Add use_rmm flag to global configuration #6656
Conversation
| @@ -8,7 +8,7 @@ dependencies: | |||
| - pyyaml | |||
| - cpplint | |||
| - pylint | |||
| - numpy | |||
| - numpy=1.19.4 | |||
There was a problem hiding this comment.
This is a workaround, as the latest shap package appears to break with NumPy 1.19.5:
https://xgboost-ci.net/blue/organizations/jenkins/xgboost/detail/PR-6656/4/pipeline/221
try:
> from .. import _cext
E ImportError: numpy.core.multiarray failed to import
There was a problem hiding this comment.
@trivialfis Have you run into this issue? What's your take on it?
There was a problem hiding this comment.
never seen it before.
What's your take on it?
@RAMitchell worked on the c extension before so maybe have better insight.
There was a problem hiding this comment.
Using shap from conda-forge also fixes the issue, but unfortunately conda-forge has an outdated version of shap.
Codecov Report
@@ Coverage Diff @@
## master #6656 +/- ##
=======================================
Coverage 81.83% 81.83%
=======================================
Files 13 13
Lines 3809 3809
=======================================
Hits 3117 3117
Misses 692 692 Continue to review full report at Codecov.
|
Closes #6297. The user should explicitly inform XGBoost if RMM is used. If RMM is used, XGBoost should not allocate from the CUB allocator.