Different seeds are giving the exact same loss on Zero 1,2 and 3 during multi gpu training [BUG]

Describe the bug
I am full-finetuning Llava using deepspeed and whenever I run it on a multi gpu setting with Zero 1,2 or 3, the loss curve is exactly the same for different seeds. I also tried to finetune a smaller part of the model using deepspeed with only 1 GPU and in this case I can get different loss curves for different seeds. But whenever it is a multi gpu setting this behavior changes. I also noticed that when I use version v0.14.4, all zero stages (1,2 and 3) have this behavior of giving the same loss for different seeds, whereas in the version v0.14.3 only zero 3 has it.

To Reproduce
Steps to reproduce the behavior:

• This is the function I use to set seed

def set_seed(seed):
    print(f"SETTING GLOBAL SEED TO {seed}")
    #pl.seed_everything(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False
    np.random.seed(seed)
    random.seed(seed)
    os.environ["PYTHONHASHSEED"] = str(seed)

• I am using this Llava repo
• This is the zero2 file I am using

{
    "fp16": {
        "enabled": "auto",
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },
    "bf16": {
        "enabled": "auto"
    },
    "train_micro_batch_size_per_gpu": "auto",
    "train_batch_size": "auto",
    "gradient_accumulation_steps": "auto",
    "zero_optimization": {
        "stage": 2,
        "overlap_comm": true,
        "contiguous_gradients": true,
        "sub_group_size": 1e9,
        "reduce_bucket_size": "auto"
    }
}

Expected behavior
I expect that using different seeds should produce different loss curves, they should not produce the exact same result

ds_report output
This is the ds_report after I allocate 5 A100 GPUs

--------------------------------------------------
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 .................. [92m[OKAY][0m
--------------------------------------------------
op name ................ installed .. compatible
--------------------------------------------------
[93m [WARNING] [0m async_io requires the dev libaio .so object and headers but these were not found.
[93m [WARNING] [0m async_io: please install the libaio-dev package with apt
[93m [WARNING] [0m If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found.
async_io ............... [93m[NO][0m ....... [93m[NO][0m
fused_adam ............. [93m[NO][0m ....... [92m[OKAY][0m
cpu_adam ............... [93m[NO][0m ....... [92m[OKAY][0m
cpu_adagrad ............ [93m[NO][0m ....... [92m[OKAY][0m
cpu_lion ............... [93m[NO][0m ....... [92m[OKAY][0m
[93m [WARNING] [0m Please specify the CUTLASS repo directory as environment variable $CUTLASS_PATH
evoformer_attn ......... [93m[NO][0m ....... [93m[NO][0m
fp_quantizer ........... [93m[NO][0m ....... [92m[OKAY][0m
fused_lamb ............. [93m[NO][0m ....... [92m[OKAY][0m
fused_lion ............. [93m[NO][0m ....... [92m[OKAY][0m
inference_core_ops ..... [93m[NO][0m ....... [92m[OKAY][0m
cutlass_ops ............ [93m[NO][0m ....... [92m[OKAY][0m
transformer_inference .. [93m[NO][0m ....... [92m[OKAY][0m
quantizer .............. [93m[NO][0m ....... [92m[OKAY][0m
ragged_device_ops ...... [93m[NO][0m ....... [92m[OKAY][0m
ragged_ops ............. [93m[NO][0m ....... [92m[OKAY][0m
random_ltd ............. [93m[NO][0m ....... [92m[OKAY][0m
[93m [WARNING] [0m sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.0
[93m [WARNING] [0m using untested triton version (2.0.0), only 1.0.0 is known to be compatible
sparse_attn ............ [93m[NO][0m ....... [93m[NO][0m
spatial_inference ...... [93m[NO][0m ....... [92m[OKAY][0m
transformer ............ [93m[NO][0m ....... [92m[OKAY][0m
stochastic_transformer . [93m[NO][0m ....... [92m[OKAY][0m
--------------------------------------------------
DeepSpeed general environment info:
torch install path ............... ['/dkfz/cluster/gpu/data/OE0612/s426n/robustness/lib/python3.10/site-packages/torch']
torch version .................... 2.0.1+cu117
deepspeed install path ........... ['/dkfz/cluster/gpu/data/OE0612/s426n/robustness/lib/python3.10/site-packages/deepspeed']
deepspeed info ................... 0.14.2, unknown, unknown
torch cuda version ............... 11.7
torch hip version ................ None
nvcc version ..................... 11.5
deepspeed wheel compiled w. ...... torch 2.0, cuda 11.7
shared memory (/dev/shm) size .... 503.84 GB

System info (please complete the following information):

• OS: CentOS Linux 7 (Core)
• 5-8 A100 GPUs on the same node
• Python 3.10.0

Start by reproducing the multi-GPU run with the provided set_seed function, Llava v1.0.1 setup, and Zero 2 configuration on the reported A100 environment. Compare behavior across DeepSpeed v0.14.3 and v0.14.4 and across Zero stages, then identify why different seeds produce identical loss curves and verify that corrected runs diverge.