New SOTA: 1.12676 BPB - 11L XSA-all(11) + GPTQ-lite + EMA + Late QAT

[gowtham0992]

New SOTA Record: val_bpb 1.12676 (3-seed mean)

Beats current SOTA (1.14276) by 0.016 nats.

3-Seed Results (8xH100 SXM, 600s)

Seed	BPB	Size
42	1.12713	15.64 MB
1337	1.12648	15.62 MB
2024	1.12667	15.88 MB
Mean	1.12676	~15.7 MB

Key Techniques

• XSA on ALL 11 layers
• GPTQ-lite optimal clip percentile search
• EMA(0.997) + Tight SWA
• Late QAT int6-all at LR scale < 0.15
• Raw binary + zstd22 serialization

Dependencies

• zstandard, flash_attn_3 (see requirements.txt)

Verified on RunPod 8xH100 SXM (official template): 1.12753 BPB

See README.md for full details.

[mohosy]

xsa on all 11 layers is bold, most ppl only do last 3 or 4. does it actualy help on the early layers too or is it just not hurting? also gptq lite clip search is a nice touch i havent seen anyone else do that yet

[gowtham0992]

xsa on all 11 layers is bold, most ppl only do last 3 or 4. does it actualy help on the early layers too or is it just not hurting? also gptq lite clip search is a nice touch i havent seen anyone else do that yet

thanks! yeah xsa on all layers actually helps, not just not hurting. its ablated to:

XSA-all(11): 1.12676 BPB, 6764 steps, 88.7ms/step XSA(4) last only: 1.13266 BPB, 6998 steps, 85.7ms/step

so ~0.006 BPB win even though its 3ms/step slower and we lose ~230 steps. early layers tend to repeat self-value patterns, xsa forces them to actually encode new info. at 11L 512d every layer counts

gptq-lite is free basically. 5 clip percentiles per row, pick min MSE. adds like 2 seconds to save and gets ~0.0006 BPB back from quant gap.

[MatoTeziTanka]

Community Review — New SOTA: 1.12676 BPB - 11L XSA-all(11) + GPTQ-lite + EMA + Late QAT

BPB: 1.12676 | Compliance: LOOKS CLEAN — pure-neural submission, no TTT/SLOT/n-gram-cache

What I found in the code (head SHA 07884e6ce1c4, file records/track_10min_16mb/2026-03-22_11L_XSA-all_GPTQ-lite_EMA_LateQAT_1.1271/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 0.08s, dim=512, layers=11, vocab=1024, code=53472 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 0.08s, dim=512, layers=11, vocab=1024, code=53472 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.