Skip to content

knowledge: add EPD provisioning cold-serving case #505

Open

Description

@Kelvin-Al

Current state Draft PR #506 is open at 255887cd3633a7e3eb59bd9bfca461641aa76f50. The Knowledge case now covers the original six-wave comparison and a nine-wave, position-balanced cold MMMU follow-up. The current head passed author-run targeted and clean-checkout installed-package checks and the GitHub declaration check. The review workflow awaits maintainer approval. Retained validation · Declaration check · Review CI

Decision needed Maintainer scope acceptance and approval to run Review CI. Formal PR approval remains pending. The requester corrected and confirmed the reported authoring configuration as GPT-6.2 Sol / xhigh. The PR declaration now records that user-provided setting. Independent execution-provenance verification is not claimed.

Next action Maintainers approve the queued review workflow. The reviewer can then inspect its output and the retained build and test artifacts for the current head. Maintainers decide scope acceptance and issue closure.

Contribution scope

User-authorized preparation of a cold EPD capacity-planning case under methods/inference-optimization-casebook. Nine changed files comprise a Markdown case, a structured source record, one casebook navigation entry, two documentation matrices, existing inventory/discovery tests, the generated casebook projection and shared provenance manifest. Runtime and benchmark execution code are outside the slice.

The case covers fixed four-H800 E+PD deployment, independently estimated stage capacity, zero-reuse cold execution checks, the original two reverse-order MMMU rounds, and a separate three-round follow-up that places every split once in each position. It retains preprocessing, transport, finite-window, failure/rerun and model/runtime identity limits.

Mathematical result and empirical scope

With positive constant per-instance capacities, one GPU per instance and scalable independent stage pools, maximize min(e * mu_E, (N - e) * mu_PD) over integer 1 <= e < N. Its increasing and decreasing branches cross at N * mu_PD / (mu_E + mu_PD). Feasible floor/ceiling evaluation finds an integer maximizer for any integer N >= 2 under that model.

Both comparisons examine the capacity inputs and model applicability for one four-H800 deployment and preprocessed MMMU workload. They do not empirically qualify other GPU budgets, cost ratios, workload types, cache policies, heterogeneous workers or transfer implementations. Changed conditions require re-estimation and a new matched check.

Direct source and artifact links

The case pins both cited evidence snapshots. Later external source changes leave these historical records unchanged. New claims require new provenance and rechecking the affected artifacts.

Completed execution and INCONCLUSIVE acceptance

The original six-wave comparison ranked 2E:2PD first in both rounds while keeping it in second position. The follow-up ran nine waves in one four-H800 allocation on the original node, with 2E:2PD ranking first once from each position. Its lead over the runner-up was 10.7%–13.8%. All nine waves completed 720 requests each with 720 visual forwards and zero EC or resident-cache hits. These are external author-reported observations. The reviewer checked the published rows and hashes.

Formal campaign acceptance remains INCONCLUSIVE. The import lacks a prospectively frozen acceptance gate, independently verified model-weight revision and complete independently checked runtime/raw-trace artifacts. The position-balanced follow-up has one observation per split and position, with limited evidence on predecessor effects and repeatability. The scope also includes preprocessed images and a filesystem connector.

A conclusive future campaign must first freeze its acceptance criteria and sampling/repetition protocol, bind exact identities, retain independently verifiable workload and traces, qualify its deployment regime and obtain independent review. No retrospective numerical gate is invented.

Validation at 255887c

133 targeted Knowledge and index tests, index freshness and targeted Ruff passed. The clean-checkout build and isolated installed read matched five committed resources and opened the EPD case through discovery. The published wheel was re-downloaded and hashed. Current-head JUnit and build evidence

Earlier validation at 455f1c8

Targeted Knowledge tests: 133 passed in 60.26s (0:01:00)
Changed Python test files: Ruff passed
git diff --check: passed
Seven published source files: retrieved and SHA-256 checked
wheel discovery: case.epd-cold-provisioning
wheel read: case body and declared source record resolved
provenance: 7 immutable published source hashes retained
ownership: one case ID, installed wheel resources, no source checkout imports

The wheel was installed into a fresh target and checked outside the checkout with isolated Python. Full-suite attempt --maxfail=5 stopped with 3 failures, 131 passes, 4 skips and 2 errors involving Windows CRLF, symlink privileges, missing FIFO support and environment-variable size. Full lint reports 12 F811 findings in four unchanged baseline files. Details and commands are in the PR. This earlier validation covered packaging and discovery at that revision.

