perf(ds4): optimize q5 fused verifier on AMD

[davide221]

Summary

• add an opt-in q=5 fused verifier for the AMD heterogeneous DS4 path
• handle two ratio-4 compressor boundaries and exact partial-rejection rollback
• keep five-row dense projections on MMVQ with qualified cache defaults
• extend the fused ROCmFP4 x4 dot path to q=5 as x4+1 on gfx1201
• add a self-contained exact-context qualification wrapper and clients

PR #565 is now merged; this branch is rebased directly on main and remains a draft while gated rejected candidates are cleaned up.

AMD validation

Hardware profile: R9700 target plus Strix Halo expert device, in-process expert parallelism, four active experts, 36 hot experts per layer, one request at a time, 128 output tokens, temperature 0.

All 25 requests in the 2K -> 4K -> 8K -> 16K -> post-16K 2K sequence produced the exact expected SHA-256:

0f785a7ffa406498aafb14553966eaed0f52220fed0f7cc016b66921d104d194

Context	prior AMD q=5 median	x4+1 hot-36 median	change
2K	64.212 tok/s	67.957 tok/s	+5.8%
4K	62.428 tok/s	65.927 tok/s	+5.6%
8K	59.267 tok/s	62.544 tok/s	+5.5%
16K	53.750 tok/s	56.335 tok/s	+4.8%
post-16K 2K	63.804 tok/s	67.458 tok/s	+5.7%

The two-second monitor observed a 30.561 GiB R9700 peak during the full hot-36 sweep. This expert budget is a hardware profile, not a portable default for smaller cards.

The final test_deepseek4_unit GPU suite completed with OK. The sweep left MMVQ width, verifier slots, and x4+1 on auto; logs confirmed width 5, nine slots, and the gfx1201 x4+1 default.

Reproduction

Run server/scripts/qualify_ds4_q5_amd.sh with TARGET_MODEL, DRAFT_MODEL, and HOTNESS_CSV. It records the manifest, exact response hashes, timings, ROCm state, and VRAM trace.

Measured A/B decisions

• retained x4+1: 67.420 tok/s hot-32 median with exact hashes
• rejected full five-column reuse: 67.217 tok/s median
• rejected forced graph replay: +0.05%, within noise
• retained hot-36 hardware profile: 67.902 tok/s isolated 2K median before the full sweep

Before marking ready

• remove or split any remaining gated kernel candidate that did not improve the full-model benchmark
• complete final review of the rebased diff against main