pub fn embed_rowdiff(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
base: DevicePtr,
served: DevicePtr,
flags: DevicePtr,
rows: u32,
h: u32,
thresh: f32,
stream: u64,
) -> Result<()>Expand description
2026-09-25: flags[r] = (max_i |base[r,i] - served[r,i]| > thresh), one
thread per row. base and served are [rows, h] BF16; flags is
[rows] u8.