v4.5 dev update. (#3153)

examples/python/CuTeDSL/blackwell/epilogue/common_dense_gemm_efc.py

@@ -573,7 +573,7 @@ class DenseGemmEFC:
             # Barriers used by the supplemental load tensor pipeline.
             c_full_mbar_ptr: cute.struct.MemRange[cutlass.Int64, self.num_c_stage]
             c_empty_mbar_ptr: cute.struct.MemRange[cutlass.Int64, self.num_c_stage]
-            tmem_dealloc_mbar_ptr: cutlass.Int64
+            tmem_dealloc_mbar: cutlass.Int64
             tmem_holding_buf: cutlass.Int32
             # (MMA, MMA_M, MMA_K, STAGE)
             sA: cute.struct.Align[
@@ -651,11 +651,11 @@ class DenseGemmEFC:
         )
         # Tensor memory dealloc barrier init
         tmem = utils.TmemAllocator(
-            storage.tmem_holding_buf,
+            storage.tmem_holding_buf.ptr,
             barrier_for_retrieve=tmem_alloc_barrier,
             allocator_warp_id=self.epilogue_warp_id[0],
             is_two_cta=self.use_2cta_instrs,
-            two_cta_tmem_dealloc_mbar_ptr=storage.tmem_dealloc_mbar_ptr,
+            two_cta_tmem_dealloc_mbar_ptr=storage.tmem_dealloc_mbar.ptr,
         )
 
         # Cluster arrive after barrier init