python.yaml

name: Python CI

on:
  push:
    branches:
      - main
  pull_request:
    types: ['opened', 'reopened', 'synchronize', 'ready_for_review']

concurrency:
  group: ${{ github.workflow }}-${{ github.ref }}
  cancel-in-progress: true

jobs:
  lint:
    # A push event cannot tell whether its branch belongs to a draft PR, so
    # branch CI is driven by pull_request events and push is limited to main.
    if: ${{ github.event_name != 'pull_request' || github.event.pull_request.draft == false }}
    runs-on: ${{ matrix.os }}
    strategy:
      fail-fast: false
      matrix:
        os: [ "ubuntu-latest" ]
        python-version: [ "3.10" ]
    steps:
      - name: Check out code
        uses: actions/checkout@v7
        with:
          fetch-depth: 0
          submodules: recursive
      - name: Set up Python environment
        uses: actions/setup-python@v4
        with:
          python-version: "3.10"
      - name: Install pre-commit
        run: pip install pre-commit
      - name: Run pre-commit
        run: pre-commit run --all-files
      - name: Set up Node.js
        uses: actions/setup-node@v1
        with:
          node-version: 20.19.0
      # ESLint and Prettier must be in `package.json`
      - name: Install Node.js dependencies
        run: cd frontend && npm ci
      - name: ESLint Check
        run: cd frontend && npx eslint .
      - name: Build frontend (static export)
        run: cd frontend && npm run build
      - name: Verify static export output
        run: |
          test -f frontend/out/index.html
          test -f frontend/out/404.html
          # Dynamic routes must emit __shell__ placeholder pages.
          find frontend/out -name '__shell__*' | grep -q .
      - name: Serve static export from backend
        run: |
          pip install fastapi httpx pytest xoscar
          pytest --noconftest -vv \
            xinference/api/tests/test_frontend_static.py \
            xinference/api/tests/test_frontend_static_real_export.py


  changes:
    if: ${{ github.event_name != 'pull_request' || github.event.pull_request.draft == false }}
    runs-on: ubuntu-latest
    outputs:
      gpu: ${{ steps.gpu.outputs.gpu }}
      groups: ${{ steps.gpu.outputs.groups }}
    steps:
      - name: Check out code
        uses: actions/checkout@v7
        with:
          fetch-depth: 0
          submodules: recursive

      - name: Detect GPU CI changes
        id: gpu
        shell: bash
        run: |
          set -uo pipefail

          if [ "${{ github.event_name }}" = "pull_request" ]; then
            base_sha="${{ github.event.pull_request.base.sha }}"
            head_sha="${{ github.event.pull_request.head.sha }}"
            # Match GitHub's PR diff: exclude commits added to the base branch
            # after the PR branch diverged.
            diff_range="$base_sha...$head_sha"
          else
            base_sha="${{ github.event.before }}"
            head_sha="${{ github.sha }}"
            diff_range="$base_sha..$head_sha"
          fi

