migrate to uv + pyproject.toml

- root pyproject.toml replaces requirements.txt and requirements-worker.txt
  (worker = root + ffmpeg-python which root already had); test deps moved
  to [dependency-groups] dev
- core/gpu/pyproject.toml replaces core/gpu/requirements.txt; uses
  [tool.uv.sources] to pin torch/torchvision and paddlepaddle-gpu to their
  CUDA index URLs, replacing the manual reinstall dance from old comments
- Dockerfiles use uv sync --frozen against uv.lock for reproducible builds;
  PATH includes /app/.venv/bin so k8s manifests' bare uvicorn/celery
  commands resolve without wrapping in uv run
- core/gpu/run.sh local mode now does uv sync + uv run python server.py;
  errors out cleanly if uv is missing

core/gpu/Dockerfile

@@ -8,8 +8,11 @@ RUN apt-get update && apt-get install -y \
 
 WORKDIR /app
 
-COPY requirements.txt .
-RUN uv pip install --system --no-cache -r requirements.txt
+ENV PATH="/app/.venv/bin:$PATH"
+
+# uv.lock is generated on first build (dev box can't reach paddlepaddle's index)
+COPY pyproject.toml ./
+RUN uv sync --no-install-project --no-dev
 
 COPY . .
 

core/gpu/pyproject.toml

@@ -0,0 +1,42 @@
+[project]
+name = "mpr-gpu"
+version = "0.1.0"
+description = "MPR remote inference server (GPU)"
+requires-python = ">=3.11"
+dependencies = [
+    "fastapi>=0.109.0",
+    "uvicorn[standard]>=0.27.0",
+    "rapidfuzz>=3.0.0",
+    "Pillow>=10.0.0",
+    "redis>=5.0.0",
+    "ultralytics>=8.0.0",
+    "paddleocr>=3.0.0",
+    "paddlepaddle-gpu==3.0.0",
+    "transformers>=4.40.0,<5",
+    "accelerate>=0.27.0",
+    "torch",
+    "torchvision",
+    "opencv-python-headless>=4.8.0",
+]
+
+# RTX 3080 / CUDA toolkit 12.8 — cu126 wheels are forward-compatible
+# (no cu128 wheels yet on either index). Mixing PyPI torch with CUDA 12.8
+# causes NCCL symbol errors, so the explicit index pins prevent uv from
+# pulling torch transitively from PyPI via ultralytics.
+[tool.uv.sources]
+torch = { index = "pytorch-cu126" }
+torchvision = { index = "pytorch-cu126" }
+paddlepaddle-gpu = { index = "paddle-cu126" }
+
+[[tool.uv.index]]
+name = "pytorch-cu126"
+url = "https://download.pytorch.org/whl/cu126"
+explicit = true
+
+[[tool.uv.index]]
+name = "paddle-cu126"
+url = "https://www.paddlepaddle.org.cn/packages/stable/cu126/"
+explicit = true
+
+[tool.uv]
+package = false

core/gpu/requirements.txt

@@ -1,31 +0,0 @@
-fastapi>=0.109.0
-uvicorn[standard]>=0.27.0
-rapidfuzz>=3.0.0
-Pillow>=10.0.0
-redis>=5.0.0
-
-# --- GPU-specific installs (mcrn: RTX 3080, CUDA toolkit 12.8) ---
-#
-# torch: must be installed from the PyTorch index, NOT from PyPI.
-# cu126 is the closest build to CUDA 12.8 (no cu128 wheel yet; cu126 is forward-compatible).
-# Install with:
-#   uv pip install --reinstall torch torchvision --index-url https://download.pytorch.org/whl/cu126
-#
-# ultralytics pulls torch as a dependency — reinstall torch after ultralytics to ensure
-# the correct CUDA build. Mixing the PyPI torch with CUDA 12.8 causes NCCL symbol errors.
-ultralytics>=8.0.0
-
-# paddlepaddle-gpu: NOT available on PyPI. Install from PaddlePaddle's package index.
-# cu126 build works on CUDA 12.8.
-# Install with:
-#   uv pip install paddlepaddle-gpu==3.0.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/
-paddleocr>=3.0.0
-
-# VLM (moondream2) — uses torch (already installed above)
-# Pinned <5: transformers 5.x broke moondream2's custom model code
-# (all_tied_weights_keys API change). Also needs accelerate for device_map.
-transformers>=4.40.0,<5
-accelerate>=0.27.0
-
-# Preprocessing (phase 12)
-opencv-python-headless>=4.8.0

core/gpu/run.sh

@@ -2,7 +2,7 @@
 # Run the inference server
 #
 # Usage:
-#   ./run.sh                  # Local (pip install -r requirements.txt first)
+#   ./run.sh                  # Local: uv sync + run server.py (auto-installs/activates .venv)
 #   ./run.sh docker           # Docker (CPU)
 #   ./run.sh docker-gpu       # Docker with GPU
 #   ./run.sh stop             # Stop Docker container
@@ -26,7 +26,12 @@ fi
 
 case "${1:-local}" in
     local)
-        python server.py
+        if ! command -v uv >/dev/null 2>&1; then
+            echo "uv not found. Install: curl -LsSf https://astral.sh/uv/install.sh | sh"
+            exit 1
+        fi
+        uv sync
+        uv run python server.py
         ;;
     docker)
         docker build -t mpr-inference .