Is there music AI without Python?

Posted by iwakawa2173@reddit | LocalLLaMA | View on Reddit | 25 comments

Is there a ready built without Python build for YuE or any other AI like LM Studio that I can load models into? This tower of Python dependencies is driving me to suicide. It took me several hours on Pinokio and YuE without Pinokio to come to the same conclusion: I need C++ packages and lots and lots of libraries and other things. Why can't I just download and run a ready build? Why do the developers just drop raw code that requires downloading 10+ gigabytes of junk to compile, which I'll never need?

Is there ready-made builds of YuE or other music AI?

P.S. I hate Python. I hate Python so much with its tower of dependencies and libraries that I'll soon start quoting AM from Harlan Ellison's story.

[-]

Such_Advantage_6949@reddit

If u hate it so much then create something of your own. If u dont have such skill then why complaining?

[-]

J0kooo@reddit

try uv for package management lmao

[-]

Minute_Attempt3063@reddit

Still uses python

Which is a pain

[-]

J0kooo@reddit

honestly never really heard "python is a pain" before

hard skill issue

[-]

Minute_Attempt3063@reddit

Maybe for you

I don't like having 150gb of the same packages over and over in different projects

Sure UV solves that.

Python is also just way slower then native.

Not to mention that with python the same model that is supposed to be in VRAM is also loaded in system ram, even if fully offloaded. No idea why, guess it's just a quirk of python

[-]

Savantskie1@reddit

I always wondered why it does this

[-]

Minute_Attempt3063@reddit

Apparently it's a skill issue XD

[-]

Savantskie1@reddit

Honestly I hate this response sometimes. Even when I know it’s the right response. But sometimes getting answers from AI is easier than from people

[-]

gefahr@reddit

Python packaging is historically atrocious, but you're right: uv solves 99% of the pain for typical use cases.

I assume most laypeople (who haven't developed Python apps) lump the two together?

[-]

satireplusplus@reddit

Ah yes, because java package management is so much better. Or npm javascript micropackage hell. Cargo (rust) is also hot shit requiring you to run some kind of bash script straight from their website, because you can't really install anything with the default distro one.

[-]

Languages_Learner@reddit

Here's music AI written in Rust: gabotechs/MusicGPT: Generate music based on natural language prompts using LLMs running locally. The repo has compiled releases for Windows, Mac and Linux.

[-]

Double_Cause4609@reddit

Because machine learning developers aren't necessarily traditional programmers. It's a completely different skill to have the mathematical and algorithmic knowledge to build a complex, state of the art system, and to write it in an optimized, low level language like C.

If you do not want to deal with Python's dependencies, either containerize it, or re-implement it in C++ yourself.

It is absurd to suggest that developers doing research, or making robust applications in industry professionally should be obligated to do extra work to save you from a process that millions of people get by doing just fine. The additional work of re-implementing the modelling code, the forward methods, and the program flow in C++ is absolutely insane, especially because the toolchains, libraries, and interfaces are all completely different.

Skill issue.

[-]

utilitycoder@reddit

Imagine how fast the AI and machine learning pipeline would be with a compiled typesafe laguage though.

[-]

Environmental-Metal9@reddit

I suspect it would have a lot to gain from being implemented in a lower level, compiled language. I’m not so sure it would gain any speed from being typesafe though. Type safety is more for devs than for the runtime, unless we’re talking about Haskell. If anyone wants to tackle live coding nanogpt in Haskell, I’d watch. That would be one hell of a monad

[-]

Rodrigo_s-f@reddit

It wouldn't do much. The libraries used for ML are written in lower level languages like C. You only call those functions from python.

For inference you can always convert the models to a binary format.

[-]

HistorianPotential48@reddit

paste the repo to chatgpt or grok 2. tell it you need a fastAPI, a dockerfile and a docker-compose. 3. pray for the best

[-]

neil_555@reddit

I just found this gem on Youtube

https://www.youtube.com/watch?v=6fhPvrLXGmU

It's a walkthrough for installing YuE on windows

[-]

neil_555@reddit

Another one here

https://www.youtube.com/watch?v=1SihhqPR0S8&t=123s

[-]

shiren271@reddit

Try comfyui. It's still python, but the package manager takes care of the dependencies.

[-]

neil_555@reddit

Hopefully the LM studio developers will add support for image generation, music and voice generation, TTS input and Speech output soon. The whole Python (and indeed having to use Linux at any level) is a total PITA. And no this isn't a skill issue it's a matter of taste!

[-]

eli_pizza@reddit

I see YuE recommends using conda, which definitely isn’t my first choice. Python is a good language but it’s not a high performance language so you are going to have a bunch of C/C++/etc stuff to install one way or the other.

Personally I would want to run the whole thing in docker. Might have to mess around to get the GPU working right but otherwise that should solve dependency hell.

[-]

cosimoiaia@reddit

Just vibecode it with chatgpt. /s

If you want off-the-shelf products don't expect experimental research advancement level.

[-]