LogisticRegression GPU fallback ignores allow_sklearn_after_onedal=False

[cakedev0]

Describe the bug

sklearnex.config_context(target_offload="gpu", allow_fallback_to_host=False, allow_sklearn_after_onedal=False) still allows sklearnex.linear_model.LogisticRegression to fall back to stock scikit-learn on host for a GPU-targeted run.

The operation completes successfully and only logs:

fallback to original Scikit-learn. All input data transferred to host for further backend computations.

This looks inconsistent with the config docs:

• allow_fallback_to_host=False should disallow GPU to host fallback.
• allow_sklearn_after_onedal=False should disallow fallback to stock scikit-learn when no accelerated version is available.

I also checked main, the dispatch path in sklearnex/_device_offload.py still appears to call the sklearn branch after writing the fallback log without checking allow_sklearn_after_onedal in that path.

To Reproduce

from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
from sklearnex import config_context
from sklearnex.linear_model import LogisticRegression

X, y = make_classification(
    n_samples=3000,
    n_features=20,
    n_informative=10,
    n_classes=2,
    random_state=42,
)
X_train, X_test, y_train, y_test = train_test_split(
    X.astype("float32"), y, test_size=0.2, random_state=0
)

with config_context(
    target_offload="gpu",
    allow_fallback_to_host=False,
    allow_sklearn_after_onedal=False,
):
    clf = LogisticRegression()
    clf.fit(X_train, y_train)
    clf.predict(X_test)

Observed behavior

The script completes without an exception. With sklearnex logging displayed, it reports fallback to stock scikit-learn on host:

INFO:sklearnex: sklearn.linear_model.LogisticRegression.fit: fallback to original Scikit-learn. All input data transferred to host for further backend computations.
INFO:sklearnex: sklearn.linear_model.LogisticRegression.predict: fallback to original Scikit-learn. All input data transferred to host for further backend computations.
INFO:sklearnex: sklearn.linear_model.LogisticRegression.decision_function: fallback to original Scikit-learn. All input data transferred to host for further backend computations.

Expected behavior

With both fallback controls set to False, this should fail instead of silently running stock scikit-learn on host.
Environment:

• OS: Linux
• Python: 3.12
• scikit-learn-intelex: distribution metadata reports 2026.0.0; sklearnex.__version__ reports 2199.9.9 (why this discrepancy btw?)
• scikit-learn: 1.8.0
• GPU visible through dpctl: Intel(R) Arc(TM) B390 GPU

[david-cortes-intel]

CC @icfaust

[syakov-intel]

Tentative direction is to drop allow_fallback_to_host and provide explicit feedback (warning? error?) once fallback happened. Also docs update required as a follow up.

[cakedev0]

The expected behavior for me would be:

• if allow_fallback_to_host is False: raise an error
• if allow_fallback_to_host is True: fallback silently? Raise a warning?

Maybe a fallback_to_host config param would be nicer, like fallback_to_host="ignore"/"warn"/"error"?

Anyway, for my usage of sklearnex (benchmarking against scikit-learn), it's important to have a way to catch when a fallback is happening, ideally with an error as it's the easier to catch.

[david-cortes-intel]

The expected behavior for me would be:

• if allow_fallback_to_host is False: raise an error
• if allow_fallback_to_host is True: fallback silently? Raise a warning?

Maybe a fallback_to_host config param would be nicer, like fallback_to_host="ignore"/"warn"/"error"?

Anyway, for my usage of sklearnex (benchmarking against scikit-learn), it's important to have a way to catch when a fallback is happening, ideally with an error as it's the easier to catch.

Note that target_offload is a rather inefficient mechanism. For a benchmark, would be better to use array API instead, which would be a more apples-to-apples comparison.

scikit-learn-intelex: distribution metadata reports 2026.0.0; sklearnex.version reports 2199.9.9 (why this discrepancy btw?)

Note that the version is not hard-coded, but is rather meant to be configured when running the setup.py script:

scikit-learn-intelex/setup.py

Line 88 in 2df0c57

sklearnex_version = (

If not specified, it should use 2199.9.9. If you see 2026.0 in the metadata, it might be because you have a previous install from pip of that release.

[cakedev0]

Note that target_offload is a rather inefficient mechanism.

Even when inputs are dpnp-gpu arrays? (or this is what you call using the array API?)

[david-cortes-intel]

Note that target_offload is a rather inefficient mechanism.

Even when inputs are dpnp-gpu arrays? (or this is what you call using the array API?)

Yes, like this:

with config_context(array_api_dispatch=True):
    ...

DPNP works without array API, but it involves data movements - see docs:
https://uxlfoundation.github.io/scikit-learn-intelex/2026.0/array_api.html

[cakedev0]

Ok I see, I missed that. I assumed array_api_dispatch was only effectively used for scikit-learn estimators supporting the array API, but actually sklearnex has a broader support on this side.

Thanks a lot for the catch!