[BUG] Zero 3 checkpoints not saving model state when load_universal=True

[Vervious]

Describe the bug
When doing model_engine.save_checkpoint, when zero is set to stage 3, and "load_universal=True" the model state files do not appear to get saved (except on rank 0).

All bf16_zero_pp_rank_X_mp_rank_00_optim_states.pt files seem correctly saved. But only a single zero_pp_rank_0_mp_rank_00_model_states.pt is saved, and none for other ranks. Both the _to_universal or _to_fp32 checkpoint scripts fail due to missing model state.

Commenting out the

[deepspeed.checkpoint]
load_universal = true

from the config fixes the problem.

To Reproduce
Steps to reproduce the behavior:

Do a multinode training run.

[deepspeed.bf16]
enabled = true

[deepspeed.optimizer]
type = "Adam"

[deepspeed.optimizer.params]
torch_adam = false
[...]

[deepspeed.zero_optimization]
stage = 3
allgather_partitions = true
overlap_comm = true
reduce_scatter = true
contiguous_gradients = true
gather_16bit_weights_on_model_save = true
use_all_reduce_for_fetch_params = false  # Wanted to set to True a la https://github.com/deepspeedai/DeepSpeed/pull/5420 due to memory peak, but has bug/memory leak
load_from_fp32_weights = false
reduce_bucket_size = 5e8

[deepspeed.checkpoint]
load_universal = true

Expected behavior
Checkpoints get saved properly.

ds_report output

[2025-09-04 21:33:26,267] [INFO] [real_accelerator.py:254:get_accelerator] Setting ds_accelerator to cuda (auto detect)
[2025-09-04 21:33:28,670] [INFO] [logging.py:107:log_dist] [Rank -1] [TorchCheckpointEngine] Initialized with serialization = False
--------------------------------------------------
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]
fused_adam ............. [NO] ....... [OKAY]
cpu_adam ............... [NO] ....... [OKAY]
cpu_adagrad ............ [NO] ....... [OKAY]
cpu_lion ............... [NO] ....... [OKAY]
dc ..................... [NO] ....... [OKAY]
 [WARNING]  Please specify the CUTLASS repo directory as environment variable $CUTLASS_PATH
evoformer_attn ......... [NO] ....... [NO]
 [WARNING]  FP Quantizer is using an untested triton version (3.4.0), only 2.3.(0, 1) and 3.0.0 are known to be compatible with these kernels
fp_quantizer ........... [NO] ....... [NO]
fused_lamb ............. [NO] ....... [OKAY]
fused_lion ............. [NO] ....... [OKAY]
 [WARNING]  gds requires the dev libaio .so object and headers but these were not found.
 [WARNING]  gds: 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.
gds .................... [NO] ....... [NO]
transformer_inference .. [NO] ....... [OKAY]
inference_core_ops ..... [NO] ....... [OKAY]
cutlass_ops ............ [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 2.8
 [WARNING]  using untested triton version (3.4.0), only 1.0.0 is known to be compatible
sparse_attn ............ [NO] ....... [NO]
spatial_inference ...... [NO] ....... [OKAY]
transformer ............ [NO] ....... [OKAY]
stochastic_transformer . [NO] ....... [OKAY]
utils .................. [NO] ....... [OKAY]
--------------------------------------------------
DeepSpeed general environment info:
torch install path ............... ['/opt/micromamba/envs/runtime/lib/python3.10/site-packages/torch']
torch version .................... 2.8.0+cu128
deepspeed install path ........... ['/opt/micromamba/envs/runtime/lib/python3.10/site-packages/deepspeed']
deepspeed info ................... 0.17.3, unknown, unknown
torch cuda version ............... 12.8
torch hip version ................ None
nvcc version ..................... 12.8
deepspeed wheel compiled w. ...... torch 0.0, cuda 0.0
shared memory (/dev/shm) size .... 1.46 TB

System info (please complete the following information):

• OS: Ubuntu 22.04.5 LTS
• GPU count and types: 8 nodes, 8 H200s each
• Interconnects (if applicable): IB
• Python version 3.10.15

Launcher context
Torch Elastic

Additional context
Commenting out the

[deepspeed.checkpoint]
load_universal = true

from the config fixes the problem.

[sfc-gh-truwase]

@Vervious thanks for the reporting this issue as well as a workaround.

1. Could you please include the stack trace of the failing _to_universal and _to_fp32 scripts?
2. Our existing example simulates a scenario of one failure/resumption. From your description, it seems this error should be reproducible by extending the example to two failures/resumptions (e.g., failing/resuming iteration-200). Are you able to test this this hypothesis?

@xylian86 FYI

[sfc-gh-truwase]

@saforem2 have you seen any issue like this?

[sfc-gh-truwase]

This appears similar to #7584

@zhengchenyu, FYI

[zhengchenyu]

I think that single zero_pp_rank_0_mp_rank_00_model_states.pt is saved is not bug. This noraml behavior was introduced in #5475. But this behavior may cause the bug you describe later

I think _to_universal or _to_fp32 checkpoint scripts fail, is not due to missing model file, but due to mode file corrupted. It will be fix by #7585.