Upstream marge conflicts fix

[knitcapcat-amd]

GitHub shows ~591 files diff because base...head includes the 128 upstream commits carried in by the merge. The actual conflict-resolution is entirely in commit b2331a1d — its "Conflict resolution" view on GitHub shows only 3 files.

Conflicts resolved

• qwen2_5_omni_multiconnector.yaml — dropped (moved to vllm_omni/deploy/ by upstream refactor [Config Refactor 2.5/N] Centralize pipeline registry vllm-project/vllm-omni#2915; mori_connector registration preserved in qwen2_5_omni_mori_intranode.yaml)
• async_omni_engine.py — kept both mori receiver_connectors path and upstream's pd_config detection; pass both to Orchestrator(...)
• orchestrator.py — kept both mori machinery (param / field / PUT logic) and upstream's PD state init + mm_features filtering

Verification

• pre-commit run --all-files passes
• 6×MI300X intra-node Qwen2.5-Omni-7B end-to-end: text + audio output correct; Mori PUT 0->1 confirmed (573KB / 120ms / 4.5 MB/s)

[gemini-code-assist]

Code Review

This pull request refactors the deployment and testing infrastructure by introducing a new 'deploy config' schema and migrating several models (Qwen3-Omni, Qwen3-TTS) to it. It adds support for new models like Ming-flash-omni 2.0, VoxCPM, and VoxCPM2, and removes external audio dependencies (librosa, sox) in favor of internal vLLM utilities. The changes also include new Claude skills for contributors, documentation for features like frame interpolation and prefix caching, and various CI pipeline updates. Reviewers identified critical issues where benchmark scripts could silently pass by falling back to empty templates on failure, and noted several instances in new examples where Omni engine instances were not properly closed or librosa was still being imported despite its removal from the project requirements.

[gemini-code-assist]

Falling back to a zeroed template when the benchmark result file is missing can hide performance regressions or execution failures in CI. Assertions against zeroed metrics (e.g., latency <= threshold) will pass incorrectly. It is better to raise an explicit error here if the benchmark fails to produce data.

    result_path = os.path.join(result_dir, result_filename)
    if not os.path.exists(result_path):
        raise FileNotFoundError(f"Benchmark result file not generated: {result_path}")

    with open(result_path, encoding="utf-8") as f:
        result = json.load(f)

[gemini-code-assist]

Similar to run_benchmark.py, falling back to a template when the result file is missing hides failures. If the benchmark script crashes or fails to generate output, the test should fail explicitly.

Suggested change

	if not tmp_result_file.exists():
	raise FileNotFoundError(f"Benchmark result file not found: {tmp_result_file}")
	with open(DIFFUSION_RESULT_TEMPLATE_PATH, encoding="utf-8") as f:
	template_payload = json.load(f)
	# Template schema is fixed and owned by this repo:
	# ``diffusion_result_template.json`` is a one-item list and metrics live at [0]["result"].
	template_metrics: dict[str, Any] = template_payload[0]["result"]
	with open(tmp_result_file, "w", encoding="utf-8") as f:
	json.dump(template_metrics, f, ensure_ascii=False, indent=2)
	print(f"Benchmark result file not generated, fallback to template: {tmp_result_file}")
	try:
	with open(tmp_result_file, encoding="utf-8") as f:
	metrics: dict[str, Any] = json.load(f)
	if not tmp_result_file.exists():
	raise FileNotFoundError(f"Benchmark result file not generated: {tmp_result_file}")
	try:
	with open(tmp_result_file, encoding="utf-8") as f:
	metrics: dict[str, Any] = json.load(f)

[gemini-code-assist]

Since librosa has been removed from requirements/common.txt in this PR, direct imports and usage of it will fail unless manually installed. Please use vllm.multimodal.media.audio.load_audio instead, which is consistent with other updates in this PR.

Suggested change

	import librosa
	from vllm.multimodal.media.audio import load_audio

[gemini-code-assist]

Replace librosa.load with load_audio to avoid dependency issues.

Suggested change

	audio_signal, sr = librosa.load(audio_path, sr=sampling_rate)
	audio_signal, sr = load_audio(audio_path, sr=sampling_rate)

[gemini-code-assist]

Replace librosa.load with load_audio to avoid dependency issues.

Suggested change

	sig, sr = librosa.load(audio_path, sr=sampling_rate)
	sig, sr = load_audio(audio_path, sr=sampling_rate)

[gemini-code-assist]

A hardcoded 30-second sleep for profiler trace flushing may be excessive or insufficient depending on the environment. Consider making this configurable or using a more robust synchronization mechanism.

[gemini-code-assist]

The Omni instance created in _run_sync is never closed, which can lead to leaked resources (e.g., orphaned stage worker processes). Please wrap the usage in a with block or call omni.close() in a finally block.

Suggested change

	omni = Omni(
	model=args.model,
	stage_configs_path=args.stage_configs_path,
	log_stats=args.log_stats,
	stage_init_timeout=args.stage_init_timeout,
	)
	with Omni(
	model=args.model,
	stage_configs_path=args.stage_configs_path,
	log_stats=args.log_stats,
	stage_init_timeout=args.stage_init_timeout,
	) as omni:

[gemini-code-assist]

The Omni instance (engine) is not closed at the end of main(), which may leak background processes. Please use a context manager.

Suggested change

	engine = Omni(
	model=args.model,
	stage_configs_path=args.stage_configs_path,
	)
	with Omni(
	model=args.model,
	stage_configs_path=args.stage_configs_path,
	) as engine: