How can we say matmul-free enables parallel computation?

[kimwin2]

I read your paper with great interest and I understand that you used GRU to eliminate the need for MatMul operations. However, I couldn't quite grasp how it enabled parallel computations. Could you please provide a more detailed explanation?

thanks in advanced.

[ridgerchu]

Hi, this is due to the nature of the parallel scan. You can find more details on that page. Mamba also utilizes this trick. In short, the modified formula satisfies the associative property, which can be easily parallelized.