Start your AI Orchestrator

The Livepeer AI network is currently in its Beta stage and is undergoing active development. Running it on the same machine as your main Orchestrator or Gateway node may cause stability issues. Please proceed with caution.

The Livepeer AI network is currently in Beta but is already integrated into the main go-livepeer software. You can run the Livepeer AI software using one of the following methods:

Docker (Recommended): The simplest and preferred method.
Pre-built Binaries: An alternative if you prefer not to use Docker.

Orchestrator Node Architecture

In the Livepeer AI network, orchestrator operations rely on two primary node types:

Orchestrator: Manages and routes incoming jobs to available compute resources.
Worker: Performs the actual computation tasks.

The simplest configuration combines both roles on a single machine, utilizing the machine’s GPUs for AI inference tasks, where the orchestrator also functions as a worker (known as a combined AI orchestrator). In this setup, capacity is limited by the available GPUs and is set as runner container count per pipeline/model_id = capacity per pipeline/model_id. For expanded scalability, operators can deploy dedicated (remote) worker nodes that connect to the orchestrator, increasing overall compute capacity. Instructions for setting up remote workers are available on the next page.

Start a Combined AI Orchestrator

Please follow the steps below to start your combined AI orchestrator node.

Retrieve the Livepeer AI Docker Image

Fetch the latest Livepeer AI Docker image from the Livepeer Docker Hub with the following command:

docker pull livepeer/go-livepeer:master

Fetch the Latest AI Runner Docker Image

The Livepeer AI network employs a containerized workflow for running AI models. Fetch the latest AI Runner image with this command:

docker pull livepeer/ai-runner:latest

Pull Pipeline-Specific Images (optional)

Next, pull any pipeline-specific images if needed. Check the pipelines documentation for more information. For example, to pull the image for the segment-anything-2 pipeline:

docker pull livepeer/ai-runner:segment-anything-2

Verify the AI Models are Available

The Livepeer AI network leverages pre-trained AI models for inference tasks. Before launching the AI Orchestrator node, verify that the weights of these models are accessible on your machine. For more information, visit the Download AI Models page.

Configure your AI Orchestrator

Confirm that the AI models are correctly set up in the aiModels.json file in the ~/.lpData/ directory. For guidance on configuring the aiModels.json file, refer to the AI Models Configuration page. The configuration file should resemble:

[
    {
        "pipeline": "text-to-image",
        "model_id": "ByteDance/SDXL-Lightning",
        "price_per_unit": 4768371,
        "warm": true
    }
]

Launch an (off-chain) AI Orchestrator

Execute the Livepeer AI Docker image using the following command:

docker run \
    --name livepeer_ai_orchestrator \
    -v ~/.lpData/:/root/.lpData/ \
    -v /var/run/docker.sock:/var/run/docker.sock \
    --network host \
    --gpus all \
    livepeer/go-livepeer:master \
    -orchestrator \
    -serviceAddr 0.0.0.0:8936 \
    -v 6 \
    -nvidia 0 \
    -aiWorker \
    -aiModels /root/.lpData/aiModels.json \
    -aiModelsDir ~/.lpData/models \
    -aiRunnerImage livepeer/ai-runner:latest # OPTIONAL

This command launches an off-chain AI Orchestrator node. While most of the commands are similar to those used when operating a Mainnet Transcoding Network Orchestrator node (explained in the go-livepeer CLI reference), there are a few Livepeer AI specific flags:

-aiWorker: This flag enables the AI Worker functionality.
-aiModels: This flag sets the path to the JSON file that contains the AI models.
-aiModelsDir: This flag indicates the directory where the AI models are stored on the host machine.
-aiRunnerImage: This optional flag specifies which version of the ai-runner image is used. Example: livepeer/ai-runner:0.0.2

Moreover, the --network host flag facilitates communication between the AI Orchestrator and the AI Runner container.Lastly, the -nvidia can be configured in a few ways. Use a comma seperated list of GPUs ie. 0,1 to activate specific GPU slots, each GPU will need it’s own config item in aiModels.json. Alternativly we can use "all" to activate all GPUs on the machine with a single model loaded in aiModels.json (Warning: If different RAM size GPUs are installed it may cause containers to fail if they have less than the required RAM).

Please note that since we use docker-out-of-docker, the aiModelsDir path should be defined as being on the host machine.

Confirm Successful Startup of the AI Orchestrator

If your Livepeer AI Orchestrator node is functioning correctly, you should see the following output:

2024/05/01 09:01:39 INFO Starting managed container gpu=0 name=text-to-image_ByteDance_SDXL-Lightning modelID=ByteDance/SDXL-Lightning
...
I0405 22:03:17.427058 2655655 rpc.go:301] Connecting RPC to uri=https://0.0.0.0:8936
I0405 22:03:17.430371 2655655 rpc.go:254] Received Ping request

Check Port Availability

To make your Livepeer AI Orchestrator node accessible from the internet, you need to configure your network settings. Ensure that port 8936 is unblocked on your machine. Additionally, consider setting up port forwarding on your router, allowing the Gateway node to be reachable from the internet.

Verify Combined AI Orchestrator Operation

Once your combined Livepeer AI Orchestrator node is running, verify that the worker is operational by sending an AI inference request directly to the ai-runner container. You can either use the Swagger UI interface or a curl command for this check.

Access the Swagger UI

Open your web browser and navigate to http://localhost:8000/docs to access the Swagger UI interface.

Initiate an Inference Request

In the Swagger UI, locate the POST /text-to-image endpoint and click the Try it out button. Use the following example JSON payload:

{
    "prompt": "A cool cat on the beach."
}

This request will instruct the AI model to generate an image based on the text in the prompt field.

Inspect the Inference Response

If the AI Orchestrator node is functioning correctly, you should receive a response similar to the following:

{
    "images": [
        {
            "url": "data:image/png;base64,iVBORw0KGgoAA...",
            "seed": 2724904334
        }
    ]
}

The url field contains the base64 encoded image generated by the AI model. To convert this image to PNG, use a base64 decoder such as Base64.guru.

AI Video

​Orchestrator Node Architecture

​Start a Combined AI Orchestrator

​Verify Combined AI Orchestrator Operation

Orchestrator Node Architecture

Start a Combined AI Orchestrator

Verify Combined AI Orchestrator Operation