Lightning + NeuralForecast on Intel GPU — Native PyTorch XPU

• Use SingleXPUStrategy(device_index=0) to run Lightning and NeuralForecast on native torch.xpu.

That's it. Tested here on Intel Arc Pro B50 / Ubuntu 25.10 and Windows 11 with torch==2.11.0+xpu, pytorch-lightning==2.6.1, and neuralforecast==3.1.6.

Choose your path

→ I just want the fix — skip to Copy one file → I want to run a working example first — skip to Run the examples

Copy one file

Copy xpu_lightning.py.

• NativeXPUAccelerator subclasses Accelerator and maps Lightning's device hooks to native torch.xpu.
• SingleXPUStrategy subclasses SingleDeviceStrategy and builds torch.device("xpu", device_index) around NativeXPUAccelerator().

Minimal Lightning usage

import pytorch_lightning as L

from xpu_lightning import SingleXPUStrategy

trainer = L.Trainer(
    strategy=SingleXPUStrategy(device_index=0),
    devices=1,
)

Minimal NeuralForecast usage

from neuralforecast import NeuralForecast
from neuralforecast.models import NHITS

from xpu_lightning import SingleXPUStrategy

model = NHITS(
    h=12,
    input_size=24,
    max_steps=20,
    strategy=SingleXPUStrategy(device_index=0),
    devices=1,
)
nf = NeuralForecast(models=[model], freq="1d")
nf.fit(df)

See minimal_lightning_xpu.py and minimal_neuralforecast_xpu.py.

Run the examples

Install

WARNING: DO NOT INSTALL intel_extension_for_pytorch HERE. This repo uses native torch.xpu only.

pip install --index-url https://download.pytorch.org/whl/xpu "torch==2.11.0"
pip install "pytorch-lightning==2.6.1" "neuralforecast==3.1.6" "polars==1.38.1" "pytest==8.4.2" "scipy==1.17.1"

Or with uv (a fast package installer):

uv pip install torch --index https://download.pytorch.org/whl/xpu
uv pip install -r requirements.txt

Run

python minimal_lightning_xpu.py
# or: uv run minimal_lightning_xpu.py

python minimal_neuralforecast_xpu.py
# or: uv run minimal_neuralforecast_xpu.py

python -m pytest -s examples/
# or: uv run pytest -s examples/

Expected output: GPU available: False, used: False is normal for native torch.xpu.

Quick check during training: model.layer.weight.device.type should be xpu, or use torch.xpu.get_device_name(0).

Why this works — the one insight

Lightning checks accelerator by name, so this is the error you hit first:

ValueError: invalid accelerator name: xpu

Passing SingleXPUStrategy(device_index=0) sidesteps that string check. Lightning gets a ready-made strategy instance and follows its native torch.device("xpu", 0) path instead.

That whole bridge lives in xpu_lightning.py. The two classes in that file are 87 lines total.

Version requirements

Package	Tested	Tested
OS	Ubuntu 25.10	Windows 11
Python	3.13.12	3.12.13
torch (xpu wheel)	2.11.0+xpu	2.11.0+xpu
pytorch_lightning	2.6.1	2.6.1
neuralforecast	3.1.6	3.1.6
polars	1.38.1	1.38.1
intel_extension_for_pytorch	not installed	not installed

These are the versions tested in this repo. Other versions may work, but they are not verified here. NeuralForecast currently publishes Python 3.10-3.13 classifiers, and Python 3.14 is not yet supported there.

Repo structure

xpu_lightning.py                ← copy this; accelerator + strategy
minimal_lightning_xpu.py        Lightning demo on xpu:0
minimal_neuralforecast_xpu.py   NeuralForecast demo on xpu:0
examples/
  test_vram_proof.py            VRAM allocation proof on xpu:0
  test_performance.py           CPU vs XPU MLP comparison
  test_lightning_xpu.py         Lightning fit on xpu:0
  test_smoke_xpu.py             XPU and class smoke test
  test_neuralforecast_xpu.py    NHITS fit and predict on xpu:0
requirements.txt                pinned versions + XPU wheel note

Context

Lightning has no built-in XPU accelerator yet; see issue #20938. PyTorch has supported Intel GPUs natively since 2.5 (October 2024). Intel EOL'd IPEX (their legacy PyTorch extension) in March 2026. This repo is the shortest working bridge.

License

MIT