Add pipeline_tag, library_name and improve model card
#2
by nielsr HF Staff - opened
README.md
CHANGED
|
@@ -1,11 +1,14 @@
|
|
| 1 |
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
datasets:
|
| 4 |
- openbmb/InfLLM-V2-data-5B
|
| 5 |
language:
|
| 6 |
- en
|
| 7 |
- zh
|
|
|
|
|
|
|
|
|
|
| 8 |
---
|
|
|
|
| 9 |
<div align="center">
|
| 10 |
|
| 11 |
<h1>NOSA: Native and Offloadable Sparse Attention</h1>
|
|
@@ -33,26 +36,23 @@ language:
|
|
| 33 |
|
| 34 |
**NOSA** is a trainable sparse attention mechanism designed for KV-cache offloading with an explicit locality constraint, paired with an inference system (**NOSI**) to realize its efficiency. It improves long-context/long-generation quality over prior offloading baselines while boosting decoding throughput by up to **5.04×** vs **FullAttn**, **1.92×** vs **InfLLMv2**, and **1.83×** vs **ShadowKV** on **1B/3B/8B** LLMs.
|
| 35 |
|
| 36 |
-
|
| 37 |
|
| 38 |
## Models
|
| 39 |
|
| 40 |
-
We train 1B, 3B, and 8B models
|
| 41 |
|
| 42 |
|Model|Link|
|
| 43 |
|:-:|:-:|
|
| 44 |
-
|NOSA-1B | [NOSA-1B](huggingface.co/openbmb/NOSA-1B) |
|
| 45 |
-
|NOSA-3B | [NOSA-3B](huggingface.co/openbmb/NOSA-3B) |
|
| 46 |
-
|NOSA-8B | [NOSA-8B](huggingface.co/openbmb/NOSA-8B) |
|
| 47 |
-
|
| 48 |
-
Please reach out to us if additional baseline models (FullAttn, InfLLMv2, or DMA) are needed. You may open an issue or contact us directly via email (our email addresses are provided in the paper).
|
| 49 |
-
|
| 50 |
-
|
| 51 |
|
|
|
|
| 52 |
|
| 53 |
## Citation
|
| 54 |
|
| 55 |
-
```
|
| 56 |
@article{huang2025nosa,
|
| 57 |
title={NOSA: Native and Offloadable Sparse Attention},
|
| 58 |
author={Huang, Yuxiang and Wang, Pengjie and Han, Jicheng and Zhao, Weilin and Su, Zhou and Sun, Ao and Lyu, Hongya and Zhao, Hengyu and Wang, Yudong and Xiao, Chaojun and Han, Xu and Liu, Zhiyuan},
|
|
|
|
| 1 |
---
|
|
|
|
| 2 |
datasets:
|
| 3 |
- openbmb/InfLLM-V2-data-5B
|
| 4 |
language:
|
| 5 |
- en
|
| 6 |
- zh
|
| 7 |
+
license: apache-2.0
|
| 8 |
+
pipeline_tag: text-generation
|
| 9 |
+
library_name: transformers
|
| 10 |
---
|
| 11 |
+
|
| 12 |
<div align="center">
|
| 13 |
|
| 14 |
<h1>NOSA: Native and Offloadable Sparse Attention</h1>
|
|
|
|
| 36 |
|
| 37 |
**NOSA** is a trainable sparse attention mechanism designed for KV-cache offloading with an explicit locality constraint, paired with an inference system (**NOSI**) to realize its efficiency. It improves long-context/long-generation quality over prior offloading baselines while boosting decoding throughput by up to **5.04×** vs **FullAttn**, **1.92×** vs **InfLLMv2**, and **1.83×** vs **ShadowKV** on **1B/3B/8B** LLMs.
|
| 38 |
|
| 39 |
+
The model was presented in the paper [NOSA: Native and Offloadable Sparse Attention](https://huggingface.co/papers/2510.13602).
|
| 40 |
|
| 41 |
## Models
|
| 42 |
|
| 43 |
+
We train 1B, 3B, and 8B models using FullAttn, InfLLMv2, DMA, and NOSA. The following NOSA models have been released on Hugging Face:
|
| 44 |
|
| 45 |
|Model|Link|
|
| 46 |
|:-:|:-:|
|
| 47 |
+
|NOSA-1B | [NOSA-1B](https://huggingface.co/openbmb/NOSA-1B) |
|
| 48 |
+
|NOSA-3B | [NOSA-3B](https://huggingface.co/openbmb/NOSA-3B) |
|
| 49 |
+
|NOSA-8B | [NOSA-8B](https://huggingface.co/openbmb/NOSA-8B) |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
|
| 51 |
+
Please reach out to the authors if additional baseline models (FullAttn, InfLLMv2, or DMA) are needed. You may open an issue on the [GitHub repository](https://github.com/thunlp/NOSA) or contact the authors directly via email.
|
| 52 |
|
| 53 |
## Citation
|
| 54 |
|
| 55 |
+
```bibtex
|
| 56 |
@article{huang2025nosa,
|
| 57 |
title={NOSA: Native and Offloadable Sparse Attention},
|
| 58 |
author={Huang, Yuxiang and Wang, Pengjie and Han, Jicheng and Zhao, Weilin and Su, Zhou and Sun, Ao and Lyu, Hongya and Zhao, Hengyu and Wang, Yudong and Xiao, Chaojun and Han, Xu and Liu, Zhiyuan},
|