Configure GitLab to access self-hosted models

DETAILS: Tier: Ultimate with GitLab Duo Enterprise - Start a trial Offering: Self-managed Status: Beta

Introduced in GitLab 17.1 with a flag named ai_custom_model. Disabled by default.

Enabled on self-managed in GitLab 17.6.

Changed to require GitLab Duo add-on in GitLab 17.6 and later.

To configure your GitLab instance to access the available self-hosted models in your infrastructure:

Use a locally hosted or GitLab.com AI gateway.
Configure your GitLab instance.
Configure the self-hosted model.
Configure the GitLab Duo features to use your self-hosted model.

Configure your GitLab instance

Prerequisites:

Upgrade to the latest version of GitLab.

To configure your GitLab instance to access the AI gateway:

::Tabs

:::TabTitle Linux package

Where your GitLab instance is installed, update the /etc/gitlab/gitlab.rb file:
```
sudo vim /etc/gitlab/gitlab.rb
```

Add and save the following environment variables:

gitlab_rails['env'] = {
  'AI_GATEWAY_URL' => '<path_to_your_ai_gateway>:<port>'
}

Run reconfigure:
```
sudo gitlab-ctl reconfigure
```

:::TabTitle Helm Chart (Kubernetes)

Add the following values to your Helm chart:
```
gitlab:
  webservice:
    extraEnv:
      AI_GATEWAY_URL: '<path_to_your_ai_gateway>:<port>'
  sidekiq:
    extraEnv:
      AI_GATEWAY_URL: '<path_to_your_ai_gateway>:<port>'
  toolbox:
    extraEnv:
      AI_GATEWAY_URL: '<path_to_your_ai_gateway>:<port>'
```
- The AI_GATEWAY_URL parameter for webservice must be externally accessible because it is given to editor extensions for direct connection to the AI gateway.
- The AI_GATEWAY_URL parameters for sidekiq and toolbox can be either externally accessible or Kubernetes internal addresses (for example, ai-gateway.gitlab.svc.cluster.local). It might be more time and resource efficient to use Kubernetes internal addresses, so the requests do not have to go through the external load balancer and the Ingress controller to re-enter the cluster.

::EndTabs

Configure the self-hosted model

Prerequisites:

You must be an administrator.

To configure a self-hosted model:

On the left sidebar, at the bottom, select Admin.
Select Self-hosted models.
- If the Self-hosted models menu item is not available, synchronize your subscription after purchase:
  1. On the left sidebar, select Subscription.
  2. In Subscription details, to the right of Last sync, select synchronize subscription ({retry}).
Select Add self-hosted model.

Complete the fields:

Deployment name: Enter a name to uniquely identify the model deployment, for example, Mixtral-8x7B-it-v0.1 on GCP.
Model family: Select the model family the deployment belongs to. Only GitLab-approved models are in this list.
Endpoint: Enter the URL where the model is hosted.
- For models hosted through vLLM, it is essential to suffix the URL with /v1.
API key: Optional. Add an API key if you need one to access the model.

Model identifier (optional): Optional. The model identifier is based on your deployment method:

Deployment method	Format	Example
vLLM	`custom_openai/<name of the model served through vLLM>`	`custom_openai/Mixtral-8x7B-Instruct-v0.1`
Bedrock	`bedrock/<model ID of the model>`	`bedrock/mistral.mixtral-8x7b-instruct-v0:1`
Azure	`azure/<model ID of the model>`	`azure/gpt-35-turbo`
Others	The field is optional

Select Create self-hosted model.

Configure GitLab Duo features to use self-hosted models

Prerequisites:

You must be an administrator.

View configured features

On the left sidebar, at the bottom, select Admin.
Select Self-hosted models.
- If the Self-hosted models menu item is not available, synchronize your subscription after purchase:
  1. On the left sidebar, select Subscription.
  2. In Subscription details, to the right of Last sync, select synchronize subscription ({retry}).
Select the AI-powered features tab.

Configure the feature to use a self-hosted model

Configure the GitLab Duo feature to send queries to the configured self-hosted model:

On the left sidebar, at the bottom, select Admin.
Select Self-hosted models.
Select the AI-powered features tab.
For the feature you want to configure, from the dropdown list, choose the self-hosted model you want to use. For example, Mistral.