Skip to content

Latest commit

 

History

History
91 lines (67 loc) · 3.06 KB

index.md

File metadata and controls

91 lines (67 loc) · 3.06 KB
Raw

Deployment Guide

As we suggest in the introduction, we have four modes of deployment:

If you want to easily get start, you can use our deployed API, the link is https://openomni.ai4wa.com, to manage the tasks.

If you want to test out the full setup locally, you can follow the guide in the All in One Local Machine section.

If you are thinking about deploy it as a product, which is fully locally within a home network, addressing the privacy issue, you can follow the guide in the Private Offline Deployment section.

If you are doing research with cluster of computing resources, or you want annotators to work on the same platform for a serious project, you can follow the guide in the Your Cloud section.

Modules

We have three components in the stack to deploy:

  • API
  • Agent
  • Client
    • Listener (Audio and Video)
    • Responder (Audio)

API

  • Required Resource
    • A server (If on cloud will require a Public IP)
    • Minimum 1 CPU and 1 GB RAM
    • Deployment method:
      • For cloud server: Docker + Docker Compose + Nginx
      • For local server: Docker + Docker Compose

Agent

  • Required Resource
    • Any high-end computational Nvidia GPU resources
      • Can be HPC Clusters
      • Can work on demand, which means, you can spin the tasks when needed
      • Can work on CPU as well, but the performance will be degraded
    • Minimum storage of 500 GB
      • This is required to store the models and the data, especially the LLM models
    • Python 3.8+

Client

  • Required Resource:
    • Hardware:
      • Microphone: To gather the audio data
      • Camera: To gather the video data
      • Speaker: To play the audio data
      • Minimum 2 GB RAM
      • Minimum 1 CPU
      • Minimum 32 GB storage
      • It can be running on a laptop, or working with a Raspberry Pi
    • Python 3.8+

Something like this

client

Storage solution

All the metadata will be communicated via the API, so here we need to think about how can we share the video and audio data between Agent/Client/API.

We have four STORAGE_SOLUTION for this four different scenarios:

  • api: audio and video data will be upload and download via api endpoint, this is for the trial on cloud.
  • volume: all the files will be shared on the same machine via the docker volume and file system, so there is no need to sync anything
  • local: all the modules will be deployed on the same local network, but different machines, so we need to sync the data between them, with rsync
  • s3: API is on your cloud, Agent is anywhere, so we will use s3 to be the storage place for the data, to make sure stable and fast access.

To switch between these four modes, all you need to do is to set the STORAGE_SOLUTION environment variable before start the API

export STORAGE_SOLUTION=api