perf(deepseek4): fix DSpark fused-verify wide-q performance cliff on gfx1151

[DeanoC]

Fixes the DSpark fused-verify wide-q performance cliff on RDNA / gfx1151 for DS4-Flash. Single concern (the cliff), split into two commits:

1. perf(deepseek4) — boundary-align the fused-verify width (deepseek4_dspark_spec.cpp): cap the verify batch to the next ratio-4 compression boundary so the boundary sits at the batch's LAST token, which the fused-verify graph's fast path assumes (deepseek4_fused_verify.inc). A mid-batch boundary is correct but ~3× slower in compute.
2. perf(cuda-hip) — MMVQ default on HIP (ggml-cuda.cu): default LUCE_MMVQ_MAX_NCOLS to the MMVQ batch ceiling (MMVQ_MAX_BATCH_SIZE=8). The prior default (3) is the measured sm_86 crossover; on RDNA the MMQ / dequant→rocBLAS-GEMM path is ~3× slower than MMVQ at small spec-verify batches, so a q=4 batch fell off MMVQ into rocBLAS HGEMM at M=4.

The two only remove the cliff together, but each is independently correctness-preserving and non-regressing.

Benchmark — before/after

config	q=4 verify	q=4 tok/s
before	311 ms	8.5
+ boundary-align (commit 1)	206 ms	9.1
+ MMVQ HIP default (commit 2)	107 ms	16.6

q=4 verify 311 → 107 ms (2.9×); wide-q now scales ~linearly (q3 = 89 ms, q4 = 107 ms) instead of collapsing. The default adaptive greedy path (q≤3, sweet spot q3 = 18.0 tok/s) and the AR single-token path are unchanged.

Methodology

• Hardware: gfx1151 (Strix Halo), ROCm 7.2.2, default power/clocks, full 128 GB unified-memory pool (GGML_CUDA_ENABLE_UNIFIED_MEMORY=1).
• Workload: DS4-Flash-ROCMFP2 + DSpark drafter, real decode path (DFLASH_DS4_SPEC=1 DFLASH_DS4_FUSED_VERIFY=1), greedy (temperature=0), 127-token generation, identical prompt across configs.
• Width control: forced-q sweep via /tmp/ds4_spec_q (adaptive width off) so each config is measured at the same q.
• Timing: built-in DFLASH_DS4_TIMING=1 per-phase cumulative means ([ds4-spec-t] TOTAL), one warmup request discarded, single loaded server per sweep (weights identical/warm across configs).
• Kernel attribution: rocprofv3 --kernel-trace (see below).

Why it was a cliff (rocprofv3 --kernel-trace)

At q=4 the dense MLA attention projections crossed the MMVQ→MMQ threshold and dispatched to rocBLAS FP16 GEMM at M=4 with 128×128 tiles: Cijk_*HB_MT128x128x32 (~34% of decode) + dequantize_block_rocmfp4_fast (13%) dominated the trace. Keeping them on MMVQ (which reads each dense weight once for all q columns) removes both. The width fix additionally keeps the ratio-4 compressor pool at the batch's last token.

Correctness

• Perf-only changes: commit 1 only narrows the speculation width; commit 2 only changes kernel selection. Neither touches numerics.
• Greedy output is bit-identical across all three configs above (same 127-token completion), verified on the DSpark decode path on gfx1151.

Notes

• Base is base/ds4-dspark-435dd26 (the DSpark branch commit) so the diff is only these two changes.
• Licensed under Apache-2.0; reproducible from public sources (ROCm + this tree, no closed-source deps).

Follow-up q-sweep — current operating point (improved draft accept, 2026-07-14)

Re-measured on the current DS4 baseline with the improved draft (higher accept rate), gfx1151, forced-q (/tmp/ds4_spec_q, adaptive off), greedy, 127-token, 2 prompts:

q_cap	verify ms	mean accept	tok/s
2	~81	~0.83	~18.3
3	~109	~1.46	~19.2
4	~138	~1.94	~18.5

q=3 remains the static optimum (~19.2 tok/s), and q=4 is now competitive (~18.5) — no longer the pre-fix cliff — confirming the wide-q path is healthy at the current operating point. q=4 still edges below q=3 on average because its verify cost grows faster than the extra accepted tokens (56 vs 52 ms/tok), though it pays off on individual high-accept steps (18.0 tok/s @accept 1.82 → 19.1 @accept 2.05).

A dynamic confidence-gated width (q=4 only on high-confidence steps) was explored to capture those steps but does not beat static q=3: the DSpark confidence head is currently over-confident and over-selects q=4, and boundary-alignment alone can't offset q=4's higher per-token verify cost. Recommendation: static q=3 pending confidence-head recalibration (a model-side follow-up).