diff --git a/dsv4/kernels/attention/fmha.py b/dsv4/kernels/attention/fmha.py
index e1b66b52..a3aa5eb5 100644
--- a/dsv4/kernels/attention/fmha.py
+++ b/dsv4/kernels/attention/fmha.py
@@ -273,9 +273,12 @@ class FmhaKernel:
             thr_smem_copy = tiled_smem_copy.get_slice(sfw_idx)
             # Debug sP shape
             print(f"[SMEM-P PROPER] sP shape: {cute.shape(sP)} rank: {len(cute.shape(sP))}")
-            # Try without group_modes first
-            tSMEM_CPYsP = thr_smem_copy.partition_D(sP)  # destination (SMEM)
+            # Try grouping ALL modes (0-4) to flatten
+            sP_flat = cute.group_modes(sP, 0, 4)
+            print(f"[SMEM-P PROPER] sP_flat shape after group_modes(0,4): {cute.shape(sP_flat)}")
+            tSMEM_CPYsP = thr_smem_copy.partition_D(sP_flat)  # destination (SMEM)
             print(f"[SMEM-P PROPER] After partition_D: tSMEM_CPYsP layout: {tSMEM_CPYsP.layout}")
+            print(f"[SMEM-P PROPER] After partition_D: tSMEM_CPYsP shape: {cute.shape(tSMEM_CPYsP)} rank: {len(cute.shape(tSMEM_CPYsP))}")
 
             row_max = -Float32.inf
             row_sum = Float32(0.0)
@@ -356,8 +359,12 @@ class FmhaKernel:
                     
                     # Partition source with QK layout
                     print(f"[SMEM-P PROPER] Before partition_S: rP_bf16_qk shape: {cute.shape(rP_bf16_qk)} layout: {rP_bf16_qk.layout}")
-                    tSMEM_CPYrP_qk = thr_smem_copy.partition_S(rP_bf16_qk)
+                    # Try flattening source too (rP_bf16_qk has 3 modes: group_modes(0,3))
+                    rP_bf16_qk_flat = cute.group_modes(rP_bf16_qk, 0, 3)
+                    print(f"[SMEM-P PROPER] rP_bf16_qk_flat shape after group_modes(0,3): {cute.shape(rP_bf16_qk_flat)}")
+                    tSMEM_CPYrP_qk = thr_smem_copy.partition_S(rP_bf16_qk_flat)
                     print(f"[SMEM-P PROPER] After partition_S: tSMEM_CPYrP_qk layout: {tSMEM_CPYrP_qk.layout}")
+                    print(f"[SMEM-P PROPER] After partition_S: tSMEM_CPYrP_qk shape: {cute.shape(tSMEM_CPYrP_qk)} rank: {len(cute.shape(tSMEM_CPYrP_qk))}")
                     
                     # Debug shapes
                     print(f"[SMEM-P PROPER] rP_bf16 shape: {cute.shape(rP_bf16)}, layout: TMEM")