Add link to paper (#1)
Browse files- Add link to paper (4848ceb47ad401827547eb58d409745f66614b6f)
Co-authored-by: Niels Rogge <nielsr@users.noreply.huggingface.co>
README.md
CHANGED
|
@@ -3,13 +3,15 @@ license: mit
|
|
| 3 |
---
|
| 4 |
|
| 5 |
# Sparge-attention model zoo
|
|
|
|
| 6 |
Welcome to Sparge-attention model zoo, this repo contains list of hyperparameters pre-tuned for branch of models.
|
| 7 |
|
|
|
|
|
|
|
| 8 |
## Naming of ckpt
|
| 9 |
The tuned ckpt is often named by following format:`${moddel name or type}_${l1}_${pv_l1}.pt`, in some cases the pv_l1 will be omitted when not choose to tune pv.
|
| 10 |
The larger l1 and pv_l1 make model more sparse, but may sacrifice output quality.
|
| 11 |
|
| 12 |
-
|
| 13 |
## Overview
|
| 14 |
|
| 15 |
| model name | tuned ckpt dir |
|
|
|
|
| 3 |
---
|
| 4 |
|
| 5 |
# Sparge-attention model zoo
|
| 6 |
+
|
| 7 |
Welcome to Sparge-attention model zoo, this repo contains list of hyperparameters pre-tuned for branch of models.
|
| 8 |
|
| 9 |
+
It was presented in the paper [SpargeAttn: Accurate Sparse Attention Accelerating Any Model Inference](https://huggingface.co/papers/2502.18137).
|
| 10 |
+
|
| 11 |
## Naming of ckpt
|
| 12 |
The tuned ckpt is often named by following format:`${moddel name or type}_${l1}_${pv_l1}.pt`, in some cases the pv_l1 will be omitted when not choose to tune pv.
|
| 13 |
The larger l1 and pv_l1 make model more sparse, but may sacrifice output quality.
|
| 14 |
|
|
|
|
| 15 |
## Overview
|
| 16 |
|
| 17 |
| model name | tuned ckpt dir |
|