Activation-profile selection: frozen Variant 1
Decision: Use the exact frozen Variant 1 corpus/profile pair for the Leanstral MXFP4 W4A8 release candidate.
Status: Final development-selection decision. The comparison was repeated across five independent, normally autotuned cold starts on the pinned clean SGLang/FlashInfer runtime. A constrained V1-centered composite was then built, calibrated as one coherent profile, and rejected by its frozen central-quality margins.
This closes calibration-data iteration for the current release. It does not replace the later untouched-quality or end-to-end performance gates.
Final conclusion: Under normal FlashInfer startup variation, V1 consistently provides the best ordinary predictive preservation. The available KL-tail improvements are real, but neither V5 nor a targeted V1-centered composite achieves them without an unacceptable central-quality cost.
What the experiment selected
The experimental variable was the 512-record calibration corpus and its resulting static per-tensor FP8 activation profile. All candidates shared:
- the same packed Leanstral MXFP4 weights and UE8M0 K/32 weight scales;
- the same static FP8 E4M3 execution contract;
- the same 100-example iterative-development pack;
- one dynamic reference within each cold start; and
- the same pinned SGLang, FlashInfer, seed, and attention backend.
The study selects the exact frozen V1 corpus/profile pair listed below. V1–V5 were evaluated as fixed profiles on the clean runtime; they were not recollected from scratch for each start. This is the correct experiment for choosing the release artifact, but it should not be described as proof that V1 is universally optimal under every possible calibration run or tactic realization.
The validation pack informed V5 and V6, so this is iterative development validation rather than an untouched final test. No scale values were spliced between profiles: each composite was recalibrated as one complete profile.
Clean replicated comparison
Each of five independent starts used a fresh ephemeral container and normal FlashInfer autotuning. All profiles were compared against the same dynamic reference within that start. The five starts, not the 500 repeated examples, are the independent units.
Central preservation
Values are mean ± sample standard deviation across five starts.
| Profile | Candidate NLL | Signed NLL delta | Mean absolute sample delta | Top-1 | Top-5 | Reference top-1 retained |
|---|---|---|---|---|---|---|
| V1 | 13.93530 ± 0.00145 | 0.04799 ± 0.00282 | 0.08660 ± 0.00144 | 0.620 ± 0.016 | 0.5316 ± 0.0059 | 0.756 ± 0.021 |
| V2 | 13.96571 ± 0.00128 | 0.07841 ± 0.00219 | 0.12148 ± 0.00181 | 0.578 ± 0.013 | 0.5472 ± 0.0033 | 0.760 ± 0.007 |
| V3 | 13.96561 ± 0.00129 | 0.07831 ± 0.00281 | 0.09765 ± 0.00085 | 0.550 ± 0.029 | 0.5304 ± 0.0070 | 0.714 ± 0.021 |
| V4 | 14.02996 ± 0.00105 | 0.14266 ± 0.00246 | 0.13983 ± 0.00045 | 0.550 ± 0.012 | 0.5328 ± 0.0054 | 0.710 ± 0.017 |
| V5 | 13.97669 ± 0.00120 | 0.08938 ± 0.00195 | 0.09480 ± 0.00038 | 0.574 ± 0.009 | 0.5456 ± 0.0036 | 0.736 ± 0.022 |
V1 won candidate NLL, signed NLL delta, mean absolute per-example NLL delta,
top-1 agreement, and top-1 flip count in all five starts. Relative to V5, V1
averaged 0.04139 lower NLL, 0.00820 lower absolute sample delta, and 4.6
more top-1 agreements per 100 examples.
KL distribution and severe tail
| Profile | Mean KL | Median KL | p90 KL | p95 KL | Maximum KL | Mean count >1 | Mean count >5 |
|---|---|---|---|---|---|---|---|
| V1 | 0.30680 ± 0.01464 | 0.06017 ± 0.00517 | 0.33877 ± 0.03457 | 0.75950 ± 0.21981 | 7.63164 ± 0.02444 | 4.2 | 3.0 |
| V2 | 0.29281 ± 0.03467 | 0.06050 ± 0.00167 | 0.33103 ± 0.01065 | 0.96258 ± 0.40017 | 5.85584 ± 0.01128 | 5.2 | 2.8 |
| V3 | 0.40588 ± 0.02479 | 0.06096 ± 0.00348 | 0.49477 ± 0.02139 | 1.45338 ± 0.67184 | 8.79102 ± 1.16882 | 5.6 | 3.2 |
| V4 | 0.30840 ± 0.02926 | 0.04335 ± 0.00220 | 0.40978 ± 0.02723 | 1.25240 ± 0.35561 | 5.95024 ± 0.13142 | 6.4 | 2.0 |
| V5 | 0.24760 ± 0.02489 | 0.05946 ± 0.00449 | 0.40532 ± 0.02252 | 0.70774 ± 0.07174 | 5.11108 ± 0.06165 | 3.4 | 1.4 |
V5's severe-tail advantage is real, not noise from a single restart. It won
mean and maximum KL in all five starts, improved all ten recurring V1 tail
failures in at least four starts, and reduced the recurring category-E rescue
substantially. It also introduced stable failures of its own and lost V1's
central NLL/top-1 behavior in every start. V1 and V5 tied on the mean number of
examples above KL 0.5 (7.2). V1 won p95 in three starts and V5 in two.
