[m_kos]

> Why is it so hard to install PyTorch, or CUDA, or libraries like FlashAttention or DeepSpeed that build against PyTorch and CUDA?

This is so true! On Windows (and WSL) it is also exacerbated by some packages requiring the use of compilers bundled with outdated Visual Studio versions, some of which are only available by manually crafting download paths. I can't wait for a better dev experience.

[giancarlostoro]

Stuff like that led me fully away from Ruby (due to Rails), which is a shame, I see videos of people chugging along with Ruby and loving it, and it looks like a fun language, but when the only way I can get a dev environment setup for Rails is using DigitalOcean droplets, I've lost all interest. It would always fail at compiling something for Rails. I would have loved to partake in the Rails hype back in 2012, but over the years the install / setup process was always a nightmare.

I went with Python because I never had this issue. Now with any AI / CUDA stuff its a bit of a nightmare to the point where you use someone's setup shell script instead of trying to use pip at all.

[awesome_dude]

Lets be honest here - whilst some experiences are better/worse than others, there doesn't seem to be a dependency management system that isn't (at least half) broken.

I use Go a lot, the journey has been

- No dependency management

- Glide

- Depmod

- I forget the name of the precursor - I just remembered, VGo

- Modules

We still have proxying, vendoring, versioning problems

Python: VirtualEnv

Rust: Cargo

Java: Maven and Gradle

Ruby: Gems

Even OS dependency management is painful - yum, apt (which was a major positive when I switched to Debian based systems), pkg (BSD people), homebrew (semi-official?)

Dependency Management is the wild is a major headache, Go (I only mention because I am most familiar with) did away with some compilation dependency issues by shipping binaries with no dependencies (meaning that it didn't matter which version of linux you built your binary for, it will run on any of the same arch linux - none of that "wrong libc" 'fun'), but you still have issues with two different people building the same binary in need of extra dependency management (vendoring brings with it caching problems - is the version in the cache up to date, will updating one version of one dependency break everything - what fun)

[ilvez]

Do I get it right that this issue is within Windows? I've never heard of the issues you describe while working with Linux.. I've seen people struggle with MacOS a bit due to brew different versions of some library or the other, mostly self compiling Ruby.

[avidphantasm]

Have you tried conda? Since the integration of mamba its solver is fast and the breadth of packages is impressive. Also, if you have to support Windows and Python with native extensions, conda is a godsend.

[viraptor]

I would recommend learning a little bit of C compilation and build systems. Ruby/Rails is about as polished as you could get for a very popular project. Maybe libyaml will be a problem once in a while if you're compiling Ruby from scratch, but otherwise this normally works without a hassle. And those skills will apply everywhere else. As long as we have C libraries, this is about as good as it gets, regardless of the language/runtime.

[nurettin]

Have you tried JRuby? It might be a bit too large for your droplet, but it has the java versions of most gems and you can produce cross-platform jars using warbler.

[chao-]

I'm surprised to hear that. Ruby was the first language in my life/career where I felt good about the dependency management and packaging solution. Even when I was a novice, I don't remember running into any problems that weren't obviously my fault (for example, installing the Ruby library for PostgreSQL before I had installed the Postgres libraries on the OS).

Meanwhile, I didn't feel like Python had reached the bare minimum for package management until Pipenv came on the scene. It wasn't until Poetry (in 2019? 2020?) that I felt like the ecosystem had reached what Ruby had back in 2010 or 2011 when bundler had become mostly stable.

[poly2it]

Have you tried Nix?

https://nixos.org

[bytehumi]

This is the right direction for Python packaging, especially for GPU-heavy workflows. Two concrete things I'm excited about: 1) curated, compatibility-tested indices per accelerator (CUDA/ROCm/CPU) so teams stop bikeshedding over torch/cu* matrixes, and 2) making metadata queryable so clients can resolve up front and install in parallel. If pyx can reduce the 'pip trial-and-error' loop for ML by shipping narrower, hardware-targeted artifacts (e.g., SM/arch-specific builds) and predictable hashes, that alone saves hours per environment. Also +1 to keeping tools OSS and monetizing the hosted service—clear separation builds trust. Curious: will pyx expose dependency graph and reverse-dependency endpoints (e.g., "what breaks if X→Y?") and SBOM/signing attestation for supply-chain checks?

[int_19h]

Given that WSL is pretty much just Linux, I don't see what relevance Visual Studio compiler versions have to it. WSL binaries are always built using Linux toolchains.

At the same time, even on Windows, libc has been stable since Win10 - that's 10 years now. Which is to say, any binary compiled by VC++ 2015 or later is C-ABI-compatible with any other such binary. The only reasons why someone might need a specific compiler version is if they are relying on some language features not supported by older ones, or because they're trying to pass C++ types across the ABI boundary, which is a fairly rare case.

[m_kos]

If you have to use, e.g., CUDA Toolkit 11.8, then you need a specific version of VS and its build tools for CUDA's VS integration to work. I don't know why exactly that is and I wish I didn't have to deal with it.

[morkalork]

This was basically the reason to use anaconda back in the day.

[setopt]

In my experience, Anaconda (including Miniconda, Micromamba, IntelPython, et al.) is still the default choice in scientific computing and machine learning.

[IHLayman]

Anaconda was a good idea until it would break apt on Ubuntu and make my job that much harder. That became the reason _not_ to use Anaconda in my book.

venv made these problems start to disappear, and now uv and Nix have closed the loop for me.

[miraculixx]

Windows is the root cause here, not pip

[miohtama]

In the past, part of the definition of an operating system was that it ships with a compiler.

[int_19h]

When was that ever a part of the definition? It was part of the early Unix culture, sure, but even many contemporary OSes didn't ship with compilers, which were a separate (and often very expensive!) piece of software.

OTOH today most Linux distros don't install any dev tools by default on a clean install. And, ironically, a clean install of Windows has .NET, which includes a C# compiler.