diff --git a/dsv4/kernels/attention/fmha.py b/dsv4/kernels/attention/fmha.py
index a3aa5eb5..952bd54d 100644
--- a/dsv4/kernels/attention/fmha.py
+++ b/dsv4/kernels/attention/fmha.py
@@ -273,10 +273,8 @@ 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 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)
+            # Don't flatten - keep original sP
+            tSMEM_CPYsP = thr_smem_copy.partition_D(sP)  # 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))}")
 
@@ -359,10 +357,7 @@ 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}")
-                    # 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)
+                    tSMEM_CPYrP_qk = thr_smem_copy.partition_S(rP_bf16_qk)
                     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))}")
                     
@@ -372,12 +367,20 @@ class FmhaKernel:
                     print(f"[SMEM-P PROPER] tSMEM_CPYrP_qk shape: {cute.shape(tSMEM_CPYrP_qk)} rank: {len(cute.shape(tSMEM_CPYrP_qk))}")
                     print(f"[SMEM-P PROPER] tSMEM_CPYsP shape: {cute.shape(tSMEM_CPYsP)} rank: {len(cute.shape(tSMEM_CPYsP))}")
                     
-                    # Attempt copy with correct layout
-                    try:
-                        cute.copy(tiled_smem_copy, tSMEM_CPYrP_qk, tSMEM_CPYsP)
-                        print(f"[SMEM-P PROPER] Copy succeeded with QK C-fragment layout")
-                    except Exception as e:
-                        print(f"[SMEM-P PROPER] Copy failed: {e}")
+                    # Manual copy instead of cute.copy (helpers are a trap)
+                    # Get sizes and iterate
+                    src_size = cute.size(tSMEM_CPYrP_qk)
+                    dst_size = cute.size(tSMEM_CPYsP)
+                    print(f"[SMEM-P PROPER] Manual copy: src_size={src_size}, dst_size={dst_size}")
+                    
+                    if src_size == dst_size:
+                        # Same number of elements - copy element by element
+                        for j in cutlass.range(src_size, vectorize=True):
+                            val = tSMEM_CPYrP_qk[j]
+                            tSMEM_CPYsP[j] = val
+                        print(f"[SMEM-P PROPER] Manual copy completed (src_size==dst_size)")
+                    else:
+                        print(f"[SMEM-P PROPER] Manual copy: size mismatch, using fallback stub")
                         # Fallback to stub for now
                         for j in cutlass.range(cute.size(sP), vectorize=True):
                             sP[j] = BFloat16(0.0)