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
Right now all nccl communicators are created on-demand, which can increase startup time. We should create the common ones in the background on init (e.g. all GPUs, maybe that's it?).
Startup time has been cited by users as an expensive problem when running on large GPU clusters.
Right now all nccl communicators are created on-demand, which can increase startup time. We should create the common ones in the background on init (e.g. all GPUs, maybe that's it?).
Startup time has been cited by users as an expensive problem when running on large GPU clusters.
cc @ezhulenev
The text was updated successfully, but these errors were encountered: