diff --git a/vllm_ascend/models/qwen2_5_vl.py b/vllm_ascend/models/qwen2_5_vl.py
index b78bcdf4..2d89d95f 100644
--- a/vllm_ascend/models/qwen2_5_vl.py
+++ b/vllm_ascend/models/qwen2_5_vl.py
@@ -70,6 +70,19 @@ class AscendQwen2_5_VisionAttention(Qwen2_5_VisionAttention):
         if self.hidden_size_per_attention_head > MIN_PAD_SIZE and self.hidden_size_per_attention_head < MAX_PAD_SIZE:
             self.hidden_size_per_attention_head = MAX_PAD_SIZE
 
+    def split_qkv(self, qkv: torch.Tensor) -> tuple[torch.Tensor, ...]:
+        # [s, b, 3 * head * head_dim]
+        seq_len, bs, _ = qkv.shape
+
+        # [s, b, 3 * head * head_dim] -> 3 * [s, b, head * head_dim]
+        q, k, v = qkv.chunk(3, dim=2)
+
+        # 3 * [s, b, head * head_dim] -> 3 * [s, b, head, head_dim]
+        new_shape = (seq_len, bs, self.num_attention_heads_per_partition,
+                     self.hidden_size_per_attention_head)
+        q, k, v = (x.view(*new_shape) for x in (q, k, v))
+        return q, k, v
+
     def forward(
         self,
         x: torch.Tensor,