TestNormNuclear does not test the nuclear norm

[lezcano]

It just tests the vector 1-norm on a 3D tensor.

The test that actually tests the nuclear norm is called TestNuclearNorm and it's commented out.

Now, all these operations are implemented in PyTorch core in terms of https://pytorch.org/docs/stable/generated/torch.linalg.matrix_norm.html, which is itself a composite function. This was added in https://github.com/pytorch/pytorch/pull/84624. aten::norm and aten::frobenius_norm and aten::nuclear_norm for BC with JIT models from previous versions, but are not currently accessible from the Python API (at least not through documented functions).

[lezcano]

Actually, in https://github.com/pytorch/pytorch/pull/84624, the dispatch was changed just for cpu, cuda, and meta. If vector_norm and matrix_norm are properly supported in XLA, this could also be changed for XLA. https://github.com/pytorch/pytorch/blob/32b2d8009a072234e6ff99127b8d96b6f41d1fff/torch/functional.py#L1489

[JackCaoG]

Hmm, if I enable the test and leave the

ExpectCounterNotChanged("aten::.*", cpp_test::GetIgnoredCounters());

which checks whthere there is any fallback, will test pass(I hope?)

[lezcano]

If they pass, uncommenting the tests LGTM

[lezcano]

But moving forward, it'd be good to have lowerings for linalg.vector_norm and the SVD so that we can add XLA to that better path in PyTorch core: https://github.com/pytorch/pytorch/blob/32b2d8009a072234e6ff99127b8d96b6f41d1fff/torch/functional.py#L1489

[JackCaoG]

Thanks, I will mark this pr as a op lowering task and assign owner accordingly.

[alanwaketan]

Maybe Meghan can take this task.