vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml : remove old quantization functions

Browse source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2024-03-08 15:20:39 +02:00
parent fb215c3832
commit 13c1cc6a9f
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
6 changed files with 100 additions and 345 deletions
Download patch file Download diff file Expand all files Collapse all files

37
examples/llava/clip.cpp
View file

@ -1919,42 +1919,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
std::vector<int64_t> hist_cur(1 << 4, 0);
switch (new_type) {
case GGML_TYPE_Q4_0: {
new_size = ggml_quantize_q4_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q4_1: {
new_size = ggml_quantize_q4_1(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q5_0: {
new_size = ggml_quantize_q5_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q5_1: {
new_size = ggml_quantize_q5_1(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q8_0: {
new_size = ggml_quantize_q8_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q2_K: {
new_size = ggml_quantize_q2_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q3_K: {
new_size = ggml_quantize_q3_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q4_K: {
new_size = ggml_quantize_q4_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q5_K: {
new_size = ggml_quantize_q5_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q6_K: {
new_size = ggml_quantize_q6_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
} break;
default: {
fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, new_type);
return false;
}
}
new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, n_elms/cur->ne[0], cur->ne[0], hist_cur.data(), nullptr);
for (size_t j = 0; j < hist_cur.size(); ++j) {
hist_all[j] += hist_cur[j];

