Non-record: Extended Compute Scaling Analysis (20K steps, 1.0960 BPB, 3 seeds (each run ~6 hours on 4xA100MIG))

[OnlyJundong]

Summary

This submission is a non-record submission. It studies how (PR #549 by @abaybektursun) scales under extended compute, removing the 10-minute wall-clock constraint. The same architecture and code are trained for 20K steps (~6 hours) on 4×A100 MIG instances (approximately 10× slower per step than 8×H100 SXM).

Results

20K steps, ~6 hours (4×A100 MIG, 3-seed comparison)

Seed	step_avg	steps	Pre-TTT bpb	Post-TTT bpb	TTT gain	Artifact
1337	828.7ms	20,000	1.1018	1.0957	-0.0061	15,077,933
42	828.8ms	20,000	1.1020	1.0962	-0.0058	15,137,145
2024	839.8ms	20,000	1.1017	1.0962	-0.0055	14,942,394
Mean	832.4ms	20,000	1.1018	1.0960 (std 0.0003)	-0.0058	15,052,491

Plots

BPB vs Steps (ASCII plot)

Power-law decay with two distinct phases: rapid early learning, then warmdown-driven final drop.

BPB
4.10 |*
     |
     |
2.50 |
     |
1.26 | *
1.23 |    *
1.22 |       *
1.20 |          *
1.19 |            * *
1.18 |                *
1.10 |                  *
     +---------+--------+-> steps (K)
     0        10       20

     |<early >|<warmdown>|
      (rapid)  (sharp drop)

Artifact Size vs Steps (ASCII plot)

Non-monotonic: grows rapidly to a peak at ~15K steps, then shrinks back below budget during warmdown.

MB
17.2 |    * * * * * *
16.0 |--------------------  16MB limit
15.1 |                  *
14.1 | 
13.1 | *
 4.6 |*
     +---------+--------+-> steps (K)
     0        10        20

     |<-fits->|<over>|<fits>|

[MatoTeziTanka]

Community Review — Non-record: Extended Compute Scaling Analysis (20K steps, 1.0960 BPB, 3 seeds (each run ~6 hours on 4xA100MIG))

BPB: 1.0960 | Compliance: LOOKS CLEAN — score-first-per-chunk TTT (legal #1416/#1423 pattern)

What I found in the code (head SHA 80311efcf92c, file records/track_non_record_16mb/2026-04-06_ExtendedCompute_20K_6hours_4xA100MIG_ScalingAnalysis/train_gpt.py):

The TTT path at line 1095 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 0.59s, dim=512, layers=11, vocab=1024, code=91502 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 0.59s, dim=512, layers=11, vocab=1024, code=91502 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.