quantize_config.json
3.1 KB · 120 lines · json Raw
1 {
2 "fp16": {},
3 "q8": {
4 "per_model_config": {
5 "model": {
6 "op_types": [
7 "Add",
8 "Concat",
9 "Conv",
10 "ConvTranspose",
11 "Div",
12 "Equal",
13 "Erf",
14 "Expand",
15 "Gather",
16 "MatMul",
17 "Mul",
18 "Pow",
19 "ReduceMean",
20 "Relu",
21 "Reshape",
22 "Resize",
23 "Shape",
24 "Slice",
25 "Softmax",
26 "Sqrt",
27 "Squeeze",
28 "Sub",
29 "Transpose",
30 "Unsqueeze",
31 "Where"
32 ],
33 "weight_type": "QUInt8"
34 }
35 },
36 "per_channel": true,
37 "reduce_range": true
38 },
39 "int8": {
40 "per_model_config": {
41 "model": {
42 "op_types": [
43 "Add",
44 "Concat",
45 "Conv",
46 "ConvTranspose",
47 "Div",
48 "Equal",
49 "Erf",
50 "Expand",
51 "Gather",
52 "MatMul",
53 "Mul",
54 "Pow",
55 "ReduceMean",
56 "Relu",
57 "Reshape",
58 "Resize",
59 "Shape",
60 "Slice",
61 "Softmax",
62 "Sqrt",
63 "Squeeze",
64 "Sub",
65 "Transpose",
66 "Unsqueeze",
67 "Where"
68 ],
69 "weight_type": "QInt8"
70 }
71 },
72 "per_channel": true,
73 "reduce_range": true
74 },
75 "uint8": {
76 "per_model_config": {
77 "model": {
78 "op_types": [
79 "Add",
80 "Concat",
81 "Conv",
82 "ConvTranspose",
83 "Div",
84 "Equal",
85 "Erf",
86 "Expand",
87 "Gather",
88 "MatMul",
89 "Mul",
90 "Pow",
91 "ReduceMean",
92 "Relu",
93 "Reshape",
94 "Resize",
95 "Shape",
96 "Slice",
97 "Softmax",
98 "Sqrt",
99 "Squeeze",
100 "Sub",
101 "Transpose",
102 "Unsqueeze",
103 "Where"
104 ],
105 "weight_type": "QUInt8"
106 }
107 },
108 "per_channel": true,
109 "reduce_range": true
110 },
111 "q4": {
112 "block_size": 32,
113 "is_symmetric": true,
114 "accuracy_level": null
115 },
116 "bnb4": {
117 "block_size": 64,
118 "quant_type": 1
119 }
120 }