Softmax Output Approximation for Activation Memory-Efficient Training of Attention-based Networks

Devs

Softmax Output Approximation for Activation Memory-Efficient Training of Attention-based Networks | Read Paper on Bytez