Guilherme Leobas

Partially fixes https://github.com/numba/numba/issues/9414

This is a follow-up of https://github.com/numba/numba/pull/8663. I only rebased the PR, removed a part of the code that wasn't covered by tests and added a release note.

Support parallel dynamic gufuncs

7

Fix: https://github.com/numba/numba/pull/9339 Moved the implementation of dynamic gufucs from `GUFunc` to `GUFuncBuilder`, deleted `gufunc.py` and added a new test class for parallel GUFuncs.

Teach dynamo about vjp

2

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #120338 * #119926 * __->__ #119405 * #118407 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @aakhundov @kadeng

[Dynamo] Fix alias issue with respect to wrapped numbers (#124731)

9

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #124774 This PR fixes an issue presented when calling `aten.alias(int)` raises a TypeError. ```python import torch import torch.autograd.forward_ad as fwAD def f(x):...

[Dynamo] Trace through `cxx_pytree` API with dynamo

2

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #121931 Internally, dynamo converts cxx_pytree calls into pytree ones cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng...

Unwrap functorch tensor if dead on metadata analysis

1

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #126894 * #126878 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang

Avoid accessing storage in wrapper tensor

2

Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * #126894 * __->__ #126878 ---- * Prevent aot autograd metadata analysis from calling `untyped_storage()` if tensor is a wrapper * Add tests to...

`@functools.wraps` graph breaks in many cases where we should be able to handle it

1

### 🐛 Describe the bug Currently, PyTorch crashes if two functions annotated with `@functools.wraps(func)` are composed together. As an example, when adding support for `torch.func.hessian` in https://github.com/pytorch/pytorch/pull/121410, `wraps` is conditionally...