nvidia-nvshmem-cu12
cuda-python<=12.9,>=12.0
cuda.core>=0.5.0
cuda.pathfinder>=1.2.3
numpy
Cython>=0.29.24

[cute]
nvidia-cutlass-dsl==4.4.2

[dev]
pytest>=6.0
pytest-mpi
black
isort

[numba]
numba-cuda[cu12]>=0.28.0
