WIP

src/llama-impl.h

@@ -224,3 +224,8 @@ struct gguf_context;
 std::string gguf_kv_to_str(const gguf_context * ctx_gguf, int i);
 
 ggml_backend_buffer_type_t llama_default_buffer_type_cpu(bool host_buffer);
+
+struct llama_split_tensor {
+    std::vector<ggml_tensor *> tensor_splits;
+    ggml_split_tensor_t        ggml;
+};