ktransformers/.github/workflows/model-e2e.yml

name: Model E2E acceptance (PR / PyPI)

# Separate from release workflows: this workflow cannot upload packages.
# Only maintainers dispatch reviewed SHAs; never execute pull_request_target code.
on:
  workflow_dispatch:
    inputs:
      mode:
        description: '验收对象：pr 为候选 wheels，pypi 为公开安装'
        type: choice
        options: [pypi, pr]
        default: pypi
      version:
        description: 'PyPI 的 ktransformers 版本；latest 按用户默认安装'
        type: string
        default: latest
      pr_repository:
        description: 'PR 所属四仓之一'
        type: choice
        options: [kvcache-ai/ktransformers, kvcache-ai/sglang, kvcache-ai/transformers, kvcache-ai/accelerate]
        default: kvcache-ai/ktransformers
      pr_number:
        description: 'PR 编号（pr 模式必填）'
        type: string
        default: ''
      pr_sha:
        description: '已审阅批准的完整 PR head SHA（pr 模式必填）'
        type: string
        default: ''
      build_run_id:
        description: '可信 main 构建流程的 run ID，包含 model-e2e-candidate artifact'
        type: string
        default: ''
  workflow_call:
    inputs:
      mode:
        type: string
        default: pypi
      version:
        type: string
        default: latest
      pr_repository:
        type: string
        default: kvcache-ai/ktransformers
      pr_number:
        type: string
        default: ''
      pr_sha:
        type: string
        default: ''
      build_run_id:
        type: string
        default: ''

permissions:
  contents: read

# Deliberately no shared Actions concurrency group: GitHub may replace an
# existing pending run. The runner queue + host-wide flock preserve waiting jobs.
jobs:
  prepare:
    if: github.repository == 'kvcache-ai/ktransformers' && github.ref == 'refs/heads/main'
    runs-on: ubuntu-24.04
    timeout-minutes: 20
    permissions:
      contents: read
      actions: read
    steps:
      - uses: actions/checkout@v4
        with:
          persist-credentials: false
      - uses: actions/setup-python@v5
        with:
          python-version: '3.12'
      - name: Freeze the approved test identity (no PR code execution)
        env:
          GH_TOKEN: ${{ github.token }}
          ENABLED: ${{ vars.KT_MODEL_E2E_ENABLED }}
          MODE: ${{ inputs.mode }}
          VERSION: ${{ inputs.version }}
          PR_REPOSITORY: ${{ inputs.pr_repository }}
          PR_NUMBER: ${{ inputs.pr_number }}
          PR_SHA: ${{ inputs.pr_sha }}
          BUILD_RUN_ID: ${{ inputs.build_run_id }}
          BUILD_WORKFLOW: ${{ vars.KT_MODEL_E2E_BUILD_WORKFLOW }}
        run: |
          python .github/model-e2e/prepare.py
          mkdir input
          mv request.json input/request.json
      - name: Download final candidates from the approved build
        if: inputs.mode == 'pr'
        uses: actions/download-artifact@v4
        with:
          name: model-e2e-candidate
          path: input/candidate
          run-id: ${{ inputs.build_run_id }}
          github-token: ${{ github.token }}
      - name: Check wheel hashes and source lock
        if: inputs.mode == 'pr'
        run: |
          python - <<'PY'
          import json, sys
          sys.path.insert(0, '.github/model-e2e')
          from contracts import verify_candidate
          verify_candidate('input/candidate', json.load(open('input/request.json')))
          PY
      - uses: actions/upload-artifact@v4
        with:
          name: model-e2e-input-${{ github.run_id }}-${{ github.run_attempt }}
          path: input/
          if-no-files-found: error
          retention-days: 7

  gpu:
    needs: prepare
    # An administrator must provision the no-credentials runner account and
    # protected environment before assigning this new label. Existing release
    # runners must NOT receive the label without the documented isolation audit.
    environment: qj5090-model-e2e
    runs-on: [self-hosted, linux, x64, gpu, kt-model-e2e]
    timeout-minutes: 1100
    permissions:
      contents: read
      actions: read
    steps:
      - uses: actions/checkout@v4
        with:
          persist-credentials: false
          path: harness
      - uses: actions/download-artifact@v4
        with:
          name: model-e2e-input-${{ github.run_id }}-${{ github.run_attempt }}
          path: input
      - name: Queue, install fresh packages, and run the three model tests
        env:
          HOST_CONFIG: ${{ vars.KT_MODEL_E2E_HOST_CONFIG || '/etc/ktransformers/model-e2e/qj5090.json' }}
          CONTROLLER_PYTHON: ${{ vars.KT_MODEL_E2E_PYTHON || 'python3.12' }}
        run: |
          set -euo pipefail
          test "$("$CONTROLLER_PYTHON" -c 'import sys; print(sys.version_info >= (3, 11))')" = True
          test_root="$(mktemp -d "$RUNNER_TEMP/kt-model-e2e.XXXXXXXX")"
          echo "E2E_OUTPUT=$test_root/output" >> "$GITHUB_ENV"
          "$CONTROLLER_PYTHON" harness/.github/model-e2e/host.py \
            --config "$HOST_CONFIG" --input "$GITHUB_WORKSPACE/input" \
            --output "$test_root/output"
      - name: Retain evidence only (not model weights, adapters, or the venv)
        if: always() && env.E2E_OUTPUT != ''
        uses: actions/upload-artifact@v4
        with:
          name: model-e2e-evidence-${{ github.run_id }}-${{ github.run_attempt }}
          path: ${{ env.E2E_OUTPUT }}/evidence/
          if-no-files-found: warn
          retention-days: 30

  report:
    if: always()
    needs: [prepare, gpu]
    runs-on: ubuntu-24.04
    timeout-minutes: 10
    permissions:
      contents: read
      actions: read
      statuses: write
    steps:
      - uses: actions/checkout@v4
        with:
          persist-credentials: false
      - uses: actions/download-artifact@v4
        if: needs.prepare.result == 'success'
        with:
          name: model-e2e-input-${{ github.run_id }}-${{ github.run_attempt }}
          path: input
      - uses: actions/download-artifact@v4
        if: needs.gpu.result == 'success' || needs.gpu.result == 'failure'
        continue-on-error: true
        with:
          name: model-e2e-evidence-${{ github.run_id }}-${{ github.run_attempt }}
          path: evidence
      - name: Summarize and report to the approved KT PR SHA
        env:
          GH_TOKEN: ${{ github.token }}
          GPU_JOB_RESULT: ${{ needs.gpu.result }}
        run: python3 .github/model-e2e/report.py