Shortcuts

distributed

Functions

all_gather_ddp_if_available

rtype

Any

distributed_available

rtype

Any

gather_all_tensors

rtype

Any

get_default_process_group_backend_for_device

rtype

Any

init_dist_connection

rtype

Any

rank_zero_only

rtype

Any

register_ddp_comm_hook

Function to register communication hook for DDP model https://pytorch.org/docs/master/ddp_comm_hooks.html.

sync_ddp

rtype

Any

sync_ddp_if_available

rtype

Any

tpu_distributed

rtype

bool

Classes

AllGatherGrad

Gathers tensors from the whole group and stacks them.

Utilities that can be used with distributed training.

class pytorch_lightning.utilities.distributed.AllGatherGrad(*args, **kwargs)[source]

Bases: torch.autograd.function.Function

Gathers tensors from the whole group and stacks them.

This implementation is copied from PyTorch.

Deprecated since version v1.8.0: This function has been deprecated in v1.8.0 in favor of torch.distributed.nn.functional.all_gather() and will be removed in v2.0.0.

static backward(ctx, *grad_output)[source]

Defines a formula for differentiating the operation with backward mode automatic differentiation (alias to the vjp function).

This function is to be overridden by all subclasses.

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computated w.r.t. the output.

Return type

Tuple[Tensor, None]

static forward(ctx, tensor, group=None)[source]

Performs the operation.

This function is to be overridden by all subclasses.

It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).

The context can be used to store arbitrary data that can be then retrieved during the backward pass. Tensors should not be stored directly on ctx (though this is not currently enforced for backward compatibility). Instead, tensors should be saved either with ctx.save_for_backward() if they are intended to be used in backward (equivalently, vjp) or ctx.save_for_forward() if they are intended to be used for in jvp.

Return type

Tensor

pytorch_lightning.utilities.distributed.register_ddp_comm_hook(model, ddp_comm_state=None, ddp_comm_hook=None, ddp_comm_wrapper=None)[source]

Function to register communication hook for DDP model https://pytorch.org/docs/master/ddp_comm_hooks.html.

Parameters
  • model (DistributedDataParallel) – DDP model

  • ddp_comm_state (Optional[object]) – state is passed to the hook and can be used to maintain and update any state information that users would like to maintain as part of the training process. Examples: error feedback in gradient compression, peers to communicate with next in GossipGrad etc.

  • ddp_comm_hook (Optional[Callable]) –

    hook(state: object, bucket: dist._GradBucket) -> torch.futures.Future

    This callable function is called once the bucket is ready. The hook can perform whatever processing is needed and return a Future indicating completion of any async work (ex: allreduce). If the hook doesn’t perform any communication, it can also just return a completed Future. The Future should hold the new value of grad bucket’s tensors. Once a bucket is ready, c10d reducer would call this hook and use the tensors returned by the Future and copy grads to individual parameters.

  • ddp_comm_wrapper (Optional[Callable]) – communication hook wrapper to support a communication hook such as FP16 compression as wrapper, which could be combined with ddp_comm_hook

Examples

>>> from torch.distributed.algorithms.ddp_comm_hooks import ( 
...     default_hooks as default,
...     powerSGD_hook as powerSGD,
...     post_localSGD_hook as post_localSGD,
... )
>>>
>>> # fp16_compress_hook for compress gradients
>>> ddp_model = ...
>>> register_ddp_comm_hook( 
...     model=ddp_model,
...     ddp_comm_hook=default.fp16_compress_hook,
... )
>>>
>>> # powerSGD_hook
>>> ddp_model = ...
>>> register_ddp_comm_hook( 
...     model=ddp_model,
...     ddp_comm_state=powerSGD.PowerSGDState(
...         process_group=None,
...         matrix_approximation_rank=1,
...         start_powerSGD_iter=5000,
...     ),
...     ddp_comm_hook=powerSGD.powerSGD_hook,
... )
>>>
>>> # post_localSGD_hook
>>> subgroup, _ = torch.distributed.new_subgroups() 
>>> ddp_model = ...
>>> register_ddp_comm_hook( 
...     model=ddp_model,
...     state=post_localSGD.PostLocalSGDState(
...         process_group=None,
...         subgroup=subgroup,
...         start_localSGD_iter=1_000,
...     ),
...     ddp_comm_hook=post_localSGD.post_localSGD_hook,
... )
>>>
>>> # fp16_compress_wrapper combined with other communication hook
>>> ddp_model = ...
>>> register_ddp_comm_hook( 
...     model=ddp_model,
...     ddp_comm_state=powerSGD.PowerSGDState(
...         process_group=None,
...         matrix_approximation_rank=1,
...         start_powerSGD_iter=5000,
...     ),
...     ddp_comm_hook=powerSGD.powerSGD_hook,
...     ddp_comm_wrapper=default.fp16_compress_wrapper,
... )
Return type

None