From 146e4f0282496e1750e856c68571e7308502dea6 Mon Sep 17 00:00:00 2001
From: biondizzle <biondizzle@gmail.com>
Date: Thu, 28 May 2026 07:46:57 +0000
Subject: [PATCH] debug: print NaN positions in test

---
 tests/unit/test_fmha_sm100_standalone.cu | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/tests/unit/test_fmha_sm100_standalone.cu b/tests/unit/test_fmha_sm100_standalone.cu
index 02c0115c..9cec4420 100644
--- a/tests/unit/test_fmha_sm100_standalone.cu
+++ b/tests/unit/test_fmha_sm100_standalone.cu
@@ -104,7 +104,10 @@ int test_kernel(const char* name, int HD_val, int sk, float scale,
     float max_diff = 0;
     int nan_count = 0;
     for(int i=0;i<B*H*HD_val;i++) {
-        if(!isfinite(ho_gpu[i]) || !isfinite(ho_ref[i])) nan_count++;
+        if(!isfinite(ho_gpu[i]) || !isfinite(ho_ref[i])) {
+            if(nan_count < 8) printf("    nan at [%d]: gpu=%f ref=%f\n", i, ho_gpu[i], ho_ref[i]);
+            nan_count++;
+        }
         else max_diff = fmaxf(max_diff, fabsf(ho_gpu[i]-ho_ref[i]));
     }
     // BF16 output has ~0.1% relative error from quantization.