From a813eadd2d2fb5d4f6179fbed860aaebfe2b3db6 Mon Sep 17 00:00:00 2001
From: Shanshan Shen <467638484@qq.com>
Date: Fri, 6 Mar 2026 14:26:37 +0800
Subject: [PATCH] [MM][Perf] Enable 2.7x faster for convolution computation
 with aclnn BatchMatMulV2 (#7017)

### What this PR does / why we need it?
Currently, we are using
https://github.com/vllm-project/vllm/blob/e2b31243c092e9f4ade5ffe4bf9a5d5ddae06ca7/vllm/model_executor/layers/conv.py#L219-L232
for convolution computation, which is used in patch embedding for VL
models.

After profiling, we find that this linear method will take about **6.87
ms**, which is much slower than just using `F.conv3d()`. In
`F.conv3d()`, it will call aclnn `BatchMatMulV2` with optimization on
Ascend NPU, which only take about **2.50 ms** and is **2.7x faster**
than linear method.

- vLLM version: v0.16.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/15d76f74e2fdb12a95ea00f0ca283acf6219a2b7
---------
Signed-off-by: shen-shanshan <467638484@qq.com>
---
 vllm_ascend/ops/conv.py | 32 ++++++++++++++++++++++++++++++++
 vllm_ascend/utils.py    |  3 +++
 2 files changed, 35 insertions(+)
 create mode 100644 vllm_ascend/ops/conv.py

diff --git a/vllm_ascend/ops/conv.py b/vllm_ascend/ops/conv.py
new file mode 100644
index 00000000..08fdb1d1
--- /dev/null
+++ b/vllm_ascend/ops/conv.py
@@ -0,0 +1,32 @@
+#
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# This file is a part of the vllm-ascend project.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+
+import torch
+from vllm.model_executor.layers.conv import Conv2dLayer, Conv3dLayer
+
+
+class AscendConv2dLayer(Conv2dLayer):
+    def forward_oot(self, x: torch.Tensor) -> torch.Tensor:
+        # Use aclnn BatchMatMulV2 for better performance on Ascend NPU.
+        return self._forward_conv(x)
+
+
+class AscendConv3dLayer(Conv3dLayer):
+    def forward_oot(self, x: torch.Tensor) -> torch.Tensor:
+        # Use aclnn BatchMatMulV2 for better performance on Ascend NPU.
+        return self._forward_conv(x)
diff --git a/vllm_ascend/utils.py b/vllm_ascend/utils.py
index da11e35b..1c52627a 100644
--- a/vllm_ascend/utils.py
+++ b/vllm_ascend/utils.py
@@ -597,6 +597,7 @@ def register_ascend_customop(vllm_config: VllmConfig | None = None):
     from vllm.model_executor.custom_op import CustomOp
 
     from vllm_ascend.ops.activation import AscendQuickGELU, AscendSiluAndMul
+    from vllm_ascend.ops.conv import AscendConv2dLayer, AscendConv3dLayer
     from vllm_ascend.ops.fused_moe.fused_moe import AscendFusedMoE, AscendSharedFusedMoE
     from vllm_ascend.ops.layernorm import AscendGemmaRMSNorm, AscendRMSNorm, AscendRMSNormGated
     from vllm_ascend.ops.linear import (
@@ -645,6 +646,8 @@ def register_ascend_customop(vllm_config: VllmConfig | None = None):
         "MMEncoderAttention": AscendMMEncoderAttention,
         "ApplyRotaryEmb": AscendApplyRotaryEmb,
         "RMSNormGated": AscendRMSNormGated,
+        "Conv2dLayer": AscendConv2dLayer,
+        "Conv3dLayer": AscendConv3dLayer,
     }
 
     # 310P: override selected ops with 310P implementations (keep minimal changes outside _310p)