Fix "Add online serving to Stable Audio Diffusion and introduce v1/audio/generate endpoint"

[ekagra-ranjan]

This adds #1255 again. It was reverted here #1789.

The issue comes from _Unsupported being added to _DiffusionServingModels. It was added because earlier OpenAIServing accessing variables not defined in _DiffusionServingModels class with self.model_config = self.models.model_config which would fail even though its not used later on. Hence, I added _Unsupported so that such assignment dont fail during init but loudly if accessed. This failure during init still fails on vllm v0.16 but is fixed in vllm v0.17

However, there are other checks in vllm-omni that rely on hasattr() to see if _DiffusionServingModels has a variable and avoids using undefined variables. These checks fail because hasattr() calls getattr() and _Unsupported is truthy.

But now the recent vllm code assigns these variables in OpenAIServing using self.engine_client instead of self.models so the _Unsupported way can be safely removed from v0.17. For it to work with v0.16, we still need

self.model_config = self._NullModelConfig()
        self.renderer = None
        self.input_processor = None
        self.io_processor = None

This is the commit which fixes it: eb765ee

The prev failing test pytest -sv tests/e2e/online_serving/test_images_generations_lora.py::test_images_generations_per_request_lora_switching now passes.

Example cmd to run Stable audio online

# online
vllm-omni serve stabilityai/stable-audio-open-1.0 \
    --host 0.0.0.0 \
    --port 8000 \
    --gpu-memory-utilization 0.9 \
    --trust-remote-code \
    --enforce-eager \
    --omni
    
# v1/audio/generate    
curl -X POST http://localhost:8000/v1/audio/generate     -H "Content-Type: application/json"     -d '{
        "input": "The sound of a dog barking",
        "audio_length": 15.0,
        "negative_prompt": "Low quality",
        "guidance_scale": 7.0,
        "num_inference_steps": 100,
        "seed": 42
    }' --output dog.wav

[hsliuustc0106]

fix conflicts, in general it looks good to me

[ekagra-ranjan]

@hsliuustc0106 - the conflicts have been resolved again. Pls have a look at the earliest!

[linyueqian]

Pushed a small follow-up (8091 port to match the Qwen3-TTS example, and registered serving/audio_generate_api.md under the OpenAI-Compatible API nav). The for_diffusion() routing concern from my earlier review is addressed — under is_pure_diffusion the engine_client is the diffusion engine and the OpenAIServing.__init__ attributes it touches are satisfied (same pattern as OmniOpenAIServingVideo.for_diffusion). LGTM.

[ekagra-ranjan]

@linyueqian thank you!
The CI failed on one task with this error: "Job never started (agent lost) (because there was a problem provisioning infrastructure to run the job)"

https://buildkite.com/vllm/vllm-omni/builds/7547/steps/canvas?sid=019db1f8-f1d7-4b2f-a377-fa025bfc8a43&tab=output

Looks harmless?