Skip to content

Record: 30ep Cosine TTT on LeakyReLU² stack (3-seed mean val_bpb=1.0781) #672

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
andrewbaggio1 wants to merge 1 commit into
base: main
Choose a base branch
from
+1,819 −0
Open

Record: 30ep Cosine TTT on LeakyReLU² stack (3-seed mean val_bpb=1.0781) #672

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
55 changes: 55 additions & 0 deletions records/track_10min_16mb/2026-03-25_CosineTTT30ep_SwiGLU_8xH100/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,55 @@
# Record: 30-Epoch Cosine TTT on LeakyReLU² Stack (3-seed mean val_bpb=1.0781)

**3-seed mean val_bpb: 1.0781** (std=0.0041) | **15.62 MB** artifact | 8xH100 SXM, 600s training + 494s TTT + 96s eval

## Results (8xH100 SXM)

| Seed | Steps | Sliding BPB (s64) | Artifact |
|------|-------|-------------------|----------|
| 1337 | 6,921 | 1.0743 | 15.62 MB |
| 42 | — | 1.0774 | 15.62 MB |
| 7 | — | 1.0825 | 15.62 MB |
| **Mean ± Std** | | **1.0781 ± 0.0041** | |

## Approach

Single change from PR #518: **TTT_EPOCHS=30** (default 30 in PR #518's code). All architecture and training hyperparameters identical to PR #518.

PR #518's architecture: 11L, d=512, 8/4 GQA heads, LeakyReLU(0.5)² MLP 3x, BigramHash(2048), SmearGate, XSA4, Partial RoPE (16 dims), LN Scale, EMA(0.997), SWA, Late QAT, OrthoInit, VE128.

TTT: 30 epochs AdamW (lr=0.0005) with cosine LR decay and per-layer LR groups (3x for mlp.proj, 0.5x for mlp.fc). DDP gradient sync across 8 GPUs.

## Timing (seed 1337)

| Phase | Time |
|-------|------|
| Training (8xH100) | 600s (wallclock capped) |
| TTT (30 epochs) | 494s |
| Sliding window eval (stride=64) | 96s |
| **Total** | **~1190s** |

Training fits in the 10-min cap. TTT + eval = 590s, within the 10-min eval budget.

## Run Command

```bash
SEED=1337 torchrun --standalone --nproc_per_node=8 train_gpt.py
```

All hyperparameters are defaults in train_gpt.py.

## Credits

- **PR #518** (architecture + cosine TTT): base submission
- **PR #481** (mrdavtan): Cosine TTT scheduling discovery
- **PR #442** (sjp611): AdamW TTT
- **PR #398** (felipe-parodi): EMA, TTT, XSA foundations

## Test plan

- [x] train_gpt.py compiles (`ast.parse` passes)
- [x] All 3 seeds: artifacts under 16 MB
- [x] 3-seed mean beats verified SOTA (1.1194) by 0.041 BPB
- [x] Training completes in under 10 min on 8xH100
- [x] TTT + eval completes in under 10 min
- [x] PR only adds files to one new folder
11 changes: 11 additions & 0 deletions records/track_10min_16mb/2026-03-25_CosineTTT30ep_SwiGLU_8xH100/submission.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,11 @@
{
"author": "Andrew Baggio",
"github_id": "andrewbaggio1",
"name": "30-Epoch Cosine TTT on LeakyReLU² Stack",
"blurb": "PR #518 architecture with TTT_EPOCHS=30. 3-seed mean val_bpb=1.0781 (std=0.0041). Single hyperparameter change from PR #518.",
"date": "2026-03-25T02:00:00Z",
"val_loss": 1.82,
"val_bpb": 1.0781,
"bytes_total": 15621761,
"bytes_code": 71793
}
Loading