Skip to main content

Self-Hosted GPU

Influgen can route part of its generation traffic to your own GPU infrastructure. This is useful when you want lower marginal cost, tighter infrastructure control, or regional routing that fits your setup better than a fully cloud-hosted path.

What self-hosting supports today​

The built-in self-hosted adapter currently supports:

  • text-to-image
  • image-to-image

It does not support every advanced workflow. In particular, motion transfer is not part of the built-in self-hosted adapter path today.

Core behavior​

The adapter maintains:

  • endpoint health state
  • request queueing
  • concurrency limits
  • circuit breaking
  • cloud fallback

Important built-in defaults include:

  • timeout: 90 seconds
  • retry count: 2
  • health TTL: 30 seconds
  • failure threshold: 3
  • circuit reset: 5 minutes

[screenshot: Self-hosted admin overview showing routing percentage, endpoint health cards, and recent routing logs]

Admin endpoints​

Self-hosted routing is administered through:

  • GET /api/v1/admin/selfhosted
  • PUT /api/v1/admin/selfhosted/routing
  • POST /api/v1/admin/selfhosted/refresh
  • POST /api/v1/admin/selfhosted/endpoints
  • PATCH /api/v1/admin/selfhosted/endpoints/{id}
  • DELETE /api/v1/admin/selfhosted/endpoints/{id}

Routing controls​

The routing config currently lets admins set:

  • enabled
  • routing_percentage

routing_percentage must stay between 0 and 100.

Use a low percentage first, validate quality and uptime, then scale routing up gradually.

Endpoint management​

Each endpoint can store:

  • name
  • URL
  • API key
  • status
  • max concurrency
  • capabilities

Influgen also tracks runtime telemetry such as:

  • health
  • last error
  • current load
  • average latency
  • consecutive failures
  • circuit-open timing

Practical rollout advice​

  1. Add one healthy endpoint.
  2. Turn routing on at a low percentage.
  3. Watch latency, failures, and quality.
  4. Keep cloud fallback enabled while confidence builds.
  5. Expand capacity only after the endpoint stays healthy under load.

When to build a custom adapter instead​

If you want self-hosted support for advanced workflows such as motion transfer or more provider-specific tuning, the better path is a custom or community adapter. See Adapter Spec.