[BUG] Running DDP with transformers integrated deepspeed get a deadlock (long time no response) when training model.

Describe the bug
I am running LLaVA code for improvements, but I added a nn.Parameter into the model and set requires_grad is True. I got a long time no response from the terminal. After I debugged train.py, I found that it may get a deadlock when running this line. There is no any response, just stuck here. When I go deeper with this line, the program is doing all_reduce in this line

https://github.com/microsoft/DeepSpeed/blob/ee7db48373bffedba9e5f5b570b7e07202b7a875/deepspeed/runtime/zero/stage_1_and_2.py#L1972

When I set requires_grad is False, everything works fine.

Expected behavior
Run properly with nn.Parameter when requires_grad is True.

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
--------------------------------------------------
 [WARNING]  async_io requires the dev libaio .so object and headers but these were not found.
 [WARNING]  async_io: please install the libaio-dev package with apt
 [WARNING]  If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found.
async_io ............... [NO] ....... [NO]
cpu_adagrad ............ [NO] ....... [OKAY]
cpu_adam ............... [NO] ....... [OKAY]
fused_adam ............. [NO] ....... [OKAY]
fused_lamb ............. [NO] ....... [OKAY]
quantizer .............. [NO] ....... [OKAY]
random_ltd ............. [NO] ....... [OKAY]
 [WARNING]  sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.0
 [WARNING]  using untested triton version (2.0.0), only 1.0.0 is known to be compatible
sparse_attn ............ [NO] ....... [NO]
spatial_inference ...... [NO] ....... [OKAY]
transformer ............ [NO] ....... [OKAY]
stochastic_transformer . [NO] ....... [OKAY]
transformer_inference .. [NO] ....... [OKAY]
--------------------------------------------------
DeepSpeed general environment info:
torch install path ............... ['/home/xiejunlin/miniconda3/envs/videollava/lib/python3.10/site-packages/torch']
torch version .................... 2.0.1
deepspeed install path ........... ['/home/xiejunlin/miniconda3/envs/videollava/lib/python3.10/site-packages/deepspeed']
deepspeed info ................... 0.9.5, unknown, unknown
torch cuda version ............... 11.8
torch hip version ................ None
nvcc version ..................... 11.5
deepspeed wheel compiled w. ...... torch 2.0, cuda 11.8

Screenshots

System info (please complete the following information):

• OS: Ubuntu 20.04
• GPU count and types: one machine with x8 A6000s each
• Interconnects: NVLink
• Python version: 3.10.2
• torch: 2.0.1
• transformers: 4.31.0

Start with the reported train.py path and inspect deepspeed/runtime/engine.py at line 2063 and deepspeed/runtime/zero/stage_1_and_2.py at line 1972. Reproduce the DDP and Transformers-integrated DeepSpeed run with the added nn.Parameter under both requires_grad settings, then determine what is needed for the all_reduce path to complete without a deadlock.