[AMD] ci: add label-gated extra-a tier (kv_canary + mock_model unit tests) (#27822)
This commit is contained in:
Michael
@@ -0,0 +1,174 @@
|
||||
name: PR Test Extra (AMD)
|
||||
# Label-gated AMD extra CI workflow — the AMD mirror of pr-test-extra.yml.
|
||||
#
|
||||
# Adds AMD runtime to a PR only when the author opts in: the PR must carry
|
||||
# BOTH `run-ci` (basic-CI prerequisite) and `run-ci-extra` (explicit opt-in).
|
||||
# The label check happens at runtime in pr-gate.yml via a live
|
||||
# `gh pr view`-style fetch, so reruns after adding the labels (e.g. via a
|
||||
# slash command) pick up the new label set — a workflow-level `if` would read
|
||||
# the frozen event payload, which never updates on rerun. The job graph also
|
||||
# runs unconditionally on workflow_dispatch / workflow_call so it can be
|
||||
# triggered manually or chained from the AMD scheduler.
|
||||
#
|
||||
# Stage: extra-a (1-gpu-small-amd). The job mirrors the container bring-up of
|
||||
# pr-test-amd.yml and dispatches `run_suite.py --hw amd --suite
|
||||
# extra-a-test-1-gpu-small-amd`. Only the mock-model / kv_canary *unit* tests
|
||||
# are onboarded so far; the canary *e2e* tests (which would land in
|
||||
# 1-/2-gpu-large) need the canary JIT kernel ported to ROCm first, so those
|
||||
# suites are intentionally not registered for AMD yet.
|
||||
|
||||
on:
|
||||
pull_request:
|
||||
# `labeled` lets the workflow re-fire when `run-ci-extra` (or `run-ci`)
|
||||
# is added after the latest push. See call-gate.if for the matching guard
|
||||
# that prevents unrelated label additions from dispatching a full run.
|
||||
types: [opened, synchronize, reopened, labeled]
|
||||
workflow_dispatch:
|
||||
inputs:
|
||||
runner_arch:
|
||||
description: 'AMD runner pool to dispatch GPU jobs to'
|
||||
required: false
|
||||
type: choice
|
||||
default: mi325
|
||||
options:
|
||||
- mi300
|
||||
- mi325
|
||||
rocm_version:
|
||||
description: 'ROCm container variant (empty = Dockerfile default; rocm720 = ROCm 7.2.0)'
|
||||
required: false
|
||||
type: choice
|
||||
default: ''
|
||||
options:
|
||||
- ''
|
||||
- rocm720
|
||||
aiter_ref:
|
||||
description: 'Override AITER commit (optional, leave empty to use Dockerfile default)'
|
||||
required: false
|
||||
type: string
|
||||
default: ''
|
||||
continue_on_error:
|
||||
description: 'Continue on error (do not fail the workflow on test failures)'
|
||||
required: false
|
||||
type: boolean
|
||||
default: false
|
||||
workflow_call:
|
||||
inputs:
|
||||
ref:
|
||||
description: 'Git ref (branch, tag, or SHA) to test. If not provided, uses the default branch.'
|
||||
required: false
|
||||
type: string
|
||||
default: ''
|
||||
rocm_version:
|
||||
description: 'ROCm container variant (empty = Dockerfile default; rocm720 = ROCm 7.2.0)'
|
||||
required: false
|
||||
type: string
|
||||
default: ''
|
||||
aiter_ref:
|
||||
description: 'Override AITER commit (optional, leave empty to use Dockerfile default)'
|
||||
required: false
|
||||
type: string
|
||||
default: ''
|
||||
continue_on_error:
|
||||
description: 'Continue on error (do not fail the workflow on test failures)'
|
||||
required: false
|
||||
type: boolean
|
||||
default: false
|
||||
|
||||
env:
|
||||
AITER_COMMIT_OVERRIDE: ${{ inputs.aiter_ref }}
|
||||
DOCKERHUB_AMD_USERNAME: ${{ secrets.DOCKERHUB_AMD_USERNAME }}
|
||||
DOCKERHUB_AMD_TOKEN: ${{ secrets.DOCKERHUB_AMD_TOKEN }}
|
||||
|
||||
concurrency:
|
||||
group: pr-test-amd-extra-${{ github.event_name }}-${{ github.head_ref || github.ref_name || 'default' }}-${{ inputs.ref || 'all' }}
|
||||
cancel-in-progress: ${{ github.event_name != 'workflow_call' }}
|
||||
|
||||
permissions:
|
||||
actions: write
|
||||
contents: read
|
||||
issues: read
|
||||
pull-requests: read
|
||||
|
||||
jobs:
