[BUG]There is a problem with asynchronous communication in zero stage2

Describe the bug There is a problem with asynchronous communication in zero stage2 by using overlap_comm.

To Reproduce Steps to reproduce the behavior: Use deepspeed zero-2 on the hugging face to train the bloomz-7b1-mt model. When you enable overlap_comm = true and control the randomness, you still find that the loss is different every time.

model: bloomz-7b1-mt zero2-config:

{
    "fp16": {
        "enabled": true,
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },

    "optimizer": {
        "type": "AdamW",
        "params": {
            "lr": "auto",
            "betas": "auto",
            "eps": "auto",
            "weight_decay": "auto"
        }
    },

    "scheduler": {
        "type": "WarmupLR",
        "params": {
            "warmup_min_lr": "auto",
            "warmup_max_lr": "auto",
            "warmup_num_steps": "auto"
        }
    },

    "zero_optimization": {
        "stage": 2,
        "offload_optimizer": {
            "device": "none",
            "pin_memory": true
        },
        "allgather_partitions": true,
        "allgather_bucket_size": 2e8,
        "overlap_comm": true,
        "reduce_scatter": true,
        "reduce_bucket_size": 2e8,
        "contiguous_gradients": true
    },

    "gradient_accumulation_steps": "auto",
    "gradient_clipping": "auto",
    "train_batch_size": "auto",
    "train_micro_batch_size_per_gpu": "auto",
    "wall_clock_breakdown": false
}

System info (please complete the following information):

OS: Ubuntu 20.04
1 machine with x8 A100s(NVIDIA A100-SXM)
Python version 3.8.13
deepspeed 0.8.1
transformers 4.26.1
torch 1.13.0a0+08820cb

Launcher context torchrun

Docker context nvcr.io/nvidia/pytorch:22.07-py3

microsoft / DeepSpeed

[BUG]There is a problem with asynchronous communication in zero stage2 #2868