Since PyTorch already supports Intel XPU via torch.xpu, but PyTorch Lightning (PL) doesn't have built-in XPU accelerator support yet, we can bypass PL's device detection logic with a minimal, clean workaround that doesn't require modifying library code. Here's how to do it for your single-XPU setup:

Step 1: Force Default Device to XPU #

First, set PyTorch's default device to your XPU at the very start of your code. This ensures all tensors (model params, training data, etc.) are created directly on the XPU:

import torch
torch.set_default_device('xpu:0')  # Use 'xpu:0' for single device

Step 2: Initialize NeuralForecast Models Normally #

You don't need to manually move models to XPU—since we set the default device, all model parameters will be loaded onto XPU automatically:

from neuralforecast import NeuralForecast
from neuralforecast.models import NBEATS

# Example model initialization (adjust params to your use case)
models = [NBEATS(input_size=24, h=12, max_steps=1000)]
nf = NeuralForecast(models=models, freq='H')

Step 3: Configure PyTorch Lightning Trainer #

The key here is to avoid PL's automatic accelerator detection (which would default to CPU). Instead, set devices=1 and skip the accelerator parameter entirely. We can also disable some optional features to avoid device-related warnings:

from pytorch_lightning import Trainer

trainer = Trainer(
    devices=1,
    max_epochs=10,
    enable_model_summary=False,  # Optional: Avoids device-related summary errors
    logger=False  # Optional: Simplifies debugging
)

Step 4: Run Training as Usual #

Pass your custom Trainer to NeuralForecast's fit method—everything will run on the XPU:

nf.fit(df=your_training_dataframe, trainer=trainer)

Verify It's Working #

To confirm your model is running on XPU, print the device of any model parameter before training:

print(next(models[0].parameters()).device)  # Should output 'xpu:0'

Why This Works #

• torch.set_default_device overrides PyTorch's default tensor creation location, so all model and data tensors live on XPU from the start.
• By omitting the accelerator parameter in Trainer, PL won't force-move tensors to CPU—it will use the device they're already on.
• NeuralForecast inherits all device settings from PyTorch and the custom Trainer, no library modifications needed.

Notes #

• Make sure your PyTorch version supports Intel XPU (PyTorch 2.1+ with Intel's XPU drivers installed).
• This is a temporary workaround until PyTorch Lightning adds native XPU support.
• No need for IPEX—we're using PyTorch's native torch.xpu interface directly.

内容的提问来源于stack exchange，提问作者Marek Ozana