pytorch  1.8.2
About: PyTorch provides Tensor computation (like NumPy) with strong GPU acceleration and Deep Neural Networks (in Python) built on a tape-based autograd system. LTS (Long Term Support) release.
  Fossies Dox: pytorch-1.8.2.tar.gz  ("unofficial" and yet experimental doxygen-generated source code documentation)  
lengths_reducer_fused_4bit_rowwise_fp16_fake_op.cc
Go to the documentation of this file.
1#include "lengths_reducer_fused_4bit_rowwise_fp16_fake_op.h"
2
3namespace caffe2 {
4
5REGISTER_CPU_OPERATOR(
6 SparseLengthsSumFused4BitRowwiseFakeFP16NNPI,
7 SparseLengthsFused4BitRowwiseFakeFP16Op<
8 CPUContext,
9 /*with_weights=*/false>);
10OPERATOR_SCHEMA(SparseLengthsSumFused4BitRowwiseFakeFP16NNPI)
11 .NumInputs(3)
12 .NumOutputs(1)
13 .ValueKeyLengthInputFillers(
14 SparseLengthsFused4BitRowwiseFakeFP16Op<CPUContext, false>::DATA,
15 SparseLengthsFused4BitRowwiseFakeFP16Op<CPUContext, false>::INDICES,
16 SparseLengthsFused4BitRowwiseFakeFP16Op<CPUContext, false>::LENGTHS)
17 .SetDoc(R"DOC(
18Performs the same operation as SparseLengthsSum, but operating on
194-bit rowwise quantized matrices with fused storage (where each row
20stores quantized values, and then 2-byte scale and 2-byte bias).
21)DOC")
22 .Input(
23 0,
24 "DATA",
25 "uint8 tensor obtained with "
26 "operator FloatToFused4BitRowwiseQuantized")
27 .Input(
28 1,
29 "INDICES",
30 "Integer vector containing indices of the first "
31 "dimension of DATA for the slices that are being aggregated")
32 .Input(
33 2,
34 "LENGTHS",
35 "Vector with the same sum of elements as the first dimension of DATA")
36 .Output(0, "output", "output")
37 .InheritOnnxSchema();
38NO_GRADIENT(SparseLengthsSumFused4BitRowwiseFakeFP16NNPI);
39
40REGISTER_CPU_OPERATOR(
41 SparseLengthsSumFused4BitRowwiseFakeFP16EmbeddingOnly,
42 SparseLengthsFused4BitRowwiseFakeFP16Op<
43 CPUContext,
44 /*with_weights=*/false,
45 /*use_fp16_for_embedding_only=*/true>);
46OPERATOR_SCHEMA(SparseLengthsSumFused4BitRowwiseFakeFP16EmbeddingOnly)
47 .NumInputs(3)
48 .NumOutputs(1)
49 .ValueKeyLengthInputFillers(
50 SparseLengthsFused4BitRowwiseFakeFP16Op<CPUContext, false, true>::DATA,
51 SparseLengthsFused4BitRowwiseFakeFP16Op<CPUContext, false, true>::
52 INDICES,
53 SparseLengthsFused4BitRowwiseFakeFP16Op<CPUContext, false, true>::
54 LENGTHS)
55 .SetDoc(R"DOC(
56Performs the same operation as SparseLengthsSum, but operating on
574-bit rowwise quantized matrices with fused storage (where each row
58stores quantized values, and then 2-byte scale and 2-byte bias).
59Convert only embedding entries using fake fp16.
60)DOC")
61 .Input(
62 0,
63 "DATA",
64 "uint8 tensor obtained with "
65 "operator FloatToFused4BitRowwiseQuantized")
66 .Input(
67 1,
68 "INDICES",
69 "Integer vector containing indices of the first "
70 "dimension of DATA for the slices that are being aggregated")
71 .Input(
72 2,
73 "LENGTHS",
74 "Vector with the same sum of elements as the first dimension of DATA")
75 .Output(0, "output", "output")
76 .InheritOnnxSchema();
77NO_GRADIENT(SparseLengthsSumFused4BitRowwiseFakeFP16EmbeddingOnly);
78
79REGISTER_CPU_OPERATOR(
80 SparseLengthsWeightedSumFused4BitRowwiseFakeFP16NNPI,
81 SparseLengthsFused4BitRowwiseFakeFP16Op<
82 CPUContext,
83 /*with_weights=*/true>);
84OPERATOR_SCHEMA(SparseLengthsWeightedSumFused4BitRowwiseFakeFP16NNPI)
85 .NumInputs(4)
86 .NumOutputs(1)
87 .WeightedValueKeyLengthInputFillers(
88 SparseLengthsFused4BitRowwiseFakeFP16Op<CPUContext, true>::DATA,
89 SparseLengthsFused4BitRowwiseFakeFP16Op<CPUContext, true>::INDICES,
90 SparseLengthsFused4BitRowwiseFakeFP16Op<CPUContext, true>::LENGTHS,
91 SparseLengthsFused4BitRowwiseFakeFP16Op<CPUContext, true>::WEIGHTS)
92 .SetDoc(R"DOC(
93Performs the same operation as SparseLengthsWeightedSum,
94but operating on 4-bit rowwise quantized matrices with fused storage
95(where each row stores quantized values, and then 2-byte scale and 2-byte bias).
