tensorplay.distributed.algorithms.model_averaging API

Runs hierarchical model averaging (hierarchical SGD).

Process groups of different sizes are organized in a hierarchy, and they average parameters by using different periods concurrently after the warm-up stage. This is an extension of PeriodicModelAverager that supports post-local SGD, which essentially only supports a two-level hierarchy: the intra-machine level and the global level, where the intra-machine level is usually embedded in post_localSGD_hook(). Similarly, the process groups within this class do not have such an intra-machine process subgroup, which should be embedded by the post-local SGD communication hook instead.

Parameters:

• period_group_size_dict – An ordered dict mapping keys of model averaging period to process group size, used for initializing process groups of different sizes in a hierarchy to average parameters concurrently.

• warmup_steps (int) – The number of warm-up steps. During this stage, model averaging is skipped.

• process_group (ProcessGroup, optional) – The overall process group containing all the processes that runs model averaging.

average_parameters(params: Iterable)[source]

Averages parameters or parameter groups of an optimizer.

Averaging only occurs if step is no less than warmup_steps and it can be divided by a period in the keys of period_process_group_dict, where step is increased by 1 at each iteration in the training loop. If step can be divided by multiple periods in the keys of period_process_group_dict, only the largest period is used, and the corresponding process group is used for averaging parameters. :param params: The parameters of a model or parameter groups of an optimizer.

Base class for all model averagers.

Parameters:

process_group – The process group to be used for all-reduce. If None, the default process group, which is created by tensorplay.distributed.init_process_group(), will be used. (default: None)

Averages parameters periodically after the warm-up stage.

This can be used for running post-local SGD, by running DistributedDataParallel (DDP) using the subgroups created by new_subgroups().

Parameters:

• period (int) – The number of steps per model averaging. Usually the period should be greater than 1 to reduce the communication cost. Otherwise, only DDP needs to be used.

• warmup_steps (int) – The number of warm-up steps. During this stage, model averaging is skipped.

• process_group – The process group to be used for all-reduce. If None, the default process group, which is created by tensorplay.distributed.init_process_group(), will be used. (default: None)

Example:

>>> # xdoctest: +SKIP("undefined variables")
>>> averager = averagers.PeriodicModelAverager(period=4, warmup_steps=100)
>>> for step in range(0, 200):
>>>    optimizer.zero_grad()
>>>    loss = loss_fn(output, labels)
>>>    loss.backward()
>>>    optimizer.step()
>>>    # Will average model parameters globally every 4 steps.
>>>    averager.average_parameters(model.parameters())

average_parameters(params: Iterable[Parameter] | Iterable[dict[str, Parameter]])[source]

Averages parameters or parameter groups of an optimizer if step is no less than warmup_steps.

Can be divided by period, where step is increased by 1 at each iteration in the training loop. :param params: The parameters of a model or parameter groups of an optimizer.