diff --git a/tests/unit/test_qk_direct.cu b/tests/unit/test_qk_direct.cu
index b205e370..fad1945a 100644
--- a/tests/unit/test_qk_direct.cu
+++ b/tests/unit/test_qk_direct.cu
@@ -118,7 +118,7 @@ test_qk_direct_kernel(float* __restrict__ out_s,
 
 int main() {
     printf("Direct QK Test (HD=%d, SK=%d)\n", HD, SK);
-    const int T = 4;
+    const int T = 1;
 
     bf16_t* h_q = (bf16_t*)calloc(128 * HD, sizeof(bf16_t));
     bf16_t* h_k = (bf16_t*)calloc(SK * HD, sizeof(bf16_t));