diff --git a/tests/fmha_v3_stage_c_example9.py b/tests/fmha_v3_stage_c_example9.py
index d2f8a54b..ba1e6fa2 100644
--- a/tests/fmha_v3_stage_c_example9.py
+++ b/tests/fmha_v3_stage_c_example9.py
@@ -323,6 +323,7 @@ class FmhaV3StageCMulti:
             # Per-tile softmax loop with online rescale.
             for kt in range(n_kv_tiles):
                 si_handle = s_cons.wait_and_advance()
+                cute.printf("SOFTMAX kt=%d row_max_before=%f row_sum_before=%f\n", kt, row_max, row_sum)
 
                 # Load S[kt]
                 tTMEM_LOADrS = cute.make_rmem_tensor(tTMEM_LOADcS.shape, self.qk_acc_dtype)
@@ -397,6 +398,7 @@ class FmhaV3StageCMulti:
                     cute.arch.fence_view_async_tmem_store()
 
                 si_handle.release()
+                cute.printf("SOFTMAX kt=%d row_max_after=%f row_sum_after=%f\n", kt, row_max, row_sum)
                 softmax_done_bar.arrive()
 
             # Wait for MMA's PV[N-1] to commit before reading O for normalize.