refine parallel_do_grad

paddle/fluid/operators/assign_op.cc

@@ -56,6 +56,7 @@ class AssignFunctor {
  private:
   void copy_tensor(const framework::LoDTensor &lod_tensor,
                    framework::LoDTensor *out) const {
+    if (lod_tensor.numel() == 0) return;
     auto &out_tensor = *out;
     TensorCopy(lod_tensor, lod_tensor.place(), dev_ctx_, &out_tensor);
     out_tensor.set_lod(lod_tensor.lod());

paddle/fluid/operators/nccl_op.cu.cc

@@ -106,6 +106,8 @@ class NCCLReduceKernel : public framework::OpKernel<T> {
     T* recvbuffer = nullptr;
     if (root == gpu_id) {
       recvbuffer = out->mutable_data<T>(ctx.GetPlace());
+    } else {
+      out->Resize(framework::make_ddim({0}));
     }
     VLOG(3) << "gpu : " << gpu_id << " invoke reduce. send " << x->numel()
             << " recv " << out->numel();

python/paddle/fluid/backward.py

@@ -248,12 +248,15 @@ def _callback_lookup_(op):
                         if o_argu in self.param_grad_names:
                             allreduce_out_name = o_argu + "__nccl_all_reduce__"
                             op_desc = _create_op_desc_(
-                                "ncclAllReduce", {
+                                "ncclReduce",
+                                {
                                     "X": [o_argu],
                                     "Communicator":
                                     ['nccl_com__do_not_change_']
-                                }, {"Out": [allreduce_out_name]},
-                                {"reduction": "ncclSum"})
+                                },
+                                {"Out": [allreduce_out_name]},
+                                {"reduction": "ncclSum",
+                                 "root": 0}, )
                             block.desc.append_op().copy_from(op_desc)
 
                             op_desc = _create_op_desc_(