pytorch

5b164251 - [with_effects] Skip over profiler.record_function_exit (#121829)

[with_effects] Skip over profiler.record_function_exit (#121829) Summary: tldr: User calls to `torch.autograd.profiler.record_function` fails when tracing with non-strict pre-dispatch export due to an effect token failure, so the solution is to skip over these operators 😅 Some user code contains calls to a `torch.autograd.profiler.record_function` context, like https://fburl.com/code/uesgknbq and https://fburl.com/code/iogbnsfw, which is used for adding user-defined events into the profiler. Currently these function calls will be skipped/removed in dynamo (https://fburl.com/code/fkf7qmai) but **non-strict pre-dispatch export** will hit these operators during tracing. However, it seems that although these operators get hit by the dispatcher, they don't actually show up in the final graph (maybe they get DCE-d). However, an issue comes up with a recent change with effect tokens (D54639390) which creates tokens if it sees a ScriptObject during tracing. The operator `torch.ops.profiler.record_function_exit` takes in a ScriptObject, so the effect tokens framework now tries to add an effect token to this operator, but results in the following error: (https://www.internalfb.com/intern/everpaste/?handle=GI-hvBknzj2ZxYkBABNzdztDxJVAbsIXAAAB, P1195258619) The reason is because this operator only gets hit during pre-dispatch, not post-dispatch tracing. During pre-dispatch tracing, we first trace using post-dispatch to collect metadata needed for functionalization, and then we do pre-dispatch tracing to construct the graph. The metadata collection phase is also when we determine what operators need effect tokens and create those tokens. However, since the operator only shows up in pre-dispatch tracing, we do create any tokens. During the actual pre-dispatch tracing to create the graph, we then run into this operator and try to get a token, but none exist, causing an error :( This PR just blocks the record_function operator from being looked at by the effect tokens framework. But a proper fix might be to have functionalization run on the pre-dispatch graph or have the operator also show up in the post-dispatch graph. But since in the PT2 stack dynamo just gets rid of this operator so that it won't show up anywhere downstream, I think we can also just ignore this operator. Test Plan: Fixed test for P1195258619 Differential Revision: D54857444 Pull Request resolved: https://github.com/pytorch/pytorch/pull/121829 Approved by: https://github.com/BoyuanFeng, https://github.com/tugsbayasgalan