由于需要跑一篇论文的代码,需要安装KeOps。这里忍不住再吐槽一次:Python开发环境的版本依赖管理太混乱了,加上各种操作系统,驱动,CUDA版本等等,太浪费生命。所以,如果要写论文,发布open source的package,最好是用大家主流用到的一些库(如PyTorch,numpy,networkx这些),免得让很多人重复实验的时候浪费时间(很多情况下你不重复实验,reviewer会吐槽),请作者和reviewer们都做个人吧!别再浪费大家的时间了。
具体到这个库,我连听都没有听说过,按照其文档进行了安装:http://www.kernel-operations.io/keops/python/installation.html
同时,上面这个页面也指出了,安装好之后要进行测试,可以运行:
import pykeops
pykeops.clean_pykeops() # just in case old build files are still present
pykeops.test_numpy_bindings() # perform the compilation
然而我这里就出错了:
[pyKeOps] Initializing build folder for dtype=float64 and lang=numpy in /home/yu/.cache/pykeops-1.5-cpython-38 ... CMake Error at /usr/share/cmake-3.16/Modules/CMakeTestCUDACompiler.cmake:46 (message):
The CUDA compiler"/usr/local/cuda-10.2/bin/nvcc"
is not able to compile a simple test program.
It fails with the following output:
Change Dir: /home/yu/.cache/pykeops-1.5-cpython-38/build-766be1f3b4/CMakeFiles/CMakeTmp
Run Build Command(s):/usr/bin/make cmTC_14c03/fast && /usr/bin/make -f CMakeFiles/cmTC_14c03.dir/build.make CMakeFiles/cmTC_14c03.dir/build
make[1]: 进入目录“/home/yu/.cache/pykeops-1.5-cpython-38/build-766be1f3b4/CMakeFiles/CMakeTmp”
Building CUDA object CMakeFiles/cmTC_14c03.dir/main.cu.o
/usr/local/cuda-10.2/bin/nvcc -ccbin=/usr/bin/c++ -x cu -c /home/yu/.cache/pykeops-1.5-cpython-38/build-766be1f3b4/CMakeFiles/CMakeTmp/main.cu -o CMakeFiles/cmTC_14c03.dir/main.cu.o
In file included from /usr/local/cuda-10.2/bin/../targets/x86_64-linux/include/cuda_runtime.h:83,
from <command-line>:
/usr/local/cuda-10.2/bin/../targets/x86_64-linux/include/crt/host_config.h:138:2: error: #error -- unsupported GNU version! gcc versions later than 8 are not supported!
138 | #error -- unsupported GNU version! gcc versions later than 8 are not supported!
| ^~~~~
make[1]: *** [CMakeFiles/cmTC_14c03.dir/build.make:66:CMakeFiles/cmTC_14c03.dir/main.cu.o] 错误 1
make[1]: 离开目录“/home/yu/.cache/pykeops-1.5-cpython-38/build-766be1f3b4/CMakeFiles/CMakeTmp”
make: *** [Makefile:121:cmTC_14c03/fast] 错误 2
呵呵呵呵呵,看这里说的“unsupported GNU version! gcc versions later than 8 are not supported!”,感觉似乎是gcc版本的问题啊!但是我用的Ubuntu 20.04里的gcc早已经降级过了。
查看一下其版本:
sudo update-alternatives --config gcc
有 3 个候选项可用于替换 gcc (提供 /usr/bin/gcc)。
选择 路径 优先级 状态
------------------------------------------------------------
* 0 /usr/bin/gcc-7 150 自动模式
1 /usr/bin/gcc-6 50 手动模式
2 /usr/bin/gcc-7 150 手动模式
3 /usr/bin/gcc-9 100 手动模式要维持当前值[*]请按<回车键>,或者键入选择的编号:
完全没有问题啊!呵呵呵。 不过看上面的错误提示,最开始是
The CUDA compiler
"/usr/local/cuda-10.2/bin/nvcc"
is not able to compile a simple test program.
