wehub-resource-sync
64 lines
1.5 KiB
C++
64 lines
1.5 KiB
C++
#pragma once
|
|
#include <MNN/MNNDefine.h>
|
|
#include <vector>
|
|
#include <cmath>
|
|
#include <algorithm>
|
|
#include <memory>
|
|
#include <string>
|
|
#include <limits>
|
|
#include <MNN/expr/ExprCreator.hpp>
|
|
|
|
namespace MNN {
|
|
namespace Quantization {
|
|
|
|
/**
|
|
* HQQ (Half-Quadratic Quantization) 量化器 C++ 实现
|
|
* 基于半二次量化算法的权重量化和编码
|
|
*/
|
|
class HQQQuantizer {
|
|
public:
|
|
struct QuantizationConfig {
|
|
int bits = 4;
|
|
int group_size = 64;
|
|
bool optimize = true;
|
|
float lp_norm = 0.7f;
|
|
float beta = 10.0f;
|
|
float kappa = 1.01;
|
|
int iters = 20;
|
|
};
|
|
|
|
struct QuantizationResult {
|
|
MNN::Express::VARP QW;
|
|
MNN::Express::VARP SZ;
|
|
QuantizationConfig config; // 配置信息
|
|
size_t elementSize = 0;
|
|
};
|
|
|
|
private:
|
|
QuantizationConfig mConfig;
|
|
|
|
public:
|
|
explicit HQQQuantizer(const QuantizationConfig& config);
|
|
|
|
/**
|
|
* 量化权重矩阵
|
|
* @param weights 输入权重数据
|
|
* @param shape 权重形状 [height, width]
|
|
* @return 量化结果
|
|
*/
|
|
QuantizationResult quantize(const std::vector<float>& weights);
|
|
|
|
/**
|
|
* 反量化权重矩阵
|
|
* @param result 量化结果
|
|
* @return 反量化后的权重
|
|
*/
|
|
MNN::Express::VARP dequantize(const QuantizationResult& result);
|
|
|
|
private:
|
|
void optimize(MNN::Express::VARP& scale, MNN::Express::VARP& zero, MNN::Express::VARP WF);
|
|
};
|
|
|
|
} // namespace Quantization
|
|
} // namespace AliNN
|