129
ggml-quants.c
View file

@ -1704,16 +1704,6 @@ void quantize_row_q2_K(const float * restrict x, void * restrict vy, int k) {
quantize_row_q2_K_reference(x, vy, k);
}
size_t ggml_quantize_q2_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) {
(void)hist; // TODO: collect histograms
for (int j = 0; j < n; j += k) {
block_q2_K * restrict y = (block_q2_K *)dst + j/QK_K;
quantize_row_q2_K_reference(src + j, y, k);
}
return (n/QK_K*sizeof(block_q2_K));
}
static float make_qkx3_quants(int n, int nmax, const float * restrict x, const float * restrict weights,
uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux,
float rmin, float rdelta, int nstep, bool use_mad) {
@ -1966,8 +1956,7 @@ static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restri
}
}
size_t quantize_q2_K(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
(void)hist;
size_t quantize_q2_K(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
size_t row_size = ggml_row_size(GGML_TYPE_Q2_K, n_per_row);
if (!quant_weights) {
quantize_row_q2_K_reference(src, dst, nrow*n_per_row);
@ -2186,16 +2175,6 @@ void quantize_row_q3_K(const float * restrict x, void * restrict vy, int k) {
quantize_row_q3_K_reference(x, vy, k);
}
size_t ggml_quantize_q3_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) {
(void)hist; // TODO: collect histograms
for (int j = 0; j < n; j += k) {
block_q3_K * restrict y = (block_q3_K *)dst + j/QK_K;
quantize_row_q3_K_reference(src + j, y, k);
}
return (n/QK_K*sizeof(block_q3_K));
}
static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restrict y, int n_per_row, const float * restrict quant_weights) {
#if QK_K != 256
(void)quant_weights;
@ -2285,8 +2264,7 @@ static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restri
#endif
}
size_t quantize_q3_K(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
(void)hist;
size_t quantize_q3_K(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
size_t row_size = ggml_row_size(GGML_TYPE_Q3_K, n_per_row);
if (!quant_weights) {
quantize_row_q3_K_reference(src, dst, nrow*n_per_row);
@ -2456,17 +2434,6 @@ void quantize_row_q4_K(const float * restrict x, void * restrict vy, int k) {
quantize_row_q4_K_reference(x, y, k);
}
size_t ggml_quantize_q4_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) {
assert(k % QK_K == 0);
(void)hist; // TODO: collect histograms
for (int j = 0; j < n; j += k) {
block_q4_K * restrict y = (block_q4_K *)dst + j/QK_K;
quantize_row_q4_K_reference(src + j, y, k);
}
return (n/QK_K*sizeof(block_q4_K));
}
static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restrict y, int n_per_row, const float * quant_weights) {
#if QK_K != 256
(void)quant_weights;
@ -2545,8 +2512,7 @@ static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restri
#endif
}
size_t quantize_q4_K(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
(void)hist;
size_t quantize_q4_K(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
size_t row_size = ggml_row_size(GGML_TYPE_Q4_K, n_per_row);
if (!quant_weights) {
quantize_row_q4_K_reference(src, dst, nrow*n_per_row);
@ -2757,17 +2723,6 @@ void quantize_row_q5_K(const float * restrict x, void * restrict vy, int k) {
quantize_row_q5_K_reference(x, y, k);
}
size_t ggml_quantize_q5_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) {
assert(k % QK_K == 0);
(void)hist; // TODO: collect histograms
for (int j = 0; j < n; j += k) {
block_q5_K * restrict y = (block_q5_K *)dst + j/QK_K;
quantize_row_q5_K_reference(src + j, y, k);
}
return (n/QK_K*sizeof(block_q5_K));
}
static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restrict y, int n_per_row, const float * quant_weights) {
#if QK_K != 256
(void)quant_weights;
@ -2866,8 +2821,7 @@ static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restri
#endif
}
size_t quantize_q5_K(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
(void)hist;
size_t quantize_q5_K(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
