Skip to content

Fix IQ1_S quantization #6287

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
wants to merge 1 commit into
base: master
Choose a base branch
from
+10 −4
Draft

Fix IQ1_S quantization #6287

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 10 additions & 4 deletions ggml-quants.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11779,6 +11779,8 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
uint16_t * index,
int8_t * shifts) {

float waux[IQ1S_BLOCK_SIZE];

const int gindex = iq2_data_index(GGML_TYPE_IQ1_S);

const uint64_t * kgrid_q2xs = iq2_data[gindex].grid;
Expand Down Expand Up @@ -11814,12 +11816,13 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
const float * xbl = x + QK_K*ibl;
float sumx2 = 0;
for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
float sigma2 = 2*sumx2/QK_K;
float sigma2 = sumx2/QK_K;

for (int ib = 0; ib < QK_K/block_size; ++ib) {
const float * xb = xbl + block_size*ib;
const float * qw = quant_weights + QK_K*ibl + block_size*ib;
for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
for (int i = 0; i < block_size; ++i) waux[i] = sqrtf(weight[i]);
float max = fabsf(xb[0]);
for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i]));
if (!max) {
Expand Down Expand Up @@ -11881,7 +11884,7 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy
if (grid_index < 0) {
all_on_grid = false;
const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
grid_index = iq1_find_best_neighbour2(neighbours, kgrid_q2xs, xb + 8*k, weight + 8*k, scale, xx, L + 8*k, NGRID_IQ1S);
grid_index = iq1_find_best_neighbour2(neighbours, kgrid_q2xs, xb + 8*k, waux + 8*k, scale, xx, L + 8*k, NGRID_IQ1S);
GGML_ASSERT(grid_index >= 0);
}
index[k] = grid_index;
Expand Down Expand Up @@ -11955,6 +11958,8 @@ static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy
uint16_t * index,
int8_t * shifts) {

float waux[IQ1M_BLOCK_SIZE];

const int gindex = iq2_data_index(GGML_TYPE_IQ1_M);

const uint64_t * kgrid_q2xs = iq2_data[gindex].grid;
Expand Down Expand Up @@ -11996,13 +12001,14 @@ static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy
const float * xbl = x + QK_K*ibl;
float sumx2 = 0;
for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
float sigma2 = 2*sumx2/QK_K;
float sigma2 = sumx2/QK_K;

for (int ib = 0; ib < QK_K/block_size; ++ib) {
const float * xb = xbl + block_size*ib;
if (quant_weights) {
const float * qw = quant_weights + QK_K*ibl + block_size*ib;
for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
for (int i = 0; i < block_size; ++i) waux[i] = sqrtf(weight[i]);
} else {
for (int i = 0; i < block_size; ++i) weight[i] = xb[i]*xb[i];
}
Expand Down Expand Up @@ -12127,7 +12133,7 @@ static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy
if (grid_index < 0) {
all_on_grid = false;
const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
grid_index = iq1_find_best_neighbour2(neighbours, kgrid_q2xs, xb + 8*k, weight + 8*k, scale, xx, L + 8*k, NGRID_IQ1S);
grid_index = iq1_find_best_neighbour2(neighbours, kgrid_q2xs, xb + 8*k, (quant_weights ? waux : weight) + 8*k, scale, xx, L + 8*k, NGRID_IQ1S);
GGML_ASSERT(grid_index >= 0);
}
index[k] = grid_index;
Expand Down