pub const DECODE_GQA_PACK_WIDTH: u32 = 6;Expand description
Query heads one packed CTA carries.
A COMPILE-TIME shape on both sides: the kernels size q_reg/o_reg
register arrays from #define PD_GQA, because a runtime bound would put
them in local memory and defeat the change. This constant is the Rust
spelling of that #define, and
cuda_sources_declare_the_pack_width_rust_dispatches_on fails if the two
drift.
6 is the Qwen3.8-27B decode ratio (nq = 24, nkv = 4). A model with any
other ratio is REFUSED by gqa_pack_shape_ok and keeps the unpacked
kernel; it is not approximated.