Lightning-AI/pytorch-lightning · critical · ImportError

PyTorch >= 2.6 requires DeepSpeed >= 0.16.0. Detected DeepSp

Error message

PyTorch >= 2.6 requires DeepSpeed >= 0.16.0. Detected DeepSpeed version: {deepspeed_version}. Please upgrade by running `pip install -U 'deepspeed>=0.16.0'`.

What it means

PyTorch 2.6 changed torch.load to default to weights_only=True, and DeepSpeed only supported that in 0.16.0+. If the env has torch>=2.6 with an older deepspeed, DeepSpeedStrategy.__init__ raises ImportError demanding an upgrade, because loading full checkpoints would otherwise break.

Source

Thrown at src/lightning/fabric/strategies/deepspeed.py:253

                per worker.

            exclude_frozen_parameters: Exclude frozen parameters when saving checkpoints.

        """
        if not _DEEPSPEED_AVAILABLE:
            raise ImportError(
                "To use the `DeepSpeedStrategy`, you must have DeepSpeed installed."
                " Install it by running `pip install -U deepspeed`."
            )

        if _TORCH_GREATER_EQUAL_2_6 and not _DEEPSPEED_GREATER_EQUAL_0_16:
            # Starting with PyTorch 2.6, `torch.load` defaults to `weights_only=True` when loading full checkpoints.
            # DeepSpeed added support for this behavior in version 0.16.0.
            import deepspeed

            deepspeed_version = deepspeed.__version__

            raise ImportError(
                f"PyTorch >= 2.6 requires DeepSpeed >= 0.16.0. "
                f"Detected DeepSpeed version: {deepspeed_version}. "
                "Please upgrade by running `pip install -U 'deepspeed>=0.16.0'`."
            )

        super().__init__(
            accelerator=accelerator,
            parallel_devices=parallel_devices,
            cluster_environment=cluster_environment,
            precision=precision,
            process_group_backend=process_group_backend,
        )
        self._backward_sync_control = None  # DeepSpeed handles gradient accumulation internally
        self._timeout: Optional[timedelta] = timeout

        self.config = self._load_config(config)
        if self.config is None:
            # User has not overridden config, set defaults

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. pip install -U 'deepspeed>=0.16.0'
2. Or pin torch<2.6 until deepspeed can be upgraded
3. Add a version-compatibility check to your environment setup so mismatches fail early with context

Example fix

# before
torch==2.6.0, deepspeed==0.15.4
# after (shell)
pip install -U 'deepspeed>=0.16.0'

import torch
from importlib.metadata import version
if tuple(int(x) for x in torch.__version__.split(".")[:2]) >= (2, 6) \
        and tuple(int(x) for x in version("deepspeed").split(".")[:2]) < (0, 16):
    raise RuntimeError("upgrade deepspeed: pip install -U 'deepspeed>=0.16.0'")

def deepspeed_torch_compatible() -> bool:
    import torch
    try:
        from importlib.metadata import version
        ds = tuple(int(x) for x in version("deepspeed").split(".")[:2])
    except Exception:
        return False
    t = tuple(int(x) for x in torch.__version__.split(".")[:2])
    return not (t >= (2, 6) and ds < (0, 16))

try:
    strategy = DeepSpeedStrategy()
except ImportError as e:
    raise RuntimeError(f"environment incompatible: {e}") from e

When it happens

Trigger: torch >= 2.6 installed together with deepspeed < 0.16.0, then constructing DeepSpeedStrategy (even if you never load a full checkpoint).

Common situations: Upgrading torch to 2.6+ in an env with a pinned old deepspeed; stale lockfiles after a base-image torch bump.

Related errors

• `precision={precision!r})` is not supported in DeepSpeed. `p
• To use the `DeepSpeedStrategy`, you must have DeepSpeed inst
• Currently only one optimizer is supported with DeepSpeed. Go
• The `{type(self).__name__}` does not support setting up the
• `{empty_init=}` is not a valid choice with `DeepSpeedStrateg

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/e3c415be16304779. Report an issue: GitHub.