[Bugfix] Stop reading the removed OmniRequestOutput.request_output accessor

[MrlixiangWE]

Purpose

Fixes #6133.

OmniRequestOutput used to hold a nested RequestOutput in a request_output field. #5146 made
it inherit RequestOutput, so the generation content lives on the object itself and the nested
field no longer exists. 21 call sites across 12 files still read it, and three nightly CUDA jobs
fail deterministically:

AttributeError: 'OmniRequestOutput' object has no attribute 'request_output'

This PR routes every one of them at the output object directly (outputs[0].images rather than
outputs[0].request_output.images).

Two things worth flagging that are not in the issue:

The blast radius is wider than the three jobs. examples/offline_inference/image_to_video/README.md:99
sits inside a python snippet that tests/examples/offline_inference/test_image_to_video.py
collects and runs, so the image-to-video Doc Test fails the same way.

Two call sites fail differently, which is why the log-scraped issue did not group them. They
guarded the access with hasattr, so instead of the AttributeError they take a branch that was
never meant to run: tests/diffusion/distributed/test_flux2_vae_patch_parallel.py now always
raises ValueError("No request_output found in OmniRequestOutput"), and
tests/diffusion/models/minimax_h3/test_minimax_h3_quantization_quality.py silently skips its
worker peak-memory update. Both guards are dropped along with the accessor.

One correction to the issue's list: more_cli_examples_003 is the bash NextStep-1.1 snippet.
Neither it, nor text_to_image.py, nor diffusion/utils/image_output.py reads the removed
accessor — #5146 already updated those — so that case is not addressed here and, if it is
still failing, it has another cause.

tests/e2e/accuracy/ltx/run_ltx_reference.py and
tests/diffusion/models/minimax_h3/test_minimax_h3_quantization_quality.py lose a few lines
rather than gaining a rename: their nested-unwrap blocks became unreachable when the nesting went
away, and the code immediately around them already reads the content off the outer object.

Regression test

tests/docs/test_example_output_accessors.py (new, core_model + cpu, following the
tests/docs/test_generate_examples.py idiom) pins both halves of the contract: the field really is
gone from the class, and no doc, example, test or tool reads it. The nested field only fails once a
real model has produced an output, which is why a GPU nightly job was the first thing to notice;
this turns that into a millisecond CPU check. #5146 is part of a series, so the next removal should
not have to wait for nightly either.

Test Plan

# regression test, red on unmodified main and green here
pytest tests/docs/test_example_output_accessors.py -q
pytest tests/docs/ -q

# adjacent output-contract suites, untouched by this PR
pytest tests/engine/test_omni_request_output.py tests/engine/test_output_processor.py -q

# the edited GPU tests still import and collect
pytest --collect-only -q tests/diffusion/distributed/test_vae_decode_parallelism.py \
    tests/diffusion/distributed/test_flux2_vae_patch_parallel.py \
    tests/diffusion/models/minimax_h3/test_minimax_h3_quantization_quality.py
python -m py_compile tests/e2e/accuracy/ltx/run_ltx_reference.py \
    tools/baselines/omni_hash_baseline.py examples/offline_inference/minicpmo/end2end.py

# the four CI-collected snippets, read back through the repo's own extractor
python -c "from tests.examples.helpers import ReadmeSnippet; ..."

pre-commit run --files <the 13 changed files>

vLLM Version: 0.27.0

vLLM-Omni Commit: dbc0dd6 (upstream main)

Test Result

The accessor is gone and the replacement is the right one, against an output built the way
diffusion/output_formatter.py builds it:

final_output_type : image
out.images        : [<PIL.Image.Image image mode=RGB size=8x8 ...>]
out.request_output: AttributeError: 'OmniRequestOutput' object has no attribute 'request_output'

The snippets the Doc Test collects, read back through ReadmeSnippet.extract_readme_snippets:

examples/offline_inference/text_to_image/README.md: 15 snippets collected, 0 still using .request_output
    quick_start_001       : images = outputs[0].images
    more_cli_examples_008 : output.images[0].save(f"{i}.jpg")
    more_cli_examples_009 : output.images[0].save(f"{i}.jpg")
examples/offline_inference/image_to_video/README.md: 17 snippets collected, 0 still using .request_output
    quick_start_001       : frames = outputs[0].images

The new test is red on unmodified main and green here:

# upstream main dbc0dd6d
pytest tests/docs/test_example_output_accessors.py -q
FAILED tests/docs/test_example_output_accessors.py::test_no_source_reads_the_nested_request_output
  ... 21 offenders across 12 files ...
1 failed, 1 passed, 14 warnings in 0.22s

# this branch
pytest tests/docs/ -q
6 passed, 14 warnings in 0.21s

Adjacent suites and collection:

pytest tests/engine/test_omni_request_output.py tests/engine/test_output_processor.py -q
41 passed, 14 warnings in 0.18s

pytest --collect-only -q <the three edited GPU tests>
7 tests collected in 1.48s
py_compile: ltx reference + hash baseline + minicpmo end2end OK

pre-commit on the 13 changed files: all hooks pass (actionlint skipped, no workflow files
touched).

Run on one A800 against upstream main dbc0dd6d, vLLM 0.27.0. The three failing nightly jobs
themselves need the CUDA runners and are covered by the Diffusion X2I Doc Test, X2V Other Accuracy
Test and Distributed Test with H100 jobs.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[vllm-omni-review-bot]

This PR appears to be related to model: minicpm.

Model owners: @y-null

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[congw729]

LGTM