diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index f69fe63f8..6cadda971 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -3978,6 +3978,10 @@ void llm_graph_context::build_pooling(
         ggml_tensor * cls_out_b,
         ggml_tensor * cls_norm) const {
     if (!cparams.embeddings) {
+        // a graph without logits still computes t_embd, so that it is never empty
+        if (res->t_logits == nullptr && res->t_embd != nullptr) {
+            ggml_build_forward_expand(gf, res->t_embd);
+        }
         return;
     }