From 3ad45fc3a839e5543cb42d2230aa7f3fbb12b66a Mon Sep 17 00:00:00 2001
From: slaren <slarengh@gmail.com>
Date: Sun, 24 Dec 2023 12:26:04 +0100
Subject: [PATCH] ensure allocations are always aligned

---
 ggml-cuda.cu | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 01a7bc82e..b5679c42f 100644
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -6657,6 +6657,10 @@ static void * ggml_cuda_pool_malloc_vmm(size_t size, size_t * actual_size) {
     int id;
     CUDA_CHECK(cudaGetDevice(&id));
 
+    // round up the allocation size to the alignment to ensure that all allocations are aligned for all data types
+    const size_t alignment = 128;
+    size = alignment * ((size + alignment - 1) / alignment);
+
     size_t avail = g_cuda_pool_size[id] - g_cuda_pool_used[id];
 
     if (size > avail) {