pub const FP8_QUANT_HOPPER_GROUPS_PER_CTA: u32 = 8;Expand description
K-groups one Hopper CTA covers: 128 threads / 16 threads per group. SSOT
with HQ_GROUPS_PER_CTA in the .cu; the kernel derives its own span from
gridDim.y, so this value only has to be the one that fills the block.