Migrate audiox examples

[zzehli]

[Feature] Migrate AudioX offline to standard task example + declare model_extras

Purpose

Closes #4540

Migrates AudioX off its bespoke examples onto the standard task example + a
declarative model_extras registration, following the reference migration in
#4225 (BagelPipeline / SenseNovaU1Pipeline).

What changed:

• New vllm_omni/model_extras/audiox.py declaring the request/response param
  contract (AUDIOX_EXTRA_BODY_PARAMS, AUDIOX_EXTRA_OUTPUT_PARAMS).
• Registered "AudioXPipeline" in vllm_omni/model_extras/registry.py, so
  AudioX knobs sent via extra_body (online) / --extra-body (offline) now flow
  through the standard apply_declared_extra_args path into
  sampling_params.extra_args instead of being dropped.
• Folded all 6 tasks (t2a/t2m/v2a/v2m/tv2a/tv2m) into the standard
  examples/offline_inference/text_to_audio/text_to_audio.py via --task +
  --video + --extra-body. Stable Audio Open behavior is preserved (fallback path
  when the model is not registered).
• Simplified the shared flags: collapsed the duplicate duration
  vocabularies onto a single --audio-length (→ audio length for Stable Audio,
  seconds_total for AudioX) and a single --audio-start (→ audio_start_in_s /
  seconds_start). Dropped the 6 AudioX-only pass-through flags
  (--seconds-start, --seconds-total, --sigma-min, --sigma-max,
  --cfg-rescale, --reference-audio); sampler/reference knobs now arrive via
  --extra-body only, matching the helios reference migration ([Feature]Migrate helios example #4569). AudioX runs
  default their output to audiox_output.wav.
• Deleted the bespoke examples/offline_inference/audiox/ dir (folded into
  the standard offline example). Kept examples/online_serving/audiox/ as a
  dedicated online example, but switched its request contract from the old
  top-level extra_args escape hatch to flattened extra_body keys routed
  through apply_declared_extra_args.
• Tests: added CPU registry unit tests in
  tests/model_extras/test_model_extras.py; updated the online e2e test to send
  sigma_* knobs through the now-declared extra_body path.
• Docs: updated recipes/audiox/AudioX.md to document the extra_body knobs
  and the standard example invocations.

The AudioX pipeline itself needs no behavioral change — it already reads every
knob from extra_args (pipeline_audiox.py:824-918).

Test Plan

Offline

Text-to-audio (t2a):

python examples/offline_inference/text_to_audio/text_to_audio.py \
  --model zhangj1an/AudioX --task t2a \
  --prompt "Fireworks burst twice, then a clock ticking." \
  --output /tmp/audiox_t2a.wav

Text-to-music (t2m):

python examples/offline_inference/text_to_audio/text_to_audio.py \
  --model zhangj1an/AudioX --task t2m \
  --prompt "Uplifting ukulele tune for a travel vlog" \
  --output audiox_t2m.wav

Text+video-to-audio (tv2a, requires --video):

python examples/offline_inference/text_to_audio/text_to_audio.py \
  --model zhangj1an/AudioX --task tv2a \
  --prompt "drum beating sound and human talking" \
  --video https://zeyuet.github.io/AudioX/static/samples/V2M/1XeBotOFqHA.mp4 \
  --output audiox_tv2a.wav

Online

Start the server in one session:

bash examples/online_serving/audiox/run_server.sh

Calls the server to generate the audio file us the sme task:

python examples/online_serving/audiox/openai_chat_client.py --task t2m     --prompt "Uplifting ukulele tune for a travel vlog"     --output t2m-online.wav

Test Result

Wasn't able to upload the wav files directly here. Format not supported by gh. Here is the zip file for all 3 results from above:
audio-combined.zip

[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.

[hsliuustc0106]

AudioX migration to standard examples looks correct.

[zzehli]

@princepride Hi ZP, can you take a look at this PR? I made the migration based on your code.

[princepride]

@zzehli Hi, DCO failed

[zzehli]

@princepride One thing about this PR, there is currently no online_serving generic client for text_to_audio. Creating and moving existing model online examples to a new client (audiox and stable audio) would touch more files than the recurrent PR scope. So I only updated the existing audiox online_serving example folder to accomodate extra_body. I can create a subsequent PR to migrate text_to_audio online_serving examples to a single client. Let me know what you think.

[princepride]

@princepride One thing about this PR, there is currently no online_serving generic client for text_to_audio. Creating and moving existing model online examples to a new client (audiox and stable audio) would touch more files than the recurrent PR scope. So I only updated the existing audiox online_serving example folder to accomodate extra_body. I can create a subsequent PR to migrate text_to_audio online_serving examples to a single client. Let me know what you think.

Yeah, you can create a subsequent PR