[Transformer][UCC] Fix async p2p ops (#1691)

* fix async ucc

* apply pep preferred way

* Revert "apply pep preferred way"

This reverts commit 84ceb08.

Author: Aidyn-A

apex/transformer/pipeline_parallel/p2p_communication.py

@@ -59,9 +59,9 @@ def _run_p2pops(
 
     need_to_sync = p2p_group.name() != default_group.name()
     reqs = []
+    ops = []
 
     if batch_p2p_comm and p2p_group.name() == "nccl":
-        ops = []
         if tensor_send_prev is not None:
             send_prev_op = torch.distributed.P2POp(
                 op=torch.distributed.isend,
@@ -140,7 +140,7 @@ def _run_p2pops(
             return (None, None, None, None, reqs)
 
         if async_comm:
-            if len(reqs) == len(ops):
+            if len(ops) == 0 or len(reqs) == len(ops):
                 tensor_send_prev_req = None if tensor_send_prev is None else reqs.pop(0)
                 tensor_recv_prev_req = None if tensor_recv_prev is None else reqs.pop(0)
                 tensor_send_next_req = None if tensor_send_next is None else reqs.pop(0)