diff --git a/api/core/model_runtime/model_providers/__base/tokenizers/gpt2_tokenzier.py b/api/core/model_runtime/model_providers/__base/tokenizers/gpt2_tokenzier.py
index 9a5c40addb..7f3c4a61e4 100644
--- a/api/core/model_runtime/model_providers/__base/tokenizers/gpt2_tokenzier.py
+++ b/api/core/model_runtime/model_providers/__base/tokenizers/gpt2_tokenzier.py
@@ -1,8 +1,6 @@
 from threading import Lock
 from typing import Any
 
-import tiktoken
-
 _tokenizer: Any = None
 _lock = Lock()
 
@@ -33,9 +31,17 @@ class GPT2Tokenizer:
             if _tokenizer is None:
                 # Try to use tiktoken to get the tokenizer because it is faster
                 #
-                _tokenizer = tiktoken.get_encoding("gpt2")
-                # base_path = abspath(__file__)
-                # gpt2_tokenizer_path = join(dirname(base_path), "gpt2")
-                # _tokenizer = TransformerGPT2Tokenizer.from_pretrained(gpt2_tokenizer_path)
+                try:
+                    import tiktoken
+
+                    _tokenizer = tiktoken.get_encoding("gpt2")
+                except Exception:
+                    from os.path import abspath, dirname, join
+
+                    from transformers import GPT2Tokenizer as TransformerGPT2Tokenizer  # type: ignore
+
+                    base_path = abspath(__file__)
+                    gpt2_tokenizer_path = join(dirname(base_path), "gpt2")
+                    _tokenizer = TransformerGPT2Tokenizer.from_pretrained(gpt2_tokenizer_path)
 
             return _tokenizer