duo-attention

DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads

Stars: 57

Visit

DuoAttention is a framework designed to optimize long-context large language models (LLMs) by reducing memory and latency during inference without compromising their long-context abilities. It introduces a concept of Retrieval Heads and Streaming Heads to efficiently manage attention across tokens. By applying a full Key and Value (KV) cache to retrieval heads and a lightweight, constant-length KV cache to streaming heads, DuoAttention achieves significant reductions in memory usage and decoding time for LLMs. The framework uses an optimization-based algorithm with synthetic data to accurately identify retrieval heads, enabling efficient inference with minimal accuracy loss compared to full attention. DuoAttention also supports quantization techniques for further memory optimization, allowing for decoding of up to 3.3 million tokens on a single GPU.

README:

DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads

[paper]

Demo

https://github.com/user-attachments/assets/b372882b-bf14-4c89-a610-22724d91a415

TL;DR

We significantly reduce both pre-filling and decoding memory and latency for long-context LLMs without sacrificing their long-context abilities.

Abstract

Deploying long-context large language models (LLMs) is essential but poses significant computational and memory challenges. Caching all Key and Value (KV) states across all attention heads consumes substantial memory. Existing KV cache pruning methods either damage the long-context capabilities of LLMs or offer only limited efficiency improvements. In this paper, we identify that only a fraction of attention heads, a.k.a, Retrieval Heads, are critical for processing long contexts and require full attention across all tokens. In contrast, all other heads, which primarily focus on recent tokens and attention sinks, referred to as Streaming Heads, do not require full attention. Based on this insight, we introduce DuoAttention, a framework that only applies a full KV cache to retrieval heads while using a light-weight, constant-length KV cache for streaming heads, which reduces both LLM's decoding and pre-filling memory and latency without compromising its long-context abilities. DuoAttention uses a lightweight, optimization-based algorithm with synthetic data to identify retrieval heads accurately. Our method significantly reduces long-context inference memory by up to 2.55x for MHA and 1.67x for GQA models while speeding up decoding by up to 2.18x and 1.50x and accelerating pre-filling by up to 1.73x and 1.63x for MHA and GQA models, respectively, with minimal accuracy loss compared to full attention. Notably, combined with quantization, DuoAttention enables Llama-3-8B decoding with 3.3 million context length on a single A100 GPU.

Installation and Usage

Environment Setup

Training and Evaluation Environment

conda create -yn duo python=3.10
conda activate duo

conda install -y git
conda install -y nvidia/label/cuda-12.4.0::cuda-toolkit
conda install -y nvidia::cuda-cudart-dev
conda install -y pytorch torchvision torchaudio pytorch-cuda=12.4 -c pytorch -c nvidia

pip install transformers accelerate sentencepiece datasets wandb accelerate sentencepiece datasets wandb zstandard matplotlib huggingface_hub
pip install tensor_parallel

pip install ninja packaging
pip install flash-attn --no-build-isolation

# LongBench evaluation
pip install seaborn rouge_score einops pandas

pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.4/

# Install DuoAttention
pip install -e .

# Install Block Sparse Streaming Attention
git clone [email protected]:mit-han-lab/Block-Sparse-Attention.git
cd Block-Sparse-Attention
python setup.py install

Demo Environment

conda create -yn duo_demo python=3.10
conda activate duo_demo

# Install DuoAttention
pip install -e .

conda install -y git
conda install -y nvidia/label/cuda-12.4.0::cuda-toolkit
conda install -y nvidia::cuda-cudart-dev

# Install QServe
git clone [email protected]:mit-han-lab/qserve.git
cd qserve
pip install -e .
pip install ninja packaging
pip install flash-attn==2.4.1 --no-build-isolation
cd kernels
python setup.py install

# Install FlashInfer
pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.3/
pip install tensor_parallel

Dataset

To download the dataset:

mkdir -p datasets
cd datasets

wget https://huggingface.co/datasets/togethercomputer/Long-Data-Collections/resolve/main/fine-tune/booksum.jsonl.zst

Model

To download models supported by DuoAttention:

mkdir -p models
cd models

# Models that DuoAttention currently supports for evaluation
huggingface-cli download togethercomputer/Llama-2-7B-32K-Instruct --local-dir Llama-2-7B-32K-Instruct
huggingface-cli download gradientai/Llama-3-8B-Instruct-Gradient-1048k --local-dir Llama-3-8B-Instruct-Gradient-1048k
huggingface-cli download gradientai/Llama-3-8B-Instruct-Gradient-4194k --local-dir Llama-3-8B-Instruct-Gradient-4194k
huggingface-cli download mistralai/Mistral-7B-Instruct-v0.2 --local-dir Mistral-7B-Instruct-v0.2
huggingface-cli download mistralai/Mistral-7B-Instruct-v0.3 --local-dir Mistral-7B-Instruct-v0.3

#  W8A8KV4 models using SmoothQuant and QServe for demo purposes
huggingface-cli download mit-han-lab/Llama-3-8B-Instruct-Gradient-1048k-w8a8kv4-per-channel --local-dir Llama-3-8B-Instruct-Gradient-1048k-w8a8kv4-per-channel
huggingface-cli download mit-han-lab/Llama-3-8B-Instruct-Gradient-4194k-w8a8kv4-per-channel --local-dir Llama-3-8B-Instruct-Gradient-4194k-w8a8kv4-per-channel

Quick Start for DuoAttention

We offer a simple one-click patch to enable DuoAttention optimization on HuggingFace models, including Llama and Mistral. Pretrained retrieval head patterns for five long-context models are available in the attn_patterns directory: Llama-2-7B-32K-Instruct, Llama-3-8B-Instruct-Gradient-1048k, Llama-3-8B-Instruct-Gradient-4194k, Mistral-7B-Instruct-v0.2, and Mistral-7B-Instruct-v0.3. If you'd like to train your own retrieval head patterns, you can use the training script provided in the scripts directory. Below is an example of how to enable DuoAttention on the Llama-3-8B-Instruct-Gradient-1048k model.

from duo_attn.utils import load_attn_pattern, sparsify_attention_heads
from duo_attn.patch import enable_duo_attention_eval
import transformers
import torch

# Load the model
model = transformers.AutoModelForCausalLM.from_pretrained(
    "models/Llama-3-8B-Instruct-Gradient-1048k",
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    attn_implementation="eager",
)

# Load the attention pattern
attn_heads, sink_size, recent_size = load_attn_pattern(
    "attn_patterns/Llama-3-8B-Instruct-Gradient-1048k/lr=0.02-reg=0.05-ctx=1000_32000-multi_passkey10"
)

# Sparsify attention heads
attn_heads, sparsity = sparsify_attention_heads(attn_heads, sparsity=0.5)

# Enable DuoAttention
enable_duo_attention_eval(
    model,
    attn_heads,
    num_recent_tokens=64,
    num_sink_tokens=256,
)

# Move model to GPU
model = model.cuda()

# Ready for inference!

Demo

After setting up the environment, you can run the following script to execute the W4A8KV4 with DuoAttention demo on the Llama-3-8B-Instruct-Gradient-4194k model. The demo is designed to run on a single A100 GPU and supports a context length of up to 3.3 million tokens.

bash scripts/run_demo.sh

Results

Retrieval Head Identification

After preparing the dataset and models, you can run the training script to identify the retrieval heads. For the models we evaluated, the corresponding attention patterns are available in the attn_patterns directory.

bash scripts/run_train.sh

Needle in a Haystack (NIAH)

DuoAttention provides comparable accuracy as full attention on the Needle-in-a-Haystack benchmark using 25% full attention ratio on the MHA model and 50% full attention ratio on the GQA model.

bash scripts/run_niah.sh

LongBench

bash scripts/run_longbench.sh

DuoAttention provides better KV budget and accuracy trade-off on LongBench benchmarks.

Efficiency

bash scripts/run_efficiency.sh

Per-token decoding latency and memory usage of DuoAttention compared to full attention across varying context sizes. DuoAttention uses a 25% retrieval head ratio for Llama-2-7B (MHA) and 50% for Llama-3-8B (GQA). DuoAttention achieves up to 2.45× memory reduction for MHA and 1.65× for GQA models, along with up to 2.13× latency reduction for MHA and 1.5× for GQA models. These reductions approach the inverse of the retrieval head ratios as context length increases. Out-of-memory (OOM) results are linearly extrapolated from measured data.

