Function moe_fp8_grouped_gemm 

pub fn moe_fp8_grouped_gemm(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    input: DevicePtr,
    weight_ptrs: DevicePtr,
    scale_ptrs: DevicePtr,
    output: DevicePtr,
    expert_offsets: DevicePtr,
    sorted_token_ids: DevicePtr,
    num_experts: u32,
    n: u32,
    k: u32,
    worklist: DevicePtr,
    total_tiles: DevicePtr,
    max_tiles: u32,
    stream: u64,
) -> Result<()>

2026-09-25: FP8 grouped GEMM for sorted MoE prefill over the work-list from moe_build_tile_worklist, which must run first on the same stream.

The kernel strides over the work-list by gridDim.x, so the grid is max_tiles (the caller’s upper bound on the tile count) clamped to 1..=16384. Extra CTAs exit at once; a grid smaller than the list is slower but still covers every tile.

input [total_tokens, K] BF16; weight_ptrs[e] points at [N, K] FP8 and scale_ptrs[e] at [N/128, K/128] FP32; output [total_expanded, N] BF16; expert_offsets [num_experts + 1]; sorted_token_ids [total_expanded], or NULL to read input row i for output row i.