Bairen Yi

@lalalapotter we expect to support the TF 2.2 release, and its release schedule is currently unknown. Support for the nightly version will come earlier, expected to be the end of...

Gentle ping for any updates.

I am wondering the exact use case you are trying to address, as we already have randomised hashing in `sklearn` ([here](http://scikit-learn.org/stable/modules/generated/sklearn.feature_extraction.FeatureHasher.html) and [there](http://scikit-learn.org/stable/modules/generated/sklearn.feature_extraction.text.HashingVectorizer.html)). Or you would like to implement this...

Same issue here, except I am using all default config and hyper parameters for Swin-B from scratch on ImageNet-1k. Gradient norm explodes quickly after several epochs.

It seems that torch-xla relies on XLA for shape inference, and PyTorch access the sizes/numel/dim after each op. If we strip out XLA, this probably need to come from somewhere...

> Some of the helpers might need to be exposed, but they're already written. Would appreciate if you could give me a pointer here to core :) > In terms...

> For size specialized traces (the easy case), asuhan advocates for directly using the size helper functions. So these size helper functions (for each and every operator in PT core)...

> Yeah, so structured kernels are trying to make it so that you can easily get the sizes and strides from PyTorch directly, rather than assuming you already have an...

I see @asuhan pushed to https://github.com/pytorch/pytorch/commit/fafb8ab1f5476418ee121f3dedb6025adc1c9e07 and corresponding changes to torch-xla side: https://github.com/pytorch/xla/tree/asuhan/xla_ltc_plugin. Nice one! I do see a TODO on shape inference: https://github.com/pytorch/pytorch/blob/lazy_tensor_staging/lazy_tensor_core/lazy_tensor_core/csrc/compiler/node_lowering.h#L19-L20. Do you plan to provide such...

@Flamefire MPI plugin needs some renovation which unfortunately is still WIP. In other words, it does not build now and we apologise for that. I will update this issue when...