diff --git a/kernel_perf_agent/kernel_opt/profiler/ncu_profiler.py b/kernel_perf_agent/kernel_opt/profiler/ncu_profiler.py
index 350c1e1..69a3a69 100644
--- a/kernel_perf_agent/kernel_opt/profiler/ncu_profiler.py
+++ b/kernel_perf_agent/kernel_opt/profiler/ncu_profiler.py
@@ -361,7 +361,10 @@ def load_ncu_metrics(
 
     # Drop the units row (first row often contains units like "%", "inst", etc.)
     if len(sub) > 0:
-        first_row_str = sub.iloc[0].astype(str).str.lower()
+        # ``.str.lower()`` propagates NaN/pd.NA cells back to float NaN, which
+        # would crash the ``tok in x`` substring check below.  ``fillna("")``
+        # makes the detection NaN-safe without changing the matching logic.
+        first_row_str = sub.iloc[0].astype(str).str.lower().fillna("")
         unit_tokens = ("%", "inst", "cycle", "block", "register", "register/thread")
         if first_row_str.apply(lambda x: any(tok in x for tok in unit_tokens)).any():
             sub = sub.iloc[1:].reset_index(drop=True)