DistServe

DistServe improves the performance of large language models (LLMs) serving by disaggregating the prefill and decoding computation. Existing LLM serving systems colocate the two phases and batch the computation of prefill and decoding across all users and requests. We find that this strategy not only leads to strong prefill-decoding interferences but also couples the resource allocation and parallelism plans for both phases. In DistServe, you can simply set the parallelism configs and scheduling strategies for the two phases and it will work just like a single instance which handles the KV-Cache communication and memory management automatically.

It utilizes a high-performance C++ Transformer inference library SwiftTransformer as the execution backend, which supports many features like model/pipeline parallelism, FlashAttention, Continuous Batching, and PagedAttention.

It supports:

GPT-2 (gpt2, gpt2-xl, ...)
OPT (facebook/opt-1.3b, facebook/opt-6.7b, ...)
LLaMA2 (meta-llama/Llama-2-7b, meta-llama/Llama-2-13b, ...)

Build && Install

# clone the project
git clone https://github.com/LLMServe/DistServe.git && cd DistServe

# setup the distserve conda environment
conda env create -f environment.yml && conda activate distserve

# clone and build the SwiftTransformer library  
git clone https://github.com/LLMServe/SwiftTransformer.git && cd SwiftTransformer && git submodule update --init --recursive
cmake -B build && cmake --build build -j$(nproc)
cd ..

# install distserve
pip install -e .

Launching

Launch Ray Cluster

DistServe relies on Ray to implement distributed workers. If you do not launch a Ray runtime in advance, it will automatically initiate a cluster consisting of all the gpus on the current node. You may need to start the Ray runtime manually in advance if you want to use multiple nodes for inference.

Run offline example

DistServe requires at least two GPUs to play with. We provide an offline inference example in examples/offline.py.

Run online example

To run online inference, you need to launch the DistServe API server, see the comments in distserve/api_server/distserve_api_server.py.

Then launch the client example in examples/online.py.

Evaluation

To reproduce all the experiments in our paper, please follow the guidance.

Citation

If you use DistServe for your research, please cite our paper:

@misc{zhong2024distserve,
      title={DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving}, 
      author={Yinmin Zhong and Shengyu Liu and Junda Chen and Jianbo Hu and Yibo Zhu and Xuanzhe Liu and Xin Jin and Hao Zhang},
      year={2024},
      eprint={2401.09670},
      archivePrefix={arXiv},
      primaryClass={cs.DC}
}

LLMServe / DistServe

readme