Pre-filling latency and memory usage of DuoAttention compared to full attention across varying pre-filling chunk sizes. DuoAttention uses a 25% retrieval head ratio for Llama-2-7B (MHA), pre-filling a context of 100K tokens, and a 50% ratio for Llama-3-8B (GQA), pre-filling a context of 320K tokens. As the pre-filling chunk size decreases, DuoAttention achieves up to 1.73× latency reduction for MHA and 1.63× for GQA models, with memory reductions up to 2.38× for MHA and 1.53× for GQA models.

DuoAttention’s decoding memory and latency vs. KV budget with a fixed context length. Memory and latency are reduced linearly when the ratio of retrieval heads is reduced. DuoAttention achieves up to 2.55× memory reduction for MHA and 1.67× for GQA models, along with up to 2.18× latency reduction for MHA and 1.50× for GQA models.

Combined with 8-bit weight and 4-bit KV cache quantization, DuoAttention can accommodate 3.3 million tokens on a single A100-80G GPU for the Llama-3-8B model.

Citation

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

@article{xiao2024duo,
        title={DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads},
        author={Xiao, Guangxuan and Tang, Jiaming and Zuo, Jingwei and Guo, Junxian and Yang, Shang and Tang, Haotian and Fu, Yao and Han, Song},
        journal={arXiv},
        year={2024}
}

For Tasks:

Click tags to check more tools for each tasks

optimize inference reduce memory usage accelerate decoding identify critical heads support quantization

For Jobs:

data scientist machine learning engineer research scientist ai architect nlp specialist

Alternative AI tools for duo-attention

Similar Open Source Tools

duo-attention

github

: 57

aimo-progress-prize

This repository contains the training and inference code needed to replicate the winning solution to the AI Mathematical Olympiad - Progress Prize 1. It consists of fine-tuning DeepSeekMath-Base 7B, high-quality training datasets, a self-consistency decoding algorithm, and carefully chosen validation sets. The training methodology involves Chain of Thought (CoT) and Tool Integrated Reasoning (TIR) training stages. Two datasets, NuminaMath-CoT and NuminaMath-TIR, were used to fine-tune the models. The models were trained using open-source libraries like TRL, PyTorch, vLLM, and DeepSpeed. Post-training quantization to 8-bit precision was done to improve performance on Kaggle's T4 GPUs. The project structure includes scripts for training, quantization, and inference, along with necessary installation instructions and hardware/software specifications.

github

: 61

ComfyUI-IF_AI_tools

ComfyUI-IF_AI_tools is a set of custom nodes for ComfyUI that allows you to generate prompts using a local Large Language Model (LLM) via Ollama. This tool enables you to enhance your image generation workflow by leveraging the power of language models.

github

: 397

NeMo-Curator

NeMo Curator is a GPU-accelerated open-source framework designed for efficient large language model data curation. It provides scalable dataset preparation for tasks like foundation model pretraining, domain-adaptive pretraining, supervised fine-tuning, and parameter-efficient fine-tuning. The library leverages GPUs with Dask and RAPIDS to accelerate data curation, offering customizable and modular interfaces for pipeline expansion and model convergence. Key features include data download, text extraction, quality filtering, deduplication, downstream-task decontamination, distributed data classification, and PII redaction. NeMo Curator is suitable for curating high-quality datasets for large language model training.

github

: 438

voicechat2

Voicechat2 is a fast, fully local AI voice chat tool that uses WebSockets for communication. It includes a WebSocket server for remote access, default web UI with VAD and Opus support, and modular/swappable SRT, LLM, TTS servers. Users can customize components like SRT, LLM, and TTS servers, and run different models for voice-to-voice communication. The tool aims to reduce latency in voice communication and provides flexibility in server configurations.

