# Tested with Python 3.12.9 on CUDA 12.x. # NOTE: flash-attn is intentionally NOT listed here, it needs --no-build-isolation. # Install it separately, see step 3 of QUICK_START.md. # --- Core model stack (versions are strict: MSA targets the transformers 4.51 API) --- torch==2.6.0 torchvision==0.21.0 transformers==4.51.3 accelerate==1.0.1 liger_kernel==0.5.10 # --- Hub / data IO --- # huggingface_hub must stay <1.0 to remain compatible with transformers 4.51.3. huggingface_hub==0.31.4 datasets==3.1.0 lmdb==1.6.2 tqdm==4.67.1 # --- Numerics --- numpy==1.26.4 pillow==11.2.1 packaging==25.0 # --- GPU monitoring: provides the `pynvml` module. # Do NOT replace with the deprecated `pynvml` package, its import path differs. nvidia-ml-py==12.575.51 # --- LLM-based answer scoring (scripts/calculate_llm_score.sh) --- openai==1.79.0