size_t row_size = ggml_row_size(GGML_TYPE_Q5_K, n_per_row);
if (!quant_weights) {
quantize_row_q5_K_reference(src, dst, nrow*n_per_row);
@ -3020,17 +2974,6 @@ void quantize_row_q6_K(const float * restrict x, void * restrict vy, int k) {
quantize_row_q6_K_reference(x, y, k);
}
size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist) {
assert(k % QK_K == 0);
(void)hist; // TODO: collect histograms
for (int j = 0; j < n; j += k) {
block_q6_K * restrict y = (block_q6_K *)dst + j/QK_K;
quantize_row_q6_K_reference(src + j, y, k);
}
return (n/QK_K*sizeof(block_q6_K));
}
static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restrict y, int n_per_row, const float * quant_weights) {
#if QK_K != 256
(void)quant_weights;
@ -3120,8 +3063,7 @@ static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restri
#endif
}
size_t quantize_q6_K(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
(void)hist;
size_t quantize_q6_K(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
size_t row_size = ggml_row_size(GGML_TYPE_Q6_K, n_per_row);
if (!quant_weights) {
quantize_row_q6_K_reference(src, dst, nrow*n_per_row);
@ -3165,9 +3107,10 @@ static void quantize_row_q4_0_impl(const float * restrict x, block_q4_0 * restri
}
}
size_t quantize_q4_0(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
size_t quantize_q4_0(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
if (!quant_weights) {
return ggml_quantize_q4_0(src, dst, nrow*n_per_row, n_per_row, hist);
quantize_row_q4_0_reference(src, dst, nrow*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);
char * qrow = (char *)dst;
@ -3209,9 +3152,10 @@ static void quantize_row_q4_1_impl(const float * restrict x, block_q4_1 * restri
}
}
size_t quantize_q4_1(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
size_t quantize_q4_1(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
if (!quant_weights) {
return ggml_quantize_q4_1(src, dst, nrow*n_per_row, n_per_row, hist);
quantize_row_q4_1_reference(src, dst, nrow*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);
char * qrow = (char *)dst;
@ -3262,9 +3206,10 @@ static void quantize_row_q5_0_impl(const float * restrict x, block_q5_0 * restri
}
}
size_t quantize_q5_0(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
size_t quantize_q5_0(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
if (!quant_weights) {
return ggml_quantize_q5_0(src, dst, nrow*n_per_row, n_per_row, hist);
quantize_row_q5_0_reference(src, dst, nrow*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);
char * qrow = (char *)dst;
@ -3314,9 +3259,10 @@ static void quantize_row_q5_1_impl(const float * restrict x, block_q5_1 * restri
}
}
size_t quantize_q5_1(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
size_t quantize_q5_1(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
if (!quant_weights) {
return ggml_quantize_q5_1(src, dst, nrow*n_per_row, n_per_row, hist);
quantize_row_q5_1_reference(src, dst, nrow*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);
char * qrow = (char *)dst;
@ -3328,6 +3274,13 @@ size_t quantize_q5_1(const float * src, void * dst, int nrow, int n_per_row, int
return nrow * row_size;
}
size_t quantize_q8_0(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
(void)quant_weights; // not used
const size_t row_size = ggml_row_size(GGML_TYPE_Q8_0, n_per_row);
quantize_row_q8_0_reference(src, dst, nrow*n_per_row);
return nrow * row_size;
}
// ====================== "True" 2-bit (de)-quantization
void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y, int k) {
@ -10698,8 +10651,7 @@ static void quantize_row_iq2_xs_impl(const float * restrict x, void * restrict v
}
}
size_t quantize_iq2_xxs(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
(void)hist;
size_t quantize_iq2_xxs(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
GGML_ASSERT(n_per_row%QK_K == 0);
int nblock = n_per_row/QK_K;
char * qrow = (char *)dst;
@ -10711,8 +10663,7 @@ size_t quantize_iq2_xxs(const float * src, void * dst, int nrow, int n_per_row,
