人工智能 / 华为昇腾 / 嘎嘎山湖

Ascend for PyTorch源码安装与构建实战

@ 14 min read

一次在华为 HiDevLab（CANN 9.0.0 + openEuler 镜像）上从源码编译 torch_npu 的完整实战记录：从源码获取、修改 op_api 日志，到 GCC 11.2.0 与 CMake 3.18.4 的连环踩坑修复，再到 Whl 打包安装与 NPU 功能验证。

环境：华为 HiDevLab 在线 IDE
CANN 9.0.0 + openEuler 镜像（aarch64）
torch_npu v2.7.1-26.0.0 / torch 2.7.0

一、环境与镜像准备

在HiDevLab创建一个CANN9.0.0-OpenEuler的镜像。

注：这个镜像的gcc版本默认只能编译2.7.1分支。

image1.png

二、获取 torch_npu 源码

进入在线IDE后，clone torch_npu代码，checkout到v2.7.1-26.0.0分支

image2.png
image3.png

下载torch_npu源码

# git clone https://gitcode.com/Ascend/pytorch.git -b v2.7.1-26.0.0 --depth 1

三、修改 CloneKernelOpApi.cpp 增加日志

在CloneKernelOpApi.cpp中增加一行日志

# cd workspace/pytorch/torch_npu/csrc/aten/ops/op_api/
# cp CloneKernelOpApi.cpp CloneKernelOpApi.cpp.backup
# vim CloneKernelOpApi.cpp

增加antares-test输出

image4.png

四、编译环境搭建（GCC / CMake）

完成torch_npu编译，并安装

# yum install -y patch libjpeg-turbo-devel dos2unix openblas git
image5.png

安装 GCC 11.2.0

# wget https://repo.huaweicloud.com/gnu/gcc/gcc-11.2.0/gcc-11.2.0.tar.gz
# rm -rf /tmp/*
# yum install bzip2
# tar -zxvf gcc-11.2.0.tar.gz
# cd gcc-11.2.0
# ./contrib/download_prerequisites
image6.png
# ./configure --enable-languages=c,c++ --disable-multilib --with-system-zlib --prefix=/usr/local/gcc11.2.0

通过grep -w processor /proc/cpuinfo|wc -l查看cpu数，示例为15，用户可自行设置相应参数。

实际显示CPU数量为640

# make -j500

报错：

make[3]: Leaving directory '/workspace/pytorch/gcc-11.2.0/host-aarch64-unknown-linux-gnu/gcc'
make[2]: *** [Makefile:4788: all-stage1-gcc] Error 2
make[2]: Leaving directory '/workspace/pytorch/gcc-11.2.0'
make[1]: *** [Makefile:27497: stage1-bubble] Error 2
make[1]: Leaving directory '/workspace/pytorch/gcc-11.2.0'
make: *** [Makefile:1009: all] Error 2

原因是缺少zlib开发组件

# yum install -y zlib-devel
# make -j500

报错：

make[4]: Leaving directory '/workspace/pytorch/gcc-11.2.0/aarch64-unknown-linux-gnu/libsanitizer/sanitizer_common'
make[3]: *** [Makefile:531: all-recursive] Error 1
make[3]: Leaving directory '/workspace/pytorch/gcc-11.2.0/aarch64-unknown-linux-gnu/libsanitizer'
make[2]: *** [Makefile:418: all] Error 2
make[2]: Leaving directory '/workspace/pytorch/gcc-11.2.0/aarch64-unknown-linux-gnu/libsanitizer'
make[1]: *** [Makefile:21270: all-target-libsanitizer] Error 2
make[1]: Leaving directory '/workspace/pytorch/gcc-11.2.0'
make: *** [Makefile:1011: all] Error 2

需要禁用libsanitizer

# cd /workspace/pytorch/gcc-11.2.0
# make distclean
# ./configure --disable-bootstrap --enable-languages=c,c++ --disable-libsanitizer --prefix=/usr/local/gcc-11.2.0
# make -j500

