transformers>=4.46.0
optimum>=1.23.0        # makes transformers’ quantizer happy
auto-gptq>=0.7.0.post0 # wheels available for CUDA 11.8/12.x
bitsandbytes>=0.43.0   # still useful for 4-bit mat-muls
accelerate>=0.29.0
sentencepiece       # tokenizer dependency
gradio>=4.28
# requirements.txt  – CPU-only
transformers>=4.46.0       # model & pipeline
torch>=2.2.0               # CPU wheel auto-selected
sentencepiece>=0.1.99      # DistilBART tokenizer
gradio>=4.28.0             # UI
optimum[onnxruntime]>=1.18.0