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, 07:30 UTC
Fix lazy import (#30858)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Harry Mellor committed
Dec 17, 2025 at 11:33 UTC
fb980eb2fdd15f81d4c5695347bdea308bb5515e
1 file changed
+6
-4
vllm/v1/structured_output/utils.py
+6
-4
@@ -20,9 +20,9 @@ from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
20
21
if TYPE_CHECKING:
22
import outlines_core as oc
23
+ import transformers.convert_slow_tokenizer as convert_slow_tokenizer
24
import transformers.file_utils as file_utils
25
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
@@ -30,8 +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
- bytes_to_unicode = LazyLoader(
34
- "bytes_to_unicode", globals(), "transformers.convert_slow_tokenizer"
33
+ convert_slow_tokenizer = LazyLoader(
34
+ "convert_slow_tokenizer", globals(), "transformers.convert_slow_tokenizer"
35
)
36
37
TokenizerLike = object
@@ -202,7 +202,9 @@ def _reduced_vocabulary(
202
A Dict of token string -> equivalent token ids
203
"""
204
205
- unicode_to_bytes = {v: k for k, v in bytes_to_unicode().items()}
205
+ unicode_to_bytes = {
206
+ v: k for k, v in convert_slow_tokenizer.bytes_to_unicode().items()
207
+ }
208
209
def convert_token_to_string(token: str) -> str:
210
string = tokenizer.convert_tokens_to_string([token])