[Split] MoRIIO READ and K3 fixes superseded by focused PRs

[whx-sjtu]

Overview

Superseded by eight focused, independently main-based draft PRs. All remaining fixes have been extracted; use the child PRs for review and integration. This PR is closed without merging, and its original branch/head remains available for history.

Claims

Independent scope	Replacement PR
Public MultiConnector completion ownership	#60635
MRV2 FULL synchronous-load context	#60636
Speculative KV output readiness	#60637
READ remote-prefill eligibility	#60638
Shared MR extent and conv/SSM READ offsets	#60639
Hybrid READ failure/timeout ownership	#60640
Atomic RDMA listener-port allocation	#60641
Kimi K3 streaming EOF text flush	#60642

The five latest PRs are based on main 194da61de1769c28bfad2b6642c60cc7eca544dd and have no commit dependency on each other or on the previous three.

Validation

Latest extraction	Independent CPU result
READ remote-prefill eligibility	99 passed
Shared MR extent and conv/SSM READ offsets	32 passed
Hybrid READ failure/timeout ownership	34 passed
Atomic RDMA listener-port allocation	5 passed
Kimi K3 streaming EOF text flush	142 passed

Each fix has a production-only negative control. Changed-file pre-commit, including mypy, passed for every branch. All eight commits cherry-pick without conflicts, and their combined suite passed 343 tests, with 1 GPU-only skip and 1 model-loading deselection. Individual counts overlap and must not be added.

The five-way composition preserves all nine files of the residual diff when applied to the same main base, except for moving the READ eligibility regression to the existing scheduler suite and reusing its fixture. Source behavior and every regression case remain accounted for.

These are CPU results. No new GPU/RDMA/model evaluation, profiling or performance experiment was dispatched for this extraction. Historical combined-stack evaluations described in this PR's discussion do not validate the new child commits in isolation.

Details

Preserved residual source: fc9f0ca96d3990cc1c82600a68b8e92f74e5eecc. The earlier combined head b5626f70f21579165f535da84fd036c131223700 also remains in history. No force-push, branch deletion or merge was needed to retire this parent.

Evidence classification is explicit in each child: READ eligibility was found in a single-machine recovery experiment; port collision has CI stderr and native-listener reproduction; shared MR and hybrid cleanup have CPU reproductions without isolated field attribution; EOF text loss is deterministic in the real parser, without original generated-token evidence attributing warmup empty responses to it.

The connector-interface discussion is isolated in #60637. Other previously separated scopes remain #59164 (sync-load zeroing) and #59441 (discovery heartbeat). Producer source retention until ACK belongs to #56465. The MR child documents integration requirements for future hybrid WRITE in #60544; the EOF child documents overlap with the broader Parser Engine migration #50229.

AI assistance was used for implementation, extraction, tests and review. All child PRs remain drafts pending final human review and exact-source runtime validation.

---

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).
• AI assistance was used during the creation of this PR.
• Design Fit: Independent fixes moved into focused child PRs.
• Testing and Validation: CPU checks pass; each child documents pending hardware/serving validation and human confirmation.
• Code Quality and Style: Child changed-file checks passed, including mypy.
• Pull Request Contents: Complete replacement index and evidence boundaries recorded.

[YukioZzz]

#57700 has been reduced to MoRIIO DSpark enablement plus the shared NIXL clipping
helper: one commit, with no scheduler, SchedulerOutput, model-state, or
forward-context changes. Its independent validation does not include these
shared-path patches.

The removed FULL pre-forward context change matches the approach here, so
#57700 will not carry a competing implementation. One distinction to preserve
in the follow-up: the removed imported-state patch also carried the actual
Mamba running-state column and covered state-only loads and conflicting CoW
destinations. Those contracts are not established by the zeroing hook alone,
and I am not treating them as already covered here. The reproductions are
retained for coordination. Pending-import prefix publication is separate
scheduler-owner work, also excluded from #57700.

[YukioZzz]

Could we extend get_sync_load_block_ids() to cover LBHNC as well as LBNHC?
LBHNC also keeps each layer’s attention page contiguous, but the current layout guard returns no exclusions for it, leaving GPU zeroing able to race a synchronous RDMA READ into the same destination.
We isolated this change with the feature patch and pre-forward context fix held constant for #57700: excluding only fully overwritten attention READ destinations restored GSM8K full / concurrency 64 from 151/1319 (11.45%) to 1277/1319 (96.82%), without additional GPU synchronization, CoW filtering, or Mamba state-index changes. This validates the isolated fix, not the complete PR.

The proposed delta is small: change the guard to not in ("LBHNC", "LBNHC") and extend the existing parameterized test with LBHNC attention-import, state-only, and no-pending-receive cases. This passed 57 related tests; restoring the original guard fails only the new LBHNC attention-import case. Independent byte-range checks also confirmed complete page coverage for both layouts.

In fact, I prefer a small follow-up PR to handle sync-load zeroing bugfix which could narrow to 40 lines with some other test code, as this PR is too big and the merging process might be slow, and keeping #57700 focused on DSpark enablement.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @whx-sjtu.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork