diff --git a/llm/server/server/data/processor.py b/llm/server/server/data/processor.py
index a470742c6e..423fe6b614 100644
--- a/llm/server/server/data/processor.py
+++ b/llm/server/server/data/processor.py
@@ -16,7 +16,7 @@
 from abc import ABC, abstractmethod
 
 from paddlenlp.transformers import Llama3Tokenizer, LlamaTokenizer
-from paddlenlp.utils.llm_utils import get_eos_token_id
+from paddlenlp.trl.llm_utils import get_eos_token_id
 from server.engine.config import Config
 from server.utils import data_processor_logger
 
@@ -282,7 +282,7 @@ def _load_tokenizer(self):
         """
         if self.config.use_hf_tokenizer:
             from transformers import AutoTokenizer
-            return AutoTokenizer.from_pretrained(self.config.model_dir, use_fast=False)
+            return AutoTokenizer.from_pretrained(self.config.model_dir, use_fast=False, vocab_file=os.path.join(self.config.model_dir, "sentencepiece.bpe.model"))
         else:
             from paddlenlp.transformers import AutoTokenizer
             return AutoTokenizer.from_pretrained(self.config.model_dir)
diff --git a/llm/server/server/engine/infer.py b/llm/server/server/engine/infer.py
index 5d1f9bd33b..655ec060e7 100644
--- a/llm/server/server/engine/infer.py
+++ b/llm/server/server/engine/infer.py
@@ -25,7 +25,7 @@
 import paddle
 import paddle.distributed as dist
 import paddle.distributed.fleet as fleet
-from paddlenlp.utils.llm_utils import get_rotary_position_embedding
+from paddlenlp.trl.llm_utils import get_rotary_position_embedding
 from paddlenlp_ops import step_paddle
 from server.data.processor import DataProcessor
 from server.engine.config import Config