feat: add multigpu support for external native models

src/backends/torch/native/native_wrapper.h

@@ -29,6 +29,35 @@
 
 #include "native_net.h"
 
+namespace std
+{
+  // from
+  // https://stackoverflow.com/questions/47496358/c-lambdas-how-to-capture-variadic-parameter-pack-from-the-upper-scope
+  // see also:
+  // https://en.cppreference.com/w/cpp/experimental/apply
+  namespace detail
+  {
+    template <class F, class Tuple, std::size_t... I>
+    constexpr decltype(auto) apply_impl(F &&f, Tuple &&t,
+                                        std::index_sequence<I...>)
+    {
+      // return std::invoke(std::forward<F>(f),
+      //                   std::get<I>(std::forward<Tuple>(t))...);
+      // Note: std::invoke is a C++17 feature
+      return std::forward<F>(f)(std::get<I>(static_cast<Tuple &&>(t))...);
+    }
+  } // namespace detail
+
+  template <class F, class Tuple>
+  constexpr decltype(auto) apply(F &&f, Tuple &&t)
+  {
+    return detail::apply_impl(
+        std::forward<F>(f), std::forward<Tuple>(t),
+        std::make_index_sequence<
+            std::tuple_size<std::decay_t<Tuple>>::value>{});
+  }
+}
+
 namespace dd
 {
   template <typename TModule>
@@ -41,6 +70,14 @@ namespace dd
     template <typename... Args>
     NativeModuleWrapper(Args &&... args) : _module(args...)
     {
+      _clone_function
+          = [args = std::make_tuple(std::forward<Args>(args)...)]() {
+              return std::apply(
+                  [](auto &&... args) {
+                    return new NativeModuleWrapper<TModule>(args...);
+                  },
+                  std::move(args));
+            };
       this->register_module("wrapped", _module);
     }
 
@@ -86,6 +123,9 @@ namespace dd
       throw MLLibInternalException(
           "NativeModuleWrapper::loss not implemented");
     }
+
+  private:
+    std::function<NativeModuleWrapper<TModule> *()> _clone_function;
   };
 }