注：本步骤耗时较长

image7.png
# make install
image8.png

配置 GCC 环境变量

export LD_LIBRARY_PATH=/usr/local/gcc-11.2.0/lib64:${LD_LIBRARY_PATH}
export CC=/usr/local/gcc-11.2.0/bin/gcc
export CXX=/usr/local/gcc-11.2.0/bin/g++
export PATH=/usr/local/gcc-11.2.0/bin:${PATH}
image9.png

如果上述操作不生效

# mv /usr/bin/gcc /usr/bin/gcc.bak
# mv /usr/bin/g++ /usr/bin/g++.bak
# mv /usr/bin/c++ /usr/bin/c++.bak
# mv /usr/bin/cc /usr/bin/cc.bak
# ln -s /usr/local/gcc-11.2.0/bin/gcc /usr/bin/gcc
# ln -s /usr/local/gcc-11.2.0/bin/g++ /usr/bin/g++
# ln -s /usr/local/gcc-11.2.0/bin/c++ /usr/bin/c++
# ln -s /usr/local/gcc-11.2.0/bin/gcc /usr/bin/cc

安装 CMake 3.18.4

# wget https://cmake.org/files/v3.18/cmake-3.18.4.tar.gz
# tar -xf cmake-3.18.4.tar.gz
# cd cmake-3.18.4/
# ./configure --prefix=/usr/local/cmake

报错

CMake Error at Utilities/cmcurl/CMakeLists.txt:485 (message):
Could not find OpenSSL.  Install an OpenSSL development package or
configure CMake with -DCMAKE_USE_OPENSSL=OFF to build without OpenSSL.
-- Configuring incomplete, errors occurred!
See also "/workspace/pytorch/cmake-3.18.4/CMakeFiles/CMakeOutput.log".
See also "/workspace/pytorch/cmake-3.18.4/CMakeFiles/CMakeError.log".
---------------------------------------------
Error when bootstrapping CMake:
Problem while running initial CMake

需要安装openssl

# yum install -y openssl-devel
# ./configure --prefix=/usr/local/cmake
make -j500

报错：

259 |   FSCONFIG_SET_PATH_EMPTY = 4,    /* Set parameter, supplying an object by (empty) path */
|   ^~~~~~~~~~~~~~~~~~~~~~~
/usr/include/linux/mount.h:102:9: error: redeclaration of enumerator 'FSCONFIG_SET_FD'
102 |         FSCONFIG_SET_FD         = 5,    /* Set parameter, supplying an object by fd */
|         ^~~~~~~~~~~~~~~
/usr/local/gcc-11.2.0/lib/gcc/aarch64-unknown-linux-gnu/11.2.0/include-fixed/sys/mount.h:261:3: note: previous definition of 'FSCONFIG_SET_FD' with type 'enum fsconfig_command'
261 |   FSCONFIG_SET_FD         = 5,    /* Set parameter, supplying an object by fd */
|   ^~~~~~~~~~~~~~~
/usr/include/linux/mount.h:103:9: error: redeclaration of enumerator 'FSCONFIG_CMD_CREATE'
103 |         FSCONFIG_CMD_CREATE     = 6,    /* Create new or reuse existing superblock */
|         ^~~~~~~~~~~~~~~~~~~
/usr/local/gcc-11.2.0/lib/gcc/aarch64-unknown-linux-gnu/11.2.0/include-fixed/sys/mount.h:263:3: note: previous definition of 'FSCONFIG_CMD_CREATE' with type 'enum fsconfig_command'
263 |   FSCONFIG_CMD_CREATE     = 6,    /* Invoke superblock creation */
|   ^~~~~~~~~~~~~~~~~~~
/usr/include/linux/mount.h:104:9: error: redeclaration of enumerator 'FSCONFIG_CMD_RECONFIGURE'
104 |         FSCONFIG_CMD_RECONFIGURE = 7,   /* Invoke superblock reconfiguration */
|         ^~~~~~~~~~~~~~~~~~~~~~~~
/usr/local/gcc-11.2.0/lib/gcc/aarch64-unknown-linux-gnu/11.2.0/include-fixed/sys/mount.h:265:3: note: previous definition of 'FSCONFIG_CMD_RECONFIGURE' with type 'enum fsconfig_command'
265 |   FSCONFIG_CMD_RECONFIGURE = 7,   /* Invoke superblock reconfiguration */
|   ^~~~~~~~~~~~~~~~~~~~~~~~
In file included from /usr/include/linux/fs.h:19,
from /workspace/pytorch/cmake-3.18.4/Utilities/cmlibarchive/libarchive/archive_read_disk_posix.c:56:
/usr/include/linux/mount.h:131:8: error: redefinition of 'struct mount_attr'
131 | struct mount_attr {
|        ^~~~~~~~~~
In file included from /workspace/pytorch/cmake-3.18.4/Utilities/cmlibarchive/libarchive/archive_read_disk_posix.c:38:
/usr/local/gcc-11.2.0/lib/gcc/aarch64-unknown-linux-gnu/11.2.0/include-fixed/sys/mount.h:219:8: note: originally defined here
219 | struct mount_attr
|        ^~~~~~~~~~
make[2]: *** [Utilities/cmlibarchive/libarchive/CMakeFiles/cmlibarchive.dir/build.make:433: Utilities/cmlibarchive/libarchive/CMakeFiles/cmlibarchive.dir/archive_read_disk_posix.c.o] Error 1
make[1]: *** [CMakeFiles/Makefile2:2162: Utilities/cmlibarchive/libarchive/CMakeFiles/cmlibarchive.dir/all] Error 2
make: *** [Makefile:182: all] Error 2

