ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-24 08:29:29 +00:00
Code Issues Packages Projects Releases Wiki Activity

Yet another

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-05-03 20:00:20 +03:00
parent 5605474193
commit 5782f1bdf0
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
ggml/src/iqk/iqk_quantize.cpp
View File

@@ -6721,12 +6721,12 @@ size_t quantize_iq1_m_r4(const float * src, void * dst, int64_t nrows, int64_t n
for (int j = 0; j < kBlockSize; ++j) weight[j] = imatrix[kBlockSize*ibl + j]*sqrt(sigma2 + xb[j]*xb[j]);
float sumwx = 0;
for (int j = 0; j < kBlockSize/2; ++j) sumwx += weight[j]*std::abs(xb[j]);
if (!sumwx) {
if (sumwx < 1e-14f) {
for (int j = 0; j < kBlockSize/2; ++j) weight[j] = sqrt(sigma2 + xb[j]*xb[j]);
}
sumwx = 0;
for (int j = kBlockSize/2; j < kBlockSize; ++j) sumwx += weight[j]*std::abs(xb[j]);
if (!sumwx) {
if (sumwx < 1e-14) {
for (int j = kBlockSize/2; j < kBlockSize; ++j) weight[j] = sqrt(sigma2 + xb[j]*xb[j]);
}
} else {