Repository navigation
Description
Current state: Issue 原始整体状态仍记录 Proposed;Batch 1 已有下述范围与分工同意记录。@Rohank143 的盘点提案现已按后续澄清把申请的 Batch 1 明确限定为 ACLGraph capture/replay、模型与 checkpoint adaptation,以及对应 owner-local metadata / index 更新。作者将 operator-tuning 从缺失改为已有 candidate,并排除本 slice;host KV offload 仅复用交叉引用。不同历史来源 pins 不再被表述为已验证兼容栈,检索案例明确尚未按 #255 执行。继 Dong1017 征询分工意见后,@jiangkuaixue123 已明确同意将上述窄 Batch 1 分给 @Rohank143,该回复当时保留来源 readset、验收负责人和报酬条件待确认。该回复是范围/分工同意,不建立支付条款,也不是交付验收。@Rohank143 又询问这是否为付费任务,并要求在开工前明确预算或奖励、支付条款与验收里程碑。Dong1017 已公开答复:本项目为无偿开源贡献,无预算、奖励或 bounty,Batch 1 也是如此;作者可以自由选择不继续。@Rohank143 已明确接受无偿贡献并愿意继续既定 Batch 1,但要求实施前确认两个方向的完整来源 readset、固定版本及验收负责人/完成标准。后续只读核对列出 ACLGraph 在 8a8e1e05 的三个既有来源文件;model/checkpoint adaptation 的历史 pin 为 bc8115ed,但仍无完整 bundled owner readset。这些是起点记录,不是新选定 readset 或已验证兼容栈;已请 @jiangkuaixue123 明确保留或重定各 owner 的版本、完整 adaptation readset 和验收 reviewer。原完成项未勾选,上述窄 Batch 1 仍未见实现产物或检索验证结果;下方原始 Ownership 文字和 GitHub assignee 状态未由本同步改写。另有 @jiangkuaixue123 提交的独立 RC 性能历史蒸馏 PR #354,合并前头为 aeb267fb5cefa1139efb2ebac2dcc2220ded5bf5,已于 2026-11-05 06:54 UTC 合入 main,squash commit 为 6fefa0108232fcf210aa7c0f2ba3fa936439bd02;作者明确将其与上述历史 ACLGraph/checkpoint-adaptation 分工区分。该 PR 现提供模型性能、融合、通信配置及历史来源筛查资源,仍为 candidate,未执行 NPU 数值/服务性能或 held-out Agent efficacy 验证。合并前 Ready CI 及 knowledge policy 已成功。此前 5360d88e 的 PASS review 已因新提交而 dismissed,不能作为当前扩大内容的批准;其当时的索引投影疑问已由 reviewer 按生成投影回归测试处置。Amy 已在 aeb267fb 提交 APPROVED / LGTM,记录其前次四项问题已解决,包括来源锚点、完整历史清单移出运行包、readset 反向链接及已安装资源覆盖。该审批未验证 NPU 数值/性能,也未在真实上游 clone 上执行锚点检查。Dong1017 的当前头限定复审已记录上述修复及 exact-head source/index receipt 无剩余阻断,明确只是非投票评论,不是正式 approval,也未验证完整上游历史或设备/Agent 效果。其后 Dong1017 已提交正式 APPROVED 记录,PR 已合并;合并后的 main lane、push declarations 及后续 declarations 均已通过。合并的是此独立候选知识切片,不代表上述窄 Batch 1 或父 Issue 完成,也不证明检索、NPU 或 Agent 效果。本同步复用已有评审、合并和托管检查记录,没有重复审查或接受父 Issue。
Decision needed: @Rohank143 已确认愿意无偿继续;待 @jiangkuaixue123 / 相关维护者确认各 owner 保留或重定的来源版本、完整 readset、验收 reviewer 及既有完成标准,并补齐责任记录。
Next action: 保留已确认的窄 Batch 1 分工,待 readset/验收安排明确后再推进该路径;分别跟踪已合并 #354 的候选知识验收边界与仍未验证的效果,不接管任一负责人。分工同意、源码蒸馏和软件 CI 不构成父 Issue 完成、检索收益或真实运行效果。
Goal and acceptance
Problem / desired outcome: 将 vLLM Ascend 的源码、官方文档及有价值的公开 PR/issue 中的经验进一步蒸馏为细粒度知识,方便优化迭代按具体问题找到适用条件、操作步骤、验证方法与失败边界,而非停留在项目简介或宽泛经验。
Completion criteria:
- 产出已有知识盘点及主题覆盖矩阵,区分已覆盖、待细化、缺失、过时和暂不纳入,并给出条目与来源链接。
- 首批主题从插件与模型适配、算子与融合、图模式、调度与 KV cache、分布式通信、量化、性能定位及常见故障中选定;根据实际来源确认范围,不要求无依据地铺满全部主题。
- 将选定主题拆成可独立检索和复用的知识单元,写明问题/触发条件、适用软硬件与版本、前置条件、具体步骤、预期产物、验证方法及失败/停止条件。
- 为关键结论保留可定位的源码、文档或 PR/issue 来源及不可变版本引用;区分源码事实、上游报告、推断建议和本地实测,不把未验证建议写成已验证优化。
- 明确 vLLM 通用机制、vLLM Ascend 插件特性与 Ascend/CANN/torch_npu 底层约束的归属,通过交叉引用复用已有内容,避免复制和相互冲突。
- 整理重复、冲突和过时条目,保留必要的版本差异与适用边界;不得将某型号 NPU 或某模型的经验无依据推广到所有环境。
- 更新既有索引、metadata 和覆盖矩阵,保持既有知识库契约;知识库正文使用英文,来源及辅助资料放在所属 owner 的资源目录。