return nrow * nblock * sizeof(block_iq2_xxs);
}
size_t quantize_iq2_xs(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
(void)hist;
size_t quantize_iq2_xs(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
GGML_ASSERT(n_per_row%QK_K == 0);
int nblock = n_per_row/QK_K;
char * qrow = (char *)dst;
@ -11154,8 +11105,7 @@ static void quantize_row_iq3_xxs_impl(int grid_size, const float * restrict x, v
}
}
size_t quantize_iq3_xxs(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
(void)hist;
size_t quantize_iq3_xxs(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
GGML_ASSERT(n_per_row%QK_K == 0);
int nblock = n_per_row/QK_K;
char * qrow = (char *)dst;
@ -11361,8 +11311,7 @@ static void quantize_row_iq3_s_impl(int block_size, const float * restrict x, vo
}
#define IQ3S_BLOCK_SIZE 32
size_t quantize_iq3_s(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
(void)hist;
size_t quantize_iq3_s(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
GGML_ASSERT(n_per_row%QK_K == 0);
int nblock = n_per_row/QK_K;
float scales[QK_K/IQ3S_BLOCK_SIZE];
@ -11392,7 +11341,7 @@ void quantize_row_iq3_s(const float * restrict x, void * restrict vy, int k) {
void quantize_row_iq3_s_reference(const float * restrict x, block_iq3_s * restrict y, int k) {
assert(k % QK_K == 0);
quantize_iq3_s(x, y, 1, k, NULL, NULL);
quantize_iq3_s(x, y, 1, k, NULL);
}
@ -11587,8 +11536,7 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
}
}
size_t quantize_iq1_s(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
(void)hist;
size_t quantize_iq1_s(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
GGML_ASSERT(n_per_row%QK_K == 0);
int nblock = n_per_row/QK_K;
char * qrow = (char *)dst;
@ -11721,8 +11669,7 @@ static void quantize_row_iq4_nl_impl(const int super_block_size, const int block
}
}
size_t quantize_iq4_nl(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
(void)hist;
size_t quantize_iq4_nl(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
GGML_ASSERT(n_per_row%QK4_NL == 0);
int nblock = n_per_row/QK4_NL;
char * qrow = (char *)dst;
@ -11752,14 +11699,13 @@ void quantize_row_iq4_nl(const float * restrict x, void * restrict vy, int k) {
void quantize_row_iq4_nl_reference(const float * restrict x, block_iq4_nl * restrict y, int k) {
assert(k % QK4_NL == 0);
quantize_iq4_nl(x, y, 1, k, NULL, NULL);
quantize_iq4_nl(x, y, 1, k, NULL);
}
size_t quantize_iq4_xs(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
size_t quantize_iq4_xs(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
#if QK_K == 64
return quantize_iq4_nl(src, dst, nrow, n_per_row, hist, quant_weights);
return quantize_iq4_nl(src, dst, nrow, n_per_row, quant_weights);
#else
(void)hist;
GGML_ASSERT(n_per_row%QK_K == 0);
int nblock = n_per_row/QK_K;
char * qrow = (char *)dst;
@ -11788,7 +11734,7 @@ void quantize_row_iq4_xs(const float * restrict x, void * restrict vy, int k) {
void quantize_row_iq4_xs_reference(const float * restrict x, block_iq4_xs * restrict y, int k) {
assert(k % QK_K == 0);
quantize_iq4_xs(x, y, 1, k, NULL, NULL);
quantize_iq4_xs(x, y, 1, k, NULL);
}
// =============================== 2.5625 bpw
@ -11961,8 +11907,7 @@ static void quantize_row_iq2_s_impl(const float * restrict x, void * restrict vy
}
}
size_t quantize_iq2_s(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
(void)hist;
size_t quantize_iq2_s(const float * src, void * dst, int nrow, int n_per_row, const float * quant_weights) {
GGML_ASSERT(n_per_row%QK_K == 0);
int nblock = n_per_row/QK_K;
char * qrow = (char *)dst;
@ -11976,7 +11921,7 @@ size_t quantize_iq2_s(const float * src, void * dst, int nrow, int n_per_row, in
void quantize_row_iq2_s_reference(const float * restrict x, block_iq2_s * restrict y, int k) {
assert(k % QK_K == 0);
quantize_iq2_s(x, y, 1, k, NULL, NULL);
quantize_iq2_s(x, y, 1, k, NULL);
}
void quantize_row_iq2_s(const float * restrict x, void * restrict vy, int k) {

