Non-record: QNA + SQWA compression thesis (8xH100 SXM)

[Abhishek8108]

Summary

Non-record submission documenting a controlled 3-run ablation on 8xH100 SXM testing whether training for the quantized artifact directly improves post-quantization BPB.

• Quantization Noise Annealing (QNA): Inject uniform noise matching int6 quant error during training, scaled by LR. Teaches weight robustness to quantization from the start.
• Stochastic Quantized Weight Averaging (SQWA): Average quantize-dequantized EMA snapshots during warmdown, ensuring the final model lives in the quant-friendly subspace.

Results

Run	Config	Sliding Window val_bpb	Quant Gap (float→int6)	Artifact Size
1	Baseline	1.1216	0.0081	15.93 MB
2	+QNA	1.1222	0.0075	15.95 MB
3	+QNA+SQWA	1.1258	0.0028	16.15 MB

Interesting negative result: Both techniques work mechanistically — QNA reduced quant gap by 7%, SQWA by 65%. But neither improved the final leaderboard metric. The bottleneck in current SOTA is float model quality, not quantization error.

Key takeaway

With a baseline quant gap of only ~0.008 BPB, the existing late QAT already handles quantization well enough. Future improvements should target the float model directly rather than compression alignment.

Files

• train_gpt.py — full script with QNA/SQWA (env-var toggleable, defaults off)
• run1_base.log, run2_qna.log, run3_qna_sqwa.log — complete training logs
• README.md — detailed write-up with implementation details and analysis
• submission.json — metadata

[MatoTeziTanka]

Community Review — Non-record: QNA + SQWA compression thesis (8xH100 SXM)

BPB: 0.008 (cache parse — may be delta/std, not val_bpb; check PR title) | Compliance: LOOKS CLEAN — score-first-per-chunk TTT (legal #1416/#1423 pattern)

What I found in the code (head SHA fe607e10da11, file records/track_non_record_16mb/2026-03-27_QNA_SQWA_CompressionThesis_8xH100/train_gpt.py):

The TTT path at line 1044 implements the score-first-per-chunk pattern: each chunk is scored under torch.no_grad() / inference_mode() before the base_model.train() + SGD adaptation runs on that same chunk, with an is_last_chunk guard so the final chunk gets no adaptation pass. This is the structural shape the legal frontier uses (PRs #1416 erichroepke, #1423 aryanbhosale).

Per Issue #402 and Issue #677, TTT is legal when each token is scored before the adapter updates on it, and that's what the code does here — chunk ci is scored under weights adapted only on chunks 0..ci-1. No prequant_ttt_adapt_adamw(val_tokens, ...) multi-epoch fine-tune, no scored-region SLOT, no target-in-key n-gram cache.

CPU smoke test (CT2038 proteus-engine, 2026-04-11): import OK in 0.06s, dim=512, layers=11, vocab=1024, code=89824 B, SMOKE_TEST_PASS

Verdict: LOOKS CLEAN.

Recommendation to @cocohearts @valerio-oai @0hq @yuzhougu-oai @notapplica: MERGE pending standard checks (3-seed validation, 16MB artifact cap, 10-min wallclock on 8×H100 SXM). The compliance picture matches the legal reference frontier and no flags were raised by the classification pass.

Auto-classification caveat: this review was drafted by the AST-based classifier against a template derived from manually-reviewed cluster PRs (#1420, #1450, #1487, #1541, #1529, #1533, #1518). If I've misread a subtlety in your eval path — e.g., multi-epoch TTT that I mistook for single-pass, or a target-in-key lookup I missed in a helper function — 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 0.06s, dim=512, layers=11, vocab=1024, code=89824 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.