- 提供代表性检索案例与预期命中条目,覆盖适用和不适用场景;完成相关结构、引用及索引测试,清楚区分内容交付、检索验证与真实运行效果。
Scope and constraints
- In scope: vLLM Ascend 相关来源盘点、细粒度蒸馏、分类归属、去重、版本/适用条件整理及检索案例。
- Out of scope: 全库目录迁移、直接修改 vLLM Ascend 实现、默认启动大规模硬件实验,以及未经实验支持的性能收益声明。
- Interfaces / limits: 遵循 docs/knowledge-base.md 与 docs/knowledge-model-policy.md;沿用现有 canonical owners 和知识准入流程。优先公开来源,不纳入凭据、私有地址或未经授权的内部记录。
Ownership and authorization
- Proposed by / actual requester: 用户在 Codex 对话中提出,Codex 代为整理提交。
- Status: Proposed
- Accountable parent / implementer / goal-acceptance owner: TBD,由维护者指定。
- Accepted decision or scoped delegation: 本次授权为提交需求 issue;未指定实施、合并或关闭权限。
Inputs and deliverables
- Sources: https://github.com/vllm-project/vllm-ascend 及其官方文档、关联公开 PR/issue;实施时冻结具体 readset 和来源版本。
- Related issues: [Tracking] SearchSpace distillation coverage and knowledge consolidation #91(知识蒸馏与 consolidation)、task: import Qwen3-TTS Ascend optimization techniques and evidence #35(已有 Qwen3-TTS Ascend 经验)、[Tracking] Skill Indexer: retrieval efficiency and accuracy #255(检索质量)、[Tracking] SkillEval evolution: local evidence, PR review and optional LLM judge #253(知识实际效果)、task: 支持每个优化迭代的知识库命中归因统计 #344(每轮知识命中归因)。本单聚焦 vLLM Ascend 内容深化,不替代这些任务的独立验收。
- Expected artifacts: 来源/覆盖清单、细化知识条目及配套资源、索引与 metadata 更新、代表性检索案例、验证报告及剩余缺口说明。
Completion and closure
- Who accepts completion and may close: 待指定的目标验收负责人。
- Remaining uncertainty / next decision: 首批主题与来源 readset、分批交付边界及负责人;未验证硬件范围须在交付中明确保留。
Activity
amy-why-3459 commentedon Nov 4, 2026
感谢整理,方向很清楚。补充几点建议,方便确定首批范围:
hardware/ascend/ascend-aclgraph-capture-replay、ascend-model-adaptation、ascend-optimization-casebook、torchnpu-device-memory-lifecycle/torchnpu-pinned-host-memory-lifecycle,以及hardware/host/host-memory-management/references/vllm-ascend-native-offload*,还有docs/reviews/issue91-vllm-ascend-host-offload.md。建议覆盖矩阵直接以这些条目和docs/reference-matrix.md、docs/skill-coverage.md为起点,标出待细化、重复和缺口的部分,避免另起一套。另外,issue 里提到需要维护者确认负责人,这一点还需要维护者来定。
Rohank143 commentedon Nov 4, 2026
📋 First-Batch Knowledge Inventory & Topic Coverage Proposal (Refs #350)
In response to the task proposal and the guidance provided by @amy-why-3459, here is the structured First-Batch Knowledge Inventory & Coverage Matrix addressing Completion Criteria 1–5:
1. Inventory of Existing vLLM Ascend Knowledge Assets in Repository
Starting from
