fix(nixl): replay strided views on inference

[S1ro1]

Summary

• require the directly transported replay prefix to remain contiguous
• split replay at the first strided operation
• reproduce the strided view and the remaining operation suffix on inference

Why

A strided tensor can share the trainer root storage without being a contiguous transfer region. Treating that view as directly transportable fragments the NIXL copy plan and can make synchronization pathologically slow.

Validation

Validated with the Laguna-XS.2 NIXL + ModelExpress Kubernetes run:

• initial 2 x TP8 pool initialized as 16 ranks
• staggered third TP8 server joined as ranks 16-23
• initial policy transfer completed in about 7.5 seconds
• first 24-rank transfer completed in about 10.7 seconds
• training continued through subsequent synchronized steps

---

Note

Medium Risk
Changes how trainer→inference weight slices are split between RDMA and local replay; incorrect boundaries could corrupt weights, but the fix is narrowly scoped to prefix eligibility in replay planning.

Overview
NIXL weight transfer now treats only contiguous same-dtype views of the trainer root as directly RDMA-able; the prefix walk in plan_tensor_replay breaks when is_contiguous() fails, not only on dtype change or non-view tensors.

Strided views that still alias root storage are no longer included in the bulk copy plan—those ops move into replay_ops and run on the inference receive arena, avoiding fragmented NIXL copy plans and slow sync (as seen with staggered TP pool joins).

^{Reviewed by Cursor Bugbot for commit c95589c. Bugbot is set up for automated code reviews on this repo. Configure here.}

[JannikSt]

Was this written by hand?

[S1ro1]

Was this written by hand?

artisanally crafted with love