[DDP] Change the --no-optimize-ddp flag to reflect the latest usage (#119437)

Compiled DDP now has 4 different optimization modes. This PR changes the Dynamo benchmark flag to reflect that change.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/119437
Approved by: https://github.com/wconstab, https://github.com/xmfan

benchmarks/dynamo/common.py

@@ -2986,9 +2986,10 @@ def parse_args(args=None):
     """,
     )
     parser.add_argument(
-        "--no-optimize-ddp",
-        action="store_true",
-        help="Disables dynamo DDPOptimizer (graph breaks). (Applies only when using --ddp benchmark mode).",
+        "--optimize-ddp-mode",
+        type=str,
+        default="ddp_optimizer",
+        help="Specify the DDP optimization mode -- the value of torch._dynamo.config.optimize_ddp.",
     )
     parser.add_argument(
         "--distributed-master-port",
@@ -3438,11 +3439,7 @@ def run(runner, args, original_dir=None):
         )
     if args.ddp:
         assert args.training, "DDP benchmark requires --training mode"
-        if args.no_optimize_ddp:
-            torch._dynamo.config.optimize_ddp = False
-        else:
-            # TODO(whc) after enabling DDPOptimizer by default this could be removed or assert
-            torch._dynamo.config.optimize_ddp = True
+        torch._dynamo.config.optimize_ddp = args.optimize_ddp_mode
         if args.only == "dlrm":
             log.error(
                 "DLRM+DDP is unsupported as it requires sharding the embedding layer separately from DDP"