Skip to content

docs: add DeepWiki navigation and clarify README guidance - #545

Merged
Dong1017 merged 2 commits into
InferMatrixLab:mainfrom
Inference1:refine-readme
Nov 21, 2026
Merged

Dong1017 merged 2 commits into
InferMatrixLab:mainfrom
Inference1:refine-readme

Conversation

@Inference1

@Inference1 Inference1 commented Nov 21, 2026 •

Copy link
Copy Markdown
Contributor

Governing Issue and delivery slice

Refs #557

This README-only follow-up builds on the documentation refresh in #533 for #531. The confirmed scope covers DeepWiki navigation, dashboard export instructions, benchmark evidence, and contribution guidance. Maintainer scope decision

Changed artifacts and evidence

  • DeepWiki. Added the project link to the top navigation. A signed-out request returned HTTP 200 and showed the InferMatrixLab/InferRSI Overview with links to Getting Started and the module layout. Project documentation
InferRSI project documentation on DeepWiki
  • Dashboard. The instructions show how to generate dashboard.html in the recorded run directory and open it in a browser. A local smoke test using a minimal run record generated the HTML in the input directory with inferrsi dashboard --input .. Dashboard guide
  • DeepSeek V4.1 Flash. The historical calibration table reports median TPOT across six run-level means. Its C1, C4 and C16 values round to 39.28, 44.07 and 56.86 ms/token, all above the 25 ms/token target. Calibration comment · Public API record
  • LFM3-Omni-2B. The retained baseline and candidate waveforms matched bit-for-bit within each tested temperature. The README percentage uses the rounded throughput means documented in the existing results guide. Upstream A/B benchmark · Throughput calculation
  • MiniCPM-o 4.5. The pinned report records the overall throughput-and-latency gate as failed and the formal speech quality gates as not run for track_s_fast_r11. It separately reports development quality measurements for track_s_fast. The README names the exact profile and links to the quality table. Same-card results
  • Contributing. Clarified that merge requires passing the required checks and approval by at least two independent reviewers, and that the issue owner can reopen the issue. Review and merge policy

Knowledge authorship declaration

Not applicable under knowledge-model-policy. The diff touches only README.md. Codex assisted the review follow-up and is credited in its commit trailer.

Checks and dependencies

  • git diff --check and the repository's PR-reference check passed.
  • Verification covered documentation, public links and source evidence, and the local dashboard smoke test. Hardware benchmarks and formal quality gates were not rerun.

@Dong1017 Dong1017 left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

PR review verdict — DO NOT PASS

Reviewed the full README-only diff for #545 from base 9466bc8d910c9fef55d4a9e23ee1a82950280c30 to head fc7fc2828e8e2d4e12ffc96592436cdd134be591. The review record shows no formal approval on this head and no outstanding changes request. The blockers below are about evidence for the new wording, not a request to rerun benchmarks.

  • README.md:100 — The DeepSeek row now identifies the three unchanged values as medians of six run-level means. I read #353 and confirmed the workload and 25 ms/token target, but the available issue text does not include the linked historical calibration comment or its six-run aggregation. The issue's current two-start protocol does not establish how those historical numbers were calculated. Cite the historical table or artifact that shows the aggregation, or keep the previous, less specific wording.
  • README.md:150 — The MiniCPM-o row changes “quality unqualified” to “formal speech quality checks were not run.” Those are not equivalent: checks could have run without qualifying the configuration. Point to the passage in the pinned same-card report that records the checks as not run for this configuration, or retain “quality unqualified.”

No runtime code changed. I did not run the dashboard, check live links, or verify CI. This comment does not approve the PR; I’ll re-read after the author pushes.


Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model gpt-6.2-sol).

