Date: 2026-07-16
Status: bitwise exact; isolated positive; not promoted
Keep the promoted MXFP4 N64 policy. N32 is a measured regression. N128 is bitwise exact and consistently faster in the isolated four-card M=2 hardware gate, but two strict candidate suites do not robustly exceed the existing 63.349928 tok/s record. No LocalMaxxing submission was made.
This is not a model, topology, speculative-depth, cache, or prompt-policy change. The frozen K160 TP4+EP, one-active-generation, target-verified MTP1 identity was retained throughout.
The exact eager-cycle profile attributes 4.1511 ms per MTP1 cycle to the routed MXFP4 verifier family. Earlier N32/N128 work had tested M=1 and exposed a graph-replay correctness race: workgroup 0 reset a global scheduler counter while other workgroups could already increment it. That history did not answer the M=2 verifier question.
XPU-kernel experiment commit
351a06a4426b77eed12f555d01da8ea186d0dc7b moves the N32/N128 reset to an
ordered queue fill before launch. The experiment is preserved on branch
codex/deepseek-v4-m2-mxfp4-policy; the promoted source remains
d15ce87d07376be53ea2d6f7ae0262ab79f7cb7b on
codex/deepseek-v4-m2-router-norm.
The reusable probe is ../scripts/probe-mxfp4-m2-policy.py; tracked raw
summaries are under
../../../data/deepseek-v4-reap-mxfp4-m2-policy-20260716/.
N32 passed all tested changed-route and graph-replay cases but regressed by 6.68-6.98 us per layer on the two sampled B70s, projecting a 0.287-0.300 ms loss across 43 MXFP4 calls. It was closed immediately.
N128 passed 48/48 bitwise-exact cases on every B70 across same-route, disjoint-route, cross-row-overlap, and adversarial route patterns. Against N64:
| Physical B70 | N64 median us | N128 median us | Projected ms saved / 43 |
|---|---|---|---|
| 0 | 219.902 | 213.342 | 0.2821 |
| 1 | 220.198 | 213.613 | 0.2831 |
| 2 | 219.136 | 213.403 | 0.2465 |
| 3 | 220.291 | 214.537 | 0.2474 |
| Mean | 219.882 | 213.724 | 0.2648 |
The isolated speedup is real and repeatable, but it clears only 53% of the 0.50 ms whole-cycle integration gate. It was integrated once because the profile identified MXFP4 as the largest genuinely open verifier family and the change was already graph-exact.
N128 candidate evidence:
/mnt/fast-ai/bench-results/deepseek-v4-flash-xpu/mtp1-m2-mxfp4-n128-candidate-20260716T0630Z
Same-binary N64 control evidence:
/mnt/fast-ai/bench-results/deepseek-v4-flash-xpu/mtp1-m2-mxfp4-n64-control-20260716T0645Z
| Policy/run | Median tok/s | p10 tok/s | Mean tok/s |
|---|---|---|---|
| N128 screen | 62.649706 | 59.577779 | 62.727082 |
| N128 confirmation | 63.628477 | 59.481507 | 62.986086 |
| N64 control 1 | 61.205692 | 58.893436 | 62.067179 |
| N64 control 2 | 63.101865 | 58.749330 | 62.641414 |
All 48 requests are fresh, cached-zero, and pass the fixed realistic-suite gate. The two N128 medians average 63.139091 tok/s; the two N64 controls average 62.153779 tok/s. However, the N64 control span is 1.896 tok/s and the N128 screen is below both the 63.349928 record and its 62.882999 support suite. The single 63.628477 row is therefore insufficient to establish a new record.
Running more suites until one exceeds the record would be benchmark shopping. The correct classification is an exact small isolated improvement whose end-to-end effect remains inside observed service variance.
The apparent 6.5803 ms dense-GEMM bucket is now decomposed by exact shape. Its largest pieces are already optimized or have a preserved negative result:
wo_a BF16 BMM; the earlier FP8 replacement lost end to end;This means the next useful lane is not another dense flag or small-N tile sweep. It must change the M=2 grouped-MXFP4 architecture enough to remove at least 0.50 ms per verifier cycle, for example by reducing scheduler/gather traffic or specializing the two-row route structure. It must first pass a four-card real-shape upper-bound gate before another TP4 service load.