docs/reference-matrix.md,docs/skill-coverage.md, and canonical skill directories:hardware/ascend/ascend-aclgraph-capture-replay:8a8e1e05ecb64a30804c99c8f362909f9a0cb5bf(compilation/acl_graph.py,test_acl_graph.py).ascend-graph-mode.hardware/ascend/ascend-model-adaptation:bc8115ed7fdc7e500320b5de70742ef049b4c1a2(.agents/skills/vllm-ascend-model-adapter/).hardware/host/host-memory-management(references/vllm-ascend-native-offload*):8a8e1e05ecb64a30804c99c8f362909f9a0cb5bf(7-file ledger, Apache-2.0 license).vllm-ascend-native-offload.mdserves as the authoritative pattern without rewriting.hardware/ascend/torchnpu-device-memory-lifecycle&torchnpu-pinned-host-memory-lifecycle:67511c53cd4c88014f6123de7097677e60e85faa(TorchNPU).hardware/ascend/ascend-operator-tuning:skill-index.json(status:candidate,empirical_validation: not_run,provenance_readsets: []). Outside Batch 1 requested slice.hardware/ascend/ascend-optimization-casebook&ascend-graph-mode:2. Topic Coverage & Status Matrix (Completion Criterion 1)
hardware/ascend/ascend-aclgraph-capture-replayvllm-ascend:compilation/acl_graph.pyhardware/ascend/ascend-model-adaptationvllm-ascend:models/hardware/host/host-memory-managementreferences/vllm-ascend-native-offload.mdhardware/ascend/ascend-operator-tuninghardware/ascend/ascend-distributed-parallelhardware/ascend/ascend-optimization-casebook3. Three-Tier Responsibility & Attribution Boundaries (Completion Criterion 5)
To prevent duplication and lexical conflicts:
repositories/vllm/): PagedAttention block abstraction, scheduler preemption, worker RPC protocol. (Does not mention NPU or CANN specifics).hardware/ascend/&hardware/host/): ACLGraph model runner, Ascend custom op dispatch, native host KV offload descriptor pool, paired weight converter.hardware/ascend/):aclmdlExecute, geExecutor, HCCL mesh scratch, TorchNPU caching allocator, H2D streams. (Does not assume vLLM context).4. Historical Source Anchors (Independent, Not a Tested Compatibility Set, Criterion 4)
vllm-project/vllm-ascend@8a8e1e05ecb64a30804c99c8f362909f9a0cb5bf(with historical documentation pinbc8115ed7fdc7e500320b5de70742ef049b4c1a2).vllm-project/vllm@ tagv0.6.3/v0.6.4.torch_npu@v2.4.0(67511c53cd4c88014f6123de7097677e60e85faa).(These are separate historical anchors and are not claimed as a single verified compatible stack until verified against an exact matching readset).
5. Proposed Retrieval Test Cases (Not Yet Executed Against #255 Evaluation Protocol)
hardware/ascend/ascend-aclgraph-capture-replay(Negative:ascend-model-adaptation).hardware/host/host-memory-management(Negative:torchnpu-device-memory-lifecycle).hardware/ascend/torchnpu-device-memory-lifecycle(Negative:ascend-aclgraph-capture-replay).Proposed Next Step & Precise Batch 1 Scope
To keep the delegation request exact:
ascend-aclgraph-capture-replay(distinct fromascend-graph-mode).ascend-model-adaptation.Host KV offload remains an existing cross-reference, and Operator tuning remains outside this slice unless maintainers request otherwise.
Could repository maintainers / @Dong1017 please confirm the exact Batch 1 boundary and authorize implementation delegation for this slice?
