Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 22:08:46 +01:00
Code Issues Packages Projects Releases Wiki Activity

ggml : minor naming changes (#8433)

Browse Source 
* ggml : minor naming changes

ggml-ci

* ggml : use PRId64 [no ci]

* ggml : revert FA K/Q names
This commit is contained in:
Georgi Gerganov 2024-07-12 10:46:02 +03:00 committed by GitHub
parent b549a1bbef
commit 370b1f7e7a
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
10 changed files with 192 additions and 189 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
examples/quantize-stats/quantize-stats.cpp
View File

@ -154,7 +154,7 @@ static void test_roundtrip_on_chunk(
} }
if (use_reference) { if (use_reference) {
qfns.from_float_reference(input_scratch, quantized_scratch, chunk_size); qfns.from_float_ref(input_scratch, quantized_scratch, chunk_size);
} else { } else {
qfns.from_float(input_scratch, quantized_scratch, chunk_size); qfns.from_float(input_scratch, quantized_scratch, chunk_size);
} }

50
ggml/include/ggml.h
View File

@ -714,9 +714,9 @@ extern "C" {
GGML_API GGML_CALL size_t ggml_nbytes (const struct ggml_tensor * tensor); GGML_API GGML_CALL size_t ggml_nbytes (const struct ggml_tensor * tensor);
GGML_API size_t ggml_nbytes_pad (const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN GGML_API size_t ggml_nbytes_pad (const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN
GGML_API GGML_CALL int ggml_blck_size(enum ggml_type type); GGML_API GGML_CALL int64_t ggml_blck_size(enum ggml_type type);
GGML_API GGML_CALL size_t ggml_type_size(enum ggml_type type); // size in bytes for all elements in a block GGML_API GGML_CALL size_t ggml_type_size(enum ggml_type type); // size in bytes for all elements in a block
GGML_API GGML_CALL size_t ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row GGML_API GGML_CALL size_t ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row
GGML_DEPRECATED( GGML_DEPRECATED(
GGML_API double ggml_type_sizef(enum ggml_type type), // ggml_type_size()/ggml_blck_size() as float GGML_API double ggml_type_sizef(enum ggml_type type), // ggml_type_size()/ggml_blck_size() as float
@ -2410,31 +2410,31 @@ extern "C" {
#endif #endif
typedef void (*ggml_to_float_t) (const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); typedef void (*ggml_to_float_t) (const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
typedef void (*ggml_vec_dot_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx, typedef void (*ggml_from_float_to_mat_t)
const void * GGML_RESTRICT y, size_t by, int nrc); (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, int64_t k, int64_t bs);
typedef void (*ggml_from_float_to_mat_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, typedef void (*ggml_vec_dot_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
int64_t k, int64_t bx); const void * GGML_RESTRICT y, size_t by, int nrc);
typedef void (*ggml_gemv_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, typedef void (*ggml_gemv_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
const void * GGML_RESTRICT y, int nr, int nc); const void * GGML_RESTRICT y, int nr, int nc);
typedef void (*ggml_gemm_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, typedef void (*ggml_gemm_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
const void * GGML_RESTRICT y, int nr, int nc); const void * GGML_RESTRICT y, int nr, int nc);
typedef struct { typedef struct {
const char * type_name; const char * type_name;
int blck_size; int64_t blck_size;
size_t type_size; int64_t blck_size_interleave; // interleave elements in blocks
bool is_quantized; size_t type_size;
ggml_to_float_t to_float; bool is_quantized;
ggml_from_float_t from_float; ggml_to_float_t to_float;
ggml_from_float_t from_float_reference; ggml_from_float_t from_float;
ggml_vec_dot_t vec_dot; ggml_from_float_t from_float_ref;
enum ggml_type vec_dot_type;
int64_t nrows; // number of rows to process simultaneously;
int64_t ncols; // number of columns to process simultaneously;
int64_t interleave_blcksize; // interleave elements in blocks of interleave_blcksize;
ggml_from_float_to_mat_t from_float_to_mat; ggml_from_float_to_mat_t from_float_to_mat;
ggml_gemv_t gemv; ggml_vec_dot_t vec_dot;
ggml_gemm_t gemm; enum ggml_type vec_dot_type;
int64_t nrows; // number of rows to process simultaneously
int64_t ncols; // number of columns to process simultaneously
ggml_gemv_t gemv;
ggml_gemm_t gemm;
} ggml_type_traits_t; } ggml_type_traits_t;
GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type); GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);

