[BUG] BFloat16 training accuracy issue with zero 0/1 (the more rank, the less accuracy) in CIFAR10

Describe the bug
When using BF16 to train CIFAR10 following https://github.com/microsoft/DeepSpeedExamples/pull/651. I encounter accuracy loss with the following conditions:

1. BF16 training
2. zero stage 0 and 1
3. world_size >= 2

To Reproduce
Steps to reproduce the behavior:

1. Use PR https://github.com/microsoft/DeepSpeedExamples/pull/651
2. Goto DeepSpeedExamples/training/cifar/
3. Change ds_config.json to change 'fp16' to 'bf16'
4. run the run_ds.sh with multiple accelerators on the system
5. Observe accuracy loss with multiple ranks. For 2 ranks, accuracy is 52%, much less than 57% of 1 rank. For 8 ranks, accuracy drops to around 43%
6. This can be observed from two CUDA cards. The 8 rank result is observed from a WIP CPU training branch. Should also be observed on 8 CUDA card system
7. When change data type to fp32 (on CPU system), there is no issue with 2 ranks or 8 ranks. When change to zero stage 2, there is no accuracy issue.

Expected behavior
BF16 train accuracy the same as single rank.

ds_report output

[2023-07-18 12:16:47,983] [INFO] [real_accelerator.py:133:get_accelerator] Setting ds_accelerator to cpu (auto detect)
[2023-07-18 12:16:48,129] [INFO] [real_accelerator.py:133:get_accelerator] Setting ds_accelerator to cpu (auto detect)
--------------------------------------------------
DeepSpeed C++/CUDA extension op report
--------------------------------------------------
NOTE: Ops not installed will be just-in-time (JIT) compiled at
      runtime if needed. Op compatibility means that your system
      meet the required dependencies to JIT install the op.
--------------------------------------------------
JIT compiled ops requires ninja
ninja .................. [OKAY]
--------------------------------------------------
op name ................ installed .. compatible
--------------------------------------------------
deepspeed_not_implemented  [NO] ....... [OKAY]
deepspeed_ccl_comm ..... [NO] ....... [OKAY]
fused_adam ............. [NO] ....... [OKAY]
--------------------------------------------------
DeepSpeed general environment info:
torch install path ............... ['/home/gma/anaconda3/envs/dscpu/lib/python3.11/site-packages/torch']
torch version .................... 2.0.1+cu117
deepspeed install path ........... ['/home/gma/DeepSpeed/deepspeed']
deepspeed info ................... 0.9.4+046afced, 046afced, master
deepspeed wheel compiled w. ...... torch 2.0

Screenshots

Finished Training
Finished Training
Finished Training
Finished Training
Finished Training
Finished Training
Finished Training
Finished Training
GroundTruth:    cat  ship  ship plane
GroundTruth:    cat  ship  ship plane
GroundTruth:    cat  ship  ship plane
Predicted:    cat plane  ship plane
Predicted:    cat plane  ship plane
GroundTruth:    cat  ship  ship planePredicted:
   cat plane  ship plane
GroundTruth:    cat  ship  ship plane
GroundTruth:    cat  ship  ship plane
Predicted:    cat plane  ship plane
Predicted:    cat plane  ship plane
Predicted:    cat plane  ship plane
GroundTruth:    cat  ship  ship plane
Predicted:    cat plane  ship plane
GroundTruth:    cat  ship  ship plane
Predicted:    cat plane  ship plane
Accuracy of the network on the 10000 test images: 43 %
Accuracy of the network on the 10000 test images: 43 %
Accuracy of the network on the 10000 test images: 43 %
Accuracy of the network on the 10000 test images: 43 %
Accuracy of the network on the 10000 test images: 43 %
Accuracy of the network on the 10000 test images: 43 %
Accuracy of the network on the 10000 test images: 43 %
Accuracy of the network on the 10000 test images: 43 %

System info (please complete the following information):

• OS: 6.4.0-rc2-2023-05-17-intel-next+ #1 SMP PREEMPT_DYNAMIC Wed May 17 15:36:48 PDT 2023 x86_64 x86_64 x86_64 GNU/Linux
• 1 machine with 2 RTX3090 cards / 1 machine with 2 SPR 48 core configured with SNC4
• Python version: 3.11.3

Launcher context
With DeepSpeed launcher

Docker context
No.

Start with DeepSpeedExamples/training/cifar/, the referenced ds_config.json, and run_ds.sh; reproduce the BF16 results across one and multiple ranks while comparing ZeRO stages 0, 1, and 2 and FP32. Trace the corresponding DeepSpeed distributed-training path, and consider the issue done when multi-rank BF16 accuracy matches the single-rank result without regressing the other configurations.