updated argument name

megatron/arguments.py

@@ -601,7 +601,7 @@ def _add_distributed_args(parser):
     group.add_argument('--use-cpu-initialization', action='store_true',
                        default=None, help='If set, affine parallel weights '
                        'initialization uses CPU' )
-    group.add_argument('--empty-unused-memory-each-iter', default=0, type=int,
+    group.add_argument('--empty-unused-memory-level', default=0, type=int,
                        choices=[0, 1, 2],
                        help='Call torch.cuda.empty_cache() each iteration '
                        '(training and eval), to reduce fragmentation.'