Review and ownership

This contribution is author-prepared. An independent reviewer checked the pinned numbers, source hashes and generated metadata, and recorded a COMMENT. Maintainer scope acceptance, Review CI and formal approval remain pending. Maintainers own acceptance and closure. The Draft PR retains the current user-confirmed author settings and their provenance limit.

Earlier resource closure and Skill Eval disposition

At 455f1c8afd2b796c2a3a40d259b3ded246e98d06, the case-specific control verifies the case ID, normalized source metadata, exact source-record bytes, failure when a required resource is missing, and content-revision invalidation after mutation. The installed wheel body and source-record bytes match their wheel entries. Checks run outside the checkout with isolated Python. That earlier check used the local-checkout dev5 wheel (SHA-256 7678eefe7ea9b00cbac0864b56bed20c3e4b6e0f48864b7307c65720bd3152ad). The clean-checkout dev1 build below supersedes it as the build and installation evidence for review.

The case states capacities in requests/s per one-GPU effective instance, measured feasible e values {1, 2, 3}, and PD count 4 - e. Preprocessing, transfer, shared contention, routing and batching have no separate expression terms. Model-scoped integer optimality and deployment-scoped observations remain distinct.

The automated Skill Eval plan at the previous head selected zero comparisons and did not execute. Missing verified ancestor pin and unavailable trusted pricing/configuration prevent automatic selection/execution. The casebook projection and shared provenance manifest are regenerated through the supported workflow. Scripted reader/package checks establish their stated behavior only. That Skill Eval did not establish model efficacy or formal service acceptance. Detailed claim dispositions and check outputs are in Draft PR #506.

Generated-projection correction and retained validation

The supported generator read committed source revision 9f89294c9898a90125d4f848fcaad2a29faa722f. The source record contains both EPD resources and the updated SKILL hash. Corrected PR head is 455f1c8afd2b796c2a3a40d259b3ded246e98d06. Its 133 passing tests include all nine index tests and an EPD free-text discovery/open regression. The generator's explicit --check reports 80 current metadata records. Installed-wheel checks also cover generated-card resource membership and the source pin.

Retained validation artifact contains head-bound JUnit and installed check script/output. These are author-executed records for independent inspection. Independent reproduction and formal campaign acceptance remain pending. The earlier unknown-attribution exception request has been withdrawn after the requester corrected and confirmed GPT-6.2 Sol / xhigh. Current policy admits that model/effort combination. These artifacts cover the earlier review head 455f1c8afd2b796c2a3a40d259b3ded246e98d06.

Earlier clean-checkout wheel evidence at 455f1c8

Build transcript, verification script and wheel tie the installed case resources to PR head 455f1c8afd2b796c2a3a40d259b3ded246e98d06. Two author-run builds yielded wheel SHA-256 8506c97a5d3bdb34cf9a80e7d07e3725ca55677c4efff47ddd0e771eb720e2e2. Installed bytes matched five committed resources and the documented discovery and read path passed. The author reran the 133 Knowledge/index tests successfully. The independent reviewer inspected the retained report and script and marked the missing build provenance resolved. Maintainer scope acceptance, review-lane CI and formal approval remain pending.

Activity

  1. Dong1017 commented on Nov 18, 2026

    @Dong1017
    Collaborator

    Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-sonnet-5-5).

    Assessment: The scope is reasonable and narrow. It adds a Knowledge-only case under inference-optimization-casebook for EPD provisioning on a fixed 4×H800 E+PD setup. It covers the min(E capacity, PD capacity) integer GPU allocation rule, cold EC execution checks, and two matched reverse-order MMMU rounds. It is docs and resource registration only, with no runtime changes. I'd treat it as acceptable in principle, with the gaps below to close before maintainer scope acceptance. This is a first-pass triage and I have not checked the draft PR, the source revision, or the external artifacts.

