[bugfix] Update modeling_llama.py so it skips keys correctly

[HDCharles]

the llama model was using past_key_value and past_key_values interchangeably which caused issues because only one of those was actually skipped in _skip_keys_device_placement when both needed to be.

This PR changes it so only past_key_values is used.

without this fix, llama + torch.compile was having issues as in the surfaced issue below. Note, though this was surfaced due to a TorchAO recipe though the bug was unrelated to torchao

i think #35763 might have caused this bug but not sure

Fixes

pytorch/ao#1705

• text models: @ArthurZucker (not sure if this is the right person)
• @hmellor @Rocketknight1 (last 2 hf people to touch this code)
• @SunMarc (previously worked on torchao hf stuff)

[jerryzh168]

cc @SunMarc can you take a look?

[HDCharles]

are these test failures actually blocking? it seems like other models just have the same issue.

[SunMarc]

This PR that was merged today actually fixes the issue that you are having (Cache is not a nn.Module anymore, so it will be skipped automatically).

[SunMarc]

Still, it would be indeed be better to only keep one of them. @ArthurZucker why do we have past_key_value and past_key_values ?