Mario Lezcano Casado comments

Results 96 comments of


Mario Lezcano Casado

Treat layout / pin_memory consistently across creation refs

@pytorchbot merge

Treat layout / pin_memory consistently across creation refs

@pytorchbot merge -f "flaky landing tests"

OpInfo: use functools.partial to decrease noise in make_tensor calls

Ahhh, nice!

Optimizer Limited Symmetric Rank 1

I think this sort of contribution would be much better suited as an external package on an external repo with its own testing and docs. As Alban said, we usually...

OpInfo: Use yield consistently instead of appending to lists

Incredible effort! It would be good to add a test to make sure that people keep using this pattern rather than the list pattern.

CUDA: Illegal memory access in `torch.linalg.solve()`

Does this still happen in `master`? This function was improved quite a bit after the 1.12 release.

CUDA: Illegal memory access in `torch.linalg.solve()`

Actually, yes, this path is also used in `master`

CUDA: Illegal memory access in `torch.linalg.solve()`

@IvanYashchuk may this one be related to https://github.com/pytorch/pytorch/issues/82894 at all?

Add an offset to torch.eye + simplify its implementation

I think there's some errors. I'll work on fixing them and then we can try to land this one.

Add an offset to torch.eye + simplify its implementation

So, I ended up introducing a number of changes, as the operation can be made pretty much "compositional". Doing so we can do without he CPU/CUDA/MPS implementations, which is great....