qsa_qprep

Function qsa_qprep 

Source
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.