354 lines
18 KiB
YAML
354 lines
18 KiB
YAML
name: Check Changes
|
|
|
|
on:
|
|
workflow_call:
|
|
inputs:
|
|
pr_head_sha:
|
|
type: string
|
|
default: ''
|
|
git_ref:
|
|
type: string
|
|
default: ''
|
|
target_stage:
|
|
type: string
|
|
default: ''
|
|
include_wheel_build:
|
|
type: boolean
|
|
default: false
|
|
run_all_tests:
|
|
type: boolean
|
|
default: false
|
|
force_continue_on_error:
|
|
type: boolean
|
|
default: false
|
|
outputs:
|
|
main_package:
|
|
value: ${{ jobs.run.outputs.main_package }}
|
|
sgl_kernel:
|
|
value: ${{ jobs.run.outputs.sgl_kernel }}
|
|
sgl_kernel_raw:
|
|
value: ${{ jobs.run.outputs.sgl_kernel_raw }}
|
|
jit_kernel:
|
|
value: ${{ jobs.run.outputs.jit_kernel }}
|
|
multimodal_gen:
|
|
value: ${{ jobs.run.outputs.multimodal_gen }}
|
|
partitions:
|
|
value: ${{ jobs.run.outputs.partitions }}
|
|
partition_model_sha:
|
|
value: ${{ jobs.run.outputs.partition_model_sha }}
|
|
b200_runner:
|
|
value: ${{ jobs.run.outputs.b200_runner }}
|
|
enable_retry:
|
|
value: ${{ jobs.run.outputs.enable_retry }}
|
|
continue_on_error:
|
|
value: ${{ jobs.run.outputs.continue_on_error }}
|
|
|
|
jobs:
|
|
run:
|
|
runs-on: ubuntu-latest
|
|
outputs:
|
|
# Use API-based detection for target_stage mode (filter-api), otherwise use dorny/paths-filter (filter)
|
|
main_package: ${{ steps.filter-api.outputs.main_package || steps.filter.outputs.main_package || steps.run-mode.outputs.run_all_tests }}
|
|
# sgl_kernel is forced to false when target_stage is set AND include_wheel_build is NOT set,
|
|
# since sgl-kernel-build-wheels normally skips in target_stage mode. When include_wheel_build
|
|
# is true, keep the real value so the wheel build runs and the target stage downloads its
|
|
# artifact (used by /rerun-stage on PRs that modify sgl-kernel/).
|
|
# This prevents CUSTOM_BUILD_SGL_KERNEL=true when the wheel artifacts aren't available.
|
|
# Note: If PR has kernel changes AND target_stage is set AND include_wheel_build is NOT set,
|
|
# the validate-target-stage step will fail.
|
|
sgl_kernel: ${{ (!inputs.target_stage || inputs.include_wheel_build) && (steps.filter-api.outputs.sgl_kernel || steps.filter.outputs.sgl_kernel) }}
|
|
# Raw sgl_kernel value before target_stage override (used for validation)
|
|
sgl_kernel_raw: ${{ steps.filter-api.outputs.sgl_kernel || steps.filter.outputs.sgl_kernel }}
|
|
jit_kernel: ${{ steps.filter-api.outputs.jit_kernel || steps.filter.outputs.jit_kernel || steps.run-mode.outputs.run_all_tests }}
|
|
multimodal_gen: ${{ steps.filter-api.outputs.multimodal_gen || steps.filter.outputs.multimodal_gen || steps.run-mode.outputs.run_all_tests }}
|
|
partitions: ${{ steps.partitions.outputs.partitions }}
|
|
partition_model_sha: ${{ steps.partition-model-sha.outputs.sha }}
|
|
b200_runner: ${{ steps.set-runner.outputs.b200_runner }}
|
|
enable_retry: ${{ steps.set-retry.outputs.enable_retry }}
|
|
continue_on_error: ${{ steps.set-continue-on-error.outputs.continue_on_error }}
|
|
steps:
|
|
- name: Checkout code
|
|
uses: actions/checkout@v4
|
|
with:
|
|
ref: ${{ inputs.pr_head_sha || inputs.git_ref || github.sha }}
|
|
|
|
- uses: ./.github/actions/check-maintenance
|
|
|
|
- name: Determine run mode
|
|
id: run-mode
|
|
run: |
|
|
