[Bugfix][TTS] Fix Audio8 voice clone reference resolution (#7815)

Signed-off-by: liuyihua <i@liuyihua.com>

tests/entrypoints/openai/test_audio8_tts_adapter.py

@@ -13,15 +13,21 @@ model (falling through to "unknown TTS model") instead of failing at startup.
 These tests pin the contract from both ends.
 """
 
+import asyncio
+from dataclasses import dataclass
+
 import pytest
 
+from vllm_omni.entrypoints.openai.protocol.audio import OpenAICreateSpeechRequest
 from vllm_omni.entrypoints.openai.tts_adapters import (
+    SpeechServingContext,
     all_tts_model_types,
     all_tts_stage_keys,
     detect_tts_model_type,
     resolve_adapter,
 )
 from vllm_omni.entrypoints.openai.tts_adapters.audio8_tts import Audio8TTSAdapter
+from vllm_omni.entrypoints.openai.tts_adapters.base import conditioning_cache_salt
 from vllm_omni.model_executor.models.audio8_tts.pipeline import AUDIO8_TTS_PIPELINE
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
@@ -29,6 +35,15 @@ pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 SLOW_AR_STAGE_KEY = "audio8_tts_slow_ar"
 
 
+@dataclass
+class _FakeSpeechServer:
+    _tts_executor: object | None = None
+
+    async def _resolve_ref_audio(self, ref_audio: str) -> tuple[list[float], int, str]:
+        assert ref_audio == "data:audio/wav;base64,UklGRg=="
+        return [0.0, 0.1], 44_100, "cache-key"
+
+
 def _pipeline_stage_keys() -> list[str]:
     return [stage.model_stage for stage in AUDIO8_TTS_PIPELINE.stages]
 
@@ -83,3 +98,24 @@ def test_apply_sampling_overrides_defaults_tts_local_seed_for_fast_ar():
     # No seed at all -> nothing injected (non-deterministic by request contract).
     out_unseeded = adapter.apply_sampling_overrides([SamplingParams(seed=None)], request)
     assert (out_unseeded[0].extra_args or {}).get("tts_local_seed") is None
+
+
+def test_build_accepts_ref_audio_resolver_cache_key(monkeypatch):
+    """The shared resolver returns a cache key in addition to audio and rate."""
+    adapter = Audio8TTSAdapter(SpeechServingContext(server=_FakeSpeechServer()))
+    request = OpenAICreateSpeechRequest(
+        input="Clone this voice.",
+        ref_audio="data:audio/wav;base64,UklGRg==",
+    )
+    monkeypatch.setattr(
+        adapter,
+        "_build_prompt",
+        lambda request, ref_audio_data, has_inline_ref_audio: {"ref_audio_data": ref_audio_data},
+    )
+
+    prepared = asyncio.run(adapter.build(request, [], has_inline_ref_audio=True))
+
+    assert prepared.prompt["ref_audio_data"] == ([0.0, 0.1], 44_100)
+    assert prepared.tts_params == {"ref_audio_cache_key": "cache-key"}
+    assert prepared.prompt["cache_salt"] == conditioning_cache_salt(request, prepared.tts_params)
+    assert prepared.prompt["cache_salt"] != conditioning_cache_salt(request, {})

vllm_omni/entrypoints/openai/tts_adapters/audio8_tts.py

@@ -97,9 +97,11 @@ class Audio8TTSAdapter(ARTTSAdapter):
         del sampling_params_list
         server = self.ctx.server
         ref_audio_data = None
+        tts_params: dict = {}
         if request.ref_audio is not None:
-            wav_list, sample_rate = await server._resolve_ref_audio(request.ref_audio)
+            wav_list, sample_rate, cache_key = await server._resolve_ref_audio(request.ref_audio)
             ref_audio_data = (wav_list, sample_rate)
+            tts_params["ref_audio_cache_key"] = cache_key
         # Prompt building tokenizes and, for voice clone, allocates tensors; keep
         # it off the event loop via the server's single-worker TTS executor.
         build_prompt = make_async(self._build_prompt, executor=server._tts_executor)
@@ -108,7 +110,6 @@ class Audio8TTSAdapter(ARTTSAdapter):
             ref_audio_data=ref_audio_data,
             has_inline_ref_audio=has_inline_ref_audio,
         )
-        tts_params: dict = {}
         prompt["cache_salt"] = conditioning_cache_salt(request, tts_params)
         return PreparedRequest(prompt=prompt, tts_params=tts_params, model_type=self.name)