Non-record: 11L Partial RoPE + XSA4 + VE128 + Tight SWA + GPTQ-lite (val_bpb=1.1804)#534

Closed

rarce wants to merge 42 commits intoopenai:mainfrom

rarce:sota-review

Commits on Mar 20, 2026

Squashed 'modded-nanogpt/' content from commit 1cafa69
rarce
committed
Merge commit '6818cb1272096e12f2d1641a595c9a02e9833d49' as 'modded-nanogpt'
rarce
committed
Squashed 'slowrun/' content from commit 7677821
rarce
committed
Merge commit '68078b7590c21c866f86b44fe9be644e1d775597' as 'slowrun'
rarce
committed
Add SOTA review knowledge base with analysis of speedrun and slowrun challenges

rarce
and
claude
committed
Add deep research on sub-100M model techniques for Parameter Golf

rarce
and
claude
committed
Update knowledge base with submission analysis and R&D directions

rarce
and
claude
committed
Update CLAUDE.md with knowledge base, frontier stack, and current state

rarce
and
claude
committed
Implement consensus stack in train_gpt_mlx.py

rarce
and
claude
committed
Add int6 quantization (step=4 rounding) to consensus stack

rarce
and
claude
committed
Port consensus stack to train_gpt.py (PyTorch/CUDA)

rarce
and
claude
committed
Add RunPod automation script (bash + curl, no dependencies)

rarce
and
claude
committed
Update CLAUDE.md with consensus stack status and RunPod workflow

rarce
and
claude
committed
Fix int6 dequantization: add weights_only=False to torch.load

rarce
and
claude
committed
Fix int6 quantization: use post-hoc step rounding like reference impl

rarce
and
claude
committed
Fix runpod.sh: match actual REST API v1 response format

rarce
and
claude
committed

Commits on Mar 21, 2026

Commits on Mar 22, 2026

Commits on Mar 23, 2026