ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-05 14:00:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

bitnet: scale is per row, not per tensor

Browse Source
This commit is contained in:
Iwan Kawrakow
2024-06-16 17:27:18 +03:00
parent 7f8901dca1
commit 4fcfcd05d1
2 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
ggml-quants.c
View File

@@ -3731,7 +3731,7 @@ void quantize_row_q8_K64_reference(const float * restrict x, block_q8_K64 * rest
for (int j = 0; j < 64; ++j) {
float ax = fabsf(x[j]);
if (ax > amax) {
amax = ax; max = x[j];
amax = ax; max = x[j];
}
}
if (!amax) {

4
iqk_mul_mat.cpp
View File

@@ -1347,12 +1347,12 @@ static void mul_mat_iq1bn_q8_K64(int n, const void * vx, size_t bx, const DataIn
typedef union { float f; uint32_t i; } scale_t;
scale_t scale;
uint16_t u = x[0].extra & 0xff;
scale.i = ((((u >> 4) | 0xf0) - 132) << 23) | ((u & 0x0f) << 19);
for (int ix = 0; ix < nrc_x; ++ix) {
x = (const block_iq1_bn *)((const char *)vx + ix*bx);
uint16_t u = x[0].extra & 0xff;
scale.i = ((((u >> 4) | 0xf0) - 132) << 23) | ((u & 0x0f) << 19);
for (int iy = 0; iy < nrc_y; ++iy) accd[iy] = _mm256_setzero_ps();