diff --git a/ggml-alloc.c b/ggml-alloc.c
index 3321f05e2..34eba3f83 100644
--- a/ggml-alloc.c
+++ b/ggml-alloc.c
@@ -386,7 +386,7 @@ static void init_view(struct ggml_allocr * alloc, struct ggml_tensor * view) {
 
     // FIXME: the view should be initialized by the owning buffer, but currently this breaks the CUDA backend
     // due to the ggml_tensor_extra_gpu ring buffer overwriting the KV cache extras
-    assert(ggml_allocr_is_measure(alloc) || view->buffer->backend == alloc->buffer->backend);
+    assert(ggml_allocr_is_measure(alloc) || !view->buffer || view->buffer->backend == alloc->buffer->backend);
     ggml_backend_buffer_init_tensor(alloc->buffer, view);
 }