Skip to content

[Task] Distill Flare preparation contracts and session-state lifetimes #148

Closed
Closed
[Task] Distill Flare preparation contracts and session-state lifetimes#148
Assignees
Dong1017
Labels
area:knowledgeKnowledge line: packs, skill library, knowledge admissionkind:taskBounded delivery task with a delegated slice

Description

@Dong1017
Dong1017
opened on Oct 25, 2026

Accepted bounded source delivery on 2026-10-29 (Asia/Shanghai): PR #217, squash 509531c681479fbdd7ddd0e8b0032239e64a44c1. Full acceptance card. Source-derived candidate / empirical not_run; model, device, service and efficacy remain unexecuted.

目标

将 NVIDIA/flare-local-inference 的准备产物边界与完整会话状态生命周期蒸馏为 InferRSI 的有来源、可检索知识。Parent: #91；相关工程 RFC: #147。

这是一项源码蒸馏任务，不是 #147 的证据适配器实现。 不运行模型/GPU、不移植 Helios、不宣称性能提升；本 issue 创建仅登记任务，尚未完成提取。

冻结范围

初始源码基线：bce032c014bae34502ba22d3304fa084d5b40afe（2026-10-25）。若实施时更新版本，须显式记录新旧差异，不能混用 ABI/状态语义。

切片 必读源码 / 文档 要回答的问题
准备与服务边界 README.md、docs/runtime-package-design.md、helios/src/runtime_package/format.rs Python准备、Rust导出/服务分别拥有何物？execution ABI 4 / prepared contract 2 如何验证？哪些句柄不能跨进程？
完整状态与资源生命周期 helios/src/cache.rs、helios/src/offload/mod.rs、docs/prefix-cache-offload.md KV page chain 与 recurrent/convolution/partial-tail 状态如何配套？publish/use/retire fence、transfer ticket 与取消 waiter 有何区别？
反例与可观测性 tests/server/test_prefix_cache_offload.py、qualify/performance.py、qualify/provenance.py restart/namespace/corrupt-record 如何防止部分恢复？lookup hit、admitted、实际 reused tokens 如何区分？客户端与 native 时钟为何不可混算？

源码链接统一从固定版本树定位。上述为初始 readset，不把文档描述或测试存在写成本地验证。

flowchart LR
    A["固定源码 + 测试/文档"] --> B["机制、前提、反例与限制"]
    B --> C["既有知识 owner 优先复用"]
    C --> D["有 pin/readset 的 candidate 知识"]
    D --> E["索引/打包检查"]
    E --> F["Tracking #91：有界来源交付"]
    G["RFC #147：离线证据适配器"] -. "单独工程任务" .-> F
Loading

产出位置与去重

  • 先检查现有 runtime/KV lifecycle、weight-loading/model-preparation owners，记录重叠与真正新增内容。
  • Flare 特有协议可拟放 src/inferrsi/data/skill_library/repositories/flare-local-inference/audit-flare-session-lifecycle/；这是拟定路径，不是现有输出。通用准备方法优先增补现有 owner，不另建同义 skill。
  • 交付机制说明、source ledger/readset、前提与反例、适用/不适用场景、许可通知；按仓库现有 schema 注册与生成投影，不手改派生文件。
  • 更新 Reference/SearchSpace 的实际覆盖与导航。来源蒸馏不等于 repo runtime descriptor、后端支持或资源效果已验证。

验收清单

  • pin、文件/符号 readset、许可证和引用关系完整；根 Apache-2.0 不自动覆盖模型/依赖权利。
  • 明确完整状态并非 KV-only；区分取消请求、transfer completion、publication 与安全回收。
  • 区分 NoMatch / CostRejected / CapacityRejected / EnqueueRejected / Admitted；上游成本阈值仅作为带前提的策略案例。
  • 至少给出恢复不完整、版本不匹配、恢复不划算或计时口径错误的可操作诊断案例。
  • 保留研究原型 SM120/SM121、GDS deferred、未独立复现等边界；不推广为通用 backend/多卡能力。
  • 知识检索、链接/锚点、注册、标准生成器和打包测试通过；保留实际命令与结果。
  • 新增/更新 skill 按 Skill Eval v1.1: automated evaluation for skill additions and updates #88 的已有流程准备评价计划；没有执行就标记 not evaluated / empirical not_run。
  • 回填 [Tracking] SearchSpace distillation coverage and knowledge consolidation #91 的具体输出与 PR；source-derived candidate 不自动晋级，未做模型/硬件实验不写 efficacy PASS。

非目标

不实现 #147、不运行服务 benchmark、不新建 offload backend，不将 Python→Rust 当性能因果证据。先完成有限切片，不追求全仓蒸馏。

Activity

  1. Dong1017
    added
    area:knowledgeKnowledge line: packs, skill library, knowledge admission
    on Oct 25, 2026
  2. Dong1017
    added a parent issue on Oct 25, 2026
  3. Dong1017
    added
    kind:taskBounded delivery task with a delegated slice
    on Oct 26, 2026

  4. Dong1017 commented on Oct 28, 2026

    @Dong1017
    Dong1017
    on Oct 28, 2026
    CollaboratorAuthor

    Claim and plan: Flare source knowledge

    Acting as the accountable Issue owner in the authorized Codex session, I am taking this bounded source-distillation task through delivery, independent review, integration and Issue-level acceptance. The user explicitly prioritized completing this Issue. Related engineering work in #147 and PRs #151/#155 remains separate.