96)DOC")
97 .Input(
98 0,
99 "DATA",
100 "uint8 tensor obtained with "
101 "operator FloatToFused4BitRowwiseQuantized")
102 .Input(
103 1,
104 "INDICES",
105 "Integer vector containing indices of the first "
106 "dimension of DATA for the slices that are being aggregated")
107 .Input(
108 2,
109 "LENGTHS",
110 "Vector with the same sum of elements as the first dimension of DATA")
111 .Input(
112 3,
113 "WEIGHTS",
114 "Vector of weights to scale rows of DATA with before reduction")
115 .Output(0, "output", "output");
116
117NO_GRADIENT(SparseLengthsWeightedSumFused4BitRowwiseFakeFP16NNPI);
118
119REGISTER_CPU_OPERATOR(
120 SparseLengthsWeightedSumFused4BitRowwiseFakeFP16EmbeddingOnly,
121 SparseLengthsFused4BitRowwiseFakeFP16Op<
122 CPUContext,
123 /*with_weights=*/true,
124 /*use_fp16_for_embedding_only=*/true>);
125OPERATOR_SCHEMA(SparseLengthsWeightedSumFused4BitRowwiseFakeFP16EmbeddingOnly)
126 .NumInputs(4)
127 .NumOutputs(1)
128 .WeightedValueKeyLengthInputFillers(
129 SparseLengthsFused4BitRowwiseFakeFP16Op<CPUContext, true, true>::DATA,
130 SparseLengthsFused4BitRowwiseFakeFP16Op<CPUContext, true, true>::
131 INDICES,
132 SparseLengthsFused4BitRowwiseFakeFP16Op<CPUContext, true, true>::
133 LENGTHS,
134 SparseLengthsFused4BitRowwiseFakeFP16Op<CPUContext, true, true>::
135 WEIGHTS)
136 .SetDoc(R"DOC(
137Performs the same operation as SparseLengthsWeightedSum,
138but operating on 4-bit rowwise quantized matrices with fused storage
139(where each row stores quantized values, and then 2-byte scale and 2-byte bias).
140Convert only embedding entries using fake fp16.
141)DOC")
142 .Input(
143 0,
144 "DATA",
145 "uint8 tensor obtained with "
146 "operator FloatToFused4BitRowwiseQuantized")
147 .Input(
148 1,
149 "INDICES",
150 "Integer vector containing indices of the first "
151 "dimension of DATA for the slices that are being aggregated")
152 .Input(
153 2,
154 "LENGTHS",
155 "Vector with the same sum of elements as the first dimension of DATA")
156 .Input(
157 3,
158 "WEIGHTS",
159 "Vector of weights to scale rows of DATA with before reduction")
160 .Output(0, "output", "output");
161
162NO_GRADIENT(SparseLengthsWeightedSumFused4BitRowwiseFakeFP16EmbeddingOnly);
163} // namespace caffe2
caffe2::CPUContext
The CPU Context, representing the bare minimum of what a Context class in Caffe2 should implement.
Definition: context.h:47
lengths_reducer_fused_4bit_rowwise_fp16_fake_op.h
caffe2
Copyright (c) 2016-present, Facebook, Inc.
Definition: blob.h:13
caffe2::REGISTER_CPU_OPERATOR
REGISTER_CPU_OPERATOR(ATen, ATenOp< CPUContext >)
caffe2::OPERATOR_SCHEMA
OPERATOR_SCHEMA(ATen)
caffe2::NO_GRADIENT
NO_GRADIENT(SparseLengthsSumFused4BitRowwiseFakeFP16NNPI)
caffe2::INDICES
true SparseLengthsFused8BitRowwiseOp< CPUContext, true >::WEIGHTS uint8 tensor obtained with INDICES
Definition: lengths_reducer_fused_8bit_rowwise_ops.cc:62
caffe2::LENGTHS
false SparseLengthsFused4BitRowwiseFakeFP16Op< CPUContext, false >::LENGTHS uint8 tensor obtained with LENGTHS
Definition: lengths_reducer_fused_4bit_rowwise_fp16_fake_op.cc:30