# Run all tests for scheduled runs and workflow_call (when ref input is provided)
|
|
# Note: github.event_name is inherited from caller, so we detect workflow_call by checking inputs.git_ref
|
|
if [[ "${{ github.event_name }}" == "schedule" || "${{ inputs.run_all_tests }}" == "true" ]]; then
|
|
echo "run_all_tests=true" >> $GITHUB_OUTPUT
|
|
echo "Run mode: ALL TESTS (schedule=${{ github.event_name == 'schedule' }}, run_all_tests=${{ inputs.run_all_tests }})"
|
|
else
|
|
echo "run_all_tests=false" >> $GITHUB_OUTPUT
|
|
echo "Run mode: FILTERED (triggered by ${{ github.event_name }})"
|
|
fi
|
|
|
|
- name: Detect file changes
|
|
id: filter
|
|
uses: dorny/paths-filter@v3
|
|
# Only use paths-filter for pull_request events (where it works correctly)
|
|
# For workflow_dispatch with target_stage, we use GitHub API in the next step
|
|
if: steps.run-mode.outputs.run_all_tests != 'true' && !inputs.target_stage
|
|
with:
|
|
filters: |
|
|
main_package:
|
|
- ".github/workflows/pr-test.yml"
|
|
- ".github/workflows/pr-gate.yml"
|
|
- ".github/actions/**"
|
|
- "python/pyproject.toml"
|
|
- "python/sglang/!(multimodal_gen)/**/!(*.md)"
|
|
- "scripts/ci/cuda/*"
|
|
- "scripts/ci/utils/*"
|
|
- "test/**/!(*.md)"
|
|
multimodal_gen:
|
|
- ".github/workflows/pr-test.yml"
|
|
- ".github/workflows/pr-test-multimodal-gen.yml"
|
|
- "python/pyproject.toml"
|
|
- "python/sglang/multimodal_gen/**/!(*.md|*.ipynb)"
|
|
- "python/sglang/jit_kernel/**"
|
|
- "python/sglang/jit_kernel/tests/diffusion/**"
|
|
- "python/sglang/jit_kernel/benchmark/diffusion/**"
|
|
- "python/sglang/cli/**"
|
|
jit_kernel:
|
|
- ".github/workflows/pr-test.yml"
|
|
- ".github/workflows/pr-test-jit-kernel.yml"
|
|
- "python/pyproject.toml"
|
|
- "python/sglang/jit_kernel/**"
|
|
sgl_kernel:
|
|
# Intentionally excludes ".github/workflows/pr-test-sgl-kernel.yml" —
|
|
# see API-side detector below for rationale.
|
|
- "sgl-kernel/**/!(*.md|THIRDPARTYNOTICES.txt|LICENSE)"
|
|
|
|
# For /rerun-stage (workflow_dispatch with target_stage), dorny/paths-filter doesn't work
|
|
# correctly because it falls back to "last commit" detection which breaks for merge commits.
|
|
# Instead, we use the GitHub API to compare the PR commit against main.
|
|
- name: Detect file changes via API (for target_stage)
|
|
id: filter-api
|
|
if: inputs.target_stage && inputs.pr_head_sha
|
|
env:
|
|
GH_TOKEN: ${{ github.token }}
|
|
run: |
|
|
echo "Detecting file changes via GitHub API for target_stage mode..."
|
|
echo "PR head SHA: ${{ inputs.pr_head_sha }}"
|
|
|
|
# Get the list of changed files by comparing PR commit against main
|
|
# This correctly handles merge commits by looking at the actual PR diff
|
|
CHANGED_FILES=$(gh api "repos/${{ github.repository }}/compare/main...${{ inputs.pr_head_sha }}" \
|
|
--jq '[.files[].filename] | .[]' 2>/dev/null || echo "")
|
|
|
|
if [ -z "$CHANGED_FILES" ]; then
|
|
echo "Warning: Could not fetch changed files from API, assuming no changes"
|
|
echo "sgl_kernel=false" >> $GITHUB_OUTPUT
|
|
echo "main_package=false" >> $GITHUB_OUTPUT
|
|
echo "jit_kernel=false" >> $GITHUB_OUTPUT
|
|
echo "multimodal_gen=false" >> $GITHUB_OUTPUT
|
|
exit 0
|
|
fi
|
|
|
|
echo "Changed files:"
|
|
echo "$CHANGED_FILES" | head -20
|
|
echo "..."
