Update api/ltx_server_refactored.py

api/ltx_server_refactored.py

@@ -272,71 +272,6 @@ class VideoService:
     # ==============================================================================
     # --- FUNÇÕES DE GERAÇÃO ATUALIZADAS E MODULARES ---
     # ==============================================================================
-
-    def _generate_single_chunk_low(
-        self, prompt, negative_prompt, 
-        height, width, num_frames, 
-        seed, ltx_configs_override=None):
-        """
-        [NÓ DE GERAÇÃO] Gera um ÚNICO chunk de latentes brutos.
-        """
-        print("\n" + "-"*20 + " INÍCIO: _generate_single_chunk_low " + "-"*20)
-        try:
-            height_padded = ((height - 1) // 8 + 1) * 8
-            width_padded = ((width - 1) // 8 + 1) * 8
-            generator = torch.Generator(device=self.device).manual_seed(seed)
-            downscale_factor = self.config.get("downscale_factor", 0.6666666)
-            vae_scale_factor = self.pipeline.vae_scale_factor
-            x_width = int(width_padded * downscale_factor)
-            downscaled_width = x_width - (x_width % vae_scale_factor)
-            x_height = int(height_padded * downscale_factor)
-            downscaled_height = x_height - (x_height % vae_scale_factor)
-    
-            all_conditions = ltx_configs_override.get("conditioning_items", [])
-            pipeline_kwargs = self.config.get("first_pass", {}).copy()
-            if ltx_configs_override:
-                print("[DEBUG] Sobrepondo configurações do LTX com valores da UI...")
-                preset = ltx_configs_override.get("guidance_preset")
-                if preset == "Customizado":
-                    try:
-                        pipeline_kwargs["guidance_scale"] = json.loads(ltx_configs_override["guidance_scale_list"])
-                        pipeline_kwargs["stg_scale"] = json.loads(ltx_configs_override["stg_scale_list"])
-                        pipeline_kwargs["guidance_timesteps"] = json.loads(ltx_configs_override["timesteps_list"])
-                    except Exception as e:
-                        print(f"  > ERRO ao parsear valores customizados: {e}. Usando Padrão.")
-                elif preset == "Agressivo":
-                    pipeline_kwargs["guidance_scale"] = [1, 2, 8, 12, 8, 2, 1]
-                    pipeline_kwargs["stg_scale"] = [0, 0, 5, 6, 5, 3, 2]
-                elif preset == "Suave":
-                    pipeline_kwargs["guidance_scale"] = [1, 1, 4, 5, 4, 1, 1]
-                    pipeline_kwargs["stg_scale"] = [0, 0, 2, 2, 2, 1, 0]
-                
-                pipeline_kwargs["num_inference_steps"] = ltx_configs_override.get("fp_num_inference_steps", pipeline_kwargs.get("num_inference_steps"))
-                pipeline_kwargs["skip_initial_inference_steps"] = ltx_configs_override.get("ship_initial_inference_steps", pipeline_kwargs.get("skip_initial_inference_steps"))
-                pipeline_kwargs["skip_final_inference_steps"] = ltx_configs_override.get("ship_final_inference_steps", pipeline_kwargs.get("skip_final_inference_steps"))
-
-            pipeline_kwargs.update({
-                "prompt": prompt, "negative_prompt": negative_prompt, "height": downscaled_height, "width": downscaled_width,
-                "num_frames": num_frames, "frame_rate": 24, "generator": generator, "output_type": "latent",
-                "conditioning_items": all_conditions if all_conditions else None,
-            })
-            
-            with torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype, enabled=self.device.type == 'cuda'):
-                latents_bruto = self.pipeline(**pipeline_kwargs).images
-                log_tensor_info(latents_bruto, f"Latente Bruto Gerado para: '{prompt[:40]}...'")
-            
-            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)
-            traceback.print_exc()
-            return None
-        finally:
-            torch.cuda.empty_cache()
-            torch.cuda.ipc_collect()
-            self.finalize(keep_paths=[])
-            
      
     def generate_narrative_low(
         self, prompt: str, negative_prompt, 
@@ -379,10 +314,6 @@ class VideoService:
             #if eco_latents_condition_overlap: eco_current_conditions_list.append(eco_latents_condition_overlap)
             #ltx_configs_override["conditioning_items"] = current_conditions
             
-            latentes_chunk = self.generate_low(
-                prompt, negative_prompt, height, width, 
-                duration, seed, eco_current_conditions_list
-            )
 
             video_path, tensor_path, final_seed = video_generation_service.generate_single_low(
                 prompt=prompt_x, negative_prompt=negative_prompt, height=height, width=width, duration=duration,
@@ -420,7 +351,6 @@ class VideoService:
                 final_latents_gpu = final_latents_cpu.to(self.device)
                 pixel_tensor = vae_manager_singleton.decode(final_latents_gpu, 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)
-        
             self.finalize(keep_paths=[video_path, tensor_path])
             return video_path, tensor_path, used_seed
         
@@ -429,11 +359,12 @@ class VideoService:
             traceback.print_exc()
             return None
         finally:
+            gc.collect()
             torch.cuda.empty_cache()
             torch.cuda.ipc_collect()
             self.finalize(keep_paths=[])
-            
-
+        
+        
     def generate_single_low(
         self, prompt: str, negative_prompt, 
         height, width, duration,
@@ -466,27 +397,22 @@ class VideoService:
             print(f"ERRO FATAL: A geração do chunk único falhou. Abortando.")
             self.finalize(keep_paths=[])
             return None, None, None
-
-        final_latents_cpu = final_latents.cpu()
-        log_tensor_info(final_latents_cpu, "Tensor de Latentes Final (CPU)")
-
-        tensor_path = os.path.join(results_dir, f"latents_single_{used_seed}.pt")
-        torch.save(final_latents_cpu, tensor_path)
-
+      
         try:
-            with torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype, enabled=self.device.type == 'cuda'):
-                final_latents_gpu = final_latents_cpu.to(self.device)
-                pixel_tensor = vae_manager_singleton.decode(final_latents_gpu, 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)
-
-            self.finalize(keep_paths=[video_path, tensor_path])
+            pixel_tensor = vae_manager_singleton.decode(final_latents, 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 = final_latents.detach().to("cpu", non_blocking=True)
+            del pixel_tensor; gc.collect(); torch.cuda.empty_cache()
+            del final_latents; gc.collect(); torch.cuda.empty_cache()
+            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:
             print("-" * 20 + f" ERRO: generate_single_low {e} " + "-"*20)
             traceback.print_exc()
             return None
         finally:
+            gc.collect()
             torch.cuda.empty_cache()
             torch.cuda.ipc_collect()
             self.finalize(keep_paths=[])