pub const DECODE_GQA_PACK_HEAD_DIM: u32 = 256;Expand description
The head dim the packed kernels are compiled for.
Same reason as the split-K siblings’ crate::attn_splitk head-dim pin:
the sources fix PD_HDIM at 256 and derive every lane’s element count from
it, so a 128- or 512-wide head would have every lane stride past its own
row.