diff --git a/xinference/core/worker.py b/xinference/core/worker.py
index 0e4b25ac3d..f87be367b5 100644
--- a/xinference/core/worker.py
+++ b/xinference/core/worker.py
@@ -1429,6 +1429,10 @@ async def launch_builtin_model(
             if isinstance(n_gpu, str) and n_gpu != "auto":
                 raise ValueError("Currently `n_gpu` only supports `auto`.")
 
+        device = kwargs.get("device")
+        if device and device.lower().startswith("cpu"):
+            n_gpu = None
+
         if peft_model_config is not None:
             if model_type in ("embedding", "rerank"):
                 raise ValueError(