So I guess what I'm wondering is, are you a python guy, or are you more like me? because for basically any of these tools, python people tell me "tool X solved all my problems" and people from my own cohort tell me "it doesn't really solve anything, it's still a mess".
If you are one of us, then I'm really listening.
I'm about the highest tier of package manager nerd you'll find out there, but despite all that, I've been struggling to create/run/manage venvs out there for ages. Always afraid of installing a pip package or some piece of python-based software (that might muck up Python versions).
I've been semi-friendly with Poetry already, but mostly because it was the best thing around at the time, and a step in the right direction.
uv has truely been a game changer. Try it out!
Python dependencies are still janky, but uv is a significant improvement over existing tools in both performance and ergonomics.
As an occasional trainer of scientists: it didn't seem to help my students.
But I'd also hesitate to say it "solves all my problems". There's plenty of python problems outside of the core focus of `uv`. For example, I think building a python package for distribution is still awkward and docs are not straightforward (for example, pointing to non-python files which I want to include was fairly annoying to figure out).
The culture that language maintains is rather hostile to maintainable development, easier to just switch to Rust and just write better code by default.
Also, tons of CAE platforms have Python bindings, so you are "forced" to work on Python. Sometimes the solution is not just "abandoning a language".
If it fits your purpose, knock yourself out, for others that may be reading: uv is great for Python dependency management on development, I still have to test it for deployment :)
I'd say Go is a better alternative if you want to replace python scripting. Less friction and much faster compilation times than Rust.
There are libraries to write SIMD in Go now, but I think the better fix is being able to autovectorize during the LLVM IR optimization stage, so its available with multiple languages.
I think LLVM has it now, its just not super great yet.
You can always drop into straight assembly if you need to as well. Go's assembler DX is quite nice after you get used to it.
Lack of types, lack of static analysis, lack of ... well, lack of everything Python doesn't provide and fights users on costs too much developer time. It is a net negative to continue pouring time and money into anything Python-based.
The sole exclusion I've seen to my social circle is those working at companies that don't directly do ML, but provide drivers/hardware/supporting software to ML people in academia, and have to try to fix their cursed shit for them.
Also, fwiw, there is no reason why Triton is Python. I dislike Triton for a lot of reasons, but its just a matmul kernel DSL, there is nothing inherent in it that has to be, or benefits from, being Python.... it takes DSL in, outputs shader text out, then has the vendor's API run it (ie, CUDA, ROCm, etc). It, too, would benefit from becoming Rust.
To say most ML people are using Rust and C couldn’t be further from the truth
Yet it was created for Python. Someone took that effort and did it. No one took that effort in Rust. End of the story of crab's superiority.
Python community is constantly creating new, great, highly usable packages that become de facto industry standards, and maintain old ones for years, creating tutorials, trainings and docs. Commercial vendors ship Python APIs to their proprietary solutions. Whereas Rust community is going through forums and social media telling them that they should use Rust instead, or that they "cheated" because those libraries are really C/C++ libraries (and BTW those should be done in Rust as well, because safety).
I wish this were broadly true.
But there's too much legacy Python sunk cost for most people though. Just so much inertia behind Python for people to abandon it and try to rebuild an extensive history of ML tooling.
I think ML will fade away from Python eventually but right now it's still everywhere.
If someone wrote a Triton impl that is all Rust instead, that would do a _lot_ of the heavy lifting on switching... most of their hard code is in Triton DSL, not in Python, the Python is all boring code that calls Triton funcs. That changes the argument on cost for a lot of people, but sadly not all.
It is happening.
It's like people just don't get it. The ML ecosystem in python didn't just spring from the ether. People wanted to interface in python badly, that's why you have all these libraries with substantial code in another language yet development didn't just shift to that language.
If python was fast enough, most would be fine to ditch the C++ backends and have everything in python, but the reverse isn't true. The C++ interface exists, and no-one is using it.
However people are definitely using it, as Android doesn't do Python, neither does ChromeOS.
It is great for learning on how to program (BASIC replacement), OS scripting tasks as Perl replacement, and embedded scripting in GUI applications.
Additionally understand PYTHONPATH, and don't mess with anything else.
All the other stuff that is supposed to fix Python issues, I never bothered with them.
Thankfully, other languages are starting to also have bindings to the same C and C++ compute libraries.
the disease is the cargo cult addiction (which Rust is full of) to micro libraries, not the language that carries 90% of all peer reviewed papers, datasets, and models published in the last decade
every major breakthrough, from AlphaFold to Stable Diffusion, ships with a Python reference implementation because that is the language researchers can read, reproduce, and extend, remove Python and you erase the accumulated, executable knowledge of an entire discipline overnight, enforcing Rust would sabotage the field more than anything
on the topic of uv, it will do more harm than good by enabling and empowering cargo cults on a systemic level
the solution has always been education, teaching juniors to value simplicity, portability and maintainability
If something doesn't work or I'm still encountering any kind of error with uv, LLMs have gotten good enough that I can just copy / paste the error and I'm very likely to zero-in on a working solution after a few iterations.
Sometimes it's a bit confusing figuring out how to run open source AI-related python projects, but the combination of uv and iterating on any errors with an LLM has so far been able to resolve all the issues I've experienced.
[1] https://github.com/astral-sh/uv