Update api/ltx_server_refactored.py

api/ltx_server_refactored.py

@@ -172,9 +172,9 @@ class VideoService:
         print(f"[DEBUG] Aplicando política de precisão: {prec}")
         if prec == "float8_e4m3fn":
             self.runtime_autocast_dtype = torch.bfloat16
-            force_promote = os.getenv("LTXV_FORCE_BF16_ON_FP8", "0") == "1"
+            force_promote = True #os.getenv("LTXV_FORCE_BF16_ON_FP8", "0") == "1"
             print(f"[DEBUG] FP8 detectado. force_promote={force_promote}")
-            if force_promote and hasattr(torch, "float8_e4m3fn"):
+            if force_promote: # and hasattr(torch, "float8_e4m3fn"):
                 try:
                     self._promote_fp8_weights_to_bf16(self.pipeline)
                 except Exception as e:
@@ -253,9 +253,6 @@ class VideoService:
         print(f"[DEBUG] Vídeo salvo em: {final_path}")
         return final_path
 
-    # ==============================================================================
-    # --- FUNÇÕES MODULARES COM A LÓGICA DE CHUNKING SIMPLIFICADA ---
-    # ==============================================================================
 
     def prepare_condition_items(
         self, items_list: List, height: int, 
@@ -272,7 +269,11 @@ class VideoService:
             conditioning_items.append(ConditioningItem(tensor, safe_frame, float(weight)))
         return conditioning_items
 
-    def _generate_low(
+
+    # ==============================================================================
+    # --- FUNÇÕES MODULARES COM A LÓGICA DE CHUNKING SIMPLIFICADA ---
+    # ==============================================================================
+    def generate_low(
         self, prompt, negative_prompt, 
         height, width, duration, seed, 
         conditioning_items=None,
@@ -310,114 +311,13 @@ class VideoService:
                 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
-        
         except Exception as e:
             pass
         finally:
             torch.cuda.empty_cache()
             torch.cuda.ipc_collect()
             self.finalize(keep_paths=[])
-            
-    # ==============================================================================
-    # --- FUNÇÕES DE GERAÇÃO ATUALIZADAS E MODULARES ---
-    # ==============================================================================
-     
-    def generate_narrative_low(
-        self, prompt: str, negative_prompt, 
-        height, width, duration,
-        seed, conditions_itens=None,
-        ltx_configs_override: dict = None):
-        
-        print("\n" + "="*80)
-        print("======           INICIANDO GERAÇÃO NARRATIVA EM CHUNKS (LOW-RES)           ======")
-        print("="*80)
-
-        prompt_list = [p.strip() for p in prompt.splitlines() if p.strip()]
-        num_chunks = len(prompt_list)
-        if num_chunks == 0: raise ValueError("O prompt está vazio ou não contém linhas válidas.")
-        
-        latentes_chunk_video = []
-        latentes_chunk_video_list = []
-        latentes_chunk = None
-        eco_latents_condition_overlap = None
-        results_dir = "/app/output"; os.makedirs(results_dir, exist_ok=True)
-        
-        for i, prompt_x in enumerate(prompt_list):
-            print(f"\n--- Gerando Chunk Narrativo {i+1}/{num_chunks}: '{prompt}' ---")
-
-            current_image_conditions = []
-            #if conditions_itens:
-            #    cond_item_original = conditions_itens[0]
-            #    if i == 0:
-            #        current_image_conditions.append(cond_item_original)
-            #    else:
-            #        cond_item_fraco = ConditioningItem(
-            #            media_item=cond_item_original.media_item, media_frame_number=0, conditioning_strength=0.1
-            #        )
-            #        current_image_conditions.append(cond_item_fraco)
-            
-            #if ltx_configs_override is None: ltx_configs_override = {}
-            
-            #eco_current_conditions_list = []
-            #if current_image_conditions: eco_current_conditions_list.extend(current_image_conditions)
-            #if eco_latents_condition_overlap: eco_current_conditions_list.append(eco_latents_condition_overlap)
-            #ltx_configs_override["conditioning_items"] = current_conditions
-            
-
-            video_path, tensor_path, used_seed = self._generate_low(
-                prompt, negative_prompt, height, 
-                width, duration, seed, 
-                conditions_itens, ltx_configs_override,
-            )
-
-            log_tensor_info(tensor_path, f"tensor_path")
-        
-      
-            #eco_latents_overlap = latentes_chunk[:, :, -4:, :, :]
-            #log_tensor_info(eco_latents_overlap, f"eco_latents_overlap")
-        
-            #eco_latents_condition_overlap = ConditioningItem(
-            #    media_item=eco_latents_overlap, media_frame_number=0, conditioning_strength=1.0
-            #)
-
-            #if i > 0:
-                #latentes_chunk_podado = latentes_chunk[:, :, 5:, :, :]
-                #latentes_chunk_video_list.append(latentes_chunk_podado)
-                #log_tensor_info(latentes_chunk_podado, "latentes_chunk_podado")
-            #else:
-            latentes_chunk_video_list.append(tensor_path)
-            log_tensor_info(tensor_path, "tensor_path")
-          
-        #final_latents_cpu = torch.cat(latentes_chunk_video_list, dim=2).cpu()
-        #log_tensor_info(final_latents_cpu, "Tensor de Latentes Final Concatenado (CPU)")
-        
-        #tensor_path = os.path.join(results_dir, f"latents_narrative_{used_seed}.pt")
-        #torch.save(final_latents_cpu, tensor_path)
-        
-        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)
-                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)
-        #    traceback.print_exc()
-        #    return None
-        #finally:
-        #    gc.collect()
-        #    torch.cuda.empty_cache()
-        #    torch.cuda.ipc_collect()
-        #    self.finalize(keep_paths=[])
-
-            
-        
+                 
     # ==============================================================================
     # --- FUNÇÃO #4: ORQUESTRADOR  (Upscaler + texturas hd) ---
     # ==============================================================================