From: Georgi Gerganov <redacted>
Date: Tue, 26 Aug 2025 14:45:17 +0000 (+0300)
Subject: graph : fix assert in memory-less build_attn (#15590)
X-Git-Tag: upstream/0.0.6527~238
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=0373486dbc0dccbdcb3b5fdd65759d88cec06196;p=pkg%2Fggml%2Fsources%2Fllama.cpp

graph : fix assert in memory-less build_attn (#15590)

ggml-ci
---

diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 6419d739..b928e9e1 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;