[hotfix] prevent nested ZeRO (#1140)

colossalai/zero/sharded_model/sharded_model_v2.py

@@ -77,6 +77,7 @@ class ShardedModelV2(nn.Module):
                  tensor_placement_policy: str = 'cuda',
                  gradient_predivide_factor: Optional[float] = 1.0,
                  reuse_fp16_shard: bool = False):
+        assert not isinstance(module, ShardedModelV2), 'Nested ShardedModelV2 is not supported.'
         super().__init__()
         self.logger = get_dist_logger()
 

colossalai/zero/sharded_optim/sharded_optim_v2.py

@@ -87,6 +87,7 @@ class ShardedOptimizerV2(ColossalaiOptimizer):
                  mp_process_group: Optional[ProcessGroup] = None,
                  verbose: bool = False) -> None:
         assert isinstance(sharded_model, ShardedModelV2), 'model must be wrapped with ShardedModel'
+        assert not isinstance(optimizer, ShardedOptimizerV2), 'Nested ShardedOptimizerV2 is not supported.'
 
         super().__init__(optimizer)
         self.shard_strategy = sharded_model.shard_strategy