返回探索

Jev 的可微 IF 与张量逻辑机制

帖子指出 Jev 的两个特殊之处:实现张量逻辑但不以此命名,使 IF 参数可微;以及不浪费生成 token 在已知上下文上。

There are two things that are special that Jev did 1. They are doing tensor logic but don’t call it that, which is that they make the args for your IF differentiable 2. They don’t waste gen tokens on context you know 1 is just telling you that true false needs a gradient to be

Jev 的可微 IF 与张量逻辑机制 1
· 0 次赞在 X 打开