perf: avoid recomputing rotary_emb for each layer in some Google and ModernBERT models

[casinca]

What does this PR do?

Following #45144 (comment)

tl;dr: self.rotary_emb was called N_layers times but only 2 calls are needed (full attn and swa) other calls were overwritting the same dict keys, which was inefficient.
This now gives a little boost depending on config.layer_types size, with the 4B model inference can be ~10% faster (see img below for a benchmark with 300 toks)

See picture below (of the comment #45144 (comment)) for details (comment link will not be directly available after resolve and will just point to the linked PR)

Click to expand image

Code Agent Policy

The Transformers repo is currently being overwhelmed by a large number of PRs and issue comments written by
code agents. We are currently bottlenecked by our ability to review and respond to them. As a result,
we ask that new users do not submit pure code agent PRs at this time.
You may use code agents in drafting or to help you diagnose issues. We'd also ask autonomous "OpenClaw"-like agents
not to open any PRs or issues for the moment.

PRs that appear to be fully agent-written will probably be closed without review, and we may block users who do this
repeatedly or maliciously.

This is a rapidly-evolving situation that's causing significant shockwaves in the open-source community. As a result,
this policy is likely to be updated regularly in the near future. For more information, please read CONTRIBUTING.md.

• I confirm that this is not a pure code agent PR.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

It was discussed with @vasqu and also @zucchini-nlp.
I kept Raushan suggestion with just set() (but can also do self.rotary_emb.layer_types it's a minor detail either way.)

[casinca]

I checked gemma4, they are correctly doing:

transformers/src/transformers/models/gemma4/modular_gemma4.py

Line 1236 in 85099df

self.unique_layer_types = set(self.config.layer_types)

But I'll check if some other Google models are also doing it the unoptimized way

[vasqu]

Careful approval, just checking with run-slow + update us if you find other models :D

[vasqu]

run-slow: gemma3

[github-actions]

Workflow Run ⚙️

This comment contains run-slow, running the specified jobs:

models: ["models/gemma3"]
quantizations: []

[casinca]

Careful approval, just checking with run-slow + update us if you find other models :D

Yep, 4 other models and upon looking, it seems safe to extend

Since it's 1 LoC change, I believe it's not worth extra PR noise, I can broaden the changes here + update title? Let me know what you'd prefer, thanks.

[vasqu]

Yes sounds good! Can you wait for the slow tests to return results before pushing? Other than that, feel free to add all of them here + lets adjust the PR title

[github-actions]

CI Results

Workflow Run ⚙️

Commit Info

Context	Commit	Description
RUN	6adc6aa0	workflow commit (merge commit)
PR	67620c54	branch commit (from PR)
main	85099df9	base commit (on main)

✅ No failing test specific to this PR 🎉 👏 !

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: gemma3, gemma3n, modernbert, modernbert_decoder, t5gemma2

[zucchini-nlp]

LGTM

[zucchini-nlp]

run-slow: gemma3, gemma3n, modernbert, modernbert_decoder, t5gemma2

[github-actions]

Workflow Run ⚙️

This comment contains run-slow, running the specified jobs:

models: ["models/gemma3", "models/gemma3n", "models/modernbert", "models/modernbert_decoder", "models/t5gemma2"]
quantizations: []

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[github-actions]

CI Results

Workflow Run ⚙️

Commit Info

Context	Commit	Description
RUN	e428da9d	workflow commit (merge commit)
PR	31bb4ddc	branch commit (from PR)
main	77c0e6e7	base commit (on main)

✅ No failing test specific to this PR 🎉 👏 !

[vasqu]

Awesome, merging this :)