See all 20 tools →
And I started running tests. I wanted to compare the effect of atlas size, so I made lots of screenshots and started looking closely. I wanted to come up with a way to recommend a specific size. I wanted to make recommendations for all the other parameters. I showed all the commands I ran.
，详情可参考服务器推荐
GLU/SwiGLU 在实际中是门控形式（two linear branches），是向量上的逐元素操作；为了在一维上可视化，我用简化的标量形式来画图 —— 把两条分支都用相同的输入值（即把 a=x, b=x），因此 GLU(x)=x∗sigmoid(x) SwiGLU(x)=x∗SiLU(x) 。这能直观展示门控机制的形状差异。。Line官方版本下载对此有专业解读
Opens in a new window