[serge] Fix 22 integration tests for model whisper failing with other (other (20), shape/size mismatch (2))

[sergereview]

Original CI failure

• Failure group: 1/1: 22 integration tests for model whisperfailing withother (other (20), shape/size mismatch (2))
• tests/models/whisper/test_modeling_whisper.py::WhisperModelIntegrationTests::test_distil_token_timestamp_generation [multi-gpu] (other, seen 7/7)

Relates to #47053

The Whisper integration tests fail with UnboundLocalError: local variable 'is_updated' referenced before assignment (and cascading errors) because WhisperAttention.forward only defines is_updated inside the isinstance(past_key_values, EncoderDecoderCache) branch. When a different cache type is used, the variable is undefined and the subsequent cross-attention reuse check crashes.

Root cause: missing initialization of is_updated before the conditional, unlike the equivalent code in bart and other encoder-decoder models.

Patch: add is_updated = False before the EncoderDecoderCache check in WhisperAttention.forward so the fallback path behaves correctly for non-EncoderDecoderCache caches.

---

This change was produced automatically by serge from a CI failure report. The patch was generated by an LLM and applied by serge; review before merging.

serge v0.1.0 · model: moonshotai/Kimi-K2.7-Code · 53 LLM turns · 60 tool calls · 111.0s · 2093703 in / 4990 out tokens

[github-actions]

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

run-slow: whisper

[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 recap

Dashboard: View test results in Grafana
Latest run: 28699482656:1
Result: failure | Jobs: 15 | Tests: 171,501 | Failures: 7 | Duration: 24h 32m