Skip to content

Non-record: 11L SwiGLU + XSA4 + EMA + U-Net + AdamW TTT (pending compute) #291

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
mohosy wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Non-record: 11L SwiGLU + XSA4 + EMA + U-Net + AdamW TTT (pending compute) #291

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
3a36c41
Non-record: 11L EMA + XSA + TTT (based on #287, pending compute)
Mar 21, 2026
c57bfbe
Update: drop TTT (negative interaction with EMA+XSA per PR #303), cle…
Mar 21, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
29 changes: 29 additions & 0 deletions records/track_10min_16mb/2026-03-20_11L_EMA_XSA_TTT_Int6/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,29 @@
# 11L EMA + XSA + TTT + Int6 MLP3x (pending compute)

## Approach

Builds on PR #287 (jfprincz, val_bpb=1.1271, 15.5MB) by adding test-time training (TTT). TTT runs during eval only so artifact size is unchanged.

### What's new over #287
- **TTT**: Full-weight SGD on val data before scoring (3 epochs, lr=0.002, momentum=0.9)
- Freezes first 2 blocks during TTT for stability
- Full DDP support

### Inherited from #287
- 11 layers, 512-dim, 8 heads, 4 KV heads (GQA)
- 3x MLP (1536 hidden), relu-squared
- EMA (decay=0.997) replacing SWA
- XSA on last 4 layers
- Int6 per-row quant + zstd-22
- SmearGate + BigramHash (2048)
- Muon WD=0.04, momentum=0.99
- FlashAttention, OrthoInit, muP

### Expected improvement
TTT has shown ~0.003-0.005 bpb improvement in PRs #254, #264, #281. Applied to #287's 1.1271 base, projected score: ~1.122-1.124.

## Checklist
- [x] Submission folder
- [x] `README.md`, `submission.json`, `train_gpt.py`
- [ ] Training log (pending compute)
- [ ] BPB score (pending compute)
10 changes: 10 additions & 0 deletions records/track_10min_16mb/2026-03-20_11L_EMA_XSA_TTT_Int6/submission.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,10 @@
{
"submitter": "mohosy",
"date": "2026-03-20",
"track": "10min_16mb",
"hardware": "8xH100 SXM",
"training_time_seconds": 600,
"val_bpb": null,
"artifact_size_bytes": null,
"notes": "11L EMA + XSA + TTT + Int6 MLP3x + SmearGate + BigramHash. Based on jfprincz #287 (1.1271) with TTT added. Pending compute."
}
Loading