[Dong1017]

Refs #91.

Adds the bounded vLLM-Ascend single CPU-tier KV-offload case to the existing host-memory owner. It distinguishes submission, tracked-job wait and normal descriptor-pool reclamation: pointer/size arrays return to the normal pool only after a successful queue-head end-event query. The four existing mechanisms and CUDA serving example remain intact. CPU dispatch still submits NPU DMA. Shutdown's pop-before-synchronize source order establishes neither retention throughout that wait nor a native safety/bug result.

The carried source is vllm-project/vllm-ascend at 8a8e1e05ecb64a30804c99c8f362909f9a0cb5bf. The seven-file immutable ledger, original 2026-10-23 inspection dates/readsets, complete 11,357-byte Apache-2.0 license, contributor/Huawei attribution and candidate / empirical_validation=not_run state are preserved. External vLLM manager/helper/ABI compatibility remains unpinned; other offload, mmap/tiering, request retirement and TorchNPU allocator behavior remain outside the case. This integration adds no new upstream inspection or runtime qualification.

This update reconciles prior PR head 9a65f7c45ac867112e6819f5dee6195684637b26 with main 65147a03e465dc838770474736a620e04d72337e, preserving both histories in actual combined source 1631c0cac62e4f0751afefa530c0f64811edfb6f. The existing eight-path canonical delta retains main's SGLang navigation, model-adapter routing, version-scoped OS qualification, authorship invalidation and opt-in selection/context. Both matrices change only the scoped host-memory row; all six nonmatrix source files retain their reviewed bytes. Source checks resolved 218 local links and passed staged whitespace.

Current checks on this combined content:

• The unchanged stock generator and matching --check passed from 1631c0cac62e4f0751afefa530c0f64811edfb6f, producing 59 records plus README. The other 58 records only refresh canonical revision/URL. The host record retains its approved semantics with current six canonical hashes/provenance; every nongenerated source entry remains unchanged in the generated child.
• All 257 indexed canonical file hashes match committed Git bytes. The registry remains 59 IDs and the combined catalog 104 records.
• Separate local processes using the actual normalized checkout reader and worktree package adapter read all five owner resources at exact revisions. The package process imports no YAML. This validates the package protocol; it is not a clean installed-wheel run.
• Actual opt-in API and CLI context reads return four complete required resources, including the license, within 49,803 bytes at the exact bundle revision, with no required-resource omissions. Without a context-byte request, Collect omits the context view.
• Generated tree: 4c6528262615d0e1614da8e3e5ec5552525ce990. Final head: d22deb3cac385e2be7f3e859426bfd47312f1c7c. All four Ready CI jobs passed at this exact head, including full checks and clean installed packages on Python 3.10/3.11. Independent final integration review found no actionable findings. Earlier source reviews and software results remain applicable only at their recorded revisions; no local full or index suite was rerun for this integration.

Native imports/tests, NPU DMA, numerical output, teardown injection, physical RAM, model/service quality, performance and agent efficacy were not run in this update. No dependencies, mocks or substitute native implementations were added. The candidate remains locally unverified; proposed cases remain proposals, and tracking-issue acceptance remains outstanding.

{
  "schema_version": 1,
  "head_sha": "d22deb3cac385e2be7f3e859426bfd47312f1c7c",
  "human_only": false,
  "material_ai_authors": [
    {"tool": "Codex", "model_id": "gpt-6.1-sol", "effort": "xhigh", "role": "author"}
  ]
}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-24T16:37:26.111628Z	d22deb3	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[Dong1017]

LGTM for the bounded source-derived candidate scope at 4a42781173ed24599bda4c5465fd0ab7f57fbc9d (base 48e0d5894117be7c57f8b4ce793c0ec986f219e8). No actionable findings.

The pinned Python worker and C++ DMA consumer support the documented submission, wait and normal-reclamation distinctions: descriptor ownership remains with the queued transfer; successful queue-head event query gates normal recycling; wait synchronizes without reclaiming. The shutdown statement correctly remains limited rather than claiming proven safety or a native bug. Existing mechanisms remain intact, and external vLLM/CANN ABI/runtime qualification is explicitly excluded. Source commit 2b2320ccbb0011697f1c49e62b22fe819f82eb53 is the parent of the projection-only refresh; unrelated generated records change only source URL/commit.

Ready CI: all four full-check/clean installed-package jobs passed on Python 3.10/3.11; current knowledge-model-policy status is successful. These are inspected GitHub CI results, not tests rerun here.

Verification was read-only source, diff, packaging-rule and projection inspection, not an independent recomputation of every source hash or native/hardware/model/agent-efficacy execution. Parallel knowledge contributions share generated discovery records: preserve all canonical changes and regenerate from the combined source commit during integration. This is a future integration requirement, not a present inventory defect. #91 remains a living tracker; LGTM is not empirical qualification or tracker acceptance.

[Dong1017]

LGTM for the bounded integration delta at d22deb3cac385e2be7f3e859426bfd47312f1c7c, reported base 65147a03e465dc838770474736a620e04d72337e. No actionable findings beyond the stated verification limits.

All five host-memory owner files, including source ledger and license, retain reviewed Git blob identities. Current-main SGLang/Ascend routing and other canonical content remain intact outside scoped matrix changes. Normal reclamation versus wait and the bounded shutdown caveat are unchanged; runtime, generator and tests match main.

Combined source 1631c0cac62e4f0751afefa530c0f64811edfb6f → head changes only generated files: other 58 records change only provenance; the host record retains reviewed semantics. GitHub reports mergeable.

All four exact-head Ready full-check/clean-wheel jobs pass on Python 3.10/3.11; per-head knowledge-model-policy succeeds. This does not clear the separate aggregate/history failure tracked in #140.

Read-only prior-review/delta/Git-identity/projection/test-diff and CI inspection. No PR code, tests, generator or native/model experiments executed, no exhaustive source-hash recomputation and no new backend/quality/performance/agent-efficacy qualification. Previous candidate/not_run and hardware-coverage limitations remain; issue acceptance and merging are separate.

[Dong1017]

Squash-merged under the owner's explicit batch instruction. Head d22deb3; merge a8e4c57.

Evidence verified before merge:

• Scoped LGTM: feat: distill Ascend offload descriptor ownership #123 (review)
• Ready CI: https://github.com/InferMatrixLab/InferRSI/actions/runs/40112304158 (four full-check/installed-package jobs passed).
• Required knowledge-model-policy SUCCESS on this exact head after trusted-main rescan: https://github.com/InferMatrixLab/InferRSI/actions/runs/40262924356/job/121378950476. The aggregate workflow may fail on other PRs; this does not claim it is globally green.
• Fresh conflict-free metadata and expected-head SHA guard. No bypass.

The current human merge instruction proceeds on scoped review, Ready tests and the actual required policy, not an assertion that every historical declarations job is green. Software/knowledge integration only; no Issue/campaign acceptance or model/hardware efficacy claim.