nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 17 Packages Projects Releases Wiki Activity

CUDA: only allocate FA tmp buffer if needed (#18564)

Browse Source
This commit is contained in:
Johannes Gäßler
2026-01-03 13:55:53 +01:00
committed by GitHub
parent 9dba9f5352
commit 0f2e42ca1d
1 changed files with 3 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/fattn-common.cuh
+3 -1
View File
@@ -918,7 +918,9 @@ void launch_fattn(
blocks_num.y = 1; blocks_num.y = 1;
blocks_num.z = 1; blocks_num.z = 1;
dst_tmp_meta.alloc(((size_t) blocks_num.x) * ncols * (2 + DV/2)); if (ntiles_total % blocks_num.x != 0) { // Fixup is only needed if the SMs work on fractional tiles.
dst_tmp_meta.alloc((size_t(blocks_num.x) * ncols * (2 + DV/2)));
}
} else { } else {
const int ntiles_KQ = (K->ne[1] + nbatch_fa - 1) / nbatch_fa; // Max. number of parallel blocks limited by tensor size. const int ntiles_KQ = (K->ne[1] + nbatch_fa - 1) / nbatch_fa; // Max. number of parallel blocks limited by tensor size.