remove increasing nccl stream for overlapping allreduce and gemm

megatron/initialize.py

@@ -176,22 +176,11 @@ def _initialize_distributed():
             else:
                 args.local_rank = device
             torch.cuda.set_device(device)
-        # Increase cuda stream priority of NCCL ops when overlapping with other ops
-        if (not args.no_async_tensor_model_parallel_allreduce and
-                args.tensor_model_parallel_size > 1):
-            from torch._C._distributed_c10d import ProcessGroupNCCL
-
-            pg_options = ProcessGroupNCCL.Options()
-            pg_options.is_high_priority_stream = True
-            pg_options._timeout = timedelta(days=7)
-        else:
-            pg_options = None
-        # Call the init process
-        torch.distributed.init_process_group(
-            backend=args.distributed_backend,
-            world_size=args.world_size, rank=args.rank,
-            timeout=timedelta(days=7),
-            pg_options=pg_options)
+    # Call the init process
+    torch.distributed.init_process_group(
+        backend=args.distributed_backend,
+        world_size=args.world_size, rank=args.rank,
+        timeout=timedelta(days=7))
 
     # Set the tensor model-parallel, pipeline model-parallel, and
     # data-parallel communicators.