# gguf-py is used from a llama.cpp checkout at ../llama.cpp (or a local gguf-py copy), not from pip
# triton ships with the torch wheel on Linux/CUDA
torch>=2.7
numpy
pyyaml