From: Georgi Gerganov <redacted>
Date: Mon, 25 Nov 2024 19:49:31 +0000 (+0200)
Subject: metal : enable mat-vec kernels for bs <= 4 (llama/10491)
X-Git-Tag: upstream/0.0.1642~139
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=f1888c9e7653b206599377751d205131dd397444;p=pkg%2Fggml%2Fsources%2Fggml

metal : enable mat-vec kernels for bs <= 4 (llama/10491)
---

diff --git a/src/ggml-metal/ggml-metal.m b/src/ggml-metal/ggml-metal.m
index ce7afc61..ae29b8b2 100644
--- a/src/ggml-metal/ggml-metal.m
+++ b/src/ggml-metal/ggml-metal.m
@@ -1935,7 +1935,7 @@ static void ggml_metal_encode_node(
 
                 // 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 = 4;
 
 #if 0
                 // the numbers below are measured on M2 Ultra for 7B and 13B models