Update hf_model.py

hf_model.py

@@ -13,7 +13,7 @@ class HFModel:
 
         try:
             log_info(f"=== Loading Model: {self.friendly_name} ===")
-            self.model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True, torch_dtype=torch.bfloat16).cuda()
+            self.model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True, torch_dtype=torch.bfloat16) #.cuda()
             self.tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
             log_info(f"=== Model Loaded Successfully: {self.friendly_name} ===")
         except Exception as e: