diff --git a/src/llm/inference_engine_enhanced.cpp b/src/llm/inference_engine_enhanced.cpp
index fb7d2be977..d03ef46289 100644
--- a/src/llm/inference_engine_enhanced.cpp
+++ b/src/llm/inference_engine_enhanced.cpp
@@ -2257,6 +2257,9 @@ bool InferenceEngineEnhanced::trySpeculativeGeneration(
             supportsTokenizerlessNativeDraftTokens(draft_plugin);
 
         if (!tok_fn_copy && !allow_native_draft_tokens) {
+            const auto draft_info = draft_plugin ? draft_plugin->getModelInfo() : std::nullopt;
+            const std::string draft_model_id = (draft_info && !draft_info->model_id.empty())
+                ? draft_info->model_id : "(unknown draft model)";
             spdlog::info("Speculative draft model '{}' has no tokenizer bridge or "
                          "known native draft-token implementation; falling back "
                          "to target generation",
diff --git a/src/llm/llama_wrapper.cpp b/src/llm/llama_wrapper.cpp
index d18aa4e41e..ff2fa2510b 100644
--- a/src/llm/llama_wrapper.cpp
+++ b/src/llm/llama_wrapper.cpp
@@ -1766,7 +1766,7 @@ std::vector<std::vector<float>> LlamaWrapper::computeTargetLogitsForTokens(
                 "Draft token out of target vocabulary range");
         }
 
-        const auto draft_token = static_cast<llama_token>(token_id);
+        auto draft_token = static_cast<llama_token>(token_id);
         const auto batch = llama_batch_get_one(&draft_token, 1);
         if (llama_decode(context_handle, batch) != 0) {
             throw std::runtime_error(