Lightning-AI/pytorch-lightning · error · ValueError

GPU parallel_devices set through

Error message

GPU parallel_devices set through {self._strategy_flag.__class__.__name__} class, but accelerator set to {self._accelerator_flag}, please choose one device type

What it means

The passed Strategy instance has `parallel_devices` whose first device is a CUDA torch.device, but the `accelerator=` flag was explicitly set to something other than "auto", "cuda", or "gpu" (e.g. "cpu" or "tpu"). The connector refuses contradictory device specifications.

Solutions

Drop the accelerator flag (it will be inferred as "cuda" from parallel_devices)
Or set parallel_devices to CPU devices if you truly want CPU execution
Ensure the strategy object matches the target hardware when reusing configs

Example fix

# before
strategy = DDPStrategy(parallel_devices=[torch.device("cuda", 0)])
fabric = Fabric(strategy=strategy, accelerator="cpu")

# after
strategy = DDPStrategy(parallel_devices=[torch.device("cuda", 0)])
fabric = Fabric(strategy=strategy)

Defensive patterns

Strategy: validation

Validate before calling

devs = getattr(strategy, "parallel_devices", None)
if devs and devs[0].type == "cuda" and accelerator not in (None, "auto", "cuda", "gpu"):
    raise SystemExit(f"GPU parallel_devices conflict with accelerator={accelerator}")

Prevention

For CPU debug runs, rebuild the strategy with CPU parallel_devices instead of forcing accelerator="cpu"
Derive the accelerator flag from the strategy's devices rather than hardcoding

When it happens

Trigger: DDPStrategy(parallel_devices=[torch.device("cuda", 0)]) with Fabric(strategy=strategy, accelerator="cpu") — a GPU-configured strategy combined with a non-GPU accelerator flag.

Common situations: Forcing accelerator="cpu" for debugging while the strategy was auto-built with CUDA devices; reusing a GPU strategy object in CPU-only tests; cluster scripts that override accelerator without rebuilding the strategy.

Related errors

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/a803aca4a212acd1. Report an issue: GitHub.

Appendix: source

Thrown at src/lightning/fabric/connector.py:291

                if self.checkpoint_io:
                    raise ValueError("checkpoint_io set through both strategy class and plugins, choose one")
                self.checkpoint_io = self._strategy_flag._checkpoint_io
            if getattr(self._strategy_flag, "cluster_environment", None):
                if self._cluster_environment_flag:
                    raise ValueError("cluster_environment set through both strategy class and plugins, choose one")
                self._cluster_environment_flag = getattr(self._strategy_flag, "cluster_environment")

            if hasattr(self._strategy_flag, "parallel_devices") and self._strategy_flag.parallel_devices:
                if self._strategy_flag.parallel_devices[0].type == "cpu":
                    if self._accelerator_flag and self._accelerator_flag not in ("auto", "cpu"):
                        raise ValueError(
                            f"CPU parallel_devices set through {self._strategy_flag.__class__.__name__} class,"
                            f" but accelerator set to {self._accelerator_flag}, please choose one device type"
                        )
                    self._accelerator_flag = "cpu"
                if self._strategy_flag.parallel_devices[0].type == "cuda":
                    if self._accelerator_flag and self._accelerator_flag not in ("auto", "cuda", "gpu"):
                        raise ValueError(
                            f"GPU parallel_devices set through {self._strategy_flag.__class__.__name__} class,"
                            f" but accelerator set to {self._accelerator_flag}, please choose one device type"
                        )
                    self._accelerator_flag = "cuda"
                self._parallel_devices = self._strategy_flag.parallel_devices

    def _check_device_config_and_set_final_flags(self, devices: Union[list[int], str, int], num_nodes: int) -> None:
        if not isinstance(num_nodes, int) or num_nodes < 1:
            raise ValueError(f"`num_nodes` must be a positive integer, but got {num_nodes}.")

        self._num_nodes_flag = num_nodes
        self._devices_flag = devices

        if self._devices_flag in ([], 0, "0"):
            accelerator_name = (
                self._accelerator_flag.__class__.__qualname__
                if isinstance(self._accelerator_flag, Accelerator)
                else self._accelerator_flag

View on GitHub (pinned to 9fed5c27d2)