@zengchuang-hw zengchuang-hw left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

NO PASS

首次独立检视:9466bc8d910c9fef55d4a9e23ee1a82950280c30…fc7fc2828e8e2d4e12ffc96592436cdd134be591。只改动 README.md,无上一轮意见。

批准前请补齐两点:

  • PR 说明仍是模板,Refs #<issue-number> 没有指向实际 Issue。请填入已有的文档 Issue,简述本次范围和实际核查结果;不用为这次小改动另做实验。
  • README.md:5 新增的 DeepWiki 是顶栏入口。我未能从公开访问核实它已经显示 InferRSI 的文档,不能据此断言链接已坏。请确认未登录访客能打开正确的项目页;若尚未索引,先移除这条入口。

正文改写与所引证据一致:Dashboard 默认写入输入目录;DeepSeek 三档数值是六次运行均值的中位数;LFM3 在各自测试温度下对比保留波形;MiniCPM 正式语音质量门未在所测配置上运行。未重跑 CLI 或性能实验。

@Dong1017 Dong1017 added the kind:docs Documentation and template delivery label Nov 21, 2026

Copy link
Copy Markdown
Collaborator

首次完整检视:9466bc8d910c9fef55d4a9e23ee1a82950280c30…fc7fc2828e8e2d4e12ffc96592436cdd134be591(1 个提交,仅 README.md)。结论针对文档措辞与所引证据是否一致,不是对底层基准的复测。当前 head 上已有 @Dong1017 的 DO NOT PASS 与 @zengchuang-hw 的 NO PASS;我逐条独立核对后,本轮不 Approve。

关于所引证据(独立核对):

  1. DeepSeek 行「六轮均值的中位」可以在已链接评论里核到。我读了 https://github.com/InferMatrixLab/InferRSI/issues/353#issuecomment-6641425447(2026-11-04T20:33:24Z)。其中表格标题即为「六轮均值的中位TPOT (ms/token)」,C1/C4/C16 为 39.282… / 44.072… / 56.862…,与正文 39.28 / 44.07 / 56.86 一致,且三档都高于 25 ms/token 目标。acceptance: DeepSeek V4.1 Flash 910C 8K/1K TPOT ≤25 ms/token at C1/C4/C16 #353 现行正文已改为后续两轮协议、不再重复该表,但 README 已经指向这条历史校准评论,因此聚合方式有出处。不必仅因 issue 正文未复述该表而打回这一句。

  2. MiniCPM-o:「overall throughput and latency target was not met」与所链 same-card 报告 中的 “Throughput-and-latency 2× gate: FAIL” 一致。正式 speech quality gates(WER / SIM / UTMOS,1088 rows × 3 seeds,C=1 与 C=32)对 headline track_s_fast_r11 为 Not run,因此 “formal speech quality checks were not run for the tested configuration” 有依据。但它与旧的 “quality unqualified” 并不完全等价:同页还有 track_s_fast(不含 r11 三项选项)的开发质量结果,绝对门槛通过,UTMOS 比参考低 0.13。请把句子收紧为明确「headline track_s_fast_r11 的正式 quality gates 未跑」,或保留 “quality unqualified”,避免读成完全没有质量数据。

合并前请补齐:

  1. PR 说明仍是模板:Refs #<issue-number>、delivery slice,以及 knowledge-model-policy 中的 <current-full-head-sha> 等占位符未填。请改成实际治理 Issue(若已有文档 Issue,用 Refs #N,不要用会自动关闭的 Closes)、简述本次范围;README 小改不必另做实验。知识声明请按政策如实填写(本 head 上 knowledge declaration policy 检查已是 success,但仍不应留下占位符)。

其他已核:

  • DeepWiki https://deepwiki.com/InferMatrixLab/InferRSI 我以未登录方式打开,页面标题与 Overview 已是 InferMatrixLab/InferRSI 的项目文档(含 Getting Started、模块布局等),不是未索引占位页。这一条我未再作为阻塞。
  • Dashboard 句与 Contributing「至少两名独立评审批准」与仓库两批准规则相符。
  • 未重跑基准,也未核验 LFM3 行所链上游 PR 的逐温度 bit-identical 原文。
  • CI:knowledge declaration policy success;ready / review lane 为 action_required。

@Inference1 Inference1 changed the title docs: add DeepWiki to README and refine some of the textual descriptions docs: add DeepWiki navigation and clarify README guidance Nov 21, 2026

@zengchuang-hw zengchuang-hw left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

NO PASS

增量复核:fc7fc2828e8e2d4e12ffc96592436cdd134be591 → 007e65b87a770b60457297031b4480dac5b8325c,上一轮检视。两个提交指向同一文件树,README 内容和上一轮验证过的正文一致。

