torch           # HF will pick a CUDA / CPU build automatically
transformers>=4.39
accelerate      # lets Transformers shard the model
sentencepiece   # <-- needed for the tokenizer
gradio==4.18