251 lines
14 KiB
JSON
251 lines
14 KiB
JSON
{
|
|
"model_config": {
|
|
"all_node_type": "int16",
|
|
"model_output_type": "float32",
|
|
"activation": {
|
|
"calibration_type": ["max"],
|
|
"num_bin": [1024, 2048, 4096],
|
|
"max_num_bin": 16384,
|
|
"max_percentile": 1.0,
|
|
"per_channel": true,
|
|
"asymmetric": [true]
|
|
},
|
|
"weight": {
|
|
"bias_correction": {
|
|
"metric": "mae"
|
|
}
|
|
},
|
|
"modelwise_search": {
|
|
"metric": "mae"
|
|
}
|
|
},
|
|
"op_config": {
|
|
"ReduceMean": {"qtype": "int16"},
|
|
"Sub": {"qtype": "int16"},
|
|
"Softmax": {"qtype": "int16"}
|
|
},
|
|
"node_config": {
|
|
"/t_embedder/Mul": {"qtype": "float32"},
|
|
"/t_embedder/Cos": {"qtype": "float32"},
|
|
"/t_embedder/Sin": {"qtype": "float32"},
|
|
"/t_embedder/Concat": {"qtype": "float32"},
|
|
"/freq_embedder/Mul": {"qtype": "float32"},
|
|
"/freq_embedder/Cos": {"qtype": "float32"},
|
|
"/freq_embedder/Sin": {"qtype": "float32"},
|
|
"/freq_embedder/Concat": {"qtype": "float32"},
|
|
"/blocks.0/attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.0/attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.0/cross_attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.0/cross_attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.0/cross_attn/k_norm/Mul_1": {"qtype": "int16"},
|
|
"/blocks.0/ffn/fc1/MatMul": {"qtype": "int16"},
|
|
"/blocks.0/ffn/act/Mul": {"qtype": "int16"},
|
|
"/blocks.0/ffn/act/Mul_1": {"qtype": "int16"},
|
|
"/blocks.0/ffn/act/Mul_2": {"qtype": "int16"},
|
|
"/blocks.0/ffn/act/Add": {"qtype": "int16"},
|
|
"/blocks.0/ffn/act/Mul_3": {"qtype": "int16"},
|
|
"/blocks.0/ffn/act/Tanh": {"qtype": "int16"},
|
|
"/blocks.0/norm1/Mul_2": {"qtype": "int16"},
|
|
"/blocks.0/cross_attn/k_norm/Div_1_reciprocal": {"qtype": "int16"},
|
|
"/blocks.0/Add": {"qtype": "int16"},
|
|
"/blocks.1/attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.1/attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.1/cross_attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.1/cross_attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.1/cross_attn/k_norm/Mul_1": {"qtype": "int16"},
|
|
"/blocks.1/ffn/fc1/MatMul": {"qtype": "int16"},
|
|
"/blocks.1/ffn/act/Mul": {"qtype": "int16"},
|
|
"/blocks.1/ffn/act/Mul_1": {"qtype": "int16"},
|
|
"/blocks.1/ffn/act/Mul_2": {"qtype": "int16"},
|
|
"/blocks.1/ffn/act/Add": {"qtype": "int16"},
|
|
"/blocks.1/ffn/act/Mul_3": {"qtype": "int16"},
|
|
"/blocks.1/ffn/act/Tanh": {"qtype": "int16"},
|
|
"/blocks.1/norm1/Mul_2": {"qtype": "int16"},
|
|
"/blocks.1/cross_attn/k_norm/Div_1_reciprocal": {"qtype": "int16"},
|
|
"/blocks.1/Add": {"qtype": "int16"},
|
|
"/blocks.2/attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.2/attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.2/cross_attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.2/cross_attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.2/cross_attn/k_norm/Mul_1": {"qtype": "int16"},
|
|
"/blocks.2/ffn/fc1/MatMul": {"qtype": "int16"},
|
|
"/blocks.2/ffn/act/Mul": {"qtype": "int16"},
|
|
"/blocks.2/ffn/act/Mul_1": {"qtype": "int16"},
|
|
"/blocks.2/ffn/act/Mul_2": {"qtype": "int16"},
|
|
"/blocks.2/ffn/act/Add": {"qtype": "int16"},
|
|
