[Perf] Skip detokenization in online beam search (#46422)
Signed-off-by: Guy Stone <guys@spotify.com> Signed-off-by: Guy Stone <guystone3@gmail.com> Co-authored-by: Claude <noreply@anthropic.com>
Guy Stone committed
Jun 23, 2026 at 18:46 UTC
3cc871aaf1554dfaebce86a6798965ec47e708da
1 file changed
+1
vllm/entrypoints/generate/beam_search/online.py
+1
@@ -61,6 +61,7 @@ class BeamSearchOnlineMixin(ABC):
61
logprobs=logprobs_num,
62
max_tokens=1,
63
temperature=temperature,
64
+ detokenize=False,
65
)
66
all_beams = [
67
BeamSearchSequence(