After seeing a few posts on here about it, I finally tried exl3 3bpw and exllamav3 for running flash next - with amazing results. On 3x3090s, 128GB DDR4: 1500 prefill, 80 tps decode On 1x5090, 128G. DDR4: 1500 prefill, 29 tps decode Both at 262k context, both with vision/spec decoding. Really impres
