Fix CI for the release branch

The pre-commit CI job on this branch never actually ran (it always died at
`pip install .`, which cannot resolve the not-yet-published
pixano-inference-client), so lint/type/format debt and a few real CI-config
problems went undetected. Fix them all:

- pre-commit (lint_and_format): drop `pip install .` — the hooks run in
  isolated environments and do not need the package installed. Exclude the
  example deployment scripts in deploy/ from ruff (force-exclude) and mypy,
  add types-requests to the mypy hook, and .prettierignore the hand-managed
  CLAUDE.md. Reformat the Phase 5/6 docs + docker-compose with Prettier, and
  the docker/models.py import order with ruff.
- OpenAPI check: move it into its own job that runs against the locked
  environment (uv sync), so it is deterministic — a fresh pip resolve pulls a
  newer FastAPI that emits a slightly different schema than the committed one.
- Ray Serve integration + backend: install torch AND torchvision from the same
  CPU index so their compiled ops stay ABI-coherent (open_clip pulled a PyPI
  torchvision against the CPU-index torch -> "operator torchvision::nms does
  not exist").
- Docker: `uv pip install .` resolves the client via [tool.uv.sources] but
  installs it editable (a path into /build), which the runtime stage's
  venv copy loses; reinstall it as a real package so it survives the copy.
- docs / publish: install the client editable before `.[docs]` (hard core dep,
  not yet on PyPI).

Author: jrabary

.github/workflows/docs.yml

@@ -49,7 +49,9 @@ jobs:
             mkdocs-material-
 
       - name: Install dependencies
-        run: pip install .[docs]
+        run: |
+          pip install -e packages/pixano-inference-client  # hard core dep, not yet on PyPI
+          pip install .[docs]
 
       - name: Publish documentation website
         run: mike deploy --push --update-aliases latest

.github/workflows/lint_and_format.yml

@@ -48,12 +48,12 @@ jobs:
           python-version: ${{ matrix.python-version }}
           cache: "pip"
 
-      # Install PyTorch and TensorFlow CPU versions manually to prevent installing CUDA
-      # Install SAM and MobileSAM manually as they cannot be included in PyPI
+      # The pre-commit hooks (ruff, mypy via mirrors, license check) run in isolated
+      # environments and do not need the project installed — so we skip `pip install .`
+      # (which could not resolve the not-yet-published pixano-inference-client anyway).
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
-          python -m pip install .
           pip install pre-commit
 
       - name: Run pre-commit

.github/workflows/publish.yml

@@ -111,7 +111,9 @@ jobs:
             mkdocs-material-
 
       - name: Install dependencies
-        run: pip install .[docs]
+        run: |
+          pip install -e packages/pixano-inference-client  # hard core dep, not yet on PyPI
+          pip install .[docs]
 
       - name: Publish documentation website
         run: mike deploy --push --update-aliases ${{ env.VERSION }} stable

.github/workflows/test_back.yml

@@ -54,7 +54,10 @@ jobs:
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
-          python -m pip install torch --index-url https://download.pytorch.org/whl/cpu
+          # Install torch AND torchvision from the same CPU index so their compiled ops stay
+          # ABI-coherent (open_clip / transformers pull torchvision; a PyPI torchvision against
+          # a CPU-index torch breaks with "operator torchvision::nms does not exist").
+          python -m pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu
           # The client is a hard core dep not yet on PyPI, so install it editable BEFORE `pip install .`.
           python -m pip install -e packages/pixano-inference-client
           python -m pip install .[test,transformers]
@@ -89,7 +92,8 @@ jobs:
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
-          python -m pip install torch --index-url https://download.pytorch.org/whl/cpu
+          # torch + torchvision from the same index (open_clip needs matching torchvision ops).
+          python -m pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu
           python -m pip install -e packages/pixano-inference-client  # hard core dep, not yet on PyPI
           python -m pip install .[test]
           python -m pip install -e examples/numpy_detector       # example entry-point plugin
@@ -168,9 +172,28 @@ jobs:
           python -m pip install -e examples/numpy_detector
           pytest tests/test_plugins.py -v
 
