Repository navigation
[Legacy host seams] Owners: re-submit only still-needed minimal contracts #1
Description
The two legacy repositories are immutable archives. Please do not reopen or merge their feature branches wholesale. If the capability is still needed against current main, submit a new, narrowly scoped PR to this fresh upstream-synchronized fork.
Candidates that still need an owner decision:
- @WMASTER123 — generation-scoped prefix-route fence from legacy Different from the output of the HF inference vllm-project/vllm#272: Draft: generation-scoped prefix route fence support intellistream/vllm-hust-legacy-20260831#272 . Re-submit only a generic cache-event/fence contract required by
vllm-hust-prefix-router. - @CubeLander — request-owned host contracts from Why does loading the vicuna-13b-v1.1 model in FastChat require 26GB of GPU memory, while VLLM requires 37GB of GPU memory? vllm-project/vllm#253/Can vllm serving clients by using multiple model instances? vllm-project/vllm#239: [Feat][Scheduler] Integrate StateHarbor request-owned runtime intellistream/vllm-hust-legacy-20260831#253 and feat(request-owned): integrate linear DSpark verification intellistream/vllm-hust-legacy-20260831#239 . The dependency-neutral protocol has been migrated to StateHarbor; re-submit only independently reviewable scheduler/worker seams, not the 83-file branch.
- @xiehanlong834-gif — generic default-off KV lifecycle event seam represented by [BugFix] Fix a bug in counting running sequences vllm-project/vllm#266: refactor(kv-offload): pluginize B134 event instrumentation (replaces #260) intellistream/vllm-hust-legacy-20260831#266 . Keep exporter/policy code in
vllm-hust-kv-transfer-observability. - @SuccinctPaul / @YWHUTER — host synchronization and pinned-output lifecycle from Multi-GPU inference and Specify which GPUs to be used during inference vllm-project/vllm#250/Question about sampler. It takes too much time vllm-project/vllm#249: fix: isolate guarded output synchronization domains for safe host batching intellistream/vllm-hust-legacy-20260831#250 and [M1] Add host gate for TP1 pinned-output arena lifecycle intellistream/vllm-hust-legacy-20260831#249 . Re-submit only if the current host still lacks the neutral lifecycle/observer seam.
- @Raing5Days — core graph admission/key seam from Raise error for long prompt vllm-project/vllm#273: feat: add full_graph_parallel (dual-stream) intellistream/vllm-hust-legacy-20260831#273 . Keep stream/capture/replay policy in
vllm-ascend-split-batch-hust.
Before opening a PR:
- verify current upstream does not already implement the behavior;
- rebase/reimplement against current
mainrather than cherry-picking the old branch; - preserve original authorship and link the legacy PR;
- add default-off equivalence and fail-closed compatibility tests;
- keep algorithms, policies, exporters, and device implementations in their extension repositories.
Already present in the current fork/upstream and therefore not requested again: EAGLE/EAGLE3/DSpark, QkNorm-RoPE, Mooncake connectors, current KV-offload/tiering infrastructure, and general hot-path/CI fixes.
Activity
YWHUTER commentedon Sep 1, 2026
@ShuhaoZhangTony owner decision completed for the @YWHUTER half of the pinned-output lifecycle item.
I audited exact fresh
mained9f108ad80f4d8f058cf35541c785452c1c48efbefore re-submitting: the native sampled-token D2H/event-wait/optionalInputBatchsecond-consumer path remains, whilevllm/v1/events.pyor an equivalent neutral typed observer outlet is absent. No open PR in this fork or upstream implements this contract.A clean current-main reimplementation is now Draft PR #3, commit
ede30a2c9a9e434d283139af9a191f323b48067a. It changes four Python files only: a default-off process-local typed outlet, seven generic async-output lifecycle payloads, guarded native call sites, and host tests. Exporter/ledger/cap/correlation/policy/device code remains out of Core. Legacy vllm-project#249/vllm-project#250/vllm-project#256/vllm-project#266 provenance and co-authors are preserved.Validation: targeted pytest 4 passed; Ruff/format, mypy 3.12, SPDX, forbidden-import, torch.cuda, typos, configuration, sign-off, and diff checks passed. No NPU run or performance claim. Parent Draft intellistream/vllm-host-control-batching#12 pins the new commit in
b15eec5; independent Core API review is pending.xiehanlong834-gif commentedon Sep 22, 2026
Owner decision for the
@xiehanlong834-gifitem (generic default-off KV lifecycleevent seam from legacy vllm-project#266): already re-submitted and under review as #6.
Fresh audit of current
main(d680ae5a) before re-submitting:vllm/v1/events.pydoes not exist onmain; there is no request/session-levellifecycle outlet at the scheduler layer.
vllm/distributed/kv_events.py(
BlockStored/BlockRemoved) reports block residency, and the worker-levelAsyncOutput*hooks cover output-token movement on the GPU path — neitherreports "this request finished / was preempted / its KV was reclaimed", and
neither carries a request identity that survives preemption.
re-submission of [BugFix] Fix a bug in counting running sequences vllm-project/vllm#266:
vllm/v1/events.py(new) plus emit points inScheduler._free_request,Scheduler._preempt_requestand the block-releasepaths, with
tests/v1/core/test_request_lifecycle_events.py.Scope is kept to what this issue asks for:
RequestFinished,RequestPreempted,RequestKvReclaimedcarry request metrics only — no CLM/DLA/experimentsemantics in core.
EventBus.enabledisFalseuntil a sink registers; everyemit site is guarded, so the default path allocates nothing.
vllm-hust-kv-transfer-observabilityandvllm-hust-clm-lifecycle, as theissue instructs.
mainrather than cherry-picking the old branch, withoriginal authorship and the legacy PR linked in the commit/PR description.
No new PR is needed for this item; #6 replaces vllm-project#266. Marking the checkbox above
can follow the review outcome.
One environment note for whoever picks up the remaining items on this list: the
vllm-ascend-hustcurrently installed on the shared node (0.19.1) no longerimports against current
main—vllm_ascend/ops/fused_moe/fused_moe.py:32does
from vllm.model_executor.layers.fused_moe.layer import FusedMoE, whichmainno longer exports. Any test that constructs a realSchedulerfails therefor that reason alone (reproduced with the unmodified upstream
tests/v1/core/test_scheduler.py::test_finish_request). Syncingvllm-ascend-husttomainunblocks scheduler-level verification for everyone.ShuhaoZhangTony commentedon Nov 8, 2026
2026-11-08 maintainer closure: this migration-era umbrella is superseded by capability-owned repositories and current-main PRs. Prefix routing, lifecycle/KV observability, Unified Communication, scheduling policies, and other seams now have their own implementation/evidence ledgers; several accepted minimal Host contracts are already in main. Unimplemented legacy branches are not approved by this closure and must never be merged wholesale. Any remaining need must be filed as a narrow current-main issue from the owning MOD with tests, default-off semantics, and rollback.