diff --git a/vllm/v1/structured_output/utils.py b/vllm/v1/structured_output/utils.py
index ae42b33f80f88..cb5ad99cfbdf7 100644
--- a/vllm/v1/structured_output/utils.py
+++ b/vllm/v1/structured_output/utils.py
@@ -21,8 +21,8 @@ from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
 if TYPE_CHECKING:
     import outlines_core as oc
     import transformers.file_utils as file_utils
-    import transformers.models.gpt2.tokenization_gpt2 as tokenization_gpt2
     import xgrammar as xgr
+    from transformers.convert_slow_tokenizer import bytes_to_unicode
 
     from vllm.tokenizers import TokenizerLike
     from vllm.v1.worker.gpu_input_batch import InputBatch
@@ -30,10 +30,8 @@ else:
     xgr = LazyLoader("xgr", globals(), "xgrammar")
     oc = LazyLoader("oc", globals(), "outlines_core")
     file_utils = LazyLoader("file_utils", globals(), "transformers.file_utils")
-    tokenization_gpt2 = LazyLoader(
-        "tokenization_gpt2",
-        globals(),
-        "transformers.models.gpt2.tokenization_gpt2",
+    bytes_to_unicode = LazyLoader(
+        "bytes_to_unicode", globals(), "transformers.convert_slow_tokenizer"
     )
 
     TokenizerLike = object
@@ -204,7 +202,7 @@ def _reduced_vocabulary(
         A Dict of token string -> equivalent token ids
     """
 
-    unicode_to_bytes = {v: k for k, v in tokenization_gpt2.bytes_to_unicode().items()}
+    unicode_to_bytes = {v: k for k, v in bytes_to_unicode().items()}
 
     def convert_token_to_string(token: str) -> str:
         string = tokenizer.convert_tokens_to_string([token])