[BUG] Expert parallel hangs at the last MoE layer

Describe the bug
I'm using DeepSpeed MoE layer to build a multi-modal LLM, I'm using Phi-3 as the base model, and replaced the MLP layer with MoE layer in DeepSpeed. However, when I enabled expert parallel, the communication hangs at the last MoE layer.

To Reproduce
Steps to reproduce the behavior:

1. Load a base Phi-3 model from HuggingFace
2. Replace self.mlp to MoE layer
3. set expert_num=4 and ep_size=4
4. run training with zero stage 2 optimizer

Expected behavior
No hangs during training

ds_report output

--------------------------------------------------
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
--------------------------------------------------
async_io ............... [NO] ....... [OKAY]
fused_adam ............. [NO] ....... [OKAY]
cpu_adam ............... [NO] ....... [OKAY]
cpu_adagrad ............ [NO] ....... [OKAY]
cpu_lion ............... [NO] ....... [OKAY]
 [WARNING]  Please specify the CUTLASS repo directory as environment variable $CUTLASS_PATH
evoformer_attn ......... [NO] ....... [NO]
fp_quantizer ........... [NO] ....... [OKAY]
fused_lamb ............. [NO] ....... [OKAY]
fused_lion ............. [NO] ....... [OKAY]
inference_core_ops ..... [NO] ....... [OKAY]
cutlass_ops ............ [NO] ....... [OKAY]
transformer_inference .. [NO] ....... [OKAY]
quantizer .............. [NO] ....... [OKAY]
ragged_device_ops ...... [NO] ....... [OKAY]
ragged_ops ............. [NO] ....... [OKAY]
random_ltd ............. [NO] ....... [OKAY]
 [WARNING]  sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.1
 [WARNING]  using untested triton version (2.1.0), only 1.0.0 is known to be compatible
sparse_attn ............ [NO] ....... [NO]
spatial_inference ...... [NO] ....... [OKAY]
transformer ............ [NO] ....... [OKAY]
stochastic_transformer . [NO] ....... [OKAY]
--------------------------------------------------
DeepSpeed general environment info:
torch install path ............... ['/root/miniconda3/lib/python3.10/site-packages/torch']
torch version .................... 2.1.2+cu121
deepspeed install path ........... ['/root/miniconda3/lib/python3.10/site-packages/deepspeed']
deepspeed info ................... 0.14.4+unknown, unknown, unknown
torch cuda version ............... 12.1
torch hip version ................ None
nvcc version ..................... 12.1
deepspeed wheel compiled w. ...... torch 2.1, cuda 12.1
shared memory (/dev/shm) size .... 94.00 GB

System info (please complete the following information):

• OS: Ubuntu 22.04
• one node with 4 RTX 3090
• Python 3.10.8
• Torch: 2.1

Additional context
I used logger to check the whole code base, the odd thing is, DeepSpeed only hangs at the last MoE layer, for Phi-3 mini model, it hangs at 32'nd decoder layer. When I hit ctrl + c, the code stopped at result.wait(), which means something is wrong with communication. During training hangs, GPU 2,3(I got 4 GPUs) have very high communication rate(11GB/s) while GPU 0, 1 got 120MB/s. I also checked the specific line of code where the training stopped, and it's the MoE module in the last decoder layer.

Reproduce the issue with Phi-3, expert_num=4, ep_size=4, ZeRO stage 2, and four GPUs. Inspect the communication around result.wait() and the MoE module in the 32nd decoder layer. Done means training completes without hanging at the last MoE layer.