Probably low on the priority list; but it might be helpful to make nvidia-cutlass-dsl optional for Windows and adjust the library to auto-detect availability of CuTeDSL (similar to what is being done for the CUDA backend) so that Windows can still use the Triton backend (ie. via PyTorch + triton-windows).
For an experiment I was able to crudely delete the cute backend and get it working. NVIDIA do not provide any binaries at the moment for CUTLASS on Windows.
Probably low on the priority list; but it might be helpful to make
nvidia-cutlass-dsloptional for Windows and adjust the library to auto-detect availability of CuTeDSL (similar to what is being done for the CUDA backend) so that Windows can still use the Triton backend (ie. via PyTorch + triton-windows).For an experiment I was able to crudely delete the cute backend and get it working. NVIDIA do not provide any binaries at the moment for CUTLASS on Windows.