pub fn fp8_quant_grid(hopper: bool, m: u32, k: u32) -> [u32; 3]Expand description
2026-09-25: Grid for one quantizer launch. Pure, so
fp8_act_quant_tests.rs checks both arms without a GPU.
M is on grid X in both arms: X allows 2^31-1 blocks, Y only 65535.
Shared arm: (M, K/128, 1), one CTA per K-group; the kernel reads the
group from blockIdx.y.
Hopper arm: (M, max(1, ceil(K/128 / 8)), 1). The kernel derives its group
span as ceil(L / gridDim.y), so any Y in 1..=L covers every group
exactly once; the Y extent is a speed choice.