From 4453c602626c6bce50b376bbb6e803d7b0131a6e Mon Sep 17 00:00:00 2001 From: wangyongjun Date: Mon, 12 Jan 2026 16:46:21 +0800 Subject: [PATCH] [bugfix]limit graph replay sync (#5761) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit ### What this PR does / why we need it? when graph mode is picewise，replay by synchronize will be effect performance, sync almost cost 250us ![123](https://github.com/user-attachments/assets/04d2a1f3-1f57-4dbb-85ce-b250f2ee7ff0) ### Does this PR introduce _any_ user-facing change? only sync when graph mode contain full mode ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/2f4e6548efec402b913ffddc8726230d9311948d --------- Signed-off-by: wangyongjun --- vllm_ascend/compilation/acl_graph.py | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/vllm_ascend/compilation/acl_graph.py b/vllm_ascend/compilation/acl_graph.py index 29ec5793..97bfb03d 100644 --- a/vllm_ascend/compilation/acl_graph.py +++ b/vllm_ascend/compilation/acl_graph.py @@ -192,12 +192,13 @@ class ACLGraphWrapper: f"got {new_input_addresses}") logger.info_once("Replaying aclgraph") - # In async scheduling or multi-threaded (MT) scenarios, it is possible that + # In async scheduling or multi-threaded (MT) scenarios when graph mode is FULL, it is possible that # the CPU's record event (from update_attn_params) for the iteration i completes # before the grph replay of iteration i-1. # To ensure proper ordering, we must call synchronize here before replaying, # so that update_attn_params only executes after the previous graph replay has fully completed. - torch.npu.synchronize() + if self.runtime_mode == CUDAGraphMode.FULL: + torch.npu.synchronize() entry.aclgraph.replay() return entry.output