PR 说明已补齐改动范围、数据来源及作者的 DeepWiki 无登录 HTTP 200 验证记录;但说明仍明确写着本轮 governing Issue 待确认。#531 已关闭,对应的是 #533 的上一轮文档刷新,并未记录本次新增导航的范围。批准前请链接本轮适用的 Issue,或记录维护者明确确认的跟进范围。

已比较精确版本及 README,复核了 #531 的范围。DeepWiki 的公开访问目前只有作者记录,本轮未独立复测;未重跑 CLI 或性能实验。仓库 About 设置不是本 PR 的文件改动,也不是本轮检视的前置项。

@Inference1

Copy link
Copy Markdown
Contributor Author

@zuiho-kai @zengchuang-hw Would you please review this

@zengchuang-hw zengchuang-hw left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

NO PASS

应本次复审请求复核。自上一轮检视以来,head 仍是 007e65b87a770b60457297031b4480dac5b8325c,文件无增量。README 改写本身没有发现新的问题;PR 正文仍说本轮对应的 Issue 待确认。我也没有找到涵盖新增 DeepWiki 导航的仓库 Issue。请补上本轮对应的 Issue 链接,或记录维护者对该跟进范围的明确确认。仓库 About 的文案建议不属于 README 改动,不是批准前置项。

本轮核对了 README、PR 范围和 Issue;未独立重测 DeepWiki 公网访问,也未运行性能实验。

Copy link
Copy Markdown
Collaborator

增量检视:fc7fc2828e8e2d4e12ffc96592436cdd134be591 → 007e65b87a770b60457297031b4480dac5b8325c(对照上次评论)。两个提交的 tree 均为 e0b4a6b501718403fdbd41f43f1c12f97e3a0ab6,README.md 与上一轮逐字节一致,没有文件增量。

PR 说明已从模板补成实际范围、数据来源链接和 DeepWiki 未登录 HTTP 200 的核验记录,knowledge 声明也改为 Not applicable。上次指出的占位符问题已解决。

仍待处理,本轮不 Approve:

  1. README.md:150 MiniCPM-o 行仍写 “The formal speech quality checks were not run for the tested configuration.” 文件未改,上次的建议仍在:请收紧为明确「headline track_s_fast_r11 的正式 quality gates 未跑」,或保留 “quality unqualified”,避免读成该配置完全没有质量数据。

  2. 说明里仍写 “A governing issue for this follow-up is pending confirmation.” Docs: refresh README items that no longer match the repository (2026-11-20 audit) #531 已关闭,对应的是 Docs: refresh README quickstart, architecture and release guidance #533 的上一轮文档刷新,并未覆盖本次新增 DeepWiki 导航等范围。批准前请链接本轮适用的 Issue,或记录维护者对该跟进范围的明确确认。当前 head 上 @zengchuang-hw 的 NO PASS 与随后应复审请求的 NO PASS 也指出了这一点。

未重跑基准或 CLI;DeepWiki 公开访问与 DeepSeek 六轮中位出处仍以上次独立核对为准,本轮未再复测。

Point directly to the six-run DeepSeek calibration and the quality status of the headline MiniCPM-o configuration. Format the dashboard instructions without trailing whitespace.

Agent-assisted-by: Codex (author)
Signed-off-by: Inference1 <3617237115@qq.com>

@zengchuang-hw zengchuang-hw left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

NO PASS

增量检视:007e65b87a770b60457297031b4480dac5b8325c → 76342a1ae2e34ac417ca63e0daf0f267f2342180(1 个提交、1 个 README)。上轮检视。独立核对了两个版本:仪表盘段落已去掉尾部空格,DeepSeek 仍指向同一六轮 TPOT 表,MiniCPM-o 已明确到 track_s_fast_r11,并直链质量状态。没有发现新内容错误。

批准前待补: 新建的 #557 已把本次 README 范围列出来,包括 DeepWiki;这解决了没有跟踪 Issue 的问题,但 Issue 仍明确等待维护者确认。请在批准范围后用普通 Refs #557 关联(不要自动关单),并把 PR 正文里的仓库 About 设置提议与这个 README-only 变更分开。我未能用外部读取独立打开 DeepWiki,不能据此称链接失效;请把已有的匿名访问证据记录到 #557。未运行 CLI 或读 CI。

@amy-why-3459 amy-why-3459 left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

PR review verdict — PASS
Reviewed base 9466bc8 and head 76342a1.
Reviewed increment 007e65b → 76342a1 (1 commit, README.md only).

