ow5

Hi! one of the contributors to the paper — we have kernels not released yet that can shave down decoding latency by >20%.

Also when we ran experiments for streaming with the current kernels, we were median ~1.3x slower at inference