Function _insertScale

transformers/llm/engine/tools/quantize_llm.cpp:215–277

}

static void _insertScale(MNN::NetT* _originalModel, std::map<int, std::pair<float, int32_t>> &_scales,
                         std::map<int, std::unique_ptr<MNN::TensorDescribeT>> &_tensorDescribes,
                         std::map<int, std::pair<float, int32_t>> tensorDescribesHasScaleIndex,
                          int featureBit, int weightBit, int blockSize) {
    float _featureClampValue = (float)((1 << (featureBit - 1)));
    auto type = MNN::DataType_DT_INT8;
    if(featureBit == 16){
        type = MNN::DataType_DT_INT16;
    }
    std::set<OpType> propagateOpTypes = { OpType_Raster, OpType_ReLU, OpType_ReLU6, OpType_Pooling,
                                              OpType_Interp, OpType_CropAndResize, OpType_ROIPooling};
    for (auto& op : _originalModel->oplists) {
        const auto opType = op->type;

        if(propagateOpTypes.find(opType) != propagateOpTypes.end()){
            bool needErase = false;
            for(int id = 0; id < op->inputIndexes.size() && needErase == false; ++id){
                auto iter = tensorDescribesHasScaleIndex.find(op->inputIndexes[id]);
                if(iter != tensorDescribesHasScaleIndex.end()){
                    needErase = true;
                }
            }
            for(int id = 0; id < op->outputIndexes.size() && needErase == false; ++id){
                auto iter = tensorDescribesHasScaleIndex.find(op->outputIndexes[id]);
                if(iter != tensorDescribesHasScaleIndex.end()){
                    needErase = true;
                }
            }
            if(needErase){
                for(int id = 0; id < op->inputIndexes.size(); ++id){
                    auto iter = _scales.find(op->inputIndexes[id]);
                    if(iter != _scales.end()){
                        _scales.erase(iter);
                    }
                }
                for(int id = 0; id < op->outputIndexes.size(); ++id){
                    auto iter = _scales.find(op->outputIndexes[id]);
                    if(iter != _scales.end()){
                        _scales.erase(iter);
                    }
                }
            }
        }
    }
    for (const auto iter :  _scales) {
        std::unique_ptr<MNN::TensorDescribeT> describe(new MNN::TensorDescribeT);
        auto index = iter.first;

        describe->index = index;
        describe->quantInfo.reset(new MNN::TensorQuantInfoT);
        describe->quantInfo->scale = iter.second.first;
        describe->quantInfo->zero = iter.second.second;
        describe->quantInfo->type = type;
        describe->quantInfo->min = -1 * _featureClampValue;
        describe->quantInfo->max = _featureClampValue - 1;
        auto dstiter = _tensorDescribes.find(index);
        if (dstiter == _tensorDescribes.end()) {
            _tensorDescribes.insert(std::make_pair(index, std::move(describe)));