,15 @@ struct server_context { queue_results.send(std::move(res)); } + // if multimodal is enabled, send an error and return false + bool ensure_no_mtmd(const int id_task) { + if (mctx) { + send_error(id_task, "This feature is not supported by multimodal", ERROR_TYPE_NOT_SUPPORTED); + return false; + } + return true; + } + void send_partial_response(server_slot & slot, const completion_token_output & tkn) { auto res = std::make_unique(); @@ -2422,7 +2459,7 @@ struct server_context { res->content = std::move(slot.generated_text); res->tokens = std::move(slot.generated_tokens); res->timings = slot.get_timings(); - res->prompt = common_detokenize(ctx, slot.prompt_tokens, true); + res->prompt = slot.prompt_tokens.detokenize(ctx, true); res->response_fields = std::move(slot.params.response_fields); res->truncated = slot.truncated; @@ -2732,6 +2769,10 @@ struct server_context { } break; case SERVER_TASK_TYPE_SLOT_SAVE: { + if (!ensure_no_mtmd(task.id)) { + break; + } + int id_slot = task.slot_action.slot_id; server_slot * slot = get_slot_by_id(id_slot); if (slot == nullptr) { @@ -2751,7 +2792,8 @@ struct server_context { std::string filename = task.slot_action.filename; std::string filepath = task.slot_action.filepath; - const size_t nwrite = llama_state_seq_save_file(ctx, filepath.c_str(), slot->id, slot->cache_tokens.data(), token_count); + const llama_tokens & tokens = slot->cache_tokens.get_text_tokens(); + const size_t nwrite = llama_state_seq_save_file(ctx, filepath.c_str(), slot->id, tokens.data(), token_count); const int64_t t_end = ggml_time_us(); const double t_save_ms = (t_end - t_start) / 1000.0; @@ -2768,6 +2810,7 @@ struct server_context { } break; case SERVER_TASK_TYPE_SLOT_RESTORE: { + if (!ensure_no_mtmd(task.id)) break; int id_slot = task.slot_action.slot_id; server_slot * slot = get_slot_by_id(id_slot); if (slot == nullptr) { @@ -2786,15 +2829,18 @@ struct server_context { std::string filename = task.slot_action.filename; std::string filepath = task.slot_action.filepath; - slot->cache_tokens.resize(slot->n_ctx); + llama_tokens tokens; + tokens.resize(slot->n_ctx); size_t token_count = 0; - size_t nread = llama_state_seq_load_file(ctx, filepath.c_str(), slot->id, slot->cache_tokens.data(), slot->cache_tokens.size(), &token_count); + size_t nread = llama_state_seq_load_file(ctx, filepath.c_str(), slot->id, tokens.data(), tokens.size(), &token_count); if (nread == 0) { - slot->cache_tokens.resize(0); + slot->cache_tokens.clear(); // KV may already been invalidated? send_error(task, "Unable to restore slot, no available space in KV cache or invalid slot save file", ERROR_TYPE_INVALID_REQUEST); break; } - slot->cache_tokens.resize(token_count); + tokens.resize(token_count); + slot->cache_tokens.clear(); + slot->cache_tokens.insert(tokens); const int64_t t_end = ggml_time_us(); const double t_restore_ms = (t_end - t_start) / 1000.0; @@ -2811,6 +2857,7 @@ struct server_context { } break; case SERVER_TASK_TYPE_SLOT_ERASE: { + if (!ensure_no_mtmd(task.id)) break; int id_slot = task.slot_action.slot_id; server_slot * slot = get_slot_by_id(id_slot); if (slot == nullptr) { @@ -2842,6 +2889,7 @@ struct server_context { res->id = task.id; queue_results.send(std::move(res)); } break; + } } @@ -2887,6 +2935,12 @@ struct server_context { continue; } + if (mctx) { + // we should never reach this because params_base.ctx_shift is automatically disabled if mmproj is loaded + // we don't support ctx_shift because an image chunk may contains multiple tokens + GGML_ABORT("not supported by multimodal"); + } + // Shift context const int n_keep = slot.params.n_keep + add_bos_token; const int n_left = slot.n_past - n_keep; @@ -2898,11 +2952,14 @@ struct server_context { llama_kv_self_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past, -n_discard); if (slot.params.cache_prompt) { - for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++) { - slot.cache_tokens[i - n_discard] = slot.cache_tokens[i]; + llama_tokens new_tokens = slot.cache_tokens.get_text_tokens(); // copy + for (size_t i = n_keep + n_discard; i < new_tokens.size(); i++) { + new_tokens[i - n_discard] = new_tokens[i]; } - slot.cache_tokens.resize(slot.cache_tokens.size() - n_discard); + new_tokens.resize(slot.cache_tokens.size() - n_discard); + slot.cache_tokens.clear(); + slot.cache_tokens.insert(new_tokens); } slot.n_past -= n_discard; @@ -2980,7 +3037,7 @@ struct server_context { SLT_INF(slot, "new prompt, n_ctx_slot = %d, n_keep = %d, n_prompt_tokens = %d\n", slot.n_ctx, slot.params.n_keep, slot.n_prompt_tokens); // print prompt tokens (for debugging) - if (1) { + /*if (1) { // first 16 tokens (avoid flooding logs) for (int i = 0; i < std::min(16, prompt_tokens.size()); i++) { SLT_DBG(slot, "prompt token %3d: %6d '%s'\n", i, prompt_tokens[i], common_token_to_piece(ctx, prompt_tokens[i]).c_str()); @@ -2990,7 +3047,7 @@ struct server_context { for (int i = 0; i < (int) prompt_tokens.size(); i++) { SLT_DBG(slot, "prompt token %3d: %6d '%s'\n", i, prompt_tokens[i], common_token_to_piece(ctx, prompt_tokens[i]).c_str()); } - } + }*/ // empty prompt passed -> release the slot and send empty response if (prompt_tokens.empty()) { @@ -3032,21 +3089,27 @@ struct server_context { // if input prompt is too big, truncate it if (slot.n_prompt_tokens >= slot.n_ctx) { + if (mctx) { + // we should never reach this + GGML_ABORT("not supported by multimodal"); + } const int n_left = slot.n_ctx - slot.params.n_keep; const int n_block_size = n_left / 2; const int erased_blocks = (slot.n_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size; + const llama_tokens & curr_tokens = slot.prompt_tokens.get_text_tokens(); llama_tokens new_tokens( - prompt_tokens.begin(), - prompt_tokens.begin() + slot.params.n_keep); + curr_tokens.begin(), + curr_tokens.begin() + slot.params.n_keep); new_tokens.insert( new_tokens.end(), - prompt_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size, - prompt_tokens.end()); + curr_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size, + curr_tokens.end()); - prompt_tokens = std::move(new_tokens); + prompt_tokens.clear(); + prompt_tokens.insert(new_tokens); slot.truncated = true; slot.n_prompt_tokens = prompt_tokens.size(); @@ -3058,13 +3121,18 @@ struct server_context { if (slot.params.cache_prompt) { // reuse any previously computed tokens that are common with the new prompt - slot.n_past = common_lcp(slot.cache_tokens, prompt_tokens); + slot.n_past = slot.cache_tokens.get_common_prefix(prompt_tokens); // reuse chunks from the cached prompt by shifting their KV cache in the new position if (params_base.n_cache_reuse > 0) { size_t head_c = slot.n_past; // cache size_t head_p = slot.n_past; // current prompt + if (mctx) { + // we should never reach this + GGML_ABORT("not supported by multimodal"); + } + SLT_DBG(slot, "trying to reuse chunks with size > %d, slot.n_past = %d\n", params_base.n_cache_reuse, slot.n_past); while (head_c < slot.cache_tokens.size() && @@ -3090,7 +3158,7 @@ struct server_context { llama_kv_self_seq_add(ctx, slot.id, head_c, head_c + n_match, kv_shift); for (size_t i = 0; i < n_match; i++) { - slot.cache_tokens[head_p + i] = slot.cache_tokens[head_c + i]; + slot.cache_tokens.set_token(head_p + i, slot.cache_tokens[head_c + i]); slot.n_past++; } @@ -3138,21 +3206,52 @@ struct server_context { // remove the non-common part from the cache slot.cache_tokens.resize(slot.n_past); + // check if we should process the image + if (slot.n_past < slot.n_prompt_tokens + && slot.prompt_tokens[slot.n_past] == LLAMA_TOKEN_NULL) { + // process the image + int32_t new_n_past; + int32_t res = slot.prompt_tokens.process_chunk(ctx, mctx, slot.n_past, slot.id, new_n_past); + int32_t n_pos = new_n_past - slot.n_past; + + if (res != 0) { + SLT_ERR(slot, "failed to process image, res = %d\n", res); + slot.release(); + send_error(slot, "failed to process image", ERROR_TYPE_SERVER); + continue; + } + + if (slot.params.cache_prompt) { + const auto & chunk = slot.prompt_tokens.find_chunk(slot.n_past); + slot.cache_tokens.push_back(chunk.get()); // copy + } + + slot.n_past += n_pos; + slot.n_prompt_tokens_processed += n_pos; + } + // add prompt tokens for processing in the current batch while (slot.n_past < slot.n_prompt_tokens && batch.n_tokens < n_batch) { + // get next token to process + llama_token cur_tok = slot.prompt_tokens[slot.n_past]; + if (cur_tok == LLAMA_TOKEN_NULL) { + break; // end of text chunk + } + // without pooling, we want to output the embeddings for all the tokens in the batch const bool need_embd = slot.task_type == SERVER_TASK_TYPE_EMBEDDING && llama_pooling_type(slot.ctx) == LLAMA_POOLING_TYPE_NONE; - common_batch_add(batch, prompt_tokens[slot.n_past], slot.n_past, { slot.id }, need_embd); - + common_batch_add(batch, cur_tok, slot.n_past, { slot.id }, need_embd); if (slot.params.cache_prompt) { - slot.cache_tokens.push_back(prompt_tokens[slot.n_past]); + slot.cache_tokens.push_back(cur_tok); } slot.n_prompt_tokens_processed++; slot.n_past++; } + // SLT_INF(slot, "new cache_tokens: %s\n", slot.cache_tokens.str().c_str()); + SLT_INF(slot, "prompt processing progress, n_past = %d, n_tokens = %d, progress = %f\n", slot.n_past, batch.n_tokens, (float) slot.n_prompt_tokens_processed / slot.n_prompt_tokens); // entire prompt has been processed @@ -3160,12 +3259,16 @@ struct server_context { slot.state = SLOT_STATE_DONE_PROMPT; GGML_ASSERT(batch.n_tokens > 0); + GGML_ASSERT((size_t) slot.n_prompt_tokens == slot.prompt_tokens.size()); common_sampler_reset(slot.smpl); // Process all prompt tokens through sampler system for (int i = 0; i < slot.n_prompt_tokens; ++i) { - common_sampler_accept(slot.smpl, prompt_tokens[i], false); + llama_token id = slot.prompt_tokens[i]; + if (id != LLAMA_TOKEN_NULL) { + common_sampler_accept(slot.smpl, id, false); + } } // extract the logits only for the last token @@ -3212,7 +3315,14 @@ struct server_context { batch.logits + i, }; - const int ret = llama_decode(ctx, batch_view); + int ret = 0; + + if (params_base.embedding || params_base.reranking) { + ret = llama_encode(ctx, batch_view); + } else { + ret = llama_decode(ctx, batch_view); + } + metrics.on_decoded(slots); if (ret != 0) { @@ -3311,6 +3421,11 @@ struct server_context { continue; } + if (mctx) { + // we should never reach this, as speculative is automatically disabled if mmproj is loaded + GGML_ABORT("not supported by multimodal"); + } + // determine the max draft that fits the current slot state int n_draft_max = slot.params.speculative.n_max; @@ -3337,7 +3452,8 @@ struct server_context { params_spec.n_reuse = llama_n_ctx(slot.ctx_dft) - slot.params.speculative.n_max; params_spec.p_min = slot.params.speculative.p_min; - llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, slot.cache_tokens, id); + const llama_tokens & cached_text_tokens = slot.cache_tokens.get_text_tokens(); + llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, cached_text_tokens, id); // keep track of total number of tokens generated in the draft slot.n_draft_total += draft.size(); @@ -3371,7 +3487,7 @@ struct server_context { slot.n_draft_accepted += ids.size() - 1; slot.cache_tokens.push_back(id); - slot.cache_tokens.insert(slot.cache_tokens.end(), ids.begin(), ids.end() - 1); + slot.cache_tokens.insert({ids.begin(), ids.end() - 1}); llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1); @@ -3894,6 +4010,7 @@ int main(int argc, char ** argv) { { "default_generation_settings", ctx_server.default_generation_settings_for_props }, { "total_slots", ctx_server.params_base.n_parallel }, { "model_path", ctx_server.params_base.model.path }, + { "modalities", json{{"vision", ctx_server.mctx != nullptr}} }, // TODO: add more in the future { "chat_template", common_chat_templates_source(ctx_server.chat_templates.get()) }, { "bos_token", common_token_to_piece(ctx_server.ctx, llama_vocab_bos(ctx_server.vocab), /* special= */ true)}, { "eos_token", common_token_to_piece(ctx_server.ctx, llama_vocab_eos(ctx_server.vocab), /* special= */ true)}, @@ -3941,9 +4058,10 @@ int main(int argc, char ** argv) { const auto handle_completions_impl = [&ctx_server, &res_error, &res_ok]( server_task_type type, json & data, - std::function is_connection_closed, + const std::vector & files, + const std::function & is_connection_closed, httplib::Response & res, - oaicompat_type oaicompat) { + oaicompat_type oaicompat) -> void { GGML_ASSERT(type == SERVER_TASK_TYPE_COMPLETION || type == SERVER_TASK_TYPE_INFILL); if (ctx_server.params_base.embedding) { @@ -3960,15 +4078,69 @@ int main(int argc, char ** argv) { // TODO: this log can become very long, put it behind a flag or think about a more compact format //SRV_DBG("Prompt: %s\n", prompt.is_string() ? prompt.get().c_str() : prompt.dump(2).c_str()); - std::vector tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true); - tasks.reserve(tokenized_prompts.size()); - for (size_t i = 0; i < tokenized_prompts.size(); i++) { + // process files + mtmd::bitmaps bitmaps; + const bool has_mtmd = ctx_server.mctx != nullptr; + { + if (!has_mtmd && !files.empty()) { + throw std::runtime_error("This server does not support multimodal"); + } + for (auto & file : files) { + mtmd::bitmap bmp(mtmd_helper_bitmap_init_from_buf(file.data(), file.size())); + if (!bmp.ptr) { + throw std::runtime_error("Failed to load image"); + } + // calculate bitmap hash (for KV caching) + std::string hash = fnv_hash(bmp.data(), bmp.nx()*bmp.ny()*3); + bmp.set_id(hash.c_str()); + bitmaps.entries.push_back(std::move(bmp)); + } + } + + // process prompt + std::vector inputs; + if (oaicompat && !prompt.is_string()) { + throw std::runtime_error("prompt must be a string"); + } + + if (oaicompat && has_mtmd) { + // multimodal + std::string prompt_str = prompt.get(); + mtmd_input_text inp_txt = { + prompt_str.c_str(), + /* add_special */ true, + /* parse_special */ true, + }; + mtmd::input_chunks chunks(mtmd_input_chunks_init()); + auto bitmaps_c_ptr = bitmaps.c_ptr(); + int32_t tokenized = mtmd_tokenize(ctx_server.mctx, + chunks.ptr.get(), + &inp_txt, + bitmaps_c_ptr.data(), + bitmaps_c_ptr.size()); + if (tokenized != 0) { + throw std::runtime_error("Failed to tokenize prompt"); + } + + server_tokens tmp(chunks, true); + inputs.push_back(std::move(tmp)); + } else { + // non-multimodal version + auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true); + for (auto & p : tokenized_prompts) { + auto tmp = server_tokens(p, ctx_server.mctx != nullptr); + inputs.push_back(std::move(tmp)); + } + } + + tasks.reserve(inputs.size()); + for (size_t i = 0; i < inputs.size(); i++) { server_task task = server_task(type); task.id = ctx_server.queue_tasks.get_new_id(); task.index = i; - task.prompt_tokens = std::move(tokenized_prompts[i]); + task.prompt_tokens = std::move(inputs[i]); task.params = server_task::params_from_json_cmpl( ctx_server.ctx, ctx_server.params_base, @@ -4050,9 +4222,11 @@ int main(int argc, char ** argv) { const auto handle_completions = [&handle_completions_impl](const httplib::Request & req, httplib::Response & res) { json data = json::parse(req.body); - return handle_completions_impl( + std::vector files; // dummy + handle_completions_impl( SERVER_TASK_TYPE_COMPLETION, data, + files, req.is_connection_closed, res, OAICOMPAT_TYPE_NONE); @@ -4060,9 +4234,11 @@ int main(int argc, char ** argv) { const auto handle_completions_oai = [&handle_completions_impl](const httplib::Request & req, httplib::Response & res) { json data = oaicompat_completion_params_parse(json::parse(req.body)); - return handle_completions_impl( + std::vector files; // dummy + handle_completions_impl( SERVER_TASK_TYPE_COMPLETION, data, + files, req.is_connection_closed, res, OAICOMPAT_TYPE_COMPLETION); @@ -4137,9 +4313,11 @@ int main(int argc, char ** argv) { tokenized_prompts[0] ); - return handle_completions_impl( + std::vector files; // dummy + handle_completions_impl( SERVER_TASK_TYPE_INFILL, data, + files, req.is_connection_closed, res, OAICOMPAT_TYPE_NONE); // infill is not OAI compatible @@ -4153,11 +4331,19 @@ int main(int argc, char ** argv) { } auto body = json::parse(req.body); - json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates.get()); + std::vector files; + json data = oaicompat_completion_params_parse( + body, + params.use_jinja, + params.reasoning_format, + ctx_server.chat_templates.get(), + ctx_server.mctx, + files); - return handle_completions_impl( + handle_completions_impl( SERVER_TASK_TYPE_COMPLETION, data, + files, req.is_connection_closed, res, OAICOMPAT_TYPE_CHAT); @@ -4166,7 +4352,14 @@ int main(int argc, char ** argv) { // same with handle_chat_completions, but without inference part const auto handle_apply_template = [&ctx_server, ¶ms, &res_ok](const httplib::Request & req, httplib::Response & res) { auto body = json::parse(req.body); - json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates.get()); + std::vector files; // dummy, unused + json data = oaicompat_completion_params_parse( + body, + params.use_jinja, + params.reasoning_format, + ctx_server.chat_templates.get(), + ctx_server.mctx, + files); res_ok(res, {{ "prompt", std::move(data.at("prompt")) }}); }; @@ -4271,7 +4464,7 @@ int main(int argc, char ** argv) { } } - std::vector tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true); + auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true); for (const auto & tokens : tokenized_prompts) { // this check is necessary for models that do not add BOS token to the input if (tokens.empty()) { @@ -4291,7 +4484,7 @@ int main(int argc, char ** argv) { task.id = ctx_server.queue_tasks.get_new_id(); task.index = i; - task.prompt_tokens = std::move(tokenized_prompts[i]); + task.prompt_tokens = server_tokens(tokenized_prompts[i], ctx_server.mctx != nullptr); // OAI-compat task.params.oaicompat = oaicompat; @@ -4385,13 +4578,14 @@ int main(int argc, char ** argv) { std::unordered_set task_ids; { std::vector tasks; - std::vector tokenized_docs = tokenize_input_prompts(ctx_server.vocab, documents, /* add_special */ false, true); + auto tokenized_docs = tokenize_input_prompts(ctx_server.vocab, documents, /* add_special */ false, true); tasks.reserve(tokenized_docs.size()); for (size_t i = 0; i < tokenized_docs.size(); i++) { + auto tmp = format_rerank(ctx_server.vocab, tokenized_query, tokenized_docs[i]); server_task task = server_task(SERVER_TASK_TYPE_RERANK); task.id = ctx_server.queue_tasks.get_new_id(); task.index = i; - task.prompt_tokens = format_rerank(ctx_server.vocab, tokenized_query, tokenized_docs[i]); + task.prompt_tokens = server_tokens(tmp, ctx_server.mctx != nullptr); tasks.push_back(std::move(task)); } diff --git a/examples/server/tests/.gitignore b/tools/server/tests/.gitignore similarity index 100% rename from examples/server/tests/.gitignore rename to tools/server/tests/.gitignore diff --git a/examples/server/tests/README.md b/tools/server/tests/README.md similarity index 96% rename from examples/server/tests/README.md rename to tools/server/tests/README.md index 652dea0382..cb87db035e 100644 --- a/examples/server/tests/README.md +++ b/tools/server/tests/README.md @@ -60,7 +60,7 @@ To run a single test: Hint: You can compile and run test in single command, useful for local developement: ```shell -cmake --build build -j --target llama-server && ./examples/server/tests/tests.sh +cmake --build build -j --target llama-server && ./tools/server/tests/tests.sh ``` To see all available arguments, please refer to [pytest documentation](https://docs.pytest.org/en/stable/how-to/usage.html) diff --git a/examples/server/tests/conftest.py b/tools/server/tests/conftest.py similarity index 100% rename from examples/server/tests/conftest.py rename to tools/server/tests/conftest.py diff --git a/examples/server/tests/pytest.ini b/tools/server/tests/pytest.ini similarity index 100% rename from examples/server/tests/pytest.ini rename to tools/server/tests/pytest.ini diff --git a/examples/server/tests/requirements.txt b/tools/server/tests/requirements.txt similarity index 100% rename from examples/server/tests/requirements.txt rename to tools/server/tests/requirements.txt diff --git a/examples/server/tests/tests.sh b/tools/server/tests/tests.sh similarity index 100% rename from examples/server/tests/tests.sh rename to tools/server/tests/tests.sh diff --git a/examples/server/tests/unit/test_basic.py b/tools/server/tests/unit/test_basic.py similarity index 100% rename from examples/server/tests/unit/test_basic.py rename to tools/server/tests/unit/test_basic.py diff --git a/examples/server/tests/unit/test_chat_completion.py b/tools/server/tests/unit/test_chat_completion.py similarity index 100% rename from examples/server/tests/unit/test_chat_completion.py rename to tools/server/tests/unit/test_chat_completion.py diff --git a/examples/server/tests/unit/test_completion.py b/tools/server/tests/unit/test_completion.py similarity index 100% rename from examples/server/tests/unit/test_completion.py rename to tools/server/tests/unit/test_completion.py diff --git a/examples/server/tests/unit/test_ctx_shift.py b/tools/server/tests/unit/test_ctx_shift.py similarity index 100% rename from examples/server/tests/unit/test_ctx_shift.py rename to tools/server/tests/unit/test_ctx_shift.py diff --git a/examples/server/tests/unit/test_embedding.py b/tools/server/tests/unit/test_embedding.py similarity index 100% rename from examples/server/tests/unit/test_embedding.py rename to tools/server/tests/unit/test_embedding.py diff --git a/examples/server/tests/unit/test_infill.py b/tools/server/tests/unit/test_infill.py similarity index 100% rename from examples/server/tests/unit/test_infill.py rename to tools/server/tests/unit/test_infill.py diff --git a/examples/server/tests/unit/test_lora.py b/tools/server/tests/unit/test_lora.py similarity index 100% rename from examples/server/tests/unit/test_lora.py rename to tools/server/tests/unit/test_lora.py diff --git a/examples/server/tests/unit/test_rerank.py b/tools/server/tests/unit/test_rerank.py similarity index 100% rename from examples/server/tests/unit/test_rerank.py rename to tools/server/tests/unit/test_rerank.py diff --git a/examples/server/tests/unit/test_security.py b/tools/server/tests/unit/test_security.py similarity index 100% rename from examples/server/tests/unit/test_security.py rename to tools/server/tests/unit/test_security.py diff --git a/examples/server/tests/unit/test_slot_save.py b/tools/server/tests/unit/test_slot_save.py similarity index 100% rename from examples/server/tests/unit/test_slot_save.py rename to tools/server/tests/unit/test_slot_save.py diff --git a/examples/server/tests/unit/test_speculative.py b/tools/server/tests/unit/test_speculative.py similarity index 100% rename from examples/server/tests/unit/test_speculative.py rename to tools/server/tests/unit/test_speculative.py diff --git a/examples/server/tests/unit/test_tokenize.py b/tools/server/tests/unit/test_tokenize.py similarity index 100% rename from examples/server/tests/unit/test_tokenize.py rename to tools/server/tests/unit/test_tokenize.py diff --git a/examples/server/tests/unit/test_tool_call.py b/tools/server/tests/unit/test_tool_call.py similarity index 100% rename from examples/server/tests/unit/test_tool_call.py rename to tools/server/tests/unit/test_tool_call.py diff --git a/tools/server/tests/unit/test_vision_api.py b/tools/server/tests/unit/test_vision_api.py new file mode 100644 index 0000000000..7cc4096f19 --- /dev/null +++ b/tools/server/tests/unit/test_vision_api.py @@ -0,0 +1,59 @@ +import pytest +from utils import * +import base64 +import requests + +server: ServerProcess + +IMG_URL_0 = "https://huggingface.co/ggml-org/tinygemma3-GGUF/resolve/main/test/11_truck.png" +IMG_URL_1 = "https://huggingface.co/ggml-org/tinygemma3-GGUF/resolve/main/test/91_cat.png" + +response = requests.get(IMG_URL_0) +response.raise_for_status() # Raise an exception for bad status codes +IMG_BASE64_0 = "data:image/png;base64," + base64.b64encode(response.content).decode("utf-8") + + +@pytest.fixture(autouse=True) +def create_server(): + global server + server = ServerPreset.tinygemma3() + + +@pytest.mark.parametrize( + "prompt, image_url, success, re_content", + [ + # test model is trained on CIFAR-10, but it's quite dumb due to small size + ("What is this:\n", IMG_URL_0, True, "(cat)+"), + ("What is this:\n", "IMG_BASE64_0", True, "(cat)+"), # exceptional, so that we don't cog up the log + ("What is this:\n", IMG_URL_1, True, "(frog)+"), + ("Test test\n", IMG_URL_1, True, "(frog)+"), # test invalidate cache + ("What is this:\n", "malformed", False, None), + ("What is this:\n", "https://google.com/404", False, None), # non-existent image + ("What is this:\n", "https://ggml.ai", False, None), # non-image data + ] +) +def test_vision_chat_completion(prompt, image_url, success, re_content): + global server + server.start(timeout_seconds=60) # vision model may take longer to load due to download size + if image_url == "IMG_BASE64_0": + image_url = IMG_BASE64_0 + res = server.make_request("POST", "/chat/completions", data={ + "temperature": 0.0, + "top_k": 1, + "messages": [ + {"role": "user", "content": [ + {"type": "text", "text": prompt}, + {"type": "image_url", "image_url": { + "url": image_url, + }}, + ]}, + ], + }) + if success: + assert res.status_code == 200 + choice = res.body["choices"][0] + assert "assistant" == choice["message"]["role"] + assert match_regex(re_content, choice["message"]["content"]) + else: + assert res.status_code != 200 + diff --git a/examples/server/tests/utils.py b/tools/server/tests/utils.py similarity index 95% rename from examples/server/tests/utils.py rename to tools/server/tests/utils.py index 4dc2062a8e..27a0f0356a 100644 --- a/examples/server/tests/utils.py +++ b/tools/server/tests/utils.py @@ -88,6 +88,7 @@ class ServerProcess: chat_template: str | None = None chat_template_file: str | None = None server_path: str | None = None + mmproj_url: str | None = None # session variables process: subprocess.Popen | None = None @@ -194,6 +195,8 @@ class ServerProcess: server_args.extend(["--chat-template", self.chat_template]) if self.chat_template_file: server_args.extend(["--chat-template-file", self.chat_template_file]) + if self.mmproj_url: + server_args.extend(["--mmproj-url", self.mmproj_url]) args = [str(arg) for arg in [server_path, *server_args]] print(f"tests: starting server with: {' '.join(args)}") @@ -379,6 +382,21 @@ class ServerPreset: server.server_reranking = True return server + @staticmethod + def tinygemma3() -> ServerProcess: + server = ServerProcess() + # mmproj is already provided by HF registry API + server.model_hf_repo = "ggml-org/tinygemma3-GGUF" + server.model_hf_file = "tinygemma3-Q8_0.gguf" + server.mmproj_url = "https://huggingface.co/ggml-org/tinygemma3-GGUF/resolve/main/mmproj-tinygemma3.gguf" + server.model_alias = "tinygemma3" + server.n_ctx = 1024 + server.n_batch = 32 + server.n_slots = 2 + server.n_predict = 4 + server.seed = 42 + return server + def parallel_function_calls(function_list: List[Tuple[Callable[..., Any], Tuple[Any, ...]]]) -> List[Any]: """ diff --git a/examples/server/themes/README.md b/tools/server/themes/README.md similarity index 100% rename from examples/server/themes/README.md rename to tools/server/themes/README.md diff --git a/examples/server/themes/buttons-top/README.md b/tools/server/themes/buttons-top/README.md similarity index 100% rename from examples/server/themes/buttons-top/README.md rename to tools/server/themes/buttons-top/README.md diff --git a/examples/server/themes/buttons-top/buttons_top.png b/tools/server/themes/buttons-top/buttons_top.png similarity index 100% rename from examples/server/themes/buttons-top/buttons_top.png rename to tools/server/themes/buttons-top/buttons_top.png diff --git a/examples/server/themes/buttons-top/favicon.ico b/tools/server/themes/buttons-top/favicon.ico similarity index 100% rename from examples/server/themes/buttons-top/favicon.ico rename to tools/server/themes/buttons-top/favicon.ico diff --git a/examples/server/themes/buttons-top/index.html b/tools/server/themes/buttons-top/index.html similarity index 100% rename from examples/server/themes/buttons-top/index.html rename to tools/server/themes/buttons-top/index.html diff --git a/examples/server/themes/wild/README.md b/tools/server/themes/wild/README.md similarity index 100% rename from examples/server/themes/wild/README.md rename to tools/server/themes/wild/README.md diff --git a/examples/server/themes/wild/favicon.ico b/tools/server/themes/wild/favicon.ico similarity index 100% rename from examples/server/themes/wild/favicon.ico rename to tools/server/themes/wild/favicon.ico diff --git a/examples/server/themes/wild/index.html b/tools/server/themes/wild/index.html similarity index 100% rename from examples/server/themes/wild/index.html rename to tools/server/themes/wild/index.html diff --git a/examples/server/themes/wild/llama_cpp.png b/tools/server/themes/wild/llama_cpp.png similarity index 100% rename from examples/server/themes/wild/llama_cpp.png rename to tools/server/themes/wild/llama_cpp.png diff --git a/examples/server/themes/wild/llamapattern.png b/tools/server/themes/wild/llamapattern.png similarity index 100% rename from examples/server/themes/wild/llamapattern.png rename to tools/server/themes/wild/llamapattern.png diff --git a/examples/server/themes/wild/wild.png b/tools/server/themes/wild/wild.png similarity index 100% rename from examples/server/themes/wild/wild.png rename to tools/server/themes/wild/wild.png diff --git a/examples/server/utils.hpp b/tools/server/utils.hpp similarity index 68% rename from examples/server/utils.hpp rename to tools/server/utils.hpp index aba2f27f9b..b8d140e3f0 100644 --- a/examples/server/utils.hpp +++ b/tools/server/utils.hpp @@ -3,7 +3,9 @@ #include "common.h" #include "log.h" #include "llama.h" +#include "arg.h" // common_remote_get_content #include "base64.hpp" +#include "mtmd.h" // increase max payload length to allow use of larger context size #define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 1048576 @@ -21,6 +23,7 @@ #include #include #include +#include #define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo" @@ -41,6 +44,8 @@ using json = nlohmann::ordered_json; #define QUE_ERR(fmt, ...) LOG_ERR("que %12.*s: " fmt, 12, __func__, __VA_ARGS__) #define QUE_DBG(fmt, ...) LOG_DBG("que %12.*s: " fmt, 12, __func__, __VA_ARGS__) +using raw_buffer = std::vector; + template static T json_value(const json & body, const std::string & key, const T & default_value) { // Fallback null to default value @@ -386,7 +391,7 @@ static inline bool is_base64(uint8_t c) { return (isalnum(c) || (c == '+') || (c == '/')); } -static inline std::vector base64_decode(const std::string & encoded_string) { +static inline raw_buffer base64_decode(const std::string & encoded_string) { int i = 0; int j = 0; int in_ = 0; @@ -396,7 +401,7 @@ static inline std::vector base64_decode(const std::string & encoded_str uint8_t char_array_4[4]; uint8_t char_array_3[3]; - std::vector ret; + raw_buffer ret; while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_])) { char_array_4[i++] = encoded_string[in_]; in_++; @@ -579,7 +584,9 @@ static json oaicompat_completion_params_parse( const json & body, /* openai api json semantics */ bool use_jinja, common_reasoning_format reasoning_format, - const struct common_chat_templates * tmpls) + const struct common_chat_templates * tmpls, + bool allow_non_text, + std::vector & out_files) { json llama_params; @@ -627,8 +634,77 @@ static json oaicompat_completion_params_parse( } } + // get input files + if (!body.contains("messages")) { + throw std::runtime_error("'messages' is required"); + } + json messages = body.at("messages"); + if (!messages.is_array()) { + throw std::runtime_error("Expected 'messages' to be an array"); + } + for (auto & msg : messages) { + json & content = msg.at("content"); + if (content.is_string() || content.is_null()) { + continue; + } + + if (!content.is_array()) { + throw std::runtime_error("Expected 'content' to be a string or an array"); + } + + for (auto & p : content) { + std::string type = json_value(p, "type", std::string()); + json image_url = json_value(p, "image_url", json::object()); + if (type == "image_url") { + if (!allow_non_text) { + throw std::runtime_error("image input is not supported by this server"); + } + + std::string url = json_value(image_url, "url", std::string()); + if (string_starts_with(url, "http")) { + // 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()); + auto res = common_remote_get_content(url, params); + if (200 <= res.first && res.first < 300) { + SRV_INF("downloaded %ld bytes\n", res.second.size()); + raw_buffer data; + data.insert(data.end(), res.second.begin(), res.second.end()); + out_files.push_back(data); + } else { + throw std::runtime_error("Failed to download image"); + } + + } else { + // try to decode base64 image + std::vector parts = string_split(url, /*separator*/ ','); + if (parts.size() != 2) { + throw std::runtime_error("Invalid image_url.url value"); + } else if (!string_starts_with(parts[0], "data:image/")) { + throw std::runtime_error("Invalid image_url.url format: " + parts[0]); + } else if (!string_ends_with(parts[0], "base64")) { + throw std::runtime_error("image_url.url must be base64 encoded"); + } else { + auto base64_data = parts[1]; + auto decoded_data = base64_decode(base64_data); + out_files.push_back(decoded_data); + } + } + + // replace this chunk with a marker + p["type"] = "text"; + p["text"] = MTMD_DEFAULT_IMAGE_MARKER; + p.erase("image_url"); + } + } + } + common_chat_templates_inputs inputs; - inputs.messages = common_chat_msgs_parse_oaicompat(body.at("messages")); + inputs.messages = common_chat_msgs_parse_oaicompat(messages); inputs.tools = common_chat_tools_parse_oaicompat(tools); inputs.tool_choice = common_chat_tool_choice_parse_oaicompat(json_value(body, "tool_choice", std::string("auto"))); inputs.json_schema = json_schema.is_null() ? "" : json_schema.dump(); @@ -642,9 +718,31 @@ static json oaicompat_completion_params_parse( throw std::runtime_error("Cannot use custom grammar constraints with tools."); } + // if the assistant message appears at the end of list, we do not add end-of-turn token + // for ex. this can be useful to modify the reasoning process in reasoning models + bool prefill_assistant_message = !inputs.messages.empty() && inputs.messages.back().role == "assistant"; + common_chat_msg last_message; + if (prefill_assistant_message) { + last_message = inputs.messages.back(); + inputs.messages.pop_back(); + + /* sanity check, max one assistant message at the end of the list */ + if (!inputs.messages.empty() && inputs.messages.back().role == "assistant"){ + throw std::runtime_error("Cannot have 2 or more assistant messages at the end of the list."); + } + + inputs.extract_reasoning = false; + inputs.add_generation_prompt = true; + } + // Apply chat template to the list of messages auto chat_params = common_chat_templates_apply(tmpls, inputs); + /* Append assistant prefilled message */ + if (prefill_assistant_message) { + chat_params.prompt += last_message.content; + } + llama_params["chat_format"] = static_cast(chat_params.format); llama_params["prompt"] = chat_params.prompt; if (!chat_params.grammar.empty()) { @@ -913,3 +1011,286 @@ static std::vector parse_lora_request( return lora; } + +// +// utils for interacting with libmtmd +// (may need to refactor in near future) +// + +/** + * server_tokens is a helper to manage the input tokens and image for the server. + * it is made this way to simplify the logic of KV cache management. + */ +struct server_tokens { + bool has_mtmd = false; + +private: // disallow accessing these members directly, risking out-of-sync + + // map a **start** position in tokens to the image chunk + std::unordered_map map_pos_to_image; + + // list of tokens + // it can include LLAMA_TOKEN_NULL, which is used to indicate a token that is not a text token + // a mtmd_input_chunk can occupy multiple tokens, one llama_token per **position** + // important: for models using mrope, an image can contain multiple tokens but will use only one **position** + llama_tokens tokens; + + // for ex. with input of 5 text tokens and 2 images: + // [0] [1] [2] [3] [4] [img0] [img0] [img0] [img1] [img1] + // pos 0 1 2 3 4 5 6 7 8 9 + // map_pos_to_image will contain: {5, img0}, {8, img1} + +public: + server_tokens() = default; + ~server_tokens() = default; + + // Prevent copying + server_tokens(const server_tokens&) = delete; + server_tokens& operator=(const server_tokens&) = delete; + + // Allow moving (usually implicitly generated if members are movable) + server_tokens(server_tokens&&) = default; + server_tokens& operator=(server_tokens&&) = default; + + // Allow accessing elements using [] operator + llama_token operator[](size_t index) { return tokens[index]; } + const llama_token& operator[](size_t index) const { return tokens[index]; } + + server_tokens(mtmd::input_chunks & mtmd_chunks, bool has_mtmd) : has_mtmd(has_mtmd) { + for (size_t i = 0; i < mtmd_chunks.size(); ++i) { + push_back(mtmd_chunks[i]); + } + } + + server_tokens(llama_tokens & tokens, bool has_mtmd) : has_mtmd(has_mtmd), tokens(tokens) {} + + // for debugging + std::string str() const { + std::ostringstream oss; + oss << "tokens: "; + for (const auto & t : tokens) { + if (t == LLAMA_TOKEN_NULL) { + oss << " "; + } else { + oss << t << " "; + } + } + oss << "\n"; + oss << "image pos: "; + for (const auto & it : map_pos_to_image) { + oss << it.first << ", "; + } + return oss.str(); + } + + const mtmd::input_chunk_ptr & find_chunk(llama_pos pos) const { + auto it = map_pos_to_image.find(pos); + if (it != map_pos_to_image.end()) { + return it->second; + } else { + throw std::runtime_error("Chunk not found"); + } + } + + void push_back(llama_token tok) { + if (tok == LLAMA_TOKEN_NULL) { + throw std::runtime_error("Invalid token"); + } + tokens.emplace_back(tok); + } + + // will create a copy of the chunk if it contains non-text data + void push_back(const mtmd_input_chunk * chunk) { + auto type = mtmd_input_chunk_get_type(chunk); + if (type == MTMD_INPUT_CHUNK_TYPE_IMAGE) { + GGML_ASSERT(has_mtmd); + auto img_tokens = mtmd_input_chunk_get_tokens_image(chunk); + const int n_pos = mtmd_image_tokens_get_n_pos(img_tokens); + llama_pos start_pos = tokens.size(); + for (int i = 0; i < n_pos; ++i) { + tokens.emplace_back(LLAMA_TOKEN_NULL); + } + mtmd::input_chunk_ptr new_chunk(mtmd_input_chunk_copy(chunk)); + map_pos_to_image[start_pos] = std::move(new_chunk); + } else if (type == MTMD_INPUT_CHUNK_TYPE_TEXT) { + size_t n_tokens; + auto text_tokens = mtmd_input_chunk_get_tokens_text(chunk, &n_tokens); + for (size_t i = 0; i < n_tokens; ++i) { + push_back(text_tokens[i]); + } + } else { + GGML_ABORT("Invalid chunk type"); + } + } + + // for compatibility with context shift and prompt truncation + void insert(const llama_tokens & inp_tokens) { + GGML_ASSERT(!has_mtmd); // only allow this if mtmd is disabled + tokens.insert(tokens.end(), inp_tokens.begin(), inp_tokens.end()); + } + + // for compatibility with speculative decoding, ctx shift, slot save/load + const llama_tokens & get_text_tokens() const { + GGML_ASSERT(!has_mtmd); // only allow this if mtmd is disabled + return tokens; + } + + // for compatibility with speculative decoding + void set_token(llama_pos pos, llama_token id) { + GGML_ASSERT(!has_mtmd); // only allow this if mtmd is disabled + tokens[pos] = id; + } + + size_t size() const { + return tokens.size(); + } + + bool empty() const { + return tokens.empty(); + } + + void clear() { + tokens.clear(); + } + + void resize(size_t n) { + GGML_ASSERT(n <= tokens.size()); + if (has_mtmd) { + // we throw an error if we try to remove a token in the middle of an image + // for ex. with input of 5 text tokens and 2 images: + // [0] [1] [2] [3] [4] [img0] [img0] [img0] [img1] [img1] + // n 1 2 3 4 5 6 7 8 9 10 + // allowed to resize ^ ^ + // disallowed to resize ^ ^ ^ + if (n > 0) { + llama_token last_token = tokens[n - 1]; + // make sure we never remove tokens in the middle of an image + if (last_token == LLAMA_TOKEN_NULL) { + find_chunk(n - 1); // will throw an error if the token is not begin-of-chunk + } + } + // remove all image chunks that are not used anymore + for (auto it = map_pos_to_image.begin(); it != map_pos_to_image.end(); ) { + llama_pos pos = it->first; + if (pos >= (llama_pos)n) { + it = map_pos_to_image.erase(it); + } else { + ++it; + } + } + } + tokens.resize(n); + } + + std::string detokenize(const llama_context * ctx, bool special) const { + llama_tokens text_tokens; + text_tokens.reserve(tokens.size()); + for (const auto & t : tokens) { + if (t != LLAMA_TOKEN_NULL) { + text_tokens.push_back(t); + } + } + return common_detokenize(ctx, text_tokens, special); + } + + size_t get_common_prefix(const server_tokens & b) const { + size_t max_idx = std::min(tokens.size(), b.tokens.size()); + for (size_t i = 0; i < max_idx; ++i) { + auto & ai = tokens[i]; + auto & bi = b.tokens[i]; + + if (ai == LLAMA_TOKEN_NULL && bi == LLAMA_TOKEN_NULL) { + GGML_ASSERT(has_mtmd); + const auto & a_chunk = find_chunk(i); + const auto & b_chunk = b.find_chunk(i); + GGML_ASSERT(a_chunk && b_chunk); + const auto * a_img = mtmd_input_chunk_get_tokens_image(a_chunk.get()); + const auto * b_img = mtmd_input_chunk_get_tokens_image(b_chunk.get()); + std::string ai_id = mtmd_image_tokens_get_id(a_img); + std::string bi_id = mtmd_image_tokens_get_id(b_img); + size_t a_pos = mtmd_image_tokens_get_n_pos(a_img); + size_t b_pos = mtmd_image_tokens_get_n_pos(b_img); + if (ai_id == bi_id && a_pos == b_pos) { + GGML_ASSERT(a_pos > 0 && "Invalid image token"); // should never happen + i += a_pos - 1; // will be +1 by the for loop + continue; + } else { + return i; + } + } else if (ai == bi) { + continue; + } else { + return i; + } + } + return max_idx; // all tokens are equal + } + + // make sure all text tokens are within the vocab range + bool validate(const struct llama_context * ctx) const { + const llama_model * model = llama_get_model(ctx); + const llama_vocab * vocab = llama_model_get_vocab(model); + const int32_t n_vocab = llama_vocab_n_tokens(vocab); + + for (size_t i = 0; i < tokens.size(); ++i) { + auto & t = tokens[i]; + if (t == LLAMA_TOKEN_NULL) { + try { + const auto & chunk = find_chunk(i); + const auto * img_tokens = mtmd_input_chunk_get_tokens_image(chunk.get()); + size_t n_pos = mtmd_image_tokens_get_n_pos(img_tokens); + i += n_pos - 1; // will be +1 by the for loop + } catch (const std::exception & e) { + return false; + } + } else if (t < 0 || t >= n_vocab) { + return false; + } + } + return true; + } + + // encode and decode the image chunk + int32_t process_chunk( + llama_context * ctx, + mtmd_context * mctx, + llama_pos n_past, + int32_t seq_id, + llama_pos & n_pos_out) { + auto it = map_pos_to_image.find(n_past); + if (it == map_pos_to_image.end()) { + throw std::runtime_error("Chunk not found"); + } + SRV_INF("%s\n", "processing image..."); + int32_t n_batch = llama_n_batch(ctx); + int64_t t0 = ggml_time_ms(); + llama_pos new_n_past = n_past; + int32_t result = mtmd_helper_eval_chunk_single(mctx, ctx, + it->second.get(), // chunk + n_past, + seq_id, + n_batch, + true, // logits last + &new_n_past); + SRV_INF("image processed in %" PRId64 " ms\n", ggml_time_ms() - t0); + if (result != 0) { + LOG_ERR("mtmd_helper_eval failed with status %d", result); + n_pos_out = n_past; + return result; + } + n_pos_out = new_n_past; + return 0; + } +}; + +// Computes FNV-1a hash of the data +static std::string fnv_hash(const uint8_t * data, size_t len) { + const uint64_t fnv_prime = 0x100000001b3ULL; + uint64_t hash = 0xcbf29ce484222325ULL; + + for (size_t i = 0; i < len; ++i) { + hash ^= data[i]; + hash *= fnv_prime; + } + return std::to_string(hash); +} diff --git a/examples/server/webui/.gitignore b/tools/server/webui/.gitignore similarity index 100% rename from examples/server/webui/.gitignore rename to tools/server/webui/.gitignore diff --git a/examples/server/webui/.prettierignore b/tools/server/webui/.prettierignore similarity index 100% rename from examples/server/webui/.prettierignore rename to tools/server/webui/.prettierignore diff --git a/examples/server/webui/eslint.config.js b/tools/server/webui/eslint.config.js similarity index 100% rename from examples/server/webui/eslint.config.js rename to tools/server/webui/eslint.config.js diff --git a/examples/server/webui/index.html b/tools/server/webui/index.html similarity index 100% rename from examples/server/webui/index.html rename to tools/server/webui/index.html diff --git a/examples/server/webui/package-lock.json b/tools/server/webui/package-lock.json similarity index 98% rename from examples/server/webui/package-lock.json rename to tools/server/webui/package-lock.json index b2e3cf94ac..2c23a7580b 100644 --- a/examples/server/webui/package-lock.json +++ b/tools/server/webui/package-lock.json @@ -21,6 +21,8 @@ "postcss": "^8.4.49", "react": "^18.3.1", "react-dom": "^18.3.1", + "react-dropzone": "^14.3.8", + "react-hot-toast": "^2.5.2", "react-markdown": "^9.0.3", "react-router": "^7.1.5", "rehype-highlight": "^7.0.2", @@ -2058,6 +2060,15 @@ "dev": true, "license": "Python-2.0" }, + "node_modules/attr-accept": { + "version": "2.2.5", + "resolved": "https://registry.npmjs.org/attr-accept/-/attr-accept-2.2.5.tgz", + "integrity": "sha512-0bDNnY/u6pPwHDMoF0FieU354oBi0a8rD9FcsLwzcGWbc8KS8KPIi7y+s13OlVY+gMWc/9xEMUgNE6Qm8ZllYQ==", + "license": "MIT", + "engines": { + "node": ">=4" + } + }, "node_modules/autoprefixer": { "version": "10.4.20", "resolved": "https://registry.npmjs.org/autoprefixer/-/autoprefixer-10.4.20.tgz", @@ -2804,6 +2815,18 @@ "node": ">=16.0.0" } }, + "node_modules/file-selector": { + "version": "2.1.2", + "resolved": "https://registry.npmjs.org/file-selector/-/file-selector-2.1.2.tgz", + "integrity": "sha512-QgXo+mXTe8ljeqUFaX3QVHc5osSItJ/Km+xpocx0aSqWGMSCf6qYs/VnzZgS864Pjn5iceMRFigeAV7AfTlaig==", + "license": "MIT", + "dependencies": { + "tslib": "^2.7.0" + }, + "engines": { + "node": ">= 12" + } + }, "node_modules/fill-range": { "version": "7.1.1", "resolved": "https://registry.npmjs.org/fill-range/-/fill-range-7.1.1.tgz", @@ -2917,6 +2940,15 @@ "url": "https://github.com/sponsors/sindresorhus" } }, + "node_modules/goober": { + "version": "2.1.16", + "resolved": "https://registry.npmjs.org/goober/-/goober-2.1.16.tgz", + "integrity": "sha512-erjk19y1U33+XAMe1VTvIONHYoSqE4iS7BYUZfHaqeohLmnC0FdxEh7rQU+6MZ4OajItzjZFSRtVANrQwNq6/g==", + "license": "MIT", + "peerDependencies": { + "csstype": "^3.0.10" + } + }, "node_modules/graceful-fs": { "version": "4.2.11", "resolved": "https://registry.npmjs.org/graceful-fs/-/graceful-fs-4.2.11.tgz", @@ -4674,6 +4706,15 @@ "node": ">=0.10.0" } }, + "node_modules/object-assign": { + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/object-assign/-/object-assign-4.1.1.tgz", + "integrity": "sha512-rJgTQnkUnH1sFw8yT6VSU3zD3sWmu6sZhIseY8VX+GRu3P6F7Fu+JNDoXfklElbLJSnc3FUQHVe4cU5hj+BcUg==", + "license": "MIT", + "engines": { + "node": ">=0.10.0" + } + }, "node_modules/optionator": { "version": "0.9.4", "resolved": "https://registry.npmjs.org/optionator/-/optionator-0.9.4.tgz", @@ -4872,6 +4913,17 @@ "url": "https://github.com/prettier/prettier?sponsor=1" } }, + "node_modules/prop-types": { + "version": "15.8.1", + "resolved": "https://registry.npmjs.org/prop-types/-/prop-types-15.8.1.tgz", + "integrity": "sha512-oj87CgZICdulUohogVAR7AjlC0327U4el4L6eAvOqCeudMDVU0NThNaV+b9Df4dXgSP1gXMTnPdhfe/2qDH5cg==", + "license": "MIT", + "dependencies": { + "loose-envify": "^1.4.0", + "object-assign": "^4.1.1", + "react-is": "^16.13.1" + } + }, "node_modules/property-information": { "version": "6.5.0", "resolved": "https://registry.npmjs.org/property-information/-/property-information-6.5.0.tgz", @@ -4938,6 +4990,46 @@ "react": "^18.3.1" } }, + "node_modules/react-dropzone": { + "version": "14.3.8", + "resolved": "https://registry.npmjs.org/react-dropzone/-/react-dropzone-14.3.8.tgz", + "integrity": "sha512-sBgODnq+lcA4P296DY4wacOZz3JFpD99fp+hb//iBO2HHnyeZU3FwWyXJ6salNpqQdsZrgMrotuko/BdJMV8Ug==", + "license": "MIT", + "dependencies": { + "attr-accept": "^2.2.4", + "file-selector": "^2.1.0", + "prop-types": "^15.8.1" + }, + "engines": { + "node": ">= 10.13" + }, + "peerDependencies": { + "react": ">= 16.8 || 18.0.0" + } + }, + "node_modules/react-hot-toast": { + "version": "2.5.2", + "resolved": "https://registry.npmjs.org/react-hot-toast/-/react-hot-toast-2.5.2.tgz", + "integrity": "sha512-Tun3BbCxzmXXM7C+NI4qiv6lT0uwGh4oAfeJyNOjYUejTsm35mK9iCaYLGv8cBz9L5YxZLx/2ii7zsIwPtPUdw==", + "license": "MIT", + "dependencies": { + "csstype": "^3.1.3", + "goober": "^2.1.16" + }, + "engines": { + "node": ">=10" + }, + "peerDependencies": { + "react": ">=16", + "react-dom": ">=16" + } + }, + "node_modules/react-is": { + "version": "16.13.1", + "resolved": "https://registry.npmjs.org/react-is/-/react-is-16.13.1.tgz", + "integrity": "sha512-24e6ynE2H+OKt4kqsOvNd8kBpV65zoxbA4BVsEOB3ARVWQki/DHzaUoC5KuON/BiccDaCCTZBuOcfZs70kR8bQ==", + "license": "MIT" + }, "node_modules/react-markdown": { "version": "9.0.3", "resolved": "https://registry.npmjs.org/react-markdown/-/react-markdown-9.0.3.tgz", @@ -5814,7 +5906,6 @@ "version": "2.8.1", "resolved": "https://registry.npmjs.org/tslib/-/tslib-2.8.1.tgz", "integrity": "sha512-oJFu94HQb+KVduSUQL7wnpmqnfmLsOA/nAh6b6EH0wCEoK0/mPeXU6c3wKDV83MkOuHPRHtSXKKU99IBazS/2w==", - "devOptional": true, "license": "0BSD" }, "node_modules/turbo-stream": { diff --git a/examples/server/webui/package.json b/tools/server/webui/package.json similarity index 96% rename from examples/server/webui/package.json rename to tools/server/webui/package.json index 6ac06b1a49..ab1b920bdc 100644 --- a/examples/server/webui/package.json +++ b/tools/server/webui/package.json @@ -24,6 +24,8 @@ "postcss": "^8.4.49", "react": "^18.3.1", "react-dom": "^18.3.1", + "react-dropzone": "^14.3.8", + "react-hot-toast": "^2.5.2", "react-markdown": "^9.0.3", "react-router": "^7.1.5", "rehype-highlight": "^7.0.2", diff --git a/examples/server/webui/postcss.config.js b/tools/server/webui/postcss.config.js similarity index 100% rename from examples/server/webui/postcss.config.js rename to tools/server/webui/postcss.config.js diff --git a/examples/server/webui/public/demo-conversation.json b/tools/server/webui/public/demo-conversation.json similarity index 100% rename from examples/server/webui/public/demo-conversation.json rename to tools/server/webui/public/demo-conversation.json diff --git a/examples/server/webui/src/App.tsx b/tools/server/webui/src/App.tsx similarity index 95% rename from examples/server/webui/src/App.tsx rename to tools/server/webui/src/App.tsx index cc4659e152..3b00a8f909 100644 --- a/examples/server/webui/src/App.tsx +++ b/tools/server/webui/src/App.tsx @@ -4,6 +4,7 @@ import Sidebar from './components/Sidebar'; import { AppContextProvider, useAppContext } from './utils/app.context'; import ChatScreen from './components/ChatScreen'; import SettingDialog from './components/SettingDialog'; +import { Toaster } from 'react-hot-toast'; function App() { return ( @@ -40,6 +41,7 @@ function AppLayout() { onClose={() => setShowSettings(false)} /> } + ); } diff --git a/examples/server/webui/src/Config.ts b/tools/server/webui/src/Config.ts similarity index 98% rename from examples/server/webui/src/Config.ts rename to tools/server/webui/src/Config.ts index dd1cc0e100..5eef608cb9 100644 --- a/examples/server/webui/src/Config.ts +++ b/tools/server/webui/src/Config.ts @@ -12,7 +12,7 @@ export const CONFIG_DEFAULT = { // Note: in order not to introduce breaking changes, please keep the same data type (number, string, etc) if you want to change the default value. Do not use null or undefined for default value. // Do not use nested objects, keep it single level. Prefix the key if you need to group them. apiKey: '', - systemMessage: 'You are a helpful assistant.', + systemMessage: '', showTokensPerSecond: false, showThoughtInProgress: false, excludeThoughtOnReq: true, diff --git a/examples/server/webui/src/components/CanvasPyInterpreter.tsx b/tools/server/webui/src/components/CanvasPyInterpreter.tsx similarity index 100% rename from examples/server/webui/src/components/CanvasPyInterpreter.tsx rename to tools/server/webui/src/components/CanvasPyInterpreter.tsx diff --git a/tools/server/webui/src/components/ChatInputExtraContextItem.tsx b/tools/server/webui/src/components/ChatInputExtraContextItem.tsx new file mode 100644 index 0000000000..ac416fa907 --- /dev/null +++ b/tools/server/webui/src/components/ChatInputExtraContextItem.tsx @@ -0,0 +1,92 @@ +import { DocumentTextIcon, XMarkIcon } from '@heroicons/react/24/outline'; +import { MessageExtra } from '../utils/types'; +import { useState } from 'react'; +import { classNames } from '../utils/misc'; + +export default function ChatInputExtraContextItem({ + items, + removeItem, + clickToShow, +}: { + items?: MessageExtra[]; + removeItem?: (index: number) => void; + clickToShow?: boolean; +}) { + const [show, setShow] = useState(-1); + const showingItem = show >= 0 ? items?.[show] : undefined; + + if (!items) return null; + + return ( +
+ {items.map((item, i) => ( +
clickToShow && setShow(i)} + > + {removeItem && ( +
+ +
+ )} + +
+ {item.type === 'imageFile' ? ( + <> + {item.name} + + ) : ( + <> +
+ +
+ +
+ {item.name ?? 'Extra content'} +
+ + )} +
+
+ ))} + + {showingItem && ( + +
+
+ {showingItem.name ?? 'Extra content'} + +
+ {showingItem.type === 'imageFile' ? ( + {showingItem.name} + ) : ( +
+ 
+                  {showingItem.content}
+
+
+ )} +
+
setShow(-1)}>
+ + )} +
+ ); +} diff --git a/examples/server/webui/src/components/ChatMessage.tsx b/tools/server/webui/src/components/ChatMessage.tsx similarity index 72% rename from examples/server/webui/src/components/ChatMessage.tsx rename to tools/server/webui/src/components/ChatMessage.tsx index 40ea74711f..08eb423526 100644 --- a/examples/server/webui/src/components/ChatMessage.tsx +++ b/tools/server/webui/src/components/ChatMessage.tsx @@ -3,7 +3,14 @@ import { useAppContext } from '../utils/app.context'; import { Message, PendingMessage } from '../utils/types'; import { classNames } from '../utils/misc'; import MarkdownDisplay, { CopyButton } from './MarkdownDisplay'; -import { ChevronLeftIcon, ChevronRightIcon } from '@heroicons/react/24/outline'; +import { + ArrowPathIcon, + ChevronLeftIcon, + ChevronRightIcon, + PencilSquareIcon, +} from '@heroicons/react/24/outline'; +import ChatInputExtraContextItem from './ChatInputExtraContextItem'; +import { BtnWithTooltips } from '../utils/common'; interface SplitMessage { content: PendingMessage['content']; @@ -85,10 +92,14 @@ export default function ChatMessage({ 'chat-end': msg.role === 'user', })} > + {msg.extra && msg.extra.length > 0 && ( + + )} +
{/* textarea for editing message */} @@ -133,59 +144,11 @@ export default function ChatMessage({ {/* render message as markdown */}
{thought && ( -
- - {isPending && isThinking ? ( - - - Thinking - - ) : ( - Thought Process - )} - -
- -
-
- )} - - {msg.extra && msg.extra.length > 0 && ( -
- - Extra content - -
- {msg.extra.map( - (extra, i) => - extra.type === 'textFile' ? ( -
- {extra.name} - 
{extra.content}
-
- ) : extra.type === 'context' ? ( -
- 
{extra.content}
-
- ) : null // TODO: support other extra types - )} -
-
+ )} setEditingContent(msg.content)} disabled={msg.content === null} + tooltipsContent="Edit message" > - ✍️ Edit - + + )} {/* assistant message */} {msg.role === 'assistant' && ( <> {!isPending && ( - + + )} )}
@@ -294,3 +259,44 @@ export default function ChatMessage({
); } + +function ThoughtProcess({ + isThinking, + content, + open, +}: { + isThinking: boolean; + content: string; + open: boolean; +}) { + return ( +
+ +
+
+ {isThinking ? ( + + + Thinking + + ) : ( + <>Thought Process + )} +
+
+
+
+ +
+
+
+ ); +} diff --git a/examples/server/webui/src/components/ChatScreen.tsx b/tools/server/webui/src/components/ChatScreen.tsx similarity index 54% rename from examples/server/webui/src/components/ChatScreen.tsx rename to tools/server/webui/src/components/ChatScreen.tsx index 29ab5ea64f..b645a494d6 100644 --- a/examples/server/webui/src/components/ChatScreen.tsx +++ b/tools/server/webui/src/components/ChatScreen.tsx @@ -1,12 +1,25 @@ -import { useEffect, useMemo, useState } from 'react'; +import { useEffect, useMemo, useRef, useState } from 'react'; import { CallbackGeneratedChunk, useAppContext } from '../utils/app.context'; import ChatMessage from './ChatMessage'; import { CanvasType, Message, PendingMessage } from '../utils/types'; -import { classNames, cleanCurrentUrl, throttle } from '../utils/misc'; +import { classNames, cleanCurrentUrl } from '../utils/misc'; import CanvasPyInterpreter from './CanvasPyInterpreter'; import StorageUtils from '../utils/storage'; import { useVSCodeContext } from '../utils/llama-vscode'; import { useChatTextarea, ChatTextareaApi } from './useChatTextarea.ts'; +import { + ArrowUpIcon, + StopIcon, + PaperClipIcon, +} from '@heroicons/react/24/solid'; +import { + ChatExtraContextApi, + useChatExtraContext, +} from './useChatExtraContext.tsx'; +import Dropzone from 'react-dropzone'; +import toast from 'react-hot-toast'; +import ChatInputExtraContextItem from './ChatInputExtraContextItem.tsx'; +import { scrollToBottom, useChatScroll } from './useChatScroll.tsx'; /** * A message display is a message node with additional information for rendering. @@ -72,24 +85,6 @@ function getListMessageDisplay( return res; } -const scrollToBottom = throttle( - (requiresNearBottom: boolean, delay: number = 80) => { - const mainScrollElem = document.getElementById('main-scroll'); - if (!mainScrollElem) return; - const spaceToBottom = - mainScrollElem.scrollHeight - - mainScrollElem.scrollTop - - mainScrollElem.clientHeight; - if (!requiresNearBottom || spaceToBottom < 50) { - setTimeout( - () => mainScrollElem.scrollTo({ top: mainScrollElem.scrollHeight }), - delay - ); - } - }, - 80 -); - export default function ChatScreen() { const { viewingChat, @@ -102,10 +97,11 @@ export default function ChatScreen() { } = useAppContext(); const textarea: ChatTextareaApi = useChatTextarea(prefilledMsg.content()); + const extraContext = useChatExtraContext(); + useVSCodeContext(textarea, extraContext); - const { extraContext, clearExtraContext } = useVSCodeContext(textarea); - // TODO: improve this when we have "upload file" feature - const currExtra: Message['extra'] = extraContext ? [extraContext] : undefined; + const msgListRef = useRef(null); + useChatScroll(msgListRef); // keep track of leaf node for rendering const [currNodeId, setCurrNodeId] = useState(-1); @@ -129,13 +125,15 @@ export default function ChatScreen() { if (currLeafNodeId) { setCurrNodeId(currLeafNodeId); } - scrollToBottom(true); + // useChatScroll will handle the auto scroll }; const sendNewMessage = async () => { const lastInpMsg = textarea.value(); - if (lastInpMsg.trim().length === 0 || isGenerating(currConvId ?? '')) + if (lastInpMsg.trim().length === 0 || isGenerating(currConvId ?? '')) { + toast.error('Please enter a message'); return; + } textarea.setValue(''); scrollToBottom(false); setCurrNodeId(-1); @@ -146,7 +144,7 @@ export default function ChatScreen() { currConvId, lastMsgNodeId, lastInpMsg, - currExtra, + extraContext.items, onChunk )) ) { @@ -154,9 +152,12 @@ export default function ChatScreen() { textarea.setValue(lastInpMsg); } // OK - clearExtraContext(); + extraContext.clearItems(); }; + // for vscode context + textarea.refOnSubmit.current = sendNewMessage; + const handleEditMessage = async (msg: Message, content: string) => { if (!viewingChat) return; setCurrNodeId(msg.id); @@ -231,10 +232,17 @@ export default function ChatScreen() { })} > {/* chat messages */} -
-
+
+
{/* placeholder to shift the message to the bottom */} - {viewingChat ? '' : 'Send a message to start'} + {viewingChat ? ( + '' + ) : ( + <> +
Send a message to start
+ + + )}
{[...messages, ...pendingMsgDisplay].map((msg) => ( ))}
{/* chat input */} -
- - - {isGenerating(currConvId ?? '') ? ( - - ) : ( - - )} -
+ stopGenerating(currConvId ?? '')} + isGenerating={isGenerating(currConvId ?? '')} + />
{canvasData?.type === CanvasType.PY_INTERPRETER && ( @@ -294,3 +275,129 @@ export default function ChatScreen() {
); } + +function ServerInfo() { + const { serverProps } = useAppContext(); + return ( +
+
+ Server Info +

+ Model: {serverProps?.model_path?.split(/(\\|\/)/).pop()} +
+ Build: {serverProps?.build_info} +
+

+
+
+ ); +} + +function ChatInput({ + textarea, + extraContext, + onSend, + onStop, + isGenerating, +}: { + textarea: ChatTextareaApi; + extraContext: ChatExtraContextApi; + onSend: () => void; + onStop: () => void; + isGenerating: boolean; +}) { + const [isDrag, setIsDrag] = useState(false); + + return ( +
+ { + setIsDrag(false); + extraContext.onFileAdded(files); + }} + onDragEnter={() => setIsDrag(true)} + onDragLeave={() => setIsDrag(false)} + multiple={true} + > + {({ getRootProps, getInputProps }) => ( +
+ {!isGenerating && ( + + )} + +
+ + + {/* buttons area */} +
+ + + {isGenerating ? ( + + ) : ( + + )} +
+
+
+ )} + +
+ ); +} diff --git a/tools/server/webui/src/components/Header.tsx b/tools/server/webui/src/components/Header.tsx new file mode 100644 index 0000000000..45775ff7a6 --- /dev/null +++ b/tools/server/webui/src/components/Header.tsx @@ -0,0 +1,88 @@ +import { useEffect, useState } from 'react'; +import StorageUtils from '../utils/storage'; +import { useAppContext } from '../utils/app.context'; +import { classNames } from '../utils/misc'; +import daisyuiThemes from 'daisyui/theme/object'; +import { THEMES } from '../Config'; +import { + Cog8ToothIcon, + MoonIcon, + Bars3Icon, +} from '@heroicons/react/24/outline'; + +export default function Header() { + const [selectedTheme, setSelectedTheme] = useState(StorageUtils.getTheme()); + const { setShowSettings } = useAppContext(); + + const setTheme = (theme: string) => { + StorageUtils.setTheme(theme); + setSelectedTheme(theme); + }; + + useEffect(() => { + document.body.setAttribute('data-theme', selectedTheme); + document.body.setAttribute( + 'data-color-scheme', + daisyuiThemes[selectedTheme]?.['color-scheme'] ?? 'auto' + ); + }, [selectedTheme]); + + return ( +
+ {/* open sidebar button */} + + +
llama.cpp
+ + {/* action buttons (top right) */} +
+
+ +
