Minor improvements.

cpp/src/interop/to_arrow_device.cu

@@ -143,7 +143,7 @@ int construct_decimals(cudf::column_view input,
   NANOARROW_RETURN_NOT_OK(initialize_array(tmp.get(), NANOARROW_TYPE_DECIMAL128, input));
 
   auto buf = detail::convert_decimals_to_decimal128<DeviceType>(input, stream, mr);
-  // Synchronize stream here as convert_decimals_to_decimal128 does not sync.
+  // Synchronize stream here to ensure the decimal128 buffer is ready.
   stream.synchronize();
   NANOARROW_RETURN_NOT_OK(set_buffer(std::move(buf), fixed_width_data_buffer_idx, tmp.get()));
 

cpp/src/io/parquet/writer_impl.cu

@@ -1674,8 +1674,6 @@ std::vector<column_view> convert_decimal_columns_and_metadata(
                 column.offset(),
                 converted_children};
     }
-    // Synchronize stream here to ensure all decimal128 buffers are ready.
-    stream.synchronize();
   };
 
   // Vector of converted column views
@@ -1689,6 +1687,9 @@ std::vector<column_view> convert_decimal_columns_and_metadata(
     std::back_inserter(converted_column_views),
     [&](auto elem) { return convert_column(thrust::get<0>(elem), thrust::get<1>(elem)); });
 
+  // Synchronize stream here to ensure all decimal128 buffers are ready.
+  stream.synchronize();
+
   return converted_column_views;
 }