root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-13 04:00:16 +00:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

i2_s to absmax

Browse Source
This commit is contained in:
Eddie-Wang1120 2024-06-15 10:10:40 +08:00
parent 7a8961fff5
commit 95dced07e4
1 changed files with 6 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
ggml-quants.c
View File

@ -3333,12 +3333,11 @@ size_t quantize_i2_s(const float * restrict src, void * restrict dst, int64_t nr
int n = nrow * n_per_row; int n = nrow * n_per_row;
// f32 -> q8 // f32 -> q8
double i2_scale = 0; double max = 0;
for (int i=0; i<n; i++) { for (int i = 0; i < n; ++i) {
if (fabs((double)(src[i])) > 1e-6) { max = MAX(max, (double)fabs((double)src[i]));
i2_scale = (double)src[i];
}
} }
double i2_scale = max;
uint8_t* q8 = (uint8_t*)dst; uint8_t* q8 = (uint8_t*)dst;
for (int i=0; i<n; i++) { for (int i=0; i<n; i++) {
@ -3363,11 +3362,9 @@ size_t quantize_i2_s(const float * restrict src, void * restrict dst, int64_t nr
} }
float* scale_ptr = (float*)((char*)i2_weight + n / 4); float* scale_ptr = (float*)((char*)i2_weight + n / 4);
for (int i=0; i<8; i++) { scale_ptr[0] = i2_scale;
scale_ptr[i] = i2_scale;
}
// 32B for scale // 32B for alignment
return nrow * row_size / 4 + 32; return nrow * row_size / 4 + 32;
} }