diff --git a/vllm_omni/benchmarks/metrics/metrics.py b/vllm_omni/benchmarks/metrics/metrics.py
index 593989c84..b68418e17 100644
--- a/vllm_omni/benchmarks/metrics/metrics.py
+++ b/vllm_omni/benchmarks/metrics/metrics.py
@@ -2,10 +2,10 @@ import warnings
 from dataclasses import dataclass
 
 import numpy as np
-from transformers import PreTrainedTokenizerBase
 from vllm.benchmarks.datasets import SampleRequest
 from vllm.benchmarks.lib.endpoint_request_func import RequestFuncOutput
 from vllm.benchmarks.serve import MILLISECONDS_TO_SECONDS_CONVERSION, TERM_PLOTLIB_AVAILABLE, BenchmarkMetrics, TaskType
+from vllm.tokenizers import TokenizerLike
 
 
 @dataclass
diff --git a/vllm_omni/benchmarks/patch/patch.py b/vllm_omni/benchmarks/patch/patch.py
index 78ddd1f17..78792e81c 100644
--- a/vllm_omni/benchmarks/patch/patch.py
+++ b/vllm_omni/benchmarks/patch/patch.py
@@ -16,7 +16,6 @@ from typing import Literal
 import aiohttp
 from pydub import AudioSegment
 from tqdm.asyncio import tqdm
-from transformers import PreTrainedTokenizerBase
 from vllm.benchmarks import datasets
 from vllm.benchmarks.datasets import SampleRequest
 from vllm.benchmarks.lib.endpoint_request_func import (
@@ -31,6 +30,7 @@ from vllm.benchmarks.lib.endpoint_request_func import (
     _validate_api_url,
 )
 from vllm.logger import init_logger
+from vllm.tokenizers import TokenizerLike
 
 logger = init_logger(__name__)
 from vllm_omni.benchmarks.data_modules.random_multi_modal_dataset import OmniRandomMultiModalDataset