Non-record: 6-Technique Stack — Catalytic Residuals + Value Residual + Gated Attention + BigramHash(10240) + 12L (val_bpb=1.1690)

[joshuaswarren]

Non-record: 6-Technique Stack (val_bpb=1.1690, 8xH100 SXM)

First submission combining 6 independently-proven architecture improvements never stacked together:

Technique	Source	Impact
Catalytic Residuals	PR #450	-0.024 bpb
Value Residual (ResFormer)	PR #413, arXiv:2410.17897	-0.015 bpb
Gated Attention	PR #413, arXiv:2505.06708	-0.003 bpb
BigramHash(10240)	PR #450	-0.070 bpb vs 2048
12 Layers	PR #450	-0.023 bpb vs 11L
3x MLP	Merged SOTA	Standard

Results

• Sliding window val_bpb: 1.1690 (stride 64)
• Post-quant roundtrip: 1.2043
• Pre-quant: 1.1911
• 6,981 steps at 85.78 ms/step
• Artifact: 15.3 MB (int6+zstd, under 16MB)
• Training: 598.8s on 8xH100 SXM

Additional techniques

OrthoInit, Muon WD=0.04, SWA (last 20%), Late QAT (threshold 0.25), logit softcap 30.0

Run

pip install sentencepiece zstandard
python3 data/cached_challenge_fineweb.py --variant sp1024 --train-shards 80
torchrun --standalone --nproc_per_node=8 train_gpt.py

All hyperparameters set as defaults.

[mohosy]

value residual + gated attention on 12 layers is alot of new stuff at once lol. did you try adding them one at a time to see which ones actualy helped or did you just yolo the whole stack