Configuration - Vast.ai mode

The configuration file of a Ponos agent in Vast.ai mode is an extension of the generic Ponos agent. Only the vast_ai section is added.

With this mode, you’ll be able to use remote GPU provided by Vast.ai

`vast_ai.api_key`

A vast.ai API key used to authenticate on the provider and use resources on your own account.

This is optional, and will default to the local settings setup by Vast.ai CLI if it’s present on the host.

`vast_ai.disk_space`

The amount of disk space in GB that will be requested when creating a new instance.

This is optional, and will default to 50 Gb.

`vast_ai.environment`

A string used to identify every Vast.ai instance created by the Ponos agent. This is important when running multiple agents on the same Vast.ai account.

`vast_ai.offer_filters`

You can filter vast.ai offers by using their own filtering syntax through the ponos-agent configuration file.

The syntax is the same as for the vast.ai CLI tool vastai search offers, except that you need to configure a list of filters (one filter per line) in the configuration file.

Available fields

We support all the fields described on the official Vast.ai documentation.

But we always set these values and you won’t be able to change these:

rentable and verified enabled
vms_enabled disabled as we only support direct Docker container execution in this mode
num_gpus set to 1, as we only support one GPU per task in Arkindex workers
disk_space requires at least the configured value in GB (default to 50Gb and more)

Do not try to override these values, as the ponos agent will not start and report an error.

Examples

Here are a few examples, written for the ponos-agent configuration:

To rent somewhat reliable RTX 3090 instances:

vast_ai:
  offer_filters:
    - reliability > 0.98
    - gpu_name: RTX_3090

To rent datacenter gpus with minimal compute_cap and total_flops:

vast_ai:
  offer_filters:
    - compute_cap > 610
    - total_flops > 5
    - datacenter=True'

To rent reliable RTX 3090 or 4090 gpus NOT in China or Vietnam:

vast_ai:
  offer_filters:
    - reliability>0.99
    - gpu_name in ["RTX 4090", "RTX 3090"]
    - geolocation notin ["CN", "VN"]

`vast_ai.registries`

List of Docker registry credentials to use when pulling private Docker images through Vast.ai instances.

`vast_ai.registries.url`

URL of the registry to authenticate.

`vast_ai.registries.username`

Username of the registry to authenticate.

`vast_ai.registries.password`

Password of the registry to authenticate.

Full configuration example

Below is an example of a YAML configuration file to start a Ponos agent in a vast_ai mode, limiting instances to a few European countries, and at least 22Gb of VRAM

# Generic Ponos agent
...

# Ponos agent in a vast.ai mode
vast_ai:
  environment: production
  api_key: XXXX
  disk_space: 100

  offer_filters:
    - geolocation in ["FR", "DE", "ES", "BE"]
    - gpu_ram > 22

  registries:
    - url: registry.gitlab.teklia.com
      username: xxx
      password: yyy