Skip to main content

Vast.ai

If you're using Vast.ai as a cloud provider for Cirun, this is how you would give Cirun access to your Vast.ai account, for it to be able to create GPU runners.

  • Go to your Vast.ai Account and generate an API key
  • Paste your API key in the "Cloud" Tab of Cirun Dashboard under the Vast.ai section

Configuration

Vast.ai runners are GPU-first. The GPU type and count are specified via instance_type, and the environment is defined by a Docker image (container mode) or a VM image (VM mode).

Container Mode

The default mode runs the runner inside a Docker container with SSH access:

runners:
- name: vastai-gpu-runner
cloud: vast_ai
instance_type: 1xRTX_4090
machine_image: nvidia/cuda:12.0.0-base-ubuntu22.04
labels:
- cirun-vastai-gpu
extra_config:
disk_space: 30

VM Mode

For workloads that need a full virtual machine instead of a container, set vm: true in extra_config and use a Vast.ai VM image:

runners:
- name: vastai-gpu-vm-runner
cloud: vast_ai
instance_type: 1xRTX_4090
machine_image: vastai/kvm:ubuntu_cli_22.04-2025-11-21
labels:
- cirun-vastai-gpu-vm
extra_config:
disk_space: 30
vm: true

Instance Type

The instance_type field specifies the number of GPUs and the GPU model in the format {count}x{gpu_model}:

instance_type: 1xRTX_4090        # 1x RTX 4090
instance_type: 4xA100_SXM # 4x A100 SXM
instance_type: 2xH100 # 2x H100
instance_type: 8xA100_SXM4_80GB # 8x A100 SXM4 80GB

Use underscores in place of spaces for GPU model names.

Machine Image

For container mode, use any Docker image (e.g. nvidia/cuda:12.0.0-base-ubuntu22.04). For VM mode, use a Vast.ai VM image (e.g. vastai/kvm:ubuntu_cli_22.04-2025-11-21).

Region

Region is optional. When specified, use a country code to restrict which datacenter the runner is placed in:

region: US

Extra Configuration

OptionTypeDescription
disk_spacenumberMinimum disk space in GB
min_ramnumberMinimum RAM in GB
vmbooleanUse VM mode instead of container mode
extra_config:
disk_space: 30
min_ram: 8
vm: true