Will Constable

I came up with a new repro that cuts dynamo, ddp, and graph-splits out of the picture. ``` import torch from torch._dynamo.utils import deepcopy_to_fake_tensor class Module(torch.nn.Module): def __init__(self): super().__init__() self.self_net_1_net_0...

I see. It seems to me that the fix for both repros would be to activate a fake mode before executing the composite op, so the scalar->tensor promotion gets fakeified....

I also seem to have a fix via https://github.com/pytorch/pytorch/pull/92986 -- do you guys think my fix should not be necessary? It seems to me that my fix is logically good...

I don't expect this repro uses cudagraphs since iiuc it is disabled by default and I did not specifically enable it. but i did not check.

Hey @bdhirsh For 'traceable collectives' we're leaning towards a design where collective apis will return a Tensor subclass that is then expected to be traced through for one purpose only:...

> One reason this is a bit harder is because functionalization is going to be turned on in aot autograd in all cases ok, i don't understand the functionalization +...

ok. i don't want to assume anyhting about the model. model could mutate. I just thought that functionalization was going to run and the issue was whether it ran before...

Well, hooks support is new, and maybe something is missing. We'll need more info to reproduce the problem. Can you share a minimal version of your script that demonstrates a...

I was able to repro this locally, but haven't had time to look into it yet. I will get to it early next week.

@johnbensnyder This should be fixed now. Closing, but please reopen if not.