1 files changed, 34 insertions, 0 deletions
diff --git a/ggml-vulkan.h b/ggml-vulkan.h
new file mode 100644
index 00000000..eb8a148e
--- /dev/null
+++ b/ggml-vulkan.h
@@ -0,0 +1,34 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_VK_NAME "Vulkan"
+
+GGML_API void ggml_vk_init(void);
+
+GGML_API void ggml_vk_preallocate_buffers_graph(struct ggml_tensor * node);
+GGML_API void ggml_vk_preallocate_buffers(void);
+GGML_API void ggml_vk_build_graph(struct ggml_tensor * node, bool last_node);
+GGML_API bool ggml_vk_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
+#ifdef GGML_VULKAN_CHECK_RESULTS
+void ggml_vk_check_results_1(struct ggml_compute_params * params, struct ggml_tensor * tensor);
+#endif
+GGML_API void ggml_vk_graph_cleanup(void);
+
+// backend API
+GGML_API GGML_CALL ggml_backend_t ggml_backend_vk_init(void);
+
+GGML_API GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend);
+
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(void);
+// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
+GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
+
+#ifdef  __cplusplus
+}
+#endif