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

Fixes for k-quants

Browse Source
This commit is contained in:
Kawrakow
2026-02-25 15:57:38 +00:00
parent daa9a2c764
commit 8feb02bde7
1 changed files with 15 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
ggml/src/ggml-quants.c
View File

@@ -1971,7 +1971,7 @@ static float make_qkx2_quants(int n, int nmax, const float * restrict x, const f
sum_x += w * x[i]; sum_x += w * x[i];
} }
if (min > 0) min = 0; if (min > 0) min = 0;
if (max == min) { if (max - min < 1e-10f) {
for (int i = 0; i < n; ++i) L[i] = 0; for (int i = 0; i < n; ++i) L[i] = 0;
*the_min = -min; *the_min = -min;
return 0.f; return 0.f;
@@ -2218,7 +2218,7 @@ static float make_qkx3_quants(int n, int nmax, const float * restrict x, const f
if (min > 0) { if (min > 0) {
min = 0; min = 0;
} }
if (max <= min) { if (max - min < 1e-10f) {
memset(L, 0, n); memset(L, 0, n);
*the_min = -min; *the_min = -min;
return 0.f; return 0.f;
@@ -2340,7 +2340,7 @@ static float make_qp_quants(int n, int nmax, const float * restrict x, uint8_t *
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {
max = MAX(max, x[i]); max = MAX(max, x[i]);
} }
if (!max) { // all zero if (max < 1e-16f) { // all zero
for (int i = 0; i < n; ++i) { L[i] = 0; } for (int i = 0; i < n; ++i) { L[i] = 0; }
return 0.f; return 0.f;
} }
@@ -2733,6 +2733,10 @@ void quantize_row_q4_K_ref(const float * restrict x, block_q4_K * restrict y, in
float av_x = sqrtf(sum_x2/32); float av_x = sqrtf(sum_x2/32);
for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]); for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
scales[j] = make_qkx2_quants(32, 15, x + 32*j, weights, L + 32*j, &mins[j], Laux, -1.f, 0.1f, 20, false); scales[j] = make_qkx2_quants(32, 15, x + 32*j, weights, L + 32*j, &mins[j], Laux, -1.f, 0.1f, 20, false);
if (isnan(scales[j])) {
printf("Oops: NaN scale\n");
GGML_ABORT("Fatal error");
}
float scale = scales[j]; float scale = scales[j];
if (scale > max_scale) { if (scale > max_scale) {
max_scale = scale; max_scale = scale;
@@ -2846,10 +2850,18 @@ static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restri
for (int l = 0; l < 32; ++l) sumw += weights[l]; for (int l = 0; l < 32; ++l) sumw += weights[l];
sw[j] = sumw; sw[j] = sumw;
scales[j] = make_qkx3_quants(32, 15, x + 32*j, weights, L + 32*j, &mins[j], Laux, -0.9f, 0.05f, 36, false); scales[j] = make_qkx3_quants(32, 15, x + 32*j, weights, L + 32*j, &mins[j], Laux, -0.9f, 0.05f, 36, false);
if (isnan(scales[j])) {
printf("%s: got NaN scale\n", __func__);
GGML_ABORT("Fatal error");
}
} }
float d_block = make_qp_quants(QK_K/32, 63, scales, Ls, sw); float d_block = make_qp_quants(QK_K/32, 63, scales, Ls, sw);
float m_block = make_qp_quants(QK_K/32, 63, mins, Lm, sw); float m_block = make_qp_quants(QK_K/32, 63, mins, Lm, sw);
if (isnan(d_block) || isnan(m_block)) {
printf("%s: d_block = %g, m_block = %g\n", __func__, (double)d_block, (double)m_block);
GGML_ABORT("Fatal error");
}
for (int j = 0; j < QK_K/32; ++j) { for (int j = 0; j < QK_K/32; ++j) {
uint8_t ls = Ls[j]; uint8_t ls = Ls[j];
uint8_t lm = Lm[j]; uint8_t lm = Lm[j];