https://github.com/andresnowak/Micro-Mojograd

https://github.com/andresnowak/Micro-Mojograd

Last synced: 27 days ago
JSON representation

• Host: GitHub
• URL: https://github.com/andresnowak/Micro-Mojograd
• Owner: andresnowak
• License: mit
• Created: 2023-09-12T00:53:27.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2023-11-25T20:16:24.000Z (about 1 year ago)
• Last Synced: 2024-02-11T18:46:15.016Z (10 months ago)
• Size: 57.6 KB
• Stars: 4
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

• awesome-mojo - Micro-Mojograd - Implementation of Karpathy's micrograd in Mojo. (🗂️ Libraries / AI)

README

        
# Micro-Mojograd

This is for now going to be a **Toy project** to learn about how do tensors work, how to make them efficient, how to do the computational graph efficiently (like joining element wise operationss or making a tensors contiguous when it used in more than one operation *use more memory but have better performance*, etc...) and the Autograd part.

When the language is a bit more mature and has complete lifetime and ownership support and Traits support, I will try to make a more complete version of this library, or maybe help another mojo project like [numojo](https://github.com/MadAlex1997/Mojo-Arrays) or [Infermo](https://github.com/TilliFe/Infermo)

And it is going to be inspired by:

-   [Micrograd](https://github.com/karpathy/micrograd)

-   [Pytorch](https://github.com/pytorch/pytorch)

-   [Tinygrad](https://github.com/tinygrad/tinygrad)

## To-do

-   [x] Tensor: using 1d memory and not using MLIR

-   [x] Tensor Matmul

-   [x] Tensor sum elementwise

-   [x] Tensor Mul elementwise

-   [x] Tensor equal operator

-   [ ] Tensor broadcasting

-   [ ] Fix problem where parallel version of element-wise operations (add, mul, eq, etc.) are slower than the sequential version.

-   [ ] Autograd: _For now doing the autogradient part is difficult, because mojo still doesn't have complete ownership and lifetime support._