Update api/ltx_server_refactored.py

api/ltx_server_refactored.py

@@ -257,7 +257,10 @@ class VideoService:
     # --- FUNÇÕES MODULARES COM A LÓGICA DE CHUNKING SIMPLIFICADA ---
     # ==============================================================================
 
-    def prepare_condition_items(self, items_list: List, height: int, width: int, num_frames: int):
+    def prepare_condition_items(
+        self, items_list: List, height: int, 
+        width: int, num_frames: int,
+    ):
         if not items_list: return []
         height_padded = ((height - 1) // 8 + 1) * 8
         width_padded = ((width - 1) // 8 + 1) * 8
@@ -269,12 +272,14 @@ class VideoService:
             conditioning_items.append(ConditioningItem(tensor, safe_frame, float(weight)))
         return conditioning_items
 
-    def _generate_low(
+    def generate_low(
         self, prompt, negative_prompt, 
         height, width, duration, seed, 
-        conditions_itens, ltx_configs_override,
+        conditioning_items=None,
+        conditions_itens=None,
+        ltx_configs_override: dict = None,
     ):
-        guidance_scale="4"
+        guidance_scale=4
         used_seed = random.randint(0, 2**32 - 1) if seed is None else int(seed)
         seed_everething(used_seed)
         FPS = 24.0
@@ -292,31 +297,27 @@ class VideoService:
         first_pass_kwargs = {
             "prompt": prompt, "negative_prompt": negative_prompt, "height": downscaled_height, "width": downscaled_width,
             "num_frames": actual_num_frames, "frame_rate": int(FPS), "generator": torch.Generator(device=self.device).manual_seed(used_seed),
-            "output_type": "latent", "conditioning_items": conditions_itens, 
+            "output_type": "latent", "conditioning_items": conditioning_items, 
+            "guidance_scale": float(guidance_scale),
             **(self.config.get("first_pass", {}))
         }
-        if True: 
-            ctx = torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype) if self.device == "cuda" else contextlib.nullcontext()
-            with ctx:
+        try: 
+            with torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype, enabled=self.device == 'cuda'):
                 latents = self.pipeline(**first_pass_kwargs).images
-                pixel_tensor = vae_manager_singleton.decode(latents, decode_timestep=float(self.config.get("decode_timestep", 0.05)))
+                pixel_tensor = vae_manager_singleton.decode(latents.clone(), decode_timestep=float(self.config.get("decode_timestep", 0.05)))
+                video_path = self._save_and_log_video(pixel_tensor, "low_res_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)
-                del latents; gc.collect(); torch.cuda.empty_cache()
-                pixel_tensor_cpu = pixel_tensor.detach().to("cpu")
-                video_path = self._save_and_log_video(pixel_tensor_cpu, "low_res_video", FPS, temp_dir, results_dir, used_seed)
-                del pixel_tensor; gc.collect(); torch.cuda.empty_cache()
-              
-        return latents_cpu
-        #except Exception as e:
-        #    traceback.print_exc()
-        #    return None
-        #finally:
-        #    torch.cuda.empty_cache()
-        #    torch.cuda.ipc_collect()
-        #    self.finalize(keep_paths=[])
-
+            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=[])
+            
     # ==============================================================================
     # --- FUNÇÕES DE GERAÇÃO ATUALIZADAS E MODULARES ---
     # ==============================================================================
@@ -363,7 +364,7 @@ class VideoService:
             #ltx_configs_override["conditioning_items"] = current_conditions
             
 
-            tensor_path =  self._generate_low(
+            video_path, tensor_path, used_seed = self._generate_low(
                 prompt, negative_prompt, height, 
                 width, duration, seed, 
                 conditions_itens, ltx_configs_override,
@@ -393,16 +394,17 @@ class VideoService:
         #tensor_path = os.path.join(results_dir, f"latents_narrative_{used_seed}.pt")
         #torch.save(final_latents_cpu, tensor_path)
         
-        if True:
+        if false:
             ctx = torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype) if self.device == "cuda" else contextlib.nullcontext()
             with ctx:
-                final_latents_video = torch.cat(latentes_chunk_video_list, dim=2)
+                #final_latents_video = torch.cat(latentes_chunk_video_list, dim=2)
                 pixel_tensor = vae_manager_singleton.decode(final_latents_video, 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)
                 tensor_path = os.path.join(results_dir, f"latents_narrative_{used_seed}.pt")
                 torch.save(final_latents_video, tensor_path)
                 del final_latents_video; gc.collect(); torch.cuda.empty_cache()
                 del pixel_tensor; gc.collect(); torch.cuda.empty_cache()
+       
         return video_path, tensor_path, used_seed
         #except Exception as e:
         #    print("-" * 20 + f" ERRO: generate_narrative_low {e} " + "-"*20)