Reactor — Deploy Real-Time Video Models

Every model deployed on Reactor runs as a Docker container. You build the image locally and push it to your private container registry.

Build your image

The reactor init template includes a Dockerfile optimized for Reactor. Build it from your project root:

docker build -t my-model:v1 .

Use CUDA base images matching your target GPU (e.g., nvidia/cuda:12.1.0-runtime-ubuntu22.04)
Install Python dependencies first to leverage Docker layer caching
Copy model code last so code changes don’t invalidate the dependency layer
Avoid including weights in the image. Upload weights separately via model weights. This keeps images small and lets you version weights independently.
Set the correct entrypoint. The Reactor runtime expects a specific startup command, which is preconfigured in the template Dockerfile.

Your registry URL is shown by reactor auth status. Tag your image with the full registry path:

docker tag my-model:v1 <your-registry>/my-model:v1

Each partner has a namespaced registry, so your images are isolated from other partners.

First, authenticate Docker with your registry:

reactor auth docker-login

Then push:

docker push <your-registry>/my-model:v1

Error	Fix
Auth credentials missing	Run `reactor auth docker-login` to refresh. Tokens expire every 12 hours.
Authorization token expired	Run `reactor auth docker-login` again.
Push is slow or timing out	Reduce image size with a `.dockerignore`. Don’t bake weights into the image.
Manifest unknown	Double-check your image tag. The registry is case-sensitive.