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 14, 22:35 UTC
[PERF] Use faster way of decode in tokenizer: avoid useless list-to-list conversion (#20000)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@centml.ai>
Vadim Gimpelson committed
Aug 2, 2025 at 12:43 UTC
58eee5f2e05b74eb2cb1a3bbda9c04df4805e4cc
1 file changed
+4
-3
vllm/transformers_utils/tokenizer.py
+4
-3
@@ -50,11 +50,12 @@ def decode_tokens(
50
`skip_special_tokens=None` means to use the backend's default
51
settings.
52
"""
53
+ decode_method = getattr(tokenizer, "_decode", tokenizer.decode)
54
if skip_special_tokens is not None:
54
- return tokenizer.decode(token_ids,
55
- skip_special_tokens=skip_special_tokens)
55
+ return decode_method(token_ids,
56
+ skip_special_tokens=skip_special_tokens)
57
57
- return tokenizer.decode(token_ids)
58
+ return decode_method(token_ids)
59
60
61
def encode_tokens(