Skip to content

Commit 918c21b

Browse files
houseroadfacebook-github-bot
authored andcommitted
Only turn on FastIncrementalDetokenizer when tokenizers >= 0.21.1
Summary: Turn off FastIncrementalDetokenizer if the tokenizers version is < 0.21.1 Reviewed By: zhewenl Differential Revision: D73638322
1 parent 7a0a9da commit 918c21b

File tree

1 file changed

+4
-1
lines changed

1 file changed

+4
-1
lines changed

vllm/v1/engine/detokenizer.py

Lines changed: 4 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,7 +1,9 @@
11
# SPDX-License-Identifier: Apache-2.0
22
from abc import ABC, abstractmethod
3+
from packaging import version
34
from typing import Optional
45

6+
import tokenizers
57
from tokenizers import Tokenizer
68
from tokenizers.decoders import DecodeStream
79
from transformers import PreTrainedTokenizerFast
@@ -43,7 +45,8 @@ def from_new_request(
4345
# No tokenizer => skipping detokenization.
4446
return IncrementalDetokenizer()
4547

46-
if isinstance(tokenizer, PreTrainedTokenizerFast):
48+
if isinstance(tokenizer, PreTrainedTokenizerFast) and \
49+
version.parse(tokenizers.__version__) >= version.parse("0.21.1"):
4750
# Fast tokenizer => use tokenizers library DecodeStream.
4851
return FastIncrementalDetokenizer(tokenizer, request)
4952

0 commit comments

Comments
 (0)