File size: 5,199 Bytes
3d1c0e1 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 |
# Copyright (c) 2025 FoundationVision
# SPDX-License-Identifier: MIT
import torch.nn as nn
import torch.distributed as dist
import infinity.models.videovae.utils.diffdist.functions as funcs
class ConsumeVariable(nn.Module):
def __init__(self, set_ones_grad=True):
"""
If set_ones_grad=True then the gradient w.r.t tensor_to_consume
is set to 1 during backprop. Otherwise, it is set to 0.
"""
super(ConsumeVariable, self).__init__()
self.set_ones_grad = set_ones_grad
def forward(self, tensor_to_consume, *tensors_to_return):
tensors_to_return = funcs.ConsumeVariableFunc.apply(
tensor_to_consume, self.set_ones_grad, *tensors_to_return)
return tensors_to_return
class Send(nn.Module):
def __init__(self, dst, group=dist.group.WORLD, tag=0):
super(Send, self).__init__()
self.dst = dst
self.group = group
self.tag = tag
def forward(self, tensor):
return funcs.SendFunc.apply(tensor, self.dst, self.group, self.tag)
class Recv(nn.Module):
def __init__(self,
src=None,
group=dist.group.WORLD,
tag=0,
next_backprop=None,
inplace=True):
super(Recv, self).__init__()
self.next_backprop = next_backprop
self.src = src
self.group = group
self.tag = tag
self.inplace = inplace
self.consume = None
if self.next_backprop is not None:
self.consume = ConsumeVariable()
def forward(self, tensor):
if self.consume:
tensor, = self.consume(self.next_backprop, tensor)
tensor, sender = funcs.RecvFunc.apply(tensor, self.src, self.group,
self.tag, self.inplace)
return tensor, sender.item()
class Broadcast(nn.Module):
def __init__(self,
src,
group=dist.group.WORLD,
next_backprop=None,
inplace=True):
super(Broadcast, self).__init__()
self.src = src
self.group = group
self.next_backprop = next_backprop
self.inplace = inplace
self.consume = None
if self.next_backprop is not None:
self.consume = ConsumeVariable()
def forward(self, tensor):
if self.consume:
tensor, = self.consume(self.next_backprop, tensor)
return funcs.BroadcastFunc.apply(tensor, self.src, self.group,
self.inplace)
class Gather(nn.Module):
def __init__(self,
dst=None,
group=dist.group.WORLD,
next_backprop=None,
inplace=True):
super(Gather, self).__init__()
self.dst = dst
self.group = group
self.next_backprop = next_backprop
self.inplace = inplace
self.consume = None
if self.next_backprop is not None:
self.consume = ConsumeVariable()
def forward(self, tensor, gather_list=None):
if self.consume:
tensor, = self.consume(self.next_backprop, tensor)
if dist.get_rank(self.group) == self.dst:
return list(
funcs.GatherFunc.apply(tensor, self.dst, self.group,
self.inplace, *gather_list))
else:
return funcs.GatherFunc.apply(tensor, self.dst, self.group,
self.inplace, None)
class Scatter(nn.Module):
def __init__(self,
src=None,
group=dist.group.WORLD,
next_backprop=None,
inplace=True):
super(Scatter, self).__init__()
self.src = src
self.group = group
self.next_backprop = next_backprop
self.inplace = inplace
self.consume = None
if self.next_backprop is not None:
self.consume = ConsumeVariable()
def forward(self, tensor, scatter_list=None):
if self.consume:
tensor, = self.consume(self.next_backprop, tensor)
if dist.get_rank(self.group) == self.src:
return funcs.ScatterFunc.apply(tensor, self.src, self.group,
self.inplace, *scatter_list)
else:
return funcs.ScatterFunc.apply(tensor, self.src, self.group,
self.inplace, None)
class AllGather(nn.Module):
def __init__(self,
group=dist.group.WORLD,
next_backprop=None,
inplace=True):
super(AllGather, self).__init__()
self.group = group
self.next_backprop = next_backprop
self.inplace = inplace
self.consume = None
if self.next_backprop is not None:
self.consume = ConsumeVariable()
def forward(self, gather_list, tensor):
if self.consume:
tensor, = self.consume(self.next_backprop, tensor)
return list(
funcs.AllGatherFunc.apply(tensor, self.group, self.inplace,
*gather_list))
|