+ + {/* theme controller is copied from https://daisyui.com/components/theme-controller/ */} +
+
+
+ +
+
    +
  • + +
    • + {THEMES.map((theme) => ( +
  • + e.target.checked && setTheme(theme)} + /> +
    • + ))} +
+
+
+
+
+ ); +} diff --git a/examples/server/webui/src/components/MarkdownDisplay.tsx b/tools/server/webui/src/components/MarkdownDisplay.tsx similarity index 94% rename from examples/server/webui/src/components/MarkdownDisplay.tsx rename to tools/server/webui/src/components/MarkdownDisplay.tsx index 5b7a725914..380dbc570a 100644 --- a/examples/server/webui/src/components/MarkdownDisplay.tsx +++ b/tools/server/webui/src/components/MarkdownDisplay.tsx @@ -11,6 +11,8 @@ import { ElementContent, Root } from 'hast'; import { visit } from 'unist-util-visit'; import { useAppContext } from '../utils/app.context'; import { CanvasType } from '../utils/types'; +import { BtnWithTooltips } from '../utils/common'; +import { DocumentDuplicateIcon, PlayIcon } from '@heroicons/react/24/outline'; export default function MarkdownDisplay({ content, @@ -81,10 +83,13 @@ const CodeBlockButtons: React.ElementType< 'display-none': !node?.position, })} > - + {canRunCode && ( )} @@ -101,16 +106,17 @@ export const CopyButton = ({ }) => { const [copied, setCopied] = useState(false); return ( - + + ); }; @@ -124,7 +130,7 @@ export const RunPyCodeButton = ({ const { setCanvasData } = useAppContext(); return ( <> - + + ); }; diff --git a/examples/server/webui/src/components/SettingDialog.tsx b/tools/server/webui/src/components/SettingDialog.tsx similarity index 99% rename from examples/server/webui/src/components/SettingDialog.tsx rename to tools/server/webui/src/components/SettingDialog.tsx index b65e73ae16..b0044d2540 100644 --- a/examples/server/webui/src/components/SettingDialog.tsx +++ b/tools/server/webui/src/components/SettingDialog.tsx @@ -196,7 +196,7 @@ const SETTING_SECTIONS: SettingSection[] = [ label: ( <> Custom JSON config (For more info, refer to{' '} - + server documentation ) diff --git a/tools/server/webui/src/components/Sidebar.tsx b/tools/server/webui/src/components/Sidebar.tsx new file mode 100644 index 0000000000..8e79e00b8d --- /dev/null +++ b/tools/server/webui/src/components/Sidebar.tsx @@ -0,0 +1,335 @@ +import { useEffect, useMemo, useState } from 'react'; +import { classNames } from '../utils/misc'; +import { Conversation } from '../utils/types'; +import StorageUtils from '../utils/storage'; +import { useNavigate, useParams } from 'react-router'; +import { + ArrowDownTrayIcon, + EllipsisVerticalIcon, + PencilIcon, + PencilSquareIcon, + TrashIcon, + XMarkIcon, +} from '@heroicons/react/24/outline'; +import { BtnWithTooltips } from '../utils/common'; +import { useAppContext } from '../utils/app.context'; +import toast from 'react-hot-toast'; + +export default function Sidebar() { + const params = useParams(); + const navigate = useNavigate(); + + const { isGenerating } = useAppContext(); + + const [conversations, setConversations] = useState([]); + const [currConv, setCurrConv] = useState(null); + + useEffect(() => { + StorageUtils.getOneConversation(params.convId ?? '').then(setCurrConv); + }, [params.convId]); + + useEffect(() => { + const handleConversationChange = async () => { + setConversations(await StorageUtils.getAllConversations()); + }; + StorageUtils.onConversationChanged(handleConversationChange); + handleConversationChange(); + return () => { + StorageUtils.offConversationChanged(handleConversationChange); + }; + }, []); + + const groupedConv = useMemo( + () => groupConversationsByDate(conversations), + [conversations] + ); + + return ( + <> + + +
+ +
+
+

Conversations

+ + {/* close sidebar button */} + +
+ + {/* new conversation button */} +
navigate('/')} + > + + New conversation +
+ + {/* list of conversations */} + {groupedConv.map((group, i) => ( +
+ {/* group name (by date) */} + {group.title ? ( + // we use btn class here to make sure that the padding/margin are aligned with the other items + + {group.title} + + ) : ( +
+ )} + + {group.conversations.map((conv) => ( + { + navigate(`/chat/${conv.id}`); + }} + onDelete={() => { + if (isGenerating(conv.id)) { + toast.error( + 'Cannot delete conversation while generating' + ); + return; + } + if ( + window.confirm( + 'Are you sure to delete this conversation?' + ) + ) { + toast.success('Conversation deleted'); + StorageUtils.remove(conv.id); + navigate('/'); + } + }} + onDownload={() => { + if (isGenerating(conv.id)) { + toast.error( + 'Cannot download conversation while generating' + ); + return; + } + const conversationJson = JSON.stringify(conv, null, 2); + const blob = new Blob([conversationJson], { + type: 'application/json', + }); + const url = URL.createObjectURL(blob); + const a = document.createElement('a'); + a.href = url; + a.download = `conversation_${conv.id}.json`; + document.body.appendChild(a); + a.click(); + document.body.removeChild(a); + URL.revokeObjectURL(url); + }} + onRename={() => { + if (isGenerating(conv.id)) { + toast.error( + 'Cannot rename conversation while generating' + ); + return; + } + const newName = window.prompt( + 'Enter new name for the conversation', + conv.name + ); + if (newName && newName.trim().length > 0) { + StorageUtils.updateConversationName(conv.id, newName); + } + }} + /> + ))} +
+ ))} +
+ Conversations are saved to browser's IndexedDB +
+
+
+ + ); +} + +function ConversationItem({ + conv, + isCurrConv, + onSelect, + onDelete, + onDownload, + onRename, +}: { + conv: Conversation; + isCurrConv: boolean; + onSelect: () => void; + onDelete: () => void; + onDownload: () => void; + onRename: () => void; +}) { + return ( +
+
+ {conv.name} +
+
+ + {/* dropdown menu */} + +
+
+ ); +} + +// WARN: vibe code below + +export interface GroupedConversations { + title?: string; + conversations: Conversation[]; +} + +// TODO @ngxson : add test for this function +// Group conversations by date +// - "Previous 7 Days" +// - "Previous 30 Days" +// - "Month Year" (e.g., "April 2023") +export function groupConversationsByDate( + conversations: Conversation[] +): GroupedConversations[] { + const now = new Date(); + const today = new Date(now.getFullYear(), now.getMonth(), now.getDate()); // Start of today + + const sevenDaysAgo = new Date(today); + sevenDaysAgo.setDate(today.getDate() - 7); + + const thirtyDaysAgo = new Date(today); + thirtyDaysAgo.setDate(today.getDate() - 30); + + const groups: { [key: string]: Conversation[] } = { + Today: [], + 'Previous 7 Days': [], + 'Previous 30 Days': [], + }; + const monthlyGroups: { [key: string]: Conversation[] } = {}; // Key format: "Month Year" e.g., "April 2023" + + // Sort conversations by lastModified date in descending order (newest first) + // This helps when adding to groups, but the final output order of groups is fixed. + const sortedConversations = [...conversations].sort( + (a, b) => b.lastModified - a.lastModified + ); + + for (const conv of sortedConversations) { + const convDate = new Date(conv.lastModified); + + if (convDate >= today) { + groups['Today'].push(conv); + } else if (convDate >= sevenDaysAgo) { + groups['Previous 7 Days'].push(conv); + } else if (convDate >= thirtyDaysAgo) { + groups['Previous 30 Days'].push(conv); + } else { + const monthName = convDate.toLocaleString('default', { month: 'long' }); + const year = convDate.getFullYear(); + const monthYearKey = `${monthName} ${year}`; + if (!monthlyGroups[monthYearKey]) { + monthlyGroups[monthYearKey] = []; + } + monthlyGroups[monthYearKey].push(conv); + } + } + + const result: GroupedConversations[] = []; + + if (groups['Today'].length > 0) { + result.push({ + title: undefined, // no title for Today + conversations: groups['Today'], + }); + } + + if (groups['Previous 7 Days'].length > 0) { + result.push({ + title: 'Previous 7 Days', + conversations: groups['Previous 7 Days'], + }); + } + + if (groups['Previous 30 Days'].length > 0) { + result.push({ + title: 'Previous 30 Days', + conversations: groups['Previous 30 Days'], + }); + } + + // Sort monthly groups by date (most recent month first) + const sortedMonthKeys = Object.keys(monthlyGroups).sort((a, b) => { + const dateA = new Date(a); // "Month Year" can be parsed by Date constructor + const dateB = new Date(b); + return dateB.getTime() - dateA.getTime(); + }); + + for (const monthKey of sortedMonthKeys) { + if (monthlyGroups[monthKey].length > 0) { + result.push({ title: monthKey, conversations: monthlyGroups[monthKey] }); + } + } + + return result; +} diff --git a/tools/server/webui/src/components/useChatExtraContext.tsx b/tools/server/webui/src/components/useChatExtraContext.tsx new file mode 100644 index 0000000000..7eeff61f5e --- /dev/null +++ b/tools/server/webui/src/components/useChatExtraContext.tsx @@ -0,0 +1,234 @@ +import { useState } from 'react'; +import { MessageExtra } from '../utils/types'; +import toast from 'react-hot-toast'; +import { useAppContext } from '../utils/app.context'; + +// Interface describing the API returned by the hook +export interface ChatExtraContextApi { + items?: MessageExtra[]; // undefined if empty, similar to Message['extra'] + addItems: (items: MessageExtra[]) => void; + removeItem: (idx: number) => void; + clearItems: () => void; + onFileAdded: (files: File[]) => void; // used by "upload" button +} + +export function useChatExtraContext(): ChatExtraContextApi { + const { serverProps } = useAppContext(); + const [items, setItems] = useState([]); + + const addItems = (newItems: MessageExtra[]) => { + setItems((prev) => [...prev, ...newItems]); + }; + + const removeItem = (idx: number) => { + setItems((prev) => prev.filter((_, i) => i !== idx)); + }; + + const clearItems = () => { + setItems([]); + }; + + const onFileAdded = (files: File[]) => { + for (const file of files) { + const mimeType = file.type; + console.debug({ mimeType, file }); + if (file.size > 10 * 1024 * 1024) { + toast.error('File is too large. Maximum size is 10MB.'); + break; + } + + if (mimeType.startsWith('image/')) { + if (!serverProps?.modalities?.vision) { + toast.error('Multimodal is not supported by this server or model.'); + break; + } + const reader = new FileReader(); + reader.onload = async (event) => { + if (event.target?.result) { + let base64Url = event.target.result as string; + + if (mimeType === 'image/svg+xml') { + // Convert SVG to PNG + base64Url = await svgBase64UrlToPngDataURL(base64Url); + } + + addItems([ + { + type: 'imageFile', + name: file.name, + base64Url, + }, + ]); + } + }; + reader.readAsDataURL(file); + } else if ( + mimeType.startsWith('video/') || + mimeType.startsWith('audio/') + ) { + toast.error('Video and audio files are not supported yet.'); + break; + } else if (mimeType.startsWith('application/pdf')) { + toast.error('PDF files are not supported yet.'); + break; + } else { + // Because there can be many text file types (like code file), we will not check the mime type + // and will just check if the file is not binary. + const reader = new FileReader(); + reader.onload = (event) => { + if (event.target?.result) { + const content = event.target.result as string; + if (!isLikelyNotBinary(content)) { + toast.error('File is binary. Please upload a text file.'); + return; + } + addItems([ + { + type: 'textFile', + name: file.name, + content, + }, + ]); + } + }; + reader.readAsText(file); + } + } + }; + + return { + items: items.length > 0 ? items : undefined, + addItems, + removeItem, + clearItems, + onFileAdded, + }; +} + +// WARN: vibe code below +// This code is a heuristic to determine if a string is likely not binary. +// It is necessary because input file can have various mime types which we don't have time to investigate. +// For example, a python file can be text/plain, application/x-python, etc. +export function isLikelyNotBinary(str: string): boolean { + const options = { + prefixLength: 1024 * 10, // Check the first 10KB of the string + suspiciousCharThresholdRatio: 0.15, // Allow up to 15% suspicious chars + maxAbsoluteNullBytes: 2, + }; + + if (!str) { + return true; // Empty string is considered "not binary" or trivially text. + } + + const sampleLength = Math.min(str.length, options.prefixLength); + if (sampleLength === 0) { + return true; // Effectively an empty string after considering prefixLength. + } + + let suspiciousCharCount = 0; + let nullByteCount = 0; + + for (let i = 0; i < sampleLength; i++) { + const charCode = str.charCodeAt(i); + + // 1. Check for Unicode Replacement Character (U+FFFD) + // This is a strong indicator if the string was created from decoding bytes as UTF-8. + if (charCode === 0xfffd) { + suspiciousCharCount++; + continue; + } + + // 2. Check for Null Bytes (U+0000) + if (charCode === 0x0000) { + nullByteCount++; + // We also count nulls towards the general suspicious character count, + // as they are less common in typical text files. + suspiciousCharCount++; + continue; + } + + // 3. Check for C0 Control Characters (U+0001 to U+001F) + // Exclude common text control characters: TAB (9), LF (10), CR (13). + // We can also be a bit lenient with BEL (7) and BS (8) which sometimes appear in logs. + if (charCode < 32) { + if ( + charCode !