Problem with setup_cuda.py install about gptq-for-llama

[farrael004]

That doesn't work either. I managed to run it in a docker image, but that's just a workaround. Unfortunately even that doesn't work for what I want to do.

[mityarko]

If you are using conda try the following:

conda install cudatoolkit=11.7 -c conda-forge
conda install cuda -c nvidia/label/cuda-11.7.0
python setup_cuda.py install

[TheTerrasque]

I'm guessing you're missing VS build toolkit. #11 (comment) - Need that for windows.

Alternatively you can use WSL2 or Docker

[farrael004]

It's installed. Maybe something from it is not in path? I wouldn't know what though.

[papuSpartan]

It's installed. Maybe something from it is not in path? I wouldn't know what though.

Open windows search by hitting Windows + S and search native. You should see x64 Native Tools Command Prompt for VS 2019 (make sure you do not select the wrong year if you already had another version installed like 2022). Also, since it's a cmd shell and not powershell you will need to run set DISTUTILS_USE_SDK=1 instead of $env:DISTUTILS_USE_SDK=1.

[DamascusGit]

That doesn't work either. I managed to run it in a docker image, but that's just a workaround. Unfortunately even that doesn't work for what I want to do.

same here. i am getting the exact same error, have tried everything and still getting this issue. i've also got VS 2019, i set DISTUTILS, i do everything and still no help.

[farrael004]

I managed to find a workaround (It requires a bit of Linux knowledge). Install Ubuntu 22.04 from the Microsoft Store, then open it and follow the installation process for linux instead. I installed cuda in this linux too, but I'm not sure if it was necessary. If you do, just make sure that the cuda installed in your windows and linux are the same version, otherwise you'll run into errors.

[DamascusGit]

I managed to find a workaround (It requires a bit of Linux knowledge). Install Ubuntu 22.04 from the Microsoft Store, then open it and follow the installation process for linux instead. I installed cuda in this linux too, but I'm not sure if it was necessary. If you do, just make sure that the cuda installed in your windows and linux are the same version, otherwise you'll run into errors.

can you speak to more detailed instructions? dying to run this.

[DamascusGit]

like, where are the linux setup instructions? is it the same instructions once i have followed the WSL setup? how do i use the VSCode build tools in linux?

[farrael004]

It's essentially the same thing, but you don't have to install build tools for it since linux comes with a C++ compiler.

[DamascusGit]

ah yeah, it didn't work for me on linux, but it worked on python 3.10 with the alternate install workaround. it gave me a "CUDA extension not installed" message, but is currently loading the sharded checkpoints anyway.

[Brawlence]

ah yeah, it didn't work for me on linux, but it worked on python 3.10 with the alternate install workaround. it gave me a "CUDA extension not installed" message, but is currently loading the sharded checkpoints anyway.

Be aware: that wheel is an old (2 weeks, lmao) version and it may not work with the current patches.

For any lost souls who look for compiled kernels, it's probably best to use: https://github.com/jllllll/GPTQ-for-LLaMa-Wheels

[qwopqwop200]

currently using triton, so I don't think this problem will occur.

[Paillat-dev]

Yes of course, but not for windows users... I's a shame