Support CPU mCuSeqlensM in Grouped Gemm?

[zfan2356]

Hello Tri Dao! I sincerely apologize for taking up your time, but I am wondering if there are any plans to support mCuSeqlensM on CPU. The reason is that DeepEP's returned Tensor are always on CPU memory, and copying them to GPU would consume a significant amount of time. Thank you very much for your consideration!

[tridao]

I haven't thought about it. You can try it out and if it works, would love to see a PR!