diff --git a/dsv4/kernels/attention/fmha.py b/dsv4/kernels/attention/fmha.py
index 67bd85e4..9c12fed9 100644
--- a/dsv4/kernels/attention/fmha.py
+++ b/dsv4/kernels/attention/fmha.py
@@ -119,14 +119,19 @@ class FmhaKernel:
             p_s = cute.slice_(self.p_smem_s,(None,None,None,0))
             gP = torch.zeros(128, self.s_k, dtype=torch.bfloat16, device='cuda')
             mgP = ct.from_dlpack(gP).mark_layout_dynamic(leading_dim=ct.get_leading_dim(gP))
-            tma_p = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileG2SOp(), mgP, p_s, self.qk_mma_tiler)
+            tma_p,mgP = cute.nvgpu.make_tiled_tma_atom_A(
+                utils.sm100.cluster_shape_to_tma_atom_A(self.cluster_shape_mn, pv_mma.thr_id),
+                mgP, p_s, self.qk_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape
+            )
         else:
             # Dummy gP and tma_p (not used, dead-code-eliminated)
             gP = torch.zeros(128, self.s_k, dtype=torch.bfloat16, device='cuda')
             mgP = ct.from_dlpack(gP).mark_layout_dynamic(leading_dim=ct.get_leading_dim(gP))
-            # Create a dummy TMA using the V SMEM layout (same structure, unused)
             v_s = cute.slice_(self.v_smem_s,(None,None,None,0))
-            tma_p = cpasync.make_tiled_tma_atom(cpasync.CopyBulkTensorTileG2SOp(), mgP, v_s, self.qk_mma_tiler)
+            tma_p,mgP = cute.nvgpu.make_tiled_tma_atom_A(
+                utils.sm100.cluster_shape_to_tma_atom_A(self.cluster_shape_mn, pv_mma.thr_id),
+                mgP, v_s, self.qk_mma_tiler, pv_mma, self.cluster_layout_vmnk.shape
+            )
         # Always create a valid mLSE tensor for the kernel.
         # CuTeDSL doesn't support None parameters in @cute.kernel.
         # For normalize=True, mLSE is unused (dead-code-eliminated by compiler).