File size: 1,168 Bytes
1b1d3e1 36c5eb1 a5e753a 36c5eb1 116873c 36c5eb1 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 |
---
license: apache-2.0
---
# BlackMamba
<img src="https://cdn-uploads.huggingface.co/production/uploads/65bc13717c6ad1994b6619e9/JdxNtwFrmEAnjJ0_MP5A3.jpeg" width="900" height="900" />
> **BlackMamba: Mixture of Experts for State-space models**\
> Quentin Anthony*, Yury Tokpanov*, Paolo Glorioso*, Beren Millidge*\
> Paper: https://arxiv.org/abs/2402.01771
<img src="https://cdn-uploads.huggingface.co/production/uploads/65bc13717c6ad1994b6619e9/aHpEc5tnCJShO2Kn0f637.png" width="900" height="900" />
## About
We provide inference code for our BlackMamba model in our github repository: https://github.com/Zyphra/BlackMamba
BlackMamba is an novel architecture which combines state-space models (SSMs) with mixture of experts (MoE). It uses [Mamba](https://arxiv.org/abs/2312.00752) as its SSM block and [switch transformer](https://arxiv.org/abs/2101.03961) as its MoE block base. BlackMamba is extremely low latency for generation and inference, providing significant speedups over all of classical transformers, MoEs, and Mamba SSM models. Additionally, due to its SSM sequence mixer, BlackMamba retains linear compuational complexity in the sequence length. |