[Bugfix][Multimodal] Scope cache hash kwargs by modality

[waizuichougou]

Purpose

Fix unnecessary multimodal processor cache misses caused by unrelated
modality options being included in every media item's cache hash.

The production trigger chain is:

request/config media_io_kwargs or mm_processor_kwargs
  -> renderer._process_multimodal()
  -> ProcessorInputs
  -> ProcessorInputs.get_mm_hashes()
  -> multimodal processor cache hit/miss check
  -> preprocessing of cache-missing media items

get_mm_hashes() currently builds one global hash-factor dictionary containing
all media_io_kwargs and all HF processor kwargs, then applies it to every
modality. Consequently, a video-only option changes an image hash. For
example, an image-only request can inherit:

--media-io-kwargs '{"video":{"num_frames":16}}'

With multimodal processor caching enabled, repeating that request needlessly
reprocesses the image. The same issue occurs in mixed image/video requests:
changing num_frames or videos_kwargs invalidates the image cache entry.
This is a real online performance and latency bug; it does not independently
change model output.

This change scopes hash factors for ordinary modalities:

• image uses only image media_io_kwargs and images_kwargs;
• video uses only video media_io_kwargs and videos_kwargs;
• audio uses only audio media_io_kwargs and audio_kwargs;
• flat HF processor kwargs remain included because they are shared.

vision_chunk remains conservative because one list can contain both images
and videos and ProcessorInputs does not currently carry each item's original
modality. It may still have extra cache misses, but this change does not create
incorrect cache reuse.

Test Plan

Run the focused multimodal processing and renderer hash tests:

python -m pytest tests/multimodal/test_processing.py -m cpu_test -q
python -m pytest tests/renderers/test_multimodal_hashes.py -m cpu_test -q

Run lint, formatting, and whitespace checks:

ruff check vllm/multimodal/processing/inputs.py \
  tests/multimodal/test_processing.py
ruff format --check vllm/multimodal/processing/inputs.py \
  tests/multimodal/test_processing.py
git diff --check

The regression tests compare identical media with changed modality-specific
options. They verify that changing video options preserves the image hash,
changing image options preserves the video hash, and video-only options are
ignored for an image-only request.

Test Result

tests/multimodal/test_processing.py -m cpu_test       98 passed
tests/renderers/test_multimodal_hashes.py -m cpu_test  1 passed
ruff check                                               passed
ruff format --check                                      passed
git diff --check                                         passed

The focused tests demonstrate the expected behavior at the hash boundary used
by the production cache. No media decoding or model inference semantics are
changed.

AI-assisted contribution

This contribution was developed with AI assistance.

---

Essential Elements of an Effective PR Description Checklist

• The purpose and production trigger chain are explained.
• The test plan includes focused test and lint commands.
• The test results include the focused regression results.
• Documentation update considered; none is needed for this internal cache
  hash fix.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[DarkLight1337]

Great!

[DarkLight1337]

/ci run

[github-actions]

✅ Triggered Buildkite CI #86823 for commit 1d9443ad005e.

[waizuichougou]

/ci retry

[github-actions]

✅ Queued 1 failed job(s) for retry in Buildkite CI #86823.