+  openapi:
+    name: OpenAPI schema in sync
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      # Generate the schema in the LOCKED environment (uv.lock) so the check is deterministic:
+      # a fresh `pip install` resolves newer FastAPI/pydantic that emit a subtly different schema
+      # than the committed docs/openapi.json (which is generated against the lockfile).
+      - name: Install uv
+        uses: astral-sh/setup-uv@v5
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
       - name: Check the committed OpenAPI schema is in sync
         run: |
-          python scripts/gen_openapi.py --check
+          uv sync --no-default-groups
+          uv run python scripts/gen_openapi.py --check
 
   client_standalone:
     name: Standalone client (no server deps)

.pre-commit-config.yaml

@@ -36,7 +36,11 @@ repos:
     rev: v1.13.0
     hooks:
       - id: mypy
-        exclude: ^(docs|tests)/
+        # deploy/ holds example deployment scripts (referencing optional plugin packages),
+        # not library code — linted like docs/tests are excluded.
+        exclude: ^(docs|tests|deploy)/
+        additional_dependencies:
+          - types-requests
   - repo: local
     hooks:
       - id: check_license_header

.prettierignore

@@ -0,0 +1,3 @@
+# AI-assistant project instructions; managed by hand, excluded from formatting (as in the
+# license-header check).
+CLAUDE.md

CHANGELOG.md

@@ -21,7 +21,7 @@ and observability.
 
 > **Upgrading:** the HTTP API and the Python client are intentionally breaking. Server operators
 > install `pixano-inference` (optionally with `[sam]`, `[clip]`, `[transformers]`); apps that
-> only *call* a server should install the new lightweight **`pixano-inference-client`** instead.
+> only _call_ a server should install the new lightweight **`pixano-inference-client`** instead.
 
 ### ⚠️ Breaking changes
 

Dockerfile

@@ -54,10 +54,12 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     # for both and their CUDA builds stay ABI-coherent (torchvision is required by
     # transformers and sam-2).
     if [ -n "${TORCH_INDEX_URL}" ]; then uv pip install torch torchvision --index-url "${TORCH_INDEX_URL}"; fi; \
-    # The lightweight client is a hard core dependency; install it from the bundled source first
-    # (it is not on PyPI at build time) so the core install below resolves it locally.
-    uv pip install ./packages/pixano-inference-client; \
+    # The lightweight client is a hard core dependency not on PyPI at build time; `uv pip install .`
+    # resolves it via [tool.uv.sources] but installs it EDITABLE (a path into /build), which breaks
+    # once the runtime stage copies only /opt/venv. So install core, then reinstall the client as a
+    # real (copied) package so it survives the multi-stage copy.
     if [ -n "${PIXANO_EXTRAS}" ]; then uv pip install ".[${PIXANO_EXTRAS}]"; else uv pip install .; fi; \
+    uv pip install --force-reinstall --no-deps ./packages/pixano-inference-client; \
     if [ "${INSTALL_SAM}" = "true" ]; then \
         uv pip install ./packages/pixano-inference-sam "sam-2 @ git+https://github.com/facebookresearch/sam2.git@${SAM2_REF}"; \
     fi; \

docker-compose.yml

@@ -31,7 +31,8 @@ services:
       # Optional: local images to serve by path (read-only). The host dir must exist and be
       # readable by uid 1000; also set PIXANO_INFERENCE_MEDIA_ROOTS above.
       # - ./media:/data/media:ro
-    command: ["--host", "0.0.0.0", "--port", "7463", "--config", "/config/models.py"]
+    command:
+      ["--host", "0.0.0.0", "--port", "7463", "--config", "/config/models.py"]
     # Reap Ray's child processes (Python is PID 1 and is not an init).
     init: true
     restart: unless-stopped

docker/models.py

@@ -10,9 +10,10 @@
 models you need. Weights are cached under /data/hf (the persistent volume).
 """
 
-from pixano_inference.configs import DeploymentConfig, ModelConfig
 from pixano_inference_sam import Sam2ImageParams
 
+from pixano_inference.configs import DeploymentConfig, ModelConfig
+
 
 models = [
     ModelConfig(