D1.5: Use 2D sC_epi layout from c_smem_s for TMA partition

dsv4/kernels/attention/fmha.py

@@ -490,15 +490,14 @@ class FmhaKernel:
 
             # Step 6: TMA store SMEM → GMEM
             # The normalized O is now in sC (written by the correction epilog).
-            # Use the same TMA store pattern as the CUTLASS FMHA reference.
-            # Partition sC and gC for the bulk TMA copy.
-            # Get 2D views of sC and gC for TMA.
-            sC_epi = sC  # sC already has c_smem_s layout, TMA can handle it
-            gC_epi = tCgC  # Use the pre-partitioned GMEM tensor
+            # The tma_c was created with (c, epi_s, epi_tile) where epi_s = select(c_smem_s, mode=[0,1]).
+            # We need to partition sC and the GMEM output for the TMA copy.
+            sC_epi_layout = cute.select(self.c_smem_s, mode=[0, 1])  # 2D SMEM layout
+            sC_epi = cute.make_tensor(sC.iterator, sC_epi_layout)  # 2D view of sC
+            gC_epi = cute.local_tile(mC, epi_tile, (Int32(0), Int32(0)))  # 2D GMEM tile
             bSG_sC, bSG_gC = cpasync.tma_partition(
                 tma_c, 0, cute.make_layout(1),
-                cute.group_modes(sC, 0, 2),
-                cute.group_modes(gC_epi, 0, 2),
+                sC_epi, gC_epi,
             )
             # One TMA store for the full output tile
             if warp_idx == self.epilogue_warp_id[0]: