X · @teortaxesTex
· X / Twitter
RT Artur Chakhvadze: Btw this is also the reason why you don’t really need diffusion. Good speculative decoding generates more tokens per step at the…
RT Artur ChakhvadzeBtw this is also the reason why you don’t really need diffusion. Good speculative decoding generates more tokens per step at the same block size, without any quality sacrificeArtur Chakhvadze: @leothecurious On coding tasks we can easily generate >10 tokens per forward pass even with moderate specula