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
this is patched in #3800 for the c10d backend on k8s. I've been using this for torchrun and seems to work well, but seems to have service object leakage when pods fail to start. The main thing is that pod hostnames don't get resolved by DNS automatically so the PR creates a headless service to fix this, but there are many other ways to achieve the same result possibly
rdvz fail to work with SkyPilot multi-node cluster (probably on k8s).
https://github.com/stas00/ml-engineering/blob/master/network/benchmarks/all_reduce_bench.py
Version & Commit info:
sky -v
: PLEASE_FILL_INsky -c
: PLEASE_FILL_INThe text was updated successfully, but these errors were encountered: