try reduce pipeline parallelism in order to reduce compute buffer sizes

2026-09-19 17:25:07 +02:00 · 2025-12-11 14:30:38 +08:00
parent 798473d867
commit b7428048fc
2 changed files with 3 additions and 2 deletions
@@ -662,7 +662,8 @@ static bool ggml_is_view_op(enum ggml_op op) {
 #endif

 #ifndef GGML_SCHED_MAX_COPIES
-#define GGML_SCHED_MAX_COPIES 4
+//kcpp reduced from 4 to 2 to try make buffer sizes smaller on multigpu
+#define GGML_SCHED_MAX_COPIES 2
 #endif

 struct ggml_backend_sched_split {