github

: 500

CompressAI-Vision

CompressAI-Vision is a tool that helps you develop, test, and evaluate compression models with standardized tests in the context of compression methods optimized for machine tasks algorithms such as Neural-Network (NN)-based detectors. It currently focuses on two types of pipeline: Video compression for remote inference (`compressai-remote-inference`), which corresponds to the MPEG "Video Coding for Machines" (VCM) activity. Split inference (`compressai-split-inference`), which includes an evaluation framework for compressing intermediate features produced in the context of split models. The software supports all the pipelines considered in the related MPEG activity: "Feature Compression for Machines" (FCM).

github

: 81

sarathi-serve

Sarathi-Serve is the official OSDI'24 artifact submission for paper #444, focusing on 'Taming Throughput-Latency Tradeoff in LLM Inference'. It is a research prototype built on top of CUDA 12.1, designed to optimize throughput-latency tradeoff in Large Language Models (LLM) inference. The tool provides a Python environment for users to install and reproduce results from the associated experiments. Users can refer to specific folders for individual figures and are encouraged to cite the paper if they use the tool in their work.

github

: 112

stable-diffusion-webui

Stable Diffusion web UI is a web interface for Stable Diffusion, implemented using Gradio library. It provides a user-friendly interface to access the powerful image generation capabilities of Stable Diffusion. With Stable Diffusion web UI, users can easily generate images from text prompts, edit and refine images using inpainting and outpainting, and explore different artistic styles and techniques. The web UI also includes a range of advanced features such as textual inversion, hypernetworks, and embeddings, allowing users to customize and fine-tune the image generation process. Whether you're an artist, designer, or simply curious about the possibilities of AI-generated art, Stable Diffusion web UI is a valuable tool that empowers you to create stunning and unique images.

github

: 137.8k

CALF

CALF (LLaTA) is a cross-modal fine-tuning framework that bridges the distribution discrepancy between temporal data and the textual nature of LLMs. It introduces three cross-modal fine-tuning techniques: Cross-Modal Match Module, Feature Regularization Loss, and Output Consistency Loss. The framework aligns time series and textual inputs, ensures effective weight updates, and maintains consistent semantic context for time series data. CALF provides scripts for long-term and short-term forecasting, requires Python 3.9, and utilizes word token embeddings for model training.

github

: 60

aircraft

Headwind Simulations A339X - A330-900neo is an open-source project aimed at creating a free Airbus A330-900neo for Microsoft Flight Simulator. The project is based on the FlyByWire System A32NX and offers a detailed simulation of the A330-941 model with various components like engines, FMS, ACAS, ATC, and more. Users can build the aircraft using Docker and node modules, and the package can be easily integrated into MSFS. The project is part of a collaborative effort with other open-source projects contributing to the aircraft's systems, cockpit, sound, and 3D parts. The repository is dual-licensed under GNU GPLv3 for textual-form source code and CC BY-NC 4.0 for artistic assets, ensuring proper usage and attribution of the content.

github

: 161

pytorch-forecasting

PyTorch Forecasting is a PyTorch-based package for time series forecasting with state-of-the-art network architectures. It offers a high-level API for training networks on pandas data frames and utilizes PyTorch Lightning for scalable training on GPUs and CPUs. The package aims to simplify time series forecasting with neural networks by providing a flexible API for professionals and default settings for beginners. It includes a timeseries dataset class, base model class, multiple neural network architectures, multi-horizon timeseries metrics, and hyperparameter tuning with optuna. PyTorch Forecasting is built on pytorch-lightning for easy training on various hardware configurations.

github

: 3.8k

contracts

AXONE Smart Contracts repository hosts Smart Contracts for the AXONE network, compatible with any Cosmos blockchains using the CosmWasm framework. It includes storage, sovereignty, and resource management oriented Smart Contracts. Each contract has different functionalities and maturity stages, with detailed tech documentation and emojis indicating maturity levels. The repository provides tools for building, testing, deploying, and interacting with Smart Contracts, along with guidelines for contributing and community engagement.

github

: 123

tensorrtllm_backend

The TensorRT-LLM Backend is a Triton backend designed to serve TensorRT-LLM models with Triton Inference Server. It supports features like inflight batching, paged attention, and more. Users can access the backend through pre-built Docker containers or build it using scripts provided in the repository. The backend can be used to create models for tasks like tokenizing, inferencing, de-tokenizing, ensemble modeling, and more. Users can interact with the backend using provided client scripts and query the server for metrics related to request handling, memory usage, KV cache blocks, and more. Testing for the backend can be done following the instructions in the 'ci/README.md' file.

