torch accelerate bitsandbytes datasets sentencepiece safetensors==0.3.0 gradio semantic-version==2.10.0 flash-attn triton colorama git+https://github.com/huggingface/transformers.git git+https://github.com/sterlind/GPTQ-for-LLaMa.git@lora_4bit git+https://github.com/sterlind/peft.git