ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-10 22:10:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

q2_K tweaks

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-07-17 16:52:39 +03:00
parent 912b74c151
commit 8c944f29c5
1 changed files with 25 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
ggml/src/ggml-quants.c
View File

@@ -2425,7 +2425,7 @@ static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restri
memset(sw, 0, QK_K/16*sizeof(float));
float sumx2 = 0;
for (int j = 0; j < QK_K; ++j) sumx2 += x[j]*x[j];
float sigma2 = 0.75f*sumx2/QK_K;
float sigma2 = 0.5f*sumx2/QK_K;
for (int j = 0; j < QK_K/16; ++j) {
const float * restrict qw = quant_weights + QK_K * i + 16*j;
for (int l = 0; l < 16; ++l) weight[l] = qw[l] * sqrtf(sigma2 + x[16*j + l]*x[16*j + l]);
@@ -2440,6 +2440,30 @@ static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restri
y[i].dmin = GGML_FP32_TO_FP16(mm);
for (int j = 0; j < QK_K/16; ++j) {
const float * restrict qw = quant_weights + QK_K * i + 16*j;
for (int l = 0; l < 16; ++l) weight[l] = qw[l] * sqrtf(sigma2 + x[16*j + l]*x[16*j + l]);
int lmin = MAX(Ls[j]-1, 0);
int lmax = MIN(Ls[j]+1,15);
int mmin = MAX(Lm[j]-1, 0);
int mmax = MIN(Lm[j]+1,15);
float best_score = INFINITY;
for (int il = lmin; il <= lmax; ++il) {
float d = dm*il;
float id = d ? 1/d : 0.f;
for (int im = mmin; im <= mmax; ++im) {
float m = mm*im;
float score = 0;
for (int ii = 0; ii < 16; ++ii) {
int q = nearest_int((x[16*j + ii] + m)*id);
q = MAX(0, MIN(3, q));
float diff = d*q - m - x[16*j + ii];
score += weight[ii] * diff * diff;
}
if (score < best_score) {
best_score = score; Ls[j] = il; Lm[j] = im;
}
}
}
float d = dm*Ls[j];
float m = mm*Lm[j];
float id = d ? 1/d : 0.f;