[UMT5] Ignore tied encoder embedding missing-key warning

[weiguangli-io]

Summary

Fixes #43992 by preventing a false missing-key report for UMT5EncoderModel when encoder.embed_tokens.weight is tied to shared.weight.

UMT5EncoderModel already declares tied weights, but loading checkpoints that only carry the shared embedding can still surface a misleading missing key warning for encoder.embed_tokens.weight. This change marks that key as ignorable-on-load for this encoder-only class, matching the tied-embedding semantics.

Changes

• src/transformers/models/umt5/modeling_umt5.py
  • Add _keys_to_ignore_on_load_missing = [r"encoder.embed_tokens.weight"] to UMT5EncoderModel.
• tests/models/umt5/test_encoder_missing_keys.py
  • Add a regression test that loads UMT5EncoderModel from a state dict missing encoder.embed_tokens.weight and asserts the key is not reported in missing_keys.

Validation

• PYTHONPATH=src python3 -m ruff check src/transformers/models/umt5/modeling_umt5.py tests/models/umt5/test_encoder_missing_keys.py
• PYTHONPATH=src /tmp/tfpycheck/bin/pytest -q tests/models/umt5/test_encoder_missing_keys.py
• PYTHONPATH=src /tmp/tfpycheck/bin/coverage run -m pytest -q tests/models/umt5/test_encoder_missing_keys.py
• PYTHONPATH=src /tmp/tfpycheck/bin/coverage report -m src/transformers/models/umt5/modeling_umt5.py

Coverage note:

• The added regression test hits the new load-ignore behavior directly; full-file coverage remains low because the model file contains many unrelated execution paths.

[github-actions]

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

run-slow: umt5

[Rocketknight1]

I think this is a duplicate of #43880!