ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 19:31:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

sanitize imatrix: iq4_nl_r4

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-08-27 14:36:00 +03:00
parent 44fc0a5eb6
commit 3ab0e41793
1 changed files with 5 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
ggml/src/iqk/iqk_quantize.cpp
View File

@@ -4738,9 +4738,13 @@ size_t quantize_iq4_nl_r4(const float * src, void * dst, int64_t nrows, int64_t
GGML_ASSERT(nrows%4 == 0); GGML_ASSERT(nrows%4 == 0);
auto row_size_nl = ggml_row_size(GGML_TYPE_IQ4_NL, n_per_row); auto row_size_nl = ggml_row_size(GGML_TYPE_IQ4_NL, n_per_row);
std::vector<char> qtmp(4*row_size_nl); std::vector<char> qtmp(4*row_size_nl);
QHelper helper(imatrix, n_per_row, 32);
auto q_func = [] (const float * x, void * vy, int n_per_row, const float * imatrix) {
quantize_iq4_nl(x, (char *)vy, 1, n_per_row, imatrix);
};
char * qrow = (char *)dst; char * qrow = (char *)dst;
for (int row = 0; row < nrows; row += 4) { for (int row = 0; row < nrows; row += 4) {
quantize_iq4_nl(src, qtmp.data(), 4, n_per_row, imatrix); helper.quantize(4, src, qtmp.data(), row_size_nl, q_func);
repack_iq4_nl(4, n_per_row, (const block_iq4_nl *)qtmp.data(), (block_iq4_nl_r4 *)qrow, false); repack_iq4_nl(4, n_per_row, (const block_iq4_nl *)qtmp.data(), (block_iq4_nl_r4 *)qrow, false);
src += 4*n_per_row; src += 4*n_per_row;
qrow += 4*row_size_nl; qrow += 4*row_size_nl;