    What looks good

    • The source is pinned to an immutable revision (epd-provisioning@3ed184e3…).
    • The formal outcome is recorded honestly as INCONCLUSIVE, and the measurement and transport/preprocessing limits are kept.
    • The issue states that acceptance and closure belong to maintainers, and that independent review is still pending.

    Gaps and risks

    1. No links in the issue. The external experiment issue, the frozen artifacts, and the draft PR are referred to but not linked. Reviewers can't verify the claims from this issue alone. Please add direct links, ideally with artifact hashes or immutable refs.
    2. "Completed" versus INCONCLUSIVE. The body says the cold serving observations are "completed", but formal acceptance is INCONCLUSIVE. The case should lead with that status and say which criterion was not met: noise, sample size, order effects, or the transport/preprocessing confound. Otherwise readers may take the allocation rule as validated.
    3. Generalization of the integer rule. The case is bound to a 4-GPU, H800, E+PD topology, a single model/workload (MMMU), and cold EC. The "applicability" section should list conditions where the rule doesn't hold. Examples are other GPU counts, a warm EC cache, a non-multimodal workload, and different E:PD cost ratios. Calling it a rule is stronger than two matched rounds support, so consider "observed/provisional heuristic" wording.
    4. Next action is already done. The "Next action" line reads as future work, but the PR is described as already drafted. Please update the issue so the state is accurate.
    5. Provenance and discovery. The issue names "required resources" and "installed discovery and provenance" validation. Please state in the PR which checks were run (resource registration, discovery from an installed package, provenance fields) and paste the output. Reviewers shouldn't have to infer it.
    6. Independent review. The case is author-prepared, so a reviewer other than the author should check the numbers against the frozen artifacts before it merges.

    Clarifying questions

    • Is the PR number open and linked to this issue, and does it touch anything outside the casebook directory and its resource registry?
    • Does the case need re-validation if the epd-provisioning source moves past 3ed184e3, or is it intentionally frozen?
    • What would turn INCONCLUSIVE into a conclusive result (more rounds, a different transport path, a controlled preprocessing baseline)? Recording that would be a useful follow-up.

    Suggested next steps

    • The author adds the links and the INCONCLUSIVE framing described above.
    • A maintainer confirms scope acceptance: a Knowledge-only case with no behavior change.
    • An independent reviewer cross-checks the cited numbers against the frozen artifacts, then reviews the PR.

    Issue acceptance and closure stay with maintainers. This comment is triage only (Dong-bot, authorized delegate).

  2. zengchuang-hw commented on Nov 18, 2026

    @zengchuang-hw
    Contributor

    独立需求与实现契约检视(基于 main 965ab5cf83399a99530949b04bb934a892ee417c):这个 Knowledge-only 方向与现有 casebook 的“保留 INCONCLUSIVE 的有边界诊断观察”一致,但目前不能据此确认实验结论或宣布交付验收完成;这也不是代维护者接受范围/关闭 Issue。

    我核对了 AGENTS、Knowledge 契约、casebook 过程/元数据、现有 admission-drain case,以及 knowledge_operations.py 的 case 发现和资源读取逻辑。建议把本 Issue 的交付检查明确为:

