A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 11, 08:45 UTC
Update where `bytes_to_unicode` is imported from (#30771)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Harry Mellor committed
Dec 16, 2025 at 16:05 UTC
e1625498f43b3c3f398e91d715f37ef42d61d8c0
1 file changed
+4
-6
vllm/v1/structured_output/utils.py
+4
-6
@@ -21,8 +21,8 @@ from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
21
if TYPE_CHECKING:
22
import outlines_core as oc
23
import transformers.file_utils as file_utils
24
- import transformers.models.gpt2.tokenization_gpt2 as tokenization_gpt2
24
import xgrammar as xgr
25
+ from transformers.convert_slow_tokenizer import bytes_to_unicode
26
27
from vllm.tokenizers import TokenizerLike
28
from vllm.v1.worker.gpu_input_batch import InputBatch
@@ -30,10 +30,8 @@ else:
30
xgr = LazyLoader("xgr", globals(), "xgrammar")
31
oc = LazyLoader("oc", globals(), "outlines_core")
32
file_utils = LazyLoader("file_utils", globals(), "transformers.file_utils")
33
- tokenization_gpt2 = LazyLoader(
34
- "tokenization_gpt2",
35
- globals(),
36
- "transformers.models.gpt2.tokenization_gpt2",
33
+ bytes_to_unicode = LazyLoader(
34
+ "bytes_to_unicode", globals(), "transformers.convert_slow_tokenizer"
35
)
36
37
TokenizerLike = object
@@ -204,7 +202,7 @@ def _reduced_vocabulary(
202
A Dict of token string -> equivalent token ids
203
"""
204
207
- unicode_to_bytes = {v: k for k, v in tokenization_gpt2.bytes_to_unicode().items()}
205
+ unicode_to_bytes = {v: k for k, v in bytes_to_unicode().items()}
206
207
def convert_token_to_string(token: str) -> str:
208
string = tokenizer.convert_tokens_to_string([token])