torch>=2.7
numpy
safetensors
jinja2
cuda-bindings
tqdm
tabulate

[:sys_platform != "win32"]
triton

[:sys_platform == "win32"]
triton-windows>=3.8.0.post29

[cu12]
nvidia-cuda-runtime-cu12
nvidia-cuda-cccl-cu12
nvidia-cuda-nvcc-cu12
nvidia-cuda-nvrtc-cu12

[cu13]
nvidia-cuda-runtime
nvidia-cuda-cccl
nvidia-cuda-nvcc
nvidia-cuda-nvrtc
