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
159
160
161
162
163
164
165
166
167
168
169
170
171
|
#include "caffe2/operators/rms_norm_op.h"
#include <array>
#include <cmath>
#include <string>
#include <tuple>
#include <vector>
#include "ATen/Parallel.h"
#include "caffe2/utils/eigen_utils.h"
#include "caffe2/utils/math/utils.h"
namespace caffe2 {
template <>
template <typename T>
bool RMSNormOp<CPUContext>::DoRunWithType() {
const auto& X = Input(0);
const auto& gamma = Input(1);
const auto& beta = Input(2);
auto* Y = Output(0, X.sizes(), at::dtype<T>());
CAFFE_ENFORCE_GE(X.dim(), 2, "RMSNorm requires input dim >= 2.");
const int canonical_axis = X.canonical_axis_index(axis_);
const std::vector<int64_t> rms_dims(
X.sizes().cbegin(), X.sizes().cbegin() + canonical_axis);
auto* rrms = Output(1, rms_dims, at::dtype<T>());
const int64_t M = X.size_to_dim(canonical_axis);
const int64_t N = X.size_from_dim(canonical_axis);
CAFFE_ENFORCE_EQ(gamma.numel(), N);
CAFFE_ENFORCE_EQ(beta.numel(), N);
const T* X_data = X.template data<T>();
const T* gamma_data = gamma.template data<T>();
const T* beta_data = beta.template data<T>();
T* Y_data = Y->template data<T>();
T* rrms_data = rrms->template data<T>();
ConstEigenArrayMap<T> X_arr(X_data, N, M);
ConstEigenVectorArrayMap<T> gamma_arr(gamma_data, N);
ConstEigenVectorArrayMap<T> beta_arr(beta_data, N);
EigenArrayMap<T> Y_arr(Y_data, N, M);
at::parallel_for(0, M, 1, [&](int64_t start, int64_t end) {
for (int64_t i = start; i < end; ++i) {
const T rrms_val =
T(1) / std::sqrt(X_arr.col(i).square().mean() + static_cast<T>(eps_));
Y_arr.col(i) = rrms_val * X_arr.col(i) * gamma_arr + beta_arr;
rrms_data[i] = rrms_val;
}
});
return true;
}
template <>
template <typename T>
void RMSNormGradientOp<CPUContext>::RMSNormBackward(
int64_t M,
int64_t N,
const T* dY,
const T* X,
const T* gamma,
const T* rrms,
T* dX) {
ConstEigenArrayMap<T> dY_arr(dY, N, M);
ConstEigenArrayMap<T> X_arr(X, N, M);
ConstEigenVectorArrayMap<T> gamma_arr(gamma, N);
EigenArrayMap<T> dX_arr(dX, N, M);
const T scale = T(1) / static_cast<T>(N);
at::parallel_for(0, M, 1, [&](int64_t start, int64_t end) {
for (int64_t i = start; i < end; ++i) {
const T ds = (dY_arr.col(i) * X_arr.col(i) * gamma_arr).sum();
const T c1 = rrms[i];
const T c2 = -scale * ds * math::utils::Cube<T>(rrms[i]);
dX_arr.col(i) = c1 * dY_arr.col(i) * gamma_arr + c2 * X_arr.col(i);
}
});
}
template <>
template <typename T>
void RMSNormGradientOp<CPUContext>::GammaBetaBackward(
int64_t M,
int64_t N,
const T* dY,
const T* X,
const T* rrms,
T* dgamma,
T* dbeta) {
math::Set<T, CPUContext>(N, T(0), dgamma, &context_);
math::Set<T, CPUContext>(N, T(0), dbeta, &context_);
ConstEigenArrayMap<T> dY_arr(dY, N, M);
ConstEigenArrayMap<T> X_arr(X, N, M);
EigenVectorArrayMap<T> dgamma_arr(dgamma, N);
EigenVectorArrayMap<T> dbeta_arr(dbeta, N);
for (int64_t i = 0; i < M; ++i) {
dgamma_arr += dY_arr.col(i) * X_arr.col(i) * rrms[i];
dbeta_arr += dY_arr.col(i);
}
}
REGISTER_CPU_OPERATOR(RMSNorm, RMSNormOp<CPUContext>);
REGISTER_CPU_OPERATOR(RMSNormGradient, RMSNormGradientOp<CPUContext>);
OPERATOR_SCHEMA(RMSNorm)
.NumInputs(3)
.NumOutputs(2)
.TensorInferenceFunction([](const OperatorDef& def,
const vector<TensorShape>& in) {
std::vector<TensorShape> out(2);
const auto input_dims_long = GetDimsVector(in[0]);
const std::vector<int> input_dims(
input_dims_long.cbegin(), input_dims_long.cend());
out[0] = CreateTensorShape(input_dims, in[0].data_type());
ArgumentHelper helper(def);
const int axis = helper.GetSingleArgument<int32_t>("axis", 1);
const int canonical_axis =
canonical_axis_index_(axis, in[0].dims().size());
const std::vector<int> rms_dims(
input_dims.cbegin(), input_dims.cbegin() + canonical_axis);
out[1] = CreateTensorShape(rms_dims, in[0].data_type());
return out;
})
.Arg(
"axis",
"(int) default to 1; Describes axis of the inputs. Defaults to one "
"because the 0th axis most likely describes the batch size")
.Arg(
"epsilon",
"(float) default to 0.001. Small value to be added to the stdev when"
" dividing out by that value. This prevents division by zero.")
.Input(
0,
"input",
"Input tensor which layer normalization will be applied to")
.Input(
1,
"gamma",
"scale tensor for elementwise_affine, the shape should be the same as "
"the dimensions of X begin from axis")
.Input(
2,
"beta",
"bias tensor for elementwise_affine, the shape should be the same as "
"the dimensions of X begin from axis")
.Output(0, "output", "Normalized values")
.Output(
1,
"rrms",
"Reciprocal of root mean square for each feature vector");
OPERATOR_SCHEMA(RMSNormGradient).NumInputs(4).NumOutputs(3);
namespace {
class GetRMSNormGradient : public GradientMakerBase {
using GradientMakerBase::GradientMakerBase;
std::vector<OperatorDef> GetGradientDefs() override {
return SingleGradientDef(
"RMSNormGradient",
"",
std::vector<std::string>{GO(0), I(0), I(1), O(1)},
std::vector<std::string>{GI(0), GI(1), GI(2)});
}
};
} // namespace
REGISTER_GRADIENT(RMSNorm, GetRMSNormGradient);
} // namespace caffe2
|
