nikitaved
nikitaved
@pytorchbot merge
I think it used to break autograd, but that might have changed, cc @albanD .
> ```python > csr = torch.sparse_csr_tensor((0, 1, 2), (0, 1), (1, 1), dtype=torch.float32, requires_grad=True) > csr2 = csr.to_sparse(layout=torch.sparse_csr).detach().requires_grad_(True) > x = torch.ones((2, 1), dtype=torch.float32) > y = torch.matmul(csr2, x) >...
@pytorchbot merge -g
Closing in favor of stack https://github.com/pytorch/pytorch/pull/94823.
@rgommers , if there is something for the GPU which is missing for the CPU in PyTorch, that is most likely because of gpu performance/functionality having much higher value... Especially...
> For example, it does not exist for a matrix `A` with `A[0,0] = 0`. With this limitation, there are very few families of matrices for which this property holds....
> The diagonal case you mentioned I don't think it's correct, but it's true that there is at least one family of matrices that I know of for which LU...
I am also curious about the dtype. Do you know whether it is float, or half/bfloat16?
@t-vi , I have updated this PR with tests.