[chore] OSS perf test, super minor (#495)

benchmarks/golden_configs/oss_mnist.py

@@ -4,7 +4,7 @@
 def get_golden_real_stats():
 
     return {
-        "reference_speed": 650,
+        "reference_speed": 660,
         "reference_memory": 1000,
         "reference_loss": 0.026,
     }

benchmarks/oss.py

@@ -150,7 +150,8 @@ def train(
 
     if optim_type == OptimType.oss_sharded_ddp:
         optimizer = OSS(params=model.parameters(), optim=OPTIM, lr=1e-4, momentum=0.9)
-        model = ShardedDDP(model, optimizer)
+        # Single node run typically, no need for reduce buckets
+        model = ShardedDDP(model, optimizer, reduce_buffer_size=0)
     else:
         device_ids = None if args.cpu else [rank]
         model = DDP(model, device_ids=device_ids, find_unused_parameters=False)  # type: ignore