del device_map

app.py

@@ -69,20 +69,20 @@ def load_lora(lora_path, base_model="decapoda-research/llama-7b-hf"):
     model = LlamaForCausalLM.from_pretrained(
         base_model,
         # load_in_8bit=True,
-        device_map=device_map,
+        # device_map=device_map,
         low_cpu_mem_usage=True,
     )
     print("Loading LoRA...")
     lora = PeftModel.from_pretrained(
         model,
         lora_path,
-        device_map=device_map,
+        # device_map=device_map,
     )
     return lora
 
 
 base_model = "decapoda-research/llama-13b-hf"
-tokenizer = LlamaTokenizer.from_pretrained(base_model, device_map=device_map)
+tokenizer = LlamaTokenizer.from_pretrained(base_model)
 # question = "如果今天是星期五, 那么后天是星期几?"
 model = load_lora(lora_path="facat/alpaca-lora-cn-13b", base_model=base_model)