zero-3 always OVERFLOW!

when I tried my code with zero-2, everything seems ok. However, when I switched to zero-3 without changing any settings, the model always print warning info of OVERFLOW like this:
[2021-07-15 17:02:39,984] [INFO] [stage3.py:2725:_overflow_clean_up] [deepscale] OVERFLOW! Rank 0 Skipping step. Attempted loss scale: 64.0, reducing to 32.0

My ds config file is :
{ "gradient_accumulation_steps": 3, "train_micro_batch_size_per_gpu": 4, "prescale_gradients": false, "zero_allow_untested_optimizer": true, "zero_optimization": { "stage":3, "contiguous_gradients": false, "allgather_bucket_size": 1.9e+08, "reduce_bucket_size": 5e+08, "overlap_comm": true, "reduce_scatter": true, "verbose": false }, "steps_per_print": 4, "gradient_clipping": 1.0, "wall_clock_breakdown": true, "optimizer": { "type": "Adam", "params": { "lr": 1e-8, "betas": [ 0.9, 0.96 ], "eps": 1.0e-8, "weight_decay": 0.01 } }, "fp16": { "enabled": true, "initial_scale_power": 8, "loss_scale_window": 100 } }

when I set the "zero_optimization": with stage=2, everything works ok. My model is about 8.3B size, and I have tried 0.4.2 and 0.4.3, but all version gives the same error.

Start by inspecting stage3.py at _overflow_clean_up and compare the supplied ZeRO stage 2 and stage 3 configurations. Reproduce the issue with the listed gradient, bucket, optimizer, and fp16 settings; done means the repeated OVERFLOW behavior is explained and its resolution is confirmed for ZeRO stage 3.