torch>=2.1.0
transformers>=4.40.0
datasets>=2.18.0
pytest>=7.0.0

# optional — enables Flash Attention 2 in GQAttention (requires CUDA + build tools)
# flash-attn>=2.8.3
