[Core] Split Omni connector model runner mixin

[natureofnature]

Purpose

Move the 2,300-line OmniConnectorModelRunnerMixin implementation under
distributed/omni_connectors/model_runner while keeping the existing worker
module as the public compatibility facade.

The implementation has two private layers with a single inheritance chain:

• omni_connector_runtime.py: connector initialization/lifecycle, shared
  request state, cleanup, rank routing, and KV transfer delegation
• omni_connector_payload_transport.py: payload caching/building,
  full-payload and async-chunk flows, TP fanout, background connector I/O,
  and scheduler feedback

These flows stay together because they share the same lock, per-request
queues, request-ID mapping, delivery state, and ordering contract.

The runner-facing Mixin import, class, methods, connector-selection helpers,
and mutable instance state remain available without changing model runners.
Connector configuration, request ordering, retries, cleanup, TP fanout, and
KV routing are unchanged.

Also remove the unused payload_span helper modules, their dedicated tests,
and the optional span-aware accumulation branch. Built-in producers do not
populate span bounds. Ordinary tensor/list accumulation, explicit overrides,
endpoint schema fields, and metadata-only payload handling remain intact.
This retires the helper import API and overlap-aware behavior for custom
payloads that explicitly supply spans.

Test Plan

vLLM Version: 0.28.0

vLLM-Omni Commit: b1235938211333d02aa35e87b5f93117ffd0dbde

Upstream Base: c704aeccae192d7f88e9138271682ddad1327a69

Test Result

• 519 passed, 2 skipped: connector/runner, schema, scheduler, engine argument,
  Qwen3 streaming helper, and KV transfer tests. Includes new accumulator
  concatenation/snapshot and override regression tests.
• Qwen3-Omni full-payload and prefix-cache E2E: 2 passed.
• Bagel image-to-image AR-to-DiT/KV E2E: 1 passed.
• Qwen3-Omni strict realtime async-chunk E2E: 1 passed; returned text and
  audio ASR transcription match exactly.
• Twelve relevant non-Mypy lint hooks passed on changed files.
• Changed transport module Mypy: 14 existing errors before and after cleanup,
  with identical normalized signatures; no new errors.
• Two independent Ultra reviewers found no blockers in the cleanup.

Real-weight E2E selectors, each run with python3 -m pytest -s -v --run-level=advanced_model:

tests/e2e/online_serving/test_qwen3_omni.py::test_mix_to_text_audio_001[default]
tests/e2e/online_serving/test_qwen3_omni.py::test_thinker_prefix_caching_audio_output[default]
tests/e2e/online_serving/test_bagel.py::test_bagel_img2img_online
tests/entrypoints/openai_api/test_qwen3_omni_realtime_websocket.py::TestQwen3OmniRealtimeWebSocket::test_streaming_audio_input_pcm_output_async_chunk[async_chunk]

This is representative regression coverage, not a full E2E suite or a
performance benchmark.

Earlier E2E run on 93a72218, before the span cleanup:

[hsliuustc0106]

This PR touches vllm_omni/distributed/, vllm_omni/worker/, tests/worker/, vllm_omni/utils/ (7 files). Based on CODEOWNERS coverage of the changed files, the most-related reviewers appear to be:

@fake0fan @Gaohan123 @princepride

Could one of you take a look when you get a chance? Thanks!

[natureofnature]

@codex review

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-08T10:24:06.431480Z	93a7221	Manual request

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Delightful!

Reviewed commit: 93a7221895

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[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 belong to: docs/design/module/omni_connector.md, docs/design/module/input_output_modality_contracts.md, docs/design/module/diffusion/parallelism.md.

Module owners: @fake0fan @princepride @xuechendi

Routing: @fake0fan via module of the changed files, module named in the PR description, semantic router, CODEOWNERS; @princepride via module of the changed files, module named in the PR description, semantic router; @xuechendi via module of the changed files, module named in the PR description, semantic router

@natureofnature, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

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

[vllm-omni-review-bot]

Omni ReviewBot triage note

Automated triage of commit b12359382113 produced:

• Priority: high. Prompt maintainer attention is suggested.

These are automated triage suggestions only — the final decision belongs to the maintainers.

[hsliuustc0106]

Reviewed 93a7221895d878e1475fefad77298c4002b939cd. I found no blocking correctness or architecture issue. The split is reasonable as an internal refactor: the public worker class and its existing runner/validation consumers remain intact, and payload construction, queues, TP fanout, and delivery state retain one shared runner instance and ordering contract. Placing this under omni_connectors/model_runner is consistent with the existing request-aware adapter layer.

One non-blocking design clarification: these are still coupled pieces of one mixin. At runtime.py:133–137, the runtime expects payload-defined hooks, including the I/O loops and full-payload flush. Initialization and cleanup call back into that subclass. Please document this host contract and the runner-adapter boundary in the class/module documentation, particularly since the move changes CODEOWNERS routing. Typed hooks would help future maintenance; a composition redesign is unnecessary for this extraction.

Static comparison preserved all 86 method names (87 definitions counting the property getter/setter separately), with 86 definitions AST-identical. The sole method change is _load_custom_func dispatching through cls, covered by the added override test. The three gate helpers and tensor-span implementation are unchanged; static parsing and diff checks passed.

Before merging, please update the test evidence to identify the exact tested commit and commands/results: the PR body currently names 603cce7a, while this review is for 93a72218. The visible successful checks cover DCO, pre-commit, wheels, and docs. I did not execute unit or device tests during this review.