Rohank143 commentedon Nov 4, 2026
Thanks for the guidance. While checking the current canonical skill library, I found that
ascend-operator-tuningalready exists and is indexed. I'll correct its coverage status to distinguish an existing candidate from a missing knowledge area.To keep the proposed delegation precise, my requested Batch 1 implementation scope is:
ascend-aclgraph-capture-replayowner (kept distinct from the separateascend-graph-modeowner).ascend-model-adaptationowner.Host KV offload will remain an existing cross-reference (
hardware/host/host-memory-management/references/vllm-ascend-native-offload.md) unless maintainers request additional refinement. Operator tuning will remain outside this requested implementation slice.I will also keep the historical source pins separate (e.g. ACLGraph @
8a8e1e05, Model Adaptation @bc8115ed) and avoid claiming a verified compatible software stack until the exact readset and version combination are confirmed. The retrieval examples are proposed cases and have not yet been executed against the #255 evaluation protocol.Could maintainers please confirm the exact Batch 1 boundary, canonical source readset, and implementation and acceptance owners before I start implementation?
Dong1017 commentedon Nov 4, 2026
Hi @jiangkuaixue123, can you share some insights when you have a chance? Do you agree to assign parts of work to @Rohank143?
Rohank143 commentedon Nov 4, 2026
I'm interested in taking responsibility for the proposed Batch 1 scope (ACLGraph refinement, Model Adaptation refinement, and metadata updates).
Before starting implementation, could you please confirm whether this is a paid assignment?
If yes, please clarify the approved budget or expected reward, payment terms, and acceptance milestones.
I'm happy to proceed once the scope, ownership, and compensation are mutually agreed upon.
Thank you.
jiangkuaixue123 commentedon Nov 4, 2026
@Dong1017 Yes, I agree to assign the proposed Batch 1 work to @Rohank143: ACLGraph capture/replay refinement, model and checkpoint adaptation refinement, and the corresponding owner-local metadata and index updates, as scoped in the clarification above.
@Rohank143 Thanks for taking this on. The source readset and acceptance owner still need to be confirmed. Your question about compensation also remains to be resolved separately; this scope approval does not establish payment terms.
Dong1017 commentedon Nov 4, 2026
Answering the compensation point @jiangkuaixue123 left open: this is an unpaid open-source project. InferRSI does not pay for contributions. There is no budget, no reward and no bounty program behind any Issue here, including this one. The scope approval above assigns work; it does not create payment terms, and no payment terms are coming.
Thank you for asking before starting rather than after. That is the right order, and it means you can decide with the answer in hand. Taking Batch 1 is entirely optional — declining it now costs you nothing and counts against you in no way.
If you do continue, the usual open-source terms apply: your commits keep your authorship and the Issue record links your work. I have given the same answer on #328 and #349, where the question about where a bounty was advertised is still open.
Rohank143 commentedon Nov 5, 2026
Thank you @Dong1017 @jiangkuaixue123 for clarifying the compensation terms.
I understand and accept that this is an unpaid open-source contribution. I would still like to proceed with the approved Batch 1 scope.
I confirm my willingness to work on:
Before implementation, could you please confirm:
Once these are confirmed, I'll proceed with the implementation and submit a PR linked to Issue #350.
Thank you for the opportunity.
Dong1017 commentedon Nov 5, 2026
@Rohank143 Thanks for confirming. I checked the existing owners at
main9dded7ad. There is a concrete difference in the retained source records:8a8e1e05ecb64a30804c99c8f362909f9a0cb5bf:vllm_ascend/compilation/acl_graph.py,docs/source/user_guide/feature_guide/graph_mode.md, andtests/ut/compilation/test_acl_graph.py, with blob hashes and the inspected scope.bc8115ed7fdc7e500320b5de70742ef049b4c1a2, but explicitly records no bundled owner readset. The existing entry citesfp8-on-npu-lessons.mdandmultimodal-ep-aclgraph-lessons.mdunder.agents/skills/vllm-ascend-model-adapter/references/. Resolving that commit does not establish that the proposed implementation files have been inspected.These are starting references, not a newly selected Batch 1 readset or a verified compatible stack. The independent RC-history work in #354 does not automatically replace either owner's historical source boundary.
@jiangkuaixue123, could you confirm whether Batch 1 should refine these retained pins or explicitly re-pin each owner, the complete model-adaptation readset, and the acceptance reviewer against this Issue's existing completion criteria? That would answer the remaining pre-implementation questions without expanding the agreed slice.
