Update api/ltx_server.py

api/ltx_server.py

@@ -254,51 +254,7 @@ class LatentConditioningItem:
     media_frame_number: int
     conditioning_strength: float
 
-
-# --- 5. CLASSE PRINCIPAL DO SERVIÇO ---
-class VideoService:
-    def __init__(self):
-        t0 = time.perf_counter()
-        print("[DEBUG] Inicializando VideoService...")
-        self.debug = os.getenv("LTXV_DEBUG", "1") == "1"
-        self.frame_log_every = int(os.getenv("LTXV_FRAME_LOG_EVERY", "8"))
-        self.config = self._load_config()
-        print(f"[DEBUG] Config carregada (precision={self.config.get('precision')}, sampler={self.config.get('sampler')})")
-        self.device = "cuda" if torch.cuda.is_available() else "cpu"
-        print(f"[DEBUG] Device selecionado: {self.device}")
-        self.last_memory_reserved_mb = 0.0
-        self._tmp_dirs = set(); self._tmp_files = set(); self._last_outputs = []
-
-        
-        self.pipeline, self.latent_upsampler = self._load_models()
-        print(f"[DEBUG] Pipeline e Upsampler carregados. Upsampler ativo? {bool(self.latent_upsampler)}")
-
-        self._apply_ltx_pipeline_patches()
-        
-        print(f"[DEBUG] Movendo modelos para {self.device}...")
-        self.pipeline.to(self.device)
-        if self.latent_upsampler:
-            self.latent_upsampler.to(self.device)
-
-        self._apply_precision_policy()
-        print(f"[DEBUG] runtime_autocast_dtype = {getattr(self, 'runtime_autocast_dtype', None)}")
-
-        # Injeta pipeline/vae no manager (impede vae=None)
-        vae_manager_singleton.attach_pipeline(
-            self.pipeline,
-            device=self.device,
-            autocast_dtype=self.runtime_autocast_dtype
-        )
-        print(f"[DEBUG] VAE manager conectado: has_vae={hasattr(self.pipeline, 'vae')} device={self.device}")
-
-        if self.device == "cuda":
-            torch.cuda.empty_cache()
-            self._log_gpu_memory("Após carregar modelos")
-
-        print(f"[DEBUG] VideoService pronto. boot_time={time.perf_counter()-t0:.3f}s")
-
-
-    def _aduc_prepare_conditioning_patch(
+def _aduc_prepare_conditioning_patch(
         self: "LTXVideoPipeline",
         conditioning_items: Optional[List[Union["ConditioningItem", "LatentConditioningItem"]]],
         init_latents: torch.Tensor,
@@ -353,6 +309,52 @@ class VideoService:
         
         return init_latents, init_pixel_coords, init_conditioning_mask, extra_conditioning_num_latents
     
+
+
+# --- 5. CLASSE PRINCIPAL DO SERVIÇO ---
+class VideoService:
+    def __init__(self):
+        t0 = time.perf_counter()
+        print("[DEBUG] Inicializando VideoService...")
+        self.debug = os.getenv("LTXV_DEBUG", "1") == "1"
+        self.frame_log_every = int(os.getenv("LTXV_FRAME_LOG_EVERY", "8"))
+        self.config = self._load_config()
+        print(f"[DEBUG] Config carregada (precision={self.config.get('precision')}, sampler={self.config.get('sampler')})")
+        self.device = "cuda" if torch.cuda.is_available() else "cpu"
+        print(f"[DEBUG] Device selecionado: {self.device}")
+        self.last_memory_reserved_mb = 0.0
+        self._tmp_dirs = set(); self._tmp_files = set(); self._last_outputs = []
+
+        
+        self.pipeline, self.latent_upsampler = self._load_models()
+        print(f"[DEBUG] Pipeline e Upsampler carregados. Upsampler ativo? {bool(self.latent_upsampler)}")
+
+        self._apply_ltx_pipeline_patches()
+        
+        print(f"[DEBUG] Movendo modelos para {self.device}...")
+        self.pipeline.to(self.device)
+        if self.latent_upsampler:
+            self.latent_upsampler.to(self.device)
+
+        self._apply_precision_policy()
+        print(f"[DEBUG] runtime_autocast_dtype = {getattr(self, 'runtime_autocast_dtype', None)}")
+
+        # Injeta pipeline/vae no manager (impede vae=None)
+        vae_manager_singleton.attach_pipeline(
+            self.pipeline,
+            device=self.device,
+            autocast_dtype=self.runtime_autocast_dtype
+        )
+        print(f"[DEBUG] VAE manager conectado: has_vae={hasattr(self.pipeline, 'vae')} device={self.device}")
+
+        if self.device == "cuda":
+            torch.cuda.empty_cache()
+            self._log_gpu_memory("Após carregar modelos")
+
+        print(f"[DEBUG] VideoService pronto. boot_time={time.perf_counter()-t0:.3f}s")
+
+
+        
         
     # Em ltx_server.py
     def _apply_ltx_pipeline_patches(self):