Skip to content

records: David Ghazaryan — MoE + BigramHash4096 (val_bpb 1.11799) #1538

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
davie2009kh wants to merge 1 commit into
base: main
Choose a base branch
from
Open

records: David Ghazaryan — MoE + BigramHash4096 (val_bpb 1.11799) #1538

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/2026-04-05_David_MoE-Bigram4096/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,55 @@
# [10min/16mb] David Ghazaryan — MoE + BigramHash4096

**val_bpb: 1.1180** (3-seed mean) | 8×H100 SXM | 600s

## Results

| Seed | val_bpb | Artifact (bytes) |
|------|---------|-----------------|
| 1337 | 1.11764880 | 15,873,596 |
| 42 | 1.11891002 | 15,893,104 |
| 2025 | 1.11742168 | 15,908,116 |
| **mean** | **1.11799350** | 15,891,605 |

## Novel Contributions

### 1. BigramHash 4096
Expanded bigram hash table from SOTA's 3072 to 4096 buckets.
Provides richer local context signal at the embedding stage.

### 2. Mixture-of-Experts MLP (first in this repo)
Replaces standard MLP with 4 experts + top-2 routing.
Same active parameters but adds expert specialisation.

## Architecture

| Component | Setting | Introduced by |
|-----------|---------|---------------|
| Layers | 11 (512d, 8 heads, 4 KV) | Baseline |
| MLP | 3x LeakyReLU(0.5)² | PR #493 |
| XSA | All 11 layers | PR #478 |
| EMA | decay=0.997 | PR #374 |
| Partial RoPE | 16/64 dims | PR #315 |
| LN Scale | 1/√(layer+1) | PR #315 |
| GPTQ | Full Hessian AR self-gen | PR #1019 |
| BigramHash | 4096 buckets dim=96 | This PR |
| MoE MLP | 4 experts top-2 | This PR |
| Compression | int6 + lzma | PR #414 |

## Requirements

```bash
pip install sentencepiece zstandard
pip install flash_attn_3 --find-links \
https://windreamer.github.io/flash-attention3-wheels/cu128_torch291
```

## Run Command

```bash
BIGRAM_VOCAB_SIZE=4096 BIGRAM_DIM=96 WARMDOWN_ITERS=4000 TARGET_MB=15.9 SEED=1337 \
torchrun --standalone --nproc_per_node=8 train_gpt.py
```

## Hardware
8× H100 80GB HBM3 (YSU HPC Cluster, YerevaNN/Eleveight AI Program)
22 changes: 22 additions & 0 deletions records/track_10min_16mb/2026-04-05_David_MoE-Bigram4096/submission.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,22 @@
{
"val_bpb": 1.1180,
"val_bpb_std": 0.000619,
"val_bpb_exact": "1.11799350",
"seeds": [1337, 42, 2025],
"per_seed_val_bpb": [
"1.11764880",
"1.11891002",
"1.11742168"
],
"per_seed_artifact_bytes": [
15873596,
15893104,
15908116
],
"artifact_bytes_mean": 15891605,
"hardware": "8xH100 SXM",
"training_time_seconds": 600,
"description": "BigramHash4096 + MoE MLP — first MoE exploration in repo",
"author": "davie2009kh",
"date": "2026-04-05"
}
Loading