== 9 && // TAB + charCode !== 10 && // LF + charCode !== 13 && // CR + charCode !== 7 && // BEL (Bell) - sometimes in logs + charCode !== 8 // BS (Backspace) - less common, but possible + ) { + suspiciousCharCount++; + } + } + // Characters from 32 (space) up to 126 (~) are printable ASCII. + // Characters 127 (DEL) is a control character. + // Characters >= 128 are extended ASCII / multi-byte Unicode. + // If they resulted in U+FFFD, we caught it. Otherwise, they are valid + // (though perhaps unusual) Unicode characters from JS's perspective. + // The main concern is if those higher characters came from misinterpreting + // a single-byte encoding as UTF-8, which again, U+FFFD would usually flag. + } + + // Check absolute null byte count + if (nullByteCount > options.maxAbsoluteNullBytes) { + return false; // Too many null bytes is a strong binary indicator + } + + // Check ratio of suspicious characters + const ratio = suspiciousCharCount / sampleLength; + return ratio <= options.suspiciousCharThresholdRatio; +} + +// WARN: vibe code below +// Converts a Base64URL encoded SVG string to a PNG Data URL using browser Canvas API. +function svgBase64UrlToPngDataURL(base64UrlSvg: string): Promise { + const backgroundColor = 'white'; // Default background color for PNG + + return new Promise((resolve, reject) => { + try { + const img = new Image(); + + img.onload = () => { + const canvas = document.createElement('canvas'); + const ctx = canvas.getContext('2d'); + + if (!ctx) { + reject(new Error('Failed to get 2D canvas context.')); + return; + } + + // Use provided dimensions or SVG's natural dimensions, with fallbacks + // Fallbacks (e.g., 300x300) are for SVGs without explicit width/height + // or when naturalWidth/Height might be 0 before full processing. + const targetWidth = img.naturalWidth || 300; + const targetHeight = img.naturalHeight || 300; + + canvas.width = targetWidth; + canvas.height = targetHeight; + + if (backgroundColor) { + ctx.fillStyle = backgroundColor; + ctx.fillRect(0, 0, canvas.width, canvas.height); + } + + ctx.drawImage(img, 0, 0, targetWidth, targetHeight); + resolve(canvas.toDataURL('image/png')); + }; + + img.onerror = () => { + reject( + new Error('Failed to load SVG image. Ensure the SVG data is valid.') + ); + }; + + // Load SVG string into an Image element + img.src = base64UrlSvg; + } catch (error) { + const message = error instanceof Error ? error.message : String(error); + const errorMessage = `Error converting SVG to PNG: ${message}`; + toast.error(errorMessage); + reject(new Error(errorMessage)); + } + }); +} diff --git a/tools/server/webui/src/components/useChatScroll.tsx b/tools/server/webui/src/components/useChatScroll.tsx new file mode 100644 index 0000000000..25ea02234a --- /dev/null +++ b/tools/server/webui/src/components/useChatScroll.tsx @@ -0,0 +1,34 @@ +import React, { useEffect } from 'react'; +import { throttle } from '../utils/misc'; + +export const scrollToBottom = (requiresNearBottom: boolean, delay?: number) => { + const mainScrollElem = document.getElementById('main-scroll'); + if (!mainScrollElem) return; + const spaceToBottom = + mainScrollElem.scrollHeight - + mainScrollElem.scrollTop - + mainScrollElem.clientHeight; + if (!requiresNearBottom || spaceToBottom < 100) { + setTimeout( + () => mainScrollElem.scrollTo({ top: mainScrollElem.scrollHeight }), + delay ?? 80 + ); + } +}; + +const scrollToBottomThrottled = throttle(scrollToBottom, 80); + +export function useChatScroll(msgListRef: React.RefObject) { + useEffect(() => { + if (!msgListRef.current) return; + + const resizeObserver = new ResizeObserver((_) => { + scrollToBottomThrottled(true, 10); + }); + + resizeObserver.observe(msgListRef.current); + return () => { + resizeObserver.disconnect(); + }; + }, [msgListRef]); +} diff --git a/examples/server/webui/src/components/useChatTextarea.ts b/tools/server/webui/src/components/useChatTextarea.ts similarity index 61% rename from examples/server/webui/src/components/useChatTextarea.ts rename to tools/server/webui/src/components/useChatTextarea.ts index 42b1281946..c2f8652031 100644 --- a/examples/server/webui/src/components/useChatTextarea.ts +++ b/tools/server/webui/src/components/useChatTextarea.ts @@ -1,35 +1,39 @@ import { useEffect, useRef, useState, useCallback } from 'react'; +import { throttle } from '../utils/misc'; // Media Query for detecting "large" screens (matching Tailwind's lg: breakpoint) const LARGE_SCREEN_MQ = '(min-width: 1024px)'; // Calculates and sets the textarea height based on its scrollHeight -const adjustTextareaHeight = (textarea: HTMLTextAreaElement | null) => { - if (!textarea) return; +const adjustTextareaHeight = throttle( + (textarea: HTMLTextAreaElement | null) => { + if (!textarea) return; - // Only perform auto-sizing on large screens - if (!window.matchMedia(LARGE_SCREEN_MQ).matches) { - // On small screens, reset inline height and max-height styles. - // This allows CSS (e.g., `rows` attribute or classes) to control the height, - // and enables manual resizing if `resize-vertical` is set. - textarea.style.height = ''; // Use 'auto' or '' to reset - textarea.style.maxHeight = ''; - return; // Do not adjust height programmatically on small screens - } + // Only perform auto-sizing on large screens + if (!window.matchMedia(LARGE_SCREEN_MQ).matches) { + // On small screens, reset inline height and max-height styles. + // This allows CSS (e.g., `rows` attribute or classes) to control the height, + // and enables manual resizing if `resize-vertical` is set. + textarea.style.height = ''; // Use 'auto' or '' to reset + textarea.style.maxHeight = ''; + return; // Do not adjust height programmatically on small screens + } - const computedStyle = window.getComputedStyle(textarea); - // Get the max-height specified by CSS (e.g., from `lg:max-h-48`) - const currentMaxHeight = computedStyle.maxHeight; + const computedStyle = window.getComputedStyle(textarea); + // Get the max-height specified by CSS (e.g., from `lg:max-h-48`) + const currentMaxHeight = computedStyle.maxHeight; - // Temporarily remove max-height to allow scrollHeight to be calculated correctly - textarea.style.maxHeight = 'none'; - // Reset height to 'auto' to measure the actual scrollHeight needed - textarea.style.height = 'auto'; - // Set the height to the calculated scrollHeight - textarea.style.height = `${textarea.scrollHeight}px`; - // Re-apply the original max-height from CSS to enforce the limit - textarea.style.maxHeight = currentMaxHeight; -}; + // Temporarily remove max-height to allow scrollHeight to be calculated correctly + textarea.style.maxHeight = 'none'; + // Reset height to 'auto' to measure the actual scrollHeight needed + textarea.style.height = 'auto'; + // Set the height to the calculated scrollHeight + textarea.style.height = `${textarea.scrollHeight}px`; + // Re-apply the original max-height from CSS to enforce the limit + textarea.style.maxHeight = currentMaxHeight; + }, + 100 +); // Throttle to prevent excessive calls // Interface describing the API returned by the hook export interface ChatTextareaApi { @@ -37,6 +41,7 @@ export interface ChatTextareaApi { setValue: (value: string) => void; focus: () => void; ref: React.RefObject; + refOnSubmit: React.MutableRefObject<(() => void) | null>; // Submit handler onInput: (event: React.FormEvent) => void; // Input handler } @@ -46,6 +51,7 @@ export interface ChatTextareaApi { export function useChatTextarea(initValue: string): ChatTextareaApi { const [savedInitValue, setSavedInitValue] = useState(initValue); const textareaRef = useRef(null); + const onSubmitRef = useRef<(() => void) | null>(null); // Effect to set initial value and height on mount or when initValue changes useEffect(() => { @@ -63,6 +69,7 @@ export function useChatTextarea(initValue: string): ChatTextareaApi { } }, [textareaRef, savedInitValue]); // Depend on ref and savedInitValue + // On input change, we adjust the height of the textarea const handleInput = useCallback( (event: React.FormEvent) => { // Call adjustTextareaHeight on every input - it will decide whether to act @@ -91,6 +98,7 @@ export function useChatTextarea(initValue: string): ChatTextareaApi { } }, ref: textareaRef, - onInput: handleInput, + refOnSubmit: onSubmitRef, + onInput: handleInput, // for adjusting height on input }; } diff --git a/examples/server/webui/src/index.scss b/tools/server/webui/src/index.scss similarity index 90% rename from examples/server/webui/src/index.scss rename to tools/server/webui/src/index.scss index a18f094542..563e7a4610 100644 --- a/examples/server/webui/src/index.scss +++ b/tools/server/webui/src/index.scss @@ -22,12 +22,15 @@ html { all: revert; } pre { - @apply whitespace-pre-wrap rounded-lg p-2; + @apply whitespace-pre-wrap rounded-lg p-2 mb-3; border: 1px solid currentColor; } p { @apply mb-2; } + hr { + @apply my-4 border-base-content/20 border-1; + } /* TODO: fix markdown table */ } @@ -35,7 +38,7 @@ html { @apply md:opacity-0 md:group-hover:opacity-100; } .btn-mini { - @apply cursor-pointer hover:shadow-md; + @apply cursor-pointer; } .chat-screen { max-width: 900px; diff --git a/examples/server/webui/src/main.tsx b/tools/server/webui/src/main.tsx similarity index 100% rename from examples/server/webui/src/main.tsx rename to tools/server/webui/src/main.tsx diff --git a/examples/server/webui/src/utils/app.context.tsx b/tools/server/webui/src/utils/app.context.tsx similarity index 94% rename from examples/server/webui/src/utils/app.context.tsx rename to tools/server/webui/src/utils/app.context.tsx index 54bb65b6e3..96cffd95ab 100644 --- a/examples/server/webui/src/utils/app.context.tsx +++ b/tools/server/webui/src/utils/app.context.tsx @@ -3,6 +3,7 @@ import { APIMessage, CanvasData, Conversation, + LlamaCppServerProps, Message, PendingMessage, ViewingChat, @@ -12,9 +13,11 @@ import { filterThoughtFromMsgs, normalizeMsgsForAPI, getSSEStreamAsync, + getServerProps, } from './misc'; import { BASE_URL, CONFIG_DEFAULT, isDev } from '../Config'; import { matchPath, useLocation, useNavigate } from 'react-router'; +import toast from 'react-hot-toast'; interface AppContextValue { // conversations and messages @@ -46,6 +49,9 @@ interface AppContextValue { saveConfig: (config: typeof CONFIG_DEFAULT) => void; showSettings: boolean; setShowSettings: (show: boolean) => void; + + // props + serverProps: LlamaCppServerProps | null; } // this callback is used for scrolling to the bottom of the chat and switching to the last node @@ -74,6 +80,9 @@ export const AppContextProvider = ({ const params = matchPath('/chat/:convId', pathname); const convId = params?.params?.convId; + const [serverProps, setServerProps] = useState( + null + ); const [viewingChat, setViewingChat] = useState(null); const [pendingMessages, setPendingMessages] = useState< Record @@ -85,6 +94,20 @@ export const AppContextProvider = ({ const [canvasData, setCanvasData] = useState(null); const [showSettings, setShowSettings] = useState(false); + // get server props + useEffect(() => { + getServerProps(BASE_URL, config.apiKey) + .then((props) => { + console.debug('Server props:', props); + setServerProps(props); + }) + .catch((err) => { + console.error(err); + toast.error('Failed to fetch server props'); + }); + // eslint-disable-next-line + }, []); + // handle change when the convId from URL is changed useEffect(() => { // also reset the canvas data @@ -260,7 +283,7 @@ export const AppContextProvider = ({ } else { console.error(err); // eslint-disable-next-line @typescript-eslint/no-explicit-any - alert((err as any)?.message ?? 'Unknown error'); + toast.error((err as any)?.message ?? 'Unknown error'); throw err; // rethrow } } @@ -377,6 +400,7 @@ export const AppContextProvider = ({ saveConfig, showSettings, setShowSettings, + serverProps, }} > {children} diff --git a/examples/server/webui/src/utils/common.tsx b/tools/server/webui/src/utils/common.tsx similarity index 56% rename from examples/server/webui/src/utils/common.tsx rename to tools/server/webui/src/utils/common.tsx index 09b08b5c97..372f464a24 100644 --- a/examples/server/webui/src/utils/common.tsx +++ b/tools/server/webui/src/utils/common.tsx @@ -36,3 +36,32 @@ export const OpenInNewTab = ({ {children} ); + +export function BtnWithTooltips({ + className, + onClick, + onMouseLeave, + children, + tooltipsContent, + disabled, +}: { + className?: string; + onClick: () => void; + onMouseLeave?: () => void; + children: React.ReactNode; + tooltipsContent: string; + disabled?: boolean; +}) { + return ( +
