Update api/ltx_server.py

api/ltx_server.py

@@ -587,55 +587,66 @@ class VideoService:
                 
                 print("[DEBUG] Multi-escala: Iniciando Passo 1 (geração de latentes base).")
                 
-                first_pass_args = self.config.get("first_pass", {}).copy()
-                first_pass_kwargs = call_kwargs.copy()
-                first_pass_kwargs.update({
-                    "guidance_scale": float(guidance_scale),
-                    "stg_scale": first_pass_args.get("stg_scale"),
-                    "rescaling_scale": first_pass_args.get("rescaling_scale"),
-                    "skip_block_list": first_pass_args.get("skip_block_list"),
-                })
-                schedule = first_pass_args.get("timesteps") or first_pass_args.get("guidance_timesteps")
-                if schedule:
-                    first_pass_kwargs["timesteps"] = schedule
-                    first_pass_kwargs["guidance_timesteps"] = schedule
-                
+                single_pass_kwargs = call_kwargs.copy()
+                first_pass_config = self.config.get("first_pass", {})
+                single_pass_kwargs.update(
+                    {
+                        "guidance_scale": float(guidance_scale),
+                        "stg_scale": first_pass_config.get("stg_scale"),
+                        "rescaling_scale": first_pass_config.get("rescaling_scale"),
+                        "skip_block_list": first_pass_config.get("skip_block_list"),
+                    }
+                )
+                schedule = first_pass_config.get("timesteps") or first_pass_config.get("guidance_timesteps")
+                if mode == "video-to-video":
+                    schedule = [0.7]; print("[INFO] Modo video-to-video (etapa única): timesteps=[0.7]")
+                if isinstance(schedule, (list, tuple)) and len(schedule) > 0:
+                    single_pass_kwargs["timesteps"] = schedule
+                    single_pass_kwargs["guidance_timesteps"] = schedule
+                print(f"[DEBUG] Single-pass: timesteps_len={len(schedule) if schedule else 0}")
+
+
                 # ==================== NOVA LÓGICA DE DIMENSÕES AQUI ====================
                 downscale_factor = self.config.get("downscale_factor", 2)
-                original_height = first_pass_kwargs["height"]
-                original_width = first_pass_kwargs["width"]
+                original_height = single_pass_kwargs["height"]
+                original_width = single_pass_kwargs["width"]
                 divisor = 24
 
                 # Calcula a altura para o primeiro passo, garantindo divisibilidade
                 target_height_p1 = original_height // downscale_factor
-                first_pass_kwargs["height"] = round(target_height_p1 / divisor) * divisor
+                single_pass_kwargs["height"] = round(target_height_p1 / divisor) * divisor
 
                 # Calcula a largura para o primeiro passo, garantindo divisibilidade
                 target_width_p1 = original_width // downscale_factor
-                first_pass_kwargs["width"] = round(target_width_p1 / divisor) * divisor
+                single_pass_kwargs["width"] = round(target_width_p1 / divisor) * divisor
                 
                 # Medida de segurança para evitar dimensões zero
-                if first_pass_kwargs["height"] == 0: first_pass_kwargs["height"] = divisor
-                if first_pass_kwargs["width"] == 0: first_pass_kwargs["width"] = divisor
+                if single_pass_kwargs["height"] == 0: first_pass_kwargs["height"] = divisor
+                if single_pass_kwargs["width"] == 0: first_pass_kwargs["width"] = divisor
                 # =======================================================================
                 
-                print(f"[DEBUG] Passo 1: Dimensões reduzidas e ajustadas para {first_pass_kwargs['height']}x{first_pass_kwargs['width']}")
-
-
-                first_pass_kwargs["output_type"] = "latent"
-
-                print(f"[DEBUG][first_pass_kwargs] {first_pass_kwargs}")
+                print(f"[DEBUG] Passo 1: Dimensões reduzidas e ajustadas para {single_pass_kwargs['height']}x{single_pass_kwargs['width']}")
                 
-                t_p1 = time.perf_counter()
+
+                print("\n[INFO] Executando pipeline promeira etapa...")
+                t_sp = time.perf_counter()
                 ctx = torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype) if self.device == "cuda" else contextlib.nullcontext()
                 with ctx:
-                    first_pass_result = self.pipeline(**first_pass_kwargs)
+                    result = self.pipeline(**single_pass_kwargs)
+                print(f"[DEBUG] single-pass tempo={time.perf_counter()-t_sp:.3f}s")
+
+                if hasattr(result, "latents"):
+                    latents = result.latents
+                elif hasattr(result, "images") and isinstance(result.images, torch.Tensor):
+                    latents = result.images
+                else:
+                    latents = result
+                #print(f"[DEBUG] Latentes (first_pass_kwargs): shape={tuple(latents.shape)}")
                     
-                latents_low_res = first_pass_result.latents if hasattr(first_pass_result, "latents") else first_pass_result
+                latents_low_res = latents
                 print(f"[DEBUG] Passo 1 concluído. Shape dos latentes de baixa resolução: {latents_low_res.shape}")
-                log_tensor_info(latents_low_res, "Latentes (Passo 1)")
                 
-                del first_pass_result, first_pass_kwargs
+                del result, single_pass_kwargs
                 gc.collect()
                 if self.device == "cuda": torch.cuda.empty_cache()