Update api/ltx_server.py

api/ltx_server.py

@@ -503,7 +503,7 @@ class VideoService:
             torch.cuda.empty_cache(); torch.cuda.reset_peak_memory_stats()
         self._log_gpu_memory("Início da Geração")
 
-        ctx = torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype) if self.device == "cuda" else contextlib.nullcontext()
+        #ctx = torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype) if self.device == "cuda" else contextlib.nullcontext()
         
         if mode == "image-to-video" and not start_image_filepath:
             raise ValueError("A imagem de início é obrigatória para o modo image-to-video")
@@ -602,32 +602,37 @@ class VideoService:
                 # 
                 first_pass_kwargs = call_kwargs.copy()
                 first_pass_kwargs.update(first_pass_args)
+
+
+                print("[DEBUG] Chamando multi_scale_pipeline...")
+                t_ms = time.perf_counter()
+                ctx = torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype) if self.device == "cuda" else contextlib.nullcontext()
                 
                 print("[DEBUG] Executando FIRST PASS (pipeline base)...")
                 with ctx:
                      result_first = self.pipeline(**first_pass_kwargs)
-            
-                latents_first = result_first.latents if hasattr(result_first, "latents") else result_first
-                print(f"[DEBUG] Latentes FIRST PASS: {tuple(latents_first.shape)}")
-            
+
+
+                if hasattr(result_first, "latents"):
+                    latents = result_first.latents
+                elif hasattr(result_first, "images") and isinstance(result.images, torch.Tensor):
+                    latents = result_first.images
+                else:
+                    latents = result_first
+                print(f"[DEBUG] Latentes (single-pass): shape={tuple(latents.shape)}")
+                latents_first = latents 
+                
                 # --- SECOND PASS ---
                 print("[DEBUG] Executando SECOND PASS (latent_upsampler)...")
                 with ctx:
-                        result_second = self.latent_upsampler(
+                    result = self.latent_upsampler(
                         latents=latents_first,
                         **second_pass_args
                     )
-            
-                latents_final = result_second.latents if hasattr(result_second, "latents") else result_second
-                print(f"[DEBUG] Latentes SECOND PASS: {tuple(latents_final.shape)}")
-            
                 
-                print("[DEBUG] Chamando multi_scale_pipeline...")
-                t_ms = time.perf_counter()
-                ctx = torch.autocast(device_type="cuda", dtype=self.runtime_autocast_dtype) if self.device == "cuda" else contextlib.nullcontext()
-                with ctx:
-                    result = multi_scale_pipeline(**multi_scale_call_kwargs)
-                print(f"[DEBUG] multi_scale_pipeline tempo={time.perf_counter()-t_ms:.3f}s")
+                #with ctx:
+                #    result = multi_scale_pipeline(**multi_scale_call_kwargs)
+                #print(f"[DEBUG] multi_scale_pipeline tempo={time.perf_counter()-t_ms:.3f}s")
 
                 if hasattr(result, "latents"):
                     latents = result.latents