4 ms·

What’s the best alternative distro for ML work? I mainly need the nvidia stack + PyTorch

by horsellama 2y ago

homarp 2y ago
arch ?
- all2 2y ago
  I managed to get a 4090 hooked up on an Arch machine. It buzzes when the LLM is generating stuff. It isn't a straight forward thing, though, you still have to dig through the deps and get your versions just right. On top of that you need to get your versions from bottom-to-top correct for everything to work. Some versions of Cuda don't play nicely with some versions of the Python modules, so take care if you go this route.
exe34 2y ago
My 3090 is plugged into an old intel pc running Debian 11. Tried Debian 12, but couldn't quite get the right version of cuda to work with both tensorflow and pytorch.
lnxg33k1 2y ago
I’m using debian sid without issues to communicate, but Ive used linux for 20 years so maybe my definition of issues might be poisoned I switched to sid from arch because their approach to software major release is dangerous (i.e. adopt plasma 6 the day of release)
rickspencer3 2y ago
I am so biased this borders on shilling, but may I suggest taking a look at openSUSE Leap Micro? https://get.opensuse.org/leapmicro/5.5/ https://get.opensuse.org/leapmicro/5.5/ We (SUSE) have extensive testing and quality control, if you are looking for something very stable. Leap is the version built out of the same bits as the enterprise version. The "Micro" part means it is smaller, transactional, and immutable. It's built for containerized and virtualized applications. You can install the nvidia OS bits in the OS, and then build a container with libraries.
bayindirh 2y ago
Debian.
FridgeSeal 2y ago
PopOS. Comes with the drivers you need, super nice and streamlined UX.
skywhopper 2y ago
Debian will be the closest in terms of functionality and organization.
logicprog 2y ago
Personally I'd just pick whatever Linux distro suits your fancy for other reasons, and then the official PyTorch docker image as a distrobox. The combination of Nvidia drivers, CUDA/CUDNN/etc libraries, and Python libraries is extremely fragile and sensitive to version bumps and stuff so I'd want it in an isolated consistent official box. This I'd what I do on ublue.

What’s the best alternative distro for ML work? I mainly need the nvidia stack + PyTorch

homarp 2y ago

arch ?

all2 2y ago
I managed to get a 4090 hooked up on an Arch machine. It buzzes when the LLM is generating stuff. It isn't a straight forward thing, though, you still have to dig through the deps and get your versions just right. On top of that you need to get your versions from bottom-to-top correct for everything to work. Some versions of Cuda don't play nicely with some versions of the Python modules, so take care if you go this route.

exe34 2y ago

My 3090 is plugged into an old intel pc running Debian 11. Tried Debian 12, but couldn't quite get the right version of cuda to work with both tensorflow and pytorch.

lnxg33k1 2y ago

I’m using debian sid without issues to communicate, but Ive used linux for 20 years so maybe my definition of issues might be poisoned I switched to sid from arch because their approach to software major release is dangerous (i.e. adopt plasma 6 the day of release)

rickspencer3 2y ago

I am so biased this borders on shilling, but may I suggest taking a look at openSUSE Leap Micro? https://get.opensuse.org/leapmicro/5.5/ https://get.opensuse.org/leapmicro/5.5/ We (SUSE) have extensive testing and quality control, if you are looking for something very stable. Leap is the version built out of the same bits as the enterprise version. The "Micro" part means it is smaller, transactional, and immutable. It's built for containerized and virtualized applications. You can install the nvidia OS bits in the OS, and then build a container with libraries.

bayindirh 2y ago

Debian.

FridgeSeal 2y ago

PopOS. Comes with the drivers you need, super nice and streamlined UX.

skywhopper 2y ago

Debian will be the closest in terms of functionality and organization.

logicprog 2y ago

Personally I'd just pick whatever Linux distro suits your fancy for other reasons, and then the official PyTorch docker image as a distrobox. The combination of Nvidia drivers, CUDA/CUDNN/etc libraries, and Python libraries is extremely fragile and sensitive to version bumps and stuff so I'd want it in an isolated consistent official box. This I'd what I do on ublue.