Read-only source-record check; no upstream semantic re-audit, tests or device runs in this reply.
Rohank143 commentedon Nov 10, 2026
Thanks @Dong1017 and @jiangkuaixue123 for the guidance.
I have implemented and submitted the candidate Batch 1 slice in PR #382:
ascend-aclgraph-capture-replay: Retains historical pin8a8e1e05ecb64a30804c99c8f362909f9a0cb5bf. Refines dynamic shape bucketing (cudagraph_capture_sizes), multi-stream synchronization rules, positional address checking limitations, and explicit Ascend error classifications (e.g.alloc sq cq fail/207008).ascend-model-adaptation: Retains historical documentation pinbc8115ed7fdc7e500320b5de70742ef049b4c1a2and resolves the missing bundled readset by addingreferences/source-notes.mdcitingfp8-on-npu-lessons.mdandmultimodal-ep-aclgraph-lessons.mdwith explicit 3-tier responsibility routing (vLLM generic engine vs. vLLM Ascend plugin vs. Ascend/CANN/torch_npu layer), FP8 load-time dequantization formula, and TP GQA KV-head replication.host-memory-managementremains an existing cross-reference,ascend-operator-tuningremains outside this slice,docs/skill-coverage.mdwas regenerated and verified with--check, and a dedicated test suite (tests/test_vllm_ascend_distill_skills.py) verifies discoverability and metadata invariants.PR: #382
cursor commentedon Nov 10, 2026
PR #382 is recorded as the candidate narrow Batch 1 implementation. Issue #350 stays open. Scope assignment is already agreed; delivery acceptance is still outstanding.
Aligned with the agreed slice (author report on #382, not a diff review):
ascend-aclgraph-capture-replaykeeps historical pin8a8e1e05ecb64a30804c99c8f362909f9a0cb5bf(the three-file ledger already on main:acl_graph.py,graph_mode.md,test_acl_graph.py).ascend-model-adaptationkeeps historical pinbc8115ed7fdc7e500320b5de70742ef049b4c1a2and addsreferences/source-notes.mdfor the previously missing bundled readset (fp8-on-npu-lessons.md,multimodal-ep-aclgraph-lessons.md), plus the 3-tier routing note.host-memory-managementstays a cross-reference;ascend-operator-tuningstays outside this slice.docs/skill-coverage.mdregenerated with--check, plustests/test_vllm_ascend_distill_skills.pyfor discoverability and metadata invariants.Still open on #350:
6fefa010) is a separate candidate slice and does not complete task: 细化 vLLM Ascend 知识蒸馏与知识库整理 #350.Next step is review on PR #382. Closing #350 waits on the readset/reviewer confirmation and the existing completion criteria. This reply does not review the PR diff.
— Dong-bot
feat: refine vLLM Ascend ACLGraph and model adaptation knowledge (Refs …
feat: refine vLLM Ascend ACLGraph and model adaptation knowledge (Refs …
knowledge: regenerate skill index for vLLM Ascend skills (Refs InferM…
feat(ascend): refine vLLM Ascend ACLGraph and model adaptation knowle…
knowledge: regenerate skill index for vLLM Ascend skills (Refs InferM…
feat(ascend): refine vLLM Ascend ACLGraph and model adaptation knowle…
knowledge: regenerate skill index for vLLM Ascend skills (Refs InferM…
feat(ascend): refine vLLM Ascend ACLGraph and model adaptation knowle…
knowledge: regenerate skill index for vLLM Ascend skills (Refs InferM…
Dong1017 commentedon Nov 17, 2026
Posted by an agent on behalf of @Dong1017 (incremental only; points already raised by others or decided by the owner are not repeated).
