Implementation of tiled attention with bf16 and circular buffers which reduces memory requirements by 4x on longer context on gemma models. #6481
Workflow file for this run
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| name: build | |
| # Trigger on push, pull request, or via manual dispatch. | |
| on: | |
| push: | |
| pull_request: | |
| types: [opened, reopened, labeled, unlabeled, synchronize] | |
| workflow_dispatch: | |
| jobs: | |
| build: | |
| runs-on: ${{ matrix.os }} | |
| name: ${{ matrix.os }} (${{ matrix.preset }}) ${{ matrix.build_type }} | |
| timeout-minutes: 30 | |
| strategy: | |
| fail-fast: false | |
| matrix: | |
| # When adding another, also add to copybara's github_check_runs. | |
| os: ['ubuntu-latest', 'macos-latest', 'windows-latest'] | |
| build_type: ['Release'] | |
| preset: ['make', 'windows'] | |
| exclude: | |
| - os: ubuntu-latest | |
| preset: windows | |
| - os: macos-latest | |
| preset: windows | |
| - os: windows-latest | |
| preset: make | |
| concurrency: | |
| group: ${{ github.workflow }}-${{ github.ref }}-${{ matrix.os }}-${{ matrix.preset }}-${{ matrix.build_type }} | |
| cancel-in-progress: true | |
| steps: | |
| - uses: actions/checkout@v6 | |
| # Set up ccache | |
| - name: ccache | |
| uses: hendrikmuhs/ccache-action@v1.2 | |
| - name: Configure CMake | |
| run: > | |
| cmake --preset ${{ matrix.preset }} | |
| -S ${{ github.workspace }} -B ${{ github.workspace }}/build | |
| -D CMAKE_BUILD_TYPE=${{ matrix.build_type }} | |
| -D CMAKE_C_COMPILER_LAUNCHER=ccache | |
| -D CMAKE_CXX_COMPILER_LAUNCHER=ccache | |
| -DCMAKE_POLICY_VERSION_MINIMUM=3.5 | |
| - name: Build | |
| run: cmake --build ${{ github.workspace }}/build --preset ${{ matrix.preset }} --config ${{ matrix.build_type }} -j 4 | |
| - name: Archive production artifacts | |
| uses: actions/upload-artifact@v6 | |
| with: | |
| name: gemma-${{ matrix.os }}-${{ matrix.preset }}-${{ matrix.build_type }} | |
| path: | | |
| ${{ github.workspace }}/build/${{ matrix.build_type }}/gemma.exe | |
| ${{ github.workspace }}/build/${{ matrix.build_type }}/libgemma.lib | |
| ${{ github.workspace }}/build/gemma | |
| ${{ github.workspace }}/build/libgemma.a | |
| bazel: | |
| runs-on: ubuntu-latest | |
| steps: | |
| - name: Harden Runner | |
| uses: step-security/harden-runner@63c24ba6bd7ba022e95695ff85de572c04a18142 # v2.7.0 | |
| with: | |
| egress-policy: audit # cannot be block - runner does git checkout | |
| - uses: actions/checkout@8ade135a41bc03ea155e62e844d188df1ea18608 # v4.0.0 | |
| - uses: bazelbuild/setup-bazelisk@b39c379c82683a5f25d34f0d062761f62693e0b2 # v3.0.0 | |
| - uses: actions/cache@d4323d4df104b026a6aa633fdb11d772146be0bf # v4.2.2 | |
| with: | |
| path: ~/.cache/bazel | |
| key: bazel-${{ runner.os }} | |
| - run: bazel build --cxxopt=-std=c++20 //:gemma --jobs=10 --show_progress_rate_limit=1 |