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 importance matrix: iq5_k, iq6_k

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-08-27 13:52:23 +03:00
parent b41e8ef6d4
commit b6f1bd68bb
1 changed files with 11 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
ggml/src/iqk/iqk_quantize.cpp
View File

@@ -3014,14 +3014,10 @@ void quantize_row_iq5_k(const float * x, void * vy, int64_t k) {
size_t quantize_iq5_k(const float * src, void * dst, int64_t nrows, int64_t n_per_row, const float * imatrix) {
GGML_ASSERT(n_per_row%QK_K == 0);
int nblock = n_per_row/QK_K;
char * qrow = (char *)dst;
for (int64_t row = 0; row < nrows; ++row) {
quantize_row_iq5_k_impl(src, (void *)qrow, n_per_row, imatrix);
src += n_per_row;
qrow += nblock*sizeof(block_iq5_k);
}
return nrows * nblock * sizeof(block_iq5_k);
QHelper helper(imatrix, n_per_row, 16);
auto row_size = ggml_row_size(GGML_TYPE_IQ5_K, n_per_row);
helper.quantize(nrows, src, dst, row_size, quantize_row_iq5_k_impl);
return nrows * row_size;
}
//
@@ -3377,19 +3373,18 @@ void quantize_row_iq6_k(const float * x, void * vy, int64_t k) {
size_t quantize_iq6_k(const float * src, void * dst, int64_t nrows, int64_t n_per_row, const float * imatrix) {
GGML_ASSERT(n_per_row%QK_K == 0);
int nblock = n_per_row/QK_K;
char * qrow = (char *)dst;
float values[128];
for (int i = 0; i < 64; ++i) {
values[i] = iq6nl_values[i];
values[i+64] = values[i] + S_IQ6K;
}
for (int64_t row = 0; row < nrows; ++row) {
quantize_row_iq6_k_impl(src, (void *)qrow, n_per_row, imatrix, values, values + 64);
src += n_per_row;
qrow += nblock*sizeof(block_iq6_k);
}
return nrows * nblock * sizeof(block_iq6_k);
auto q_func = [values] (const float * x, void * vy, int n_per_row, const float * imatrix) {
quantize_row_iq6_k_impl(x, vy, n_per_row, imatrix, values, values + 64);
};
auto row_size = ggml_row_size(GGML_TYPE_IQ6_K, n_per_row);
QHelper helper(imatrix, n_per_row, 16);
helper.quantize(nrows, src, dst, row_size, q_func);
return nrows * row_size;
}
namespace {