|
||||
# =============================================== PR Gate ====================================================
|
||||
# Runtime live-fetch label gate (mirrors pr-test-extra.yml's call-gate):
|
||||
# requires both `run-ci` and `run-ci-extra`. A failure here cascades to
|
||||
# every test job via `needs`, so a PR without the labels ends in one red
|
||||
# ~30s gate job plus a row of skipped jobs instead of consuming AMD runners.
|
||||
#
|
||||
# The job-level `if` only filters the `labeled` event type so that adding an
|
||||
# unrelated label doesn't dispatch a full run; the actual label-presence
|
||||
# gate is enforced at runtime inside pr-gate.yml.
|
||||
call-gate:
|
||||
if: |
|
||||
github.event_name != 'pull_request' ||
|
||||
github.event.action != 'labeled' ||
|
||||
github.event.label.name == 'run-ci' ||
|
||||
github.event.label.name == 'run-ci-extra'
|
||||
uses: ./.github/workflows/pr-gate.yml
|
||||
with:
|
||||
require-run-ci: true
|
||||
require-run-ci-extra: true
|
||||
secrets: inherit
|
||||
|
||||
# =============================================== extra-a (1-gpu-small) ===============================================
|
||||
# Single unpartitioned job: the 21 onboarded unit tests total ~233s, so the
|
||||
# expensive per-job setup (container bring-up + sgl-kernel ROCm build + dep
|
||||
# install, several minutes) dominates. Partitioning would multiply that
|
||||
# setup across scarce AMD GPUs to shave only a couple minutes of test time,
|
||||
# so one GPU running the whole suite sequentially is the better trade.
|
||||
extra-a-test-1-gpu-small-amd:
|
||||
name: ${{ format('extra-a-test-1-gpu-small-amd{0} (linux-{1}-1gpu-sglang)', inputs.rocm_version && format('-{0}', inputs.rocm_version) || '', inputs.runner_arch || 'mi325') }}
|
||||
needs: [call-gate]
|
||||
if: ${{ !cancelled() && needs.call-gate.result == 'success' }}
|
||||
runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || 'mi325') }}
|
||||
steps:
|
||||
- name: Checkout code
|
||||
uses: actions/checkout@v4
|
||||
with:
|
||||
ref: ${{ inputs.ref || github.sha }}
|
||||
|
||||
- name: Ensure VRAM is clear
|
||||
run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
|
||||
|
||||
- name: Start CI container
|
||||
# `rocm_version` (e.g. rocm720) selects an alternate ROCm container; empty uses the Dockerfile default.
|
||||
run: bash scripts/ci/amd/amd_ci_start_container.sh ${{ inputs.rocm_version && format('--rocm-version {0}', inputs.rocm_version) || '' }}
|
||||
env:
|
||||
GITHUB_WORKSPACE: ${{ github.workspace }}
|
||||
|
||||
- name: Install dependencies
|
||||
run: bash scripts/ci/amd/amd_ci_install_dependency.sh
|
||||
|
||||
- name: Run test
|
||||
timeout-minutes: 45
|
||||
run: |
|
||||
bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite extra-a-test-1-gpu-small-amd --timeout-per-file 1800 ${{ inputs.continue_on_error == true && '--continue-on-error' || '' }}
