4D mask documentation updates

[poedator]

following #27539 this PR adds updates to transformers documentation to reflect possibility of utilizing 4D masks.

Plan:

• add updates for Llama model docstring(s)
• identify other models that can use 4D masks in present form (which requires ability to accept custom position_ids argument) and updating their docstrings. Classes that need updates:
  • Falcon Model
  • [TODO identify more]
• update code comments that may need corrections, like cases where the mask may be either 2D or 4D now. one example is based on this comment by @shentianxiao

Update 20.12.2023:
to find out which models require docstring changes, I scanned all model classes in transformers insing inspect.

• excluded tf and jax classes
• excluded models without position_ids argument in .forward() - can't use 4D mask effectively
• excluded models that do not use _prepare_4d_attention_mask method - need different code change to use 4D mask
• excluded multi-modal models (clip, clvp, vit, bark, git)

what is left is LlamaModel, FalconModel and XGLMModel.

cc @ArthurZucker

[ArthurZucker]

Feel free to ping me for a review whenever this is ready 🤗

[poedator]

Feel free to ping me for a review whenever this is ready 🤗

@ArthurZucker , I only identified 3 applicable model classes and made changes. Please check my logic in classes selection in my big first message above.

[ArthurZucker]

#28132 might be relevant to you! I'll make sure 4d can still be used but will make things a lot easier I think.
LGTM otherwise!

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[alex-hh]

Has something like this or equivalent been merged? I don't see much documentation for 4D masks but would find useful!

[ArthurZucker]

Yep sorry, it's supported for some models.

transformers/src/transformers/models/bloom/modeling_bloom.py

Line 65 in 18527bd

A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.

doc is a bit scarce, feel free to open a pr to add this!