Update api/ltx_server_refactored_complete.py

api/ltx_server_refactored_complete.py

@@ -283,6 +283,14 @@ class VideoService:
         downscaled_height = self._align(int(height_padded * downscale_factor), vae_scale_factor)
         downscaled_width = self._align(int(width_padded * downscale_factor), vae_scale_factor)
 
+        for item in original_conditioning_items:
+            moved_tensor = item.media_item.to(self.main_device, dtype=self.runtime_autocast_dtype)
+            processed_conditioning_items.append(ConditioningItem(
+                media_item=moved_tensor,
+                media_frame_number=item.media_frame_number,
+                conditioning_strength=item.conditioning_strength
+            ))
+
         # 1. Começa com a configuração padrão
         first_pass_config = self.config.get("first_pass", {}).copy()
         
@@ -351,6 +359,39 @@ class VideoService:
         video_path = self._save_and_log_video(pixel_tensor, f"{base_filename}_{seed}")
         return str(video_path), str(final_latents_path), seed
 
+
+
+    @log_function_io
+    def prepare_condition_items(self, items_list: List, height: int, width: int, num_frames: int) -> List[Condition-ingItem]:
+        """
+        [CORRIGIDO] Prepara ConditioningItems, mas mantém os tensores na CPU.
+        O movimento para a GPU será tratado posteriormente.
+        """
+        if not items_list: return []
+        height_padded, width_padded = self._align(height), self._align(width)
+        padding_values = calculate_padding(height, width, height_padded, width_padded)
+        
+        conditioning_items = []
+        for media_item, frame, weight in items_list:
+            if isinstance(media_item, str):
+                # Carrega a imagem e aplica padding, mas mantém na CPU.
+                tensor = load_image_to_tensor_with_resize_and_crop(media_item, height, width)
+                tensor = torch.nn.functional.pad(tensor, padding_values)
+                # O tensor permanece na CPU aqui.
+            elif isinstance(media_item, torch.Tensor):
+                # Se for um tensor (como o de overlap), apenas garante que está na CPU.
+                tensor = media_item.cpu()
+            else:
+                logging.warning(f"Unknown conditioning media type: {type(media_item)}. Skipping.")
+                continue
+            
+            safe_frame = max(0, min(int(frame), num_frames - 1))
+            conditioning_items.append(ConditioningItem(tensor, safe_frame, float(weight)))
+            
+        self._log_conditioning_items(conditioning_items)
+        return conditioning_items
+
+    
     @log_function_io
     def prepare_condition_items(self, items_list: List, height: int, width: int, num_frames: int) -> List[ConditioningItem]:
         """[UNIFIED] Prepares ConditioningItems from a mixed list of file paths and tensors."""