[Non-Record] Depth Recurrence + SwiGLU + Mixed Quantization

[Shakil281]

Summary

Non-record experimental submission exploring three novel architectural modifications on top of the PR #549 stack (LeakyReLU + Legal TTT + Parallel Muon):

• Depth Recurrence (K=2): Loop 11 unique layers twice for 22 effective layers with shared weights. U-Net skip connections activate only on the final pass. Trades compute for parameters — the freed budget can widen the model.
• SwiGLU Activation: Replaces LeakyReLU(0.5)² with gated linear units using SiLU. The up projection is 2x wider (split into gate + value). Standard in modern LLMs (LLaMA, Mistral, Gemma).
• Mixed int5/int6 Quantization: Per-layer quantization sensitivity — int6 for critical layers, int5 for middle MLP layers. Saves ~5-8% artifact space.

Expected Impact

Modification	Expected BPB Delta
Depth Recurrence (K=2)	-0.002 to -0.005
SwiGLU replacing LeakyReLU²	-0.001 to -0.003
Mixed int5/int6 quant	-0.000 to -0.001

Status

• Untested on 8×H100 (pending compute grant approval)
• Code modifications are complete and structurally verified
• Submitted as architectural proposal for community exploration

Test plan

• Validate on 8×H100 SXM with 3 seeds (1337, 42, 2025)
• Confirm artifact fits under 16MB
• Measure throughput impact of depth recurrence (2x layer passes)
• Compare SwiGLU vs LeakyReLU² BPB in isolation
• Test mixed quant vs uniform int6

🤖 Generated with Claude Code

[MatoTeziTanka]

Community Review — [Non-Record] Depth Recurrence + SwiGLU + Mixed Quantization

BPB: (not parsed — see PR title) | Compliance: LOOKS CLEAN — score-first-per-chunk TTT (legal #1416/#1423 pattern)

What I found in the code (head SHA 91c8ee68c899, file records/track_non_record_16mb/2026-03-25_DepthRecurrence_SwiGLU_AdvancedTTT_MixedQuant/train_gpt.py):

The TTT path at line 1104 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 0.04s, dim=512, layers=11, vocab=1024, code=91628 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 0.04s, dim=512, layers=11, vocab=1024, code=91628 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.