Pytorch CI failing

[ricardoV94]

PyTorch CI started failing in a recent PR. I guess they did some breaking changes. Strangely enough it can do import torch.compiler?

Other things are failing like it not liking numpy dtypes?

Closes #1723

[ricardoV94]

It seems to be installing a rather old version of pytorch?

  pytorch                    2.0.1         py3.11_cpu_0           pytorch    

So maybe that's why. It is currently at 2.9.0.

CC @maresb @Ch0ronomato

[maresb]

@ricardoV94, it's caused by 8a6e407

[ricardoV94]

I've lost the context for that mkl thing

[maresb]

I'm pretty sure you're safe reverting that upper bound.

[ricardoV94]

Odd error:

FAILED tests/link/pytorch/test_extra_ops.py::test_pytorch_Unique_axis[1] - AssertionError: expected size 3==3, stride 1==2 at dim=0; expected size 2==2, stride 3==1 at dim=1
Error in op: torch.ops.aten.unique_dim.default
This error most often comes from a incorrect fake (aka meta) kernel for a custom op.

[ricardoV94]

It stopped automagically failing in other CI. I'll try to rerun without the Pin anyway

[ricardoV94]

2 tests fail in the most recent pytorch:

• FAILED tests/link/pytorch/test_extra_ops.py::test_pytorch_Unique_axis[1] - AssertionError: expected size 3==3, stride 1==2 at dim=0; expected size 2==2, stride 3==1 at dim=1
  Error in op: torch.ops.aten.unique_dim.default

• FAILED tests/link/pytorch/test_nlinalg.py::test_eig - AssertionError: expected size 3==3, stride 1==3 at dim=0; expected size 3==3, stride 3==1 at dim=1
  Error in op: torch.ops.aten.linalg_eig.default

May be worth checking this so we can remove the MKL pin? Or is it indeed related to it?

[maresb]

I'm guessing that this happens in practice, and you're only seeing it by busting the cache.

Unpinned dependencies are such a massive headache. I really need to finish that pixi PR. 😞