mirror of
https://github.com/jart/cosmopolitan.git
synced 2025-01-31 11:37:35 +00:00
5a4cf9560f
This change makes quantized models (e.g. q4_0) go 10% faster on Macs however doesn't offer much improvement for Intel PC hardware. This change syncs llama.cpp 699b1ad7fe6f7b9e41d3cb41e61a8cc3ea5fc6b5 which recently made a breaking change to nearly all its file formats without any migration. Since that'll break hundreds upon hundreds of models on websites like HuggingFace llama.com will support both file formats because llama.com will never ever break the GGJT file format
23 lines
949 B
C
23 lines
949 B
C
#ifndef COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q4_2_H_
|
|
#define COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q4_2_H_
|
|
#include "third_party/ggml/fp16.h"
|
|
#if !(__ASSEMBLER__ + __LINKER__ + 0)
|
|
COSMOPOLITAN_C_START_
|
|
|
|
#define V1_QK4_2 16
|
|
typedef struct {
|
|
ggml_fp16_t d; // delta
|
|
uint8_t qs[V1_QK4_2 / 2]; // nibbles / quants
|
|
} block_v1_q4_2;
|
|
|
|
void dequantize_row_v1_q4_2(const void* restrict, float* restrict, int);
|
|
void quantize_row_v1_q4_2(const float* restrict, void* restrict, int);
|
|
size_t ggml_quantize_v1_q4_2(const float*, void*, int, int, int64_t*);
|
|
void ggml_vec_dot_v1_q4_2_q8_0(const int, float* restrict, const void* restrict,
|
|
const void* restrict);
|
|
void quantize_row_v1_q4_2_reference(const float* restrict,
|
|
block_v1_q4_2* restrict, int);
|
|
|
|
COSMOPOLITAN_C_END_
|
|
#endif /* !(__ASSEMBLER__ + __LINKER__ + 0) */
|
|
#endif /* COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q4_2_H_ */
|