gradio>=3.30
transformers>=4.30
torch
huggingface-hub
accelerate
# Optional (only if you use adapters/PEFT): peft>=0.3.0
# Optional for 8-bit inference (may fail in some Spaces): bitsandbytes