double n_batch size when pipeline parallel is enabled, keep u_batch the same

2026-09-19 09:15:18 +02:00 · 2026-03-21 11:22:10 +08:00
parent 3113e3a643
commit 07327b6c10
2 changed files with 7 additions and 3 deletions
@@ -670,7 +670,7 @@ static bool ggml_is_view_op(enum ggml_op op) {

 #ifndef GGML_SCHED_MAX_COPIES
 //kcpp can reduce this if you want to try make buffer sizes smaller on multigpu
-#define GGML_SCHED_MAX_COPIES 4
+#define GGML_SCHED_MAX_COPIES 2
 #endif

 struct ggml_backend_sched_split {