          full=false
          llm=false; embedding=false; image=false; audio=false

          # Fall back to a full run when there is no usable diff base (new
          # branch, or a force push whose previous head no longer exists).
          if [ -z "$base_sha" ] || [[ "$base_sha" =~ ^0+$ ]] \
              || ! git diff --name-only "$diff_range" > changed-files.txt 2>/dev/null; then
            full=true
          else
            # Map each changed path to the GPU test group(s) it affects.
            # Paths matching no pattern (docs, examples, web UI, other
            # workflows, ...) do not need GPU CI at all.
            while IFS= read -r file; do
              case "$file" in
                xinference/model/audio/*) audio=true ;;
                xinference/model/image/*) image=true ;;
                xinference/model/embedding/*|xinference/model/rerank/*) embedding=true ;;
                xinference/model/llm/*|xinference/model/scheduler/*|xinference/core/tests/test_continuous_batching.py) llm=true ;;
                # Vendored model code is audio-only today (cosyvoice, f5_tts,
                # matcha, fish_speech, ...).
                xinference/thirdparty/*) audio=true ;;
                # The web UI never affects GPU tests.
                xinference/ui/*) ;;
                # Tests under the shared runtime packages exercise CPU logic
                # that the ubuntu/macos/windows CI already covers; editing a
                # test file does not change the GPU launch/inference path, so
                # it needs no GPU CI. (test_continuous_batching is matched by
                # the llm rule above and still triggers the llm group.)
                xinference/core/tests/*|xinference/api/tests/*|xinference/client/tests/*) ;;
                # OAuth2 / auth and the admin & launch-history routers are pure
                # HTTP/auth logic with no model inference path; the non-GPU CI
                # covers them. Keep them out of GPU CI entirely.
                xinference/api/oauth2/*|xinference/api/routers/admin.py|xinference/api/routers/launch_history.py) ;;
                # Per-family HTTP routers map to a single GPU group instead of
                # forcing a full run.
                xinference/api/routers/llm.py) llm=true ;;
                xinference/api/routers/embeddings.py|xinference/api/routers/rerank.py) embedding=true ;;
                xinference/api/routers/images.py) image=true ;;
                xinference/api/routers/audio.py) audio=true ;;
                # Every GPU test launches models through a real RESTful API
                # server subprocess and the xinference/client/ package (see
                # conftest.py's `setup` fixture), so changes to either affect
                # every family. The same fixture starts the test cluster through
                # xinference/deploy/, and xinference/core/ is the actor runtime
                # shared by every model family. Any of these affects every group.
                # This also covers restful_api.py (the launch/inference
                # endpoints), dependencies.py, and the models/videos routers.
                xinference/core/*|xinference/api/*|xinference/client/*|xinference/deploy/*) full=true; break ;;
                # video/ and flexible/ have no GPU test group at all: no
                # group below runs a single test from either package, so a full
                # run costs every GPU minute while covering none of the change.
                # Their tests (video/tests/, flexible/tests/) run in the
                # ubuntu/macos/windows CI, which is where they are covered.
                xinference/model/video/*|xinference/model/flexible/*) ;;
                # Remaining model families with no dedicated group, plus
                # top-level files directly under xinference/model/ (batch.py,
                # cache_manager.py, core.py, custom.py, utils.py) which are
                # shared by every family; play it safe and run every group.
                xinference/model/*) full=true; break ;;
                # The rest of the package (deploy/, docs, examples, other
                # top-level modules) is not exercised by any GPU test.
                .github/workflows/python.yaml|.github/requirements/model-ci-constraints.txt|pyproject.toml|build_backend.py|build_web.py) full=true; break ;;
              esac
            done < changed-files.txt
          fi

          groups=""
          if [ "$full" = true ]; then
            groups="llm embedding image audio"
          else
            [ "$llm" = true ] && groups="$groups llm"
            [ "$embedding" = true ] && groups="$groups embedding"
            [ "$image" = true ] && groups="$groups image"
            [ "$audio" = true ] && groups="$groups audio"
            groups="${groups# }"
          fi

          gpu=true
          [ -z "$groups" ] && gpu=false
          echo "gpu=$gpu" >> "$GITHUB_OUTPUT"
          echo "groups=$groups" >> "$GITHUB_OUTPUT"
          echo "GPU CI required: $gpu (groups: ${groups:-none})"

  build_test_job:
    runs-on: ${{ matrix.os }}
    needs: lint
    env:
      CONDA_ENV: test
    defaults:
      run:
        shell: bash -l {0}
    strategy:
      fail-fast: false
      matrix:
        os: [ "ubuntu-latest", "macos-latest", "windows-latest" ]
        python-version: [ "3.10", "3.11", "3.12", "3.13", "3.14" ]
        module: [ "xinference" ]
        exclude:
          - { os: macos-latest, python-version: "3.11" }
          - { os: macos-latest, python-version: "3.12" }
          - { os: macos-latest, python-version: "3.13" }
          - { os: windows-latest, python-version: "3.11" }
          - { os: windows-latest, python-version: "3.12" }
          - { os: windows-latest, python-version: "3.13" }
        include:
          - { os: macos-latest, module: metal, python-version: "3.10" }

    steps:
      - name: Check out code
        uses: actions/checkout@v7
        with:
          fetch-depth: 0
          submodules: recursive

      - name: Set up conda ${{ matrix.python-version }}
        uses: conda-incubator/setup-miniconda@v3
        with:
          python-version: ${{ matrix.python-version }}
          activate-environment: ${{ env.CONDA_ENV }}

      # Important for Python 3.12 and newer.
      - name: Update pip and setuptools
        if: ${{ matrix.python-version == '3.12' || matrix.python-version == '3.13' || matrix.python-version == '3.14' }}
        run: |
          python -m pip install -U pip "setuptools<82"

      - name: Install PyTorch for Python 3.13 and newer
        if: ${{ matrix.python-version == '3.13' || matrix.python-version == '3.14' }}
        run: |
          python -m pip install torch torchvision torchaudio

