From 716e960a808e525ecec45fc701cb7fbe607de06a Mon Sep 17 00:00:00 2001
From: Georgi Gerganov <ggerganov@gmail.com>
Date: Wed, 3 Apr 2024 15:25:47 +0300
Subject: [PATCH] metal : pad shared memory to 16 bytes

---
 ggml-metal.m | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/ggml-metal.m b/ggml-metal.m
index 51a5fab3a..419d8b9e5 100644
--- a/ggml-metal.m
+++ b/ggml-metal.m
@@ -2417,7 +2417,10 @@ static enum ggml_status ggml_metal_graph_compute(
                             ne00_padded *= 2;
                         }
 
-                        const int mem_size = ne00_padded*sizeof(int32_t);
+                        // Metal kernels require the buffer size to be multiple of 16 bytes
+                        // https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/1443142-setthreadgroupmemorylength
+                        const int mem_size = GGML_PAD(ne00_padded*sizeof(int32_t), 16);
+
                         id<MTLComputePipelineState> pipeline = nil;
 
                         switch (order) {