68
ggml/src/ggml-aarch64.c
View File

@ -20,19 +20,19 @@
// Functions to create the interleaved data layout formats // Functions to create the interleaved data layout formats
// interleave 4 block_q4_0s in blocks of interleave_blcksize // interleave 4 block_q4_0s in blocks of blck_size_interleave
// returns an interleaved block_q4_0x4 // returns an interleaved block_q4_0x4
// in the interleaved block_q4_0x4, place deltas for 4 block_q4_0 blocks // in the interleaved block_q4_0x4, place deltas for 4 block_q4_0 blocks
// first, then interleave quants from 4 block_q4_0s in blocks of interleave_blcksize // first, then interleave quants from 4 block_q4_0s in blocks of blck_size_interleave
// //
// - in : an array of block_q4_0 pointers // - in : an array of block_q4_0 pointers
// - interleave_blcksize : the block_q4_0 quants bytes are interleaved in blocks of // - blck_size_interleave : the block_q4_0 quants bytes are interleaved in blocks of
// interleave_blcksize bytes // blck_size_interleave bytes
// - xor_mask : the mask to convert the nibbles in block_q4_0 quants bytes // - xor_mask : the mask to convert the nibbles in block_q4_0 quants bytes
// from bias offset form to pure sign form (this saves subtract // from bias offset form to pure sign form (this saves subtract
// operations durin unpacking) // operations durin unpacking)
// //
static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int interleave_blcksize, unsigned int xor_mask) { static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
block_q4_0x4 out; block_q4_0x4 out;
for (int i = 0; i < 4; i++) { for (int i = 0; i < 4; i++) {
@ -40,9 +40,9 @@ static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int interleave_b
} }
for (int i = 0; i < QK4_0 * 2; i++) { for (int i = 0; i < QK4_0 * 2; i++) {
int src_offset = (i / (4 * interleave_blcksize)) * interleave_blcksize; int src_offset = (i / (4 * blck_size_interleave)) * blck_size_interleave;
int src_id = (i % (4 * interleave_blcksize)) / interleave_blcksize; int src_id = (i % (4 * blck_size_interleave)) / blck_size_interleave;
src_offset += (i % interleave_blcksize); src_offset += (i % blck_size_interleave);
out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask; out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
} }
@ -50,11 +50,11 @@ static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int interleave_b
return out; return out;
} }
// interleave 8 block_q4_0s in blocks of interleave_blcksize // interleave 8 block_q4_0s in blocks of blck_size_interleave
// returns an interleaved block_q4_0x8 // returns an interleaved block_q4_0x8
// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks // in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
// first, then interleave quants from 8 block_q4_0s in blocks of interleave_blcksize // first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int interleave_blcksize, unsigned int xor_mask) { static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
block_q4_0x8 out; block_q4_0x8 out;
for (int i = 0; i < 8; i++) { for (int i = 0; i < 8; i++) {
@ -62,9 +62,9 @@ static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int interleave_b
} }
for (int i = 0; i < QK4_0 * 4; i++) { for (int i = 0; i < QK4_0 * 4; i++) {
int src_offset = (i / (8 * interleave_blcksize)) * interleave_blcksize; int src_offset = (i / (8 * blck_size_interleave)) * blck_size_interleave;
int src_id = (i % (8 * interleave_blcksize)) / interleave_blcksize; int src_id = (i % (8 * blck_size_interleave)) / blck_size_interleave;
src_offset += (i % interleave_blcksize); src_offset += (i % blck_size_interleave);
out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask; out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
} }
@ -135,7 +135,7 @@ void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int64_t k)
} }
#else #else
// scalar // scalar
const int interleave_blcksize = 4; const int blck_size_interleave = 4;
float srcv[4][QK8_0]; float srcv[4][QK8_0];
float id[4]; float id[4];
@ -155,12 +155,12 @@ void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int64_t k)
} }
for (int j = 0; j < QK8_0 * 4; j++) { for (int j = 0; j < QK8_0 * 4; j++) {
int src_offset = (j / (4 * interleave_blcksize)) * interleave_blcksize; int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
int src_id = (j % (4 * interleave_blcksize)) / interleave_blcksize; int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
src_offset += (j % interleave_blcksize); src_offset += (j % blck_size_interleave);
float x0 = srcv[src_id][src_offset] * id[src_id]; float x0 = srcv[src_id][src_offset] * id[src_id];
y[i].qs[j] = roundf(x0);; y[i].qs[j] = roundf(x0);
} }
} }
#endif #endif
@ -253,7 +253,7 @@ void quantize_q8_0_4x8(const float * restrict x, void * restrict vy, int64_t k)
} }
#else #else
// scalar // scalar
const int interleave_blcksize = 8; const int blck_size_interleave = 8;
float srcv[4][QK8_0]; float srcv[4][QK8_0];
float id[4]; float id[4];
@ -273,26 +273,30 @@ void quantize_q8_0_4x8(const float * restrict x, void * restrict vy, int64_t k)
} }
for (int j = 0; j < QK8_0 * 4; j++) { for (int j = 0; j < QK8_0 * 4; j++) {
int src_offset = (j / (4 * interleave_blcksize)) * interleave_blcksize; int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
int src_id = (j % (4 * interleave_blcksize)) / interleave_blcksize; int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
src_offset += (j % interleave_blcksize); src_offset += (j % blck_size_interleave);
float x0 = srcv[src_id][src_offset] * id[src_id]; float x0 = srcv[src_id][src_offset] * id[src_id];
y[i].qs[j] = roundf(x0);; y[i].qs[j] = roundf(x0);
} }
} }
#endif #endif
} }
void quantize_mat_q8_0(const float * restrict x, void * restrict vy, int64_t nrow, int64_t n_per_row, int64_t interleave_blcksize) { void quantize_mat_q8_0(const float * restrict x, void * restrict vy, int64_t nrow, int64_t n_per_row, int64_t blck_size_interleave) {
assert(nrow == 4); assert(nrow == 4);
UNUSED(nrow); UNUSED(nrow);
if (interleave_blcksize == 4) quantize_q8_0_4x4(x, vy, n_per_row); if (blck_size_interleave == 4) {
else if (interleave_blcksize == 8) quantize_q8_0_4x8(x, vy, n_per_row); quantize_q8_0_4x4(x, vy, n_per_row);
else assert(false); } else if (blck_size_interleave == 8) {
quantize_q8_0_4x8(x, vy, n_per_row);
} else {
assert(false);
}
} }
static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, int nrows_interleaved, int interleave_blcksize) { static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, int nrows_interleaved, int blck_size_interleave) {
assert(n_per_row % QK4_0 == 0); assert(n_per_row % QK4_0 == 0);
const int nb = n_per_row / QK4_0; const int nb = n_per_row / QK4_0;
@ -311,15 +315,15 @@ static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict ds
for (int64_t x = 0; x < nb; x++) { for (int64_t x = 0; x < nb; x++) {
for (int i = 0; i < nrows_interleaved; i++ ) { for (int i = 0; i < nrows_interleaved; i++ ) {
quantize_row_q4_0_reference(src + b + i * n_per_row + x * QK4_0, (block_q4_0 *) dst_tmp + i, QK4_0); quantize_row_q4_0_ref(src + b + i * n_per_row + x * QK4_0, (block_q4_0 *) dst_tmp + i, QK4_0);
} }
if (nrows_interleaved == 8) { if (nrows_interleaved == 8) {
*(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, interleave_blcksize, 0x88); *(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave, 0x88);
out_ptr = (block_q4_0x8 *) out_ptr + 1; out_ptr = (block_q4_0x8 *) out_ptr + 1;
} }
else if (nrows_interleaved == 4) { else if (nrows_interleaved == 4) {
*(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, interleave_blcksize, 0x88); *(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave, 0x88);
out_ptr = (block_q4_0x4 *) out_ptr + 1; out_ptr = (block_q4_0x4 *) out_ptr + 1;
} }
} }

14
ggml/src/ggml-aarch64.h
View File

@ -16,7 +16,7 @@ extern "C" {
void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t interleave_blcksize); void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization") // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
@ -24,14 +24,14 @@ size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT d
size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
// GEMV // GEMV
void ggml_gemv_q4_0_4x4_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_4x8_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_8x8_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
// GEMM // GEMM
void ggml_gemm_q4_0_4x4_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x8_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_8x8_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
#ifdef __cplusplus #ifdef __cplusplus
} }

