[Test] Add e2e tests for Qwen3-TTS speech endpoint

[linyueqian]

PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.

Purpose

Add e2e tests for Qwen3-TTS /v1/audio/speech endpoint. The existing unit tests used mocks that didn't match real behavior, allowing bugs like #1159 to slip through undetected.

Also add doc for tts.

Test Plan

pytest tests/e2e/online_serving/test_qwen3_tts.py -v -s

Test Result

Verified the test correctly catches the multimodal_output bug:

• Without fix (main branch):
  FAILED tests/e2e/online_serving/test_qwen3_tts.py::TestQwen3TTSCustomVoice::test_speech_english_basic
  Server returns {"error":{"message":"TTS model did not produce audio output."}} instead of WAV audio.

• With fix:
  PASSED - Valid WAV audio returned (87KB, 24000Hz mono)

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft.

BEFORE SUBMITTING, PLEASE READ https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md (anything written below this line will be removed by GitHub Actions)

[hsliuustc0106]

could you please also add the api server endpoint in doc?

[Copilot]

Pull request overview

This PR adds comprehensive e2e tests for the Qwen3-TTS model's /v1/audio/speech endpoint. The tests were created in response to bug #1159, where unit tests with mocks didn't catch real behavior issues. The new tests verify actual model inference without mocks.

Changes:

• Adds e2e tests for Qwen3-TTS CustomVoice and VoiceDesign models
• Tests /v1/audio/speech, /v1/audio/voices, and /v1/models endpoints
• Includes regression test for multimodal_output bug that caused "TTS model did not produce audio output" error

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[linyueqian]

could you please also add the api server endpoint in doc?

OK. I have added that doc.

[congw729]

Please add markers as suggested; more usage could be checked at #577

[linyueqian]

Please add markers as suggested; more usage could be checked at #577

Thank you! I have added them.

[hsliuustc0106]

fix ci please

[linyueqian]

fix ci please

Need to merge #1203 first. The test catches that bug so it's expected to fail without it.

[hsliuustc0106]

fix ci please

Need to merge #1203 first. The test catches that bug so it's expected to fail without it.

merged, fix ci please

[linyueqian]

@hsliuustc0106 Should be fixed now.

[zhumingjue138]

We plan to set the running time (timeout_in_minutes) of daily CI to 10 minutes, Could we preserve only the core code changes in this PR to reduce CI runtime, and move the remaining changes to the nightly CI (.buildkite/test-nightly.yaml)?

[linyueqian]

We plan to set the running time (timeout_in_minutes) of daily CI to 10 minutes, Could we preserve only the core code changes in this PR to reduce CI runtime, and move the remaining changes to the nightly CI (.buildkite/test-nightly.yaml)?

The full test suite only takes ~6 min so it should fit within the 10 min limit. I've reduced the timeout accordingly. Happy to split if you still think it's needed though.

[congw729]

Hi, I noticed that you removed the markers to pass the CI test. The error already fixed after #577 is merged. You can add those markers back and try again : )

[linyueqian]

@congw729 Added back @hardware_test, @pytest.mark.core_model and @pytest.mark.omni on all tests.

@hsliuustc0106 CI is passing now. One thing I ran into: record_audio_generated_frames added in the recently merged #891 crashes when the audio output tensor is 0-dim (scalar), since shape[0] raises IndexError. I wrapped it in try/except so metrics errors don't break speech generation.

[linyueqian]

Also cc #1289 which fixes the same metrics bug. Would be good to merge this test PR so future updates like this get caught in CI.

[Gaohan123]

LGTM. Thanks!