"/blocks.2/ffn/act/Mul_3": {"qtype": "int16"},
|
|
"/blocks.2/ffn/act/Tanh": {"qtype": "int16"},
|
|
"/blocks.2/norm1/Mul_2": {"qtype": "int16"},
|
|
"/blocks.2/cross_attn/k_norm/Div_1_reciprocal": {"qtype": "int16"},
|
|
"/blocks.2/Add": {"qtype": "int16"},
|
|
"/blocks.3/attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.3/attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.3/cross_attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.3/cross_attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.3/cross_attn/k_norm/Mul_1": {"qtype": "int16"},
|
|
"/blocks.3/ffn/fc1/MatMul": {"qtype": "int16"},
|
|
"/blocks.3/ffn/act/Mul": {"qtype": "int16"},
|
|
"/blocks.3/ffn/act/Mul_1": {"qtype": "int16"},
|
|
"/blocks.3/ffn/act/Mul_2": {"qtype": "int16"},
|
|
"/blocks.3/ffn/act/Add": {"qtype": "int16"},
|
|
"/blocks.3/ffn/act/Mul_3": {"qtype": "int16"},
|
|
"/blocks.3/ffn/act/Tanh": {"qtype": "int16"},
|
|
"/blocks.3/norm1/Mul_2": {"qtype": "int16"},
|
|
"/blocks.3/cross_attn/k_norm/Div_1_reciprocal": {"qtype": "int16"},
|
|
"/blocks.3/Add": {"qtype": "int16"},
|
|
"/blocks.4/attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.4/attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.4/cross_attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.4/cross_attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.4/cross_attn/k_norm/Mul_1": {"qtype": "int16"},
|
|
"/blocks.4/ffn/fc1/MatMul": {"qtype": "int16"},
|
|
"/blocks.4/ffn/act/Mul": {"qtype": "int16"},
|
|
"/blocks.4/ffn/act/Mul_1": {"qtype": "int16"},
|
|
"/blocks.4/ffn/act/Mul_2": {"qtype": "int16"},
|
|
"/blocks.4/ffn/act/Add": {"qtype": "int16"},
|
|
"/blocks.4/ffn/act/Mul_3": {"qtype": "int16"},
|
|
"/blocks.4/ffn/act/Tanh": {"qtype": "int16"},
|
|
"/blocks.4/norm1/Mul_2": {"qtype": "int16"},
|
|
"/blocks.4/cross_attn/k_norm/Div_1_reciprocal": {"qtype": "int16"},
|
|
"/blocks.4/Add": {"qtype": "int16"},
|
|
"/blocks.5/attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.5/attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.5/cross_attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.5/cross_attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.5/cross_attn/k_norm/Mul_1": {"qtype": "int16"},
|
|
"/blocks.5/ffn/fc1/MatMul": {"qtype": "int16"},
|
|
"/blocks.5/ffn/act/Mul": {"qtype": "int16"},
|
|
"/blocks.5/ffn/act/Mul_1": {"qtype": "int16"},
|
|
"/blocks.5/ffn/act/Mul_2": {"qtype": "int16"},
|
|
"/blocks.5/ffn/act/Add": {"qtype": "int16"},
|
|
"/blocks.5/ffn/act/Mul_3": {"qtype": "int16"},
|
|
"/blocks.5/ffn/act/Tanh": {"qtype": "int16"},
|
|
"/blocks.5/norm1/Mul_2": {"qtype": "int16"},
|
|
"/blocks.5/cross_attn/k_norm/Div_1_reciprocal": {"qtype": "int16"},
|
|
"/blocks.5/Add": {"qtype": "int16"},
|
|
"/blocks.6/attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.6/attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.6/cross_attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.6/cross_attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.6/cross_attn/k_norm/Mul_1": {"qtype": "int16"},
|
|
"/blocks.6/ffn/fc1/MatMul": {"qtype": "int16"},
|
|
"/blocks.6/ffn/act/Mul": {"qtype": "int16"},
|
|
"/blocks.6/ffn/act/Mul_1": {"qtype": "int16"},
|
|
"/blocks.6/ffn/act/Mul_2": {"qtype": "int16"},
