Upload folder using huggingface_hub (#3)
Browse files- Upload folder using huggingface_hub (ea53a377dff8ee0965c0646917454d9b8330b258)
- Create quantize_config.json (51b8e8e55a2d98f91d3fff9f747dc88ee2e88afd)
onnx/decoder_model_merged_quantized.onnx
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:9b8ad5f1ac3be007c6d086f36534fdde2a9e5968cc638042e8d224d1d40dbf69
|
3 |
+
size 425840661
|
onnx/decoder_model_quantized.onnx
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:5753a2747a49ad9273d743e905a8099c096c5e97f23c3a152393df9095e392d8
|
3 |
+
size 424500149
|
onnx/decoder_with_past_model_quantized.onnx
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:1e5efdd8061f833bad572628096e3bf4f7678f07a18764231f400e4c2959710f
|
3 |
+
size 373134168
|
onnx/encodec_decode_quantized.onnx
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f8107ad02e637778e217393b53bf924a81a0fad97834e1211f2b72f7c9c65aae
|
3 |
+
size 59737547
|
onnx/encodec_encoder_quantized.onnx
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:cd3c0c8a3d18041393ae7402eb4cbaff86a1c7f795b4dc304ec2350093be5171
|
3 |
+
size 8096878
|
onnx/text_encoder_quantized.onnx
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:34ee66cb1402b39b7aef78a4fafdf5be841831f52ca671402a45976ef73fa9ac
|
3 |
+
size 109954656
|
quantize_config.json
ADDED
@@ -0,0 +1,196 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"q8": {
|
3 |
+
"per_model_config": {
|
4 |
+
"text_encoder": {
|
5 |
+
"op_types": [
|
6 |
+
"Abs",
|
7 |
+
"Add",
|
8 |
+
"Cast",
|
9 |
+
"Concat",
|
10 |
+
"Constant",
|
11 |
+
"ConstantOfShape",
|
12 |
+
"Div",
|
13 |
+
"Gather",
|
14 |
+
"Greater",
|
15 |
+
"Less",
|
16 |
+
"Log",
|
17 |
+
"MatMul",
|
18 |
+
"Min",
|
19 |
+
"Mul",
|
20 |
+
"Pow",
|
21 |
+
"Range",
|
22 |
+
"ReduceMean",
|
23 |
+
"Relu",
|
24 |
+
"Reshape",
|
25 |
+
"Shape",
|
26 |
+
"Softmax",
|
27 |
+
"Sqrt",
|
28 |
+
"Sub",
|
29 |
+
"Transpose",
|
30 |
+
"Unsqueeze",
|
31 |
+
"Where"
|
32 |
+
],
|
33 |
+
"weight_type": "QInt8"
|
34 |
+
},
|
35 |
+
"decoder_with_past_model": {
|
36 |
+
"op_types": [
|
37 |
+
"Add",
|
38 |
+
"Cast",
|
39 |
+
"Concat",
|
40 |
+
"Constant",
|
41 |
+
"Div",
|
42 |
+
"Erf",
|
43 |
+
"Gather",
|
44 |
+
"MatMul",
|
45 |
+
"Mul",
|
46 |
+
"Pow",
|
47 |
+
"Range",
|
48 |
+
"ReduceMean",
|
49 |
+
"Reshape",
|
50 |
+
"Shape",
|
51 |
+
"Softmax",
|
52 |
+
"Sqrt",
|
53 |
+
"Sub",
|
54 |
+
"Transpose",
|
55 |
+
"Unsqueeze"
|
56 |
+
],
|
57 |
+
"weight_type": "QInt8"
|
58 |
+
},
|
59 |
+
"decoder_model": {
|
60 |
+
"op_types": [
|
61 |
+
"Add",
|
62 |
+
"Cast",
|
63 |
+
"Concat",
|
64 |
+
"Constant",
|
65 |
+
"ConstantOfShape",
|
