ml.trainers.mixins.data_parallel

Defines a trainer mixin for data and model parallelism.

This defines how to wrap the model when launching multi-GPU or multi-node jobs. There are two wrappers:

• DistributedDataParallel (DDP)

• FullyShardedDataParallel (FSDP)

DDP is the default wrapper unless conf.parallel.use_fsdp is set to True. DDO runs each model replica on a single GPU processing a subset of the batch, and then synchronizes gradients across all GPUs. FSDP supports more complex sharding of the model across GPUs and nodes, and also supports CPU offloading.

class ml.trainers.mixins.data_parallel.TaskModel(task: TaskT, model: ModelT)

Bases: Module, Generic[ModelT, TaskT, Batch, Loss]

Initializes internal Module state, shared by both nn.Module and ScriptModule.

forward(batch: Batch, state: State) → Loss

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class ml.trainers.mixins.data_parallel.ParallelConfig(use_fsdp: bool = False, cpu_offload: bool = False, sharding_strategy: torch.distributed.fsdp.api.ShardingStrategy = <ShardingStrategy.HYBRID_SHARD: 4>, sync_module_states: bool = True)

Bases: object

use_fsdp: bool = False

cpu_offload: bool = False

sharding_strategy: ShardingStrategy = 4

sync_module_states: bool = True

ml.trainers.mixins.data_parallel.ddp(model: Module, cfg: ParallelConfig) → DistributedDataParallel

ml.trainers.mixins.data_parallel.fsdp(model: Module, cfg: ParallelConfig) → FullyShardedDataParallel

ml.trainers.mixins.data_parallel.dp(model: T, cfg: ParallelConfig) → T | DistributedDataParallel | FullyShardedDataParallel

Wraps a model for data parallel training, if necessary.

Parameters:

• model – The model to wrap.

• cfg – The model configuration.

Returns:

The wrapped model.

class ml.trainers.mixins.data_parallel.TrainerParallelConfig(name: str = '???', exp_name: str = '${ml.exp_name:null}', exp_dir: str = '???', log_dir_name: str = 'logs', use_double_weight_precision: bool = False, checkpoint: ml.trainers.base.CheckpointConfig = <factory>, parallel: ml.trainers.mixins.data_parallel.ParallelConfig = <factory>)

Bases: BaseTrainerConfig

parallel: ParallelConfig

class ml.trainers.mixins.data_parallel.ParallelMixin(config: TrainerConfigT)

Bases: BaseTrainer[ParallelConfigT, ModelT, TaskT]

Defines a trainer mixin for fully sharded data parallel models.