94
ggml/src/ggml-quants.c
View File

@ -658,7 +658,7 @@ static inline __m128i packNibbles( __m256i bytes ) {
#endif //__loongarch_asx #endif //__loongarch_asx
// reference implementation for deterministic creation of model files // reference implementation for deterministic creation of model files
void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int64_t k) { void quantize_row_q4_0_ref(const float * restrict x, block_q4_0 * restrict y, int64_t k) {
static const int qk = QK4_0; static const int qk = QK4_0;
assert(k % qk == 0); assert(k % qk == 0);
@ -696,11 +696,11 @@ void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict
} }
void quantize_row_q4_0(const float * restrict x, void * restrict y, int64_t k) { void quantize_row_q4_0(const float * restrict x, void * restrict y, int64_t k) {
quantize_row_q4_0_reference(x, y, k); quantize_row_q4_0_ref(x, y, k);
} }
void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int64_t k) { void quantize_row_q4_1_ref(const float * restrict x, block_q4_1 * restrict y, int64_t k) {
const int qk = QK4_1; const int qk = QK4_1;
assert(k % qk == 0); assert(k % qk == 0);
@ -738,10 +738,10 @@ void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict
} }
void quantize_row_q4_1(const float * restrict x, void * restrict y, int64_t k) { void quantize_row_q4_1(const float * restrict x, void * restrict y, int64_t k) {
quantize_row_q4_1_reference(x, y, k); quantize_row_q4_1_ref(x, y, k);
} }
void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int64_t k) { void quantize_row_q5_0_ref(const float * restrict x, block_q5_0 * restrict y, int64_t k) {
static const int qk = QK5_0; static const int qk = QK5_0;
assert(k % qk == 0); assert(k % qk == 0);
@ -786,10 +786,10 @@ void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict
} }
void quantize_row_q5_0(const float * restrict x, void * restrict y, int64_t k) { void quantize_row_q5_0(const float * restrict x, void * restrict y, int64_t k) {
quantize_row_q5_0_reference(x, y, k); quantize_row_q5_0_ref(x, y, k);
} }
void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int64_t k) { void quantize_row_q5_1_ref(const float * restrict x, block_q5_1 * restrict y, int64_t k) {
const int qk = QK5_1; const int qk = QK5_1;
assert(k % qk == 0); assert(k % qk == 0);
@ -834,11 +834,11 @@ void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict
} }
void quantize_row_q5_1(const float * restrict x, void * restrict y, int64_t k) { void quantize_row_q5_1(const float * restrict x, void * restrict y, int64_t k) {
quantize_row_q5_1_reference(x, y, k); quantize_row_q5_1_ref(x, y, k);
} }
// reference implementation for deterministic creation of model files // reference implementation for deterministic creation of model files
void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int64_t k) { void quantize_row_q8_0_ref(const float * restrict x, block_q8_0 * restrict y, int64_t k) {
assert(k % QK8_0 == 0); assert(k % QK8_0 == 0);
const int nb = k / QK8_0; const int nb = k / QK8_0;
@ -1144,12 +1144,12 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k)
#else #else
GGML_UNUSED(nb); GGML_UNUSED(nb);
// scalar // scalar
quantize_row_q8_0_reference(x, y, k); quantize_row_q8_0_ref(x, y, k);
#endif #endif
} }
// reference implementation for deterministic creation of model files // reference implementation for deterministic creation of model files
void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int64_t k) { void quantize_row_q8_1_ref(const float * restrict x, block_q8_1 * restrict y, int64_t k) {
assert(QK8_1 == 32); assert(QK8_1 == 32);
assert(k % QK8_1 == 0); assert(k % QK8_1 == 0);
const int nb = k / QK8_1; const int nb = k / QK8_1;
@ -1508,7 +1508,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k)
#else #else
GGML_UNUSED(nb); GGML_UNUSED(nb);
// scalar // scalar
quantize_row_q8_1_reference(x, y, k); quantize_row_q8_1_ref(x, y, k);
#endif #endif
} }
@ -1899,7 +1899,7 @@ static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t *
//========================- 2-bit (de)-quantization //========================- 2-bit (de)-quantization
void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int64_t k) { void quantize_row_q2_K_ref(const float * restrict x, block_q2_K * restrict y, int64_t k) {
assert(k % QK_K == 0); assert(k % QK_K == 0);
const int nb = k / QK_K; const int nb = k / QK_K;
@ -2002,7 +2002,7 @@ void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int6
} }
void quantize_row_q2_K(const float * restrict x, void * restrict vy, int64_t k) { void quantize_row_q2_K(const float * restrict x, void * restrict vy, int64_t k) {
quantize_row_q2_K_reference(x, vy, k); quantize_row_q2_K_ref(x, vy, k);
} }
static float make_qkx3_quants(int n, int nmax, const float * restrict x, const float * restrict weights, static float make_qkx3_quants(int n, int nmax, const float * restrict x, const float * restrict weights,
@ -2226,7 +2226,7 @@ static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restri
size_t quantize_q2_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { size_t quantize_q2_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
size_t row_size = ggml_row_size(GGML_TYPE_Q2_K, n_per_row); size_t row_size = ggml_row_size(GGML_TYPE_Q2_K, n_per_row);
if (!quant_weights) { if (!quant_weights) {
quantize_row_q2_K_reference(src, dst, (int64_t)nrow*n_per_row); quantize_row_q2_K_ref(src, dst, (int64_t)nrow*n_per_row);
} }
else { else {
char * qrow = (char *)dst; char * qrow = (char *)dst;
@ -2241,7 +2241,7 @@ size_t quantize_q2_K(const float * restrict src, void * restrict dst, int64_t nr
//========================= 3-bit (de)-quantization //========================= 3-bit (de)-quantization
void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int64_t k) { void quantize_row_q3_K_ref(const float * restrict x, block_q3_K * restrict y, int64_t k) {
assert(k % QK_K == 0); assert(k % QK_K == 0);
const int nb = k / QK_K; const int nb = k / QK_K;
@ -2368,7 +2368,7 @@ void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int6
} }
void quantize_row_q3_K(const float * restrict x, void * restrict vy, int64_t k) { void quantize_row_q3_K(const float * restrict x, void * restrict vy, int64_t k) {
quantize_row_q3_K_reference(x, vy, k); quantize_row_q3_K_ref(x, vy, k);
} }
static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restrict y, int64_t n_per_row, const float * restrict quant_weights) { static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restrict y, int64_t n_per_row, const float * restrict quant_weights) {
@ -2458,7 +2458,7 @@ static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restri
size_t quantize_q3_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { size_t quantize_q3_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
size_t row_size = ggml_row_size(GGML_TYPE_Q3_K, n_per_row); size_t row_size = ggml_row_size(GGML_TYPE_Q3_K, n_per_row);
if (!quant_weights) { if (!quant_weights) {
quantize_row_q3_K_reference(src, dst, (int64_t)nrow*n_per_row); quantize_row_q3_K_ref(src, dst, (int64_t)nrow*n_per_row);
} }
else { else {
char * qrow = (char *)dst; char * qrow = (char *)dst;
@ -2473,7 +2473,7 @@ size_t quantize_q3_K(const float * restrict src, void * restrict dst, int64_t nr
// ====================== 4-bit (de)-quantization // ====================== 4-bit (de)-quantization
void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int64_t k) { void quantize_row_q4_K_ref(const float * restrict x, block_q4_K * restrict y, int64_t k) {
assert(k % QK_K == 0); assert(k % QK_K == 0);
const int nb = k / QK_K; const int nb = k / QK_K;
@ -2572,7 +2572,7 @@ void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int6
void quantize_row_q4_K(const float * restrict x, void * restrict vy, int64_t k) { void quantize_row_q4_K(const float * restrict x, void * restrict vy, int64_t k) {
assert(k % QK_K == 0); assert(k % QK_K == 0);
block_q4_K * restrict y = vy; block_q4_K * restrict y = vy;
quantize_row_q4_K_reference(x, y, k); quantize_row_q4_K_ref(x, y, k);
} }
static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restrict y, int64_t n_per_row, const float * quant_weights) { static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restrict y, int64_t n_per_row, const float * quant_weights) {
@ -2651,7 +2651,7 @@ static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restri
size_t quantize_q4_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { size_t quantize_q4_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
size_t row_size = ggml_row_size(GGML_TYPE_Q4_K, n_per_row); size_t row_size = ggml_row_size(GGML_TYPE_Q4_K, n_per_row);
if (!quant_weights) { if (!quant_weights) {
quantize_row_q4_K_reference(src, dst, (int64_t)nrow*n_per_row); quantize_row_q4_K_ref(src, dst, (int64_t)nrow*n_per_row);
} }
else { else {
char * qrow = (char *)dst; char * qrow = (char *)dst;
@ -2666,7 +2666,7 @@ size_t quantize_q4_K(const float * restrict src, void * restrict dst, int64_t nr
// ====================== 5-bit (de)-quantization // ====================== 5-bit (de)-quantization
void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int64_t k) { void quantize_row_q5_K_ref(const float * restrict x, block_q5_K * restrict y, int64_t k) {
assert(k % QK_K == 0); assert(k % QK_K == 0);
const int64_t nb = k / QK_K; const int64_t nb = k / QK_K;
@ -2783,7 +2783,7 @@ void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int6
void quantize_row_q5_K(const float * restrict x, void * restrict vy, int64_t k) { void quantize_row_q5_K(const float * restrict x, void * restrict vy, int64_t k) {
assert(k % QK_K == 0); assert(k % QK_K == 0);
block_q5_K * restrict y = vy; block_q5_K * restrict y = vy;
quantize_row_q5_K_reference(x, y, k); quantize_row_q5_K_ref(x, y, k);
} }
static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restrict y, int64_t n_per_row, const float * quant_weights) { static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restrict y, int64_t n_per_row, const float * quant_weights) {
@ -2882,7 +2882,7 @@ static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restri
size_t quantize_q5_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { size_t quantize_q5_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
size_t row_size = ggml_row_size(GGML_TYPE_Q5_K, n_per_row); size_t row_size = ggml_row_size(GGML_TYPE_Q5_K, n_per_row);
if (!quant_weights) { if (!quant_weights) {
quantize_row_q5_K_reference(src, dst, (int64_t)nrow*n_per_row); quantize_row_q5_K_ref(src, dst, (int64_t)nrow*n_per_row);
} }
else { else {
char * qrow = (char *)dst; char * qrow = (char *)dst;
@ -2897,7 +2897,7 @@ size_t quantize_q5_K(const float * restrict src, void * restrict dst, int64_t nr
// ====================== 6-bit (de)-quantization // ====================== 6-bit (de)-quantization
void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int64_t k) { void quantize_row_q6_K_ref(const float * restrict x, block_q6_K * restrict y, int64_t k) {
