Lightning-AI/pytorch-lightning · error · ValueError

Device should be CUDA, got {device} instead.

Error message

Device should be CUDA, got {device} instead.

What it means

Trainer._init_debugging_flags validates fast_dev_run: when passed as an int it must be >= 0. Negative integers such as fast_dev_run=-1 are rejected immediately at Trainer construction because a negative run count is meaningless.

Source

Thrown at src/lightning/fabric/accelerators/cuda.py:36

from typing_extensions import override

from lightning.fabric.accelerators.accelerator import Accelerator
from lightning.fabric.accelerators.registry import _AcceleratorRegistry
from lightning.fabric.utilities.rank_zero import rank_zero_info


class CUDAAccelerator(Accelerator):
    """Accelerator for NVIDIA CUDA devices."""

    @override
    def setup_device(self, device: torch.device) -> None:
        """
        Raises:
            ValueError:
                If the selected device is not of type CUDA.
        """
        if device.type != "cuda":
            raise ValueError(f"Device should be CUDA, got {device} instead.")
        _check_cuda_matmul_precision(device)
        torch.cuda.set_device(device)

    @override
    def teardown(self) -> None:
        _clear_cuda_memory()

    @staticmethod
    @override
    def parse_devices(devices: Union[int, str, list[int]]) -> Optional[list[int]]:
        """Accelerator device parsing logic."""
        from lightning.fabric.utilities.device_parser import _parse_gpu_ids

        return _parse_gpu_ids(devices, include_cuda=True)

    @staticmethod
    @override
    def get_parallel_devices(devices: list[int]) -> list[torch.device]:

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. Set fast_dev_run to a valid value: True, False, or a non-negative int like 1 or 5
2. If computing it dynamically, clamp: fast_dev_run = max(0, int(value))
3. Use fast_dev_run=0 or False to disable it instead of -1

Example fix

# before
trainer = Trainer(fast_dev_run=-1)

# after
trainer = Trainer(fast_dev_run=False)  # or 1, 5, True, 0

def sanitize_fast_dev_run(v):
    if isinstance(v, bool):
        return v
    if isinstance(v, int):
        if v < 0:
            raise ValueError("fast_dev_run must be >= 0")
        return v
    return bool(v)

def is_valid_fast_dev_run(v) -> bool:
    return v is None or isinstance(v, bool) or (isinstance(v, int) and v >= 0)

When it happens

Trigger: Constructing Trainer(fast_dev_run=-1) or any negative int; computing fast_dev_run programmatically (e.g. from a config or CLI arg) where a subtraction/default yields a negative value.

Common situations: Sweep/search configs generating fast_dev_run from expressions; typos; porting scripts where fast_dev_run was derived from a dataset size that can be 0 or negative; passing a float like -1.0 is not caught here (only int is checked) but negatives should be avoided regardless.

Understand the failure class

Background: "Invalid value" and "allowed values are" config errors: what your library rejected and how to fix it — this error's family across 41 libraries.

Related errors

• You requested to find {num_devices} devices but there are no
• `max_epochs` must be a non-negative integer or -1. You passe
• `max_steps` must be a non-negative integer or -1 (infinite s
• No `{step_name}()` method defined to run `Trainer.{trainer_m
• f"`Trainer(barebones=True, fast_dev_run={fast_dev_run!r})` w

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/c05d80e01a22c038. Report an issue: GitHub.