|
||||
|
||||
# =============================================== aggregator ====================================================
|
||||
# Single fan-in job so branch protection / notifications depend on one job
|
||||
# rather than every matrix leg. Fails if any dependent failed or was
|
||||
# cancelled; `skipped` (e.g. PR without the opt-in labels) is not a failure.
|
||||
pr-test-amd-extra-finish:
|
||||
needs:
|
||||
[
|
||||
call-gate,
|
||||
extra-a-test-1-gpu-small-amd,
|
||||
]
|
||||
if: always()
|
||||
runs-on: ubuntu-latest
|
||||
steps:
|
||||
- name: Check all dependent job statuses
|
||||
run: |
|
||||
json_needs='${{ toJson(needs) }}'
|
||||
job_names=$(echo "$json_needs" | jq -r 'keys_unsorted[]')
|
||||
for job in $job_names; do
|
||||
result=$(echo "$json_needs" | jq -r --arg j "$job" '.[$j].result')
|
||||
echo "$job: $result"
|
||||
if [[ "$result" == "failure" || "$result" == "cancelled" ]]; then
|
||||
echo "The above jobs failed."
|
||||
exit 1
|
||||
fi
|
||||
done
|
||||
echo "All jobs completed successfully"
|
||||
exit 0
|
||||
@@ -203,6 +203,26 @@ jobs:
|
||||
- "python/pyproject_rocm.toml"
|
||||
- "python/pyproject_other.toml"
|
||||
|
||||
# =============================================== extra (scheduled) ====================================================
|
||||
# ROCm 7.2 mirror of pr-test-amd.yml's `call-pr-test-amd-extra`: chain the
|
||||
# label-gated AMD extra tier into this workflow's daily schedule, but in a
|
||||
# ROCm 7.2 container (`rocm_version: rocm720`). On `schedule` (and
|
||||
# run_all_tests dispatch) the extra suite runs on `main` without the
|
||||
# `run-ci-extra` label (pr-gate.yml only enforces labels on pull_request
|
||||
# events). Targeted dispatches (target_stage set) are excluded. Not wired
|
||||
# into any finish aggregator so the base rocm720 run never depends on it.
|
||||
call-pr-test-amd-extra-rocm720:
|
||||
if: |
|
||||
(github.event_name == 'schedule' || inputs.run_all_tests == true) &&
|
||||
!(inputs.target_stage || inputs.target_stage_select)
|
||||
uses: ./.github/workflows/pr-test-amd-extra.yml
|
||||
with:
|
||||
ref: ${{ inputs.pr_head_sha || inputs.ref || '' }}
|
||||
rocm_version: rocm720
|
||||
aiter_ref: ${{ inputs.aiter_ref }}
|
||||
continue_on_error: true
|
||||
secrets: inherit
|
||||
|
||||
# =============================================== sgl-kernel ====================================================
|
||||
sgl-kernel-unit-test-amd-rocm720:
|
||||
needs: [check-changes]
|
||||
|
||||
@@ -187,6 +187,25 @@ jobs:
|
||||
- "python/pyproject_rocm.toml"
|
||||
- "python/pyproject_other.toml"
|
||||
|
||||
# =============================================== extra (scheduled) ====================================================
|
||||
# Chain the label-gated AMD extra tier into the scheduled run, mirroring
|
||||
# pr-test.yml's `call-pr-test-extra`. On `schedule` (and run_all_tests
|
||||
# dispatch) the extra suite runs on `main` without needing the
|
||||
# `run-ci-extra` label (pr-gate.yml only enforces labels on pull_request
|
||||
# events). Targeted /rerun-stage dispatches (target_stage set) are excluded.
|
||||
# Not added to `pr-test-amd-finish` so the base AMD gate never depends on
|
||||
# the opt-in extra suite.
|
||||
call-pr-test-amd-extra:
|
||||
if: |
|
||||
(github.event_name == 'schedule' || inputs.run_all_tests == true) &&
|
||||
!(inputs.target_stage || inputs.target_stage_select)
|
||||
uses: ./.github/workflows/pr-test-amd-extra.yml
|
||||
with:
|
||||
ref: ${{ inputs.pr_head_sha || inputs.ref || '' }}
|
||||
aiter_ref: ${{ inputs.aiter_ref }}
|
||||
continue_on_error: true
|
||||
secrets: inherit
|
||||
|
||||
# =============================================== sgl-kernel ====================================================
|
||||
sgl-kernel-unit-test-amd:
|
||||
name: ${{ format('sgl-kernel-unit-test-amd (linux-{0}-1gpu-sglang)', inputs.runner_arch || 'mi325') }}
|
||||
|
||||
Reference in New Issue
Block a user