+ +
+ ); +} diff --git a/examples/server/webui/src/utils/llama-vscode.ts b/tools/server/webui/src/utils/llama-vscode.ts similarity index 69% rename from examples/server/webui/src/utils/llama-vscode.ts rename to tools/server/webui/src/utils/llama-vscode.ts index c45b0d3973..0ad8f8042e 100644 --- a/examples/server/webui/src/utils/llama-vscode.ts +++ b/tools/server/webui/src/utils/llama-vscode.ts @@ -1,6 +1,6 @@ -import { useEffect, useState } from 'react'; -import { MessageExtraContext } from './types'; +import { useEffect } from 'react'; import { ChatTextareaApi } from '../components/useChatTextarea.ts'; +import { ChatExtraContextApi } from '../components/useChatExtraContext.tsx'; // Extra context when using llama.cpp WebUI from llama-vscode, inside an iframe // Ref: https://github.com/ggml-org/llama.cpp/pull/11940 @@ -15,11 +15,10 @@ interface SetTextEvData { * window.postMessage({ command: 'setText', text: 'Spot the syntax error', context: 'def test()\n return 123' }, '*'); */ -export const useVSCodeContext = (textarea: ChatTextareaApi) => { - const [extraContext, setExtraContext] = useState( - null - ); - +export const useVSCodeContext = ( + textarea: ChatTextareaApi, + extraContext: ChatExtraContextApi +) => { // Accept setText message from a parent window and set inputMsg and extraContext useEffect(() => { const handleMessage = (event: MessageEvent) => { @@ -27,18 +26,25 @@ export const useVSCodeContext = (textarea: ChatTextareaApi) => { const data: SetTextEvData = event.data; textarea.setValue(data?.text); if (data?.context && data.context.length > 0) { - setExtraContext({ - type: 'context', - content: data.context, - }); + extraContext.clearItems(); + extraContext.addItems([ + { + type: 'context', + name: 'Extra context', + content: data.context, + }, + ]); } textarea.focus(); + setTimeout(() => { + textarea.refOnSubmit.current?.(); + }, 10); // wait for setExtraContext to finish } }; window.addEventListener('message', handleMessage); return () => window.removeEventListener('message', handleMessage); - }, [textarea]); + }, [textarea, extraContext]); // Add a keydown listener that sends the "escapePressed" message to the parent window useEffect(() => { @@ -52,9 +58,5 @@ export const useVSCodeContext = (textarea: ChatTextareaApi) => { return () => window.removeEventListener('keydown', handleKeyDown); }, []); - return { - extraContext, - // call once the user message is sent, to clear the extra context - clearExtraContext: () => setExtraContext(null), - }; + return {}; }; diff --git a/examples/server/webui/src/utils/misc.ts b/tools/server/webui/src/utils/misc.ts similarity index 65% rename from examples/server/webui/src/utils/misc.ts rename to tools/server/webui/src/utils/misc.ts index 87f55b2af9..ba760e83bb 100644 --- a/examples/server/webui/src/utils/misc.ts +++ b/tools/server/webui/src/utils/misc.ts @@ -1,6 +1,11 @@ // @ts-expect-error this package does not have typing import TextLineStream from 'textlinestream'; -import { APIMessage, Message } from './types'; +import { + APIMessage, + APIMessageContentPart, + LlamaCppServerProps, + Message, +} from './types'; // ponyfill for missing ReadableStream asyncIterator on Safari import { asyncIterator } from '@sec-ant/readable-stream/ponyfill/asyncIterator'; @@ -57,19 +62,47 @@ export const copyStr = (textToCopy: string) => { */ export function normalizeMsgsForAPI(messages: Readonly) { return messages.map((msg) => { - let newContent = ''; + if (msg.role !== 'user' || !msg.extra) { + return { + role: msg.role, + content: msg.content, + } as APIMessage; + } + + // extra content first, then user text message in the end + // this allow re-using the same cache prefix for long context + const contentArr: APIMessageContentPart[] = []; for (const extra of msg.extra ?? []) { if (extra.type === 'context') { - newContent += `${extra.content}\n\n`; + contentArr.push({ + type: 'text', + text: extra.content, + }); + } else if (extra.type === 'textFile') { + contentArr.push({ + type: 'text', + text: `File: ${extra.name}\nContent:\n\n${extra.content}`, + }); + } else if (extra.type === 'imageFile') { + contentArr.push({ + type: 'image_url', + image_url: { url: extra.base64Url }, + }); + } else { + throw new Error('Unknown extra type'); } } - newContent += msg.content; + // add user message to the end + contentArr.push({ + type: 'text', + text: msg.content, + }); return { role: msg.role, - content: newContent, + content: contentArr, }; }) as APIMessage[]; } @@ -78,13 +111,19 @@ export function normalizeMsgsForAPI(messages: Readonly) { * recommended for DeepsSeek-R1, filter out content between and tags */ export function filterThoughtFromMsgs(messages: APIMessage[]) { + console.debug({ messages }); return messages.map((msg) => { + if (msg.role !== 'assistant') { + return msg; + } + // assistant message is always a string + const contentStr = msg.content as string; return { role: msg.role, content: msg.role === 'assistant' - ? msg.content.split('').at(-1)!.trim() - : msg.content, + ? contentStr.split('').at(-1)!.trim() + : contentStr, } as APIMessage; }); } @@ -126,3 +165,25 @@ export const cleanCurrentUrl = (removeQueryParams: string[]) => { }); window.history.replaceState({}, '', url.toString()); }; + +export const getServerProps = async ( + baseUrl: string, + apiKey?: string +): Promise => { + try { + const response = await fetch(`${baseUrl}/props`, { + headers: { + 'Content-Type': 'application/json', + ...(apiKey ? { Authorization: `Bearer ${apiKey}` } : {}), + }, + }); + if (!response.ok) { + throw new Error('Failed to fetch server props'); + } + const data = await response.json(); + return data as LlamaCppServerProps; + } catch (error) { + console.error('Error fetching server props:', error); + throw error; + } +}; diff --git a/examples/server/webui/src/utils/storage.ts b/tools/server/webui/src/utils/storage.ts similarity index 96% rename from examples/server/webui/src/utils/storage.ts rename to tools/server/webui/src/utils/storage.ts index 1dfc9d9799..505693e927 100644 --- a/examples/server/webui/src/utils/storage.ts +++ b/tools/server/webui/src/utils/storage.ts @@ -116,6 +116,16 @@ const StorageUtils = { }); return conv; }, + /** + * update the name of a conversation + */ + async updateConversationName(convId: string, name: string): Promise { + await db.conversations.update(convId, { + name, + lastModified: Date.now(), + }); + dispatchConversationChange(convId); + }, /** * if convId does not exist, throw an error */ diff --git a/examples/server/webui/src/utils/types.ts b/tools/server/webui/src/utils/types.ts similarity index 78% rename from examples/server/webui/src/utils/types.ts rename to tools/server/webui/src/utils/types.ts index 0eb774001e..ba673dd943 100644 --- a/examples/server/webui/src/utils/types.ts +++ b/tools/server/webui/src/utils/types.ts @@ -48,7 +48,10 @@ export interface Message { children: Message['id'][]; } -type MessageExtra = MessageExtraTextFile | MessageExtraContext; // TODO: will add more in the future +export type MessageExtra = + | MessageExtraTextFile + | MessageExtraImageFile + | MessageExtraContext; export interface MessageExtraTextFile { type: 'textFile'; @@ -56,12 +59,32 @@ export interface MessageExtraTextFile { content: string; } +export interface MessageExtraImageFile { + type: 'imageFile'; + name: string; + base64Url: string; +} + export interface MessageExtraContext { type: 'context'; + name: string; content: string; } -export type APIMessage = Pick; +export type APIMessageContentPart = + | { + type: 'text'; + text: string; + } + | { + type: 'image_url'; + image_url: { url: string }; + }; + +export type APIMessage = { + role: Message['role']; + content: string | APIMessageContentPart[]; +}; export interface Conversation { id: string; // format: `conv-{timestamp}` @@ -89,3 +112,14 @@ export interface CanvasPyInterpreter { } export type CanvasData = CanvasPyInterpreter; + +// a non-complete list of props, only contains the ones we need +export interface LlamaCppServerProps { + build_info: string; + model_path: string; + n_ctx: number; + modalities?: { + vision: boolean; + }; + // TODO: support params +} diff --git a/examples/server/webui/src/vite-env.d.ts b/tools/server/webui/src/vite-env.d.ts similarity index 100% rename from examples/server/webui/src/vite-env.d.ts rename to tools/server/webui/src/vite-env.d.ts diff --git a/examples/server/webui/tailwind.config.js b/tools/server/webui/tailwind.config.js similarity index 100% rename from examples/server/webui/tailwind.config.js rename to tools/server/webui/tailwind.config.js diff --git a/examples/server/webui/tsconfig.app.json b/tools/server/webui/tsconfig.app.json similarity index 100% rename from examples/server/webui/tsconfig.app.json rename to tools/server/webui/tsconfig.app.json diff --git a/examples/server/webui/tsconfig.json b/tools/server/webui/tsconfig.json similarity index 100% rename from examples/server/webui/tsconfig.json rename to tools/server/webui/tsconfig.json diff --git a/examples/server/webui/tsconfig.node.json b/tools/server/webui/tsconfig.node.json similarity index 100% rename from examples/server/webui/tsconfig.node.json rename to tools/server/webui/tsconfig.node.json diff --git a/examples/server/webui/vite.config.ts b/tools/server/webui/vite.config.ts similarity index 98% rename from examples/server/webui/vite.config.ts rename to tools/server/webui/vite.config.ts index b8a0f03d97..366df3b751 100644 --- a/examples/server/webui/vite.config.ts +++ b/tools/server/webui/vite.config.ts @@ -71,6 +71,7 @@ export default defineConfig({ server: { proxy: { '/v1': 'http://localhost:8080', + '/props': 'http://localhost:8080', }, headers: { 'Cross-Origin-Embedder-Policy': 'require-corp', diff --git a/examples/tokenize/CMakeLists.txt b/tools/tokenize/CMakeLists.txt similarity index 100% rename from examples/tokenize/CMakeLists.txt rename to tools/tokenize/CMakeLists.txt diff --git a/examples/tokenize/tokenize.cpp b/tools/tokenize/tokenize.cpp similarity index 100% rename from examples/tokenize/tokenize.cpp rename to tools/tokenize/tokenize.cpp diff --git a/examples/tts/CMakeLists.txt b/tools/tts/CMakeLists.txt similarity index 100% rename from examples/tts/CMakeLists.txt rename to tools/tts/CMakeLists.txt diff --git a/examples/tts/README.md b/tools/tts/README.md similarity index 96% rename from examples/tts/README.md rename to tools/tts/README.md index 4509763c65..557014aebb 100644 --- a/examples/tts/README.md +++ b/tools/tts/README.md @@ -45,7 +45,7 @@ $ popd This model file is PyTorch checkpoint (.ckpt) and we first need to convert it to huggingface format: ```console -(venv) python examples/tts/convert_pt_to_hf.py \ +(venv) python tools/tts/convert_pt_to_hf.py \ models/WavTokenizer-large-speech-75token/wavtokenizer_large_speech_320_24k.ckpt ... Model has been successfully converted and saved to models/WavTokenizer-large-speech-75token/model.safetensors @@ -105,7 +105,7 @@ $ source venv/bin/activate And then run the python script using: ```conole -(venv) python ./examples/tts/tts-outetts.py http://localhost:8020 http://localhost:8021 "Hello world" +(venv) python ./tools/tts/tts-outetts.py http://localhost:8020 http://localhost:8021 "Hello world" spectrogram generated: n_codes: 90, n_embd: 1282 converting to audio ... audio generated: 28800 samples diff --git a/examples/tts/convert_pt_to_hf.py b/tools/tts/convert_pt_to_hf.py similarity index 100% rename from examples/tts/convert_pt_to_hf.py rename to tools/tts/convert_pt_to_hf.py diff --git a/examples/tts/tts-outetts.py b/tools/tts/tts-outetts.py similarity index 100% rename from examples/tts/tts-outetts.py rename to tools/tts/tts-outetts.py diff --git a/examples/tts/tts.cpp b/tools/tts/tts.cpp similarity index 100% rename from examples/tts/tts.cpp rename to tools/tts/tts.cpp