XXXXyu/bitnet_b1_58-3B-vlut-gguf

bitnetb158-3B-vlut-gguf

This repository contains state-of-the-art ternary-packed versions of bitnet_b1_58-3B in GGUF format, optimized for efficient on-device inference using the Vec-LUT method.

Key Features

• —🎯 SOTA Compression: Achieves BPW (bits per weight) as low as 1.60 through lossless sub-2-bit ternary packing.
• —⚡ SOTA Performance: Delivers superior throughput (4.2x speedup) in parallel inference scenarios via vector lookup table (LUT).
• —🔌 Drop-in Ready: Seamless integration with vlut.cpp for immediate deployment on edge devices.

Available Model Variants

Models are named as ggml-model-{PACKING}_{TILE}.gguf:

Selection Guide

• —BPW vs. Speed: I1_V achieves lower memory usage but may not always outperform I2_V in speed.
• —Tiling Trade-off: Tiled variants (tile size > 1) deliver higher throughput but require larger cache capacity.
• —Starting Point: Use I1_V_2 or I2_V_4 as a starting point.

For detailed tiling parameter analysis, see Evaluation.md and the paper.

Usage

Prerequisites

Install vlut.cpp (these models require vlut.cpp, not vanilla llama.cpp):

git clone https://github.com/Cipherxzc/vlut.cpp.git
cd vlut.cpp
cmake -B build && cmake --build build --config Release -j4

Download & Run

# Download the recommended variant, e.g., I2_V_4
hf download <repo_id> \
  ggml-model-I2_V_4.gguf --local-dir ./models

# Run parallel inference
./build/bin/llama-batched \
  -m ./models/ggml-model-I2_V_4.gguf \
  -p "I believe the meaning of life is" \
  -np 32 -n 16 -t 1 --temp 0.5 --repeat-penalty 1.5

# Benchmark performance
./build/bin/llama-bench \
  -m ./models/ggml-model-I2_V_4.gguf \
  -t 1 -p 128 -n 0

For comprehensive usage instructions, refer to the vlut.cpp Quick Start Guide.

Citation

If you use these models, please cite our paper:

@article{li2025veclut,
  title={Vec-LUT: Vector Table Lookup for Parallel Ultra-Low-Bit LLM Inference on Edge Devices},
  author={Li, Xiangyu and Yin, Chengyu and Wang, Weijun and Wei, Jianyu and Cao, Ting and Liu, Yunxin},
  journal={arXiv preprint arXiv:2512.06443},
  year={2025},
  url={https://arxiv.org/abs/2512.06443}
}