From cd652acb65b5c900069c21f5056a10f75e7de4eb Mon Sep 17 00:00:00 2001 From: zhaozx-cn <59479021+zhaozx-cn@users.noreply.github.com> Date: Thu, 13 Nov 2025 11:29:37 +0800 Subject: [PATCH] [BugFix] Fix kv_no_split not contiguous (#3711) allgather need contiguous data, split operation return uncontiguous data. - vLLM version: v0.11.0rc3 - vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0 Signed-off-by: zhaozx-cn --- vllm_ascend/attention/mla_v1.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/vllm_ascend/attention/mla_v1.py b/vllm_ascend/attention/mla_v1.py index 4044126..c0e175c 100644 --- a/vllm_ascend/attention/mla_v1.py +++ b/vllm_ascend/attention/mla_v1.py @@ -1166,6 +1166,8 @@ class AscendMLAImpl(MLAAttentionImpl): dim=-1, ) q_c = self.q_a_layernorm(q_c) + # allgather need contiguous data + kv_no_split = kv_no_split.contiguous() else: q_c = hidden_states kv_no_split = self.kv_a_proj_with_mqa(hidden_states)[0]