This repo is a alternative solution for standard self-attention. It uses tree routing with learnable functions for cheaper self-attention while not compromising accuracy.
-
Updated
Oct 11, 2026 - Python
This repo is a alternative solution for standard self-attention. It uses tree routing with learnable functions for cheaper self-attention while not compromising accuracy.
Lightweight System One model for edge devices: tree-masked multi-question inference on frozen Ternary-Bonsai-4B (MLX/PyTorch)
Tree-based speculative decoding benchmarked against linear under equal verification budget — branch factor, depth, and draft quality sweep with fair node-level comparison.
Tree-masked multi-question inference for the Qwen3-4B System One LoRA (PyTorch / MLX)
To associate your repository with the tree-attention topic, visit your repo's landing page and select "manage topics."