PR #40251: [Bugfix] Forward mm_processor_kwargs in offline generate APIs

• Repository: vllm-project/vllm
• Author: wuyingjun-lucky
• State: open | merged: False
• Link: https://github.com/vllm-project/vllm/pull/40251

Description (problem / solution / changelog)

Summary

Fixes #40134.

Propagate mm_processor_kwargs through the offline completion path so multimodal overrides work consistently across the public offline APIs:

• add mm_processor_kwargs to LLM.generate()
• add mm_processor_kwargs to LLM.enqueue()
• thread the override through _run_completion(), _add_completion_requests(), _preprocess_cmpl(), and _preprocess_cmpl_one()
• pass the override into renderer.render_cmpl(..., prompt_extras=...)
• only pass prompt_extras when mm_processor_kwargs is explicitly provided, so prompt-local mm_processor_kwargs are not overwritten by None
• add focused regression tests for the public API forwarding and renderer handoff

This matches the existing LLM.chat() behavior and fixes the current mismatch between the offline API surface and multimodal processing expectations.

Why this is not duplicating an existing PR

Duplicate checks run against vllm-project/vllm on 2026-04-18:

• gh issue view 40134 --repo vllm-project/vllm --comments
• gh pr list --repo vllm-project/vllm --state open --search '40134 in:body' --limit 20
• gh pr list --repo vllm-project/vllm --state open --search 'mm_processor_kwargs offline generate enqueue' --limit 20

The issue exists and has no comments claiming ownership, and the two PR searches returned no open PRs for this fix. I also checked adjacent multimodal PRs during local triage; the open ones I found were addressing different problems.

Test Plan

/home/wuyingjun/llm/vllm/.venv/bin/python -m pytest tests/entrypoints/llm/test_mm_processor_kwargs.py -v
/home/wuyingjun/llm/vllm/.venv/bin/python -m py_compile vllm/entrypoints/llm.py tests/entrypoints/llm/test_mm_processor_kwargs.py
/home/wuyingjun/llm/vllm/.venv/bin/pre-commit run --files vllm/entrypoints/llm.py tests/entrypoints/llm/test_mm_processor_kwargs.py

Test Result

• pytest: 6 passed in the targeted regression file
• py_compile: passed
• pre-commit: passed for all hooks on the changed files, including ruff, mypy, SPDX, and repository-specific validation hooks

AI assistance disclosure

AI assistance was used to draft and implement this change, then the changes and targeted validation were reviewed locally before submission.

Changed files

• tests/entrypoints/llm/test_mm_processor_kwargs.py (added, +173/-0)
• vllm/entrypoints/llm.py (modified, +31/-3)