Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

CI XLA failing test: test_ema_final_weight_distrib_single_device_xla

See original GitHub issue

tests/ignite/handlers/test_ema_handler.py::test_ema_final_weight_distrib_single_device_xla

https://github.com/pytorch/ignite/pull/2180/checks?check_run_id=3471518836#step:9:147

_test_ema_final_weight(get_dummy_model(), device=device, ddp=True)

...

actual = tensor([[4.0625, 4.0625]], device='xla:0', dtype=torch.float64)
expected = tensor([[4.0625, 4.0625]], device='xla:0', dtype=torch.float64)

> Operator __and__ is only supported for integer or boolean type tensors, got: f64[1,2]{1,0}

cc @sandylaker

Issue Analytics

State:
Created 2 years ago
Comments:5 (3 by maintainers)

Top GitHub Comments

1reaction

sandylakercommented, Aug 31, 2021

@vfdev-5 Thanks for the cc. I will create a PR to fix it as soon as possible.

0reactions

sandylakercommented, Aug 31, 2021

@sandylaker good question. Seems like the cast happens inside torch.testing.assert_allclose: https://github.com/pytorch/ignite/pull/2182/checks?check_run_id=3473653855#step:9:206

@vfdev-5 So a floating point will casted to torch.float64 anyway, according to torch.testing. But I think the problem might be caused by torch_xla. Please have a look at my comments in the PR #2182 .