past_key_value(s) name inconsistency causing problems

[HDCharles]

System Info

• transformers version: 4.50.0.dev0
• Platform: Linux-6.4.3-0_fbk14_zion_2601_gcd42476b84e9-x86_64-with-glibc2.34
• Python version: 3.12.9
• Huggingface_hub version: 0.28.1
• Safetensors version: 0.5.2
• Accelerate version: 1.4.0
• Accelerate config: not found
• DeepSpeed version: not installed
• PyTorch version (GPU?): 2.6.0.dev20241112+cu121 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using distributed or parallel set-up in script?: no
• Using GPU in script?: yes
• GPU type: NVIDIA H100

Who can help?

@ArthurZucker probably others

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

run https://huggingface.co/docs/transformers/main/en/quantization/torchao

Expected behavior

no error

---

this error is related to #36289

a bunch of models use past_key_value and past_key_values interchangeably, this causes issues since kwarg names are hardcoded to be skipped by the _skip_keys_device_placement attribute. This causes issues any time torch.compile is used with a model that has this issue.
The above PR fixes the issue for llama but other models like src/transformers/models/moonshine/modeling_moonshine.py
src/transformers/models/mistral/modeling_mistral.py
src/transformers/models/emu3/modeling_emu3.py
...etc, have the same issue which is actually breaking CI for that PR.

this is also the cause of pytorch/ao#1705 which is where this was first surfaced.

is there a reason for these two names to be used instead of just one? If not it seems like they should be entirely consolidated to avoid such issues, if so, then _skip_keys_device_placement needs to include both across all models or