cog

Why we included this project

Shipping a trained model as a real service is often the messiest part of ML. Cog automates that step: you describe the environment in a cog.yaml file, write your inference logic as a Python runner, and it builds a Docker image with the CUDA, cuDNN, and Python versions that match your framework, sidestepping the usual dependency and driver mismatches. Because it reads your input and output type annotations, it can also generate an OpenAPI schema and a fast Rust/Axum HTTP server, so you get a working REST endpoint without writing API code by hand. If you're an ML engineer or a small team looking for a dependable way to get from a local prototype to a container that runs anywhere Docker does, including your own infrastructure or Replicate, this tool removes a lot of friction.