Mason Chang

www.masonchang.com

Results 4 issues of


                                            Mason Chang

Prototype bucketing for dynamic shapes in Dynamo Bridge

DO NOT SUBMIT, just for visibility.

DO_NOT_MERGE

Automatically move CUDA non XLA Tensors to XLA Device and back to CUDA device

2

comment

Currently only works for inference. The assumptions don't hold for training with Autograd yet.

Replicate test dynamic shapes models with the equivalent version of torch.compile

Try to cache unsupported ops to reduce compilation time

DO NOT SUBMIT Testing and want to get benchmarks first.