pr-sglang.yml

name: PR - SGLang

on:
  pull_request:
    branches: [main]
    types: [opened, reopened, synchronize]
    paths:
      - "**sglang**"

permissions:
  contents: read
  pull-requests: read

env:
  # CI Image configuration
  CONTAINER_TYPE: "general"
  FRAMEWORK: "sglang"
  SGLANG_VERSION: "0.5.5"
  PYTHON_VERSION: "py312"
  CUDA_VERSION: "cu129"
  OS_VERSION: "ubuntu22.04"
  # Prod Image configuration
  PROD_SAGEMAKER_IMAGE: sglang:0.5-gpu-py312
  # CI environment configuration
  FORCE_COLOR: "1"
  TEST_ARTIFACTS_DIRECTORY: "/test_artifacts/sglang"

jobs:
  pr-validation:
    uses: ./.github/workflows/reusable/pr-validation.yml
    with:
      workflow-id: ${{ github.event.pull_request.number }}
      concurrency-group: "${{ github.workflow }}-pr-validation"
      concurrency-cancel-in-progress: true
      required-permission-level: "write"
      python-version: "3.12"
      run-pre-commit: true
      pre-commit-args: "--all-files"
      change-filters: |
        build-change:
          - "docker/sglang/**"
          - "scripts/sglang/**"
          - "scripts/common/**"
          - "scripts/telemetry/**"
          - ".github/workflows/pr-sglang*"
        test-change:
          - "test/sglang/**"

  read_input:
    runs-on: ubuntu-latest
    needs: [pr-validation]
    outputs:
      sglang-version: ${{ env.SGLANG_VERSION }}
      python-version: ${{ env.PYTHON_VERSION }}
      cuda-version: ${{ env.CUDA_VERSION }}
      os-version: ${{ env.OS_VERSION }}
      container-type: ${{ env.CONTAINER_TYPE }}
      framework: ${{ env.FRAMEWORK }}
      prod-sagemaker-image: ${{ env.PROD_SAGEMAKER_IMAGE }}
      test-artifacts-directory: ${{ env.TEST_ARTIFACTS_DIRECTORY }}
    steps:
      - name: Setup variables
        run: |
          echo "Setting up environment variables as job outputs"
          echo "SGLANG_VERSION=${{ env.SGLANG_VERSION }}"
          echo "PYTHON_VERSION=${{ env.PYTHON_VERSION }}"
          echo "CUDA_VERSION=${{ env.CUDA_VERSION }}"
          echo "OS_VERSION=${{ env.OS_VERSION }}"
          echo "CONTAINER_TYPE=${{ env.CONTAINER_TYPE }}"
          echo "FRAMEWORK=${{ env.FRAMEWORK }}"
          echo "PROD_SAGEMAKER_IMAGE=${{ env.PROD_SAGEMAKER_IMAGE }}"
          echo "TEST_ARTIFACTS_DIRECTORY=${{ env.TEST_ARTIFACTS_DIRECTORY }}"

