File size: 4,237 Bytes
21bbd75 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 | # Executive summary
## Conclusion
On a fixed governance stack on one RTX 5070 Ti, we measured what
quantization regime — QAT Q4_0 versus PTQ Q4_K_M — does to governed
routing quality, in a 2×2 design over Gemma 4 12B (dense; a clean
same-base pair) and a 26B-class MoE released pair (25.8B PTQ vs 25.2B
QAT; regime confounded with a possible base revision). The answer: **on
this stack the QAT premium is not a scalar — it differs in sign across
the two released pairs** — and regime choice is a *behavioral* change
even where quality scores barely move.
| Cell (Core-500, N=500) | route_correctness | mean latency |
|---|---:|---:|
| 12B dense, PTQ Q4_K_M | **0.760** | 2,482 ms |
| 12B dense, QAT Q4_0 | 0.748 | 2,626 ms |
| 26B MoE, PTQ Q4_K_M | 0.738 | 3,252 ms † |
| 26B MoE, QAT Q4_0 | 0.756 | **1,613 ms** |
† the 26B PTQ artifact does not fit the 16 GB card and runs ~26/74
CPU/GPU-split; its latency reflects offload and is excluded from all
conclusions.
- **Within-pair regime effects are marginal and oppositely signed**:
12B dense −0.012 under QAT (McNemar p=0.070); 26B pair +0.018
(p=0.078). Smoke-100 agrees in sign for both.
- **The interaction — an exploratory, single-run headline — survives
clustering-robust analysis**: delta-of-deltas +0.030, stem-clustered
bootstrap 95% CI [+0.010, +0.052] (Core-500 is 100 stems × 5
paraphrases; naive task-iid CI [+0.008, +0.052]). Aggregate averaging
cancels the two effects and hides the structure. No equivalence test
was run on the cell aggregates; one cross-model contrast (12B-PTQ vs
26B-PTQ) is nominally significant (p=0.035).
- **The interaction concentrates in the volatile-current family**, whose
four cells form a crossed pattern (12B: 0.5314→0.4971; 26B:
0.4971→0.5314; the numeric identity of the crossed counts is
coincidence, not a tested effect).
- **Safety-critical rates: 0.000 in 7 of 8 runs**; one over-verification
event in 26B-PTQ Core-500 (0.008). With events this rare the data
bound, rather than establish, regime-invariance.
- **Regime is a behavioral change**: only 23–32 % of same-task
temperature-0 outputs are byte-identical across regimes (median first
divergence 46–73 characters), against same-configuration repeat
baselines of **100/100 in all four cells** — including the
CPU-offloaded PTQ MoE cell, across run windows — and a 100/100
`num_ctx` byte-identity control on the 12B side. Hash-based caching,
audit replay, and golden-file tests do not survive a
quantization-regime swap.
## Practical implications
1. Choose quantization regimes per released artifact pair and per task
family, not by recipe reputation; aggregate "quality parity" can
conceal offsetting family-level shifts.
2. On this stack, the QAT 26B artifact is best-or-tied on the two hardest
families while being the fastest memory-fitting cell — consistent with
its same-day adoption as a production binding (a partially circular
check: two of the four cells are that decision's own runs; the
anti-diagonal is the independent evidence). The dense-tier fallback is
better served by the **PTQ** 12B (clean-pair QAT effect ≤ 0).
3. Treat any quantization swap as a behavioral migration with full
re-validation, not a drop-in.
## Honest limits
The 26B released pair may embed a base revision (25.8B vs 25.2B) and its
expert configuration is unverifiable from runtime metadata; the PTQ 26B
cell ran partially on CPU — so regime, revision, and compute path are
confounded on that side; the 12B pair is clean. All inferential results
are exploratory: no preregistration, single run per cell, ≥5 inferential
quantities without multiplicity correction, paraphrase-clustered rows
(effective N nearer 100 than 500). One GPU, one runtime (Ollama 0.30.6),
one governance stack, routing tasks only; answer-content quality is
unmeasured. The general "QAT beats PTQ" direction is vendor-documented
prior art; the intended contribution is the governed-routing task class,
the two-scale paired design exposing the signed interaction, and the
regime-to-regime exact-output divergence measurement (the last is
plausibly adjacent to community cross-quantization regression reports).
|