Jeffrey Wan

Thanks for the report @awgu, this overhead is because the type of the IValue needs to be manually inferred in the custom Function case because there's no schema for us...

@pytorchbot merge

Modifying autograd kernel generation signature would mean that the dispatcher would need to understand mutable tensor list returns, which we want to avoid. Python binding level should refer to something...

Public API sounds okay, but preferably with a clear story around when we'd use this over existing APIs that do similar "forbid_in_autograd" things like `.detach()` or setting `.requires_grad=False`. (We also...

> We still planning to do factory function support? Did I miss it in this PR stack :p Still planned, yes. Should be easy to do with the new infra...

> I'm still curious whether we can fix this in prim.layout.default or something, but the short-term fix seems okay Yup the root cause seems to be with `prim.layout.default` in that...

@pytorchbot merge

Error log: ``` 2024-04-12T01:43:36.6700909Z ____________________ TestAOTAutograd.test_set__not_allowed _____________________ 2024-04-12T01:43:36.6701501Z Traceback (most recent call last): 2024-04-12T01:43:36.6702082Z File "functorch/test_aotdispatch.py", line 540, in test_set__not_allowed 2024-04-12T01:43:36.6702891Z inp = [torch.ones(3, 3, requires_grad=True), torch.ones(3, 3, requires_grad=True)] 2024-04-12T01:43:36.6703843Z...

This test was recently introduced in https://github.com/pytorch/pytorch/pull/122981 cc @bdhirsh

> Came to this issue from a code reference and following up out of curiosity as to what's left to do here. Do failing tests, slow or not, need to...