diff --git a/sdk/plugins/llama_cpp/src/vlm.cpp b/sdk/plugins/llama_cpp/src/vlm.cpp
index 7a344db86..98431b313 100644
--- a/sdk/plugins/llama_cpp/src/vlm.cpp
+++ b/sdk/plugins/llama_cpp/src/vlm.cpp
@@ -89,18 +89,19 @@ int32_t LlamaVlm::create(const geniex_VlmCreateInput* input) {
 
     // Initialize vision context if mmproj_path provided
     if (input->mmproj_path) {
-        ggml_backend_dev_t htp_device = device == Device::NPU ? ggml_backend_dev_by_name("HTP0") : nullptr;
-
         mtmd_context_params mparams = mtmd_context_params_default();
-        mparams.use_gpu             = device == Device::GPU || htp_device != nullptr;
-        mparams.device              = htp_device;
-        mparams.print_timings       = false;
-        mparams.n_threads           = 4;
+        mparams.use_gpu             = false;
+        if (!selection->empty()) {
+            mparams.use_gpu = true;
+            mparams.device  = selection->data()[0];
+        }
+        mparams.print_timings = false;
+        mparams.n_threads     = 4;
         // Zack TODO: elegant fix this error:  no member named 'verbosity' in 'mtmd_context_params'
         // mparams.verbosity           = GGML_LOG_LEVEL_ERROR;
 
         this->ctx_vision = mtmd_init_from_file(input->mmproj_path, this->model, mparams);
-        if (!this->ctx_vision && htp_device) {
+        if (!this->ctx_vision && !selection->empty()) {
             GENIEX_LOG_WARN("mtmd failed to initialize the vision encoder on HTP; falling back to CPU");
             mparams.use_gpu  = false;
             mparams.device   = nullptr;