fix nccl dist train bug

7 years ago · d76ebd7853
parent 88fa9c2ec7
commit d76ebd7853
1 changed files with 4 additions and 0 deletions
--- a/paddle/fluid/operators/gen_nccl_id_op.cc
+++ b/paddle/fluid/operators/gen_nccl_id_op.cc
@ -67,6 +67,10 @@ class GenNCCLIdOp : public framework::OperatorBase {
      client->AsyncSendVar(ep, dev_ctx, *scope, NCCL_ID_VARNAME);
    }
    client->Wait();
+    for (auto& ep : endpoint_list) {
+      client->AsyncSendBatchBarrier(ep);
+    }
+    client->Wait();
    VLOG(3) << "sending completed...";
  }