assert(k % QK_K == 0); assert(k % QK_K == 0);
const int64_t nb = k / QK_K; const int64_t nb = k / QK_K;
@ -3001,7 +3001,7 @@ void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int6
void quantize_row_q6_K(const float * restrict x, void * restrict vy, int64_t k) { void quantize_row_q6_K(const float * restrict x, void * restrict vy, int64_t k) {
assert(k % QK_K == 0); assert(k % QK_K == 0);
block_q6_K * restrict y = vy; block_q6_K * restrict y = vy;
quantize_row_q6_K_reference(x, y, k); quantize_row_q6_K_ref(x, y, k);
} }
static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restrict y, int64_t n_per_row, const float * quant_weights) { static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restrict y, int64_t n_per_row, const float * quant_weights) {
@ -3091,7 +3091,7 @@ static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restri
size_t quantize_q6_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { size_t quantize_q6_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
size_t row_size = ggml_row_size(GGML_TYPE_Q6_K, n_per_row); size_t row_size = ggml_row_size(GGML_TYPE_Q6_K, n_per_row);
if (!quant_weights) { if (!quant_weights) {
quantize_row_q6_K_reference(src, dst, (int64_t)nrow*n_per_row); quantize_row_q6_K_ref(src, dst, (int64_t)nrow*n_per_row);
} }
else { else {
char * qrow = (char *)dst; char * qrow = (char *)dst;
@ -3108,7 +3108,7 @@ static void quantize_row_q4_0_impl(const float * restrict x, block_q4_0 * restri
static_assert(QK4_0 == 32, "QK4_0 must be 32"); static_assert(QK4_0 == 32, "QK4_0 must be 32");
if (!quant_weights) { if (!quant_weights) {
quantize_row_q4_0_reference(x, y, n_per_row); quantize_row_q4_0_ref(x, y, n_per_row);
return; return;
} }
@ -3134,7 +3134,7 @@ static void quantize_row_q4_0_impl(const float * restrict x, block_q4_0 * restri
size_t quantize_q4_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { size_t quantize_q4_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
if (!quant_weights) { if (!quant_weights) {
quantize_row_q4_0_reference(src, dst, (int64_t)nrow*n_per_row); quantize_row_q4_0_ref(src, dst, (int64_t)nrow*n_per_row);
return nrow * ggml_row_size(GGML_TYPE_Q4_0, n_per_row); return nrow * ggml_row_size(GGML_TYPE_Q4_0, n_per_row);
} }
size_t row_size = ggml_row_size(GGML_TYPE_Q4_0, n_per_row); size_t row_size = ggml_row_size(GGML_TYPE_Q4_0, n_per_row);
@ -3151,7 +3151,7 @@ static void quantize_row_q4_1_impl(const float * restrict x, block_q4_1 * restri
static_assert(QK4_1 == 32, "QK4_1 must be 32"); static_assert(QK4_1 == 32, "QK4_1 must be 32");
if (!quant_weights) { if (!quant_weights) {
quantize_row_q4_1_reference(x, y, n_per_row); quantize_row_q4_1_ref(x, y, n_per_row);
return; return;
} }
@ -3179,7 +3179,7 @@ static void quantize_row_q4_1_impl(const float * restrict x, block_q4_1 * restri
size_t quantize_q4_1(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { size_t quantize_q4_1(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
if (!quant_weights) { if (!quant_weights) {
quantize_row_q4_1_reference(src, dst, (int64_t)nrow*n_per_row); quantize_row_q4_1_ref(src, dst, (int64_t)nrow*n_per_row);
return nrow * ggml_row_size(GGML_TYPE_Q4_1, n_per_row); return nrow * ggml_row_size(GGML_TYPE_Q4_1, n_per_row);
} }
size_t row_size = ggml_row_size(GGML_TYPE_Q4_1, n_per_row); size_t row_size = ggml_row_size(GGML_TYPE_Q4_1, n_per_row);
@ -3196,7 +3196,7 @@ static void quantize_row_q5_0_impl(const float * restrict x, block_q5_0 * restri
static_assert(QK5_0 == 32, "QK5_0 must be 32"); static_assert(QK5_0 == 32, "QK5_0 must be 32");
if (!quant_weights) { if (!quant_weights) {
quantize_row_q5_0_reference(x, y, n_per_row); quantize_row_q5_0_ref(x, y, n_per_row);
return; return;
} }
@ -3233,7 +3233,7 @@ static void quantize_row_q5_0_impl(const float * restrict x, block_q5_0 * restri
size_t quantize_q5_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { size_t quantize_q5_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
if (!quant_weights) { if (!quant_weights) {
quantize_row_q5_0_reference(src, dst, (int64_t)nrow*n_per_row); quantize_row_q5_0_ref(src, dst, (int64_t)nrow*n_per_row);
return nrow * ggml_row_size(GGML_TYPE_Q5_0, n_per_row); return nrow * ggml_row_size(GGML_TYPE_Q5_0, n_per_row);
} }
size_t row_size = ggml_row_size(GGML_TYPE_Q5_0, n_per_row); size_t row_size = ggml_row_size(GGML_TYPE_Q5_0, n_per_row);
@ -3250,7 +3250,7 @@ static void quantize_row_q5_1_impl(const float * restrict x, block_q5_1 * restri
static_assert(QK5_1 == 32, "QK5_1 must be 32"); static_assert(QK5_1 == 32, "QK5_1 must be 32");
if (!quant_weights) { if (!quant_weights) {
quantize_row_q5_1_reference(x, y, n_per_row); quantize_row_q5_1_ref(x, y, n_per_row);
return; return;
} }
@ -3286,7 +3286,7 @@ static void quantize_row_q5_1_impl(const float * restrict x, block_q5_1 * restri
size_t quantize_q5_1(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { size_t quantize_q5_1(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
if (!quant_weights) { if (!quant_weights) {
quantize_row_q5_1_reference(src, dst, (int64_t)nrow*n_per_row); quantize_row_q5_1_ref(src, dst, (int64_t)nrow*n_per_row);
return nrow * ggml_row_size(GGML_TYPE_Q5_1, n_per_row); return nrow * ggml_row_size(GGML_TYPE_Q5_1, n_per_row);
} }
size_t row_size = ggml_row_size(GGML_TYPE_Q5_1, n_per_row); size_t row_size = ggml_row_size(GGML_TYPE_Q5_1, n_per_row);
@ -3302,7 +3302,7 @@ size_t quantize_q5_1(const float * restrict src, void * restrict dst, int64_t nr
size_t quantize_q8_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { size_t quantize_q8_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
(void)quant_weights; // not used (void)quant_weights; // not used
const size_t row_size = ggml_row_size(GGML_TYPE_Q8_0, n_per_row); const size_t row_size = ggml_row_size(GGML_TYPE_Q8_0, n_per_row);
quantize_row_q8_0_reference(src, dst, (int64_t)nrow*n_per_row); quantize_row_q8_0_ref(src, dst, (int64_t)nrow*n_per_row);
return nrow * row_size; return nrow * row_size;
} }
@ -3590,7 +3590,7 @@ void dequantize_row_iq4_xs(const block_iq4_xs * restrict x, float * restrict y,
//===================================== Q8_K ============================================== //===================================== Q8_K ==============================================
void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int64_t k) { void quantize_row_q8_K_ref(const float * restrict x, block_q8_K * restrict y, int64_t k) {
assert(k % QK_K == 0); assert(k % QK_K == 0);
const int64_t nb = k / QK_K; const int64_t nb = k / QK_K;
@ -3641,7 +3641,7 @@ void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int6
} }
void quantize_row_q8_K(const float * restrict x, void * restrict y, int64_t k) { void quantize_row_q8_K(const float * restrict x, void * restrict y, int64_t k) {
quantize_row_q8_K_reference(x, y, k); quantize_row_q8_K_ref(x, y, k);
} }
//===================================== Dot ptoducts ================================= //===================================== Dot ptoducts =================================
@ -13530,10 +13530,10 @@ size_t quantize_iq3_xxs(const float * restrict src, void * restrict dst, int64_t
void quantize_row_iq3_xxs(const float * restrict x, void * restrict vy, int64_t k) { void quantize_row_iq3_xxs(const float * restrict x, void * restrict vy, int64_t k) {
assert(k % QK_K == 0); assert(k % QK_K == 0);
block_iq3_xxs * restrict y = vy; block_iq3_xxs * restrict y = vy;
quantize_row_iq3_xxs_reference(x, y, k); quantize_row_iq3_xxs_ref(x, y, k);
} }
void quantize_row_iq3_xxs_reference(const float * restrict x, block_iq3_xxs * restrict y, int64_t k) { void quantize_row_iq3_xxs_ref(const float * restrict x, block_iq3_xxs * restrict y, int64_t k) {
assert(k % QK_K == 0); assert(k % QK_K == 0);
quantize_row_iq3_xxs_impl(256, x, y, k, NULL); quantize_row_iq3_xxs_impl(256, x, y, k, NULL);
} }
@ -13746,10 +13746,10 @@ size_t quantize_iq3_s(const float * restrict src, void * restrict dst, int64_t n
void quantize_row_iq3_s(const float * restrict x, void * restrict vy, int64_t k) { void quantize_row_iq3_s(const float * restrict x, void * restrict vy, int64_t k) {
assert(k % QK_K == 0); assert(k % QK_K == 0);
block_iq3_s * restrict y = vy; block_iq3_s * restrict y = vy;
quantize_row_iq3_s_reference(x, y, k); quantize_row_iq3_s_ref(x, y, k);
} }
void quantize_row_iq3_s_reference(const float * restrict x, block_iq3_s * restrict y, int64_t k) { void quantize_row_iq3_s_ref(const float * restrict x, block_iq3_s * restrict y, int64_t k) {
assert(k % QK_K == 0); assert(k % QK_K == 0);
quantize_iq3_s(x, y, 1, k, NULL); quantize_iq3_s(x, y, 1, k, NULL);
} }
@ -14487,7 +14487,7 @@ void quantize_row_iq4_nl(const float * restrict x, void * restrict vy, int64_t k
} }
} }
void quantize_row_iq4_nl_reference(const float * restrict x, block_iq4_nl * restrict y, int64_t k) { void quantize_row_iq4_nl_ref(const float * restrict x, block_iq4_nl * restrict y, int64_t k) {
assert(k % QK4_NL == 0); assert(k % QK4_NL == 0);
quantize_row_iq4_nl(x, y, k); quantize_row_iq4_nl(x, y, k);
} }
@ -14515,10 +14515,10 @@ size_t quantize_iq4_xs(const float * restrict src, void * restrict dst, int64_t
void quantize_row_iq4_xs(const float * restrict x, void * restrict vy, int64_t k) { void quantize_row_iq4_xs(const float * restrict x, void * restrict vy, int64_t k) {
assert(k % QK_K == 0); assert(k % QK_K == 0);
block_iq4_xs * restrict y = vy; block_iq4_xs * restrict y = vy;
quantize_row_iq4_xs_reference(x, y, k); quantize_row_iq4_xs_ref(x, y, k);
} }
void quantize_row_iq4_xs_reference(const float * restrict x, block_iq4_xs * restrict y, int64_t k) { void quantize_row_iq4_xs_ref(const float * restrict x, block_iq4_xs * restrict y, int64_t k) {
assert(k % QK_K == 0); assert(k % QK_K == 0);
quantize_iq4_xs(x, y, 1, k, NULL); quantize_iq4_xs(x, y, 1, k, NULL);
} }
@ -14705,7 +14705,7 @@ size_t quantize_iq2_s(const float * restrict src, void * restrict dst, int64_t n
return nrow * nblock * sizeof(block_iq2_s); return nrow * nblock * sizeof(block_iq2_s);
} }
void quantize_row_iq2_s_reference(const float * restrict x, block_iq2_s * restrict y, int64_t k) { void quantize_row_iq2_s_ref(const float * restrict x, block_iq2_s * restrict y, int64_t k) {
assert(k % QK_K == 0); assert(k % QK_K == 0);
quantize_iq2_s(x, y, 1, k, NULL); quantize_iq2_s(x, y, 1, k, NULL);
} }
@ -14713,7 +14713,7 @@ void quantize_row_iq2_s_reference(const float * restrict x, block_iq2_s * restri
void quantize_row_iq2_s(const float * restrict x, void * restrict vy, int64_t k) { void quantize_row_iq2_s(const float * restrict x, void * restrict vy, int64_t k) {
assert(k % QK_K == 0); assert(k % QK_K == 0);
block_iq2_s * restrict y = vy; block_iq2_s * restrict y = vy;
quantize_row_iq2_s_reference(x, y, k); quantize_row_iq2_s_ref(x, y, k);
} }
static bool validate_float(float f, size_t i) { static bool validate_float(float f, size_t i) {

