[Bugfix] Fix Fish Speech and CosyVoice3 online serving - missing is_comprehension and broken model detection (#2058)

Signed-off-by: linyueqian <linyueqian@outlook.com>

vllm_omni/entrypoints/openai/serving_speech.py

@@ -144,7 +144,8 @@ class OmniOpenAIServingSpeech(OpenAIServing, AudioMixin):
         self._tts_stage = self._find_tts_stage()
         self._is_tts = self._tts_stage is not None
         self._is_fish_speech = (
-            self._tts_stage is not None and getattr(self._tts_stage, "model_stage", None) == "fish_speech_slow_ar"
+            self._tts_stage is not None
+            and getattr(getattr(self._tts_stage, "engine_args", None), "model_stage", None) == "fish_speech_slow_ar"
         )
         self._fish_speech_tokenizer = None
 

vllm_omni/model_executor/stage_configs/cosyvoice3.yaml

@@ -5,6 +5,7 @@
 
 stage_args:
   - stage_id: 0
+    is_comprehension: true
     runtime:
       devices: 0
     engine_args:

vllm_omni/model_executor/stage_configs/fish_speech_s2_pro.yaml

@@ -2,6 +2,7 @@ async_chunk: true
 stage_args:
   - stage_id: 0
     stage_type: llm
+    is_comprehension: true
     runtime:
       devices: "0"
       max_batch_size: 16