Force workspace usage with MPI backends on MNNVL communicators when fabric allocated workspace is available.#108
Merged
romerojosh merged 3 commits intomainfrom Mar 4, 2026
Merged
Conversation
…ors when fabric allocated workspace is available. Signed-off-by: Josh Romero <joshr@nvidia.com>
Signed-off-by: Josh Romero <joshr@nvidia.com>
Collaborator
Author
|
/build |
|
🚀 Build workflow triggered! View run |
|
✅ Build workflow passed! View run |
Signed-off-by: Josh Romero <joshr@nvidia.com>
Collaborator
Author
|
/build |
|
🚀 Build workflow triggered! View run |
|
✅ Build workflow passed! View run |
romerojosh
changed the title
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
This PR detects and disables some of the transpose shortcut paths (e.g. alltoall directly from/to user input/output buffers bypassing the workspace) for MPI backends in situations where:
CUDECOMP_ENABLE_CUMEM=1)The reason for this is that at the current time, MPI communication over MNNVL involving non-fabric allocated buffers uses a staging protocol to route communication via internal fabric-allocated buffers which is a bit less efficient than communication using already fabric allocated buffers, similar to the treatment of managed memory allocations. The shortcut paths present MPI with non-fabric allocated user input/output buffers which trigger this less efficient path, negating the benefits of the shortcut.
Similarly logic is applied to halo communication cases that would normally bypass work space staging.