|
#pragma once |
|
#pragma GCC diagnostic ignored "-Wstrict-aliasing" |
|
|
|
#include <cstddef> |
|
#include <cstdint> |
|
#include <vector> |
|
|
|
namespace fastertransformer { |
|
|
|
enum class QuantType { INT8_WEIGHT_ONLY, PACKED_INT4_WEIGHT_ONLY }; |
|
|
|
int get_bits_in_quant_type(QuantType quant_type); |
|
|
|
void preprocess_weights(int8_t *preprocessed_quantized_weight, |
|
const int8_t *row_major_quantized_weight, size_t rows, |
|
size_t cols, bool is_int4, int arch); |
|
|
|
template<typename ComputeType, typename WeightType> |
|
void symmetric_quantize(int8_t* processed_quantized_weight, |
|
ComputeType* scale_ptr, |
|
const WeightType* input_weight_ptr, |
|
const std::vector<size_t>& shape, |
|
QuantType quant_type); |
|
|
|
|
|
template<typename ComputeType, typename WeightType> |
|
void symmetric_quantize(int8_t* processed_quantized_weight, |
|
int8_t* unprocessed_quantized_weight, |
|
ComputeType* scale_ptr, |
|
const WeightType* input_weight_ptr, |
|
const std::vector<size_t>& shape, |
|
QuantType quant_type); |
|
} |
|
|
