[Bugfix] Sync main into dev/migrate-MR-v2 and fix build errors

[Sy0307]

Background

Build #7230 on dev/migrate-MR-v2 surfaced three regressions, and the branch had also drifted behind main by 17 commits. This PR does both: sync main into the branch and fix the three failures.

Changes

1. Merge origin/main into dev/migrate-MR-v2

Resolves:

• Import conflict in vllm_omni/core/sched/omni_generation_scheduler.py (HEAD added import os, main added from __future__ import annotations — keep both).
• Modify/delete of five vllm_omni/model_executor/stage_configs/*.yaml files. Main migrated them to vllm_omni/deploy/*.yaml under PR [Config Refactor][2/N] Pipeline + Deploy Config Schema #2383's schema refactor. Dev's only change to those files was adding stop_token_ids / detokenize to default_sampling_params; those values are carried over into the new deploy/qwen3_tts.yaml and deploy/qwen3_omni_moe.yaml.

This pull in also resolves CI failure #5 (simple-unit-test: stage_configs property has no setter) — PR #2884 on main already fixed the FakeAsyncOmniClass fixture.

2. [BugFix] Add Qwen2_5Omni to test_init_model_state expected set

Fixes CI failure #3 (modelrunner-v2-unit-test: test_omni_architectures_set_contains_expected). The expected set hardcoded in the test was out of sync with _OMNI_ARCHITECTURES.

3. [BugFix] Fix MTP buffer size mismatch for Omni Talker models

Fixes CI failure #2 (full-moon-omni-star-doc-test-with-h100: size of tensor a (2048) must match size of tensor b (1024)).

OmniModelState allocated its MTP and static inputs_embeds buffers with self.inputs_embeds_size (= hf_text_config.hidden_size = 2048 for Qwen3-Omni Thinker). But Talker stages replace embed_tokens with codec_embedding whose output dim is 1024. Probe the real dim once at init via model.embed_input_ids(dummy) and use that for buffer allocation.

4. [BugFix] Propagate finished_req_ids for already_finished_reqs

Fixes CI failure #4 (qwen3-tts-base-e2e-test-modelrunner-v2: Orchestrator thread crashed / No free indices).

The already_finished_reqs branch in OmniGenerationScheduler.schedule() only removed requests from the running queue but never added them to self.finished_req_ids. So the worker never got the finished signal, never released the corresponding req_state slots, and the next new request hit AssertionError: No free indices in req_states.add_request. Propagate to both self.finished_req_ids and self.finished_req_ids_dict to match upstream _free_request behavior.

Test Plan

• tests/worker_v2/test_init_model_state.py — 5/5 passed
• tests/core/sched/test_generation_scheduler_{finish_condition,restore}.py — 11/11 passed
• tests/worker_v2/ tests/core/sched/ full — 101/101 passed
• Qwen3-TTS 0.6B server + 5 concurrent /v1/audio/speech requests — all OK
• Qwen3-TTS 0.6B server + 3×10 concurrent stress — 30/30 OK, no regressions
• Full CI (wait for build)

CI Failure Mapping

Build #7230 job	Root cause	Fix in this PR
full-moon-diffusion-x2v-star-accuracy-test	pre-existing flaky test	Already skipped by PR #2883 on main (merged in)
full-moon-omni-star-doc-test-with-h100	MTP buffer 2048 vs 1024	Commit 3 (MTP)
modelrunner-v2-unit-test	test expected set stale	Commit 2 (test)
qwen3-tts-base-e2e-test-modelrunner-v2	No free indices race	Commit 4 (scheduler)
simple-unit-test	stage_configs @property setter	Fixed on main by #2884 (merged in)

[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.

[Sy0307]

@tzhouam PTAK and DCO error is due to merge so plz dismiss them.