nGPT on the Hypersphere: Making Normalized Transformers Work at 16MB (Research)

[DbBested]

Research contribution (not a record submission). I wanted to see if nGPT could be made competitive under Parameter Golf constraints after PR #831 dismissed it. Turns out it can — and the investigation turned up some findings that might be useful to others.

val_bpb: 1.1502 (sliding window, 8xH200, 560s)

Findings

• PR Research: Why Novel Architectures Fail at 16MB — Throughput-Quantization Co-optimization #831's failure was three init bugs, not architectural incompatibility. Fixing zero-init projections, adding learnable logit scaling, and not normalizing embeddings takes nGPT from 1.69 to 1.27 BPB.
• torch.compile has a precision compounding bug with sequential L2 normalization. Inductor fuses through float() casts; across 86 normalize calls the bf16 errors compound catastrophically. Fix: wrap normalize in an opaque autograd.Function via allow_in_graph. Generalizes beyond nGPT. (Large numeric divergence for torch compile vs eager in bf16 pytorch/pytorch#168126)
• Post-dequant renormalization cuts the quantization gap from +0.35 to +0.008 BPB. Three lines of code.
• nGPT's compression advantage vanishes at full training length — undertrained weights compress well because they're close to orthogonal init, not because of the architecture.
• Paper design choices (signed alpha, s_z scaling) hurt at 5000 steps — the optimizer doesn't have time to exploit the extra degrees of freedom.
• TTT is incompatible with renorm dequantization — produces NaN at every LR tested.

Hardware

MIT ORCD cluster, 8xH200 SXM, CUDA 12.4, PyTorch 2.6. 560s wallclock.

Full writeup with code, tables, and ablations in the README.

[MatoTeziTanka]

Community Review — nGPT on the Hypersphere: Making Normalized Transformers Work at 16MB (Research)

BPB: 1.1502 | Compliance: LOOKS CLEAN — score-first-per-chunk TTT (legal #1416/#1423 pattern)

What I found in the code (head SHA 673af4d636eb, file records/track_10min_16mb/2026-03-29_nGPT_Hypersphere_L2NormalizeHP/train_gpt.py):

The TTT path at line 1158 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 4.91s, dim=512, layers=11, vocab=1024, code=118566 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 4.91s, dim=512, layers=11, vocab=1024, code=118566 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.