12L QAT Int4-MLP + Int6-Attn (Non-record)

[Meirzhan05]

12L QAT Int4-MLP + Int6-Attn (Non-record Submission)

Summary

• Adds mixed-precision Quantization-Aware Training (int4 for MLP, int6 for attention) via Straight-Through Estimator from the start of training
• Uses the ~3MB byte savings from int4 MLP compression to fund a 12th transformer layer (vs 11 in SOTA)
• Halves the sliding window eval stride from 64 → 32 for better per-token context

3-Seed Validation Results

Seed	Post-EMA BPB	Post-Quant BPB	Artifact Size	Wall Time
42	1.1593	1.2234	12,972,053 B	600s
314	1.1603	1.2386	12,951,841 B	600s
999	1.1598	1.2255	12,970,681 B	600s
Mean	1.1598	1.2292	12.96 MB
Std	0.0005	0.0082

Note: This is a non-record submission. The post-quantization BPB (1.2292) does not beat the naive baseline (1.2244) due to a large quantization gap (~0.069 BPB). The QAT approach with int4 MLP quantization introduces too much distortion at this model scale.

Changes

Component	SOTA #1 (1.1194)	This PR
Layers	11	12
MLP precision	int6	int4 (QAT)
Attn precision	int6	int6 (QAT)
Eval stride	64	32
QAT	Late (last ~10%)	Full (from step 0)

Implementation

QAT is applied directly in MLP.forward and CausalSelfAttention.forward on the banked weight tensors via _fake_quantize_ste (row-wise scale, STE gradient). Clip ranges: mlp_clip=7 (int4), attn_clip=31 (int6). Post-training quantization uses GPTQ-lite clip search with the same ranges.

All techniques from PR #549 are inherited: LeakyReLU(0.5)², Legal Score-First TTT, Parallel Muon + Parameter Banking, XSA on last 4 layers, Partial RoPE (16/64 dims), LN Scale 1/sqrt(layer+1), EMA (0.997), SmearGate, BigramHash(2048), Value Embedding.

Test Plan

• Verify artifact size < 16MB after LZMA compression (12.96 MB mean)
• Verify training completes in ≤ 600s on 8xH100 SXM (600s all seeds)
• Run 3-seed validation (seeds 42, 314, 999)
• Confirm val_bpb improvement over baseline — not achieved (non-record)

Hardware

• 8x H100 SXM 80GB (RunPod)
• PyTorch 2.9.1+cu128
• Flash Attention 3

[MatoTeziTanka]

Community Review — 2L QAT Int4-MLP + Int6-Attn

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 8dd90897ab7b, file records/track_10min_16mb/2026-03-25_QAT_Int4MLP_12L/train_gpt.py):

The TTT path at line 1098 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=12, vocab=1024, code=91173 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=12, vocab=1024, code=91173 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.

[Meirzhan05]

@MatoTeziTanka
3-seed validation complete on 8×H100 SXM (RunPod):

Seed	Post-Quant BPB	Artifact	Wall Time
42	1.2234	12.97 MB	600s
314	1.2386	12.95 MB	600s
999	1.2255	12.97 MB	600s
Mean	1.2292	12.96 MB

All seeds pass artifact (<16MB) and wallclock (≤600s) requirements. Post-quant BPB does not beat the naive baseline (1.2244) due to the int4 MLP quantization gap (~0.069 BPB), so marking this as a non-record submission.