Some problems about pytorch extension

Reference from ：【 Environmental Science 】pytorch choice cuda The order of 【 About cudatoolkit and /usr/local/cuda】

                cuda and cudatoolkit

One ：cuda Expand

        involves cuda Expand , of cuda Source code （.cu file ） The compiled part needs to use NVIDIA Provided by the official website CUDAtoolkit Tools , I like to use it in some projects ninja To compile the source code , Pay attention to ninja need cuda The version of at least 10.0 above .（ and , When there is build.ninja This file , It indicates that the author has written the compiled version for you , Directly enter ：ninja You can perform compilation ）. The ultimate goal of the extension is to generate dynamic link libraries .so file

Two ：cuda Version selection problem

pytorch choice cuda The order of versions is divided into ：

1： Search order at compile time ;

     ①： Search for environment variables CUDA_HOME or CUDA_PATH（ But in general , These two environment variables do not exist ）（ User and environment variable settings ~/.bashrc, System level environment variable settings /etc/profile; After modification, you need to use source Make changes , Otherwise, the modified content will not work ）

     ②：/usr/local/cuda（ This is actually a soft connection , That is to say, it is not the real one we need cuda, It's a goal that points to us cuda Link to , So we can easily change the version we need by modifying it ）

     ③：which nvcc The parent directory of （which nvcc Will be in the environment variable PATH In looking for ,PATH Generally in ~/.bashrc Set in file , There may be more than one PATH Both contain nvcc The situation of , At this time, choose the first PATH Included in NVCC, It can be done by echo $PATH Check the setting of environment variables .）, In general which nvcc The result is /usr/bin/nvcc, there nvcc The default point is the system default cuda, Generally, the version is relatively low , It can be modified to point to the version we need

Be careful ,nvcc -V and nvidia-smi Coming out cuda Version inconsistency , A corresponding drive API, A corresponding operation API, Let's see nvcc -V

     ④： If none of the above exists , be torch.utils.cpp_extension.CUDA_HOME by None, Will use conda Installed cudatoolkit, Its path is cudart The parent directory of the library file directory （ This may be through conda Installed cudatoolkit, General direct use conda install cudatoolkit, It's here that I found cuda Library ）.
 

2： Search order when executing （ Search .so The order of the documents ）： see Reference resources

3：pytorch Used at compile time cuda The version is available through ：torch.version.cuda Check it out.

      And at run time cuda The version is available through torch.utils.cpp_extension.CUDA_HOME Check it out. . Reference from cuda and cudatoolkit

3、 ... and ：cuda/cudatoolkit/nvcc

Reference from ：cuda and cudatoolkit