Recent research, such as BitNet, is paving the way for a new era of 1-bitLarge 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 isternary {-1, 0, 1}. It matches the full-precision (i.e., FP16 or BF16)Transformer LLM with the same model size and training tokens in terms of bothperplexity and end-task performance, while being significantly morecost-effective in terms of latency, memory, throughput, and energy consumption.More profoundly, the 1.58-bit LLM defines a new scaling law and recipe fortraining new generations of LLMs that are both high-performance andcost-effective. Furthermore, it enables a new computation paradigm and opensthe door for designing specific hardware optimized for 1-bit LLMs.
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)