Added Benchmark for Rotary Decode Kernel + Performance Speed Up for Rotary Kernel #280
This file contains 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] | |
| merge_group: | |
| branches: [main_perf] | |
| types: [checks_requested] | |
| push: | |
| branches: [main_perf, micmelesse/upstream_pr] | |
| concurrency: | |
| group: ${{ github.ref }} | |
| cancel-in-progress: true | |
| permissions: read-all | |
| jobs: | |
| Runner-Preparation-AMD: | |
| runs-on: ubuntu-latest | |
| timeout-minutes: 30 | |
| outputs: | |
| matrix-HIP: ${{ steps.set-matrix.outputs.matrix-HIP }} | |
| steps: | |
| - name: Prepare runner matrix | |
| id: set-matrix | |
| run: | | |
| if [ x"${{ github.repository }}" == x"ROCm/flash-attention" ]; then | |
| echo '::set-output name=matrix-HIP::[["self-hosted", "rocm"]]' | |
| else | |
| echo '::set-output name=matrix-HIP::[["ubuntu-latest"]]' | |
| fi | |
| Integration-Tests-AMD: | |
| needs: Runner-Preparation-AMD | |
| if: needs.Runner-Preparation-AMD.outputs.matrix-HIP != '' | |
| runs-on: ${{ matrix.runner }} | |
| strategy: | |
| matrix: | |
| runner: ${{fromJson(needs.Runner-Preparation-AMD.outputs.matrix-HIP)}} | |
| container: | |
| image: rocm/pytorch:rocm6.0.2_ubuntu22.04_py3.10_pytorch_2.1.2 | |
| options: --device=/dev/kfd --device=/dev/dri --security-opt seccomp=unconfined --group-add video --user root | |
| steps: | |
| - name: Checkout | |
| uses: actions/checkout@v4 | |
| - name: Install Triton | |
| run: | | |
| pip uninstall -y triton | |
| pip install matplotlib pandas pytest | |
| git clone https://github.com/triton-lang/triton | |
| cd triton | |
| git checkout 3ca2f498e98ed7249b82722587c511a5610e00c4 | |
| pip install --verbose -e python | |
| cd .. | |
| - name: Build | |
| run: | | |
| export FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE" | |
| python setup.py install | |
| - name: Flash Attention Tests Using Reference Impl | |
| run: | | |
| export FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE" | |
| export FLASH_ATTENTION_TRITON_AMD_REF=1 | |
| pytest tests/test_flash_attn_triton_amd.py | |
| - name: Flash Attention Tests | |
| run: | | |
| export FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE" | |
| pytest tests/test_flash_attn_triton_amd.py | |
| - name: AMD Tests | |
| run: | | |
| export FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE" | |
| pytest -v -s flash_attn/flash_attn_triton_amd/test.py | |
| - name: AMD Bench | |
| run: | | |
| export FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE" | |
| python flash_attn/flash_attn_triton_amd/bench.py | |
| - name: AMD Bench with Autotune | |
| run: | | |
| export FLASH_ATTENTION_TRITON_AMD_ENABLE="TRUE" | |
| export FLASH_ATTENTION_TRITON_AMD_AUTOTUNE=1 | |
| python flash_attn/flash_attn_triton_amd/bench.py |