Incremental note: the Batch 1 PR #382 (
Refs #350), last recorded here as a "candidate", was merged on 2026-11-16 11:20 UTC as6c7e2379, approved by @Dong1017 andcursor[bot]. The status block still says delivery acceptance is outstanding, so Batch 1 acceptance (or remaining gaps) should be recorded here. The other inventory items stay outside Batch 1 as already agreed.Scope: PR state and approvals via the GitHub API. I did not re-review the merged content.
zengchuang-hw commentedon Nov 17, 2026
Independent scope/readiness review of #350 against InferRSI main
61303fa8150472dfbda62ec8d77908bc05c73f3b.The distillation goal is useful, but the starting point is no longer an empty library: the historical model-adaptation and ACLGraph owners coexist with a release-specific
repositories/vllm-ascendowner. I would use an inventory/delta acceptance contract, not count existing entries as newly delivered work or infer whole-issue completion from a small passing test subset.Concrete correction needed in the current baseline: ascend-model-adaptation/SKILL.md lines 112–122 requires exactly four feature statuses and forbids “not tested.” None represents an applicable feature with incomplete evidence, a blocked test, or an observed failure whose cause is unresolved. For example, if real-weight loading succeeds but multimodal execution has not run, “verified,” “framework unsupported,” “checkpoint missing,” and “N/A” are all unsupported conclusions. Preserve an explicit unverified/blocked state and evidence/reason (or separate applicability, support and validation fields); add a case that proves partial execution does not force a false categorical claim. This is a content-contract gap, not an observed NPU runtime defect.
Retrieval evidence needs a negative/version-boundary dimension. The five distillation queries prove expected positive top-5 hits. The RC tests add positive routes, source/metadata and history-helper checks. They do not establish refusal or correct interpretation when hardware, checkpoint or version is outside a procedure's evidence. The catalog reader filters declared fields and token occurrences, not runtime compatibility. Add cases with expected owner/source revision and expected refusal/missing-evidence output, including an A2 deployment presented to the RC procedure and an untested multimodal checkpoint presented to the historical adaptation procedure. Keep discoverability separate from applicability and downstream efficacy; this does not require changing the engine or retrieval architecture.
Suggested bounded completion record: for each selected topic, record covered/refine/missing/stale/excluded, existing owner and resource, exact upstream readset, supported stack boundary, proposed delta, positive and inapplicable cases, and validation result. Preserve separate historical adaptation, runtime-wrapper and RC pins rather than combining them into a purportedly tested stack. Reuse the existing generic-vLLM/plugin/CANN routing and link across owners rather than duplicate mechanisms. Content and matrix/index checks, retrieval usefulness, and native runtime effects should have separate outcomes.
Direct source spot-check: I read upstream vllm_ascend/init.py at 3b31886237ed65c435e0965b10e8002ff5769766: connector, model-loader and profiling hooks call the process-local global-patch guard, while platform registration returns the platform class string. This supports the current plugin-registration procedure's distinction between distribution discovery and engine-process initialization; it is not proof of a working installed deployment.
Checks actually run on the pinned InferRSI checkout:
PYTHONPATH=src python -m pytest tests/test_vllm_ascend_distill_skills.py tests/test_vllm_ascend_rc_skill.py -o addopts='' -qusing an existing test environment: 20 passed. The default interpreter lacked pytest; the reported successful run used the existing environment. I also inspected the two knowledge contracts, relevant owner procedures/source notes, matrix entries, and catalog search implementation. No full-suite/CI validation, exhaustive upstream-source audit, NPU/model/service experiment or held-out efficacy evaluation was performed. No code changes or completion/implementation commitment are implied by this review.Rohank143 commentedon Nov 17, 2026
Incremental progress update on Batch 1 acceptance and technical remediation (Refs #350, follow-up to Batch 1 PR #382 merged as
6c7e2379):Thank you @Dong1017 and @zengchuang-hw for the review and clear acceptance contract.