  # ======================================================
  # =============== SGLang SageMaker jobs ================
  # ======================================================
  build-sglang-image:
    needs: [pr-validation, read_input]
    if: needs.pr-validation.outputs.build-change == 'true'
    uses: ./.github/workflows/reusable/build-container-image.yml
    with:
      image-name: "ci"
      image-tag: "sglang-${{ needs.read_input.outputs.sglang-version }}-gpu-${{ needs.read_input.outputs.python-version }}-${{ needs.read_input.outputs.cuda-version }}-${{ needs.read_input.outputs.os-version }}-sagemaker-pr-${{ github.event.pull_request.number }}"
      dockerfile-path: "docker/sglang/Dockerfile"
      build-context: "."
      build-target: "sglang-sagemaker"
      base-image: "lmsysorg/sglang:v${{ needs.read_input.outputs.sglang-version }}-${{ needs.read_input.outputs.cuda-version }}-amd64"
      build-args: |
        {
          "CACHE_REFRESH": "${{ github.run_number }}",
          "BASE_IMAGE": "lmsysorg/sglang:v${{ needs.read_input.outputs.sglang-version }}-${{ needs.read_input.outputs.cuda-version }}-amd64",
          "CONTAINER_TYPE": "${{ needs.read_input.outputs.container-type }}",
          "FRAMEWORK": "${{ needs.read_input.outputs.framework }}",
          "FRAMEWORK_VERSION": "${{ needs.read_input.outputs.sglang-version }}"
        }
      workflow-id: ${{ github.event.pull_request.number }}
      runs-on: '["codebuild-runner-${{ github.run_id }}-${{ github.run_attempt }}", "fleet:x86-build-runner", "buildspec-override:true"]'
      concurrency-group: "${{ github.workflow }}-build-sglang-image"
      concurrency-cancel-in-progress: true
      registry-type: "ecr"
      enable-cache: true
      cache-from: "registry"
      cache-to: "inline"
      push-image: true
      cleanup-image: true
      pre-build-script: ".github/scripts/buildkitd.sh"
    secrets:
      AWS_ACCOUNT_ID: ${{ vars.CI_AWS_ACCOUNT_ID }}
      AWS_REGION: ${{ vars.AWS_REGION }}

  set-test-environment:
    needs: [pr-validation, read_input, build-sglang-image]
    uses: ./.github/workflows/reusable/set-test-environment.yml
    with:
      environment-name: "sglang"
      build-job-result: ${{ needs.build-sglang-image.result }}
      ci-image-uri: ${{ needs.build-sglang-image.outputs.image-uri }}
      prod-image-name: ${{ needs.read_input.outputs.prod-sagemaker-image }}
      workflow-id: ${{ github.event.pull_request.number }}
      change-conditions: '["${{ needs.pr-validation.outputs.build-change }}", "${{ needs.pr-validation.outputs.test-change }}"]'
      concurrency-group: "${{ github.workflow }}-set-test-environment"
      concurrency-cancel-in-progress: true
    secrets:
      CI_AWS_ACCOUNT_ID: ${{ vars.CI_AWS_ACCOUNT_ID }}
      PROD_AWS_ACCOUNT_ID: ${{ vars.PROD_AWS_ACCOUNT_ID }}
      AWS_REGION: ${{ vars.AWS_REGION }}

  sglang-local-benchmark-test:
    needs: [read_input, set-test-environment, build-sglang-image]
    if: success()
    uses: ./.github/workflows/reusable/test-execution.yml
    with:
      test-name: "sglang-local-benchmark"
      test-category: "benchmark"
      execution-mode: "container"
      container-execution-style: "external"
      aws-account-id: ${{ needs.set-test-environment.outputs.aws-account-id }}
      image-uri: ${{ needs.set-test-environment.outputs.image-uri }}
      workflow-id: ${{ github.event.pull_request.number }}
      run-id: ${{ github.run_id }}
      run-attempt: ${{ github.run_attempt }}
      runs-on: '["codebuild-runner-${{ github.run_id }}-${{ github.run_attempt }}", "fleet:x86-g6xl-runner", "buildspec-override:true"]'
      concurrency-group: "${{ github.workflow }}-sglang-local-benchmark-test"
      concurrency-cancel-in-progress: true
      container-gpus: "all"
      container-volumes: '["${HOME}/.cache/huggingface:/root/.cache/huggingface", "${{ needs.read_input.outputs.test-artifacts-directory }}/dataset:/dataset"]'
      container-ports: '["30000:30000"]'
      container-environment: '["SM_SGLANG_MODEL_PATH=Qwen/Qwen3-0.6B", "SM_SGLANG_REASONING_PARSER=qwen3", "SM_SGLANG_HOST=127.0.0.1", "SM_SGLANG_PORT=30000"]'
      container-environment-secrets: '["HF_TOKEN"]'
      pre-setup-commands: |
        mkdir -p ${{ needs.read_input.outputs.test-artifacts-directory }}/dataset
        if [ ! -f ${{ needs.read_input.outputs.test-artifacts-directory }}/dataset/ShareGPT_V3_unfiltered_cleaned_split.json ]; then
            echo "Downloading ShareGPT dataset..."
            wget -P ${{ needs.read_input.outputs.test-artifacts-directory }}/dataset https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
        else
            echo "ShareGPT dataset already exists. Skipping download."
        fi
      container-startup-delay: 120
      show-container-logs: true
      test-commands: |
        python3 -m sglang.bench_serving \
        --backend sglang \
        --host 127.0.0.1 --port 30000 \
        --num-prompts 1000 \
        --model Qwen/Qwen3-0.6B \
        --dataset-name sharegpt \
        --dataset-path /dataset/ShareGPT_V3_unfiltered_cleaned_split.json
    secrets:
      AWS_REGION: ${{ vars.AWS_REGION }}

