[Perf] Skip detokenization in offline beam search (#50333)
Signed-off-by: samuelkim7 <samuelmwkim@gmail.com>
Myeongwon Kim committed
Aug 11, 2026 at 01:17 UTC
8977ea8895b19023f729db474d0f609e5bbac144
1 file changed
+2
vllm/entrypoints/generate/beam_search/offline.py
+2
@@ -119,6 +119,7 @@ class BeamSearchOfflineMixin(OfflineInferenceMixin):
119
logprobs=2 * beam_width,
120
max_tokens=1,
121
temperature=temperature,
122
+ detokenize=False,
123
skip_clone=True, # Internal beam search, safe to skip clone
124
)
125
instances: list[BeamSearchInstance] = []
@@ -443,6 +444,7 @@ class BeamSearchOfflineMixin(OfflineInferenceMixin):
444
logprobs=base_params.logprobs,
445
max_tokens=1,
446
temperature=base_params.temperature,
447
+ detokenize=False,
448
allowed_token_ids=(
449
allowed_ids
450
if len(allowed_ids) <= _MAX_NUM_ALLOWED_TOKEN_IDS