Previous blockers on 007e65b are addressed in this increment and in the updated PR body:

  • MiniCPM-o now says formal speech quality gates for track_s_fast_r11 were not run, and links to the report's #numerical-and-quality-status. I re-read that section: “Throughput-and-latency 2× gate: FAIL” and “Speech quality gates … Not run for track_s_fast_r11,” with a separate development result for track_s_fast.
  • The PR is linked with plain Refs #557. I confirmed that README-only scope as a maintainer on #557, and recorded the signed-out DeepWiki check there. The About suggestion is no longer in the delivery slice; it is a separate maintainer follow-up in the PR body.
  • The dashboard sentence no longer has trailing whitespace.

DeepSeek still points at the same six-run median TPOT comment I already checked. CI — knowledge declaration policy is green on this head. The ready/review lane is action_required because this is a fork PR; that is not a README wording issue. I did not rerun the dashboard CLI or benchmarks, and I have not independently re-checked the LFM3 upstream bit-identical wording.

Accuracy/performance comparison evidence is not applicable (documentation-only change, no new experiments). @zengchuang-hw's remaining NO PASS items on this head (maintainer confirmation of #557, DeepWiki evidence on #557, and splitting About out of the delivery slice) are now resolved; that review is COMMENTED, not CHANGES_REQUESTED. Earlier DO NOT PASS / NO PASS reviews were on previous heads.

@Inference1

Copy link
Copy Markdown
Contributor Author

@Dong1017 @zengchuang-hw @amy-why-3459, could you independently re-review the README-only diff from 9466bc8d910c9fef55d4a9e23ee1a82950280c30 to 76342a1ae2e34ac417ca63e0daf0f267f2342180? Current diff

  • The PR uses Refs #557. The authorized scope decision is recorded, and the repository About proposal was removed from the PR description. Scope decision
  • docs: track README navigation and evidence follow-up #557 records the date and method for the signed-out DeepWiki checks. The project Overview returned HTTP 200, and the Getting Started page opened in a separate anonymous request. A local inferrsi dashboard --input . smoke test wrote dashboard.html in the input directory. Verification record
  • The DeepSeek row now labels the six-run median TPOT table directly, and the MiniCPM-o row names track_s_fast_r11 and links the pinned quality status. The LFM3 report confirms matching retained waveforms within each temperature. Calibration table · MiniCPM-o quality status · LFM3 benchmark
  • git diff --check and the PR-reference check passed. GitHub shows both PR commits as verified with the required DCO sign-off. Commits

Please record your verdict as a formal GitHub review of this head using the repository's review format.

Posted by Codex on behalf of Inference1.

@zengchuang-hw zengchuang-hw left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

NO PASS

证据增量复审:76342a1ae2e34ac417ca63e0daf0f267f2342180 未变。上一轮的范围阻塞已解决:#557 已记录范围确认,PR 正文已用 Refs #557 关联,并移除了 About 设置建议。README 其余文字和来源的核对结论不变。

仅剩已接受的匿名访问验收项。 #557 记录了作者以匿名 GET 看到项目正文的结果;我用独立网页读取和匿名 HTTPS 请求复核时,分别遇到工具无法访问及连接超时。这不是链接失效的证据,也不否定作者的记录。请在 #557 补一份可独立核对的匿名访问留存(例如移除敏感头后保留请求方式、最终 URL、HTTP 状态、时间及项目页面标题),或提供可复核的公开结果。没有额外的措辞或代码阻塞,未运行 CLI 或查看 CI。

@Dong1017 Dong1017 left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

PR review verdict — PASS
Reviewed base 9466bc8 and head 76342a1 (README.md only, full diff seen).

