You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Either:
A) Value is based on available threads on system. Should be fixable with n_jobs=-1 (except perhaps on cluster, see below).
B) Set through a command line argument (with value 1 or option A as default).
Additional context
Pre-1.3.0 versions of XGBoost have a bug and should use something like this instead.
XGBoost uses cgroup to determine if CPU limit should be set lower than actual system resources (when using n_jobs=-1). Not sure if this works nicely together with Slurm. Something like this might be needed for that instead. But either way would require some testing on the cluster no matter what solution is chosen.
The text was updated successfully, but these errors were encountered:
- Fixes an issue where the amount of threads that XGBoost could use to train was hardcoded. The user can now set that amount themselves, although limited to at least 1.
Closes#112
Describe the bug
Number of parallel threads used by XGBoost is hardcoded instead of checking n threads available on system:
Expected behavior
Either:
A) Value is based on available threads on system. Should be fixable with
n_jobs=-1
(except perhaps on cluster, see below).B) Set through a command line argument (with value
1
or option A as default).Additional context
Pre-1.3.0 versions of XGBoost have a bug and should use something like this instead.
XGBoost uses cgroup to determine if CPU limit should be set lower than actual system resources (when using
n_jobs=-1
). Not sure if this works nicely together with Slurm. Something like this might be needed for that instead. But either way would require some testing on the cluster no matter what solution is chosen.The text was updated successfully, but these errors were encountered: