From 1ddc262558fe736f4e046dc6d203eefbf0bbe5d8 Mon Sep 17 00:00:00 2001
From: Pascal <admin@serveurperso.com>
Date: Thu, 15 Oct 2026 06:38:12 +0200
Subject: [PATCH] graph: keep the graph non empty without embeddings (#30522)

An embedding-only graph has no logits, and build_pooling() returns
early when embeddings are off, so since the models no longer expand
t_embd the graph ends up empty and its inputs are never allocated.
Expand t_embd in that case.
---
 src/llama-graph.cpp | 4 ++++
 1 file changed, 4 insertions(+)

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;
     }