From 6d5ac3c8d5aae1831edf17c2d5cda11abe025bb3 Mon Sep 17 00:00:00 2001
From: Yonghye Kwon <developer.0hye@gmail.com>
Date: Thu, 29 Oct 2026 14:04:28 +0900
Subject: [PATCH] tests : fix comp_padding CUDA graph reservation (#31344)

Assisted-by: Codex
---
 tests/llama-context/cases-kv.cpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/tests/llama-context/cases-kv.cpp b/tests/llama-context/cases-kv.cpp
index 708b6cbc4..d5d5e1f09 100644
--- a/tests/llama-context/cases-kv.cpp
+++ b/tests/llama-context/cases-kv.cpp
@@ -782,7 +782,8 @@ struct case_comp_padding : ctx_test {
             ctx_opts opts;
             opts.n_ctx      = 256*n_seq;
             opts.n_batch    = 64;
-            opts.n_ubatch   = 64;
+            // Reserve the DSV4.1 small-batch attention graph when n_seq == 1.
+            opts.n_ubatch   = 32;
             opts.n_seq_max  = n_seq;
             opts.kv_unified = n_seq == 1;