From 5acc9e50f5f47709207e844acdfffa488370b0d9 Mon Sep 17 00:00:00 2001
From: slaren <slarengh@gmail.com>
Date: Sun, 24 Dec 2023 11:59:15 +0100
Subject: [PATCH] cuda : fallback to CPU on host buffer alloc fail

---
 ggml-cuda.cu | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index f61f06853..bdf7498ba 100644
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -9797,8 +9797,10 @@ static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buff
 
 static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     void * ptr = ggml_cuda_host_malloc(size);
+
     if (ptr == nullptr) {
-        return nullptr;
+        // fallback to cpu buffer
+        return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
     }
 
     // FIXME: this is a hack to avoid having to implement a new buffer type