34
ggml/src/ggml-quants.h
View File

@ -12,25 +12,25 @@ extern "C" {
#endif #endif
// Quantization // Quantization
void quantize_row_q4_0_reference(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k); void quantize_row_q4_0_ref(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_1_reference(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k); void quantize_row_q4_1_ref(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_0_reference(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k); void quantize_row_q5_0_ref(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_1_reference(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k); void quantize_row_q5_1_ref(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_0_reference(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k); void quantize_row_q8_0_ref(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_1_reference(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k); void quantize_row_q8_1_ref(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
void quantize_row_q2_K_reference(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k); void quantize_row_q2_K_ref(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
void quantize_row_q3_K_reference(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k); void quantize_row_q3_K_ref(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_K_reference(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k); void quantize_row_q4_K_ref(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_K_reference(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k); void quantize_row_q5_K_ref(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k);
void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k); void quantize_row_q6_K_ref(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k); void quantize_row_q8_K_ref(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k);
void quantize_row_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k); void quantize_row_iq3_xxs_ref(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl * GGML_RESTRICT y, int64_t k); void quantize_row_iq4_nl_ref (const float * GGML_RESTRICT x, block_iq4_nl * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_xs_reference (const float * GGML_RESTRICT x, block_iq4_xs * GGML_RESTRICT y, int64_t k); void quantize_row_iq4_xs_ref (const float * GGML_RESTRICT x, block_iq4_xs * GGML_RESTRICT y, int64_t k);
void quantize_row_iq3_s_reference (const float * GGML_RESTRICT x, block_iq3_s * GGML_RESTRICT y, int64_t k); void quantize_row_iq3_s_ref (const float * GGML_RESTRICT x, block_iq3_s * GGML_RESTRICT y, int64_t k);
void quantize_row_iq2_s_reference (const float * GGML_RESTRICT x, block_iq2_s * GGML_RESTRICT y, int64_t k); void quantize_row_iq2_s_ref (const float * GGML_RESTRICT x, block_iq2_s * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);

