# GUI-G2-3B + CCF inference requirements. # Pinned to versions known to work together with the published model # weights. Newer versions probably work too but haven't been tested # end-to-end. # Core inference stack torch>=2.4.0 transformers==4.49.0 qwen-vl-utils==0.0.14 Pillow>=10.0.0 # Optional but strongly recommended for >2x speedup on big images # (skip if you're on a GPU without flash-attn support; predict.py has # --attn-impl sdpa as a fallback). # Install separately: pip install flash-attn==2.7.0.post2 --no-build-isolation