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

Minor PPL improvement via a block scale fudge factor

Browse source
This commit is contained in:
Iwan Kawrakow 2024-02-20 13:10:46 +02:00
parent 5691fecd06
commit 76aff093b4
1 changed files with 1 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

3
ggml-quants.c
View file

@ -10881,7 +10881,7 @@ static void quantize_row_iq3_xs_impl(int grid_size, const float * restrict x, vo
}
float d = max_scale/31;
dh[0] = GGML_FP32_TO_FP16(d);
dh[0] = GGML_FP32_TO_FP16(d * 1.0125f); // small improvement via this fudge factor
float id = 1/d;
//float sumqx = 0, sumq2 = 0;
for (int ib = 0; ib < QK_K/32; ++ib) {
@ -10903,7 +10903,6 @@ static void quantize_row_iq3_xs_impl(int grid_size, const float * restrict x, vo
const int8_t * signs = keven_signs_q2xs + 8*((scales_and_signs[ib] >> 7*(k/2)) & 127) + 4*(k%2);
const float * xk = xb + 4*k;
const float * wk = weight + 4*k;
//const uint8_t * grid = (const uint8_t *)(kgrid_q3xs + q3[8*ib+k]);
int idx = q3[8*ib+k];
if (grid_size == 512) idx |= ((h << (8-k)) & 256);
const uint8_t * grid = (const uint8_t *)(kgrid_q3xs + idx);