fix offload_dir error

utils/inference.py

@@ -45,12 +45,18 @@ def load_tokenizer_and_model(base_model, adapter_model, load_8bit=True):
             )
     else:
         model = LlamaForCausalLM.from_pretrained(
-            base_model, device_map={"": device}, low_cpu_mem_usage=True, torch_dtype=torch.float16
+            base_model, 
+            device_map={"": device}, 
+            low_cpu_mem_usage=True, 
+            torch_dtype=torch.float16,
+            offload_folder="."
         )
         if adapter_model is not None:
             model = PeftModel.from_pretrained(
                 model,
-                adapter_model
+                adapter_model,
+                torch_dtype=torch.float16,
+                offload_folder="."
             )
 
     model.eval()