In reply to @justinrestivo:matrix.org
If so, can I see your config?

The nixos config is just

hardware.opengl.enable = true;
services.xserver.videoDrivers = [ "nvidia" ];

The nixpkgs config for building pytorch is { cudaSupport = true; } plus optional cudaCapabilities and cudaEnableForwardCompat.

and what corresponding versions of cuda/cdnn etc are being used.

It's pretty flexible, you can build pretty much against whichever version

In reply to @trexd:matrix.org
I do the above and it works for me justinrestivo

Any chance you could run the following and see if it segfaults for you? Would help me identify if it's my system setup or something else. It'll just bring in the python tooling and run a self contained script

git clone git@github.com:DieracDelta/detypstify.git && cd detypstify && git checkout reproducible && cd train && nix develop .#mvp -c python3 main.py

In reply to @trexd:matrix.org
I do the above and it works for me justinrestivo

Any chance you could run the following and see if it core dumps for you? Would help me identify if it's my system setup or something else. It'll just bring in the python tooling and run a self contained script

git clone git@github.com:DieracDelta/detypstify.git && cd detypstify && git checkout reproducible && cd train && nix develop .#mvp -c python3 main.py

In reply to @trexd:matrix.org
I do the above and it works for me justinrestivo

The backtrace is kinda strange, but coming from libcudnn

0x00007ffff7772e5d in __strlen_avx2 () from /nix/store/anlf335xlh41yjhm114swi87406mq5pw-glibc-2.38-44/lib/libc.so.6
(gdb) bt
#0  0x00007ffff7772e5d in __strlen_avx2 () from /nix/store/anlf335xlh41yjhm114swi87406mq5pw-glibc-2.38-44/lib/libc.so.6
#1  0x00007ff9189340dc in ?? ()
   from /nix/store/pq6sfashimm70zdzlh5gdzv3cl05f6xv-cudatoolkit-11-cudnn-8.9.1-lib/lib/libcudnn_cnn_train.so.8

In reply to @trexd:matrix.org
I do the above and it works for me justinrestivo

Any chance you could run the following and see if it core dumps for you? Would help me identify if it's my system setup or something else. It'll just bring in the python tooling and run a self contained script. Cleanup can be done by removing the directory and gcing

git clone git@github.com:DieracDelta/detypstify.git && cd detypstify && git checkout reproducible && cd train && nix develop .#mvp -c python3 main.py

In reply to @justinrestivo:matrix.org

The backtrace is kinda strange, but coming from libcudnn

0x00007ffff7772e5d in __strlen_avx2 () from /nix/store/anlf335xlh41yjhm114swi87406mq5pw-glibc-2.38-44/lib/libc.so.6
(gdb) bt
#0  0x00007ffff7772e5d in __strlen_avx2 () from /nix/store/anlf335xlh41yjhm114swi87406mq5pw-glibc-2.38-44/lib/libc.so.6
#1  0x00007ff9189340dc in ?? ()
   from /nix/store/pq6sfashimm70zdzlh5gdzv3cl05f6xv-cudatoolkit-11-cudnn-8.9.1-lib/lib/libcudnn_cnn_train.so.8

In reply to @zopieux:matrix.zopi.eu
Hi folks, from the readme it seems like the only recommended flake usage is to use inputs nixpkgs-unstable & SomeoneSerge/nixpkgs-unfree, however I'm not sure I understand how this plays with the flake lock being updated every once in a while by the bot. I'd like to make sure I'm using a nixpkgs & unfree rev that has the packages I care about in the cachix. Is there some magic I'm missing?

Is there anything the community can help with this rehaul?

Yes, I'm sure the community can. Somebody has got to push this (I so far have been struggling with some unrelated stuff so I couldn't): write up an opencollective proposal, write a new proposal to reconsider the hydra situation, &c. One could also finance those who work on this for a living: https://nixos.org/community/teams/cuda/

Is there anything the community can help with this rehaul?

Yes, I'm sure the community can. Somebody has got to push this (I so far have been struggling with some unrelated stuff so I couldn't): write up an opencollective proposal, write a new proposal to reconsider the hydra situation, &c. One could also finance those who already work on this for a living: https://nixos.org/community/teams/cuda/