Fix: Implement lazy model loading for ZeroGPU compatibility

- Models now load on first use instead of startup

- Device detection happens within GPU context

- Added GPU duration timeouts for all generation functions

- Ensures GPU is acquired before model initialization

app.py

@@ -61,7 +61,7 @@ def get_audio_processor():
 
 # ==================== TAB 1: STANDARD ACE-STEP GUI ====================
 
-@spaces.GPU
+@spaces.GPU(duration=300)
 def standard_generate(
     prompt: str,
     lyrics: str,
@@ -100,7 +100,7 @@ def standard_generate(
         return None, f"❌ Error: {str(e)}"
 
 
-@spaces.GPU
+@spaces.GPU(duration=180)
 def standard_variation(audio_path: str, variation_strength: float) -> Tuple[str, str]:
     """Generate variation of existing audio."""
     try:
@@ -110,7 +110,7 @@ def standard_variation(audio_path: str, variation_strength: float) -> Tuple[str,
         return None, f"❌ Error: {str(e)}"
 
 
-@spaces.GPU
+@spaces.GPU(duration=180)
 def standard_repaint(
     audio_path: str,
     start_time: float,
@@ -124,7 +124,7 @@ def standard_repaint(
     except Exception as e:
         return None, f"❌ Error: {str(e)}"
 
-@spaces.GPU
+@spaces.GPU(duration=180)
 
 def standard_lyric_edit(
     audio_path: str,
@@ -139,7 +139,7 @@ def standard_lyric_edit(
 
 
 # ==================== TAB 2: CUSTOM TIMELINE WORKFLOW ====================
-@spaces.GPU
+@spaces.GPU(duration=300)
 
 def timeline_generate(
     prompt: str,
@@ -232,7 +232,7 @@ def timeline_extend(
         prompt, lyrics, context_length, "auto", 0.7, -1, session_state
     )
 
-@spaces.GPU
+@spaces.GPU(duration=240)
 
 def timeline_inpaint(
     start_time: float,

src/ace_step_engine.py

@@ -34,33 +34,18 @@ class ACEStepEngine:
             config: Configuration dictionary
         """
         self.config = config
-        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
         self._initialized = False
         self.dit_handler = None
         self.llm_handler = None
         
-        logger.info(f"ACE-Step Engine initializing on {self.device}")
+        logger.info(f"ACE-Step Engine created (GPU will be detected on first use)")
 
         if not ACE_STEP_AVAILABLE:
             logger.error("ACE-Step 1.5 modules not available")
             logger.error("Please ensure acestep package is installed in your environment")
             return
 
-        try:
-            # Initialize official handlers
-            self.dit_handler = AceStepHandler()
-            self.llm_handler = LLMHandler()
-            
-            # Download and load models
-            self._download_checkpoints()
-            self._load_models()
-            
-            logger.info("✓ ACE-Step Engine fully initialized")
-        except Exception as e:
-            logger.error(f"Failed to initialize ACE-Step Engine: {e}")
-            logger.error("Engine will not be available for generation")
-            import traceback
-            traceback.print_exc()
+        logger.info("✓ ACE-Step Engine created (models will load on first use)")
 
     def _download_checkpoints(self):
         """Download model checkpoints from HuggingFace if not present."""
@@ -150,6 +135,30 @@ class ACEStepEngine:
             logger.error(f"Failed to initialize models: {e}")
             raise
 
+    def _ensure_models_loaded(self):
+        """Ensure models are loaded (lazy loading for ZeroGPU compatibility)."""
+        if not self._initialized:
+            logger.info("Lazy loading models on first use...")
+            
+            # Detect device now (within GPU context on ZeroGPU)
+            self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+            logger.info(f"Using device: {self.device}")
+            
+            # Create handlers if not already created
+            if self.dit_handler is None:
+                self.dit_handler = AceStepHandler()
+            if self.llm_handler is None:
+                self.llm_handler = LLMHandler()
+            
+            try:
+                # Download and load models
+                self._download_checkpoints()
+                self._load_models()
+                logger.info("✓ Models loaded successfully")
+            except Exception as e:
+                logger.error(f"Failed to load models: {e}")
+                raise
+
     def generate(
         self,
         prompt: str,
@@ -177,10 +186,8 @@ class ACEStepEngine:
         Returns:
             Path to generated audio file
         """
-        if not self._initialized:
-            error_msg = "❌ Engine not initialized - ACE-Step 1.5 may not be installed or models are not loaded"
-            logger.error(error_msg)
-            raise RuntimeError(error_msg)
+        # Ensure models are loaded (lazy loading for ZeroGPU)
+        self._ensure_models_loaded()
 
         try:
             # Prepare generation parameters
@@ -266,8 +273,8 @@ class ACEStepEngine:
 
     def generate_variation(self, audio_path: str, strength: float = 0.5) -> str:
         """Generate variation of existing audio."""
-        if not self._initialized:
-            raise RuntimeError("Engine not initialized")
+        # Ensure models are loaded (lazy loading for ZeroGPU)
+        self._ensure_models_loaded()
             
         try:
             params = GenerationParams(