[bri-prism]

What

• Opt-in split-K for the PQ2_0 one-column ffn_down mat-vec on Hopper, behind GGML_CUDA_PQ2_SPLITK=1 (default off).
• The PQ2_0 multi-column mat-vec is parameterized by rows per group, warps per CTA, split-K factor and unroll; the default instantiation keeps the existing loop and reduction order.

Why

• At K x M = 12288 x 4096 and 17408 x 5120 the one-column launch leaves warps idle; two warps per row over halves of K keep more of them busy.

How

• With the switch on, each row's K range is split between two warps whose partial sums are added in warp order through shared memory. This changes the float summation order, so output is not bit-identical (mean KLD about 2e-4 against the default, at the level of ordinary batching differences). Default off: launch path and results are unchanged.

Testing

• Switch unset or 0: the SASS of all default instantiations is byte-identical to the base build, and greedy output is identical on 6 prompts on two PQ2_0 models.
• test-backend-ops MUL_MAT passes with the switch unset and on, plus 4 extra cases on the split-K shapes.
• H200 tg128 with the switch on, alternating runs: +2.8% and +2.6% on the two models.