Add a script for tuning flash attention kernels #1658
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 Offline Tests | |
on: | |
workflow_dispatch: | |
pull_request: | |
branches: [triton-mlir] | |
merge_group: | |
branches: [triton-mlir] | |
types: [checks_requested] | |
push: | |
branches: [triton-mlir] | |
concurrency: | |
group: ${{ github.ref }} | |
cancel-in-progress: ${{ github.ref != 'refs/heads/triton-mlir' }} | |
permissions: read-all | |
env: | |
TRITON_BUILD_WITH_CLANG_LLD: "TRUE" | |
TRITON_USE_ASSERT_ENABLED_LLVM: "TRUE" | |
TRITON_DISABLE_LINE_INFO: 1 | |
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/triton" ]; then | |
echo '::set-output name=matrix-HIP::[["self-hosted", "rocm.gfx90a"]]' | |
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 }} | |
timeout-minutes: 30 | |
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: Clear cache | |
run: | | |
rm -rf ~/.triton | |
mkdir -p ~/.triton | |
ls -alh ~/.triton | |
- name: Update PATH | |
run: | | |
echo "/opt/rocm/llvm/bin" >> $GITHUB_PATH | |
- name: Install pip dependencies | |
run: | | |
python3 -m pip install --upgrade pip | |
python3 -m pip install lit matplotlib pandas | |
- name: Install Triton | |
run: | | |
echo "PATH is '$PATH'" | |
pip uninstall -y triton | |
cd python | |
pip install -v -e . | |
- name: Run Perf Kernels Unit Tests | |
run: | | |
pytest -vvv ./python/perf-kernels/flash-attention.py | |
- name: Run Perf Kernels Benchmark | |
run: | | |
python ./python/perf-kernels/flash-attention.py | |
- name: Run Core Tests | |
run: | | |
pytest -n 32 ./python/test/unit/language/test_core_amd.py | |
- name: Run lit tests | |
run: | | |
python3 -m pip install lit | |
cd python | |
LIT_TEST_DIR="build/$(ls build | grep -i cmake)/test" | |
if [ ! -d "${LIT_TEST_DIR}" ]; then | |
echo "Could not find '${LIT_TEST_DIR}'" ; exit -1 | |
fi | |
lit -v "${LIT_TEST_DIR}" | |
- name: Run CXX unittests | |
run: | | |
cd python | |
cd "build/$(ls build | grep -i cmake)" | |
ctest |