Bhavya Bahl comments

Results 26 comments of


                                            Bhavya Bahl

Runtime is already initialized. Do not use the XLA ' RuntimeError: Runtime is already initialized. Do not use the XLA device before calling xmp.spawn.

Can you share more details about your notebook? It seems like this error occurs due to accessing some torch_xla code before calling `xmp.spawn`

`F.embedding_bag(..., mode='max')` yields different results than PyTorch eager.

I can take a look at this.

Unnecessary FP64 cast for `padding_idx` in `EmbeddingDenseBackward`

I don't know the reason for casting `padding_idx` either. On the surface, it looks unnecessary but I am not sure if it was done to avoid some XLA compiler error....

Failing `torch_xla._XLAC._xla_custom_call()` with `RuntimeError: Bad StatusOr access: UNIMPLEMENTED: No registered implementation for custom call to my_lib.my_op.default for platform CUDA`

@lsy323 Do you know if StableHLO export works for custom calls?

Publish Python 3.12 Torch XLA wheels (non-nightly)

@metrizable We are going to have 3.12 wheels for 2.8 stable release in a week. It would be extra work for us to go back and add 3.12 wheels for...

pytorch-xla and lib-tpu version mismatch with latest nightly image

@jialei777 Is this fixed after upgrading to 0827 nightly?