processor

Module Contents

PackageProcessor

Provide more flexible distributed tensor communication functions based on

class PackageProcessor

Bases: object

Provide more flexible distributed tensor communication functions based on torch.distributed.send() and torch.distributed.recv().

PackageProcessor defines the details of point-to-point package communication.

EVERYTHING is torch.Tensor in FedLab.

static send_package(package, dst)

Three-segment tensor communication pattern based on torch.distributed

Pattern is shown as follows:

1.1 sender: send a header tensor containing slice_size to receiver

1.2 receiver: receive the header, and get the value of slice_size and create a buffer for incoming slices of content

2.1 sender: send a list of slices indicating the size of every content size.

2.2 receiver: receive the slices list.

3.1 sender: send a content tensor composed of a list of tensors.

3.2 receiver: receive the content tensor, and parse it to obtain slices list using parser function

static recv_package(src=None)

Three-segment tensor communication pattern based on torch.distributed

Pattern is shown as follows:

1.1 sender: send a header tensor containing slice_size to receiver

1.2 receiver: receive the header, and get the value of slice_size and create a buffer for incoming slices of content

2.1 sender: send a list of slices indicating the size of every content size.

2.2 receiver: receive the slices list.

3.1 sender: send a content tensor composed of a list of tensors.

3.2 receiver: receive the content tensor, and parse it to obtain slices list using parser function