[Perf] Optimize VoxCPM2 first-request latency via startup warmup (#3424)

Signed-off-by: Dan250124 <416947747@qq.com>

vllm_omni/entrypoints/openai/api_server.py

@@ -916,6 +916,10 @@ async def omni_init_app_state(
         engine_client, state.openai_serving_models, request_logger=request_logger, model_name=model_name
     )
 
+    # Warm up speech pipeline (CUDA Graph capture, torch.compile) so the first
+    # real user request is fast instead of paying a 100s compilation tax.
+    await state.openai_serving_speech.warmup()
+
     state.openai_serving_audio_generate = OmniOpenAIServingAudioGenerate(
         engine_client, state.openai_serving_models, request_logger=request_logger, model_name=model_name
     )