Add Hybrid Depth-Recurrent Transformer submission

[tobiascanavesi]

Hybrid Depth-Recurrent Transformer

Testing this new architecture that solves the int8 quantization compounding problem in depth-recurrent transformers.

Key Insight

Standard depth-recurrence shares all weights across loop iterationsm int8 rounding errors compound on every loop (0.40 BPB gap). The hybrid keeps precision-sensitive layers near input/output as unique weights, while only the bulk middle layers are shared and looped.
Result: quantization gap reduced from 0.40 to near-zero (-0.004 BPB).

Architecture

• 1 unique entry layer + 4 shared blocks × 5 loops + 1 unique exit layer = 22 effective layers from 6 weight blocks
• U-Net skip connections across full effective depth
• Per-virtual-layer scalars (attn_scale, mlp_scale, resid_mix, q_gain)

Techniques

• FP16 tied embedding passthrough during int8 quantization
• Sliding window evaluation (stride=64, seq_len=1024)
• Decoupled Muon weight decay (0.02)
• Overtone spectral embedding init (SVD power-law shaping)
• Phase-transition residual mixing initialization

Preliminary Results (2×H100)

Seed	val_bpb	Steps	Artifact
1337	1.3323	954	14.2 MB

8×H100 run pending, expecting significant improvement with full compute.

Reproduce

WARMDOWN_ITERS=2500 MATRIX_LR=0.03 SCALAR_LR=0.03 TIED_EMBED_LR=0.04 torchrun --nproc_per_node=8 train_gpt.py

[MatoTeziTanka]

Community Review — Add Hybrid Depth-Recurrent Transformer submission

BPB: 0.004 (cache parse — may be delta/std, not val_bpb; check PR title) | Compliance: LOOKS CLEAN — pure-neural submission, no TTT/SLOT/n-gram-cache

What I found in the code (head SHA c821a8c4d174, file records/track_10min_16mb/2026-03-21_HybridDepthRecurrent/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 0.03s, dim=768, layers=22, vocab=1024, code=58154 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 0.03s, dim=768, layers=22, vocab=1024, code=58154 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.