TemporalMesh Transformer: 29.4 PPL at 48% compute — dynamic graph attention + adaptive exit gates (open-source, 226 tests)
#39
by vigneshwar234 - opened
TemporalMesh Transformer (TMT) — New efficient transformer architecture
Sharing TMT, an open-source PyTorch transformer that jointly solves three problems no other architecture addresses together:
Three core problems → five innovations:
- 🕸 Mesh Attention: kNN graph rebuilt per-layer from cosine similarity → O(S·k) vs O(S²)
- ⏱ Temporal Decay: learned multiplicative attenuation post-softmax (not additive like ALiBi)
- ⚡ Adaptive Depth Routing: per-token exit gate, punctuation exits layer 2, rare words layer 12
- 🔀 Dual-Stream FFN: syntax + semantic parallel streams, sigmoid fusion
- 🧠 EMA Memory Anchors: 16 persistent fast-weight vectors, cross-sequence recall
Results (120M params, WikiText-2):
| Model | PPL ↓ | Compute |
|---|---|---|
| Vanilla Transformer | 42.1 | 100% |
| Longformer | 39.6 | 62% |
| RWKV | 33.1 | 50% |
| Mamba | 31.8 | 55% |
| Full TMT | 29.4 | 48% |
Superadditive effect: combined gain = 12.7 PPL vs 8.6 from summing components individually.
📄 Paper: https://zenodo.org/records/20287390
💻 Code + 226 tests: https://github.com/vignesh2027/TemporalMesh-Transformer
🎮 Live demo: https://huggingface.co/spaces/vigneshwar234/TemporalMesh-Transformer-Demo
🤗 Model: https://huggingface.co/vigneshwar234/TemporalMesh-Transformer