pub fn qsa_qprep_rows(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
qk: DevicePtr,
q_norm_w: DevicePtr,
q_out: DevicePtr,
rows: u32,
first_pos: u32,
qkw: u32,
n_heads: u32,
hd: u32,
rot: u32,
theta: f32,
eps: f32,
stream: u64,
) -> Result<()>Expand description
Stage 2: per-row q prep for a contiguous selective row range.