flask
flask-cors
torch>=2.0.0
transformers>=4.36.0
accelerate
safetensors
datasets
sqlite3
datetime
bitsandbytes  # Optional, only needed for 4-bit quantized models
💡 Notes:
If you are not using 4-bit quantization, you can remove bitsandbytes.

For performance on CUDA-enabled GPUs, you should still manually install the correct CUDA-compatible version of torch, e.g.:
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
Replace cu118 with your actual CUDA version (cu121, cu117, etc).