    1. 可追溯来源:补充 draft PR 的直接链接、epd-provisioning 的完整仓库/固定提交路径,以及外部实验 Issue、冻结原始结果与校验值的可访问指针。目前 Issue 正文只有提交号,无法独立定位并核对来源内容。区分“作者已完成的观察”“独立核对过的原始结果”“未重跑”;未重新计算的摘要哈希不要标为 verified。
    2. 不要把容量模型写成服务增益结论:明确 min(E capacity, PD capacity) 两项的单位、工作负载条件、整数可行域(固定四张 H800 的分配约束)及未纳入的传输/预处理成本。把理论分配规则、cold EC 执行观察和两轮反序匹配测量分开呈现;没有完整冻结质量/SLO/测量证据时,保留正式 INCONCLUSIVE,不外推为任意 GPU/模型配置的最优策略。
    3. 资源闭包按实际 reader 验证:现有 reader 从 owner 的 cases/*.json 发现 kind: case,按 content 和 required_resources 装载文件;它不会因为 Markdown 写了相对链接就自动把目标纳入 case bundle。除自身 JSON basename 外,正文必需的本地来源记录/附件也应列入 required_resources,避免在 checkout 可读、规范化 bundle 内断链。外部冻结证据仍留在原 owner,不冒充本地保留资源。
    4. 交付测试分层:用最终 PR revision 验证独立 case 身份读取、必需资源完整/缺失失败、来源元数据与内容 revision 一致,以及脱离 checkout 的已安装包发现/资源字节一致。现有 fresh-install 测试是 opt-in,不能把普通 smoke 绿灯当成已经跑过安装验证;这些测试即使通过,也不等于 H800 服务性能验收。

    本轮只做规范与实现静态核对,没有读取其他人的检视意见,没有执行仓库代码、安装测试或 GPU 实验,也未取得本案冻结实验原文。以上是可落地的交付/证据边界,不是对尚未核实的 draft PR 或实验结果的批准。

  3. david6666666 commented on Nov 18, 2026

    @david6666666
    Collaborator

    Could the summary name the metric behind “ranked first”? In the frozen wave results, 2E:2PD led on central-window PD completions/s and lowered p90 TTFT at this offered load, but p90 TPOT was worse than 1E:3PD in both rounds (18.81 vs. 17.56 ms; 18.48 vs. 17.95 ms). Naming that tradeoff here would keep the result from reading like an overall latency or SLO win.

  4. zengchuang-hw commented on Nov 18, 2026

    @zengchuang-hw
    Contributor

    针对正文 2026-11-18 19:04:30 UTC 更新的独立证据复核,承接本人此前 Issue 意见及 PR #506 当前版本检视。

    有实质进展:当前 head 的可检查验证材料已补齐,不能再把它描述为“仅有旧版本证据/clean-wheel 进行中”。 我读取了固定版本 88c63700757fb9d2aa1fd262df249128835d63a3 下完整的 report.json、verify.py、README 和 JUnit:

    • JUnit 有 133 个 test case,报告零失败、错误和跳过,包含 EPD closure/query 以及 index freshness 等控制。README 将其归属到 255887cd3633a7e3eb59bd9bfca461641aa76f50;这是作者执行并保留的测试证据,不是我的独立重跑。
    • 构建记录实际测得该 HEAD 和 tree,记录空的构建前工作区状态、固定构建依赖、fresh wheel 构建、五份 Git/checkout/wheel/installed 资源对比,以及 checkout 外隔离安装后的真实发现/读取路径。脚本与报告相互对应;记录的 wheel 为 inferrsi-0.1.dev1+g255887cd3-py3-none-any.whl,SHA-256 42d1d956f6210bd28c26ef7f1925f7b555df9e55968559ab5b0382cf5b27555e。因此我上一轮要求保留新 head 可检查验证/构建链记录这一缺口已解决,不再保持为缺失记录阻塞。

    正文也已把原六波与新九波观察分开,补齐两个固定来源及中断样本链接,解决上一轮的摘要同步建议。新九波表格、五份来源哈希及生成资源哈希的独立核对已记录于上述 PR 检视;本次不重复主张新的设备验证。

    仍不代表验收或批准: 当前 head 的 ready/review CI 仍为 action_required,不是已通过测试;声明状态为 success,但不替代 CI。维护者范围接受与正式 campaign 资格仍分开处理,保持 INCONCLUSIVE。这里没有发现新的确定内容缺陷,也没有代维护者关闭 Issue。

    本轮只读检查正文增量与固定证据,未读取他人检视意见,未执行仓库代码、测试、构建/安装或 GPU 任务,也未下载独立哈希核验二进制 wheel。报告的构建与安装结果仍明确标为作者执行;不把留存记录误称为独立实验复现。

  5. added
    kind:taskBounded delivery task with a delegated slice
    area:knowledgeKnowledge line: packs, skill library, knowledge admission
    on Nov 19, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    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