Don't put kvcache on GPU in stateful mode

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

@@ -140,7 +140,7 @@ static enum ggml_status ggml_backend_openvino_buffer_init_tensor(ggml_backend_bu
 
     // Put kvcache on device memory for GPU (NPU memory is too small even for kvcache)
     if (buffer->usage == GGML_BACKEND_BUFFER_USAGE_ANY && strncmp(tensor->name, "cache_", 6) == 0 && !ctx->is_remote &&
-        ggml_openvino_get_device_name() == "GPU") {
+        ggml_openvino_get_device_name() == "GPU" && !getenv("GGML_OPENVINO_STATEFUL_EXECUTION")) {
         GGML_ASSERT(ctx->tensor_extras.empty());
         auto device = ctx->device;
         auto size = ctx->size;