      - name: Install NumPy 1.x where required
        if: ${{ startsWith(matrix.os, 'windows') && matrix.python-version == '3.10' }}
        run: |
          python -m pip install "numpy<2"


      - name: Install dependencies
        env:
          MODULE: ${{ matrix.module }}
          OS: ${{ matrix.os }}
          # Tests never touch the web UI; skip the npm build that the build
          # backend would otherwise run on editable installs.
          NO_WEB_UI: 1
        run: |
          if [ "$OS" == "ubuntu-latest" ]; then
            sudo rm -rf /usr/share/dotnet
            sudo rm -rf /opt/ghc
            sudo rm -rf "/usr/local/share/boost"
            sudo rm -rf "$AGENT_TOOLSDIRECTORY"
          fi
          if [ "$MODULE" != "metal" ]; then
            export PIP_CONSTRAINT="$PWD/.github/requirements/model-ci-constraints.txt"
          fi
          pip install -e ".[dev]"
          pip install "xllamacpp>=0.2.0"
          if [ "$MODULE" == "metal" ]; then
            conda install -c conda-forge "ffmpeg<7"
            pip install "mlx>=0.22.0"
            pip install mlx-lm
            pip install "mlx-vlm>=0.3.4"
            pip install mlx-whisper
            # F5-TTS owns a model-specific MLX pin in its virtualenv recipe.
            # Keep it out of the host env so the integration test exercises
            # that isolated dependency set instead of inheriting this one.
            pip install qwen-vl-utils!=0.0.9
            pip install tomli
          else
            pip install transformers
            pip install attrdict
            pip install "timm>=0.9.16"
            if [ "${{ matrix.python-version }}" != "3.13" ] && [ "${{ matrix.python-version }}" != "3.14" ]; then
              pip install torch torchvision
            fi
            pip install accelerate
            pip install sentencepiece
            pip install transformers_stream_generator
            pip install bitsandbytes
            pip install "sentence-transformers>=5.1.1,<6"
            pip install modelscope
            pip install diffusers
            pip install protobuf
            pip install "FlagEmbedding==1.3.5"
            pip install "tenacity>=8.2.0,<8.4.0"
            pip install "jinja2==3.1.2"
            pip install jj-pytorchvideo
            pip install qwen-vl-utils!=0.0.9
            pip install datamodel_code_generator
            pip install jsonschema
            pip install -r .github/requirements/model-ci-constraints.txt
            python -c "from datasets import Dataset; from transformers import AutoModel; from peft import PeftModel; from sentence_transformers import SentenceTransformer, CrossEncoder; from FlagEmbedding import BGEM3FlagModel; from diffusers import DiffusionPipeline"
          fi
        working-directory: .

      - name: Clean up disk
        if: |
          (startsWith(matrix.os, 'ubuntu'))
        run: |
          sudo rm -rf /usr/share/dotnet
          sudo rm -rf /usr/local/lib/android
          sudo rm -rf /opt/ghc
          sudo apt-get clean
          sudo rm -rf /var/lib/apt/lists/*
          df -h

      - name: Fix SSL on Windows
        if: startsWith(matrix.os, 'windows')
        shell: bash
        run: |
          echo "activate conda env"

          source $CONDA/etc/profile.d/conda.sh || true
          conda activate $CONDA_ENV || true

          python -V
          which python

          echo "before: $SSL_CERT_FILE"

          python -m pip install --quiet certifi || true

          SSL_CERT_FILE=$(python -c "import certifi,os;print(os.path.normpath(certifi.where()))")

          export SSL_CERT_FILE
          export REQUESTS_CA_BUNDLE=$SSL_CERT_FILE
          export CURL_CA_BUNDLE=$SSL_CERT_FILE

          echo "after: $SSL_CERT_FILE"
          echo "SSL_CERT_FILE=$(python -c 'import certifi;print(certifi.where())')" >> $GITHUB_ENV

