diff --git a/src/llama-model.h b/src/llama-model.h
index 1fcbce16..2b34de95 100644
--- a/src/llama-model.h
+++ b/src/llama-model.h
@@ -422,7 +422,7 @@ struct llama_model {
 
     size_t max_nodes(int n_tokens) const {
         auto n_tensors = tensors_by_name.size();
-        if (split_mode == LLAMA_SPLIT_MODE_GRAPH) n_tensors *= devices.size();
+        if (split_mode == LLAMA_SPLIT_MODE_GRAPH && !devices.empty()) n_tensors *= devices.size();
         if (arch == LLM_ARCH_QWEN3NEXT || arch == LLM_ARCH_QWEN35MOE) {
             return std::max<size_t>(n_tokens * 40, 32u * n_tensors);
         }