From 34c2996ab8832f3a475dd908d34864648b846a82 Mon Sep 17 00:00:00 2001
From: Zhu Yi Lin <116337067+GDzhu01@users.noreply.github.com>
Date: Mon, 20 Oct 2025 19:53:32 +0800
Subject: [PATCH] [main] v_proj combining transpose and matmul (#3545)

### What this PR does / why we need it?

v_proj combining transpose and matmul

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?

CI passed with new added/existing test.

- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

Signed-off-by: GDzhu1 <809721801@qq.com>
---
 vllm_ascend/attention/mla_v1.py | 13 +++++++------
 1 file changed, 7 insertions(+), 6 deletions(-)

diff --git a/vllm_ascend/attention/mla_v1.py b/vllm_ascend/attention/mla_v1.py
index 2196858..f88e01a 100644
--- a/vllm_ascend/attention/mla_v1.py
+++ b/vllm_ascend/attention/mla_v1.py
@@ -559,12 +559,13 @@ class AscendMLAImpl(MLAAttentionImpl):
         self.speculative_config = vllm_config.speculative_config
 
     def _v_up_proj(self, x):
-        # Convert from (B, N, L) to (N, B, L)
-        x = x.view(-1, self.num_heads, self.kv_lora_rank).transpose(0, 1)
-        # Multiply (N, B, L) x (N, L, V) -> (N, B, V)
-        x = torch.bmm(x, self.W_UV)
-        # Convert from (N, B, V) to (B, N * V)
-        x = x.transpose(0, 1).reshape(-1, self.num_heads * self.v_head_dim)
+        x = x.view(-1, self.num_heads, self.kv_lora_rank)
+        x = torch_npu.npu_transpose_batchmatmul(x,
+                                                self.W_UV,
+                                                perm_x1=[1, 0, 2],
+                                                perm_x2=[0, 1, 2],
+                                                perm_y=[1, 0, 2])
+        x = x.reshape(-1, self.num_heads * self.v_head_dim)
         return x
 
     # Return `ql_nope`, `q_pe`