原书出处
Designing Large Language Model Applications
Token-level early exit In this approach, different tokens from the same sequence can exit at different layers. This is more complex to implement than sequence-level early exit. Similar to sequence-level early exit techniques, you can implement binary classifiers to decide whether to exit at a particular layer, but this happens at each token at each layer, instead of the entire sequence. For more details on
原书定位详情
来源:a19c7933-4eda-4176-aebd-30fe7e1ad292
正文位置:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch09.html
EPUB SHA-256:1b4d2347fec2f949fcddfbaf4fad597a1ae20887a096cfae64c59a619b4781bf