jeffhataws
jeffhataws
Looks like there's python 3.12/3.13 version are available now in nightlies for torch-xla 2.8: https://github.com/pytorch/xla/issues/8959
Actually, I checked and it is not yet working. I updated and reopened https://github.com/pytorch/xla/issues/8959
Hi @yaochengji just checking if we will have this for torch-xla v2.8.
- Original PR: [Enable bucketized all-reduce for gradients #7216](https://github.com/pytorch/xla/pull/7216) - Reason to backport: Parity with Neuron branch r2.1_aws_neuron - Backport link: WIP
We suspect it could be a PJRT async issue, similar to this note in the original PR: https://github.com/pytorch/xla/pull/8849#issuecomment-2749554914
> The impact, here, would be that we could see performance regressions. Basically, we were skipping creating another tensor (i.e. copying) whenever the tensor was already a contiguous tensor on...
"TRACE" is not in the enum list. Has the list changed? Or the "TRACE" usage is committed before the enum list change?
@qihqi do you have updates for this issue? Looks like now we see this issue when we use xp.Trace.
> [@rpsilva-aws](https://github.com/rpsilva-aws) [@jeffhataws](https://github.com/jeffhataws) Would it also be okay to remove `XLA_USE_32BIT_LONG`? We will need to keep that until we have the feature request https://github.com/pytorch/pytorch/issues/141994 in Pytorch resolved.
I need to check on the readiness of deprecating XLA_USE_FP16 and XLA_DOWNCAST_FP16