      - name: Test with pytest
        env:
          MODULE: ${{ matrix.module }}
          PYTORCH_MPS_HIGH_WATERMARK_RATIO: 1.0
          PYTORCH_MPS_LOW_WATERMARK_RATIO: 0.2
          XFORMERS_FORCE_DISABLE_TRITON: 1
          TORCH_DISABLE_FLASH_ATTENTION: 1
        run: |
          set -e

          run_pytest() {
            pytest --timeout=3000 \
              -W ignore::PendingDeprecationWarning \
              --cov-config=pyproject.toml \
              --cov-report=xml \
              --cov=xinference \
              "$@"
          }

          if [ "$MODULE" == "metal" ]; then
            status=0
            mkdir -p test-results
            python -m pip freeze > test-results/metal-host-packages.txt
            for test_file in \
              xinference/model/llm/mlx/tests/test_mlx.py \
              xinference/model/llm/mlx/tests/test_xavier.py \
              xinference/model/audio/tests/test_whisper_mlx.py \
              xinference/model/audio/tests/test_f5tts_mlx.py \
              xinference/model/llm/mlx/tests/test_distributed_model.py; do
              group=$(basename "$test_file" .py)
              echo "::group::Metal tests: $group"
              # Keep separate processes for model environments, and run every
              # group even if an earlier one fails. Preserve the failure status.
              run_pytest --junitxml="test-results/$group.xml" "$test_file" || status=1
              echo "::endgroup::"
            done
            [ "$status" -eq 0 ]
          else
            run_pytest \
              -vv \
              --ignore xinference/core/tests/test_continuous_batching.py \
              --ignore xinference/model/image/tests/test_stable_diffusion.py \
              --ignore xinference/model/image/tests/test_got_ocr2.py \
              --ignore xinference/model/audio/tests \
              --ignore xinference/model/embedding/tests/test_integrated_embedding.py \
              --ignore xinference/model/llm/transformers/tests/test_tensorizer.py \
              --ignore xinference/model/llm/tests/test_llm_model.py \
              --ignore xinference/model/llm/vllm \
              --ignore xinference/model/llm/sglang \
              --ignore xinference/client/tests/test_client.py \
              --ignore xinference/client/tests/test_async_client.py \
              --ignore xinference/model/llm/mlx \
              xinference
            run_pytest \
              benchmark/tests/test_benchmark_pd.py \
              benchmark/tests/test_benchmark_sglang_pd.py \
              benchmark/tests/test_analyze_pd_profile.py \
              xinference/model/llm/sglang/tests/test_xavier.py \
              xinference/model/llm/sglang/tests/test_xavier_pd.py \
              xinference/model/llm/sglang/tests/test_pd.py \
              xinference/model/llm/sglang/tests/test_xavier_gpu_lifecycle.py \
              xinference/model/llm/sglang/tests/test_gc_lifecycle.py \
              xinference/model/llm/sglang/tests/test_runtime.py \
              xinference/model/llm/vllm/tests/test_distributed_executor_v1.py \
              xinference/model/llm/vllm/tests/test_executor_routing.py \
              xinference/model/llm/vllm/tests/test_speculative_config.py \
              xinference/model/llm/vllm/xavier/test/test_profiling.py \
              xinference/model/llm/vllm/xavier/test/test_actor_loop.py \
              xinference/model/llm/vllm/tests/test_pd.py \
              xinference/model/llm/vllm/xavier/test/test_block_tracker_reuse.py \
              xinference/model/llm/vllm/xavier/test/test_remote_kvcache_manager.py \
              xinference/model/llm/vllm/xavier/test/test_v1_engine.py \
              xinference/model/llm/vllm/xavier/test/test_gpu_transfer.py \
              xinference/model/llm/vllm/xavier/test/test_kv_schema.py \
              xinference/model/llm/vllm/xavier/test/test_cross_engine.py \
              xinference/model/llm/vllm/xavier/test/test_direct_handoff.py \
              xinference/model/llm/vllm/xavier/test/test_direct_history.py \
              xinference/model/llm/vllm/xavier/test/test_tiered_snapshot.py \
              xinference/model/llm/vllm/xavier/test/test_request_transfer.py \
              xinference/model/llm/vllm/xavier/test/test_transfer_blocks.py \
              xinference/model/llm/vllm/xavier/test/test_review_regressions.py
          fi
        working-directory: .

      - name: Upload Metal test results
        if: ${{ always() && matrix.module == 'metal' }}
        uses: actions/upload-artifact@v4
        with:
          name: metal-test-results-${{ matrix.python-version }}
          path: test-results/
          if-no-files-found: ignore

  gpu_test_job:
    name: build_test_job (gpu-t4, gpu, 3.12)
    runs-on: gpu-t4
    needs: [lint, changes]
    # Runners are billed per-minute: skip when no GPU-relevant path changed,
    # and only run for PRs and pushes to main (a push to a PR branch already
    # triggers the pull_request run).
    if: ${{ needs.changes.outputs.gpu == 'true' && (github.event_name == 'pull_request' || github.ref == 'refs/heads/main') }}
    timeout-minutes: 90
    env:
      CONDA_ENV: test
      # GPU tests never touch the web UI; skipping the npm build saves
      # minutes. The API server only warns when the UI bundle is absent.
      NO_WEB_UI: 1
    defaults:
      run:
        shell: bash -l {0}
    steps:
      - name: Check out code
        uses: actions/checkout@v7
        with:
          fetch-depth: 0
          submodules: recursive

