MindSpore报错Kernel launch failed, msg: Acl compile and execute failed, op_type_:AvgPool3D

经验分享 Tech Blogs 模型训练-Model Training
1

1 系统环境

硬件环境(Ascend/GPU/CPU): Ascend910
MindSpore版本: mindspore=2.3.0
执行模式（PyNative/ Graph）:不限
Python版本: Python=3.8
操作系统平台: linux

2 报错信息

2.1 问题描述

执行如下脚本时报错

2.2 脚本信息

x = ops.randn((1,216,64,64), dtype=ms.float32)
out = ops.avg_pool2d(x, 2, stride=2, ceil_mode=True)
print(out.shape)

2.3 报错信息

[ERROR] KERNEL (110510, ffff9e802ebe, python) :2024-@9-19-10:20:53.977.250 [mindspore/ccsrc/plugin/device/ascend/kernel/acl/acl_kernel_mod.cc:261] Launch] Kernel launch failed, msg: Acl compile and execute failed, op_type_:AvgPoo0l13D复制

3 根因分析

报错详细信息如下

[ERROR] KERNEL(100,fffe38f9f120,python):2024-10-28-14:49:50.966.185 [mindspore/ccsrc/plugin/device/ascend/kernel/acl/acl_kernel_mod.cc:260] Launch] Kernel launch failed, msg: Acl compile and execute failed, op_type_:AvgPool3D

上面的报错是调用cann的AvgPool3D报错。这该是和cann对接的问题，是mindspore实现的问题。
avg_pool2d的实现如下，最终调用了AvgPool3D算子。而avg_pool3d在Ascend下同样存在问题。在CPU下是正常的。

if not isinstance(input_x, (Tensor, Tensor_)):  
    raise TypeError("For avg_pool2d, the input input_x must be tensor")  

kernel_size = _check_avgpool_2d_kernel_size(kernel_size)  
stride = _check_avgpool_2d_stride(stride)  
padding = _check_avgpool_2d_padding(padding)  
_check_avg_pool2d_type_and_value(ceil_mode, count_include_pad, divisor_override)  
squeeze_op = _get_cache_prim(P.Squeeze)(2)  
avg_pool_op = _get_cache_prim(P.AvgPool3D)(kernel_size=kernel_size,  
                                            strides=stride,  
                                            pad_mode='pad',  
                                            pad=padding,  
                                            ceil_mode=ceil_mode,  
                                            count_include_pad=count_include_pad,  
                                            divisor_override=divisor_override)  
input_x = expand_dims_(input_x, 2)  
input_x = avg_pool_op(input_x)  
input_x = squeeze_op(input_x)  
return input_x

4 解决方案

除了ops的avg_pool2d，nn层也同样有mindspore.nn.AvgPool2d的实现。

注意ceil_mode=True也不能设置。因为底层对应cann只支持输入为NCHW格式。

import mindspore as ms  
from mindspore import ops  
import numpy as np  
pool = ms.nn.AvgPool2d(kernel_size=2, stride=2)  
x = ops.randn((1,216,64,64), dtype=ms.float32)  
output = pool(x)  
print(output.shape)
【MindSpore报错解决地图】常见报错问题和解决方案（持续更新）