torch>=2.4.0
safetensors>=0.4.0
tokenizers>=0.20.0
huggingface_hub>=0.26.0
einops>=0.8.0
numpy>=1.26.0

# flash-linear-attention is auto-installed by inference.py on first run.