Starstruk Logo MarkStarstruk

0.2

Netw0rkNoob/VulnClaw0.2Mar 1, 2024by jafioti

AI Summary

This release introduces significant additions to the framework's capabilities, including Convolutional layers (Conv1D, Conv2d) and a Simple Pooling implementation. It also adds support for the Mistral model, improves performance via f16 matmul optimizations, and implements Q8 Weight Quantization.

Key Highlights

  • Conv1D and Conv2d modules added
  • Mistral model support
  • Q8 Weight Quantization implemented
  • Swish activation function fixed
  • f16 matmul improvements for longer prompts

New Features

  • Simple Pooling implementation
  • Conv1D module
  • Conv2d
  • Mistral model support
  • Q8 Weight Quantization
  • Fixed Swish activation function

Full Release Notes

## What's Changed
Everything.

And also:
* Simple Pooling implementation by @TheSeamau5 in https://github.com/jafioti/luminal/pull/4
* Conv1D module by @TheSeamau5 in https://github.com/jafioti/luminal/pull/7
* Conv2d by @TheSeamau5 in https://github.com/jafioti/luminal/pull/8
* Add support for Mistral by @TheSeamau5 in https://github.com/jafioti/luminal/pull/9
* Shell script to download mistral that actually works by @TheSeamau5 in https://github.com/jafioti/luminal/pull/10
* Fix swish by @TheSeamau5 in https://github.com/jafioti/luminal/pull/11
* Minor improvement to f16 matmul, Longer prompt and token generation for testing by @TheSeamau5 in https://github.com/jafioti/luminal/pull/12
* Small Improvements to main by @TheSeamau5 in https://github.com/jafioti/luminal/pull/17
* Q8 Weight Quantization by @jafioti in https://github.com/jafioti/luminal/pull/24


**Full Changelog**: https://github.com/jafioti/luminal/compare/0.1...0.2