66 |
+
"Div",
|
67 |
+
"Equal",
|
68 |
+
"Erf",
|
69 |
+
"Expand",
|
70 |
+
"Gather",
|
71 |
+
"Less",
|
72 |
+
"MatMul",
|
73 |
+
"Mul",
|
74 |
+
"Pow",
|
75 |
+
"Range",
|
76 |
+
"ReduceMean",
|
77 |
+
"Reshape",
|
78 |
+
"Shape",
|
79 |
+
"Slice",
|
80 |
+
"Softmax",
|
81 |
+
"Sqrt",
|
82 |
+
"Squeeze",
|
83 |
+
"Sub",
|
84 |
+
"Transpose",
|
85 |
+
"Unsqueeze",
|
86 |
+
"Where"
|
87 |
+
],
|
88 |
+
"weight_type": "QInt8"
|
89 |
+
},
|
90 |
+
"encodec_decode": {
|
91 |
+
"op_types": [
|
92 |
+
"Add",
|
93 |
+
"Cast",
|
94 |
+
"Ceil",
|
95 |
+
"Concat",
|
96 |
+
"Constant",
|
97 |
+
"ConstantOfShape",
|
98 |
+
"Conv",
|
99 |
+
"ConvTranspose",
|
100 |
+
"Div",
|
101 |
+
"Elu",
|
102 |
+
"Gather",
|
103 |
+
"LSTM",
|
104 |
+
"Pad",
|
105 |
+
"Reshape",
|
106 |
+
"Shape",
|
107 |
+
"Slice",
|
108 |
+
"Split",
|
109 |
+
"Squeeze",
|
110 |
+
"Sub",
|
111 |
+
"Transpose",
|
112 |
+
"Unsqueeze"
|
113 |
+
],
|
114 |
+
"weight_type": "QUInt8"
|
115 |
+
},
|
116 |
+
"encodec_encoder": {
|
117 |
+
"op_types": [
|
118 |
+
"Add",
|
119 |
+
"ArgMax",
|
120 |
+
"Cast",
|
121 |
+
"Ceil",
|
122 |
+
"Concat",
|
123 |
+
"ConcatFromSequence",
|
124 |
+
"Constant",
|
125 |
+
"ConstantOfShape",
|
126 |
+
"Conv",
|
127 |
+
"Div",
|
128 |
+
"Elu",
|
129 |
+
"Equal",
|
130 |
+
"Expand",
|
131 |
+
"Gather",
|
132 |
+
"Identity",
|
133 |
+
"If",
|
134 |
+
"LSTM",
|
135 |
+
"LessOrEqual",
|
136 |
+
"Loop",
|
137 |
+
"MatMul",
|
138 |
+
"Max",
|
139 |
+
"Mul",
|
140 |
+
"Neg",
|
141 |
+
"Pad",
|
142 |
+
"Pow",
|
143 |
+
"ReduceL2",
|
144 |
+
"ReduceMin",
|
145 |
+
"ReduceSum",
|
146 |
+
"Reshape",
|
147 |
+
"ScatterND",
|
148 |
+
"SequenceEmpty",
|
149 |
+
"SequenceInsert",
|
150 |
+
"Shape",
|
151 |
+
"Slice",
|
152 |
+
"Squeeze",
|
153 |
+
"Sub",
|
154 |
+
"Transpose",
|
155 |
+
"Unsqueeze",
|
156 |
+
"Where"
|
157 |
+
],
|
158 |
+
"weight_type": "QUInt8"
|
159 |
+
},
|
160 |
+
"decoder_model_merged": {
|
161 |
+
"op_types": [
|
162 |
+
"Add",
|
163 |
+
"Cast",
|
164 |
+
"Concat",
|
165 |
+
"Constant",
|
166 |
+
"ConstantOfShape",
|
167 |
+
"Div",
|
168 |
+
"Equal",
|
169 |
+
"Erf",
|
170 |
+
"Expand",
|
171 |
+
"Gather",
|
172 |
+
"If",
|
173 |
+
"Less",
|
174 |
+
"MatMul",
|
175 |
+
"Mul",
|
176 |
+
"Pow",
|
177 |
+
"Range",
|
178 |
+
"ReduceMean",
|
179 |
+
"Reshape",
|
180 |
+
"Shape",
|
181 |
+
"Slice",
|
182 |
+
"Softmax",
|
183 |
+
"Sqrt",
|
184 |
+
"Squeeze",
|
185 |
+
"Sub",
|
186 |
+
"Transpose",
|
187 |
+
"Unsqueeze",
|
188 |
+
"Where"
|
189 |
+
],
|
190 |
+
"weight_type": "QInt8"
|
191 |
+
}
|
192 |
+
},
|
193 |
+
"per_channel": false,
|
194 |
+
"reduce_range": false
|
195 |
+
}
|
196 |
+
}
|