diff --git a/tests/unit/test_fmha_sm100_standalone.cu b/tests/unit/test_fmha_sm100_standalone.cu
index 0487b1a6..f5b92d38 100644
--- a/tests/unit/test_fmha_sm100_standalone.cu
+++ b/tests/unit/test_fmha_sm100_standalone.cu
@@ -101,8 +101,10 @@ int test_kernel(const char* name, int HD_val, int sk, float scale,
     for (int i = 0; i < B*H*HD_val; i++) ho_gpu[i] = bf16_to_f32_cpu(hob[i]);
 
     float cos = cosine_sim(ho_gpu, ho_ref, B*H*HD_val);
+    float max_diff = 0;
+    for(int i=0;i<B*H*HD_val;i++) max_diff = fmaxf(max_diff, fabsf(ho_gpu[i]-ho_ref[i]));
     int pass = cos > 0.999f;
-    printf("  %s hd=%d s_k=%d: cos %.6f %s\n", name, HD_val, sk, cos, pass ? "✅" : "❌");
+    printf("  %s hd=%d s_k=%d: cos %.6f max_diff %.6f %s\n", name, HD_val, sk, cos, max_diff, pass ? "✅" : "❌");
 
     if (!pass) {
         printf("    GPU[:4] = %.6f %.6f %.6f %.6f\n", ho_gpu[0], ho_gpu[1], ho_gpu[2], ho_gpu[3]);