github

: 664

HEC-Commander

HEC-Commander Tools is a suite of python notebooks developed with AI assistance for water resource engineering workflows, providing automation for HEC-RAS and HEC-HMS through Jupyter Notebooks. It contains automation scripts for HEC-HMS, HEC-RAS, and DSS, along with miscellaneous tools. The repository also includes blog posts, ChatGPT assistants, and presentations related to H&H modeling and water resources workflows. Developed to support Region 4 of the Louisiana Watershed Initiative by Fenstermaker.

github

: 53

MiniCheck

MiniCheck is an efficient fact-checking tool designed to verify claims against grounding documents using large language models. It provides a sentence-level fact-checking model that can be used to evaluate the consistency of claims with the provided documents. MiniCheck offers different models, including Bespoke-MiniCheck-7B, which is the state-of-the-art and commercially usable. The tool enables users to fact-check multi-sentence claims by breaking them down into individual sentences for optimal performance. It also supports automatic prefix caching for faster inference when repeatedly fact-checking the same document with different claims.

github

: 80

vision-llms-are-blind

This repository contains the code and data for the paper 'Vision Language Models Are Blind'. It explores the limitations of large language models with vision capabilities (VLMs) in performing basic visual tasks that are easy for humans. The repository presents benchmark results showcasing the poor performance of state-of-the-art VLMs on tasks like counting line intersections, identifying circles, letters, and shapes, and following color-coded paths. The research highlights the challenges faced by VLMs in understanding visual information accurately, drawing parallels to myopia and blindness in human vision.

github

: 96

For similar tasks

duo-attention

github

: 57

edgen

Edgen is a local GenAI API server that serves as a drop-in replacement for OpenAI's API. It provides multi-endpoint support for chat completions and speech-to-text, is model agnostic, offers optimized inference, and features model caching. Built in Rust, Edgen is natively compiled for Windows, MacOS, and Linux, eliminating the need for Docker. It allows users to utilize GenAI locally on their devices for free and with data privacy. With features like session caching, GPU support, and support for various endpoints, Edgen offers a scalable, reliable, and cost-effective solution for running GenAI applications locally.

github

: 279

easydist

EasyDist is an automated parallelization system and infrastructure designed for multiple ecosystems. It offers usability by making parallelizing training or inference code effortless with just a single line of change. It ensures ecological compatibility by serving as a centralized source of truth for SPMD rules at the operator-level for various machine learning frameworks. EasyDist decouples auto-parallel algorithms from specific frameworks and IRs, allowing for the development and benchmarking of different auto-parallel algorithms in a flexible manner. The architecture includes MetaOp, MetaIR, and the ShardCombine Algorithm for SPMD sharding rules without manual annotations.

github

: 70

Awesome_LLM_System-PaperList

Since the emergence of chatGPT in 2022, the acceleration of Large Language Model has become increasingly important. Here is a list of papers on LLMs inference and serving.

github

: 157

LLM-Viewer

LLM-Viewer is a tool for visualizing Language and Learning Models (LLMs) and analyzing performance on different hardware platforms. It enables network-wise analysis, considering factors such as peak memory consumption and total inference time cost. With LLM-Viewer, users can gain valuable insights into LLM inference and performance optimization. The tool can be used in a web browser or as a command line interface (CLI) for easy configuration and visualization. The ongoing project aims to enhance features like showing tensor shapes, expanding hardware platform compatibility, and supporting more LLMs with manual model graph configuration.

github

: 210

ServerlessLLM

ServerlessLLM is a fast, affordable, and easy-to-use library designed for multi-LLM serving, optimized for environments with limited GPU resources. It supports loading various leading LLM inference libraries, achieving fast load times, and reducing model switching overhead. The library facilitates easy deployment via Ray Cluster and Kubernetes, integrates with the OpenAI Query API, and is actively maintained by contributors.

