llama.cpp fork: custom radix-select top-k kernel replacing cub::DeviceTopK (DSV4 prefill optimization) + ds4docs
%22%2F%3E%3Ctext%20x%3D%22320%22%20y%3D%22250%22%20font-family%3D%22ui-sans-serif%2Csystem-ui%2Csans-serif%22%20font-size%3D%22200%22%20font-weight%3D%22800%22%20fill%3D%22rgba(255%2C255%2C255%2C0.95)%22%20text-anchor%3D%22middle%22%3EW%3C%2Ftext%3E%3Ctext%20x%3D%22320%22%20y%3D%22335%22%20font-family%3D%22ui-monospace%2Cmonospace%22%20font-size%3D%2226%22%20fill%3D%22rgba(255%2C255%2C255%2C0.8)%22%20text-anchor%3D%22middle%22%3Evektorprime%2Fworking_ds4_speed%3C%2Ftext%3E%3C%2Fsvg%3E)