Sliding Window block classification logic #747
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: AMD Perf Kernel Tests | |
| on: | |
| workflow_dispatch: | |
| pull_request: | |
| branches: [main_perf] | |
| concurrency: | |
| group: ${{ github.workflow }}-${{ github.ref }} | |
| cancel-in-progress: true | |
| jobs: | |
| Integration-Tests-AMD: | |
| runs-on: ${{ matrix.runner }} | |
| strategy: | |
| matrix: | |
| runner: [linux-mi300-gpu-1] | |
| fail-fast: false # disables failing the entire job when one matrix entry fails | |
| timeout-minutes: 720 # self hosted runners can run jobs for longer than the default of 360 minutes | |
| container: | |
| image: rocm/pytorch:latest | |
| options: --device=/dev/kfd --device=/dev/dri --security-opt seccomp=unconfined --shm-size 16G --group-add video --user root | |
| steps: | |
| - name: Checkout | |
| uses: actions/checkout@v4 | |
| - name: Show Device Info | |
| run: | | |
| rocminfo | grep gfx | |
| - name: Uninstall Triton | |
| run: | | |
| pip uninstall -y triton | |
| rm -rf ~/.triton | |
| rm -rf ./triton/python/build | |
| - name: Install Triton | |
| run: | | |
| pip install triton==3.3.0 | |
| - name: Show Triton version | |
| run: | | |
| pip show triton | |
| - name: Build | |
| run: | | |
| FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE" python setup.py install | |
| - name: Install dependencies for bench and misc | |
| run: | | |
| pip install matplotlib pandas tabulate | |
| - name: AMD Internal Tests | |
| run: | | |
| FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE" FLASH_ATTENTION_TRITON_AMD_AUTOTUNE=0 pytest flash_attn/flash_attn_triton_amd/test.py | |
| - name: Flash Attention Tests | |
| run: | | |
| FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE" FLASH_ATTENTION_TRITON_AMD_AUTOTUNE=0 pytest -n 8 tests/test_flash_attn_triton_amd.py | |
| - name: AMD Bench | |
| run: | | |
| python flash_attn/flash_attn_triton_amd/bench.py -benchmark_fn flash_attn_func | |
| python flash_attn/flash_attn_triton_amd/bench.py -benchmark_fn flash_attn_varlen_func | |
| python flash_attn/flash_attn_triton_amd/bench.py -benchmark_fn flash_attn_with_kvcache |