change fp16 to bf16 since no GPU

utils/inference.py

@@ -48,14 +48,14 @@ def load_tokenizer_and_model(base_model, adapter_model, load_8bit=True):
             base_model, 
             device_map={"": device}, 
             low_cpu_mem_usage=True, 
-            torch_dtype=torch.float16,
+            torch_dtype=torch.bfloat16,
             offload_folder="."
         )
         if adapter_model is not None:
             model = PeftModel.from_pretrained(
                 model,
                 adapter_model,
-                torch_dtype=torch.float16,
+                torch_dtype=torch.bfloat16,
                 offload_folder="."
             )