删除权重矩阵的一些行和列,让 LLAMA-2 70B 的参数量减少 25%,模型还能保持 99% 的零样本任务性能,同时计算效率大大提升。这就是微软 SliceGPT 的威力。
![图片](https://image.jiqizhixin.com/uploads/editor/670fce7f-9cb8-40c3-96a5-d5b5e753fe58/640.png)
论文标题:SLICEGPT: COMPRESS LARGE LANGUAGE MODELS BY DELETING ROWS AND COLUMNS 论文链接:https://arxiv.org/pdf/2401.15024.pdf
![图片](https://image.jiqizhixin.com/uploads/editor/33a0bb43-66a8-41b1-b819-0a4c6f6bc44e/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/2fc30553-1a7c-4ffc-8e4b-ed493d90293b/640.png)
注意,向量 x 乘以 Q 不会改变向量的 norm,因为在这项工作中,Q 的维度总是与 transformer D 的嵌入维度相匹配。
![图片](https://image.jiqizhixin.com/uploads/editor/f5bbfdce-3c88-4f10-a40f-7b64d720b680/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/b1675a71-6c6f-44b2-bd77-80ddf1963dc0/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/ef3a20d3-2ffa-4096-a737-fbfcc7bd3203/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/3a7d4f65-95f6-435b-bf03-8460e1081682/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/25c42e3f-3644-4fcc-9e22-6f2f92689ae9/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/aa8cbe7f-8621-4767-a5fb-79f2d411558e/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/2febec1d-8615-460c-b562-4c23c07f24b4/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/2e240e76-e93f-4e00-be80-08afd9e64c4d/640.jpeg)
![图片](https://image.jiqizhixin.com/uploads/editor/ac7de626-4f3c-4f39-a37c-4ce32a04b802/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/08a14ce9-dacf-470f-ab46-07cc3647b27b/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/6b600964-eb79-4de3-8f01-b05a71eb1221/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/a9b96fe9-57e5-4589-ab56-0a44149b0830/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/ffde93f1-50fb-4b8c-8e85-62c22385a9a3/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/b4ffe312-3406-4e2c-b7f9-3ee535aca58a/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/04436d5a-05ff-441b-892d-cadfc16b8ec8/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/6cad9813-4b63-4341-b4aa-d10b8f5b0e56/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/b6abe76b-2a88-4d8e-846a-09a663a99b1d/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/e0856f9a-4362-4e39-8209-5e50cfa04e88/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/a7dae25e-88b8-4515-ae14-2c13da8834e0/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/09b8f0de-de2f-4022-95ec-2a46a4c5ae2e/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/569589e2-61da-4b8f-8e97-f315b98291d3/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/96ed3b41-9074-46a9-9df1-c4cbf8158ffb/640.jpeg)
![图片](https://image.jiqizhixin.com/uploads/editor/e47462ff-76dc-47b8-abca-e413d3113775/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/bf9742c4-9b5d-468c-a7ba-fe980f87733b/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/a36131d6-6760-4ab4-86de-d87b370a3fbe/640.png)