comm_helpers.py

import collections
import logging
import math
import sys
import copy

import torch
import torch.distributed as dist
import functools

def flatten_tensors(tensors):
    """
    Reference: https://github.com/facebookresearch/stochastic_gradient_push

    Flatten dense tensors into a contiguous 1D buffer. Assume tensors are of
    same dense type.
    Since inputs are dense, the resulting tensor will be a concatenated 1D
    buffer. Element-wise operation on this buffer will be equivalent to
    operating individually.
    Arguments:
        tensors (Iterable[Tensor]): dense tensors to flatten.
    Returns:
        A 1D buffer containing input tensors.
    """
    if len(tensors) == 1:
        return tensors[0].view(-1).clone()
    flat = torch.cat([t.view(-1) for t in tensors], dim=0)
    return flat


def unflatten_tensors(flat, tensors):
    """
    Reference: https://github.com/facebookresearch/stochastic_gradient_push

    View a flat buffer using the sizes of tensors. Assume that tensors are of
    same dense type, and that flat is given by flatten_dense_tensors.
    Arguments:
        flat (Tensor): flattened dense tensors to unflatten.
        tensors (Iterable[Tensor]): dense tensors whose sizes will be used to
            unflatten flat.
    Returns:
        Unflattened dense tensors with sizes same as tensors and values from
        flat.
    """
    outputs = []
    offset = 0
    for tensor in tensors:
        numel = tensor.numel()
        outputs.append(flat.narrow(0, offset, numel).view_as(tensor))
        offset += numel
    return tuple(outputs)

def communicate(tensors, communication_op, attention=False):
    """
    Reference: https://github.com/facebookresearch/stochastic_gradient_push

    Communicate a list of tensors.
    Arguments:
        tensors (Iterable[Tensor]): list of tensors.
        communication_op: a method or partial object which takes a tensor as
            input and communicates it. It can be a partial object around
            something like torch.distributed.all_reduce.
    """
    flat_tensor = flatten_tensors(tensors)
    communication_op(tensor=flat_tensor)
    if attention:
        return tensors/flat_tensor
    for f, t in zip(unflatten_tensors(flat_tensor, tensors), tensors):
        t.set_(f)

def SyncAllreduce(model, rank, size):
    '''
    Inputs:
        model: (x^i) local neural net model at i-th worker node
        anchor_model: (z^1=z^2=...=z^m=z) local copy of auxiliary variable
        rank: (i) worker index
        size: (m) total number of workers
        group: worker group
    Output:
        return void, change in-place
    Formula:
        x_new = sum_i x_i / size
    '''
    communication_op = functools.partial(dist.all_reduce)
    params_list = []
    for param in model.parameters():
        param.data.div_(float(size))
        params_list.append(param.data)

    communicate(params_list, communication_op)


def communicate_1(tensors, communication_op, group, attention=False):
    """
    Reference: https://github.com/facebookresearch/stochastic_gradient_push

    Communicate a list of tensors.
    Arguments:
        tensors (Iterable[Tensor]): list of tensors.
        communication_op: a method or partial object which takes a tensor as
            input and communicates it. It can be a partial object around
            something like torch.distributed.all_reduce.
    """
    flat_tensor = flatten_tensors(tensors)
    communication_op(tensor=flat_tensor, group=group)
    if attention:
        return tensors/flat_tensor
    for f, t in zip(unflatten_tensors(flat_tensor, tensors), tensors):
        t.set_(f)

def SyncAllreduce_1(model, rank, size,group):
    communication_op = functools.partial(dist.all_reduce)
    params_list = []
    for param in model.parameters():
        param.data.div_(float(size))
        params_list.append(param.data)

    communicate_1(params_list, communication_op,group=group)

def SyncAllreduce_2(model, rank, size, ue_list):
    communication_op = functools.partial(dist.all_reduce)
    params_list = []
    ue_list_set = set(ue_list)
    if rank in ue_list_set:
        for param in model.parameters():
            param.data.div_(float(len(ue_list)))
            params_list.append(param.data)
    else:
        for param in model.parameters():
            param.data.mul_(0.0)
            params_list.append(param.data)


    communicate(params_list, communication_op)