quantization.h

Go to the documentation of this file.

#pragma once
 
#include <cstddef>
#include <cstdint>
#include <limits>
#include <vector>
 
#include "ggml_types.h"
#include "gguf_parser.h"  // Include for GGML_QK_K
 
// Define RESTRICT macro based on compiler
#ifdef _MSC_VER
#define RESTRICT
#else
#define RESTRICT __restrict__
#endif
 
// Forward declarations
struct block_q2_K;
struct block_q3_K;
struct block_q4_K;
struct block_q6_K;
 
float fp16_to_fp32(uint16_t h, bool is_gguf_scale_field = false);
 
uint16_t fp32_to_fp16(float f);
 
#pragma pack(push, 1)
 
struct block_q4_K {
    uint16_t d;          
    uint16_t dmin;       
    uint8_t scales[12];  
    uint8_t qs[GGML_QK_K / 2]; 
};
static_assert(sizeof(block_q4_K) == 2 + 2 + 12 + 128, "Size mismatch for standard block_q4_K");
 
struct block_q6_K {
    uint8_t ql[GGML_QK_K / 2];     
    uint8_t qh[GGML_QK_K / 4];     
    int8_t scales[GGML_QK_K / 16]; 
    uint16_t d;                    
};
static_assert(sizeof(block_q6_K) == 128 + 64 + 16 + 2, "Size mismatch for block_q6_K");
 
struct block_q2_K {
    uint16_t d;                    
    uint16_t dmin;                 
    uint8_t scales[GGML_QK_K / 16]; 
    uint8_t qs[GGML_QK_K / 4];     
};
static_assert(sizeof(block_q2_K) == 2 + 2 + 16 + 64, "Size mismatch for block_q2_K");
 
struct block_q3_K {
    uint8_t hmask[GGML_QK_K / 8]; 
    uint8_t qs[GGML_QK_K / 4];    
    uint8_t scales[12];           
    uint16_t d;                   
    uint16_t dmin;                
};
static_assert(sizeof(block_q3_K) == 32 + 64 + 12 + 2 + 2, "Size mismatch for block_q3_K");
 
struct block_q8_K {
    uint16_t d;                      
    int8_t qs[GGML_QK_K];           
    int16_t bsums[GGML_QK_K / 16];  
};
 
struct block_q8_0 {
    uint16_t d;              
    int8_t qs[GGML_QK8_0];  
};
static_assert(sizeof(block_q8_0) == sizeof(uint16_t) + GGML_QK8_0, "Size mismatch for block_q8_0");
 
#pragma pack(pop)
 
const char* ggml_type_name(GGMLType type);
 
size_t ggml_type_size(GGMLType type);
 
size_t ggml_type_block_size(GGMLType type);
 
void dequantize_q2_k(const void* q_data, float* f_data,
                     int num_weights_in_block,
                     bool log_details_for_this_block = false);
 
void dequantize_q4_k_m(const block_q4_K* qblock, float* RESTRICT output_f32,
                       int num_elements, bool log_this_block = false);
 
void dequantize_q6_k(const block_q6_K* qblock, float* RESTRICT output_f32,
                     int num_elements, bool log_this_block = false);
 
void dequantize_vector_q6k_to_f32(const std::vector<block_q6_K>& q_weights,
                                  std::vector<float>& f32_weights,
                                  size_t total_num_elements,
                                  int log_first_n_blocks = 0);
 
void dequantize_q3_k(const void* q_data, float* f_data,
                     int num_weights_in_block);
 
void handle_i8_tensor(const void* i8_data, float* f_data, size_t num_elements);
 
void quantize_q4_k_m(const float* f_data, void* q_data, int num_elements);
 
void quantize_q6_k(const float* f_data, void* q_data, int num_elements);
 
std::vector<block_q8_K> quantize_fp32_to_q8_K(const std::vector<float>& f_data);
 
float vec_dot_q6_k_q8_k_cpu(int n, const std::vector<block_q6_K>& x,
                            const std::vector<block_q8_K>& y,
                            bool log_this_call);
 
void matvec_q6k_q8k_cpu(const std::vector<block_q6_K>& mat_q6k,
                        const std::vector<block_q8_K>& vec_q8k,
                        std::vector<float>& out_f32, int rows, int cols,
                        bool log_calls);
 
float vec_dot_q4_k_q8_k_cpu(int n, const std::vector<block_q4_K>& x_vec,
                            const std::vector<block_q8_K>& y_vec,
                            bool log_this_call);
 
void matvec_q4k_q8k_cpu(const std::vector<block_q4_K>& mat_q4k,
                        const std::vector<block_q8_K>& vec_q8k,
                        std::vector<float>& out_f32, int rows, int cols,
                        bool log_calls);
 
void dequantize_q8_0_block(const block_q8_0* qblock, float* output);
 
void dequantize_vector_q4k_to_f32(const std::vector<block_q4_K>& q_weights,
                                  std::vector<float>& f32_weights,
                                  size_t total_num_elements,
                                  int log_first_n_blocks = 0);
 
void dequantize_vector_q8_0_to_f32(const std::vector<block_q8_0>& q_weights,
                                   std::vector<float>& f32_weights,
                                   size_t total_num_elements,
                                   int log_first_n_blocks = 0);