ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-25 15:44:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

iq4_kt: minor tweaks

Browse Source
This commit is contained in:
Iwan Kawrakow
2024-11-11 19:59:15 +02:00
parent e9ced1bbe6
commit de7fe92833
2 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-cuda/convert.cu
View File

@@ -423,7 +423,7 @@ static __global__ void dequantize_block_iq4_kt(const void * __restrict__ vx, dst
const uint16_t * ql = (const uint16_t *)x[i].ql;
uint32_t idx1 = ql[2*ib+0] + 4096;
uint32_t idx2 = ql[2*ib+1] + 4096;
const float dl = scale * x[i].scales[ib/8] * 31.75f;
const float dl = scale * x[i].scales[ib/8] * 31.75f * 1.01f;
uint32_t s[2];
const half * h = (const half *)s;
for (int j = 0; j < 4; ++j) {

2
ggml/src/ggml-cuda/dmmv.cu
View File

@@ -181,7 +181,7 @@ static __global__ void dequantize_mul_mat_vec_iq4_kt(const void * __restrict__ v
if (row > nrows) return;
const float * dptr = (const float *)((const char *)vx + row*row_size);
const float d = *dptr * 31.75f * 1.015f;
const float d = *dptr * 31.75f * 1.01f;
const block_iq4_kt * x = (const block_iq4_kt *)(dptr + 1);
const int num_blocks_per_row = ncols / QK_K;