Record: Coprime-Stride Loader + Full GPTQ + XSA-all — val_bpb 1.1133 (3-seed mean)

[Bortlesboat]

Summary

• val_bpb: 1.1133 (3-seed mean, std 0.0001)
• Artifact: ~15.89 MB (all seeds under 16,000,000 bytes)
• Eval time: ~85s (no TTT, sliding window stride=64)
• Built on PR #549 by @abaybektursun and PR #1060 by @dexhunter

3-Seed Results

Seed	Sliding BPB	Artifact
1337	1.1133	15,899,687
42	1.1132	15,881,359
999	1.1133	15,892,371
Mean +/- Std	1.1133 +/- 0.0001

What's New

1. GPTQ Reserve Optimization: Reduced calibration reserve from 14s to 9s (actual calibration takes ~8.4s), recovering ~55 extra training steps
2. FA3/FA2 Graceful Fallback: try/except import for flash_attn_interface with fallback to flash_attn

Stack

• Coprime-stride multi-shard data pipeline (PR Memmap multi-shard data pipeline + GPU prefetch + LeakyReLU² + Legal TTT + Parallel Muon #726 style)
• Full Hessian GPTQ with Cholesky error compensation
• XSA on all 11 layers
• BigramHash(2816x112), SmearGate, Partial RoPE(16d), LN Scale
• EMA(0.997), Parallel Muon + Parameter Banking
• FA3 Hopper, ~91ms/step, ~6,500 steps

Compliance

• Standard F.cross_entropy scoring (softmax, sum=1)
• No TTT, no mixer, no eval-built adaptation
• Artifact < 16,000,000 bytes (all 3 seeds)
• Training < 600s, eval < 600s
• Causal sliding-window evaluation (stride=64)

See README.md for full details.

[Bortlesboat]

Superseded by #1169 (better score). Closing.

[MatoTeziTanka]

Community Review — Record: Coprime-Stride Loader + Full GPTQ + XSA-all — val_bpb 1.1133 (3-seed mean)

BPB: 1.1133 | Compliance: LOOKS CLEAN — score-first-per-chunk TTT (legal #1416/#1423 pattern)

What I found in the code (head SHA 7eb6f2410602, file records/track_10min_16mb/2026-03-29_V16_CoprimeSWA_EMA_1.1136/train_gpt.py):

The TTT path at line 1128 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 0.05s, dim=512, layers=11, vocab=1024, code=96935 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 0.05s, dim=512, layers=11, vocab=1024, code=96935 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.