gpetters94

> I have recently pushed some fixes in this diff: #83092 > > You will need to rebase on top of it. The easiest way to do so would be...

@silvasean I'm picking this one up - how would that logic generalize to TupleConstruct chains? For example, in OPT the final return is of a prim.TupleConstruct, but the arguments of...

If we're just looking for the tensor output then `self.forward()[0]` should work. I'm working on a refactor of the model to see if we can get that to pass.

It's a tensor that is the "last hidden state" as per [this](https://huggingface.co/docs/transformers/model_doc/opt#transformers.OPTModel.forward). I'm not familiar with how NLP models work, so I don't know if the other returns are something...

PR's up at #858.

> @gpetters94 Could you please resolve the conflicts? Done.

It's holding up U-Net support, but it isn't the only thing blocking that.

@ZihengJiang @silvasean I made a small change in the shape logic (I forgot that two of the dims needed transposing) so I need to make another [PR](https://github.com/pytorch/pytorch/pull/83557) in Pytorch that...

@vivekkhandelwal1 @silvasean The upstream shape code is merged, so this one should be good to go.

@vivekkhandelwal1 Done.