|
|
|
|
# Check for sgl-kernel changes
|
|
# Note: edits to .github/workflows/pr-test-sgl-kernel.yml are intentionally
|
|
# NOT considered sgl-kernel changes. That filter line used to be included
|
|
# so workflow refactors got retested, but in practice it only catches the
|
|
# workflow's *consumers* (test job definitions), not the wheel build steps
|
|
# themselves — and gating sgl_kernel=true on it forces a 20-30 min wheel
|
|
# rebuild + the stage-a-test-1-gpu-small gate for pure CI-yaml edits that
|
|
# can't actually affect kernel behavior. PRs that touch wheel-build logic
|
|
# in scripts/ci/cuda/ or sgl-kernel/ still trigger correctly.
|
|
if echo "$CHANGED_FILES" | grep -qE "^sgl-kernel/"; then
|
|
echo "sgl_kernel=true" >> $GITHUB_OUTPUT
|
|
echo "Detected sgl-kernel changes"
|
|
else
|
|
echo "sgl_kernel=false" >> $GITHUB_OUTPUT
|
|
fi
|
|
|
|
# Check for main_package changes (excluding multimodal_gen, jit_kernel/diffusion, jit_kernel/tests/diffusion, jit_kernel/benchmark/diffusion, cli)
|
|
# Note: Need to filter out multimodal_gen and diffusion-related paths before checking, not pipe grep -q output
|
|
MAIN_PKG_FILES=$(echo "$CHANGED_FILES" | grep -E "^(python/sglang/|python/pyproject\.toml|scripts/ci/cuda/|scripts/ci/utils/|test/|\.github/workflows/pr-test\.yml|\.github/workflows/pr-gate\.yml|\.github/actions/)" | grep -v -E "^(python/sglang/multimodal_gen/|python/sglang/jit_kernel/diffusion/|python/sglang/jit_kernel/tests/diffusion/|python/sglang/jit_kernel/benchmark/diffusion/|python/sglang/cli/)" || true)
|
|
if [ -n "$MAIN_PKG_FILES" ]; then
|
|
echo "main_package=true" >> $GITHUB_OUTPUT
|
|
echo "Detected main_package changes"
|
|
else
|
|
echo "main_package=false" >> $GITHUB_OUTPUT
|
|
fi
|
|
|
|
# Check for jit_kernel changes
|
|
if echo "$CHANGED_FILES" | grep -qE "^(python/sglang/jit_kernel/|python/pyproject\.toml|\.github/workflows/pr-test\.yml|\.github/workflows/pr-test-jit-kernel\.yml)"; then
|
|
echo "jit_kernel=true" >> $GITHUB_OUTPUT
|
|
echo "Detected jit_kernel changes"
|
|
else
|
|
echo "jit_kernel=false" >> $GITHUB_OUTPUT
|
|
fi
|
|
|
|
# Check for multimodal_gen changes, including diffusion-specific jit_kernel coverage
|
|
if echo "$CHANGED_FILES" | grep -qE "^(python/sglang/multimodal_gen/|python/sglang/cli/|python/sglang/jit_kernel/diffusion/|python/sglang/jit_kernel/tests/diffusion/|python/sglang/jit_kernel/benchmark/diffusion/|python/pyproject\.toml|\.github/workflows/pr-test\.yml|\.github/workflows/pr-test-multimodal-gen\.yml)"; then
|
|
echo "multimodal_gen=true" >> $GITHUB_OUTPUT
|
|
echo "Detected multimodal_gen changes"
|
|
else
|
|
echo "multimodal_gen=false" >> $GITHUB_OUTPUT
|
|
fi
|
|
|
|
- name: Determine full-parallel mode
|
|
id: parallel-mode
|
|
env:
