LogQuantKV

Official Implementation of LogQuant: Log-Distributed 2-Bit Quantization of KV Cache with Superior Accuracy Preservation.

how to use

requirements

• torch>=2.3
• transformers>=4.42
• quanto==0.2.0

pip install -U transformers quanto==0.2.0

implemented methods

• LogQuant: total full precision tokens memory cost will be [3*window_length]
• PartialLogQuant: total full precision tokens memory cost will be [2*window_length]

use with huggingface models

• example of LogQuant

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from src.LogQuant import QuantoLogQuantizedCache, LogQuantizedCacheConfig

device = "cuda:0" if torch.cuda.is_available() else "cpu"
model_name = "Qwen/Qwen1.5-7B-Chat"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map=device,
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

config = LogQuantizedCacheConfig(
            backend="quanto",
            nbits=2,
            window_length=4,
            compute_dtype="auto",
            device=device,
        )
cache = QuantoLogQuantizedCache(config)

messages = [
    {
        "role": "system",
        "content": "You are a helpful assistant."
    },
    {
        "role": "user", 
        "content": "如果把脏话都说出来了，那么嘴是不是就干净了"
    }]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)

output = model.generate(input_ids, max_new_tokens=128, past_key_values=cache)

print(tokenizer.decode(output[0], skip_special_tokens=True))
'''
system
You are a helpful assistant.
user
如果把脏话都说出来了，那么嘴是不是就干净了
assistant
不，把脏话或不尊重的语言说出口并不会让嘴巴变得干净。恰恰相反，这通常会显示出粗俗、不礼貌或者情绪失控。真正的“干净”是通过言语和行为展现出尊重、理解、宽容和礼貌。即使在压力或冲突的情况下，我们也应该努力控制自己的言辞，用更建设性的方式表达我们的想法。
'''

Citation

If you find LogQuant useful or relevant to your project and research, please kindly cite our paper:

@inproceedings{
han2025logquant,
title={LogQuant: Log-Distributed 2-Bit Quantization of {KV} Cache with Superior Accuracy Preservation},
author={CHEN Han and Zicong Jiang and Zining Zhang and Bingsheng He and Luo Pingyi and Mian Lu and Yuqiang Chen},
booktitle={Sparsity in LLMs (SLLM): Deep Dive into Mixture of Experts, Quantization, Hardware, and Inference},
year={2025},
url={https://openreview.net/forum?id=E26zsSXT3J}
}