Softmax
activation
Softmax over attention scores; numerically stable form
公式
FLOPs
5 * batch * heads * seq * seqBytes
2 * batch * heads * seq * seq使用此算子的模型 (0)
尚无模型在算子拆解中引用此算子。
Softmax over attention scores; numerically stable form
5 * batch * heads * seq * seq2 * batch * heads * seq * seq尚无模型在算子拆解中引用此算子。