root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-28 12:24:35 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

Fix misaligned memory access in Q4_1 kernel

Browse Source
This commit is contained in:
Iwan Kawrakow 2023-07-28 17:12:27 +03:00
parent e5d23f2e7e
commit 72af25998c
1 changed files with 3 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
ggml-cuda.cu
View File

@ -1368,7 +1368,9 @@ static __device__ __forceinline__ float vec_dot_q4_1_q8_1(
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
const block_q4_1 * bq4_1 = (const block_q4_1 *) vbq;
const int vi = *((int *) &bq4_1->qs[sizeof(int) * (iqs + 0)]);
int vi;
memcpy(&vi, &bq4_1->qs[sizeof(int) * (iqs + 0)], sizeof(vi));
//const int vi = *((int *) &bq4_1->qs[sizeof(int) * (iqs + 0)]);
const int ui0 = *((int *) &bq8_1->qs[sizeof(int) * (iqs + 0)]);
const int ui1 = *((int *) &bq8_1->qs[sizeof(int) * (iqs + QI4_1)]);