Update api/ltx_server.py

api/ltx_server.py

@@ -544,6 +544,13 @@ class VideoService:
             output_video_path = os.path.join(temp_dir, f"output_{used_seed}.mp4")
             final_output_path = None
 
+            vae_manager_singleton.attach_pipeline(
+                self.pipeline,
+                device=self.device,
+                autocast_dtype=self.runtime_autocast_dtype
+            )
+
+
             print("[DEBUG] Decodificando bloco de latentes com VAE → tensor de pixels...")
             # Se desejar “desocupar” a GPU antes do decode, pode-se mover p/ CPU e limpar:
             # latents_cpu = latents.detach().to("cpu", non_blocking=True); torch.cuda.empty_cache(); torch.cuda.ipc_collect(); latents = latents_cpu.to(self.device)