Config for MMCV 2.x with pytorch and cuda

Recently I have encountered a big(as for me) problem in configuring mmcv for training and running Segformer on the machine with relatively new version of pytorch and cuda.

Problem

Initially, cuda:13.2 and pytorch:nightly are configured in my machine, which are not compatible with mmcv:1.x, throwing error about setuptool.

Attempts

A natural consideration is to degrade pytorch and cuda so that they can work with mmcv:1.x,  however, turning out in vein. The reason is that there's no suitable version for pytorch to be compatible with mmcv:1.x.

 

Then another natural thinking is obviously installing mmcv:2.x. Unfortunately, the installation work subsequently, though, an error was popped onto my screen

cannot find mmcv.runner

 

The underlying reason is evident that the api implementation of 2.x is quite different from 1.x. The only way to fix the problem is to transfer the api in my source code to ones that can work in 2.x ,leading to a great amount of work.

Solution

My colleague, TingYang, solved this problem in an elegant but arduous way. Working for up to 6 hours, she has tryed several combinations and found a solution.

 

 

This combination can run mmcv:2.x without the error of mmcv.runner. It is amazing but useful to some degree. So I post here, hoping it can help someone.