fix: PDF ingestion fails on Linux due to CUDA torchvision mismatch

The desktop CI workflow only pre-installed CPU-only torch but not
torchvision. When docling pulled in torchvision from PyPI, it got the
CUDA variant whose _C.so links against libtorch_cuda.so/libc10_cuda.so
which don't exist in CPU-only torch, causing every PDF page to fail.

- Add torch/torchvision as explicit deps pinned to PyTorch CPU index
  via [tool.uv.sources] in pyproject.toml (works for all builds)
- Add torch/torchvision to PACKAGES_WITH_BINARIES in PyInstaller spec
  so their .so files land where LD_LIBRARY_PATH is set on Linux
- Remove redundant CPU PyTorch install step from CI workflow
- Switch Dockerfile from uv pip install to uv sync (uses lockfile)

Author: itstauq

.github/workflows/desktop-publish.yml

@@ -58,10 +58,6 @@ jobs:
       - name: Setup uv
         uses: astral-sh/setup-uv@v4
 
-      - name: Install CPU-only PyTorch (Linux)
-        if: matrix.platform == 'ubuntu-22.04'
-        run: uv pip install torch --index-url https://download.pytorch.org/whl/cpu --system
-
       - name: Install backend dependencies
         run: uv pip install -e "backend/.[build,libs]" --system
 

Dockerfile

@@ -29,21 +29,16 @@ RUN apk update && apk add --no-cache \
 
 WORKDIR /app
 
-COPY backend/pyproject.toml backend/README.md ./backend/
+# Copy lockfile, pyproject.toml, and minimal source for uv sync
+COPY backend/pyproject.toml backend/uv.lock backend/README.md ./backend/
 COPY backend/syft_space/__init__.py ./backend/syft_space/
 
 ENV UV_COMPILE_BYTECODE=1 UV_LINK_MODE=copy
 
-# 1. Install base deps (fast, frequently cached)
-RUN uv venv /app/.venv --python python${PYTHON_VERSION} && \
-    uv pip install --python /app/.venv/bin/python -e "./backend"
-
-# 2. Install CPU-only PyTorch (avoids 4GB+ CUDA packages)
-RUN uv pip install --python /app/.venv/bin/python \
-    torch torchvision --index-url https://download.pytorch.org/whl/cpu
-
-# 3. Install heavy optional deps (docling will use existing torch)
-RUN uv pip install --python /app/.venv/bin/python -e "./backend[libs]"
+# Install all deps (including libs) from lockfile.
+# CPU-only torch/torchvision is enforced by [tool.uv.sources] in pyproject.toml.
+ENV UV_PROJECT_ENVIRONMENT=/app/.venv
+RUN cd backend && uv sync --extra libs --no-install-project
 
 # ============================================================================
 # Stage 2: Production
@@ -75,6 +70,9 @@ COPY --from=backend-builder /app/.venv /app/.venv
 # Copy backend source
 COPY backend/ ./backend/
 
+# Install the project itself (source-only, deps already in venv)
+RUN /app/.venv/bin/pip install --no-deps -e ./backend
+
 # Copy pre-built frontend (must run: cd frontend && bun run build)
 COPY frontend/dist ./frontend/dist
 

backend/pyproject.toml

@@ -64,6 +64,8 @@ libs = [
     "docling>=2.75.0",
     "docling-core>=2.65.2",
     "chromadb>=0.5.0",
+    "torch>=2.10.0",
+    "torchvision>=0.25.0",
 ]
 
 # All development dependencies (includes test, lint, pre-commit)
@@ -125,5 +127,12 @@ skips = [
 where = ["."]  # Find packages in current directory (backend/)
 include = ["syft_space*"]  # Include syft_space and sub-packages
 
+[[tool.uv.index]]
+name = "pytorch-cpu"
+url = "https://download.pytorch.org/whl/cpu"
+explicit = true
+
 [tool.uv.sources]
 syft-accounting-sdk = { git = "https://github.com/OpenMined/accounting-sdk.git", rev = "main" }
+torch = { index = "pytorch-cpu" }
+torchvision = { index = "pytorch-cpu" }

backend/syft-space-backend.spec

@@ -41,6 +41,8 @@ PACKAGES_WITH_BINARIES = [
     'chromadb',
     'chromadb_rust_bindings',
     'docling_parse',
+    'torch',
+    'torchvision',
 ]
 
 a = Analysis(