Record: Fused Triton MLP + Full GPTQ + Coprime Loader + XSA-all + BH2816 (val_bpb 1.1116)

[barneywohl]

Record Submission

Author: @barneywohl
Date: 2026-03-30
val_bpb: 1.1116 ± 0.0005 (3-seed mean)

Results (8×H100 SXM)

Seed	Sliding BPB	Artifact
1337	1.1110	15,982,859
42	1.1121	15,981,083
2024	1.1118	15,982,475
Mean ± Std	1.1116 ± 0.0005

Improvement over SOTA

• vs merged SOTA (1.1194): -0.0078 nats (requirement: ≥0.005) ✅
• vs PR Record: 1.1122 BPB — Coprime-Stride Loader + Full GPTQ + XSA-all (3-seed mean) #1060 (1.1122): -0.0006 nats

Stack

1. Fused Triton MLP — custom kernel for leaky_relu(x,0.5).square(), saves 1.8ms/step
2. Full Hessian GPTQ — Cholesky + actorder + 5-way clip sweep
3. Coprime-stride loader — multi-shard diversity with memmap
4. XSA on all 11 layers — exclusive self-attention everywhere
5. BigramHash(2816×112) — enlarged bigram features
6. fullgraph=True torch.compile

Built on PR #549 by @abaybektursun with techniques from PRs #726, #634, #1019, #287.

See records folder for full README, logs, and reproducible script.

[MatoTeziTanka]

Community Review — Record: Fused Triton MLP + Full GPTQ + Coprime Loader + XSA-all + BH2816 (val_bpb 1.1116)

BPB: 1.1116 | Compliance: LOOKS CLEAN — score-first-per-chunk TTT (legal #1416/#1423 pattern)

What I found in the code (head SHA fe30417220b1, file records/track_10min_16mb/2026-03-30_FusedMLP_FullGPTQ_CoprimeLoader_XSA11_BH2816/train_gpt.py):

The TTT path at line 1173 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 0.13s, dim=512, layers=11, vocab=1024, code=100911 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 0.13s, dim=512, layers=11, vocab=1024, code=100911 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.