GPU Container Job
A GPU Container job is CosmicAC's most general-purpose job type. This page explains what it is, how GPU hardware is exposed to the VM, and how the shell connection works.
What a GPU Container job is
Despite the name, a CosmicAC GPU Container is not a Docker container in the traditional sense. A GPU Container job is a KubeVirt virtual machine (VM) that Kubernetes schedules and manages. The term "container" describes the job type and the shell interface it exposes, not the underlying isolation mechanism.
How GPU hardware is exposed
Kubernetes uses device plugins to expose GPU hardware to scheduled jobs. When Kubernetes schedules a GPU Container job, the device plugin makes the requested GPUs accessible to the VM.
Shell access
Shell access connects the CosmicAC CLI to an agent running inside the VM over Hyperswarm SSH, a peer-to-peer SSH tunnel. Closing the shell does not stop the job.
Environment control
GPU Container jobs provide full control over the environment inside the VM. The operating system, packages, scripts, and system configuration are all configurable.
This contrasts with Managed Inference, where CosmicAC manages the serving layer.