      - name: Set up conda 3.12
        uses: conda-incubator/setup-miniconda@v3
        with:
          python-version: "3.12"
          activate-environment: ${{ env.CONDA_ENV }}

      - name: Restore pip package cache
        uses: actions/cache@v4
        with:
          path: ~/.cache/pip
          key: gpu-pip-${{ hashFiles('.github/workflows/python.yaml', '.github/requirements/model-ci-constraints.txt', 'pyproject.toml') }}
          restore-keys: gpu-pip-

      # funasr downloads its auxiliary models (fsmn-vad, ct-punc, cam++) from
      # ModelScope, whose CDN is very slow from GitHub's US runners (~15min
      # per run). They are only ~1.5GB in total, so cache them. Main models
      # come from HuggingFace (fast) and are intentionally not cached — all
      # of them together would blow the 10GB per-repo cache limit.
      # NOTE: xinference redirects MODELSCOPE_CACHE to ~/.xinference/modelscope
      # (see get_xinference_home in xinference/constants.py), so that is the
      # directory to cache — ~/.cache/modelscope is never created.
      - name: Restore ModelScope model cache
        uses: actions/cache@v4
        with:
          path: ~/.xinference/modelscope
          key: gpu-modelscope-${{ hashFiles('xinference/model/audio/models/*.json') }}
          restore-keys: gpu-modelscope-

      - name: Update pip and setuptools
        run: |
          python -m pip install -U pip "setuptools<82"

      - name: Install dependencies
        run: |
          export PIP_CONSTRAINT="$PWD/.github/requirements/model-ci-constraints.txt"
          pip install -U -e ".[audio,dev]"
          pip install -U "openai>1"
          pip install -U modelscope
          pip install -U gguf
          pip install -U uv
          pip install -U sse_starlette
          pip install -U xoscar
          pip install -U "python-jose[cryptography]"
          pip install -U "passlib[bcrypt]"
          pip install -U "aioprometheus[starlette]"
          pip install -U "pynvml"
          pip install transformers
          pip install "funasr==1.2.7"
          conda install -y -c conda-forge "pynini==2.1.5" "ffmpeg<7"
          pip install -U "nemo_text_processing<1.1.0"
          pip install -U omegaconf~=2.3.0
          pip install -U "WeTextProcessing<1.0.4"
          pip install -U librosa
          pip install -U xxhash
          pip install -U "ChatTTS==0.2.4"
          pip install -U HyperPyYAML
          pip uninstall -y matcha-tts
          pip install -U 'onnxruntime-gpu==1.22.0; sys_platform == "linux"'
          pip install -U openai-whisper
          pip install -U "torch==2.7.0" "torchaudio==2.7.0" "torchvision==0.22.0"
          pip install -U "loguru"
          pip install -U "natsort"
          pip install -U "loralib"
          pip install -U "ormsgpack"
          pip uninstall -y opencc
          pip uninstall -y "faster_whisper"
          pip install -U accelerate
          pip install -U verovio
          pip install -U cachetools
          pip install -U silero-vad
          pip install -U pydantic
          pip install -U "diffusers==0.35.1"
          pip install -U onnx
          pip install -U onnxconverter_common
          pip install -U torchdiffeq
          pip install -U "x_transformers>=1.31.14"
          pip install -U pypinyin
          pip install -U tomli
          pip install -U vocos
          pip install -U jieba
          pip install -U soundfile
          pip install tensorizer
          pip install -U "sentence-transformers==5.1.1"
          pip install -U "FlagEmbedding==1.3.5"
          pip install -U "peft>=0.18.0"
          pip install "xllamacpp>=0.2.0" --index-url https://xorbitsai.github.io/xllamacpp/whl/cu124 --extra-index-url https://pypi.org/simple
          pip install hf_transfer
          # espeak-ng (TTS phonemization) is not packaged on conda-forge;
          # install it from apt.
          sudo apt-get update -qq
          sudo apt-get install -y -qq espeak-ng
          # Resolve the shared stack together, including transitive datasets.
          pip install "numpy>=2" -r .github/requirements/model-ci-constraints.txt
          python -c "from datasets import Dataset; from transformers import AutoModel; from peft import PeftModel; from sentence_transformers import SentenceTransformer, CrossEncoder; from FlagEmbedding import BGEM3FlagModel; from diffusers import DiffusionPipeline"
        working-directory: .

