剪枝Pruning

删掉模型里不重要的连接，让它变小变快。Delete the unimportant connections in a model to make it smaller and faster.

把权重接近零的连接去掉，模型更小、推理更快。

论文实测：对预训练模型做结构化剪枝，精度损失可以很小。

Removing weights close to zero makes the model smaller and inference faster.

Measured in the paper: structured pruning of a pretrained model can lose very little accuracy.

它和量化是两条互补的压缩路线。本地跑不动的模型，有时靠剪枝救回来。It is a compression route complementary to quantisation. A model that will not fit locally is sometimes rescued by pruning.

去 Models 图谱：看这个模型有哪些量化版、按你的显存该下哪一档Models atlas: which quantisations exist and which tier fits your VRAM ↗

回到名词地图看第5站的其他 29 个词 →Back to the map: the other 29 terms in station 5 →

• 剪枝有效性研究论文To prune, or not to prune (paper) ↗