|
|
GH_TOKEN: ${{ github.token }}
|
|
run: |
|
|
# `full=true` lifts the matrix-fanout throttle so each suite's
|
|
# max_parallel = size. Conditions (matching the prior set-parallel
|
|
# step exactly):
|
|
# 1. Scheduled cron run.
|
|
# 2. pull_request event with the `high priority` label.
|
|
# 3. workflow_dispatch with target_stage set (i.e. /rerun-stage)
|
|
# whose underlying PR carries `high priority`. The labels
|
|
# aren't on the dispatch payload, so look them up via API:
|
|
# try SHA -> /pulls (works for fork PRs), fall back to
|
|
# branch name -> gh pr list (works for non-fork PRs).
|
|
FULL=false
|
|
if [[ "${{ github.event_name }}" == "schedule" ]]; then
|
|
FULL=true
|
|
echo "Scheduled run -> full parallelism"
|
|
elif [[ "${{ github.event_name }}" == "pull_request" && "${{ contains(github.event.pull_request.labels.*.name, 'high priority') }}" == "true" ]]; then
|
|
FULL=true
|
|
echo "high priority PR -> full parallelism"
|
|
elif [[ -n "${{ inputs.target_stage }}" ]]; then
|
|
LABELS=""
|
|
PR_HEAD_SHA="${{ inputs.pr_head_sha }}"
|
|
if [[ -n "$PR_HEAD_SHA" ]]; then
|
|
LABELS=$(gh api "repos/${{ github.repository }}/commits/${PR_HEAD_SHA}/pulls" \
|
|
--jq '.[0].labels[].name' 2>/dev/null || true)
|
|
fi
|
|
if [[ -z "$LABELS" ]]; then
|
|
LABELS=$(gh pr list --head "${{ github.ref_name }}" --repo "${{ github.repository }}" \
|
|
--json labels --jq '.[0].labels[].name' 2>/dev/null || true)
|
|
fi
|
|
echo "PR labels (via API): ${LABELS:-"(none)"}"
|
|
if echo "$LABELS" | grep -Fxq "high priority"; then
|
|
FULL=true
|
|
echo "high priority PR (via API) -> full parallelism"
|
|
fi
|
|
fi
|
|
echo "full=$FULL" >> "$GITHUB_OUTPUT"
|
|
|
|
- name: Resolve sglang-ci-stats SHA
|
|
id: partition-model-sha
|
|
env:
|
|
GH_TOKEN: ${{ github.token }}
|
|
run: |
|
|
# Pin all shards to one immutable commit so dispatch and every
|
|
# runtime LPT use the same model snapshot. Soft fail -> static.
|
|
SHA=$(gh api repos/sgl-project/sglang-ci-stats/commits/main --jq '.sha' 2>/dev/null || true)
|
|
if [[ -n "$SHA" ]]; then
|
|
echo "Pinned sglang-ci-stats@$SHA"
|
|
echo "sha=$SHA" >> "$GITHUB_OUTPUT"
|
|
else
|
|
echo "::warning::Could not resolve sglang-ci-stats SHA; using in-source est_time"
|
|
echo "sha=" >> "$GITHUB_OUTPUT"
|
|
fi
|
|
|
|
- name: Fetch live partition model
|
|
if: steps.partition-model-sha.outputs.sha != ''
|
|
run: |
|
|
# SHA resolved -> require fetch (curl --retry). Soft fallback
|
|
# would risk cross-shard LPT divergence on transient curl flake.
|
|
rm -f /tmp/partition-model.json
|
|
URL="https://raw.githubusercontent.com/sgl-project/sglang-ci-stats/${{ steps.partition-model-sha.outputs.sha }}/model.json"
|
|
curl --fail --silent --show-error --max-time 15 --retry 3 --retry-delay 2 \
|
|
"$URL" -o /tmp/partition-model.json
|
|
echo "Fetched partition-model.json ($(wc -c < /tmp/partition-model.json) bytes)"
|
|
|
|
- name: Compute partitions
|
|
id: partitions
|
|
run: |
|
|
# Emit a single JSON output `partitions` keyed by suite name with
|
|
# {size, arr, max_parallel} fields per suite. Replaces the prior
|
|
# full/low max-parallel presets; `--full-parallel` keeps the
|
|
# `high priority` PR / scheduled cron escape hatch.
|
|
# See scripts/ci/utils/compute_partitions.py.
|
|
python3 scripts/ci/utils/compute_partitions.py \
|
|
--full-parallel ${{ steps.parallel-mode.outputs.full }} \
|
|
--partition-model-file /tmp/partition-model.json \
|
|
>> "$GITHUB_OUTPUT"
|
|
|
|
- name: Set B200 runner tag
|
|
id: set-runner
|
|
run: |
|
|
# Use kernel-build runner only when sgl_kernel changes are detected AND we're not in target_stage mode
|
|
# (target_stage skips wheel builds, so we can't use custom kernels)
|
|
# Use API-based detection (filter-api) for target_stage mode, otherwise use dorny/paths-filter (filter)
|
|
sgl_kernel="${{ steps.filter-api.outputs.sgl_kernel || steps.filter.outputs.sgl_kernel }}"
|
|
target_stage="${{ inputs.target_stage }}"
|
|
if [[ "$sgl_kernel" == "true" && -z "$target_stage" ]]; then
|
|
echo "b200_runner=4-gpu-b200-kernel" >> $GITHUB_OUTPUT
|
|
else
|
|
echo "b200_runner=4-gpu-b200" >> $GITHUB_OUTPUT
|
|
fi
|
|
|
|
- name: Enable retry for CI
|
|
id: set-retry
|
|
run: |
|
|
echo "enable_retry=true" >> $GITHUB_OUTPUT
|
|
echo "Retry logic enabled for CI"
|
|
|
|
- name: Set continue-on-error for full test runs
|
|
id: set-continue-on-error
|
|
run: |
|
|
if [[ "${{ steps.run-mode.outputs.run_all_tests }}" == "true" || "${{ inputs.force_continue_on_error }}" == "true" ]]; then
|
|
echo "continue_on_error=true" >> $GITHUB_OUTPUT
|
|
echo "Full test run or force flag detected, enabling continue-on-error to run all tests"
|
|
else
|
|
echo "continue_on_error=false" >> $GITHUB_OUTPUT
|
|
echo "Filtered run, continue-on-error disabled"
|
|
fi
|
|
|
|
- name: Validate target_stage with kernel changes
|
|
# Fail only when PR has sgl-kernel changes AND the caller didn't opt into include_wheel_build.
|
|
# include_wheel_build=true means sgl-kernel-build-wheels will run alongside the target stage
|
|
# (see the sgl_kernel output and sgl-kernel-build-wheels if-conditions above/below), so it's
|
|
# safe to proceed.
|
|
if: inputs.target_stage && !inputs.include_wheel_build && (steps.filter-api.outputs.sgl_kernel == 'true' || steps.filter.outputs.sgl_kernel == 'true')
|
|
run: |
|
|
echo "::error::Cannot use /rerun-stage when PR has sgl-kernel changes without include_wheel_build."
|
|
echo "::error::The sgl-kernel-build-wheels job is skipped in target_stage mode by default, but this PR modifies sgl-kernel/ files."
|
|
echo "::error::The slash-command handler should have set include_wheel_build=true automatically; falling back to /tag-and-rerun-ci."
|
|
echo ""
|
|
echo "ERROR: Cannot use /rerun-stage when PR has sgl-kernel changes without include_wheel_build."
|
|
echo ""
|
|
echo "This PR modifies files in sgl-kernel/, which requires building custom kernel wheels."
|
|
echo "Running the target stage without rebuilding the kernel would use the wrong (PyPI)"
|
|
echo "version of sgl-kernel instead of your changes."
|
|
echo ""
|
|
echo "The /rerun-stage handler sets include_wheel_build=true automatically when it detects"
|
|
echo "sgl-kernel/ changes on the PR. If you see this error, the handler may be outdated."
|
|
echo ""
|
|
echo "Alternatives:"
|
|
echo " /tag-and-rerun-ci - Re-run the full workflow including kernel builds"
|
|
echo " /rerun-ci - Re-run the full workflow"
|
|
echo ""
|
|
exit 1
|
|
|
|
- name: Show filter results in summary (table)
|
|
run: |
|
|
{
|
|
echo "## Change Detection"
|
|
echo ""
|
|
echo "| Component | Changed |"
|
|
echo "|-------------------|---------|"
|
|
echo "| main_package | ${{ steps.filter-api.outputs.main_package || steps.filter.outputs.main_package || steps.run-mode.outputs.run_all_tests }} |"
|
|
echo "| sgl_kernel (raw) | ${{ steps.filter-api.outputs.sgl_kernel || steps.filter.outputs.sgl_kernel }} |"
|
|
echo "| sgl_kernel (used) | ${{ (!inputs.target_stage || inputs.include_wheel_build) && (steps.filter-api.outputs.sgl_kernel || steps.filter.outputs.sgl_kernel) }} |"
|
|
echo "| jit_kernel | ${{ steps.filter-api.outputs.jit_kernel || steps.filter.outputs.jit_kernel || steps.run-mode.outputs.run_all_tests }} |"
|
|
echo "| multimodal_gen | ${{ steps.filter-api.outputs.multimodal_gen || steps.filter.outputs.multimodal_gen || steps.run-mode.outputs.run_all_tests }} |"
|
|
echo "| target_stage | ${{ inputs.target_stage || '(none)' }} |"
|
|
echo "| detection_method | ${{ inputs.target_stage && 'GitHub API' || 'dorny/paths-filter' }} |"
|
|
echo "| b200_runner | ${{ steps.set-runner.outputs.b200_runner }} |"
|
|
echo "| enable_retry | ${{ steps.set-retry.outputs.enable_retry }} |"
|
|
echo "| continue_on_error | ${{ steps.set-continue-on-error.outputs.continue_on_error }} |"
|
|
} >> $GITHUB_STEP_SUMMARY
|