|
|
"/blocks.6/ffn/act/Add": {"qtype": "int16"},
|
|
"/blocks.6/ffn/act/Mul_3": {"qtype": "int16"},
|
|
"/blocks.6/ffn/act/Tanh": {"qtype": "int16"},
|
|
"/blocks.6/norm1/Mul_2": {"qtype": "int16"},
|
|
"/blocks.6/cross_attn/k_norm/Div_1_reciprocal": {"qtype": "int16"},
|
|
"/blocks.6/Add": {"qtype": "int16"},
|
|
"/blocks.7/attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.7/attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.7/cross_attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.7/cross_attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.7/cross_attn/k_norm/Mul_1": {"qtype": "int16"},
|
|
"/blocks.7/ffn/fc1/MatMul": {"qtype": "int16"},
|
|
"/blocks.7/ffn/act/Mul": {"qtype": "int16"},
|
|
"/blocks.7/ffn/act/Mul_1": {"qtype": "int16"},
|
|
"/blocks.7/ffn/act/Mul_2": {"qtype": "int16"},
|
|
"/blocks.7/ffn/act/Add": {"qtype": "int16"},
|
|
"/blocks.7/ffn/act/Mul_3": {"qtype": "int16"},
|
|
"/blocks.7/ffn/act/Tanh": {"qtype": "int16"},
|
|
"/blocks.7/norm1/Mul_2": {"qtype": "int16"},
|
|
"/blocks.7/cross_attn/k_norm/Div_1_reciprocal": {"qtype": "int16"},
|
|
"/blocks.7/Add": {"qtype": "int16"},
|
|
"/blocks.8/attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.8/attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.8/cross_attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.8/cross_attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.8/cross_attn/k_norm/Mul_1": {"qtype": "int16"},
|
|
"/blocks.8/ffn/fc1/MatMul": {"qtype": "int16"},
|
|
"/blocks.8/ffn/act/Mul": {"qtype": "int16"},
|
|
"/blocks.8/ffn/act/Mul_1": {"qtype": "int16"},
|
|
"/blocks.8/ffn/act/Mul_2": {"qtype": "int16"},
|
|
"/blocks.8/ffn/act/Add": {"qtype": "int16"},
|
|
"/blocks.8/ffn/act/Mul_3": {"qtype": "int16"},
|
|
"/blocks.8/ffn/act/Tanh": {"qtype": "int16"},
|
|
"/blocks.8/norm1/Mul_2": {"qtype": "int16"},
|
|
"/blocks.8/cross_attn/k_norm/Div_1_reciprocal": {"qtype": "int16"},
|
|
"/blocks.8/Add": {"qtype": "int16"},
|
|
"/blocks.9/attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.9/attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.9/cross_attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.9/cross_attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.9/cross_attn/k_norm/Mul_1": {"qtype": "int16"},
|
|
"/blocks.9/ffn/fc1/MatMul": {"qtype": "int16"},
|
|
"/blocks.9/ffn/act/Mul": {"qtype": "int16"},
|
|
"/blocks.9/ffn/act/Mul_1": {"qtype": "int16"},
|
|
"/blocks.9/ffn/act/Mul_2": {"qtype": "int16"},
|
|
"/blocks.9/ffn/act/Add": {"qtype": "int16"},
|
|
"/blocks.9/ffn/act/Mul_3": {"qtype": "int16"},
|
|
"/blocks.9/ffn/act/Tanh": {"qtype": "int16"},
|
|
"/blocks.9/norm1/Mul_2": {"qtype": "int16"},
|
|
"/blocks.9/cross_attn/k_norm/Div_1_reciprocal": {"qtype": "int16"},
|
|
"/blocks.9/Add": {"qtype": "int16"},
|
|
"/blocks.10/attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.10/attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.10/cross_attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.10/cross_attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.10/cross_attn/k_norm/Mul_1": {"qtype": "int16"},
|
|
"/blocks.10/ffn/fc1/MatMul": {"qtype": "int16"},
|
|
"/blocks.10/ffn/act/Mul": {"qtype": "int16"},
|
|
