fnmodel / requirements.txt
aeb56
Workaround flash-attn: create fake module with PyTorch fallback attention
b705945
# Core ML dependencies
torch>=2.1.0
transformers>=4.56.0
accelerate>=0.34.0
sentencepiece>=0.1.99
tiktoken>=0.8.0
einops>=0.7.0
triton>=3.0.0
# Flash Linear Attention (required by Kimi model)
git+https://github.com/sustcsonglin/flash-linear-attention.git@main
# Evaluation
lm-eval>=0.4.0
# UI
gradio==4.19.2
# Utils
safetensors>=0.4.0
protobuf>=3.20.0