CPU/GPU interoperability POC

[viclafargue]

[beckernick]

Gave this a test locally. It's really smooth! Quick comment:

• When dispatching linear regression to CPU, the weights/bias attributes are not propagated back up to the cuML estimator -- they're accessible only via the underlying sk_model_. Ideally, cuML users won't need to think about the distinction between PGU and CPU dispatched models. They can access attributes in the same way regardless of hardware.

import cuml
from cuml.common.device_selection import using_device_type, set_global_device_type
from sklearn.datasets import make_regression

X, y = make_regression(
    n_samples=100000,
    noise=50,
    n_features=100,
)

set_global_device_type("cpu")

clf = cuml.linear_model.LinearRegression()
clf.fit(X,y)
print(clf.score(X, y))
print(clf.coef_, clf.intercept_)
0.9451533686273299
None None

[divyegala]

rerun tests

[dantegd]

rerun tests

[dantegd]

rerun tests

[dantegd]

@viclafargue @divyegala could you fix the copyright year so that CI can run a set of tests?:

python/cuml/common/array_descriptor.py:2 Issue: Current year not included in the copyright header

[cjnolet]

rerun tests

[dantegd]

rerun tests

[cjnolet]

I'm waiting on the raft packages to deploy to the rapidsai-nightly channel and then we should be able to rerun this PR.

[viclafargue]

I'm waiting on the raft packages to deploy to the rapidsai-nightly channel and then we should be able to rerun this PR.

Should we remove the latest commit?

[cjnolet]

Should we remove the latest commit?

Yes. Though it might be easier to revert the commit so you don't have to rebase / force push.

[cjnolet]

JFYI: We're still waiting for pylibraft / raft-dask to be deployed to conda, which could take another 25-40mins. I reverted the last commit in the meantime.

[cjnolet]

rerun tests

[viclafargue]

rerun tests

[codecov-commenter]

Codecov Report

Base: 79.26% // Head: 79.30% // Increases project coverage by +0.04% :tada:

Coverage data is based on head (495c676) compared to base (d65121a). Patch coverage: 95.12% of modified lines in pull request are covered.

Additional details and impacted files

@@               Coverage Diff                @@
##           branch-22.10    #4874      +/-   ##
================================================
+ Coverage         79.26%   79.30%   +0.04%     
================================================
  Files               181      183       +2     
  Lines             11540    11618      +78     
================================================
+ Hits               9147     9214      +67     
- Misses             2393     2404      +11     

Flag	Coverage Δ
dask	46.10% <58.53%> (+0.07%)	:arrow_up:
non-dask	68.77% <93.90%> (+0.11%)	:arrow_up:

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
python/cuml/common/array_descriptor.py	100.00% <ø> (ø)
python/cuml/common/input_utils.py	93.85% <50.00%> (-0.39%)	:arrow_down:
python/cuml/internals/global_settings.py	91.89% <86.36%> (-8.11%)	:arrow_down:
python/cuml/common/__init__.py	100.00% <100.00%> (ø)
python/cuml/common/device_selection.py	100.00% <100.00%> (ø)
python/cuml/common/memory_utils.py	82.35% <100.00%> (+2.47%)	:arrow_up:
python/cuml/dask/common/base.py	94.44% <100.00%> (+0.03%)	:arrow_up:
python/cuml/experimental/common/__init__.py	100.00% <100.00%> (ø)
python/cuml/explainer/common.py	88.57% <100.00%> (+0.16%)	:arrow_up:
python/cuml/testing/utils.py	89.19% <100.00%> (+0.07%)	:arrow_up:
... and 1 more

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

:umbrella: View full report at Codecov.
:loudspeaker: Do you have feedback about the report comment? Let us know in this issue.

[cjnolet]

@viclafargue webare past code freeze for 22.10 and decided to push this to 22.12. Can you base this ok the 22.12 branch?

[dantegd]

rerun tests

[dantegd]

@gpucibot merge