Skip to content

Latest commit

 

History

History
84 lines (65 loc) · 2.31 KB

README.md

File metadata and controls

84 lines (65 loc) · 2.31 KB
Raw

Preble

Preble is a load balancer for effecient prefix caching systems. PrePrint release at https://arxiv.org/abs/2407.00023

Installation

You can install the package using pip:

Code Structure

The multi_node directory contains the code for running as a separate abstraction layer to SGLang/vLLM in a distributed setting. This code is responsible for coordinating and managing the execution of the distributed system.

Editable Installation

pip3 install -e .
pip install -e "python[all]"
pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.3/

Regular Pip Installation:

pip3 install preble
pip install git+https://github.com/wuklab/preble.git#egg=preble[all]
pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.3/

We release a custom version of sglang that supports chunked prefill

Programatically starting the server

We can support providing a list of runtime urls

from preble.main import start_server

start_server(
    runtime_selection_policy="custom",
    runtime_urls="http://127.0.0.1:30000/generate,http://127.0.0.1:30001/generate",
    host='127.0.0.1',
    port=8000,
    model="mistralai/Mistral-7B-v0.1"
)

We can also support dynamically loading the models to seperate cuda devices

from preble.main import start_server_and_load_models

start_server_and_load_models(
    model_name="mistralai/Mistral-7B-v0.1",
    devices=[0, 1],
    host="127.0.0.1",
    port=8000
)

The server can be run via:

python3 multi_node/server/server.py <server/deploy_and_run>
  • server runs the server given a list of urls
  • deploy_and_run generates two endpoints

CLI Configuration

    runtime_selection_policy: The policy to select the runtime (e.g., custom, round_robin).
    runtime_urls: Comma-separated list of runtime URLs.
    host: The host address for the server.
    port: The port number for the server.
    model: The model to be used (e.g., mistralai/Mistral-7B-v0.1).

Citation And Acknowledgment

The code is forked of sglang

pypi build and install instructions

Currently uploaded at: python setup.py bdist_wheel twine upload --repository testpypi dist/* --verbose python3 -m pip install --index-url https://test.pypi.org/simple/ preble

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.