35
ggml-quants.h
View file

@ -333,23 +333,24 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const
//
// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
//
size_t quantize_iq2_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq2_xs (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq2_s (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq3_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq1_s (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq4_nl (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq4_xs (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq3_s (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_q2_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_q3_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_q4_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_q5_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_q6_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_q4_0 (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_q4_1 (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_q5_0 (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_q5_1 (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
size_t quantize_iq2_xxs(const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_iq2_xs (const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_iq2_s (const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_iq3_xxs(const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_iq1_s (const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_iq4_nl (const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_iq4_xs (const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_iq3_s (const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q2_K (const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q3_K (const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q4_K (const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q5_K (const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q6_K (const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q4_0 (const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q4_1 (const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q5_0 (const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q5_1 (const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
size_t quantize_q8_0 (const float * src, void * dst, int nrows, int n_per_row, const float * imatrix);
void iq2xs_init_impl(enum ggml_type type);
void iq2xs_free_impl(enum ggml_type type);

38
ggml-vulkan.cpp
View file

@ -4104,43 +4104,7 @@ static void ggml_vk_test_transfer(ggml_backend_vk_context * ctx, size_t ne, bool
static void ggml_vk_quantize_data(const float * from, void * to, size_t ne, ggml_type quant) {
std::vector<int64_t> hist_cur(1 << 4, 0);
switch(quant) {
case GGML_TYPE_F32:
memcpy(to, from, sizeof(float) * ne);
break;
case GGML_TYPE_Q4_0:
ggml_quantize_q4_0(from, to, ne, ne, hist_cur.data());
break;
case GGML_TYPE_Q4_1:
ggml_quantize_q4_1(from, to, ne, ne, hist_cur.data());
break;
case GGML_TYPE_Q5_0:
ggml_quantize_q5_0(from, to, ne, ne, hist_cur.data());
break;
case GGML_TYPE_Q5_1:
ggml_quantize_q5_1(from, to, ne, ne, hist_cur.data());
break;
case GGML_TYPE_Q8_0:
ggml_quantize_q8_0(from, to, ne, ne, hist_cur.data());
break;
case GGML_TYPE_Q2_K:
ggml_quantize_q2_K(from, to, ne, ne, hist_cur.data());
break;
case GGML_TYPE_Q3_K:
ggml_quantize_q3_K(from, to, ne, ne, hist_cur.data());
break;
case GGML_TYPE_Q4_K:
ggml_quantize_q4_K(from, to, ne, ne, hist_cur.data());
break;
case GGML_TYPE_Q5_K:
ggml_quantize_q5_K(from, to, ne, ne, hist_cur.data());
break;
case GGML_TYPE_Q6_K:
ggml_quantize_q6_K(from, to, ne, ne, hist_cur.data());
break;
default:
GGML_ASSERT(false);
}
gml_quantize_chunk(quant, from, to, 0, 1, ne, hist_cur.data(), nullptr);
}
static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_type quant) {