Mean KL was therefore not used as a soothing single-number selector. Central preservation, tail distributions, recurring named failures, categories, lengths, routing, fallback, and range diagnostics were kept separate.
Constrained composite test
Variant 6 tested whether V5-like D/E tail improvements could be added without giving up V1's central quality. It retained 464 V1 records and made 48 deterministic D/E short-or-medium substitutions from V5. It was recalibrated as one coherent 512-record profile.
Across five fresh normally autotuned starts:
| Profile | Candidate NLL | Top-1 | Mean KL | Maximum KL | p95 KL |
|---|---|---|---|---|---|
| V1 | 13.93502 | 0.618 | 0.30513 | 7.64528 | 0.66222 |
| V5 | 13.97607 | 0.564 | 0.23958 | 5.61991 | 0.64364 |
| V6 | 13.96375 | 0.576 | 0.24836 | 5.95601 | 0.72812 |
V6 improved mean KL in every start, but its NLL was worse than V1 in every
start and its top-1 agreement was lower in every start. Its V6-minus-V1 NLL
mean was +0.028728; the one-sided 99% lower paired-t bound was +0.025366,
above the frozen +0.01 rejection margin. Its top-1 mean was -0.042; the
one-sided 99% upper bound was -0.020152, below the frozen -0.01 margin.
Starts 2–5 alone reach the same rejection. V6 was rejected and V1 retained.
The central loss margins were frozen before the comparison; the reject-only
futility look was formalized after start 1, so it is not presented as a formal
study-wide type-I-error guarantee. The independent starts 2–5 sensitivity makes
the practical decision unchanged.
Routing, fallback, and range interpretation
Expert coverage was high for all profiles. Across the replicated study, V1
covered a mean 4,504 / 4,608 layer/expert pairs (97.7431%); V5 covered
4,496.2 / 4,608 (97.5738%). Fallback routing exposure was below 3.5e-7
of assignments and did not explain the tail.
V1 averaged 1,525 validation-overflow-risk scale values and V5 averaged 1,332.2. This is a range diagnostic, not an objective. Lower risk counts did not imply better central model behavior, and full-mode maxima do not prove per-example clipping. Seam-adjacent coverage remains desirable when the behavioral evidence supports it.
Frozen receipts
source runtime:
SGLang bbb5702a3c7faba97ab425b15808cbe5f3ab6f7c
FlashInfer Python/cubin/JIT 0.6.14/0.6.14/0.6.14+cu130
runtime image sha256:f74a9c7d37ac724e30bc4841999e15a8fddf87c1e5bab2165041000879f4f152
Variant 1 corpus:
inputs/variant-1/calibration_samples.jsonl
sha256 57e1e10fce8a6a62cef14e58e26109065104fe109c4075c6b6f302726918ed57
records 512
Variant 1 activation profile:
runs/mxfp4-calibration/variant-1/activation_profile.json
sha256 5480257595178836022477080bc6132e2accff2d6e26c959b092e57a0f911531
ModelOpt 0.37.0
keys 288
finite positive values 9,432
development-validation pack:
validation/validation.jsonl
sha256 8cdbf9cfcc2b9c1bcb305c14f979799be6a7b9efad54dd4e5a281f00c152ccdf
replicated V1-V5 aggregate:
runs/clean-upstream-main-study/gate2-fi0614/replicated-autotuned-profile-study/aggregate.json
sha256 2a20473486a7a84efb462aff12b28c2319ce86a45ca0ead2d74a919ced270324
Variant 6 corpus/profile:
corpus sha256 6b1a712d8ed90b7403b9ffbc3f71a73411eff5c920b3043a2ea17cb255e11856
profile sha256 7f426c18d4ee09d07c4687fca82bce63f7bbdd9c427f230cefb9465213d569ce
Variant 6 futility report:
runs/clean-upstream-main-study/gate2-fi0614/constrained-composite-fresh-study/futility-report.json
sha256 2a1e17844838bf950eec3fe157a029132d608f67fcd811e0fd1bf8b363d0c717
The detailed reports are docs/replicated-autotuned-profile-study.md and
docs/constrained-v1-composite-result.md.
Scope of the decision
This evidence is sufficient to freeze V1 for the release candidate. It does not yet establish final generalization, acceptable untouched-pack quality, or an end-to-end speed advantage over W4A16. Those are later release gates and must use the clean pinned runtime. Reopen corpus design only if new evidence identifies a reproducible release-blocking failure that the frozen V1 artifact cannot tolerate.
The result also closes simple record-substitution composite development. A new attempt would need a genuinely different mechanism—such as layer-specific calibration, constrained scale optimization, or another activation-quantization scheme—not another unprincipled mixture of the same records.