[guygir]

Overview

When /render is called with return_mm_kwargs=false, skip encoder-tensor
construction if the processor can still emit exact token_ids, mm_hashes,
and mm_placeholders. Complements vllm-project#60728 and vllm-project#46722; it does not change
return_mm_kwargs=true.

Production code under vllm/ (this PR only, base main): +227 / −10 (approx).

Claims

• Routing-only /render can avoid materializing encoder tensors.
• The fast path is fail-closed: custom or tensor-dependent processors keep
  full preprocessing.
• Chat, completion, Messages, and Responses rendering pass
  routing_metadata_only=not return_mm_kwargs.

Validation

Local/cluster tests on the rebased commit. Result: 7 passed.

pytest -q tests/multimodal/test_processing.py -k routing_metadata \
  tests/entrypoints/scale_out/render/test_render.py -k \
  'return_mm_kwargs or test_render_responses_returns_generate_request_without_stored_state'

CI should cover tests/multimodal (misc.yaml) and
tests/entrypoints/scale_out (entrypoints.yaml).

This PR is the generic path only. Family speedups are in the follow-up
adapter PRs.

Details

return_mm_kwargs=false already drops tensors from the HTTP response
(vllm-project#60728), but MultiModalProcessor.apply() can still build them. Cache-aware
routers never use those tensors.

apply_routing_metadata() reuses existing prompt updates when they do not
read processed kwargs. Tensor-dependent updates, custom apply, custom
prompt finalization, encoder-decoder processors, and direct kwargs-data
access fall back to full apply().

No model-runner, scheduler, or KV-cache code changes. Default generate
behavior is unchanged.

AI assistance was used to implement and validate this change. A human
submitter reviewed the diff.

---

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).

• AI assistance was used during the creation of this PR.

• Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.

• Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.

• Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.

• Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.