vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

iq3s_mult: quantization tuning

Browse source
This commit is contained in:
Iwan Kawrakow 2024-03-03 11:32:53 +02:00
parent 5b9c8785fa
commit 8b713a987e
1 changed files with 5 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

6
ggml-quants.c
View file

@ -12092,7 +12092,11 @@ static void quantize_row_iq3_s_impl(int block_size, const float * restrict x, vo
} }
float d = max_scale/31; float d = max_scale/31;
y[ibl].d = GGML_FP32_TO_FP16(d * 1.025f); //1.033f); #ifdef IQ3S_SLOW_MULT
y[ibl].d = GGML_FP32_TO_FP16(d * 1.025f);
#else
y[ibl].d = GGML_FP32_TO_FP16(d * 1.030f);
#endif
float id = 1/d; float id = 1/d;
for (int ib = 0; ib < QK_K/block_size; ib += 2) { for (int ib = 0; ib < QK_K/block_size; ib += 2) {
int l1 = nearest_int(0.5f*(id*scales[ib+0]-1)); int l1 = nearest_int(0.5f*(id*scales[ib+0]-1));