try remote_code=False and device_map=auto

utils/models.py

@@ -159,16 +159,16 @@ def run_inference(model_name, context, question):
         if "bitnet" in model_name.lower():
             bitnet_model = BitNetForCausalLM.from_pretrained(
                 model_name,
-                device_map="cuda",
+                device_map="auto",
                 torch_dtype=torch.bfloat16,
-                trust_remote_code=True,
+                #trust_remote_code=True,
             )
             pipe = pipeline(
                 "text-generation",
                 model=bitnet_model,
                 tokenizer=tokenizer,
-                device_map="cuda",
-                trust_remote_code=True,
+                device_map="auto",
+                #trust_remote_code=True,
                 torch_dtype=torch.bfloat16,
                 model_kwargs={
                     "attn_implementation": "eager",