pub fn qsa_qprep(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
q_in: DevicePtr,
q_norm_w: DevicePtr,
q_out: DevicePtr,
n_heads: u32,
hd: u32,
rot: u32,
pos: u32,
theta: f32,
eps: f32,
stream: u64,
) -> Result<()>Expand description
One decode query: per head, RMSNorm*(1+w) + partial rope at pos -> FP32.