diff --git a/backend/loader.py b/backend/loader.py
index 104cd115..de74d2d0 100644
--- a/backend/loader.py
+++ b/backend/loader.py
@@ -379,7 +379,7 @@ def load_huggingface_component(guess, component_name, lib_name, cls_name, repo_p
 
             override_dtype = backend.args.dynamic_args.forge_unet_storage_dtype
             if override_dtype is torch.int8:
-                override_dtype = None
+                override_dtype = torch.bfloat16 if memory_management.should_use_bf16(load_device) else None
                 try_int8 = True
 
             if guess.nunchaku:
diff --git a/backend/operations.py b/backend/operations.py
index 17513ad5..5f5a7072 100644
--- a/backend/operations.py
+++ b/backend/operations.py
@@ -404,7 +404,7 @@ class ForgeOperationsInt8(ForgeOperations):
             self.compute_dtype = torch.bfloat16
             self.lora_patches = []  # List of (down_scaled, up, start, size) set by INT8ModelPatcher
 
-            self.parameters_manual_cast = True
+            self.parameters_manual_cast = current_dtype != self.compute_dtype
 
         def reset_parameters(self):
             return None