Search results
bitnet.cpp is the official inference framework for 1-bit LLMs (e.g., BitNet b1.58). It offers a suite of optimized kernels, that support fast and lossless inference of 1.58-bit models on CPU (with NPU and GPU support coming next).
By applying 1.58-bit quantization to convolutional neural networks and building upon transformative research in the field, this project extends beyond simple implementation to create a living document that evolves with ongoing experimentation.
18 wrz 2024 · BitNet is a special transformers architecture that represents each parameter with only three values: (-1, 0, 1), offering a extreme quantization of just 1.58 ( l o g 2 (3) log_2(3) l o g 2 (3)) bits per parameter. However, it requires to train a model from scratch.
28 lut 2024 · Recent research, such as BitNet, is paving the way for a new era of 1-bit Large Language Models (LLMs). In this work, we introduce a 1-bit LLM variant, namely BitNet b1.58, in which every single parameter (or weight) of the LLM is ternary {-1, 0, 1}.
9 mar 2024 · 1.58 BitNet uses low-precision binary weights and quantized activations to 8 bits, and high-precision for optimizer states and gradient functions during training. It can be represented as a...
26 mar 2024 · Unlike its predecessor, BitNet b1.58 replaces the conventional nn.Linear layers with BitLinear layers, leveraging 1.58-bit weights and 8-bit activations.
7 wrz 2024 · Reducing the precision of model weights from 32-bit floats to 8-bit integers, or even 1-bit binary numbers is called quantization. This article aims to demystify 1.58-bit large language models with an easily accessible overview based on a literature review.
