CPU-only PyTorch:
CPU-only PyTorch: two-step Dockerfile install eliminates 6.8GB GPU waste. easyocr/docling-ibm-models pull torch which defaults to GPU version (nvidia 4.3GB + triton 639MB + cuda 106MB). Fix: install torch first from CPU index, then requirements. RUN uv pip install —system —no-cache torch torchvision —index-url https://download.pytorch.org/whl/cpu → then RUN uv pip install —system —no-cache -r requirements.txt. Applied to SharePoint MCP: 8.62GB → 2.05GB.