ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-22 14:14:32 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix division by zero bug (#349)

Browse Source 
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-04-26 09:19:43 +02:00
committed by GitHub
parent 86be28d5bd
commit 815307d3bd
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.c
View File

@@ -21790,7 +21790,7 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa
const struct ggml_tensor * k = node->src[1];
if (q->ne[1] == 1 && q->ne[3] == 1 && q->ne[2]/k->ne[2] > 1 && n_tasks > 1 && k->ne[1]/32 > 1) {
if (k->ne[2] > 1) {
int nk = 32 * (k->ne[2]*k->ne[1]/(32*n_tasks));
int nk = MAX(1, 32 * (k->ne[2]*k->ne[1]/(32*n_tasks)));
int nstep_k = k->ne[2]*k->ne[1]/nk;
size_t result_size = (Dv + 16)*q->ne[2]/k->ne[2]*sizeof(float);
size_t size = nstep_k*result_size;

2
ggml/src/iqk/iqk_flash_attn.cpp
View File

@@ -154,7 +154,7 @@ extern "C" IQK_API bool iqk_flash_attn_noalibi(int type_q, int type_mask, float
}
if (neq3 == 1 && rk2 > 1 && rk2 == rv2 && neq1 == 1 && nth >= 1 && nek2*nek1 >= 32*nth) {
int nk = 32 * (nek2*nek1/(32*nth));
int nk = std::max(1, 32 * (nek2*nek1/(32*nth)));
int nkk = (nek1 + nk - 1)/nk;
int nstep_k = nek2*nkk;
auto result_size = (Dv + 16)*rk2*sizeof(float);