feat: HyperCLOVAX-SEED-Omni-8B full pipeline (vision decoder + S2S/T2V E2E)

[KilJaeeun]

Summary

Builds on vllm-project#869 (HyperCLOVAX audio decoder) to complete the full 3-stage HyperCLOVAX-SEED-Omni-8B pipeline with E2E validated Speech-to-Speech and Text-to-Vision.

Commits

1. feat: vision pipeline, thinker model, stage config (길재은)

• diffusion/models/hyperclovax_vision/: HyperCLOVAX vision diffusion pipeline
• model_executor/models/hcx_omni/: HCX Omni thinker model
• model_executor/stage_configs/hcx_omni.yaml: 3-stage pipeline config
• model_executor/stage_input_processors/hyperclovax_seed_omni.py: thinker→decoder token routing
• examples/, tests/: client demo, e2e and unit tests

2. fix: async fan-out topology, serving pipeline, vLLM 0.18.0 compat

• async_omni.py: redesign _process_sequential_results for fan-out topology — Stage-0 independently routes to Stage-1 (vision) AND Stage-2 (audio) via engine_input_source
• serving_chat.py: add _stage0_is_llm guard to prevent GLM-Image path clobbering HCX Omni multimodal inputs
• vLLM 0.18.0 API alignment across entrypoints and workers

3. fix: diffusion IPC, audio/vision decoder E2E fixes

• diffusion/ipc.py, diffusion_engine.py, diffusion_worker.py: IPC stability for multi-stage diffusion
• pipeline_hyperclovax_audio.py: finetuned decoder path, speaker embedding fallback
• registry.py, request.py: HCX Omni diffusion registration

E2E Test Results

Scenario	Input	Output	Result
Speech-to-Speech	440 Hz sine wave (1s, 16kHz WAV)	11.84s / 568KB WAV (BigVGAN, 24kHz)	✅ PASS
Text-to-Vision	text prompt	768×768 PNG (diffusion, 50 steps)	✅ PASS

[KilJaeeun]

Test Plan Results

Verified against a running 3-stage HyperCLOVAX-SEED-Omni-8B server (hcx_3stage.yaml, TP=2 on GPUs 4–5 + decoders on GPUs 6–7).

#	Test	Result
1	Server health (GET /health)	✅ HTTP 200
2	Speech-to-Speech (audio input → audio output)	✅ 568,364 bytes WAV
3	Text-to-Vision (text prompt → image output)	✅ 189,767 bytes PNG
4	Unit tests (pytest tests/unit/, 12 tests)	✅ 12/12 PASS (82% coverage)

S2S

Choice[0]: content='입력 모달리티와 사용자의 의도를 우선 고려하여 답변 형식을 정의합니다...' audio=False
Choice[1]: audio=True  →  568,364 bytes saved to /tmp/s2s_output.wav
[PASS] Speech-to-Speech

T2V

Total choices: 2
Image output: 189,767 bytes → saved to /tmp/t2v_output.png
[PASS] Text-to-Vision

Unit Tests

collected 12 items
tests/unit/model_executor/test_hcx_omni_processing.py ..........  [100%]

Name                                                                    Stmts   Miss  Cover
-------------------------------------------------------------------------------------------
vllm_omni/model_executor/stage_input_processors/hyperclovax_seed_omni   55     10    82%
-------------------------------------------------------------------------------------------
TOTAL                                                                     55     10    82%

12 passed in 0.08s