Skip to content

fix(build): suppress MLX folding warning after flags #154

fix(build): suppress MLX folding warning after flags

fix(build): suppress MLX folding warning after flags #154

Workflow file for this run

# Triton AOT vendored-artifact sync.
#
# Two jobs:
# 1. drift-check (GitHub-hosted, no GPU): pure sha256 compare of
# triton_kernels/*.py vs every vendored MANIFEST. Runs on any push/PR that
# touches the kernels, the vendored tree, or the AOT cmake. FAILS on drift,
# so a kernel change cannot merge while shipping stale cubins.
# 2. regen-pr (SELF-HOSTED GPU runner, label [self-hosted, gpu]): actually
# reruns scripts/regen-triton-aot.sh and opens a sync PR with the refreshed
# artifacts, then executes them on CUDA. Generation is target-pinned; runtime
# validation requires the self-hosted GPU runner (e.g. dgx.casa).
# Trigger manually (workflow_dispatch) or via the weekly schedule; it only
# opens a PR when the drift check says the tree is stale.
name: triton-aot-sync
on:
push:
branches: [main]
paths:
- '.github/workflows/triton-aot-sync.yml'
- 'CMakeLists.txt'
- 'triton_kernels/**'
- 'src/vt/cuda/triton_aot_vendored/**'
- 'cmake/TritonAOT.cmake'
- 'cmake/TritonAOTKernels.cmake'
- 'cmake/DumpTritonAOTContract.cmake'
- 'scripts/regen-triton-aot.sh'
- 'scripts/check-triton-aot-drift.sh'
- 'scripts/triton-aot-compile.py'
- 'tests/scripts/test_triton_aot_drift.sh'
pull_request:
paths:
- '.github/workflows/triton-aot-sync.yml'
- 'CMakeLists.txt'
- 'triton_kernels/**'
- 'src/vt/cuda/triton_aot_vendored/**'
- 'cmake/TritonAOT.cmake'
- 'cmake/TritonAOTKernels.cmake'
- 'cmake/DumpTritonAOTContract.cmake'
- 'scripts/regen-triton-aot.sh'
- 'scripts/check-triton-aot-drift.sh'
- 'scripts/triton-aot-compile.py'
- 'tests/scripts/test_triton_aot_drift.sh'
schedule:
- cron: '17 5 * * 1' # weekly drift audit (Mon 05:17 UTC)
workflow_dispatch: {}
permissions:
contents: write
pull-requests: write
jobs:
drift-check:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- name: Check vendored Triton AOT artifacts vs kernel sources
run: bash scripts/check-triton-aot-drift.sh
regen-pr:
# Only meaningful with a GPU runner registered; scheduled/manual only so
# ordinary pushes never queue on a runner that may not exist.
if: github.event_name == 'workflow_dispatch' || github.event_name == 'schedule'
runs-on: [self-hosted, gpu]
needs: []
timeout-minutes: 60
steps:
- uses: actions/checkout@v4
- name: Skip if already in sync
id: drift
run: |
if bash scripts/check-triton-aot-drift.sh; then
echo "stale=false" >> "$GITHUB_OUTPUT"
else
echo "stale=true" >> "$GITHUB_OUTPUT"
fi
- name: Regenerate vendored artifacts
if: steps.drift.outputs.stale == 'true'
run: flock -w 7200 /tmp/gpu bash scripts/regen-triton-aot.sh
- name: Verify regenerated artifact contract and vendored consumer
if: steps.drift.outputs.stale == 'true'
run: |
bash scripts/check-triton-aot-drift.sh
cmake -S . -B build-triton-aot-consumer \
-DVLLM_CPP_CUDA=ON \
-DVLLM_CPP_CUDA_ARCHITECTURES=121a \
-DVLLM_CPP_TRITON=ON \
-DVLLM_CPP_TRITON_REGEN=OFF \
-DVLLM_CPP_BUILD_TESTS=ON \
-DVLLM_CPP_BUILD_EXAMPLES=OFF \
-DVLLM_CPP_SERVER=OFF
cmake --build build-triton-aot-consumer --target vllm test_ops_gdn -j2
flock -w 7200 /tmp/gpu \
build-triton-aot-consumer/tests/test_ops_gdn
- name: Open sync PR
if: steps.drift.outputs.stale == 'true'
env:
GH_TOKEN: ${{ github.token }}
run: |
set -euo pipefail
git config user.name "triton-aot-sync[bot]"
git config user.email "[email protected]"
branch="sync/triton-aot-$(date -u +%Y%m%d-%H%M)"
git checkout -b "$branch"
git add src/vt/cuda/triton_aot_vendored/
git commit -F - <<'MSG'
build: sync vendored Triton AOT artifacts with triton_kernels sources (automated regen)
Automated regen via scripts/regen-triton-aot.sh after drift was detected by
scripts/check-triton-aot-drift.sh. Review the MANIFEST + artifact diff, then
merge; CI's drift-check gates that the result is coherent.
FOLLOWING_AGENTS_PROTOCOL
Assisted-by: triton-aot-sync:workflow [GitHubActions]
MSG
git push origin "$branch"
gh pr create --base main --head "$branch" \
--title "build: sync vendored Triton AOT artifacts (automated regen)" \
--body "Automated regen of \`src/vt/cuda/triton_aot_vendored/\` after \`scripts/check-triton-aot-drift.sh\` detected drift vs \`triton_kernels/*.py\`. Review the MANIFEST diff (triton/ptxas/toolkit pins + per-kernel source hashes) and the artifact churn, then merge. Generated on a self-hosted GPU runner via \`scripts/regen-triton-aot.sh\`."