fix bug of example

gradio_app.py

@@ -64,6 +64,9 @@ def ensure_examples_generated(model_inference_obj) -> List[str]:
         # Generate the example
         print(f">>> Generating example motion: {example['prompt']}")
         try:
+            # Force CPU device for example generation at startup
+            # This is necessary for Hugging Face Zero GPU environment where GPU
+            # is only available inside @spaces.GPU decorated functions
             html_content, fbx_files = model_inference_obj.run_inference(
                 text=example["prompt"],
                 seeds_csv=example["seeds"],
@@ -73,6 +76,7 @@ def ensure_examples_generated(model_inference_obj) -> List[str]:
                 original_text=example["prompt"],
                 output_dir=example_dir,
                 output_filename=example_filename,
+                device="cpu",  # Force CPU for startup example generation
             )
             print(f">>> Example '{example_filename}' generated successfully!")
             generated_examples.append(example_filename)

hymotion/utils/gradio_runtime.py

@@ -300,27 +300,21 @@ class ModelInference:
 
         return _MODEL_CACHE
 
-    def run_inference(self, *args, **kwargs):
+    def run_inference(self, *args, device: str = None, **kwargs):
         """
-        Run DepthAnything3 model inference on images.
+        Run model inference for motion generation.
         Args:
-            target_dir: Directory containing images
-            apply_mask: Whether to apply mask for ambiguous depth classes
-            mask_edges: Whether to mask edges
-            filter_black_bg: Whether to filter black background
-            filter_white_bg: Whether to filter white background
-            process_res_method: Method for resizing input images
-            show_camera: Whether to show camera in 3D view
-            selected_first_frame: Selected first frame filename
-            save_percentage: Percentage of points to save (0-100)
-            infer_gs: Whether to infer 3D Gaussian Splatting
+            device: Device to run inference on. If None, auto-detect.
+                    Use "cpu" to force CPU inference (e.g., when not in @spaces.GPU context).
         Returns:
-            Tuple of (prediction, processed_data)
+            Tuple of (html_content, fbx_files)
         """
         print(f"[{self.__class__.__name__}] Running inference...")
-        # Device check
-        device = "cuda" if torch.cuda.is_available() else "cpu"
+        # Device check - allow explicit override for Zero GPU compatibility
+        if device is None:
+            device = "cuda" if torch.cuda.is_available() else "cpu"
         device = torch.device(device)
+        print(f"[{self.__class__.__name__}] Using device: {device}")
 
         # Initialize model if needed - get model instance (not stored in self)
         model = self.initialize_model(device)