Skip to content

Record: GatedDeltaNet FLA + Score-First TTT + Brotli — val_bpb 1.00980 (3-seed mean) #1711

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

Closed
aamodbhatt wants to merge 1 commit into from
Closed

Record: GatedDeltaNet FLA + Score-First TTT + Brotli — val_bpb 1.00980 (3-seed mean) #1711

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
59 changes: 59 additions & 0 deletions records/track_10min_16mb/2026-04-18_GDN_FLA_ScoreFirstTTT/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,59 @@
# GatedDeltaNet FLA + Score-First TTT + Brotli

**val_bpb: 1.00980** (3-seed mean, std 0.0015) | **~15.6 MB** | 8xH100 SXM

## Results

| Seed | Steps | EMA BPB | Pre-TTT BPB | **Post-TTT BPB** | TTT Gain | Artifact |
|------|-------|---------|-------------|-----------------|----------|----------|
| 1337 | 2,400 | 0.998608 | 1.01720 | **1.00803** | -0.00917 | 15,595,190 |
| 42 | 2,410 | 1.001194 | 1.02054 | **1.01069** | -0.00986 | 15,602,610 |
| 2025 | 2,408 | 1.001260 | 1.01933 | **1.01067** | -0.00866 | 15,608,600 |
| **Mean** | **2,406** | **1.000354** | **1.01902** | **1.00980 (std 0.0015)** | **-0.00923** | |

## What Changed

- **GatedDeltaNet (FLA)** K_KVShare_Wider architecture from PR #1687 — O(n) linear attention replacing softmax
- **Brotli-11 compression** instead of zstd-22 — saves ~900KB, all artifacts well under 16MB
- **Score-first TTT** (SGD lr=0.005, momentum=0.9, 3 epochs, freeze first 2 blocks) adapted from PR #461

## Architecture

| Component | Setting |
|-----------|---------|
| Layers | 10 GDN (544d, 8H) |
| KV Sharing | Stride 2 |
| MLP | 3x width |
| BigramHash | 3072 x 112 + trigram |
| Weight avg | EMA(0.997) + SWA(every 50) |
| Late QAT | Threshold 0.15 |
| Quantization | Int6 matrices + Int8 embeddings |
| Compression | Brotli quality 11 |
| Optimizer | Muon (matrices) + Adam (scalars/embeds) |

## TTT Protocol

Legal score-first TTT per Issue #1017:

1. Val tokens split into 32K-token chunks
2. For each chunk: **SCORE** under `torch.inference_mode()`, then **TRAIN** on already-scored tokens
3. SGD(lr=0.005, momentum=0.9), 3 epochs, freeze first 2 blocks, cosine LR decay
4. Last chunk scored but never trained on

## Run Command

```bash
pip install flash-linear-attention==0.4.2 fla-core==0.4.2

ARCH_MODE=K TTT_ENABLED=1 TTT_LR=0.005 TTT_EPOCHS=3 \
TTT_CHUNK_TOKENS=32768 TTT_FREEZE_BLOCKS=2 TTT_MOMENTUM=0.9 \
TTT_BATCH_SEQS=32 TTT_GRAD_CLIP=1.0 \
SEED=1337 MAX_WALLCLOCK_SECONDS=600 \
torchrun --standalone --nproc_per_node=8 train_gpt.py
```

## Credits

- **GatedDeltaNet architecture**: PR #1687 by @resouer
- **Flash Linear Attention**: @sustcsonglin (fla-core 0.4.2)
- **TTT recipe**: PR #461 by @Christopher-Lee-McClendon
Loading