编译报错 could not convert ‘output_sizes’ from ‘SmallVector<[...],8>’ to ‘SmallVector<[...],32>

op_plugin编译错误，错误信息如下
op-plugin/build/pytorch/third_party/op-plugin/op_plugin/ops/base_ops/opapi/SwiGluKernelNpuOpApi.cpp:30:12: error: could not convert ‘output_sizes’ from ‘SmallVector<[...],8>’ to ‘SmallVector<[...],32>’
30 | return output_sizes;
| ^~~~~~~~~~~~
| |
| SmallVector<[...],8>

torch 2.1 1.11都试过，均出现这个错误

已解决，gcc版本需要使用10.2.0

gcc-11.4.0遇到相同的报错，退回10.5.0后可以修复：

apt-get install gcc-10 g++-10
update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 10
update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 10

Can this issue be solved in the implementation of SwiGluKernelNpuOpApi.cpp? @佳威

It is quite complicated for Ubuntu 22.04 users to downgrade to gcc/g++ 10 using conda..

PS. Changing the template parameter to 8 also works.

How to change the template parameter to 8?I can not find this parameter.

对gcc版本有强制要求，能不能在README里写清楚？各种尝试半天，最后翻issue发现是gcc版本问题。

氪服一下

Ascend / op-plugin

内容风险标识

评论 (6)

Ascend / op-plugin .gitee-modal { width: 500px !important; }

内容风险标识