ScalingConfig

class ray.train.ScalingConfig(trainer_resources: dict | None = None, num_workers: int | Tuple[int, int] = 1, use_gpu: bool | Domain | Dict[str, List] = False, resources_per_worker: Dict | Domain | Dict[str, List] | None = None, placement_strategy: str | Domain | Dict[str, List] = 'PACK', accelerator_type: str | None = None, label_selector: Dict[str, str] | List[Dict[str, str]] | None = None, use_tpu: bool = False, topology: str | None = None, elastic_resize_monitor_interval_s: float = 60.0)

Bases: ScalingConfig

Configuration for scaling training.

Parameters:

• num_workers – The number of workers (Ray actors) to launch. Each worker will reserve 1 CPU by default. The number of CPUs reserved by each worker can be overridden with the resources_per_worker argument. If the number of workers is 0, the training function will run in local mode, meaning the training function runs in the same process. To enable elasticity, provide a (min_workers, max_workers) tuple of ints.

• elastic_resize_monitor_interval_s – While the worker group is healthy, consider resizing the worker group every elastic_resize_monitor_interval_s seconds.

• use_gpu – If True, training will be done on GPUs (1 per worker). Defaults to False. The number of GPUs reserved by each worker can be overridden with the resources_per_worker argument.

• resources_per_worker –

  If specified, the resources defined in this Dict is reserved for each worker. Define the "CPU" and "GPU" keys (case-sensitive) to override the number of CPU or GPUs used by each worker.

  Accepts the same resource keys that Ray uses for scheduling tasks and actors (see Resources):

  • "CPU": number of logical CPUs per worker.

  • "GPU": number of logical GPUs per worker. Prefer setting use_gpu=True (which reserves 1 GPU per worker) and only override this key when you need a different per-worker count.

  • "TPU": number of logical TPUs per worker, when use_tpu=True.

  • "memory": heap memory reserved per worker, in bytes (for example, "memory": 1e9 reserves 1 GB per worker).

  • Any custom resource name configured on your cluster (for example, "special_hardware": 1).

  Keys are case-sensitive: use "CPU", "GPU", and "TPU" (uppercase), and "memory" (lowercase).

• placement_strategy – The placement strategy to use for the placement group of the Ray actors. See Placement Group Strategies for the possible options.

• label_selector – A list of label selectors for Ray Train worker placement. If a single label selector is provided, it will be applied to all Ray Train workers. If a list is provided, it must be the same length as the max number of Ray Train workers.

• accelerator_type – [Experimental] If specified, Ray Train will launch the training coordinator and workers on the nodes with the specified type of accelerators. See the available accelerator types. Ensure that your cluster has instances with the specified accelerator type or is able to autoscale to fulfill the request. This field is required when use_tpu is True and num_workers is greater than 1.

• use_tpu – [Experimental] If True, training will be done on TPUs (1 TPU VM per worker). Defaults to False. The number of TPUs reserved by each worker can be overridden with the resources_per_worker argument. This arg enables SPMD execution of the training workload.

• topology – [Experimental] If specified, Ray Train will launch the training coordinator and workers on nodes with the specified topology. Topology is auto-detected for TPUs and added as Ray node labels. This arg enables SPMD execution of the training workload. This field is required when use_tpu is True and num_workers is greater than 1.

Methods

as_placement_group_factory	Returns a PlacementGroupFactory to specify resources for Tune.
from_placement_group_factory	Create a ScalingConfig from a Tune's PlacementGroupFactory

Attributes

accelerator_type
additional_resources_per_worker	Resources per worker, not including CPU or GPU resources.
elastic_resize_monitor_interval_s
elasticity_enabled
label_selector
max_workers
min_workers
num_cpus_per_worker	The number of CPUs to set per worker.
num_gpus_per_worker	The number of GPUs to set per worker.
num_tpus_per_worker	The number of TPUs to set per worker.
num_workers
placement_strategy
resources_per_worker
topology
total_resources	Map of total resources required for training.
trainer_resources
use_gpu
use_tpu