pub fn moe_expert_down_act_fp8_grouped(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
geometry: Fp8GroupedGeometry,
act: DevicePtr,
down_w: DevicePtr,
down_s: DevicePtr,
output: DevicePtr,
expert_offsets: DevicePtr,
active_experts: DevicePtr,
active_count: DevicePtr,
sh_act: DevicePtr,
sh_down: &Fp8Weight,
sh_down_out: DevicePtr,
n: u32,
k: u32,
cap: u32,
num_tokens: u32,
stream: u64,
) -> Result<()>Expand description
2026-09-26: Grouped FP8 down over the SiLU product, launched with geometry
(FP8_GROUPED_DOWN_SCALAR reads FP32 products, FP8_GROUPED_DOWN_TC BF16).
k is the intermediate width; rows of act and output are sorted
positions, rows of sh_act and sh_down_out tokens.