In response to @zengchuang-hw's review, we have addressed the two concrete contract gaps identified and submitted PR #482 (
feat: permit unverified feature status and add boundary retrieval tests (Refs #350)):1. Technical Remediation Delivered in PR #482
⏳ unverified / blockedfeature status inascend-model-adaptation/SKILL.md:⏳ unverified / blockedstatus, requiring documented evidence or blocker reasons (e.g. real-weight loading succeeded, but multimodal visual execution is unverified due to lack of NPU execution).verified,framework unsupported,checkpoint missing, orN/A).tests/test_vllm_ascend_distill_skills.py:test_model_adaptation_unverified_feature_status_contract: Validates that presenting an untested multimodal checkpoint triggers an explicit refusal point (Stop on unverified multimodal visual encoder output) rather than forcing categorical claims.test_distillation_skills_negative_retrieval_and_scope_refusal: Validates catalog query isolation and refusal against out-of-scope topics (operator kernel tuning assembly), cross-backend refusal (metal graph capture), and refusal points for untested architectures.docs/skill-coverage.mdviapython tools/build_skill_coverage.py(verified with--check)._provenance.jsonviapython tools/build_skill_index.py --source-revision 6282cc181825a76b053ffa15559866f0e9bf97b8(verified with--check).2. Bounded Completion & Inventory Acceptance Record
Per @zengchuang-hw's suggestion, below is the inventory/delta acceptance contract separating historical pins, candidate release pins, and excluded topics:
hardware/ascend/ascend-aclgraph-capture-replay/8a8e1e05ecb64a30804c99c8f362909f9a0cb5bf(acl_graph.py,graph_mode.md,test_acl_graph.py)tests/test_vllm_ascend_distill_skills.py(positive ACLGraph query match)hardware/ascend/ascend-model-adaptation/bc8115ed7fdc7e500320b5de70742ef049b4c1a2(SKILL.md,references/source-notes.md)⏳ unverified / blockedstate with blocker reason; added bundled readset notestest_model_adaptation_unverified_feature_status_contract(untested multimodal checkpoint yields refusal point)repositories/vllm-ascend/vllm-ascend-release-candidate/aeb267fb5cefa1139efb2ebac2dcc2220ded5bf5(merged in #354)tests/test_vllm_ascend_rc_skill.py(15 tests passing)operator kernel tuning assemblycorrectly excluded3. Separation of Pins & Verification Scope
8a8e1e05for ACLGraph,bc8115edfor adaptation) and release-candidate pin (aeb267fb) remain strictly separate and are not combined into a single untested monolithic stack.PYTHONPATH=src python -m pytest tests/test_vllm_ascend_distill_skills.py tests/test_vllm_ascend_rc_skill.py tests/test_skill_index.py tests/test_skill_coverage_view.py: 49 passed in 96.62s.python tools/build_skill_coverage.py --check: Passed.python tools/build_skill_index.py --source-revision 6282cc181825a76b053ffa15559866f0e9bf97b8 --check: Passed (80 metadata records).python -m ruff check: Clean (0 errors).fix(knowledge): permit unverified feature status and add boundary ret…
zengchuang-hw commentedon Nov 17, 2026
回应直接审查请求 6944287419(更新时间 2026-11-17T09:37:33Z)。
限定结论:当前存在可定位、已注册的知识交付材料,但本次证据不足以确认 #350 整体验收完成。 本评论明确回答该请求;不把本人更早的评论视为自动覆盖。需求依据是隔离提取的原始 Goal/Scope/AC,不是该请求正文、其他评审、作者转述或 PR482 的实现/测试声明。
本次检查固定于 main
a02ef1863e29de4510272e65ae69341b1ef83f22:repositories/vllm-ascend与hardware/ascend/ascend-model-adaptation,二者均在 skill-index.json 注册,metadata 状态均为 candidate。前者覆盖插件注册、算子融合、通信、模型性能等分主题资源;后者的章节覆盖版本职责、checkpoint 调试、FP8、KV replication、graph、multimodal、feature-status 等。这里只确认材料组织,不据标题宣称对应技术知识正确或完整。3b31886237ed65c435e0965b10e8002ff5769766、明确 paths 与references/sources.mdreadset;其 Markdown 内检测到 159 个不同的完整 SHA GitHub blob/tree 链接。这证明有可定位的引用基础,不等于这些引用已逐项读证并支持每条主张。我未重新获取该 upstream revision,也不以引用数量衡量知识质量。_skill_documents把 skill/metadata 构造成可检索条目;_why_unreadable检查 read/material;现有 Bundle.manifest 仍要求 entrypoint、声明资源及依赖摘要闭合。因此“已注册”“文本有命中”“读入口成立”“来源主张有依据”“设备上有效”必须分开验收。针对原始目标,我的验收建议是:
实做范围:元数据/本人结果去重、固定 main 文件树与 index/metadata、上述两组 Markdown 相对目标静态核对、检索与读取边界 AST 检查;已读 AGENTS/review skill 无相关变化。未做:整个知识库逐条语义审查、上游引用逐项验证、实际查询回归、pytest/index生成器执行、硬件有效性或性能实验。未改代码、合并、关闭 Issue、调用付费模型/设备或恢复停止任务。单个 PR 的技术或状态修订也不能代替本 Issue 的完整交付验收。