From 55a4b5ac407637e6782b880b430319a4479521f2 Mon Sep 17 00:00:00 2001 From: zouyida2052 Date: Wed, 22 Oct 2025 21:52:57 +0800 Subject: [PATCH] unify logic between aclgraph and torchair (#3560) ### What this PR does / why we need it? unify logic between aclgraph and torchair for mtp spec decode - vLLM version: v0.11.0rc3 - vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0 --------- Signed-off-by: zouyida2052 --- vllm_ascend/spec_decode/mtp_proposer.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vllm_ascend/spec_decode/mtp_proposer.py b/vllm_ascend/spec_decode/mtp_proposer.py index a9509544..2a11731a 100644 --- a/vllm_ascend/spec_decode/mtp_proposer.py +++ b/vllm_ascend/spec_decode/mtp_proposer.py @@ -502,7 +502,7 @@ class MtpProposer(Proposer): # prepare next mtp inputs # mtp>1: prefill skip or decode skip last loop - if with_prefill and self.torchair_graph_enabled: + if with_prefill: for _ in range(self.num_speculative_tokens - 1): draft_token_ids_list.append(draft_token_ids) if step == self.num_speculative_tokens - 1 or with_prefill: