kolinko
Compilation out of MacOS
Hi, i must first say your project is very good, its very useful in essence and i guess we have a ton of stuff to explore.
I am not a Swift person but i know swift is very limited to MacOS ecosystem
do you have currently any interest in making that available for other OSes? Why have you choosen the Swift for this project?
Hi Propfend, It’s not so much about an OS as about hardware - I chose Apple Silicon / MX, so people can run it on their laptops. Making it run on NVidia GPUs would require rewriting the whole thing in CUDA.
Implementing it on CUDA was on my todo list, just didn’t have the time to do it yet.
The core of the algorithm itself is just a few functions in Metal - porting them to CUDA shouldn’t be too difficult if anyone knows Nvidia’s architecture, especially with the help of chatgpt ;)
If you’re up for a challenge, we can collaborate here - but this would require writing the core algorithm in C++/Cuda first. Since this is a new approach to vector-matrix multiplication, my experience was that higher level frameworks hinder more than help - they have assumptions about data and operation structures that you need to hack around to implement the thing efficiently. That’s why I couldn’t use Python’s MLx, nor pytorch.
but this would require writing the core algorithm in C++/Cuda first. Since this is a new approach to vector-matrix multiplication, my experience was that higher level frameworks hinder more than help.
Yes, i think using something like C++/Rust would be a better choice. I am up to collaboration. I have some AMD GPU and Windows if that can be some help as well.
How can we start, another branch/fork?