diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 237673f10c..87a3846fdb 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -54,7 +54,7 @@ def _wrap_fast_inference(generate, device_type, dtype, model):
         kwargs["pad_token_id"] = kwargs.pop("pad_token_id", model_eos_token_id)
 
         # Autocasted
-        with torch.autocast(device_type = device_type, dtype = dtype):
+        with torch.autocast(device_type = model.device.type, dtype = dtype):
             output = generate(*args, **kwargs)
         pass
         return output