Use CPU-only torch for Docker builds to reduce image size

[adamtheturtle]

Summary

Configure uv to install CPU-only torch from PyTorch's dedicated index via [tool.uv] configuration in pyproject.toml. Update the Dockerfile to use uv sync (which respects tool.uv sources) instead of uv pip install, and upgrade uv from 0.1.44 to 0.5.0 for source configuration support. This reduces the Docker image size by ~3GB by removing unnecessary CUDA dependencies.

Changes

• pyproject.toml: Add [tool.uv] with index configuration for CPU-only torch from https://download.pytorch.org/whl/cpu
• Dockerfile:
  • Upgrade uv from 0.1.44 to 0.5.0
  • Switch from uv pip install --editable . to uv sync --no-cache
  • Remove manual venv creation (handled by uv sync)
  • Add UV_PROJECT_ENVIRONMENT to direct uv to the container's venv location

Impact

• Docker image is ~3GB smaller (no CUDA dependencies)
• Build time is faster (smaller downloads)
• Behavior unchanged: torch CPU variant works fine for scoring/evaluation in the Flask server

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes dependency resolution and Docker build/install flow for torch/torchvision, which could cause build failures or runtime incompatibilities if the CPU wheels or uv config behave differently across environments.

Overview
Configures uv to install torch and torchvision from PyTorch’s CPU-only wheel index via new [tool.uv] entries in pyproject.toml, and adds torchvision as an explicit dependency (with a deptry ignore) so it can be routed to that index.

Updates the Flask server Docker build to set UV_PROJECT_ENVIRONMENT and use uv sync --no-cache (instead of uv pip install --editable .) when installing into the container venv.

^{Written by Cursor Bugbot for commit 9f484ea. This will update automatically on new commits. Configure here.}