Fedora 40 Looks At Packaging Its Own PyTorch

[phoronix]

Phoronix: Fedora 40 Looks At Packaging Its Own PyTorch

While on Fedora and other Linux distributions it can be as easy as running "pip3 install torch" or similar for deploying the PyTorch machine learning framework, Fedora 40 is looking at packaging PyTorch on its own for enhancing the Fedora Linux user experience...

https://www.phoronix.com/news/Fedora-40-Own-PyTorch-Proposal

[Jabberwocky]

Whatever you do please just keep it open source.

It's not just as easy as running "pip3 install torch".

Microsoft is maintaining it's own PyTorch. AMD is doing the same. Nvidia also... everyone forks clang, everyone forks llvm.

It's a mess with different obscure --index-url hardcoded into install blogs or requirements.txt that becomes a maintenance nightmare.

[Serafean]

Compute is currently in the state 3D was when D3D, glide, quickdraw and OpenGL started battling it out : a mess...

[V1tol]

Compute is currently in the state 3D was when D3D, glide, quickdraw and OpenGL started battling it out : a mess...

I hope Vulkan will save everyone. Right now I am scared on amount of dependencies Arch's python-pytorch-* wants to install on my machine - from 15 to 26 gigs depending on runtime

[klapaucius]

with the Pip-based installation building locally on the system, there is chances it could fail and not be consistent from machine-to-machine. Thus the hope is that telling users to sudo dnf install python-torch will be more consistent and easier.

That's why there are Anaconda/Miniconda, conda etc, in order to specify consistent environments with well-defined versions of the packages. It's wasteful because it creates self-contained distributions which duplicates the local packages, but the configuration can be shared with other people, maybe not using Fedora at all.

[spicfoo]

That's why there are Anaconda/Miniconda, conda etc, in order to specify consistent environments with well-defined versions of the packages. It's wasteful because it creates self-contained distributions which duplicates the local packages, but the configuration can be shared with other people, maybe not using Fedora at all.

You can still share with other people regardless of their distro. That's why things like Podman and distrobox exists but sourcing the content from native packages makes it easier to have a clear BOM and reproducible builds of software that goes into those containers.

[atmartens]

That's why there are Anaconda/Miniconda, conda etc, in order to specify consistent environments with well-defined versions of the packages. It's wasteful because it creates self-contained distributions which duplicates the local packages, but the configuration can be shared with other people, maybe not using Fedora at all.

I gave up on Anaconda/Miniconda/Mamba etc. & switched to Python envs. Works with pip and is way more well behaved!

[curfew]

I hope Vulkan will save everyone. Right now I am scared on amount of dependencies Arch's python-pytorch-* wants to install on my machine - from 15 to 26 gigs depending on runtime

Umm... The base package plus all its dependencies comes down to 780 MB download size and 5.6 GB install size which sounds pretty alright given that all the included packages seem "useful" and not e.g. older versions of system libraries.

The Cuda-optimized package takes a lot more space but that's because Cuda alone adds about 5 GB and then there's older GCC and so on being downloaded as well.

[boeroboy]

It's not just as easy as running "pip3 install torch".

Maybe it's just me but pip is such a mess lately I can't even do that right now. They keep renaming packages and deps and I can't even install it right now. I would love to have stable packaging. I can't even convert models right now without it.

> pip3 install torch
Defaulting to user installation because normal site-packages is not writeable
ERROR: Could not find a version that satisfies the requirement torch (from versions: none)
ERROR: No matching distribution found for torch

[V1tol]

Umm... The base package plus all its dependencies comes down to 780 MB download size and 5.6 GB install size which sounds pretty alright given that all the included packages seem "useful" and not e.g. older versions of system libraries.

The Cuda-optimized package takes a lot more space but that's because Cuda alone adds about 5 GB and then there's older GCC and so on being downloaded as well.

Anyone not masochistic needs GPU in pytorch. CUDA becomes 4GB/16GB (download and installed) and ROCM is 2GB/26GB on my machine.