Repository navigation
Support per-node RDMA ports and reject ambiguous fabric discovery - #66
Merged
Merged
Conversation
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Distributed launches now use explicit per-node RDMA devices and socket interfaces. Ambiguous discovery fails before any model or container changes; explicit rail lists remain exact, including asymmetric device names.
Validation: focused regression tests and static checks passed. Both existing port-1 rails between the second Spark pair passed bounded RDMA transfers; updated SparkGLM and SparkQwen launchers accepted the explicit rails and rejected ambiguous discovery and the backchannel. Existing first-pair containers were unchanged.
The Atlas GLM recipe consumes cluster.nodes[nodeId].fabricHca; NVIDIA Sync refresh preserves explicit node overrides.