diff --git a/ggml-qnn.cpp b/ggml-qnn.cpp
index 072003e1d76b8..84549ddbf4353 100644
--- a/ggml-qnn.cpp
+++ b/ggml-qnn.cpp
@@ -3074,12 +3074,15 @@ ggml_backend_buffer_type_t ggml_backend_qnn_buffer_type(size_t device) {
         return nullptr;
     }
 
+    static ggml_backend_qnn_buffer_type_context ggml_backend_qnn_buffer_type_contexts[GGML_QNN_MAX_DEVICES];
     static ggml_backend_buffer_type ggml_backend_qnn_buffer_types[GGML_QNN_MAX_DEVICES];
 
     static bool ggml_backend_qnn_buffer_type_initialized = false;
 
     if (!ggml_backend_qnn_buffer_type_initialized) {
-        for (int i = 0; i < GGML_QNN_MAX_DEVICES; i++) {
+        for (size_t i = 0; i < GGML_QNN_MAX_DEVICES; i++) {
+            auto &context = ggml_backend_qnn_buffer_type_contexts[i];
+            context = { i, std::string(GGML_QNN_NAME) + std::to_string(i) };
             ggml_backend_qnn_buffer_types[i] = {
                 /* .iface   = */ {
                     /* .get_name         = */ ggml_backend_qnn_buffer_type_name,
@@ -3090,7 +3093,7 @@ ggml_backend_buffer_type_t ggml_backend_qnn_buffer_type(size_t device) {
                     /* .supports_backend = */ ggml_backend_qnn_buffer_type_supports_backend,
                     /* .is_host          = */ ggml_backend_qnn_buffer_is_host
                 },
-                /* .context = */ new ggml_backend_qnn_buffer_type_context { device, GGML_QNN_NAME + std::to_string(device) },
+                /* .context = */ &context,
             };
         }
         ggml_backend_qnn_buffer_type_initialized = true;