[fix][whisper]: fix max_new_tokens handling

[eustlb]

What does this PR do?

PR #44130 changed how the generation loop enforces max_length: it is now the strict total decoder length (forced/initial tokens + generated content), whereas the old cache_position based approach effectively didn't count one position.

Whisper prepends forced decoder tokens before the transcript ([<|startoftranscript|>, <|lang|>, <|transcribe|>, <|notimestamps|>]). So when a user passes max_length=20 they mean "20 content tokens," and _set_max_new_tokens_and_length is supposed to inflate the limit to leave room for that forced prefix. Two defects broke this:

1. The compensation was computed but discarded.
   The function computed max_length = min(generation_config.max_length + num_initial_tokens, …) but a refactor Fix whisper kwargs and generation config #30018 dropped the line that wrote it back. That refactor
   reworked _set_max_new_tokens_and_length from returning a kwargs dict (which set kwargs["max_length"]) to mutating generation_config in place. The max_new_tokens branch got its in-place replacement, but the max_length branch's equivalent (generation_config.max_length = max_length) was never carried over. It became dead code, harmless under the old loop, actively harmful once [generate] Completely stop relying on cache_position to prepare inputs #44130 made max_length strict.

2. An off-by-one hidden in num_initial_tokens.
   The compensation used decoder_input_ids.shape[-1] - 1. That -1 existed to cancel the old loop's free (uncounted) position, so the net inflation still came out to the full prefix length. PR [generate] Completely stop relying on cache_position to prepare inputs #44130 removed that quirk, so the -1 now under-counts the prefix by one, leaving the output one token short even after re-adding the assignment.

TD;DR: #30018 planted latent dead code, #44130 activated it

Fixes:

• test_modeling_whisper.py::WhisperModelIntegrationTests::test_tiny_generation
• test_modeling_whisper.py::WhisperModelIntegrationTests::test_tiny_en_generation

[github-actions]

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

run-slow: whisper

[eustlb]

run-slow: whisper

[github-actions]

Workflow Run ⚙️

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

models: ["models/whisper"]
quantizations: []

[github-actions]

CI Dashboard: View test results in Grafana

[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	fe761868	workflow commit (merge commit)
PR	5c10ba10	branch commit (from PR)
main	1048e9af	base commit (on main)

✅ No failing test specific to this PR 🎉 👏 !

[vasqu]

Wow thanks for catching this. It's really hard to keep up with custom generation loops 😢