      - name: Test with pytest
        env:
          XFORMERS_FORCE_DISABLE_TRITON: 1
          TORCH_DISABLE_FLASH_ATTENTION: 1
          # Speed up model downloads from the HuggingFace hub; runners are
          # ephemeral so every run downloads the main models from scratch.
          HF_HUB_ENABLE_HF_TRANSFER: 1
          # Pin the model source instead of relying on locale auto-detection;
          # ModelScope is very slow from US-hosted runners.
          XINFERENCE_MODEL_SRC: huggingface
        run: |
          # One pytest process per selected group: one cold start per family
          # instead of one per file, while an import error or crash in one
          # family cannot take down the others. No --cov: nothing in this
          # repo consumes the coverage report. A failing group does not stop
          # the remaining groups, so one run reports every failure.
          #
          # On a failure, retry only the failed tests once via --last-failed.
          # test_continuous_batching hits an intermittent KV-cache batching
          # bug (~1 in 4 runs, tracked separately) that is a same-process
          # race, not environment flakiness; each pytest invocation still
          # gets fresh model/server subprocesses from the `setup` fixture, so
          # a rerun is not just re-hitting stale state.
          status=0
          run_pytest() {
            local rc=0
            pytest --timeout=3000 \
              -W ignore::PendingDeprecationWarning \
              "$@" || rc=$?
            if [ "$rc" -ne 0 ]; then
              echo "::warning::pytest exited with code $rc for group '$group'; retrying failed tests once"
              rc=0
              pytest --timeout=3000 \
                -W ignore::PendingDeprecationWarning \
                --last-failed \
                "$@" || rc=$?
            fi
            echo "group '$group' pytest exit code: $rc"
            if [ "$rc" -ne 0 ]; then
              echo "::warning::pytest exited with code $rc for group '$group' after retry"
              status=1
            fi
          }

          for group in ${{ needs.changes.outputs.groups }}; do
            echo "::group::GPU tests: $group"
            case "$group" in
              llm)
                run_pytest \
                  xinference/core/tests/test_continuous_batching.py \
                  xinference/model/llm/tests/test_llm_model.py \
                  xinference/model/llm/transformers/tests/test_tensorizer.py
                ;;
              embedding)
                # test_integrated_embedding must run first: the other three
                # tests launch models in per-model virtualenvs, which leak
                # their site-packages into the process environment, and models
                # launched afterwards in the same pytest process import
                # packages from the wrong env (seen with both the
                # Qwen3-VL-Embedding and Qwen3-VL-Reranker virtualenvs).
                run_pytest \
                  xinference/model/embedding/tests/test_integrated_embedding.py \
                  xinference/model/embedding/tests/test_qwen3_vl_engine_params.py \
                  xinference/model/embedding/vllm/tests/test_vllm_embedding.py \
                  xinference/model/rerank/tests/test_qwen3_vl_reranker_virtualenv.py
                ;;
              image)
                # test_got_ocr2 is intentionally omitted: it self-skips on
                # transformers >= 4.48.0 and this job pins 4.53.2, so it never
                # runs a single case here.
                run_pytest \
                  xinference/model/image/tests/test_stable_diffusion.py
                ;;
              audio)
                # test_cosyvoice is intentionally omitted: all of its cases
                # are @pytest.mark.skip'd (diffusers on this image is
                # incompatible), so it never runs anything here.
                run_pytest \
                  xinference/model/audio/tests/test_whisper.py \
                  xinference/model/audio/tests/test_funasr.py \
                  xinference/model/audio/tests/test_chattts.py \
                  xinference/model/audio/tests/test_f5tts.py \
                  xinference/model/audio/tests/test_melotts.py \
                  xinference/model/audio/tests/test_kokoro.py \
                  xinference/model/audio/tests/test_fish_speech.py \
                  xinference/model/audio/tests/test_megatts.py
                ;;
            esac
            echo "::endgroup::"
          done
          echo "aggregate pytest status: $status"
          # Finish with a plain test instead of an explicit `exit`: two runs
          # with every pytest group green still ended in exit code 1 through
          # the login shell's explicit-exit path on this runner image.
          [ "$status" -eq 0 ]
        working-directory: .