diff --git a/vllm_ascend/attention/mla_v1.py b/vllm_ascend/attention/mla_v1.py
index adb19cce..64f72c65 100644
--- a/vllm_ascend/attention/mla_v1.py
+++ b/vllm_ascend/attention/mla_v1.py
@@ -1579,6 +1579,8 @@ 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]