github

: 199

sarathi-serve

github

: 112

aphrodite-engine

Aphrodite is the official backend engine for PygmalionAI, serving as the inference endpoint for the website. It allows serving Hugging Face-compatible models with fast speeds. Features include continuous batching, efficient K/V management, optimized CUDA kernels, quantization support, distributed inference, and 8-bit KV Cache. The engine requires Linux OS and Python 3.8 to 3.12, with CUDA >= 11 for build requirements. It supports various GPUs, CPUs, TPUs, and Inferentia. Users can limit GPU memory utilization and access full commands via CLI.

github

: 1.1k

For similar jobs

Awesome_Mamba

Awesome Mamba is a curated collection of groundbreaking research papers and articles on Mamba Architecture, a pioneering framework in deep learning known for its selective state spaces and efficiency in processing complex data structures. The repository offers a comprehensive exploration of Mamba architecture through categorized research papers covering various domains like visual recognition, speech processing, remote sensing, video processing, activity recognition, image enhancement, medical imaging, reinforcement learning, natural language processing, 3D recognition, multi-modal understanding, time series analysis, graph neural networks, point cloud analysis, and tabular data handling.

github

: 125

unilm

The 'unilm' repository is a collection of tools, models, and architectures for Foundation Models and General AI, focusing on tasks such as NLP, MT, Speech, Document AI, and Multimodal AI. It includes various pre-trained models, such as UniLM, InfoXLM, DeltaLM, MiniLM, AdaLM, BEiT, LayoutLM, WavLM, VALL-E, and more, designed for tasks like language understanding, generation, translation, vision, speech, and multimodal processing. The repository also features toolkits like s2s-ft for sequence-to-sequence fine-tuning and Aggressive Decoding for efficient sequence-to-sequence decoding. Additionally, it offers applications like TrOCR for OCR, LayoutReader for reading order detection, and XLM-T for multilingual NMT.

github

: 19.6k

llm-app-stack

LLM App Stack, also known as Emerging Architectures for LLM Applications, is a comprehensive list of available tools, projects, and vendors at each layer of the LLM app stack. It covers various categories such as Data Pipelines, Embedding Models, Vector Databases, Playgrounds, Orchestrators, APIs/Plugins, LLM Caches, Logging/Monitoring/Eval, Validators, LLM APIs (proprietary and open source), App Hosting Platforms, Cloud Providers, and Opinionated Clouds. The repository aims to provide a detailed overview of tools and projects for building, deploying, and maintaining enterprise data solutions, AI models, and applications.

github

: 929

awesome-deeplogic

Awesome deep logic is a curated list of papers and resources focusing on integrating symbolic logic into deep neural networks. It includes surveys, tutorials, and research papers that explore the intersection of logic and deep learning. The repository aims to provide valuable insights and knowledge on how logic can be used to enhance reasoning, knowledge regularization, weak supervision, and explainability in neural networks.

github

: 214

Awesome-LLMs-on-device

Welcome to the ultimate hub for on-device Large Language Models (LLMs)! This repository is your go-to resource for all things related to LLMs designed for on-device deployment. Whether you're a seasoned researcher, an innovative developer, or an enthusiastic learner, this comprehensive collection of cutting-edge knowledge is your gateway to understanding, leveraging, and contributing to the exciting world of on-device LLMs.

github

: 747

duo-attention

github

: 57

weave

Weave is a toolkit for developing Generative AI applications, built by Weights & Biases. With Weave, you can log and debug language model inputs, outputs, and traces; build rigorous, apples-to-apples evaluations for language model use cases; and organize all the information generated across the LLM workflow, from experimentation to evaluations to production. Weave aims to bring rigor, best-practices, and composability to the inherently experimental process of developing Generative AI software, without introducing cognitive overhead.

github

: 685

agentcloud

AgentCloud is an open-source platform that enables companies to build and deploy private LLM chat apps, empowering teams to securely interact with their data. It comprises three main components: Agent Backend, Webapp, and Vector Proxy. To run this project locally, clone the repository, install Docker, and start the services. The project is licensed under the GNU Affero General Public License, version 3 only. Contributions and feedback are welcome from the community.

github

: 500