Horace He comments

Results 242 comments of


                                            Horace He

libtorch

Currently, there are no plans. Do you have a particular use case for this?

Make fx tracing use input tensor copies to prevent side effects

Hi, thanks for the PR :) Most of the work on this tracer is now going on here: https://github.com/pytorch/pytorch/blob/master/torch/fx/experimental/proxy_tensor.py However, about the specific proposal in this PR.... I don't know...

Batching rule for GRU cell

I suspect that would be more likely, yeah.

Integrate CUDAGraphs with AOTAutograd

cc: @eellison (this issue could also go on the TorchDynamo repo)

added list clearing FX codegen for backwards in AOTAutograd.

Note: I've currently not bothered to land this PR, as Torchscript seems to keep all of the input activations alive. We've verified that this reduces memory usage as expected if...

Add decomposition for `aten.native_batch_norm_backward` op

I think there are still some tests failing. For example, I don't believe `push_back` is a method on python lists :) https://github.com/pytorch/functorch/pull/675/files#diff-8e71d16136c7deeb8104d8d5bb0dd794c208f38856f9ced3448ab02f3e34c43aR522 To run the tests, run ``` pytest test/test_ops.py::TestDecompositionOpInfoCUDA...