Update api/ltx_server_refactored.py

api/ltx_server_refactored.py

@@ -480,6 +480,56 @@ class VideoService:
         return video_path, latents_path, used_seed
 
 
+    def __init__(self):
+        """Inicializa o serviço com 4 workers especializados."""
+        t0 = time.perf_counter()
+        print("[INFO] Inicializando VideoService com 4 Workers...")
+        
+        # Configuração para 4 GPUs
+        self.multi_gpu_enabled = GPU_CONFIG["enable_multi_gpu"] and torch.cuda.device_count() >= 4
+        
+        if self.multi_gpu_enabled:
+            self.transformer_devices = [f"cuda:{gpu}" for gpu in GPU_CONFIG["transformer_workers"]]
+            self.vae_devices = [f"cuda:{gpu}" for gpu in GPU_CONFIG["vae_workers"]]
+            self.current_transformer_idx = 0
+            self.current_vae_idx = 0
+            
+            print(f"[INFO] Configuração 4-Workers:")
+            print(f"       Transformer Workers: {self.transformer_devices}")
+            print(f"       VAE Workers: {self.vae_devices}")
+        else:
+            self.device_ltx = self.device_vae = "cuda" if torch.cuda.is_available() else "cpu"
+            print("[INFO] Usando configuração single-GPU")
+            
+        self.config = self._load_config("ltxv-13b-0.9.8-distilled-fp8.yaml")
+        self.pipeline, self.latent_upsampler = self._load_models_from_hub()
+        self._setup_4gpu_workers()
+        
+        self.runtime_autocast_dtype = self._get_precision_dtype()
+        
+        # Configurar VAE managers para todas as GPUs VAE
+        self.vae_managers = []
+        if self.multi_gpu_enabled:
+            for vae_device in self.vae_devices:
+                # Usar o mesmo VAE manager singleton mas configurar para dispositivos diferentes
+                manager = type(vae_manager_singleton)()  # Nova instância
+                manager.attach_pipeline(
+                    self.pipeline,
+                    device=vae_device,
+                    autocast_dtype=self.runtime_autocast_dtype
+                )
+                self.vae_managers.append(manager)
+        else:
+            vae_manager_singleton.attach_pipeline(
+                self.pipeline,
+                device=self.device_vae,
+                autocast_dtype=self.runtime_autocast_dtype
+            )
+        
+        self._tmp_dirs = set()
+        RESULTS_DIR.mkdir(exist_ok=True)
+        print(f"[INFO] VideoService 4-Workers pronto. Tempo: {time.perf_counter()-t0:.2f}s")
+        
 
 
     def _set_generation_environment(self):