"/blocks.10/ffn/act/Mul_1": {"qtype": "int16"},
|
|
"/blocks.10/ffn/act/Mul_2": {"qtype": "int16"},
|
|
"/blocks.10/ffn/act/Add": {"qtype": "int16"},
|
|
"/blocks.10/ffn/act/Mul_3": {"qtype": "int16"},
|
|
"/blocks.10/ffn/act/Tanh": {"qtype": "int16"},
|
|
"/blocks.10/norm1/Mul_2": {"qtype": "int16"},
|
|
"/blocks.10/cross_attn/k_norm/Div_1_reciprocal": {"qtype": "int16"},
|
|
"/blocks.10/Add": {"qtype": "int16"},
|
|
"/blocks.11/attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.11/attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.11/cross_attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.11/cross_attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.11/cross_attn/k_norm/Mul_1": {"qtype": "int16"},
|
|
"/blocks.11/ffn/fc1/MatMul": {"qtype": "int16"},
|
|
"/blocks.11/ffn/act/Mul": {"qtype": "int16"},
|
|
"/blocks.11/ffn/act/Mul_1": {"qtype": "int16"},
|
|
"/blocks.11/ffn/act/Mul_2": {"qtype": "int16"},
|
|
"/blocks.11/ffn/act/Add": {"qtype": "int16"},
|
|
"/blocks.11/ffn/act/Mul_3": {"qtype": "int16"},
|
|
"/blocks.11/ffn/act/Tanh": {"qtype": "int16"},
|
|
"/blocks.11/norm1/Mul_2": {"qtype": "int16"},
|
|
"/blocks.11/cross_attn/k_norm/Div_1_reciprocal": {"qtype": "int16"},
|
|
"/blocks.11/Add": {"qtype": "int16"},
|
|
"/blocks.12/attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.12/attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.12/cross_attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.12/cross_attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.12/cross_attn/k_norm/Mul_1": {"qtype": "int16"},
|
|
"/blocks.12/ffn/fc1/MatMul": {"qtype": "int16"},
|
|
"/blocks.12/ffn/act/Mul": {"qtype": "int16"},
|
|
"/blocks.12/ffn/act/Mul_1": {"qtype": "int16"},
|
|
"/blocks.12/ffn/act/Mul_2": {"qtype": "int16"},
|
|
"/blocks.12/ffn/act/Add": {"qtype": "int16"},
|
|
"/blocks.12/ffn/act/Mul_3": {"qtype": "int16"},
|
|
"/blocks.12/ffn/act/Tanh": {"qtype": "int16"},
|
|
"/blocks.12/norm1/Mul_2": {"qtype": "int16"},
|
|
"/blocks.12/cross_attn/k_norm/Div_1_reciprocal": {"qtype": "int16"},
|
|
"/blocks.12/Add": {"qtype": "int16"},
|
|
"/blocks.13/attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.13/attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.13/cross_attn/MatMul": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.13/cross_attn/MatMul_1": {"InputType0": "int16", "InputType1": "int16"},
|
|
"/blocks.13/cross_attn/k_norm/Mul_1": {"qtype": "int16"},
|
|
"/blocks.13/ffn/fc1/MatMul": {"qtype": "int16"},
|
|
"/blocks.13/ffn/act/Mul": {"qtype": "int16"},
|
|
"/blocks.13/ffn/act/Mul_1": {"qtype": "int16"},
|
|
"/blocks.13/ffn/act/Mul_2": {"qtype": "int16"},
|
|
"/blocks.13/ffn/act/Add": {"qtype": "int16"},
|
|
"/blocks.13/ffn/act/Mul_3": {"qtype": "int16"},
|
|
"/blocks.13/ffn/act/Tanh": {"qtype": "int16"},
|
|
"/blocks.13/norm1/Mul_2": {"qtype": "int16"},
|
|
"/blocks.13/cross_attn/k_norm/Div_1_reciprocal": {"qtype": "int16"},
|
|
"/blocks.13/Add": {"qtype": "int16"},
|
|
"/blocks.13/norm3/Div_1_reciprocal": {"qtype": "int16"},
|
|
"/final_layer/ffn_final/act/Mul_1": {"qtype": "int16"},
|
|
"/final_layer/ffn_final/act/Mul_2 ": {"qtype": "int16"},
|
|
"/final_layer/norm_final/Div_1_reciprocal": {"qtype": "float32"}
|
|
}
|
|
} |