It fails with the following output:
会不会有可能和CUDA的版本有关系?再仔细检查一下,分别运行:
gcc --version
nvcc --version
cmake --version
nvidia-smi
发现我的驱动已经变成了CUDA Version: 11.2。呵呵,我自己从来没改过,应该是Ubuntu自动升级的结果吧。怀疑是CUDA compiler和驱动的版本不一致造成的。还是参考网上的教程:https://blog.csdn.net/ashome123/article/details/105822040/
尝试升级一下CUDA的版本:https://developer.nvidia.com/cuda-toolkit-archive,CUDA升级的过程我也没截图,我没有选择安装驱动,因为用nvidia-smi看到的驱动版本似乎比它提供的还要高些。安装完成之后会说,这是一个不完整的安装:
===========
= Summary =
===========Driver: Not Selected
Toolkit: Installed in /usr/local/cuda-11.2/
Samples: Installed in /home/yu/, but missing recommended librariesPlease make sure that
- PATH includes /usr/local/cuda-11.2/bin
- LD_LIBRARY_PATH includes /usr/local/cuda-11.2/lib64, or, add /usr/local/cuda-11.2/lib64 to /etc/ld.so.conf and run ldconfig as rootTo uninstall the CUDA Toolkit, run cuda-uninstaller in /usr/local/cuda-11.2/bin
***WARNING: Incomplete installation! This installation did not install the CUDA Driver. A driver of version at least 460.00 is required for CUDA 11.2 functionality to work.
To install the driver using this installer, run the following command, replacing <CudaInstaller> with the name of this run file:
sudo <CudaInstaller>.run --silent --driver
感觉这个应该无所谓吧?安装完后再运行:
nvcc --version
发现还是10.2,这个也很好理解,用网友介绍的方法更新一下:https://www.cnblogs.com/laosan007/p/12612446.html,即可。
然后再尝试运行:
import pykeops
pykeops.clean_pykeops() # just in case old build files are still present
pykeops.test_numpy_bindings() # perform the compilation
发现不报CUDA的错了!所以说明前面的报错的确和CUDA Driver和显卡驱动版本不一致有关!
但是还是报了下面的错,呵呵呵呵,这个也能预想到:
CMake 3.18 or higher is required. You are running version 3.16.3
Call Stack (most recent call first):
CMakeLists.txt:29 (include)
不过这个报错挺让人无语的。 所以还是怪自己没有仔细看所有的版本要求:http://www.kernel-operations.io/keops/python/installation.html#requirements
只是简单看了看Gcc,CUDA,PyTorch的版本,正如网友所说:https://blog.csdn.net/weixin_39702483/article/details/112539037
安装pykeops遇到了很多很多坑,对于官网的要求,我的建议是,即使是(optional)也必须遵守,因为他们之所以写上,是因为很多时候他们只在那些环境里测试过(不要问我是怎么发现的TAT)
我就想问问,作为一个python lib,谁给你们的脸让你们要求cmake的版本??
我之前是用apt安装的cmake,感觉这个和最新版差得好多啊!尝试了一下再次运行:
sudo apt install cmake
发现apt中的最新版的确是3.16.3:
cmake 已经是最新版 (3.16.3-1ubuntu1)。
这个也太老了吧,按照网友的介绍:https://www.cnblogs.com/yanqingyang/p/12731855.html
手动安装cmake,但是没想还有报错:
Could NOT find OpenSSL, try to set the path to OpenSSL root folder in the system variable OPENSSL_ROOT_DIR (missing: OPENSSL_CRYPTO_LIBRARY OPENSSL_INCLUDE_DIR)
我去!!!系统中怎么可能没有openssl呢,搜了一下,看到https://www.jianshu.com/p/55249a0272ad,发现是应该输入:
sudo apt-get install libssl-dev
然后再重新安装:
./bootstrap
make
sudo make install
再运行一下:
cmake --version
发现确实更新了:
cmake version 3.21.1
CMake suite maintained and supported by Kitware (kitware.com/cmake).
呵呵呵,然后再尝试一下测试pykeops,发现这次终于不报错了,所以说明至少把这个库安装好了:
[pyKeOps] Compiling pybind11 template libKeOps_template_7a64278704 in /home/yu/.cache/pykeops-1.5-cpython-38 ... done.
Done.pyKeOps with numpy bindings is working!
只能说,由于版本/依赖等带来的问题太让人觉得恶心,其实学术界也有些研究,例如ICSE每年都有论文说这个事情,但是现实往往让人很抓狂,不知道什么时候会有一些一劳永逸的解决方案?