nikitaved

Results 69 comments of nikitaved

`mul`: sparse-dense + sparse-sparse with 0-dims support take 2.

Thank you, @kit1980 , for your comments! Unless there is anything else, can I merge it on green? @bhosmer , could you please stamp it just to be sure we...

`mul`: sparse-dense + sparse-sparse with 0-dims support take 2.

@kit1980 Ed had an issue with merging his contribution marked as sparse unless sparse people approved that, hence the concern. Thanks for taking care of this! Let me try to...

`mul`: sparse-dense + sparse-sparse with 0-dims support take 2.

@pytorchbot merge

csr <-> csc, csc <-> csc, bsr <-> bsc, bsc <-> bsc, bsr <-> bsr conversions

@amjames , I have done the extension. Could you please have another look?

resize_as_sparse support all compressed layouts

@amjames , is this PR being tested somewhere in the stack?

Remove deprecated torch.lstsq

@de-gozaru , I have had some issues with the determinism of linalg.lstsq while working on its forward/backward ADs. If my memory serves me well, it was the case with only...

Remove deprecated torch.lstsq

It makes sense to play with a driver. Not all of them are non-deterministic.

sparse mm/addmm enable dense x csc, csc x dense and simplify layout check logic.

Shall I stamp it as well?

Improve `bsr @ strided` performance in `baddmm` for `bfloat16/half` with Triton kernels.

@cpuhrsch , the CI failures are real. These happen at import with `from torch._dynamo.utils import has_triton`.

Improve `bsr @ strided` performance in `baddmm` for `bfloat16/half` with Triton kernels.

Yes, they are reproducible. I have not yet figured out the root cause.