  sglang-frontend-test:
    needs: [read_input, build-sglang-image, set-test-environment]
    if: success()
    uses: ./.github/workflows/reusable/test-execution.yml
    with:
      test-name: "sglang-frontend"
      test-category: "frontend"
      aws-account-id: ${{ needs.set-test-environment.outputs.aws-account-id }}
      image-uri: ${{ needs.set-test-environment.outputs.image-uri }}
      workflow-id: ${{ github.event.pull_request.number }}
      run-id: ${{ github.run_id }}
      run-attempt: ${{ github.run_attempt }}
      external-repo: "sgl-project/sglang"
      external-repo-ref: "v${{ needs.read_input.outputs.sglang-version }}"
      external-repo-path: "sglang_source"
      runs-on: '["codebuild-runner-${{ github.run_id }}-${{ github.run_attempt }}", "fleet:x86-g6exl-runner", "buildspec-override:true"]'
      concurrency-group: "${{ github.workflow }}-sglang-frontend-test"
      concurrency-cancel-in-progress: true
      container-gpus: "all"
      container-entrypoint: "/bin/bash"
      container-volumes: '["${HOME}/.cache/huggingface:/root/.cache/huggingface", "./sglang_source:/workdir"]'
      container-environment: '[]'
      container-environment-secrets: '["HF_TOKEN"]'
      container-workdir: "/workdir"
      setup-commands: |
        bash scripts/ci/ci_install_dependency.sh
      test-commands: |
        nvidia-smi
        # Frontend Test
        cd /workdir/test/lang
        python3 run_suite.py --suite per-commit
    secrets:
      AWS_REGION: ${{ vars.AWS_REGION }}

  sglang-sagemaker-endpoint-test:
    needs: [set-test-environment]
    if: |
      always() && !failure() && !cancelled() &&
      needs.set-test-environment.result == 'success'
    uses: ./.github/workflows/reusable/test-execution.yml
    with:
      test-name: "sglang-sagemaker-endpoint"
      test-category: "integration"
      execution-mode: "local"
      workflow-id: ${{ github.event.pull_request.number }}
      run-id: ${{ github.run_id }}
      run-attempt: ${{ github.run_attempt }}
      runs-on: '["codebuild-runner-${{ github.run_id }}-${{ github.run_attempt }}", "fleet:default-runner", "buildspec-override:true"]'
      concurrency-group: "${{ github.workflow }}-sglang-sagemaker-endpoint-test"
      concurrency-cancel-in-progress: false
      ecr-authenticate: false
      local-setup-commands: |
        uv venv
        source .venv/bin/activate
        uv pip install -r test/requirements.txt
        uv pip install -r test/sglang/sagemaker/requirements.txt
      local-test-commands: |
        source .venv/bin/activate
        cd test/
        python3 -m pytest -vs -rA --image-uri ${{ needs.set-test-environment.outputs.image-uri }} sglang/sagemaker
    secrets:
      AWS_REGION: ${{ vars.AWS_REGION }}