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, 17:30 UTC
[Bugfix] Cache added_vocab to avoid per-token overhead (#30351)
Signed-off-by: limingliang <limingliang@stepfun.com> Co-authored-by: limingliang <limingliang@stepfun.com>
Mingliang Li committed
Dec 10, 2025 at 12:05 UTC
d007387aa742c25f60d9b35bc103cbaf753114c8
1 file changed
+4
-2
vllm/tokenizers/deepseekv32.py
+4
-2
@@ -17,6 +17,8 @@ class DeepseekV32Tokenizer(HfTokenizer):
17
self.name_or_path = (
18
tokenizer.name_or_path if hasattr(tokenizer, "name_or_path") else ""
19
)
20
+ self._added_vocab = self.tokenizer.get_added_vocab()
21
+ self._added_vocab_size = len(self._added_vocab)
22
23
@classmethod
24
def from_pretrained(
@@ -98,7 +100,7 @@ class DeepseekV32Tokenizer(HfTokenizer):
100
101
def __len__(self) -> int:
102
# </think> is an added token in DeepseekV32 tokenizer
101
- return self.vocab_size + len(self.get_added_vocab())
103
+ return self.vocab_size + self._added_vocab_size
104
105
def __call__(
106
self,
@@ -120,7 +122,7 @@ class DeepseekV32Tokenizer(HfTokenizer):
122
return self.tokenizer.get_vocab()
123
124
def get_added_vocab(self) -> dict[str, int]:
123
- return self.tokenizer.get_added_vocab()
125
+ return self._added_vocab.copy()
126
127
def encode(
128
self,