From 857ddac4e535fd93b0dd56e72916c2870ea2d50c Mon Sep 17 00:00:00 2001 From: Jakub Zakrzewski Date: Wed, 14 Oct 2026 21:01:01 +0200 Subject: [PATCH] [Perf][MRV2] Skip empty draft-token resets on request admission (#61640) Signed-off-by: Jakub Zakrzewski Co-authored-by: OpenAI Codex Co-authored-by: Nick Hill --- vllm/v1/worker/gpu/states.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/vllm/v1/worker/gpu/states.py b/vllm/v1/worker/gpu/states.py index 16260b0ddb..0a8604e866 100644 --- a/vllm/v1/worker/gpu/states.py +++ b/vllm/v1/worker/gpu/states.py @@ -114,7 +114,8 @@ class RequestState: self.num_computed_tokens_np[req_idx] = num_computed_tokens self.num_computed_tokens.stage_write_elem(req_idx, num_computed_tokens) - self.draft_tokens[req_idx].zero_() + if self.num_speculative_steps > 0: + self.draft_tokens[req_idx].zero_() def apply_staged_writes(self) -> None: self.prompt_len.copy_to_uva()