[BUG] Unit Test TestFp8ComposabilityAcrossZero.test[fp16] failure

The DeepSpeed unit test TestFp8ComposabilityAcrossZero.test[fp16] fails with assertion error:

_________________________________________ TestFp8ComposabilityAcrossZero.test[fp16] __________________________________________
multiprocessing.pool.RemoteTraceback:
"""
Traceback (most recent call last):
  File "/usr/lib/python3.12/multiprocessing/pool.py", line 125, in worker
    result = (True, func(*args, **kwds))
                    ^^^^^^^^^^^^^^^^^^^
  File "/usr/lib/python3.12/multiprocessing/pool.py", line 51, in starmapstar
    return list(itertools.starmap(args[0], args[1]))
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/root/not_working/DeepSpeed/tests/unit/common.py", line 325, in _dist_run
    raise e
  File "/root/not_working/DeepSpeed/tests/unit/common.py", line 317, in _dist_run
    self.run(**self._fixture_kwargs)
  File "/root/not_working/DeepSpeed/tests/unit/common.py", line 473, in run
    self._current_test(**fixture_kwargs)
  File "/root/not_working/DeepSpeed/tests/unit/runtime/half_precision/test_fp8.py", line 98, in test
    assert (all_equal)
            ^^^^^^^^^
AssertionError
"""

This regression started occurring after the following commit: https://github.com/deepspeedai/DeepSpeed/commit/889f0ead27435cd7755a73a9ba27e0913ab3c548

To Reproduce

1. git clone https://github.com/deepspeedai/DeepSpeed.git
2. cd DeepSpeed
3. pip install .
4. cd tests
5. TORCH_EXTENSIONS_DIR=./torch-extensions pytest -s --color=yes --durations=0 --verbose unit/runtime/half_precision/test_fp8.py::TestFp8ComposabilityAcrossZero::test[fp16]

ds_report output

[2025-09-22 21:35:15,755] [INFO] [real_accelerator.py:260:get_accelerator] Setting ds_accelerator to cuda (auto detect)
[2025-09-22 21:35:18,407] [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 (3.3.1), 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 (3.3.1), 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 ............... ['/usr/local/lib/python3.12/dist-packages/torch']
torch version .................... 2.8.0a0+34c6371d24.nv25.08
deepspeed install path ........... ['/usr/local/lib/python3.12/dist-packages/deepspeed']
deepspeed info ................... 0.17.6+66ad2780, 66ad2780, master
torch cuda version ............... 13.0
torch hip version ................ None
nvcc version ..................... 13.0
deepspeed wheel compiled w. ...... torch 2.8, cuda 13.0
shared memory (/dev/shm) size .... 64.00 MB
  [WARNING] /dev/shm size might be too small, if running in docker increase to at least --shm-size='1gb'
  [WARNING] see more details about NCCL requirements: https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/troubleshooting.html#sharing-data

System info (please complete the following information):

• OS: Ubuntu 24.04.2 LTS
• GPU count and types: 8 NVIDIA H100 80GB HBM3
• Python version: Python 3.12.3

Docker context
docker run --gpus all -it --name DeepSpeed_F8_UT nvcr.io/nvidia/pytorch:25.08-py3

Start by running tests/unit/runtime/half_precision/test_fp8.py::TestFp8ComposabilityAcrossZero::test[fp16] using the reported Docker and GPU environment. Read the failing test at tests/unit/runtime/half_precision/test_fp8.py:98 and compare behavior before and after commit 889f0ead27435cd7755a73a9ba27e0913ab3c548. Done means the regression is addressed and this test passes.