๐ฅ Level up your model training w/ GaLore + Transformers for SOTA results on consumer-grade hardware!
โฌ๏ธ 82.5% less optimizer state memory footprint without performance degradation by expressing the gradient weight matrix as low rank.
๐ฉ๐ฟโ๐ป Install via pip install transformers>=4.39.0 galore-torch. #ProudlyGpuPoor
The integration of GaLore into the training of large language models (LLMs) marks a significant advancement in the field of deep learning, particularly in terms of memory efficiency and the democratization of AI research. By allowing for the training of billion-parameter models on consumer-grade hardware, reducing memory footprint in optimizer states, and leveraging advanced projection matrix techniques, GaLore opens new horizons for researchers and practitioners with limited access to high-end computational resources.
Under the hood there we have many other improvements, due to extensive maintenance activity, community contributions by super active + knowledgable volunteers โจ ๐ and the official sponsorship by Hugging Face that makes all this possible ๐ค โค๏ธ ๐
We would greatly appreciate any further community contributions, be it to help with refactorings, exterminating flaky tests, writing doc-strings, tutorials, new features. Don't be shy, just contact us and we see where this leads us: https://github.com/TimDettmers/bitsandbytes/discussions
Please let us know what you think: Your feedback is essential to us, and we would greatly appreciate any insights you have on how we can further enhance it or even better be happy to merge your contributions, filling in some blanks: Especially doc-strings are still a big topic and there several placeholder that would be super helpful to have filled in. Please post your feedback here: https://github.com/TimDettmers/bitsandbytes/discussions/1090
Since taking over maintenance together with Younes Belkada and since Hugging Face graciously agreed to support the library, we've already made enormous strides and community contributions have sprung back to life: It's so motivating to have so many knowledgeable contributors that often invest extensive free-time and bring their unique ideas to the table.
A notable example are our ongoing efforts to enable cross-platform support, including Intel, Apple Silicon, AMD, and Windows. Simultaneously, we're working diligently to streamline community contributions in BNB, making the process more accessible for everyone. A heartfelt thank you to all who have contributed thus far!
With HuggingFace's committed to supporting bitsandbytes going forward, we're sure to promptly respond to and integrate additional community contributions.
Looking forward to growing bitsandbytes further as part of the FOSS community: pushing forward the state of the art in democratization of AI!