# Versions the release was validated with. infer_cpu.py relies on DynamicCache.batch_repeat_interleave (transformers 4.x).
torch==2.8.0
transformers==4.57.1
pyarrow>=14
numpy>=1.26
huggingface_hub>=0.26
# Optional, GPU only: flash-attn==2.8.3 (infer_gpu.py falls back to SDPA without it)