111
ggml/src/ggml.c
View File

@ -592,7 +592,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = false, .is_quantized = false,
.to_float = (ggml_to_float_t) ggml_fp16_to_fp32_row, .to_float = (ggml_to_float_t) ggml_fp16_to_fp32_row,
.from_float = (ggml_from_float_t) ggml_fp32_to_fp16_row, .from_float = (ggml_from_float_t) ggml_fp32_to_fp16_row,
.from_float_reference = (ggml_from_float_t) ggml_fp32_to_fp16_row, .from_float_ref = (ggml_from_float_t) ggml_fp32_to_fp16_row,
.vec_dot = (ggml_vec_dot_t) ggml_vec_dot_f16, .vec_dot = (ggml_vec_dot_t) ggml_vec_dot_f16,
.vec_dot_type = GGML_TYPE_F16, .vec_dot_type = GGML_TYPE_F16,
.nrows = 1, .nrows = 1,
@ -604,7 +604,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_q4_0, .to_float = (ggml_to_float_t) dequantize_row_q4_0,
.from_float = quantize_row_q4_0, .from_float = quantize_row_q4_0,
.from_float_reference = (ggml_from_float_t) quantize_row_q4_0_reference, .from_float_ref = (ggml_from_float_t) quantize_row_q4_0_ref,
.vec_dot = ggml_vec_dot_q4_0_q8_0, .vec_dot = ggml_vec_dot_q4_0_q8_0,
.vec_dot_type = GGML_TYPE_Q8_0, .vec_dot_type = GGML_TYPE_Q8_0,
#if defined (__ARM_FEATURE_MATMUL_INT8) #if defined (__ARM_FEATURE_MATMUL_INT8)
@ -620,7 +620,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_q4_1, .to_float = (ggml_to_float_t) dequantize_row_q4_1,
.from_float = quantize_row_q4_1, .from_float = quantize_row_q4_1,
.from_float_reference = (ggml_from_float_t) quantize_row_q4_1_reference, .from_float_ref = (ggml_from_float_t) quantize_row_q4_1_ref,
.vec_dot = ggml_vec_dot_q4_1_q8_1, .vec_dot = ggml_vec_dot_q4_1_q8_1,
.vec_dot_type = GGML_TYPE_Q8_1, .vec_dot_type = GGML_TYPE_Q8_1,
#if defined (__ARM_FEATURE_MATMUL_INT8) #if defined (__ARM_FEATURE_MATMUL_INT8)
@ -636,7 +636,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = false, .is_quantized = false,
.to_float = NULL, .to_float = NULL,
.from_float = NULL, .from_float = NULL,
.from_float_reference = NULL, .from_float_ref = NULL,
.vec_dot = NULL, .vec_dot = NULL,
.vec_dot_type = GGML_TYPE_COUNT, .vec_dot_type = GGML_TYPE_COUNT,
.nrows = 1, .nrows = 1,
@ -648,7 +648,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = false, .is_quantized = false,
.to_float = NULL, .to_float = NULL,
.from_float = NULL, .from_float = NULL,
.from_float_reference = NULL, .from_float_ref = NULL,
.vec_dot = NULL, .vec_dot = NULL,
.vec_dot_type = GGML_TYPE_COUNT, .vec_dot_type = GGML_TYPE_COUNT,
.nrows = 1, .nrows = 1,
@ -660,7 +660,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_q5_0, .to_float = (ggml_to_float_t) dequantize_row_q5_0,
.from_float = quantize_row_q5_0, .from_float = quantize_row_q5_0,
.from_float_reference = (ggml_from_float_t) quantize_row_q5_0_reference, .from_float_ref = (ggml_from_float_t) quantize_row_q5_0_ref,
.vec_dot = ggml_vec_dot_q5_0_q8_0, .vec_dot = ggml_vec_dot_q5_0_q8_0,
.vec_dot_type = GGML_TYPE_Q8_0, .vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1, .nrows = 1,
@ -672,7 +672,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_q5_1, .to_float = (ggml_to_float_t) dequantize_row_q5_1,
.from_float = quantize_row_q5_1, .from_float = quantize_row_q5_1,
.from_float_reference = (ggml_from_float_t) quantize_row_q5_1_reference, .from_float_ref = (ggml_from_float_t) quantize_row_q5_1_ref,
.vec_dot = ggml_vec_dot_q5_1_q8_1, .vec_dot = ggml_vec_dot_q5_1_q8_1,
.vec_dot_type = GGML_TYPE_Q8_1, .vec_dot_type = GGML_TYPE_Q8_1,
.nrows = 1, .nrows = 1,
@ -684,7 +684,8 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_q8_0, .to_float = (ggml_to_float_t) dequantize_row_q8_0,
.from_float = quantize_row_q8_0, .from_float = quantize_row_q8_0,
.from_float_reference = (ggml_from_float_t) quantize_row_q8_0_reference, .from_float_ref = (ggml_from_float_t) quantize_row_q8_0_ref,
.from_float_to_mat = quantize_mat_q8_0,
.vec_dot = ggml_vec_dot_q8_0_q8_0, .vec_dot = ggml_vec_dot_q8_0_q8_0,
.vec_dot_type = GGML_TYPE_Q8_0, .vec_dot_type = GGML_TYPE_Q8_0,
#if defined (__ARM_FEATURE_MATMUL_INT8) #if defined (__ARM_FEATURE_MATMUL_INT8)
@ -692,7 +693,6 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
#else #else
.nrows = 1, .nrows = 1,
#endif #endif
.from_float_to_mat = quantize_mat_q8_0,
}, },
[GGML_TYPE_Q8_1] = { [GGML_TYPE_Q8_1] = {
.type_name = "q8_1", .type_name = "q8_1",
@ -700,7 +700,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.type_size = sizeof(block_q8_1), .type_size = sizeof(block_q8_1),
.is_quantized = true, .is_quantized = true,
.from_float = quantize_row_q8_1, .from_float = quantize_row_q8_1,
.from_float_reference = (ggml_from_float_t) quantize_row_q8_1_reference, .from_float_ref = (ggml_from_float_t) quantize_row_q8_1_ref,
.vec_dot_type = GGML_TYPE_Q8_1, .vec_dot_type = GGML_TYPE_Q8_1,
.nrows = 1, .nrows = 1,
}, },
@ -711,7 +711,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_q2_K, .to_float = (ggml_to_float_t) dequantize_row_q2_K,
.from_float = quantize_row_q2_K, .from_float = quantize_row_q2_K,
.from_float_reference = (ggml_from_float_t) quantize_row_q2_K_reference, .from_float_ref = (ggml_from_float_t) quantize_row_q2_K_ref,
.vec_dot = ggml_vec_dot_q2_K_q8_K, .vec_dot = ggml_vec_dot_q2_K_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K, .vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1, .nrows = 1,
@ -723,7 +723,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_q3_K, .to_float = (ggml_to_float_t) dequantize_row_q3_K,
.from_float = quantize_row_q3_K, .from_float = quantize_row_q3_K,
.from_float_reference = (ggml_from_float_t) quantize_row_q3_K_reference, .from_float_ref = (ggml_from_float_t) quantize_row_q3_K_ref,
.vec_dot = ggml_vec_dot_q3_K_q8_K, .vec_dot = ggml_vec_dot_q3_K_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K, .vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1, .nrows = 1,
@ -735,7 +735,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_q4_K, .to_float = (ggml_to_float_t) dequantize_row_q4_K,
.from_float = quantize_row_q4_K, .from_float = quantize_row_q4_K,
.from_float_reference = (ggml_from_float_t) quantize_row_q4_K_reference, .from_float_ref = (ggml_from_float_t) quantize_row_q4_K_ref,
.vec_dot = ggml_vec_dot_q4_K_q8_K, .vec_dot = ggml_vec_dot_q4_K_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K, .vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1, .nrows = 1,
@ -747,7 +747,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_q5_K, .to_float = (ggml_to_float_t) dequantize_row_q5_K,
.from_float = quantize_row_q5_K, .from_float = quantize_row_q5_K,
.from_float_reference = (ggml_from_float_t) quantize_row_q5_K_reference, .from_float_ref = (ggml_from_float_t) quantize_row_q5_K_ref,
.vec_dot = ggml_vec_dot_q5_K_q8_K, .vec_dot = ggml_vec_dot_q5_K_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K, .vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1, .nrows = 1,
@ -759,7 +759,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_q6_K, .to_float = (ggml_to_float_t) dequantize_row_q6_K,
.from_float = quantize_row_q6_K, .from_float = quantize_row_q6_K,
.from_float_reference = (ggml_from_float_t) quantize_row_q6_K_reference, .from_float_ref = (ggml_from_float_t) quantize_row_q6_K_ref,
.vec_dot = ggml_vec_dot_q6_K_q8_K, .vec_dot = ggml_vec_dot_q6_K_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K, .vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1, .nrows = 1,
@ -771,7 +771,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_iq2_xxs, .to_float = (ggml_to_float_t) dequantize_row_iq2_xxs,
.from_float = NULL, .from_float = NULL,
.from_float_reference = NULL, .from_float_ref = NULL,
.vec_dot = ggml_vec_dot_iq2_xxs_q8_K, .vec_dot = ggml_vec_dot_iq2_xxs_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K, .vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1, .nrows = 1,
@ -783,7 +783,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_iq2_xs, .to_float = (ggml_to_float_t) dequantize_row_iq2_xs,
.from_float = NULL, .from_float = NULL,
.from_float_reference = NULL, .from_float_ref = NULL,
.vec_dot = ggml_vec_dot_iq2_xs_q8_K, .vec_dot = ggml_vec_dot_iq2_xs_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K, .vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1, .nrows = 1,
@ -795,7 +795,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_iq3_xxs, .to_float = (ggml_to_float_t) dequantize_row_iq3_xxs,
.from_float = quantize_row_iq3_xxs, .from_float = quantize_row_iq3_xxs,
.from_float_reference = (ggml_from_float_t)quantize_row_iq3_xxs_reference, .from_float_ref = (ggml_from_float_t)quantize_row_iq3_xxs_ref,
.vec_dot = ggml_vec_dot_iq3_xxs_q8_K, .vec_dot = ggml_vec_dot_iq3_xxs_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K, .vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1, .nrows = 1,
@ -807,7 +807,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_iq3_s, .to_float = (ggml_to_float_t) dequantize_row_iq3_s,
.from_float = quantize_row_iq3_s, .from_float = quantize_row_iq3_s,
.from_float_reference = (ggml_from_float_t)quantize_row_iq3_s_reference, .from_float_ref = (ggml_from_float_t)quantize_row_iq3_s_ref,
.vec_dot = ggml_vec_dot_iq3_s_q8_K, .vec_dot = ggml_vec_dot_iq3_s_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K, .vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1, .nrows = 1,
@ -819,7 +819,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_iq2_s, .to_float = (ggml_to_float_t) dequantize_row_iq2_s,
.from_float = quantize_row_iq2_s, .from_float = quantize_row_iq2_s,
.from_float_reference = (ggml_from_float_t)quantize_row_iq2_s_reference, .from_float_ref = (ggml_from_float_t)quantize_row_iq2_s_ref,
.vec_dot = ggml_vec_dot_iq2_s_q8_K, .vec_dot = ggml_vec_dot_iq2_s_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K, .vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1, .nrows = 1,
@ -831,7 +831,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_iq1_s, .to_float = (ggml_to_float_t) dequantize_row_iq1_s,
.from_float = NULL, .from_float = NULL,
.from_float_reference = NULL, .from_float_ref = NULL,
.vec_dot = ggml_vec_dot_iq1_s_q8_K, .vec_dot = ggml_vec_dot_iq1_s_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K, .vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1, .nrows = 1,
