diff --git a/dsv4/kernels/attention/fmha.py b/dsv4/kernels/attention/fmha.py
index 4e371c5a..1c3f4449 100644
--- a/dsv4/kernels/attention/fmha.py
+++ b/dsv4/kernels/attention/fmha.py
@@ -268,7 +268,9 @@ class FmhaKernel:
             )
             tiled_smem_copy = cute.make_tiled_copy_C(smem_copy_atom, qk_mma)
             thr_smem_copy = tiled_smem_copy.get_slice(sfw_idx)
-            sP_2d = cute.group_modes(sP, 0, 3)  # flatten to 2D for copy
+            print(f"[SMEM-P DEBUG] sP shape: {cute.shape(sP)}")
+                    sP_2d = cute.group_modes(sP, 0, 3)
+                    print(f"[SMEM-P DEBUG] sP_2d shape: {cute.shape(sP_2d)}")  # flatten to 2D for copy
             tSMEM_CPYsP = thr_smem_copy.partition_D(sP_2d)  # destination (SMEM)
 
             row_max = -Float32.inf