182
ggml.c
View file

@ -20159,133 +20159,6 @@ void ggml_quantize_free(void) {
ggml_critical_section_end();
}
size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist) {
assert(k % QK4_0 == 0);
const int nb = k / QK4_0;
for (int b = 0; b < n; b += k) {
block_q4_0 * restrict y = (block_q4_0 *) dst + b/QK4_0;
quantize_row_q4_0_reference(src + b, y, k);
for (int i = 0; i < nb; i++) {
for (int j = 0; j < QK4_0; j += 2) {
const uint8_t vi0 = y[i].qs[j/2] & 0x0F;
const uint8_t vi1 = y[i].qs[j/2] >> 4;
hist[vi0]++;
hist[vi1]++;
}
}
}
return (n/QK4_0*sizeof(block_q4_0));
}
size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist) {
assert(k % QK4_1 == 0);
const int nb = k / QK4_1;
for (int b = 0; b < n; b += k) {
block_q4_1 * restrict y = (block_q4_1 *) dst + b/QK4_1;
quantize_row_q4_1_reference(src + b, y, k);
for (int i = 0; i < nb; i++) {
for (int j = 0; j < QK4_1; j += 2) {
const uint8_t vi0 = y[i].qs[j/2] & 0x0F;
const uint8_t vi1 = y[i].qs[j/2] >> 4;
hist[vi0]++;
hist[vi1]++;
}
}
}
return (n/QK4_1*sizeof(block_q4_1));
}
size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist) {
assert(k % QK5_0 == 0);
const int nb = k / QK5_0;
for (int b = 0; b < n; b += k) {
block_q5_0 * restrict y = (block_q5_0 *)dst + b/QK5_0;
quantize_row_q5_0_reference(src + b, y, k);
for (int i = 0; i < nb; i++) {
uint32_t qh;
memcpy(&qh, &y[i].qh, sizeof(qh));
for (int j = 0; j < QK5_0; j += 2) {
const uint8_t vh0 = ((qh & (1u << (j/2 + 0 ))) >> (j/2 + 0 )) << 4;
const uint8_t vh1 = ((qh & (1u << (j/2 + 16))) >> (j/2 + 12));
// cast to 16 bins
const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2;
const uint8_t vi1 = ((y[i].qs[j/2] >> 4) | vh1) / 2;
hist[vi0]++;
hist[vi1]++;
}
}
}
return (n/QK5_0*sizeof(block_q5_0));
}
size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist) {
assert(k % QK5_1 == 0);
const int nb = k / QK5_1;
for (int b = 0; b < n; b += k) {
block_q5_1 * restrict y = (block_q5_1 *)dst + b/QK5_1;
quantize_row_q5_1_reference(src + b, y, k);
for (int i = 0; i < nb; i++) {
uint32_t qh;
memcpy(&qh, &y[i].qh, sizeof(qh));
for (int j = 0; j < QK5_1; j += 2) {
const uint8_t vh0 = ((qh & (1u << (j/2 + 0 ))) >> (j/2 + 0 )) << 4;
const uint8_t vh1 = ((qh & (1u << (j/2 + 16))) >> (j/2 + 12));
// cast to 16 bins
const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2;
const uint8_t vi1 = ((y[i].qs[j/2] >> 4) | vh1) / 2;
hist[vi0]++;
hist[vi1]++;
}
}
}
return (n/QK5_1*sizeof(block_q5_1));
}
size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist) {
assert(k % QK8_0 == 0);
const int nb = k / QK8_0;
for (int b = 0; b < n; b += k) {
block_q8_0 * restrict y = (block_q8_0 *)dst + b/QK8_0;
quantize_row_q8_0_reference(src + b, y, k);
for (int i = 0; i < nb; i++) {
for (int j = 0; j < QK8_0; ++j) {
const int8_t vi = y[i].qs[j];
hist[vi/16 + 8]++;
}
}
}
return (n/QK8_0*sizeof(block_q8_0));
}
bool ggml_quantize_requires_imatrix(enum ggml_type type) {
return
type == GGML_TYPE_IQ2_XXS ||
@ -20293,8 +20166,15 @@ bool ggml_quantize_requires_imatrix(enum ggml_type type) {
type == GGML_TYPE_IQ1_S;
}
size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start,
int nrows, int n_per_row, int64_t * hist, const float * imatrix) {
size_t ggml_quantize_chunk(
enum ggml_type type,
const float * src,
void * dst,
int start,
int nrows,
int n_per_row,
int64_t * hist,
const float * imatrix) {
ggml_quantize_init(type); // this is noop if already initialized
size_t result = 0;
int n = nrows * n_per_row;
@ -20305,7 +20185,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_q4_0(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
result = quantize_q4_0(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_Q4_1:
@ -20314,7 +20194,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_q4_1(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
result = quantize_q4_1(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_Q5_0:
@ -20323,7 +20203,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_q5_0(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
result = quantize_q5_0(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_Q5_1:
@ -20332,14 +20212,17 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_q5_1(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
result = quantize_q5_1(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_Q8_0:
{
GGML_ASSERT(start % QK8_0 == 0);
block_q8_0 * block = (block_q8_0*)dst + start / QK8_0;
result = ggml_quantize_q8_0(src + start, block, n, n, hist);
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_q8_0(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_Q2_K:
{
@ -20347,7 +20230,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_q2_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
result = quantize_q2_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_Q3_K:
@ -20356,7 +20239,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_q3_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
result = quantize_q3_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_Q4_K:
@ -20365,7 +20248,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_q4_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
result = quantize_q4_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_Q5_K:
@ -20374,7 +20257,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_q5_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
result = quantize_q5_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_Q6_K:
@ -20383,7 +20266,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_q6_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
result = quantize_q6_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_IQ2_XXS:
@ -20393,7 +20276,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
GGML_ASSERT(imatrix);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_iq2_xxs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
result = quantize_iq2_xxs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_IQ2_XS:
@ -20403,7 +20286,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
GGML_ASSERT(imatrix);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_iq2_xs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
result = quantize_iq2_xs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_IQ3_XXS:
@ -20412,7 +20295,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_iq3_xxs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
result = quantize_iq3_xxs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_IQ3_S:
@ -20421,7 +20304,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_iq3_s(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
result = quantize_iq3_s(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_IQ2_S:
@ -20430,7 +20313,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_iq2_s(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
result = quantize_iq2_s(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_IQ1_S:
@ -20439,7 +20322,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_iq1_s(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
result = quantize_iq1_s(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
case GGML_TYPE_IQ4_NL:
@ -20451,7 +20334,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_iq4_nl(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
result = quantize_iq4_nl(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
#if QK_K != 64
@ -20461,7 +20344,7 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
GGML_ASSERT(start % n_per_row == 0);
size_t start_row = start / n_per_row;
size_t row_size = ggml_row_size(type, n_per_row);
result = quantize_iq4_xs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
result = quantize_iq4_xs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, imatrix);
GGML_ASSERT(result == row_size * nrows);
} break;
#endif
@ -20480,6 +20363,9 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
default:
assert(false);
}
GGML_UNUSED(hist); // TODO: poppulate
return result;
}

24
ggml.h
View file

@ -2194,25 +2194,19 @@ extern "C" {
GGML_API void ggml_quantize_init(enum ggml_type type);
GGML_API void ggml_quantize_free(void);
// TODO: these would probably get removed in favor of the more general ggml_quantize_chunk
GGML_API size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);
// some quantization type cannot be used without an importance matrix
GGML_API bool ggml_quantize_requires_imatrix(enum ggml_type type);
// calls ggml_quantize_init internally (i.e. can allocate memory)
GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst,
int start, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
GGML_API size_t ggml_quantize_chunk(
enum ggml_type type,
const float * src,
void * dst,
int start,
int nrows,
int n_per_row,
int64_t * hist,
const float * imatrix);
//
// gguf