QAT x SWA Ablation: SWA sabotages QAT (-3.64 mBPB, 3-seed validated)

[alexanderaperry-arch]

Leaderboard-relevant ablation: SWA and QAT are antagonistic

Systematic 2×2 factorial (QAT on/off × SWA on/off) on the PR #180 stack. 3-seed validated on 8xH100, all runs under 10 min and 16MB.

Result

Config	QAT	SWA	Mean BPB (3 seeds)	Delta vs Control
no_swa_qat	Yes	No	1.14018	-3.64 mBPB
control	No	Yes	1.14382	baseline
qat_snap70	Yes	Yes	1.14468	+0.86 mBPB
no_swa	No	No	1.14486	+1.04 mBPB

Why this matters

Every QAT submission in this competition (#117, #139, smeargate_ortho) also used SWA — and every one underperformed non-QAT entries. Our ablation shows why: SWA's checkpoint averaging dilutes the quantization-boundary alignment that QAT works to achieve. Combining them is worse than either alone.

The fix is simple: remove SWA when using QAT. This alone yields ~3.6 mBPB.

Actionable for competitors

• If you're using SWA + QAT together, drop SWA
• QAT alone is 3.5x more effective than SWA alone for quantization quality
• Training val_bpb is misleading for QAT — post-quantization BPB is the metric that matters
• QAT weights need ~10% magnitude pruning (vs 3%) to fit under 16MB — they compress worse

Open question

Top entries now use EMA instead of SWA. Nagel et al. (2022) proposes EMA to stabilize QAT, but our short-horizon results suggest averaging mechanisms in general may conflict with QAT under tight wallclock constraints. EMA × QAT interaction is untested.

Validation

• Seeds: 42, 1337, 2024 (std ±0.00044)
• Artifact: 15,787,003 bytes (158KB under limit)
• Hardware: 8xH100 SXM, 600s wallclock cap
• Based on PR Record: 10L Int5-MLP + BigramHash(10240) + SWA(0.4) + WD=0.04 (val_bpb=1.1428, mean 3 seeds) #180 (10L/512d/MLP3x, int5-MLP/int6-attn, zstd-22)

[MatoTeziTanka]

Community Review — QAT x SWA Ablation: SWA sabotages QAT (-3.64 mBPB, 3-seed validated)

BPB: 1.1402 | Compliance: LOOKS CLEAN — pure-neural submission, no TTT/SLOT/n-gram-cache

What I found in the code (head SHA ae1d2c25a4cf, file records/track_10min_16mb/2026-03-28_QAT_SWA_Ablation/train_gpt.py):

Static code review found no TTT adaptation function, no SLOT optimization loop, no n-gram-cache class, and no pre-quant val-token fine-tune. The eval path uses the standard sliding-window stride-64 pattern. The submission is a pure-neural architecture iteration on the standard SP1024/SP4096/SP8192 baseline.

CPU smoke test (CT2038 proteus-engine, 2026-04-11): import OK in 0.03s, dim=512, layers=10, vocab=1024, code=54948 B, SMOKE_TEST_PASS

Verdict: LOOKS CLEAN.

Recommendation to @cocohearts @valerio-oai @0hq @yuzhougu-oai @notapplica: MERGE pending the usual record-track checks (3-seed validation, under-16MB artifact cap, ≤600s train + ≤600s eval on 8×H100 SXM). No compliance flags from the classification pass — this looks like a clean pure-neural iteration on the standard baseline.

Auto-classification caveat: this review was drafted by the AST-based classifier. If there's a non-standard eval mechanism (logit postprocessing, hedge mixing, etc.) that I missed because it's factored into a helper file or a non-standard function name, please flag it and I'll re-run the audit manually.

---

Reviewed by @MatoTeziTanka — The Agora. CPU smoke test (CT2038 proteus-engine, 2026-04-11): import OK in 0.03s, dim=512, layers=10, vocab=1024, code=54948 B, SMOKE_TEST_PASS. Classification via deterministic AST-based classify_prs.py (pattern bank derived from ~65 manually-reviewed PRs earlier in the 2026-04-11 sweep). This review was auto-drafted from a template and spot-checked before posting — if the template misread your code, please call it out so I can iterate the classifier.