@ivanfioravanti Hi Ivan, nice to meet you.
I just ran Laguna S 2.1 oQ2e on a clean 64GB M1 Max, so I thought I'd share.
Results: quality 0.967; 16K prefill 115.72 tok/s; decode 25.39 tok/s; 38.47GB peak; no swap.
This was with the pinned mlx-vlm setup—not oMLX 0.5.3:
https://t.co/AKnZToA16Z