ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-23 22:54:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

sanitize imatrix: iq5_ks

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-08-27 14:24:35 +03:00
parent 63efad8978
commit 44fc0a5eb6
1 changed files with 5 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
ggml/src/iqk/iqk_quantize.cpp
View File

@@ -4289,14 +4289,13 @@ size_t quantize_iq5_ks(const float * src, void * dst, int64_t nrows, int64_t n_p
constexpr int kBlockSize = 32;
GGML_ASSERT(n_per_row%QK_K == 0);
auto row_size = ggml_row_size(GGML_TYPE_IQ5_KS, n_per_row);
char * qrow = (char *)dst;
float weight[kBlockSize];
std::vector<float> all_scales(n_per_row/kBlockSize);
for (int64_t row = 0; row < nrows; ++row) {
quantize_row_iq5_ks_impl(QK_K, kBlockSize, n_per_row, src, qrow, all_scales.data(), weight, iq5nl_values, imatrix, 5);
src += n_per_row;
qrow += row_size;
}
QHelper helper(imatrix, n_per_row, kBlockSize);
auto q_func = [&all_scales, &weight, block_size = kBlockSize] (const float * x, void * vy, int n_per_row, const float * imatrix) {
quantize_row_iq5_ks_impl(QK_K, block_size, n_per_row, x, (char *)vy, all_scales.data(), weight, iq5nl_values, imatrix, 5);
};
helper.quantize(nrows, src, dst, row_size, q_func);
return nrows * row_size;
}