distilvit / quantize_config.json
tarekziade's picture
Upload 16 files
61f54c3 verified
raw
history blame
3.11 kB
{
"per_channel": false,
"reduce_range": false,
"per_model_config": {
"decoder_with_past_model": {
"op_types": [
"Sqrt",
"Split",
"Unsqueeze",
"Tanh",
"Gemm",
"MatMul",
"Cast",
"Div",
"Pow",
"Add",
"Constant",
"Reshape",
"ConstantOfShape",
"Range",
"Slice",
"Where",
"Concat",
"Shape",
"Squeeze",
"ReduceMean",
"Softmax",
"Gather",
"Sub",
"Mul",
"Transpose"
],
"weight_type": "QInt8"
},
"decoder_model": {
"op_types": [
"Sqrt",
"Split",
"Unsqueeze",
"Tanh",
"Gemm",
"MatMul",
"Cast",
"Div",
"Pow",
"Add",
"Constant",
"Reshape",
"ConstantOfShape",
"Range",
"Slice",
"Where",
"Concat",
"Shape",
"Squeeze",
"ReduceMean",
"Softmax",
"Gather",
"Sub",
"Mul",
"Transpose"
],
"weight_type": "QInt8"
},
"encoder_model": {
"op_types": [
"Sqrt",
"Unsqueeze",
"Equal",
"MatMul",
"Div",
"Pow",
"Add",
"Erf",
"Constant",
"Reshape",
"ConstantOfShape",
"Slice",
"Where",
"Conv",
"Concat",
"Shape",
"ReduceMean",
"Softmax",
"Gather",
"Expand",
"Mul",
"Sub",
"Transpose"
],
"weight_type": "QUInt8"
},
"decoder_model_merged": {
"op_types": [
"Sqrt",
"Split",
"If",
"Unsqueeze",
"Tanh",
"Gemm",
"MatMul",
"Cast",
"Div",
"Pow",
"Add",
"Constant",
"Reshape",
"ConstantOfShape",
"Range",
"Slice",
"Where",
"Concat",
"Shape",
"Squeeze",
"ReduceMean",
"Softmax",
"Gather",
"Sub",
"Mul",
"Transpose"
],
"weight_type": "QInt8"
}
}
}