Compare this to the way something like maven/gradle handles this and you have to wonder WTF is going on here.
Maybe your various LLM libraries are pinning different versions of Torch?
Different Python versions each need their own separate Torch binaries as well.
At least with uv you don't end up with separate duplicate copies of PyTorch in each of the virtual environments for each of your different projects!
Found this the hard way. Something to do with breakage in ABI perhaps. Was looking at the way python implements extensions the other day. Very weird.
There is a "stable ABI" which is a subset of the full ABI, but no requirement to stick to it. The ABI effectively changes with every minor Python version - because they're constantly trying to improve the Python VM, which often involves re-working the internal representations of built-in types, etc. (Consider for example the improvements made to dictionaries in Python 3.6 - https://docs.python.org/3/whatsnew/3.6.html#whatsnew36-compa... .) Of course they try to make proper abstracted interfaces for those C structs, but this is a 34 year old project and design decisions get re-thought all the time and there are a huge variety of tiny details which could change and countless people with legacy code using deprecated interfaces.
The bytecode also changes with every minor Python version (and several times during the development of each). The bytecode file format is versioned for this reason, and .pyc caches need to be regenerated. (And every now and then you'll hit a speed bump, like old code using `async` as an identifier which subsequently becomes a keyword. That hit TensorFlow once: https://stackoverflow.com/questions/51337939 .)
Was some kind of FFI using dlopen and sharing memory across the vm boundary ever considered in the past, instead of having to compile extensions alongside a particular version of python?
I remember seeing some ffi library, probably on pypi. But I don't think it is part of standard python.
To my understanding, though, it's less performant. And you still need a stable ABI layer to call into. FFI can't save you if the C code decides in version N+1 that it expects the "memory shared across the vm boundary" to have a different layout.
Yes, it's essentially that: CPython doesn't guarantee exact ABI stability between versions unless the extension (and its enclosing package) intentionally build against the stable ABI[1].
The courteous thing to do in the Python packaging ecosystem is to build "abi3" wheels that are stable and therefore don't need to be duplicated as many times (either on the index or on the installing client). Torch doesn't build these wheels for whatever reason, so you end up with multiple slightly different but functionally identical builds for each version of Python you're using.
TL;DR: This happens because of an interaction between two patterns that Python makes very easy: using multiple Python versions, and building/installing binary extensions. In a sense, it's a symptom of Python's success: other ecosystems don't have these problems because they have far fewer people running multiple configurations simultaneously.
I am planning to shift some of my stuff to pypy (so a "fast" python exists, kind of). But some dependencies can be problematic, I have heard.
(Recent positive developments in Python’s interpreted performance have subverted this informal tendency.)
You already had billions of lines of Java and JS code that HAD to be sped up. So they had no alternative. If python had gone down the same route, speeding it up without caveats would have been that much easier.
In other languages that didn't happen and you don't have anywhere near as good scientific/ML packages as a result.
Improving Python performance has been a topic as far back as 2008 when I attended my first PyCon. A quick detour on Python 3 because there is some historical revisionism because many online people weren't around in the earlier days.
Back then the big migration to Python 3 was in front of the community. The timeline concerns that popped up when Python really got steam in the industry between 2012 and 2015 weren't as huge a concern. You can refer to Guido's talks from PyCon 2008 and 2009 if they are available somewhere to get the vibe on the urgency. Python is impactful because it changes the language and platform while requiring a massive amount of effort.
Back to perf. Around 2008, there was a feeling that an alternative to CPython might be the future. Candidates included IronPython, Jython, and PyPy. Others like Unladen Swallow wanted to make major changes to CPython (https://peps.python.org/pep-3146/).
Removing the GIL was another direction people wanted to take because it seemed simpler in a way. This is a well researched area with David Beazley having many talks like this oldie (https://www.youtube.com/watch?v=ph374fJqFPE). The idea is much older (https://dabeaz.blogspot.com/2011/08/inside-look-at-gil-remov...).
All of these alternative implementations of Python from this time period have basically failed at the goal of replacing CPython. IronPython was a Python 2 implmentation and updating to Python 3 while trying grow to challenge CPython was impossible. Eventually, Microsoft lost interest and that was that. Similar things happened for the others.
GIL removal was a constant topic from 2008 until recently. Compatibility of extensions was a major concern causing inertia and the popularity meant even more C/C++/Rust code relying on a GIL. The option to disable (https://peps.python.org/pep-0703/) only happened because the groundwork was eventually done properly to help the community move.
The JVM has very clearly defined interfaces and specs similar to the CLR which make optimization viable. JS doesn't have the compatibility concerns.
That was just a rough overview but many of the stories of Python woes miss a lot of this context. Many discussions about perf over the years have descended into a GIL discussion without any data to show the GIL would change performance. People love to talk about it but turn out to be IO-bound when you profile code.
It was expected that extension maintainers would respond negatively to this. In many cases it presents a decision: do I port this to the new platform, or move away from Python completely? You have to remember, the impactful decisions leading us down this path were closer to 2008 than today when dropping Python or making it the second option to help people migrate, would have been viable for a lot of these extensions. There was also a lot of potential for people to follow a fork of the traditional CPython interpreter.
There were no great options because there are many variables to consider. Perf is only one of them. Pushing ahead only on perf is hard when it's unclear if it'll actually impact people in the way they think it will when they can't characterize their actual perf problem beyond "GIL bad".
For pypy it's in a weird spot as the things it does fast are the ones you'd usually just offload to a module implemented in C
(TIOBE's methodology is a bit questionable though, as far as I can tell it's almost entirely based on how many search engine hits they get for "X programming". https://www.tiobe.com/tiobe-index/programminglanguages_defin...)
Python got lucky with the machine learning community using Python. (Thank you TensorFlow and PyTorch, and the SciPy community for saving Python.)
Well, the extensions are going to complicate this a lot.
A fast JIT interpreter cannot reach into a library and do its magic there the way HotSpot/V8 can with native Java/JS code.
(I don't know if this is the reason in Torch's case or not, but I know from experience that it's the reason for many other popular Python packages.)