[Bugfix]: Clean up chunked prefill logging when using whisper

[simondanielsson]

Purpose

Closes #25071.

Test Plan

1. When using whisper:

vllm serve openai/whisper-large-v3

logs should no longer mention "Chunked prefill is enabled with ...":

(APIServer pid=3140911) INFO 09-17 12:37:08 [scheduler.py:222] Chunked prefill is enabled with max_num_batched_tokens=8192.
(APIServer pid=3140911) INFO 09-17 12:37:10 [__init__.py:2790] Encoder-decoder models do not support chunked prefill nor prefix caching; disabling both.

Expecting simply

(APIServer pid=3140911) INFO 09-17 12:37:10 [__init__.py:2790] Encoder-decoder models do not support chunked prefill nor prefix caching; disabling both.

2. Should lead to no changes to the SchedulerConfig nor VllmConfig. Verify with new tests.

Test Result

1. Command:

• Tested on GPU: L4.
• Output from "test" command:

(vllm) danielssonsimon@XXXXXX:~/code/vllm$ vllm serve openai/whisper-large-v3
INFO 09-17 18:43:30 [__init__.py:216] Automatically detected platform cuda.
(APIServer pid=49917) INFO 09-17 18:43:33 [api_server.py:1813] vLLM API server version 0.10.2rc3.dev169+ge3db5ebb6.d20250917
(APIServer pid=49917) INFO 09-17 18:43:33 [utils.py:328] non-default args: {'model_tag': 'openai/whisper-large-v3', 'model': 'openai/whisper-large-v3'}
(APIServer pid=49917) INFO 09-17 18:43:42 [__init__.py:707] Resolved architecture: WhisperForConditionalGeneration
(APIServer pid=49917) `torch_dtype` is deprecated! Use `dtype` instead!
(APIServer pid=49917) INFO 09-17 18:43:42 [__init__.py:1762] Using max model len 448
(APIServer pid=49917) INFO 09-17 18:43:43 [scheduler.py:197] Encoder-decoder models do not support chunked prefill nor prefix caching; disabling both.
Fetching 1 files: 100%|█████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 11915.64it/s]

2. New tests pass locally.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @simondanielsson.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[hmellor]

If this already exists in ModelConfig, why duplicate it here?

[simondanielsson]

True, we likely don't want to store it here as well.

Would an InitVar be sufficient here?

[hmellor]

The InitVar solution works.

However, in other cases like this (where two sibling configs interact) I've tended to perform those interactions in the parent's __post_init__, VllmConfig in this case. Would that work in this case?

[russellb]

That's where I had it before this change, but we end up with a confusing log message about features being enabled coming from the SchedulerConfig's post_init before VllmConfig's post_init fixed it and disabled them.

[simondanielsson]

Another option would be to perform the Chunked prefill is enabled... log in the VllmConfig, but not sure it makes sense to put it there

[hmellor]

Ah I see, thankn you for explaining. Let's stick with the initvar

[russellb]

lgtm, as long as @hmellor is ok with your change to use an InitVar. Thanks!

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @simondanielsson.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @simondanielsson.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[hmellor]

Sorry we should have merged this already. @simondanielsson can you please update the branch and we'll enable auto-merge?

[simondanielsson]

@russellb seeing CI is failing now in an (to me seemingly) unrelated place - are you more familiar with the kv connectors and can see directly what is causing this err? Otherwise I'll debug this after the weekend :) Thanks!

[2025-09-27T09:55:31Z] FAILED v1/kv_connector/unit/test_offloading_connector.py::test_offloading_connector - AssertionError: Expected 'lookup' to not have been called. Called 1 times.
--
  | [2025-09-27T09:55:31Z] Calls: [call(<itertools.islice object at 0x7f08ddc4fbf0>)].
  | [2025-09-27T09:55:31Z]
  | [2025-09-27T09:55:31Z] pytest introspection follows:
  | [2025-09-27T09:55:31Z]
  | [2025-09-27T09:55:31Z] Args:
  | [2025-09-27T09:55:31Z] assert (<itertools.i...f08ddc4fbf0>,) == ()
  | [2025-09-27T09:55:31Z]
  | [2025-09-27T09:55:31Z]   Left contains one more item: <itertools.islice object at 0x7f08ddc4fbf0>
  | [2025-09-27T09:55:31Z]
  | [2025-09-27T09:55:31Z]   Full diff:
  | [2025-09-27T09:55:31Z]   - ()
  | [2025-09-27T09:55:31Z]   + (
  | [2025-09-27T09:55:31Z]   +     <itertools.islice object at 0x7f08ddc4fbf0>,
  | [2025-09-27T09:55:31Z]   + )
  | [2025-09-27T09:55:31Z] !!!!!!!!!!!!!!!!!!!!!!!!!! stopping after 1 failures !!!!!!!!!!!!!!!!!!!!!!!!!!!
  | [2025-09-27T09:55:31Z] ============= 1 failed, 20 passed, 2 warnings in 86.19s (0:01:26) ==============
  | [2025-09-27T09:55:37Z] 🚨 Error: The command exited with status 1