JackCaoG
JackCaoG
Ah, it is actually caused by some of the ops fall back to CPU ``` Counter: aten::_local_scalar_dense Value: 1 Counter: aten::count_nonzero.dim_IntList Value: 1 Counter: aten::nonzero Value: 2 Counter: aten::prelu_backward Value:...
We are trying to do the experimental release for dynamic shape at the end of this year. The context is we already have a way to make `nonzero` op to...
@wonjoolee95 can you fix the linter?
Thanks for reporting, If you could provide a small repo we might be able to take a look. I am guessing the issue is the outputs is on CPU but...
Windows or Linux shouldn't make difference. Depending on how which actual device you want to run on you can do `GPU_NUM_DEVICES=1` or any number will trigger GPU mode, ``` export...
I would start with printing out the tensor type.. my best guess is tensor is actually on CPU. You need to manully move tensor or model to xla device. Check...
`LsbMask` is the error I see locally
You can build with `BUILD_CPP_TESTS=0` to get around this issue.
We don't run into the issue if we build pt/xla test on a newer [docker images](https://github.com/pytorch/xla#-available-images-and-wheels) we published, but it still fails in my local environment. This seems to be...
Local build passed for me if I set `DEBUG=0` which sets the flag in https://github.com/pytorch/xla/blob/master/setup.py#L330