Init the process group with a load-scaled timeout for sharded loading - #48228
Merged
Conversation
Sharded loading of a TB-scale checkpoint takes tens of minutes and ranks finish far apart (measured: >10 minutes spread on GLM-4.6, 714 GB over 16 ranks), so the fastest ranks' first collective dies on the default 10-minute NCCL watchdog. Use a 2-hour timeout when transformers initializes the group itself.
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
Contributor
CI recapDashboard: View test results in Grafana |
ArthurZucker
approved these changes
Sep 1, 2026
ArthurZucker
left a comment
ArthurZucker
left a comment
Collaborator
There was a problem hiding this comment.
LGTM ty for finding this
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do?
When transformers initializes torch.distributed itself (the
DistributedConfigloading path), the process group gets the default 10-minute NCCL watchdog. Too short for sharded loading of large checkpoints:The failure is nasty to diagnose: 0% GPU, all ranks CPU-spinning, no output — it looks like a hang.
Fix: use a 2-hour timeout when transformers creates the group. Users who init the group themselves keep their own timeout.
Note the measurements above are GLM-4.6 (357B, 665 GiB) at 16 ranks — mid-range for this loading path. Larger checkpoints (GLM-5.2 is 1.4 TiB), more nodes, higher-latency filesystems, and configurations with more rank-to-rank variance all widen the spread further.
Found while training GLM-4.6/GLM-5.2 through
Trainer(#48204, which carries the same commit).Part of the loading-performance work tracked in #48239.