Update handler.py

handler.py

@@ -14,7 +14,12 @@ class EndpointHandler:
     def __init__(self, path=""):
         model_name = "openbmb/MiniCPM-V-2_6-int4"
         self.tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
-        self.model = AutoModel.from_pretrained(model_name, trust_remote_code=True).eval()
+        self.model = AutoModel.from_pretrained(
+            model_name,
+            trust_remote_code=True,
+            # Explicitly disable 4-bit loading
+            device_map="auto"
+        ).eval()
 
     def __call__(self, data):
         image_input = data.get("image")