Add past_key_values as inputs to TorchExportableModuleWithStaticCache forward

[justinchuby]

Feature request

With #38879, exporting with static cache as input should be supported in

transformers/src/transformers/integrations/executorch.py

Line 288 in 797860c

def forward(self, input_ids: torch.Tensor, cache_position: torch.Tensor):

. This helps the exported program preserve proper input order for downstream conversion (e.g. torch.onnx.export).

Related

• Export Huggingface models with StaticCache pytorch/pytorch#155862

cc @xadupre @yuanyao-nv @guangy10

Motivation

Better preservation of exported program signatures

Your contribution

Yes

[sandeepyadav1478]

Hey @justinchuby

I see you have pushed an MR already, I guess work been done. Lmk, if I could help anyway :-)

[justinchuby]

It has not been done yet. The PR referenced is a prerequisite.

[sandeepyadav1478]

It has not been done yet. The PR referenced is a prerequisite.

Hey @justinchuby, lmk, if you want me help/contribution anyway.
Will happy to work. :-)