pub fn fp8_quant_grid(hopper: bool, m: u32, k: u32) -> [u32; 3]Expand description
Grid for one quantizer launch. PURE, so both arms are testable without a
GPU (fp8_act_quant_tests.rs).
M goes on grid X in both arms — its limit is 2^31-1, where grid Y stops at
65535, and MoE total_expanded exceeds 65535.
Shared arm: (M, K/128), one CTA per K-group, which is what the shared
kernel indexes with blockIdx.y.
Hopper arm: (M, ceil(K/128 / 8)). The kernel re-derives its own group span
as ceil(L / gridDim.y) rather than assuming 8, so this Y extent is a
PERFORMANCE choice and not a correctness contract — any Y in 1..=L covers
the same groups exactly once.