Update api/ltx_server_refactored.py

api/ltx_server_refactored.py

@@ -342,12 +342,23 @@ class VideoService:
         }
         
         # --- NÓ 1.3: CHAMADA AO PIPELINE ---
-        with torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype, enabled=self.device.type == 'cuda'):
-            latents_bruto = self.pipeline(**first_pass_kwargs).images
-            log_tensor_info(latents_bruto, f"Latente Bruto Gerado para: '{prompt[:40]}...'")
+        try: 
+            with torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype, enabled=self.device.type == 'cuda'):
+                latents_bruto = self.pipeline(**first_pass_kwargs).images
+                latents_cpu_bruto = latents_bruto.detach().to("cpu")
+                tensor_path_cpu = os.path.join(results_dir, f"latents_low_res_{used_seed}.pt")
+                torch.save(latents_cpu_bruto, tensor_path_cpu)
+                log_tensor_info(latents_bruto, f"Latente Bruto Gerado para: '{prompt[:40]}...'")
+        
+            print("-" * 20 + " FIM: _generate_single_chunk_low " + "-"*20)
+            return tensor_path_cpu
         
-        print("-" * 20 + " FIM: _generate_single_chunk_low " + "-"*20)
-        return latents_bruto
+        except Exception as e:
+            print("-" * 20 + f" ERRO: _generate_single_chunk_low {e} " + "-"*20)
+        finally:
+            torch.cuda.empty_cache()
+            torch.cuda.ipc_collect()
+            self.finalize(keep_paths=[])
 
     # ==============================================================================
     # --- FUNÇÃO #2: ORQUESTRADOR NARRATIVO (MÚLTIPLOS PROMPTS) ---
@@ -439,16 +450,22 @@ class VideoService:
         final_latents = torch.cat(latentes_chunk_video, dim=2)
         log_tensor_info(final_latents, "Tensor de Latentes Final Concatenado")
         
-        tensor_path = os.path.join(results_dir, f"latents_narrative_{used_seed}.pt")
-        torch.save(final_latents.cpu(), tensor_path)
-        
-        with torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype, enabled=self.device.type == 'cuda'):
-            pixel_tensor = vae_manager_singleton.decode(final_latents, decode_timestep=float(ltx_configs_override.get("decode_timestep", 0.05) if ltx_configs_override else 0.05))
-            video_path = self._save_and_log_video(pixel_tensor, "narrative_video", FPS, temp_dir, results_dir, used_seed)
+        try: 
+            with torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype, enabled=self.device.type == 'cuda'):
+                pixel_tensor = vae_manager_singleton.decode(final_latents.clone(), decode_timestep=float(self.config.get("decode_timestep", 0.05)))
+                video_path = self._save_and_log_video(pixel_tensor, "narrative_video", FPS, temp_dir, results_dir, used_seed)
+                latents_cpu = latents.detach().to("cpu")
+                tensor_path = os.path.join(results_dir, f"latents_low_res_{used_seed}.pt")
+                torch.save(latents_cpu, tensor_path)
+            return video_path, tensor_path, used_seed
         
-        self.finalize(keep_paths=[video_path, tensor_path])
-        return video_path, tensor_path, used_seed
-
+        except Exception as e:
+            pass
+        finally:
+            torch.cuda.empty_cache()
+            torch.cuda.ipc_collect()
+            self.finalize(keep_paths=[])
+            
     # ==============================================================================
     # --- FUNÇÃO #3: ORQUESTRADOR SIMPLES (PROMPT ÚNICO) ---
     # ==============================================================================
@@ -487,16 +504,24 @@ class VideoService:
         print("\n--- Finalizando Geração Simples: Salvando e decodificando ---")
         log_tensor_info(final_latents, "Tensor de Latentes Final")
 
-        tensor_path = os.path.join(results_dir, f"latents_single_{used_seed}.pt")
-        torch.save(final_latents.cpu(), tensor_path)
-
-        with torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype, enabled=self.device.type == 'cuda'):
-            pixel_tensor = vae_manager_singleton.decode(final_latents, decode_timestep=float(ltx_configs_override.get("decode_timestep", 0.05) if ltx_configs_override else 0.05))
-            video_path = self._save_and_log_video(pixel_tensor, "single_video", FPS, temp_dir, results_dir, used_seed)
-
-        self.finalize(keep_paths=[video_path, tensor_path])
-        return video_path, tensor_path, used_seed
+        try: 
+            with torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype, enabled=self.device.type == 'cuda'):
+                pixel_tensor = vae_manager_singleton.decode(final_latents.clone(), decode_timestep=float(self.config.get("decode_timestep", 0.05)))
+                video_path = self._save_and_log_video(pixel_tensor, "single_video", FPS, temp_dir, results_dir, used_seed)
+                latents_cpu = latents.detach().to("cpu")
+                tensor_path = os.path.join(results_dir, f"latents_single_{used_seed}.pt")
+                torch.save(latents_cpu, tensor_path)
+            return video_path, tensor_path, used_seed
+        
+        except Exception as e:
+            pass
+        finally:
+            torch.cuda.empty_cache()
+            torch.cuda.ipc_collect()
+            self.finalize(keep_paths=[])
+            
 
+        
     # ==============================================================================
     # --- FUNÇÃO #4: ORQUESTRADOR  (Upscaler + texturas hd) ---
     # ==============================================================================