Lightning-AI/pytorch-lightning · error · MisconfigurationException

No supported gpu backend found!

Error message

No supported gpu backend found!

What it means

When Trainer(accelerator='gpu') is chosen, the connector probes for a usable GPU backend (MPS, then CUDA). If neither MPS nor CUDA reports available, it raises this MisconfigurationException because 'gpu' cannot resolve to a concrete backend.

Source

Thrown at src/lightning/pytorch/trainer/connectors/accelerator_connector.py:341

                else self._accelerator_flag
            )
            raise MisconfigurationException(
                f"`Trainer(devices={self._devices_flag!r})` value is not a valid input"
                f" using {accelerator_name} accelerator."
            )

    @staticmethod
    def _choose_auto_accelerator() -> str:
        """Choose the accelerator type (str) based on availability."""
        return _select_auto_accelerator()

    @staticmethod
    def _choose_gpu_accelerator_backend() -> str:
        if MPSAccelerator.is_available():
            return "mps"
        if CUDAAccelerator.is_available():
            return "cuda"
        raise MisconfigurationException("No supported gpu backend found!")

    def _set_parallel_devices_and_init_accelerator(self) -> None:
        if isinstance(self._accelerator_flag, Accelerator):
            self.accelerator: Accelerator = self._accelerator_flag
        else:
            self.accelerator = AcceleratorRegistry.get(self._accelerator_flag)
        accelerator_cls = self.accelerator.__class__

        if not accelerator_cls.is_available():
            available_accelerator = [
                acc_str
                for acc_str in self._accelerator_types
                if AcceleratorRegistry[acc_str]["accelerator"].is_available()
            ]
            raise MisconfigurationException(
                f"`{accelerator_cls.__qualname__}` can not run on your system"
                " since the accelerator is not available. The following accelerator(s)"
                " is available and can be passed into `accelerator` argument of"

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. Verify CUDA availability with torch.cuda.is_available(); fix driver/toolkit installation or GPU visibility (CUDA_VISIBLE_DEVICES)
2. Use accelerator='auto' so Lightning falls back to CPU when no GPU exists
3. Explicitly run on CPU: Trainer(accelerator='cpu')

Example fix

# before
trainer = Trainer(accelerator="gpu", devices=1)
# after
accel = "gpu" if torch.cuda.is_available() else "cpu"
trainer = Trainer(accelerator=accel, devices=1)

import torch
if not (torch.backends.mps.is_available() or torch.cuda.is_available()):
    accelerator = "cpu"
else:
    accelerator = "gpu"
trainer = Trainer(accelerator=accelerator)

def gpu_backend_available() -> bool:
    import torch
    return torch.cuda.is_available() or getattr(torch.backends, "mps", None) and torch.backends.mps.is_available()

from lightning.pytorch.utilities.exceptions import MisconfigurationException
try:
    trainer = Trainer(accelerator="gpu")
except MisconfigurationException as e:
    if "No supported gpu backend" in str(e):
        trainer = Trainer(accelerator="cpu")
    else:
        raise

When it happens

Trigger: Trainer(accelerator='gpu') on a machine with no NVIDIA GPU/CUDA toolkit or Apple Silicon GPU; broken CUDA installs where CUDAAccelerator.is_available() is False.

Common situations: Running GPU training in CI containers without CUDA; CUDA driver/toolkit mismatch after system updates; Apple/AMD machines without proper backends.

Related errors

• GPUs requested but none are available.
• str(_TRANSFORMER_ENGINE_AVAILABLE)
• Cannot re-initialize CUDA in forked subprocess. To use CUDA
• Lightning can't create new processes if CUDA is already init
• GPUs should be a list

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/6b00a5f385ea6880. Report an issue: GitHub.