diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 6419d739bd8a2..b928e9e16ead8 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -1376,7 +1376,7 @@ ggml_tensor * llm_graph_context::build_attn(
 
     // [TAG_NO_CACHE_PAD]
     // TODO: if ubatch.equal_seqs() == true, we can split the three tensors below into ubatch.n_seqs_unq streams
-    assert(!ubatch.equal_seqs());
+    assert(!ubatch.equal_seqs() || (k_cur->ne[3] == 1 && k_cur->ne[3] == ubatch.n_seqs_unq));
 
     ggml_tensor * q = q_cur;
     ggml_tensor * k = k_cur;