pub fn gated_rms_norm_prefill(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
input: DevicePtr,
gate: DevicePtr,
weight: &DenseWeight,
output: DevicePtr,
heads_per_token: u32,
head_dim: u32,
eps: f32,
num_actual_tokens: u32,
input_token_stride: u32,
gate_token_stride: u32,
stream: u64,
) -> Result<()>Expand description
Batched gated RMS norm for prefill: all (head, actual_token) pairs in one launch.
Grid: (heads_per_token, num_actual_tokens, 1) Block: (min(head_dim, 1024), 1, 1)