原因是gcc 11.2.0 安装时，fixincludes 机制会把”当时”系统的 sys/mount.h 复制一份”修复版”放进自己的 include-fixed 目录；

而当前环境的系统内核头文件（/usr/include/linux/mount.h）比 gcc 11 发布时新，多了 FSCONFIG_SET_STRING、struct mount_attr 等新定义。

现在系统默认编译器是 gcc 11，编译任何间接包含 linux/mount.h 的程序（CMake 的 libarchive、之前的 libsanitizer 都是）都会撞上这份旧快照。

修复：把冲突的 include-fixed 文件移走

# cd /usr/local/gcc-11.2.0/lib/gcc/aarch64-unknown-linux-gnu/11.2.0/include-fixed/
# ls sys/

把冲突的 mount.h 备份移走（不要直接删，留个后路）

# mv sys/mount.h sys/mount.h.bak
# make -j500
# make install
image10.png
# ln -s /usr/local/cmake/bin/cmake /usr/bin/cmake
# cmake --version
image11.png

五、编译 torch_npu Whl 包并安装

编译生成 torch_npu Whl 安装包

# bash ci/build.sh --python=3.10

报错：

./ci/build.sh: line 106: python3.10: command not found

先确认pip源

# cd ~/.pip
# vim pip.conf
[global]
#以华为源为例，请根据实际情况进行替换。
index-url = https://mirrors.huaweicloud.com/repository/pypi/simple
trusted-host = mirrors.huaweicloud.com
timeout = 120

最后直接用3.12的Python编译

# python3.12 -m pip install torch==2.7.0
# export MAX_JOBS=100
# bash ./ci/build.sh --python=3.12
# pip install --upgrade torch_npu-2.7.1.post6+git422f6d9-cp312-cp312-linux_aarch64.whl
image12.png

六、功能验证

安装验证

# python3 -c "import torch;import torch_npu; a = torch.randn(3, 4).npu(); print(a + a);"
image13.png

运行tensor.clone()，观察日志输出

# cat > test/test.py << 'EOF'
import torch
import torch_npu
x = torch.randn(2, 2).npu()
y = x.clone()
print(y)
EOF
# python3.12 /workspace/pytorch/test/test.py
image14.png
# TAGS
嘎嘎山湖 人工智能 华为昇腾
Antares

>_ Antares