torch>=2.0
numpy
tqdm

[:platform_system == "Linux"]
triton>=3.6

[:platform_system == "Windows"]
triton-windows>=3.6

[bench]
scikit-learn<1.8,>=1.5
tqdm
matplotlib

[cutedsl]

[cutedsl:platform_system == "Linux"]
nvidia-cutlass-dsl<4.7,>=4.4

[dev]
pytest>=7
scikit-learn<1.8,>=1.5
tqdm
flashlib[cutedsl]
