resources:
    reservations:
      devices:
        - capabilities: ["nvidia-compute"]
```

##### `driver`

A different driver for the reserved device(s) can be requested using `driver` field. The value is specified as a string.

```yml
deploy:
  resources:
    reservations:
      devices:
        - capabilities: ["nvidia-compute"]
          driver: nvidia
```

##### `count`

If `count` is set to `all` or not specified, Compose reserves all devices that satisfy the requested capabilities. Otherwise, Compose reserves at least the number of devices specified. The value is specified as an integer.

```yml
deploy:
  resources:
    reservations:
      devices:
        - capabilities: ["tpu"]
          count: 2
```

`count` and `device_ids` fields are exclusive. Compose returns an error if both are specified.

##### `device_ids`

If `device_ids` is set, Compose reserves devices with the specified IDs provided they satisfy the requested capabilities. The value is specified as a list of strings.

```yml
deploy:
  resources:
    reservations:
      devices:
        - capabilities: ["gpu"]
          device_ids: ["GPU-f123d1c9-26bb-df9b-1c23-4a731f61d8c7"]
```

`count` and `device_ids` fields are exclusive. Compose returns an error if both are specified.

##### `options`

Driver specific options can be set with `options` as key-value pairs.

```yml
deploy:
  resources:
    reservations:
      devices:
        - capabilities: ["gpu"]
          driver: gpuvendor
          options:
            virtualization: false
```

### `restart_policy`

`restart_policy` configures if and how to restart containers when they exit. If `restart_policy` is not set, Compose considers the `restart` field set by the service configuration.

- `condition`. When set to:
  - `none`, containers are not automatically restarted regardless of the exit status.
  - `on-failure`, the container is restarted if it exits due to an error, which manifests as a non-zero exit code.
  - `any` (default), containers are restarted regardless of the exit status. 
- `delay`: How long to wait between restart attempts, specified as a [duration](extension.md#specifying-durations). The default is 0, meaning restart attempts can occur immediately.
- `max_attempts`: The maximum number of failed restart attempts allowed before giving up. (Default: unlimited retries.)
A failed attempt only counts toward `max_attempts` if the container does not successfully restart within the time defined by `window`.