Update api/ltx_server.py

api/ltx_server.py

@@ -242,24 +242,6 @@ def log_tensor_info(tensor, name="Tensor"):
     print("------------------------------------------\n")
 
 
-    @torch.no_grad()
-    def _upsample_latents_internal(self, latents: torch.Tensor) -> torch.Tensor:
-        """
-        Lógica extraída diretamente da LTXMultiScalePipeline para upscale de latentes.
-        """
-        if not self.latent_upsampler:
-            raise ValueError("Latent Upsampler não está carregado.")
-
-        # Garante que os modelos estejam no dispositivo correto
-        self.latent_upsampler.to(self.device)
-        self.pipeline.vae.to(self.device)
-        print(f"[DEBUG-UPSAMPLE] Shape de entrada: {tuple(latents.shape)}")
-        latents = un_normalize_latents(latents, self.pipeline.vae, vae_per_channel_normalize=True)
-        upsampled_latents = self.latent_upsampler(latents)
-        upsampled_latents = normalize_latents(upsampled_latents, self.pipeline.vae, vae_per_channel_normalize=True)
-        print(f"[DEBUG-UPSAMPLE] Shape de saída: {tuple(upsampled_latents.shape)}")
-        
-        return upsampled_latents
 
 
 
@@ -453,6 +435,29 @@ class VideoService:
                 pass
         print(f"[DEBUG] FP8→BF16: params_promoted={p_cnt}, buffers_promoted={b_cnt}")
 
+
+
+    @torch.no_grad()
+    def _upsample_latents_internal(self, latents: torch.Tensor) -> torch.Tensor:
+        """
+        Lógica extraída diretamente da LTXMultiScalePipeline para upscale de latentes.
+        """
+        if not self.latent_upsampler:
+            raise ValueError("Latent Upsampler não está carregado.")
+
+        # Garante que os modelos estejam no dispositivo correto
+        self.latent_upsampler.to(self.device)
+        self.pipeline.vae.to(self.device)
+        print(f"[DEBUG-UPSAMPLE] Shape de entrada: {tuple(latents.shape)}")
+        latents = un_normalize_latents(latents, self.pipeline.vae, vae_per_channel_normalize=True)
+        upsampled_latents = self.latent_upsampler(latents)
+        upsampled_latents = normalize_latents(upsampled_latents, self.pipeline.vae, vae_per_channel_normalize=True)
+        print(f"[DEBUG-UPSAMPLE] Shape de saída: {tuple(upsampled_latents.shape)}")
+        
+        return upsampled_latents
+        
+
+    
     def _apply_precision_policy(self):
         prec = str(self.config.get("precision", "")).lower()
         self.runtime_autocast_dtype = torch.float32