@ -843,7 +843,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_iq1_m, .to_float = (ggml_to_float_t) dequantize_row_iq1_m,
.from_float = NULL, .from_float = NULL,
.from_float_reference = NULL, .from_float_ref = NULL,
.vec_dot = ggml_vec_dot_iq1_m_q8_K, .vec_dot = ggml_vec_dot_iq1_m_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K, .vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1, .nrows = 1,
@ -855,7 +855,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_iq4_nl, .to_float = (ggml_to_float_t) dequantize_row_iq4_nl,
.from_float = quantize_row_iq4_nl, .from_float = quantize_row_iq4_nl,
.from_float_reference = (ggml_from_float_t)quantize_row_iq4_nl_reference, .from_float_ref = (ggml_from_float_t)quantize_row_iq4_nl_ref,
.vec_dot = ggml_vec_dot_iq4_nl_q8_0, .vec_dot = ggml_vec_dot_iq4_nl_q8_0,
.vec_dot_type = GGML_TYPE_Q8_0, .vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1, .nrows = 1,
@ -867,7 +867,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = true, .is_quantized = true,
.to_float = (ggml_to_float_t) dequantize_row_iq4_xs, .to_float = (ggml_to_float_t) dequantize_row_iq4_xs,
.from_float = quantize_row_iq4_xs, .from_float = quantize_row_iq4_xs,
.from_float_reference = (ggml_from_float_t)quantize_row_iq4_xs_reference, .from_float_ref = (ggml_from_float_t)quantize_row_iq4_xs_ref,
.vec_dot = ggml_vec_dot_iq4_xs_q8_K, .vec_dot = ggml_vec_dot_iq4_xs_q8_K,
.vec_dot_type = GGML_TYPE_Q8_K, .vec_dot_type = GGML_TYPE_Q8_K,
.nrows = 1, .nrows = 1,
@ -886,7 +886,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
.is_quantized = false, .is_quantized = false,
.to_float = (ggml_to_float_t) ggml_bf16_to_fp32_row, .to_float = (ggml_to_float_t) ggml_bf16_to_fp32_row,
.from_float = (ggml_from_float_t) ggml_fp32_to_bf16_row, .from_float = (ggml_from_float_t) ggml_fp32_to_bf16_row,
.from_float_reference = (ggml_from_float_t) ggml_fp32_to_bf16_row, .from_float_ref = (ggml_from_float_t) ggml_fp32_to_bf16_row,
.vec_dot = (ggml_vec_dot_t) ggml_vec_dot_bf16, .vec_dot = (ggml_vec_dot_t) ggml_vec_dot_bf16,
.vec_dot_type = GGML_TYPE_BF16, .vec_dot_type = GGML_TYPE_BF16,
.nrows = 1, .nrows = 1,
@ -894,48 +894,48 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
[GGML_TYPE_Q4_0_4_4] = { [GGML_TYPE_Q4_0_4_4] = {
.type_name = "q4_0_4x4", .type_name = "q4_0_4x4",
.blck_size = QK4_0, .blck_size = QK4_0,
.blck_size_interleave = 4,
.type_size = sizeof(block_q4_0), .type_size = sizeof(block_q4_0),
.is_quantized = true, .is_quantized = true,
.to_float = NULL, .to_float = NULL,
.from_float = NULL, .from_float = NULL,
.from_float_reference = NULL, .from_float_ref = NULL,
.vec_dot = NULL, .vec_dot = NULL,
.vec_dot_type = GGML_TYPE_Q8_0, .vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1, .nrows = 1,
.ncols = 4, .ncols = 4,
.interleave_blcksize = 4,
.gemv = ggml_gemv_q4_0_4x4_q8_0, .gemv = ggml_gemv_q4_0_4x4_q8_0,
.gemm = ggml_gemm_q4_0_4x4_q8_0, .gemm = ggml_gemm_q4_0_4x4_q8_0,
}, },
[GGML_TYPE_Q4_0_4_8] = { [GGML_TYPE_Q4_0_4_8] = {
.type_name = "q4_0_4x8", .type_name = "q4_0_4x8",
.blck_size = QK4_0, .blck_size = QK4_0,
.blck_size_interleave = 8,
.type_size = sizeof(block_q4_0), .type_size = sizeof(block_q4_0),
.is_quantized = true, .is_quantized = true,
.to_float = NULL, .to_float = NULL,
.from_float = NULL, .from_float = NULL,
.from_float_reference = NULL, .from_float_ref = NULL,
.vec_dot = NULL, .vec_dot = NULL,
.vec_dot_type = GGML_TYPE_Q8_0, .vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1, .nrows = 1,
.ncols = 4, .ncols = 4,
.interleave_blcksize = 8,
.gemv = ggml_gemv_q4_0_4x8_q8_0, .gemv = ggml_gemv_q4_0_4x8_q8_0,
.gemm = ggml_gemm_q4_0_4x8_q8_0, .gemm = ggml_gemm_q4_0_4x8_q8_0,
}, },
[GGML_TYPE_Q4_0_8_8] = { [GGML_TYPE_Q4_0_8_8] = {
.type_name = "q4_0_8x8", .type_name = "q4_0_8x8",
.blck_size = QK4_0, .blck_size = QK4_0,
.blck_size_interleave = 8,
.type_size = sizeof(block_q4_0), .type_size = sizeof(block_q4_0),
.is_quantized = true, .is_quantized = true,
.to_float = NULL, .to_float = NULL,
.from_float = NULL, .from_float = NULL,
.from_float_reference = NULL, .from_float_ref = NULL,
.vec_dot = NULL, .vec_dot = NULL,
.vec_dot_type = GGML_TYPE_Q8_0, .vec_dot_type = GGML_TYPE_Q8_0,
.nrows = 1, .nrows = 1,
.ncols = 8, .ncols = 8,
.interleave_blcksize = 8,
.gemv = ggml_gemv_q4_0_8x8_q8_0, .gemv = ggml_gemv_q4_0_8x8_q8_0,
.gemm = ggml_gemm_q4_0_8x8_q8_0, .gemm = ggml_gemm_q4_0_8x8_q8_0,
} }
@ -3115,7 +3115,7 @@ size_t ggml_nbytes_pad(const struct ggml_tensor * tensor) {
return GGML_PAD(ggml_nbytes(tensor), GGML_MEM_ALIGN); return GGML_PAD(ggml_nbytes(tensor), GGML_MEM_ALIGN);
} }
GGML_CALL int ggml_blck_size(enum ggml_type type) { GGML_CALL int64_t ggml_blck_size(enum ggml_type type) {
return type_traits[type].blck_size; return type_traits[type].blck_size;
} }
@ -12192,15 +12192,14 @@ static void ggml_compute_forward_mul_mat(
const enum ggml_type type = src0->type; const enum ggml_type type = src0->type;
enum ggml_type const vec_dot_type = type_traits[type].vec_dot_type; enum ggml_type const vec_dot_type = type_traits[type].vec_dot_type;
ggml_from_float_t const from_float_to_vec_dot = type_traits[vec_dot_type].from_float; ggml_from_float_t const from_float = type_traits[vec_dot_type].from_float;
int64_t const vec_dot_num_rows = type_traits[type].nrows; ggml_from_float_to_mat_t const from_float_to_mat = type_traits[vec_dot_type].from_float_to_mat;
int64_t const matmul_num_cols = type_traits[type].ncols; int64_t const vec_dot_num_rows = type_traits[type].nrows;
int64_t const interleave_blcksize = type_traits[type].interleave_blcksize; int64_t const matmul_num_cols = type_traits[type].ncols;
ggml_from_float_to_mat_t const from_float_to_mat int64_t const blck_size_interleave = type_traits[type].blck_size_interleave;
= type_traits[vec_dot_type].from_float_to_mat; ggml_gemv_t const gemv = type_traits[type].gemv;
ggml_gemv_t const gemv = type_traits[type].gemv; ggml_gemm_t const gemm = type_traits[type].gemm;
ggml_gemm_t const gemm = type_traits[type].gemm;
GGML_ASSERT(ne0 == ne01); GGML_ASSERT(ne0 == ne01);
GGML_ASSERT(ne1 == ne11); GGML_ASSERT(ne1 == ne11);
@ -12264,14 +12263,14 @@ UseGgmlGemm1:;
for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) { for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) {
from_float_to_mat((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), from_float_to_mat((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11),
(void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1), (void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
4, ne10, interleave_blcksize); 4, ne10, blck_size_interleave);
} }
i11_processed = ne11 - ne11 % 4; i11_processed = ne11 - ne11 % 4;
} }
for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) { for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) {
from_float_to_vec_dot((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11),
(void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1), (void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
ne10); ne10);
} }
} }
} }
@ -12355,7 +12354,7 @@ UseGgmlGemm2:;
int64_t src0_start = (ith * ne01) / nth; int64_t src0_start = (ith * ne01) / nth;
int64_t src0_end = ((ith + 1) * ne01) / nth; int64_t src0_end = ((ith + 1) * ne01) / nth;
src0_start = (src0_start % matmul_num_cols) ? src0_start + matmul_num_cols - (src0_start % matmul_num_cols): src0_start; src0_start = (src0_start % matmul_num_cols) ? src0_start + matmul_num_cols - (src0_start % matmul_num_cols): src0_start;
src0_end = (src0_end % matmul_num_cols) ? src0_end + matmul_num_cols - (src0_end % matmul_num_cols): src0_end; src0_end = (src0_end % matmul_num_cols) ? src0_end + matmul_num_cols - (src0_end % matmul_num_cols): src0_end;
if (src0_start >= src0_end) return; if (src0_start >= src0_end) return;
// If there are more than three rows in src1, use gemm; otherwise, use gemv. // If there are more than three rows in src1, use gemm; otherwise, use gemv.
@ -12413,11 +12412,11 @@ static void ggml_compute_forward_mul_mat_id(
const bool src1_cont = ggml_is_contiguous(src1); const bool src1_cont = ggml_is_contiguous(src1);
ggml_vec_dot_t const vec_dot = type_traits[type].vec_dot; ggml_vec_dot_t const vec_dot = type_traits[type].vec_dot;
enum ggml_type const vec_dot_type = type_traits[type].vec_dot_type; enum ggml_type const vec_dot_type = type_traits[type].vec_dot_type;
ggml_from_float_t const from_float_to_vec_dot = type_traits[vec_dot_type].from_float; ggml_from_float_t const from_float = type_traits[vec_dot_type].from_float;
int64_t const matmul_num_cols = type_traits[type].ncols; int64_t const matmul_num_cols = type_traits[type].ncols;
ggml_gemv_t const gemv = type_traits[type].gemv; ggml_gemv_t const gemv = type_traits[type].gemv;
// we don't support permuted src0 or src1 // we don't support permuted src0 or src1
GGML_ASSERT(nb00 == ggml_type_size(type)); GGML_ASSERT(nb00 == ggml_type_size(type));
@ -12458,9 +12457,9 @@ static void ggml_compute_forward_mul_mat_id(
for (int64_t i13 = 0; i13 < ne13; ++i13) { for (int64_t i13 = 0; i13 < ne13; ++i13) {
for (int64_t i12 = 0; i12 < ne12; ++i12) { for (int64_t i12 = 0; i12 < ne12; ++i12) {
for (int64_t i11 = ith; i11 < ne11; i11 += nth) { for (int64_t i11 = ith; i11 < ne11; i11 += nth) {
from_float_to_vec_dot((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11),
(void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1), (void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
ne10); ne10);
} }
} }
} }
@ -21063,8 +21062,8 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
(int64_t) info->ne[3]; (int64_t) info->ne[3];
if (ne % ggml_blck_size(info->type) != 0) { if (ne % ggml_blck_size(info->type) != 0) {
fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%d)\n", fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%" PRId64 ")\n",
__func__, info->name.data, (int)info->type, ggml_type_name(info->type), ne, ggml_blck_size(info->type)); __func__, info->name.data, (int) info->type, ggml_type_name(info->type), ne, ggml_blck_size(info->type));
fclose(file); fclose(file);
gguf_free(ctx); gguf_free(ctx);
return NULL; return NULL;

