Update model/openllama.py

model/openllama.py

@@ -221,6 +221,8 @@ class OpenLLAMAPEFTModel(nn.Module):
         # finally:
         #     print(self.llama_model.hf_device_map)
         self.llama_model = get_peft_model(self.llama_model, peft_config)
+        delta_ckpt = torch.load(args['delta_ckpt_path'])
+        self.llama_model.load_state_dict(delta_ckpt, strict=False)
         self.llama_model.print_trainable_parameters()
 
         self.llama_tokenizer = LlamaTokenizer.from_pretrained(vicuna_ckpt_path, use_fast=False, torch_dtype=torch.bfloat16, device_map='auto', offload_folder="offload2", offload_state_dict = True)