yaoshiang comments

Results 15 comments of


                                            yaoshiang

TPU test flake: SIGSEGV in train_decoder_only_eager_spmd_data_parallel.py

Good news is that I was able to repro this in isolation on an isolated machine, a v6e-4 using v2-alpha-tpuv6e.

Torchvision bounding boxes do not match the images, becuase the bboxes are from the pre-cropped, pre-resized version.

I'll look into it. If I'm able to generate valid boxes, how do you recommend I store them? I could do what we did with CelebA and put them on...

Torchvision bounding boxes do not match the images, becuase the bboxes are from the pre-cropped, pre-resized version.

The issue here is that there are actually two sets of images: uncropped "in the wild" and cropped. The identity and binary attributes obviously don't change. For landmarks, there are...

torch.linalg.lstsq issues on GPU/TPU

Thanks for sharing this issue. Can you share the line that triggers the exception? I assume the call to `lstsq`. Can you also run it with E and F directly...

torch.linalg.lstsq issues on GPU/TPU

Thanks so much. Looks like an issue with the longer series of ops during compilation. Can you try this as a potential workaround with xla.step(): diff = torch.randn(8,3,1,requires_grad=True,device=device) A =...

torch.linalg.lstsq issues on GPU/TPU

I ran it on pt-cuda and it worked successfully.

torch.linalg.lstsq issues on GPU/TPU

> [@yaoshiang](https://github.com/yaoshiang) what version of torch are you using? I tested this on 2.4 GPU and 2.6 TPU with the same error. Also - what variant of the code are...

torch.linalg.lstsq issues on GPU/TPU

BTW are you in touch with Ela Jamali?

The "Pytorch/XLA overview" is very long, goes into advanced topics, and is overall intimidating for new users.

I think a revamp of this is probably best handled as part of Chris Jones's larger look at docs.

Abstract xla::OpSharding

Is there a possibility to not make API changes but instead allow setting a flag for when we want PJRT to know about less than the global view? On a...