4
tests/test-double-float.cpp
View File

@ -14,7 +14,7 @@
#pragma GCC diagnostic push #pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wdouble-promotion" #pragma GCC diagnostic ignored "-Wdouble-promotion"
// ggml.c::quantize_row_q4_0_reference // ggml.c::quantize_row_q4_0_ref
inline static uint8_t round_orig(float v0) { return ((int8_t) (round(v0))) + 8; } inline static uint8_t round_orig(float v0) { return ((int8_t) (round(v0))) + 8; }
// ggml.c::ggml_silu_f32 // ggml.c::ggml_silu_f32
@ -24,7 +24,7 @@ inline static float silu_orig(float x) {
#pragma GCC diagnostic pop #pragma GCC diagnostic pop
// ggml.c::quantize_row_q4_0_reference // ggml.c::quantize_row_q4_0_ref
inline static uint8_t round_float(float v0) { return (int8_t)roundf(v0) + 8; } inline static uint8_t round_float(float v0) { return (int8_t)roundf(v0) + 8; }
// ggml.c::ggml_silu_f32 // ggml.c::ggml_silu_f32

2
tests/test-quantize-fns.cpp
View File

@ -60,7 +60,7 @@ static float reference_quantization_error(ggml_type_traits_t & qfns, size_t test
qfns.from_float(test_data, tmp_q.data(), test_size); qfns.from_float(test_data, tmp_q.data(), test_size);
qfns.to_float(tmp_q.data(), tmp_out.data(), test_size); qfns.to_float(tmp_q.data(), tmp_out.data(), test_size);
qfns.from_float_reference(test_data, tmp_q.data(), test_size); qfns.from_float_ref(test_data, tmp_q.data(), test_size);
qfns.to_float(tmp_q.data(), tmp_out_ref.data(), test_size); qfns.to_float(tmp_q.data(), tmp_out_ref.data(), test_size);
return array_rmse(tmp_out.data(), tmp_out_ref.data(), test_size); return array_rmse(tmp_out.data(), tmp_out_ref.data(), test_size);

2
tests/test-quantize-perf.cpp
View File

@ -285,7 +285,7 @@ int main(int argc, char * argv[]) {
for (size_t size : params.test_sizes) { for (size_t size : params.test_sizes) {
printf(" %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024)); printf(" %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
auto quantize_fn = [&](void) -> float { auto quantize_fn = [&](void) -> float {
qfns.from_float_reference(test_data1, test_q1, size); qfns.from_float_ref(test_data1, test_q1, size);
return test_q1[0]; return test_q1[0];
}; };
size_t quantized_size = ggml_row_size(type, size); size_t quantized_size = ggml_row_size(type, size);