pub fn topk_gpu(
device: &Device,
queue: &Queue,
logits: &[f32],
k: usize,
block_size: usize,
) -> Vec<TopKItem>Expand description
Reduce logits to its global top-K on the GPU. block_size is elements per workgroup
(clamped to TOPK_BLOCK_SIZE, the WGSL var<workgroup> cap). Blocking (native readback).