diff --git a/dsv4/kernels/attention/fmha.py b/dsv4/kernels/attention/fmha.py
index b4657739..94007426 100644
--- a/dsv4/kernels/attention/fmha.py
+++ b/dsv4/kernels/attention/fmha.py
@@ -268,6 +268,7 @@ 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)
             tSMEM_CPYsP = thr_smem_copy.partition_D(sP_2d)  # destination (SMEM)
 
             row_max = -Float32.inf