pub fn dense_gemm_ba_gates_prefill_hopper(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
input: DevicePtr,
ba_weight: &DenseWeight,
a_log: DevicePtr,
dt_bias: DevicePtr,
gate_out: DevicePtr,
m: u32,
n: u32,
k: u32,
k_stride: u32,
gate_stride: u32,
nv: u32,
vheads_per_group: u32,
stream: u64,
) -> Result<()>Expand description
Launch dense_gemm_ba_gates_prefill_hopper — one CTA per token.
Same ABI as the parent, deliberately: the two differ in grid and in
nothing else a caller can see, which is what lets ba_gates_pick return a
handle rather than a closure.