From 8977ea8895b19023f729db474d0f609e5bbac144 Mon Sep 17 00:00:00 2001 From: Myeongwon Kim <65876994+samuelkim7@users.noreply.github.com> Date: Tue, 11 Aug 2026 01:17:38 +0200 Subject: [PATCH] [Perf] Skip detokenization in offline beam search (#50333) Signed-off-by: samuelkim7 --- vllm/entrypoints/generate/beam_search/offline.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/vllm/entrypoints/generate/beam_search/offline.py b/vllm/entrypoints/generate/beam_search/offline.py index e1bceb5d23..7814d4d44e 100644 --- a/vllm/entrypoints/generate/beam_search/offline.py +++ b/vllm/entrypoints/generate/beam_search/offline.py @@ -119,6 +119,7 @@ class BeamSearchOfflineMixin(OfflineInferenceMixin): logprobs=2 * beam_width, max_tokens=1, temperature=temperature, + detokenize=False, skip_clone=True, # Internal beam search, safe to skip clone ) instances: list[BeamSearchInstance] = [] @@ -443,6 +444,7 @@ class BeamSearchOfflineMixin(OfflineInferenceMixin): logprobs=base_params.logprobs, max_tokens=1, temperature=base_params.temperature, + detokenize=False, allowed_token_ids=( allowed_ids if len(allowed_ids) <= _MAX_NUM_ALLOWED_TOKEN_IDS