Implemented proper work with multiple threads

[olegkkruglov]

Changes proposed in this pull request:

• Add patching for _FuncWrapper class from sklearn.utils.parallel to properly call patched config_context
• Fix patching for get_config to allow run patched version in sklearn.utils.parallel
• Remove unnecessary queue propagation via config in BaseSVC._fit_proba which causes test failure after get_config patch fix

[samir-nasibli]

In the next PRs please don't mix implementing common feature and bug fix. But for now we can leave it as is.
Better disable the test and fix it on separate branch :)

[olegkkruglov]

/intelci: run

[samir-nasibli]

/intelci: run

[olegkkruglov]

/intelci: run

[olegkkruglov]

/intelci: run

[samir-nasibli]

@olegkkruglov please rebase your branches and run intelci

[olegkkruglov]

/intelci: run

[samir-nasibli]

/intelci: run

[ethanglaser]

Any additional tests/examples for this functionality? Or is it covered by existing?

[samir-nasibli]

You can use this reproducer for the test:

import numpy as np
from sklearnex import patch_sklearn, config_context
from sklearn.datasets import make_classification
from sklearn.ensemble import BaggingClassifier
patch_sklearn()

from sklearn.svm import SVC

X, y = make_classification(
    n_samples=1000,
    n_features=4,
    n_informative=2,
    n_redundant=0,
    random_state=0,
    shuffle=False,
)

with config_context(target_offload="gpu"):
    ExtraTreesClassifier(max_depth=2, random_state=0).fit(X, y)
        # decision_function
    ensemble = BaggingClassifier(
        SVC(decision_function_shape="ovr"), n_jobs=3, random_state=0
    ).fit(X, y)

The result is:

Intel(R) Extension for Scikit-learn* enabled (https://github.com/intel/scikit-learn-intelex)
INFO:sklearnex: sklearn.svm.SVC.fit: running accelerated version on GPU
INFO:sklearnex: sklearn.svm.SVC.fit: running accelerated version on GPU
INFO:sklearnex: sklearn.svm.SVC.fit: running accelerated version on GPU
INFO:sklearnex: sklearn.svm.SVC.fit: running accelerated version on GPU
INFO:sklearnex: sklearn.svm.SVC.fit: running accelerated version on GPU
INFO:sklearnex: sklearn.svm.SVC.fit: running accelerated version on GPU
INFO:sklearnex: sklearn.svm.SVC.fit: running accelerated version on GPU
INFO:sklearnex: sklearn.svm.SVC.fit: running accelerated version on GPU
INFO:sklearnex: sklearn.svm.SVC.fit: running accelerated version on GPU
INFO:sklearnex: sklearn.svm.SVC.fit: running accelerated version on GPU

On master it fallbacks on CPU. So your branch works correct with n_jobs enabled in this case.

[ethanglaser]

On master it fallbacks on CPU. So your branch works correct with n_jobs enabled in this case.

Perhaps adding something similar with allow_fallback_to_host=False flag added to config_context would be a good test of whether things are working properly?

[samir-nasibli]

On master it fallbacks on CPU. So your branch works correct with n_jobs enabled in this case.

Perhaps adding something similar with allow_fallback_to_host=False flag added to config_context would be a good test of whether things are working properly?

Good point, we can update logger for that as well.

[Alexsandruss]

@Mergifyio rebase

[mergify]

rebase

❌ Base branch update has failed

Git reported the following error:

Rebasing (1/7)
Auto-merging sklearnex/dispatcher.py
CONFLICT (content): Merge conflict in sklearnex/dispatcher.py
error: could not apply 284f7a7... Implemented proper work with multiple threads
hint: Resolve all conflicts manually, mark them as resolved with
hint: "git add/rm <conflicted_files>", then run "git rebase --continue".
hint: You can instead skip this commit: run "git rebase --skip".
hint: To abort and get back to the state before "git rebase", run "git rebase --abort".
Could not apply 284f7a7... Implemented proper work with multiple threads

[Alexsandruss]

/intelci: run

[ethanglaser]

Suggested change

	with config_context(target_offload="gpu"):
	with config_context(target_offload="gpu", allow_fallback_to_host=False):

[ethanglaser]

Actually maybe this modification isn't necessary, but I am a bit confused by the test - when would dpctl be available but no GPU? Thanks for adding the test.

[Alexsandruss]

dpctl is not only for gpu devices. For example, CI has dpctl installed without gpu in azure pipelines used instances.

[Alexsandruss]

/intelci: run

[Alexsandruss]

/intelci: run

[Alexsandruss]

/intelci: run

[Alexsandruss]

http://intel-ci.intel.com/ee5150fa-6720-f13c-870e-a4bf010d0e2e

[samir-nasibli]

Please attach GPU CI job as well.

[ethanglaser]

Please attach GPU CI job as well.

http://intel-ci.intel.com/ee51704d-b860-f1a2-a5d9-a4bf010d0e2e

[napetrov]

@Alexsandruss should example will be updated https://github.com/intel/scikit-learn-intelex/blob/master/examples/sklearnex/n_jobs.py

[Alexsandruss]

@Alexsandruss should example will be updated https://github.com/intel/scikit-learn-intelex/blob/master/examples/sklearnex/n_jobs.py

Will be in next PR with n_jobs parameter update.

[samir-nasibli]

Do we have proper tracker for this issues?