[fix:]: initialize ModernBertUnpaddedRotaryEmbedding when loading from a checkpoint

[emapco]

What does this PR do?

This fix improves accuracy and prevents existing ModernBert models from yielding NaN tensors. Previously, when a pretrained modernbert model is loaded, the ModernBertUnpaddedRotaryEmbedding's inv_freq non-persistent buffer is not initialized resulting in NaN loss and logits.

Without fix:

• ModernBertForSequenceClassification ModernBertModelIntegrationTest.test_inference_sequence_classification_flash_attention_2 rarely returns NaN tensors but instead fails due to inaccurate logits
• ModernBertModelIntegrationTest.test_inference_sequence_classification_flash_attention_2_modernbert_base always fails due to NaN logits.

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?

@ArthurZucker @Cyrilvallez

[github-actions]

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

run-slow: modernbert

[Copilot]

Pull request overview

This PR fixes a critical bug in ModernBERT models where loading from a checkpoint with flash_attention_2 results in NaN tensors due to an uninitialized inv_freq buffer in ModernBertUnpaddedRotaryEmbedding. The fix ensures proper initialization during model loading and adds comprehensive tests to verify the bug is resolved.

Key changes:

• Initializes ModernBertUnpaddedRotaryEmbedding's non-persistent inv_freq buffer and cache attributes in _init_weights
• Adds two new integration tests specifically for flash_attention_2 implementation to ensure NaN values are not produced

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
src/transformers/models/modernbert/modular_modernbert.py	Adds initialization logic for ModernBertUnpaddedRotaryEmbedding in _init_weights method to compute and register the inv_freq buffer and initialize cache attributes
src/transformers/models/modernbert/modeling_modernbert.py	Auto-generated from modular file - contains identical initialization logic as modular_modernbert.py
tests/models/modernbert/test_modeling_modernbert.py	Adds two new slow tests for flash_attention_2 sequence classification with NaN checks to verify the fix works for both tiny-random and base models