My earlier DO NOT PASS items (on fc7fc28) are resolved:

  • DeepSeek row (README.md:~103): it now says the values are medians of the six run-level mean TPOT measurements at C1/C4/C16. It links the historical #353 calibration comment directly. amy-why-3459 read that comment and confirmed the 39.28 / 44.07 / 56.86 values and the six-run aggregation, and all three values are above the 25 ms/token target.
  • MiniCPM-o row (README.md:~153): "quality unqualified" is replaced by "The overall throughput and latency target was not met. Formal speech quality gates for track_s_fast_r11 were not run." It links the pinned #numerical-and-quality-status section. The wording is scoped to that profile, so it does not read as "no quality data at all". amy re-read that section and found it consistent.

Scope and governance:

  • The PR uses plain Refs #557. #557 records the README-only scope, and that scope was confirmed by me as Dong-bot (authorized delegate) and by amy-why-3459. Repository About metadata is outside the diff and the PR scope.
  • The other README edits match #557's accepted clauses. They are the DeepWiki nav link, the dashboard text (inferrsi dashboard --input . writes dashboard.html in that directory, then open it in a browser), the LFM3 per-temperature bit-for-bit wording, and the Contributing clause (required checks, two independent approvals, issue owner accepts or reopens).
  • The diff touches no runtime code, so no benchmark or accuracy comparison applies.

Formal review state on this head: amy-why-3459 APPROVED. zengchuang-hw's remaining item is a COMMENTED "NO PASS" on this head. It asks only for an independently checkable anonymous-access record for DeepWiki. They state there are no wording or code blockers. #557 records the date, method (anonymous curl GET), HTTP 200 and page title, and amy independently opened the page signed out. I treat that as sufficient for a README nav link. The link is also easy to remove if DeepWiki ever stops serving it.

Not verified: I did not run the dashboard CLI, open DeepWiki, or re-read the upstream LFM3 and MiniCPM-o passages. The LFM3 and DeepWiki points rest on the author's record in #557 and amy's check. I did not check CI. amy reports the knowledge-declaration check is green and the review lane is action_required because this is a fork PR.


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

@Inference1

Copy link
Copy Markdown
Contributor Author

@zengchuang-hw,已按你在 76342a1 上的复审意见补上可复核的留存。两次匿名 HTTPS GET 分别访问 DeepWiki 项目 Overview 和 Getting Started,均返回 HTTP 200。留存含请求时间、最终 URL、状态码、页面标题、HTML 哈希和复核命令。 匿名访问留存

amy-why-3459 也记录了独立的未登录浏览器核对。请复核这份留存是否满足最后一项验收。 独立核对记录

Codex 代表 Inference1 发布。

@zengchuang-hw zengchuang-hw left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

PASS

同一提交 76342a1ae2e34ac417ca63e0daf0f267f2342180 的证据增量复审,无新代码;回应本次请求。上轮剩余的 DeepWiki 留存已补齐。独立对照了 README 和匿名 GET 留存:请求排除了认证和 Cookie,README 目标与记录的最终 URL 一致;Overview 和 Getting Started 均记录 HTTP 200、对应项目的标题、时间与响应体哈希,并给出了复核命令。这是对留存及其与改动一致性的核对,不声称我这边重新取得了 HTTP 200 或重新计算了响应体哈希。

#557 已确认 README-only 范围;仪表盘说明和三项指标来源沿用前轮源码与资料核对。未重新运行 dashboard CLI 或查看 CI。本批准仅针对该 README 改动,不替代 Issue 最终验收或合入判断。

@cursor cursor Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

PR review verdict — PASS
Reviewed base 9466bc8d910c9fef55d4a9e23ee1a82950280c30 and head 76342a1ae2e34ac417ca63e0daf0f267f2342180.

Technical / third-party verify only (cursor App collaborator permission none — not write+ eligible / dual-tip / merge-ready).

Checked

  • README-only: DeepWiki nav link; dashboard open wording; DeepSeek TPOT / LFM3 / MiniCPM evidence wording; contributing merge bar phrasing.
  • Tip already has Dong1017 + amy + zeng Approves after #557 README-only scope confirm; declarations / knowledge-model-policy green on tip.

@Dong1017
Dong1017 merged commit 6226ed0 into InferMatrixLab:main Nov 21, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

kind:docs Documentation and template delivery

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants