-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathsoftmax.py
executable file
·149 lines (126 loc) · 6.09 KB
/
softmax.py
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
import torch
import mhalib
###########################################################################################
class FastSoftmaxFunction(torch.autograd.Function):
@staticmethod
def forward(cxt, input, dim, batch, seqlen, heads, stream, sync, timers):
if timers: timers['start_fprop'].record()
mhalib.FastSoftmaxFprop(input, batch, seqlen, heads, stream, sync)
if timers: timers['stop_fprop'].record()
cxt.save_for_backward(input,seqlen)
cxt.dim = dim
cxt.batch = batch
cxt.heads = heads
cxt.stream = stream
cxt.sync = sync
cxt.timers = timers
return input
@staticmethod
def backward(cxt, grad_output):
output, seqlen, = cxt.saved_tensors
dim = cxt.dim
batch = cxt.batch
heads = cxt.heads
if cxt.timers: cxt.timers['start_dgrad'].record()
mhalib.FastSoftmaxBprop(output, grad_output, batch, seqlen, heads, cxt.stream, cxt.sync)
if cxt.timers: cxt.timers['stop_dgrad'].record()
return grad_output, None, None, None, None, None, None, None
class FastSoftmax(torch.nn.Module):
def __init__(self, dim=None, stream=True, sync=True, timer=False):
super(FastSoftmax, self).__init__()
self.dim = dim
self.stream = stream
self.sync = sync
if timer:
self.timers = {'start_fprop':torch.cuda.Event(enable_timing=True),
'start_dgrad':torch.cuda.Event(enable_timing=True),
'stop_fprop':torch.cuda.Event(enable_timing=True),
'stop_dgrad':torch.cuda.Event(enable_timing=True)}
else:
self.timers = None
def forward(self, input, batch, seqlen, heads):
return FastSoftmaxFunction.apply(input, self.dim, batch, seqlen, heads, self.stream, self.sync, self.timers)
###########################################################################################
class FastMaskSoftmaxFunction(torch.autograd.Function):
@staticmethod
def forward(cxt, input, mask, dim, batch, seqlen, heads, stream, sync, timers):
if timers: timers['start_fprop'].record()
mhalib.FastMaskSoftmaxFprop(input, mask, batch, seqlen, heads, stream, sync)
if timers: timers['stop_fprop'].record()
cxt.save_for_backward(input,seqlen)
cxt.dim = dim
cxt.batch = batch
cxt.heads = heads
cxt.stream = stream
cxt.sync = sync
cxt.timers = timers
return input
@staticmethod
def backward(cxt, grad_output):
output, seqlen, = cxt.saved_tensors
dim = cxt.dim
batch = cxt.batch
heads = cxt.heads
if cxt.timers: cxt.timers['start_dgrad'].record()
mhalib.FastSoftmaxBprop(output, grad_output, batch, seqlen, heads, cxt.stream, cxt.sync)
if cxt.timers: cxt.timers['stop_dgrad'].record()
return grad_output, None, None, None, None, None, None, None, None, None, None, None
class FastMaskSoftmax(torch.nn.Module):
def __init__(self, dim=None, stream=True, sync=True, timer=False):
super(FastMaskSoftmax, self).__init__()
self.dim = dim
self.stream = stream
self.sync = sync
if timer:
self.timers = {'start_fprop':torch.cuda.Event(enable_timing=True),
'start_dgrad':torch.cuda.Event(enable_timing=True),
'stop_fprop':torch.cuda.Event(enable_timing=True),
'stop_dgrad':torch.cuda.Event(enable_timing=True)}
else:
self.timers = None
def forward(self, input, mask, batch, seqlen, heads):
return FastMaskSoftmaxFunction.apply(input, mask, self.dim, batch, seqlen, heads, self.stream, self.sync, self.timers)
###########################################################################################
class FastMaskSoftmaxDropoutFunction(torch.autograd.Function):
@staticmethod
def forward(cxt, input, mask, dim, batch, seqlen, heads, dropout_prob, stream, sync, timers, is_training):
if timers: timers['start_fprop'].record()
output, dropout_mask, = mhalib.FastMaskSoftmaxDropoutFprop(input, mask, batch, seqlen, heads, dropout_prob, stream, sync, is_training)
if timers: timers['stop_fprop'].record()
cxt.save_for_backward(input,dropout_mask,seqlen)
cxt.dim = dim
cxt.batch = batch
cxt.heads = heads
cxt.dropout_prob = dropout_prob
cxt.stream = stream
cxt.sync = sync
cxt.timers = timers
return output
@staticmethod
def backward(cxt, grad_output):
output, dropout_mask, seqlen, = cxt.saved_tensors
dim = cxt.dim
batch = cxt.batch
heads = cxt.heads
dropout_prob = cxt.dropout_prob
if cxt.timers: cxt.timers['start_dgrad'].record()
mhalib.FastMaskSoftmaxDropoutBprop(output, grad_output, dropout_mask, batch, seqlen, heads, dropout_prob, cxt.stream, cxt.sync)
if cxt.timers: cxt.timers['stop_dgrad'].record()
return grad_output, None, None, None, None, None, None, None, None, None, None, None, None, None
class FastMaskSoftmaxDropout(torch.nn.Module):
def __init__(self, dim=None, dropout_prob=None, stream=True, sync=True, timer=False):
super(FastMaskSoftmaxDropout, self).__init__()
self.dim = dim
self.dropout_prob = dropout_prob
self.stream = stream
self.sync = sync
if timer:
self.timers = {'start_fprop':torch.cuda.Event(enable_timing=True),
'start_dgrad':torch.cuda.Event(enable_timing=True),
'stop_fprop':torch.cuda.Event(enable_timing=True),
'stop_dgrad':torch.cuda.Event(enable_timing=True)}
else:
self.timers = None
def forward(self, input, mask, batch, seqlen, heads, is_training):
return FastMaskSoftmaxDropoutFunction.apply(input, mask, self.dim, batch, seqlen, heads, self.dropout_prob, self.stream, self.sync, self.timers, is_training)
###########################################################################################