Update video_service.py

video_service.py

@@ -311,8 +311,11 @@ class VideoService:
 
         return pipeline, latent_upsampler
 
-    # Precisão: promove FP8->BF16 e define dtype de autocast
+    # Precisão: promove FP8->BF16 e define dtype de autocast (versão segura)
     def _promote_fp8_weights_to_bf16(self, module):
+        # Só promova se for realmente um nn.Module; Pipelines não são nn.Module
+        if not isinstance(module, torch.nn.Module):
+            return
         f8 = getattr(torch, "float8_e4m3fn", None)
         if f8 is None:
             return
@@ -329,24 +332,32 @@ class VideoService:
                     b.data = b.data.to(torch.bfloat16)
             except Exception:
                 pass
-
+    
     def _apply_precision_policy(self):
         prec = str(self.config.get("precision", "")).lower()
         self.runtime_autocast_dtype = torch.float32
         if prec == "float8_e4m3fn":
-            # FP8 experimental: promove pesos para BF16 e padroniza autocast em BF16
-            if hasattr(torch, "float8_e4m3fn"):
-                self._promote_fp8_weights_to_bf16(self.pipeline)
-                if self.latent_upsampler:
-                    self._promote_fp8_weights_to_bf16(self.latent_upsampler)
+            # FP8: kernels nativos da LTX podem estar ativos; por padrão, não promover pesos
             self.runtime_autocast_dtype = torch.bfloat16
+            force_promote = os.getenv("LTXV_FORCE_BF16_ON_FP8", "0") == "1"
+            if force_promote and hasattr(torch, "float8_e4m3fn"):
+                # Promove apenas módulos reais; ignora objetos Pipeline
+                try:
+                    self._promote_fp8_weights_to_bf16(self.pipeline)
+                except Exception:
+                    pass
+                try:
+                    if self.latent_upsampler:
+                        self._promote_fp8_weights_to_bf16(self.latent_upsampler)
+                except Exception:
+                    pass
         elif prec == "bfloat16":
             self.runtime_autocast_dtype = torch.bfloat16
         elif prec == "mixed_precision":
             self.runtime_autocast_dtype = torch.float16
         else:
             self.runtime_autocast_dtype = torch.float32
-
+    
     def _prepare_conditioning_tensor(self, filepath, height, width, padding_values):
         tensor = load_image_to_tensor_with_resize_and_crop(filepath, height, width)
         tensor = torch.nn.functional.pad(tensor, padding_values)