ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-26 17:20:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix overflow in offset calculation in mmq

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-12-12 13:22:02 +00:00
parent b74fb479af
commit 50fbde85dc
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-cuda/mmq.cuh
View File

@@ -3875,7 +3875,7 @@ static __device__ void mul_mat_q_process_tile(
const int * y = (const int *) yc + jt*(mmq_x*sizeof(block_q8_1_mmq)/sizeof(int)); const int * y = (const int *) yc + jt*(mmq_x*sizeof(block_q8_1_mmq)/sizeof(int));
for (int kb0 = kb0_start; kb0 < kb0_stop; kb0 += blocks_per_iter) { for (int kb0 = kb0_start; kb0 < kb0_stop; kb0 += blocks_per_iter) {
load_tiles(x + stride01*it*mmq_y, tile_x, kb0, tile_x_max_i, stride01); load_tiles(x + int64_t(stride01)*it*mmq_y, tile_x, kb0, tile_x_max_i, stride01);
{ {
const int * by0 = y + stride11*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 0*sizeof(block_q8_1_mmq)/sizeof(int)); const int * by0 = y + stride11*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 0*sizeof(block_q8_1_mmq)/sizeof(int));