gpu support for dgx spark (and its partner devices)

[busFred]

I'm on Lenovo ThinkStation PGX (NVidia DGX Spark's Lenovo equivalent); I have xgboost.__version__ = '3.1.2' installed via pip install xgboost. I tried device="cuda" and device="gpu" and both yield the same warning and fallback to cpu.

import xgboost as xgbst
import torch as th
x = th.randn((279, 580))
y = th.randn((279, 1))
xgbr = xgbst.XGBRegressor(device="cuda")
xgbr.fit(x.numpy(), y.numpy())

/home/hungtien/miniconda3/envs/tmplafa/lib/python3.13/site-packages/xgboost/training.py:199: UserWarning: [13:55:49] WARNING: /workspace/src/context.cc:207: Device is changed from GPU to CPU as we couldn't find any available GPU on the system.
bst.update(dtrain, iteration=i, fobj=obj)

xgbr = xgbst.XGBRegressor(device="gpu")
xgbr.fit(x.numpy(), y.numpy())

/home/hungtien/miniconda3/envs/tmplafa/lib/python3.13/site-packages/xgboost/training.py:199: UserWarning: [13:59:45] WARNING: /workspace/src/context.cc:207: Device is changed from GPU to CPU as we couldn't find any available GPU on the system.
bst.update(dtrain, iteration=i, fobj=obj)

[busFred]

might be related to #11563 but i'm not sure

[trivialfis]

Yeah, you found the right issue tracker.

On the other hand, we need to make some additional adaptation for DGX spark to improve performance, XGBoost is not developed with coherent memory in mind.