fix(openvino): explicit ov::Tensor frees in ggml_backend_openvino_free

Author: thedanhoffman

ggml/src/ggml-openvino/ggml-openvino.cpp

@@ -603,6 +603,15 @@ bool ggml_backend_buft_is_openvino_host(ggml_backend_buffer_type_t buft) {
 
 static void ggml_backend_openvino_free(ggml_backend_t backend) {
     ggml_backend_openvino_context * ctx = (ggml_backend_openvino_context *) backend->context;
+
+    if (ctx->runtime_context) {
+        auto r_ctx = std::static_pointer_cast<ov_runtime_context>(ctx->runtime_context);
+        r_ctx->backend_count--;
+        if (r_ctx->backend_count == 0) {
+            r_ctx->clear_caches();
+        }
+    }
+
     delete ctx;
     delete backend;
 }
@@ -672,6 +681,7 @@ GGML_BACKEND_API ggml_backend_t ggml_backend_openvino_init(int device) {
     std::shared_ptr<ov_runtime_context> r_ctx = std::static_pointer_cast<ov_runtime_context>(ctx->runtime_context);
     r_ctx->device = ggml_openvino_get_device_name();
     r_ctx->stateful = is_stateful_enabled() && !ggml_openvino_is_npu();
+    r_ctx->backend_count++;
 
     ggml_backend_t openvino_backend = new ggml_backend{
         /* .guid      = */ ggml_backend_openvino_guid(),

ggml/src/ggml-openvino/utils.h

@@ -59,11 +59,21 @@ struct ov_runtime_context {
     //      Simultanous stateful inference request support to be added.
     size_t stateful_kv_size;
     std::map<std::string, std::string> kv_state_input_name_map;
+    int backend_count;
 
     ov_runtime_context() :
         device("CPU"),
         stateful(false),
-        stateful_kv_size(0) {}
+        stateful_kv_size(0),
+        backend_count(0) {}
+
+    void clear_caches() {
+        decoder_cache.clear();
+        infer_request_cache.clear();
+        infer_request_cache_prefill.clear();
+        ov_input_names_cache.clear();
+        ov_output_names_cache.clear();
+    }
 };
 
 enum ggml_status ov_graph_compute(struct ggml_cgraph * cgraph, ggml_backend_t backend);