Lightning-AI/pytorch-lightning · error · ValueError

You requested to find {num_devices} devices but there are no

Error message

You requested to find {num_devices} devices but there are no visible CUDA devices on this machine.

What it means

_determine_batch_limits validates percentage-like Trainer arguments (limit_train_batches, limit_val_batches, limit_test_batches, limit_predict_batches, val_check_interval, overfit_batches). A value must be either a fraction in [0.0, 1.0] or an integral count (> 1 and a whole number). Anything else (e.g. 1.5, 2.5, -1) raises this MisconfigurationException.

Source

Thrown at src/lightning/fabric/accelerators/cuda.py:106

    tests for each GPU on the system until the target number of usable devices is found.

    A subset of GPUs on the system might be used by other processes, and if the GPU is configured to operate in
    'exclusive' mode (configurable by the admin), then only one process is allowed to occupy it.

    Args:
        num_devices: The number of devices you want to request. By default, this function will return as many as there
            are usable CUDA GPU devices available.

    Warning:
        If multiple processes call this function at the same time, there can be race conditions in the case where
        both processes determine that the device is unoccupied, leading into one of them crashing later on.

    """
    if num_devices == 0:
        return []
    visible_devices = _get_all_visible_cuda_devices()
    if not visible_devices:
        raise ValueError(
            f"You requested to find {num_devices} devices but there are no visible CUDA devices on this machine."
        )
    if num_devices > len(visible_devices):
        raise ValueError(
            f"You requested to find {num_devices} devices but this machine only has {len(visible_devices)} GPUs."
        )

    available_devices = []
    unavailable_devices = []

    for gpu_idx in visible_devices:
        try:
            torch.tensor(0, device=torch.device("cuda", gpu_idx))
        except RuntimeError:
            unavailable_devices.append(gpu_idx)
            continue

        available_devices.append(gpu_idx)

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. Use a fraction 0.0-1.0 for percentages (0.8, not 80)
2. Use a whole integer for exact batch counts (limit_train_batches=100)
3. Round fractional counts: int(count) or math.floor before passing

Example fix

# before
trainer = Trainer(limit_train_batches=2.5)

# after
trainer = Trainer(limit_train_batches=0.5)  # half the batches
# or
trainer = Trainer(limit_train_batches=2)  # exactly 2 batches

def validate_batches(v: float, name: str) -> Union[int, float]:
    if 0 <= v <= 1:
        return v
    if v > 1 and float(v).is_integer():
        return int(v)
    raise ValueError(f"{name}={v!r} must be in [0.0, 1.0] or an int")

def is_valid_batch_limit(v) -> bool:
    return isinstance(v, (int, float)) and (0 <= v <= 1 or (v > 1 and float(v).is_integer()))

When it happens

Trigger: Trainer(limit_train_batches=2.5); Trainer(overfit_batches=-1); Trainer(val_check_interval=1.2) treated through this path; any float > 1 that is not a whole number, or a negative number.

Common situations: Confusing fraction vs count semantics (thinking 2.0 means '2x the dataset'); config files with floats from sweeps; converting '80%' to 80 instead of 0.8; arithmetic producing fractional counts like len(loader)/2 when odd.

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

• Device should be CUDA, got {device} instead.
• `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
• You requested to find {num_devices} devices but this machine

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/c2725ddd9854503f. Report an issue: GitHub.