metal : try to improve batched decoding

2023-12-01 21:47:42 +02:00 · 2023-12-01 21:47:42 +02:00 · 3cb1c348b3
commit 3cb1c348b3
parent 3e68df8616
1 changed files with 1 additions and 1 deletions
--- a/ggml-metal.m
+++ b/ggml-metal.m
@ -1083,7 +1083,7 @@ void ggml_metal_graph_compute(

                            // find the break-even point where the matrix-matrix kernel becomes more efficient compared
                            // to the matrix-vector kernel
-                            int ne11_mm_min = 1;
+                            int ne11_mm_min = src0t == GGML_TYPE_F16 ? 1 : 16;

 #if 0
                            // the numbers below are measured on M2 Ultra for 7B and 13B models