sgl-project--sglang/.github/workflows/pr-test-amd-extra.yml

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. Each job mirrors the container bring-up of pr-test-amd.yml
# and dispatches `run_suite.py --hw amd --suite extra-a-test-{config}-amd`:
#   - 1-gpu-small: mock-model / kv_canary unit + single-GPU canary e2e tests
#   - 1-gpu-large: single-GPU model e2e tests (quant fp8kv-triton,
#                  sessions streaming-session, spec standalone triton,
#                  kv_canary self-bench)
#   - 2-gpu-large: multi-GPU (TP/PP/PD) mock-model + kv_canary e2e tests
# kv_canary e2e is registered to the same extra-a stage as its CUDA siblings
# (it now exercises the ROCm canary kernels end-to-end).

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:
  # Mirror pr-test-amd.yml: schedule/dispatch runs key on run_id (unique group,
  # never cancel each other); PR runs share a per-branch group so pushes cancel
  # stale runs. (In a reusable workflow github.event_name is the originating
  # event, never workflow_call, so the guard keys off the real events.)
  group: pr-test-amd-extra-${{ (github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && format('full-{0}', github.run_id) || github.head_ref || github.ref_name || inputs.ref || 'default' }}
  cancel-in-progress: ${{ github.event_name == 'pull_request' }}

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' || '' }}

  # =============================================== extra-a (1-gpu-large) ===============================================
  # Single-GPU lora / spec / quant model e2e tests. Runs on the same 1-gpu
  # pool as small (AMD GPUs carry enough VRAM that "large" here is a CUDA
  # memory-tier label, not a separate AMD runner pool).
  extra-a-test-1-gpu-large-amd:
    name: ${{ format('extra-a-test-1-gpu-large-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
        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: 60
        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-large-amd --timeout-per-file 1800 ${{ inputs.continue_on_error == true && '--continue-on-error' || '' }}

  # =============================================== extra-a (2-gpu-large) ===============================================
  # Multi-GPU TP / PP / PD mock-model + kv_canary e2e tests. Mirrors CUDA's
  # extra-a 2-gpu-large stage; runs on the 2-GPU AMD pool.
  extra-a-test-2-gpu-large-amd:
    name: ${{ format('extra-a-test-2-gpu-large-amd{0} (linux-{1}-2gpu-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}-2gpu-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
        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: 60
        run: |
          bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite extra-a-test-2-gpu-large-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,
        extra-a-test-1-gpu-large-amd,
        extra-a-test-2-gpu-large-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