Repository navigation
[ckpt] feat: minimal Core transfer option for NIXL - #1
Draft
samnordmann wants to merge 2 commits into
Draft
samnordmann wants to merge 2 commits into
samnordmann wants to merge 2 commits into
Conversation
Snapshot of the current endpoint-transfer prototype for review in the author's fork. Native GPU performance parity and full end-to-end framework validation remain open; this is not an upstream-ready change. Assisted-by: OpenAI Codex Signed-off-by: Samuel Nordmann <snordmann@nvidia.com>
This was referenced Sep 17, 2026
Preserve the existing NIXL ring, descriptors, CUDA barriers and ZMQ protocol. Remove new wire/lifetime machinery and keep one focused READ-path contract test. Assisted-by: OpenAI Codex Signed-off-by: Samuel Nordmann <snordmann@nvidia.com>
samnordmann
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What / design
Add
transfer_api="torch"to the NIXL checkpoint engine, using Core and the shared NIXL shim. Default stays native. Keep ring topology, bucket buffers, CUDA barriers, native wire descriptors and ZMQ control messages unchanged.This is not the mixed-dtype alignment fix in upstream #7318; current NIXL/checkpoint PRs were checked for overlap.
Tests / limits
pytest -q tests/checkpoint_engine/test_nixl_transfer_api.py: 1 passed under an isolated CPU harness stubbing unavailable Ray/CuPy/runtime modules and native calls. Exercises the actual READ path and existing completion tag. No GPU/ring or performance result.Success-path prototype only: active cancellation/failure recovery is deferred. AI-assisted fork draft; human line-by-line review and native validation required before upstream submission.