[xyzzing]

Per the offer in the #944 thread: the gfx110x fp16 (rocWMMA) qsa-select scorer arm, as the gfx110x branch of the existing STRATA_SELECT_WMMA opt-in, alongside the bf16 hand-rolled kernel that 0.1.40 ships.

What it is. One commit: the fp16 rocWMMA scorer kernel for gfx110x (rocwmma.hpp, 16x16x16, fp16 operands / fp32 accumulate) behind the same STRATA_SELECT_WMMA env, plus its parity harness src/kernels/qsa_select_wmma_parity.cpp (fp64 oracle + the bitwise warp host model, pre-declared bounds, negative controls). The bf16 kernel (gfx12 + the other gfx11 layouts) and the warp kernel stay the fallbacks everywhere else — the default path (env absent) is byte-identical to main; on cards without the rocWMMA headers the arm compiles out and the host entry refuses.

Why fp16 here. Paired A/B on gfx1100 (2026-10-06, same engine build, arms differing only in the select kernel): the fp16 arm's qsa-select phase ran 1.64x faster than the shipped bf16 arm at 128K (measured on the fork's 0.1.40 refresh line; the fork carries this exact kernel in production). The re-land on this base measured parity-neutral overall vs our previous production reads (x1.0009 on the canonical stream tail).

Gates on this base. Full HIP build green; qsa_select_wmma_parity all cases PASS (worst fp64-oracle bound utilisation 11.0%, one-hot/coordinate-coded adversarial case bitwise vs its host model, 0 differing selection ids, negative control fails as designed).

Prepared with an AI engineering agent (GLM-5.3 / Z.ai) under human direction; every number is from our own recorded runs.

[Niko1221]

Thank you. The qsa-select phase is now covered by the exact thread scorer (#1943, bitwise equal scores, no rocWMMA dependency, 3.6-4x on the phase), and your own end-to-end number on this base is neutral, so we are not adding a rounded fp16 arm next to it. If the fp16 arm beats the exact scorer end to end on a gfx110x we would reconsider.

[xyzzing]

Measured on gfx1100 (build 0.1.45, 128K prompt, RX 7900 XTX): the fork's fp16 gfx110x arm runs the QSA select phase in 1,714 ms, the exact thread scorer in 4,320 ms, and the default radix in 9,859-10,440 ms. End to end prefill ties within noise (2,066 vs 2,055 tok/s) because the select is a small share of the 128K prompt; the 262K tier was not tested. Your condition to reconsider was an end-to-end win for the fp16 arm: the phase win is 2.5x but the total is a tie at 128K, so we accept the decision and close. The fork keeps the arm (carried since 700c026) and will re-test if the exact scorer's dispatch changes.