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

Fix KT quantization yet again

Browse Source
This commit is contained in:
Kawrakow
2026-02-25 14:20:22 +00:00
parent c77ec4b8b8
commit ab3c601269
1 changed files with 15 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
ggml/src/iqk/iqk_quantize.cpp
View File

@@ -8733,6 +8733,11 @@ void quantize_row_iq1_kt_impl(const float * x, void * vy, int n_per_row, const f
float ax = std::abs(xb[j]); float ax = std::abs(xb[j]);
amax = std::max(amax, ax); amax = std::max(amax, ax);
} }
if (amax < 1e-16f) {
scales[ib] = 0.0f;
for (int ig = 0; ig < Q::kNg; ++ig) all_idx[(ibl*Q::kSuperBlockSize + ib*Q::kBlockSize)/Q::kGroupSize + ig] = 0;
continue;
}
float scale_0 = std::max(90.f, 124.f*amax/amax_row); float scale_0 = std::max(90.f, 124.f*amax/amax_row);
quantizer.find_best_match( amax/scale_0, xb, weight, best_idx); quantizer.find_best_match( amax/scale_0, xb, weight, best_idx);
auto [dp, score_p] = quantizer.find_best_scale(xb, weight, best_idx); auto [dp, score_p] = quantizer.find_best_scale(xb, weight, best_idx);
@@ -8998,6 +9003,11 @@ void quantize_row_iq2_kt_impl(const float * x, void * vy, int n_per_row, const f
float ax = std::abs(xb[j]); float ax = std::abs(xb[j]);
amax = std::max(amax, ax); amax = std::max(amax, ax);
} }
if (amax < 1e-16f) {
scales[ib] = 0.0f;
for (int ig = 0; ig < Q::kNg; ++ig) all_idx[(ibl*Q::kSuperBlockSize + ib*Q::kBlockSize)/Q::kGroupSize + ig] = 0;
continue;
}
float scale_0 = std::max(90.f, 124.f*amax/amax_row); float scale_0 = std::max(90.f, 124.f*amax/amax_row);
quantizer.find_best_match( amax/scale_0, xb, weight, best_idx); quantizer.find_best_match( amax/scale_0, xb, weight, best_idx);
auto [dp, score_p] = quantizer.find_best_scale(xb, weight, best_idx); auto [dp, score_p] = quantizer.find_best_scale(xb, weight, best_idx);
@@ -9289,8 +9299,10 @@ void quantize_row_iq3_kt_impl(const float * x, void * vy, int n_per_row, const f
xaux[j] = ax; xaux[j] = ax;
amax = std::max(amax, ax); amax = std::max(amax, ax);
} }
scales[ib] = 0; if (amax < 1e-16f) {
if (!amax) continue; scales[ib] = 0.0f;
continue;
}
//quantizer.find_best_match(amax/96.f, xaux, weight, best_idx+Q::kNg); //quantizer.find_best_match(amax/96.f, xaux, weight, best_idx+Q::kNg);
//scales[ib] = quantizer.find_best_scale(xaux, weight, best_idx+Q::kNg).first; //scales[ib] = quantizer.find_best_scale(xaux, weight, best_idx+Q::kNg).first;
@@ -9577,7 +9589,7 @@ void quantize_row_iq4_kt_impl(const float * x, void * vy, int n_per_row, const f
float ax = std::abs(xaux[j]); float ax = std::abs(xaux[j]);
amax = std::max(amax, ax); amax = std::max(amax, ax);
} }
if (!amax) { if (amax < 1e-16f) {
scales[ib] = 0; scales[ib] = 0;
continue; continue;
} }