    I will retain the initial NVIDIA/flare-local-inference source baseline bce032c014bae34502ba22d3304fa084d5b40afe and make one focused contribution PR:

    1. Read the preparation/package ABI, cache/offload lifecycle, server tests and qualification/provenance sources listed above; retain file/symbol references and license boundaries.
    2. Compare existing preparation, runtime and KV knowledge owners. Add the Flare-specific complete-state protocol and diagnostics under its appropriate owner, reusing existing generic routes.
    3. Cover all five restore-admission outcomes, complete KV/recurrent/convolution/tail state, cancellation versus transfer completion/publication/reclamation, and version/completeness/cost/timing counterexamples. Preserve the research-prototype and deferred-GDS limits.
    4. Register source-derived candidate knowledge, update the two coverage matrices, and generate discovery projections using the repository's native tools. Prepare the Skill Eval v1.1: automated evaluation for skill additions and updates #88 evaluation plan with empirical validation not_run and efficacy not evaluated.
    5. Verify retrieval, links/anchors, generated sources and clean installed packages; obtain independent review of the final full head and all applicable Ready CI before squash integration. Then backfill [Tracking] SearchSpace distillation coverage and knowledge consolidation #91 and check every acceptance item before posting the closure summary and closing this source-only Issue.

    The material author uses actual gpt-6.1-sol / max because current material-admission rules exclude ultra; independent review uses a separate gpt-6.1-sol / ultra instance. No model/GPU execution, upstream installation, paid evaluation, serving benchmark or runtime-adapter implementation is included.

    Next action: freeze the source readset and finalize the owner/gap map against current main before constructing the knowledge bundle.

  5. Dong1017 commented on Oct 28, 2026

    @Dong1017
    Dong1017
    on Oct 28, 2026
    CollaboratorAuthor

    Accepted: bounded Flare source knowledge

    flowchart LR
    A["Pinned NVIDIA source and license"] --> B["Existing preparation owner"]
    A --> C["New complete-session audit"]
    B --> D["Native discovery and installed reads"]
    C --> D
    D --> E["Accepted source candidate"]
    E -.-> F["Device and efficacy not run"]
    Loading
    • Delivered: PR #217, squash 509531c681479fbdd7ddd0e8b0032239e64a44c1: Flare preparation boundaries, complete continuation/KV lifecycle, five admission outcomes, actionable diagnostics, source ledger/license and an explicit old/new loading evaluation plan. Tracking #91 is backfilled and remains open.
    • Verified: Independent final-head LGTM; Ready CI has both InferRSI full suites passing 2355 tests, 11 skips and 183 subtests, plus two fresh isolated installed-package checks confirming 121 records. Static SkillEval plan/publish, lint, declaration and knowledge-model-policy pass.
    • Limits: Accepted as source-derived candidate / empirical not_run. No Flare model, GPU, server, performance or skill-effectiveness run; automatic behavioral comparisons/calls remain zero. The research SM120/SM121, deferred-GDS, multi-card and separate model/dependency-rights boundaries remain. This does not accept new [RFC] Flare beyond distillation: offline qualification evidence adapter #147 runtime behavior.
    Full source-scope acceptance record

    The accountable Issue owner accepts this bounded delivery under the existing claim and delegation.

    • Reviewed PR head: 9b413a0016dfbbaa9e73b19174988ff9c95cc02d; assessed/tested main: f054de30f881e9d523667820997e5fb33fd7b161. The actual squash tree 1daedb6c0ac8e70235c8fd570941416e3e4a3e17 equals the approved head tree.
    • NVIDIA/local-inference, formerly NVIDIA/flare-local-inference: bce032c014bae34502ba22d3304fa084d5b40afe, actual tree f657cae770c66df9d6b47f9dbe8e40e14dc6fefa. The 20-file ledger separates CODE/DOC/TEST-SOURCE; root Apache-2.0 does not extend model or dependency rights.
    • Material author: actual gpt-6.1-sol / max; independent reviewer: actual gpt-6.1-sol / ultra, review only. AREX methods were adapted to native InferRSI ownership and readers; no DisCo execution is claimed.
    Criterion Accepted evidence
    Pin/readset/license Owner and ledger retain exact file identities, symbols and notices.
    Complete state and lifetime Procedure/reference distinguish endpoint continuation state from ancestor KV pages, waiter cancellation from scheduler-owned tickets, and completion/publication/use/retirement/reclamation.
    Five outcomes NoMatch, CostRejected, CapacityRejected, EnqueueRejected and Admitted stay distinct; the conditional cost example uses additional saved tokens and does not turn admission counters into completed reuse.
    Actionable diagnostics Incomplete state, namespace/version/corruption and permission-read errors, cancellation, grouped SSE events and cross-clock errors have evidence-backed next observations and safe boundaries.
    Prototype boundaries Research hardware scope, GDS deferred and unexecuted native/model/multi-card effects remain explicit.
    Discovery/generators/package Canonical source 8011318e0c1ba1b92fe35f28c149874e6280614b; current input set 488 produces 73 discovery records/74 files and a 146-input coverage view. All 106 incoming owner files and 185 fixtures are preserved. Linux stock checks and real wheel reads pass. The final query-limit test fix retains exact cardinality/navigation assertions and changes no generator input/output.
    Existing #88 plan Plan has N/new-session, matched complete old/new loading bundles, retained loading/warmup regression cases and separate discovery probes. It is prepared, not evaluated; uncovered behavioral paths are not a PASS.
    Tracker backfill #91 records the exact PR, revision, owners, source pin and candidate/not_run limits; its ongoing intake remains open.

    Prior failed working-copy/coverage checks, stale hidden-PR-reference failures and the obsolete 100-result test cap were retained as historical failures, not reused as current clearance. The source/resource effect is unqualified; software delivery is the accepted scope.

  6. Dong1017
    closed this as completedon Oct 28, 2026
  7. Dong1017
    self-assigned this
    on Oct 28, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

area:knowledgeKnowledge line: packs, skill library, knowledge admissionkind:taskBounded delivery task with a delegated slice

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions