Performance: Adjust session timeout and GPU memory limit; minim voice pre-caching and improve singleton instance management

api/src/core/model_config.py

@@ -8,7 +8,7 @@ class ONNXCPUConfig(BaseModel):
     
     # Session pooling
     max_instances: int = Field(4, description="Maximum concurrent model instances")
-    instance_timeout: int = Field(300, description="Session timeout in seconds")
+    instance_timeout: int = Field(60, description="Session timeout in seconds")
     
     # Runtime settings
     num_threads: int = Field(8, description="Number of threads for parallel operations")
@@ -27,7 +27,7 @@ class ONNXGPUConfig(ONNXCPUConfig):
     
     # CUDA settings
     device_id: int = Field(0, description="CUDA device ID")
-    gpu_mem_limit: float = Field(0.7, description="Fraction of GPU memory to use")
+    gpu_mem_limit: float = Field(0.5, description="Fraction of GPU memory to use")
     cudnn_conv_algo_search: str = Field("EXHAUSTIVE", description="CuDNN convolution algorithm search")
     
     # Stream management

api/src/inference/model_manager.py

@@ -118,14 +118,8 @@ class ModelManager:
             # Initialize model with warmup voice
             await self.load_model(model_path, warmup_voice, backend_type)
 
-            # Pre-cache common voices in background
-            common_voices = ['af', 'af_bella', 'af_sky', 'af_nicole']
-            for voice_name in common_voices:
-                try:
-                    await voice_manager.load_voice(voice_name, device=backend.device)
-                    logger.debug(f"Pre-cached voice {voice_name}")
-                except Exception as e:
-                    logger.warning(f"Failed to pre-cache voice {voice_name}: {e}")
+            # Only pre-cache default voice to avoid memory bloat
+            logger.info(f"Using {settings.default_voice} as warmup voice")
 
             # Get available voices count
             voices = await voice_manager.list_voices()

api/src/inference/voice_manager.py

@@ -195,8 +195,9 @@ class VoiceManager:
         }
 
 
-# Global singleton instance
+# Global singleton instance and lock
 _manager_instance = None
+_manager_lock = asyncio.Lock()
 
 
 async def get_manager(config: Optional[VoiceConfig] = None) -> VoiceManager:
@@ -209,6 +210,14 @@ async def get_manager(config: Optional[VoiceConfig] = None) -> VoiceManager:
         VoiceManager instance
     """
     global _manager_instance
+    
+    # Fast path - return existing instance
+    if _manager_instance is not None:
+        return _manager_instance
+        
+    # Slow path - create new instance with lock
+    async with _manager_lock:
+        # Double-check pattern
         if _manager_instance is None:
             _manager_instance = VoiceManager(config)
         return _manager_instance

web/app.js

@@ -41,7 +41,7 @@ class KokoroPlayer {
             container: this.elements.waveContainer,
             width: this.elements.waveContainer.clientWidth,
             height: 80,
-            style: '"ios9"',
+            style: 'ios9',
             // color: '#6366f1',    
             speed: 0.02,
             amplitude: 0.7,