Add eight AgentX-era glossary terms and refresh nine definitions / 新增八条 AgentX 相关术语并更新九条释义

[functionstackx]

Summary

I read the 104 commits merged since 2026-08-11 and pulled out the concepts that became user-visible in that window but had no glossary entry, plus the entries whose text now describes behavior the dashboard no longer has. Both languages updated in the same change.

Glossary is 54 entries → 62.

New terms (8)

Term	Category	Why now
E2E Normalized Interactivity	Benchmark metrics	Shipped as an experimental agentic x-axis with its own FAQ answer and help link (#847, #848)
Tokens per dollar	Benchmark metrics	Now the default y-axis on the inference charts, with input/output and RMB variants (#770, #787, #793)
Energy per token	Benchmark metrics	Query energy and % TDP axes landed, split across all-in provisioned vs measured telemetry (#734, #735)
Context parallelism (PCP / DCP)	Parallelism	DCP and PCP degrees are surfaced in point tooltips and parallelism labels (#745)
KV cache offload	Serving	Offload halo on the chart, offload type/engine and hit rates in the point detail view (#729, #818, #760)
KV cache manager	Software	kv_offload_backend is recorded run metadata; Mooncake / LMCache / HiCache sit under the engines
KV-aware routing	Serving	Router layer is part of the recipe label (Dynamo, llm-d, Mooncake ATOMesh) and is the AgentX bottleneck story
TileRT	Software	A first-class framework label, deliberately retained in best-per-SKU views (#843)

Each entry follows the existing shape: plain-English gloss, definition, mechanism, why it matters, and an InferenceX-specific benchmarkContext that says how to read it on this dashboard rather than restating the definition.

Refreshed definitions (9)

Corrections, not cosmetic edits:

• speculative-decoding said InferenceX "distinguishes MTP-enabled and disabled curves". Since Merge mixed speculative decoding points on agentic curves #695 that is only true for fixed-sequence scenarios; agentic curves treat spec decode as point-level metadata and merge the points. Also records the fixed acceptance-length methodology, which exists because replayed AgentX content is synthetic.
• kv-cache and prefix-caching both said prefix caching is disabled on random data. Still true for fixed-sequence, but AgentX is the deliberate exception: reuse is the property being measured, and hit rate is a reported per-point quantity.
• iso-interactivity now records the rule from fix(inference): derive $/M tok and J/token from throughput instead of splining them / 每 token 成本与能耗改为按吞吐量推导而非直接插值 #726 — the frontier is built on throughput vs interactivity, and $/M tok and J/token are derived from the interpolated throughput rather than splined separately, because splining a per-chip constant over throughput breaks the identity between knots.
• pareto-frontier now explains that best-per-SKU views merge permitted optimizations into one curve per model/SKU/engine, so adjacent points on a line can differ in spec decode, disaggregation, or offload.
• input-output-sequence-length now covers agentic runs, where there is no single ISL/OSL pair to quote and the point view plots a fitted lognormal with percentiles (feat(inference): plot agentic ISL/OSL distributions as lognormal / 将智能体 ISL/OSL 分布改为对数正态呈现 #810, feat: show resident ISL and OSL percentiles #811).
• interactivity, performance-per-dollar, tokens-per-megawatt cross-link to the new metrics that supersede or complement them on the current axes.

Notes for reviewers

• No new files or routes. The glossary index, per-slug pages, and sitemap all derive from getAllGlossaryEntries() / getAllZhGlossaryEntries(), so /glossary/* and /zh/glossary/* pick up all eight terms automatically in both languages.
• articleSlugs on the new entries point only at posts that actually discuss the term; most cite the AgentX v3 port, which is where these concepts are documented at length.
• Not done here, deliberately: the FAQ entry for E2E Normalized Interactivity (Add normalized interactivity FAQ help link / 为归一化交互性添加 FAQ 帮助链接 #848) and the AgentX telemetry tutorial could both link their new glossary pages. That is a cross-linking change to other surfaces and belongs in its own PR.

Verification

• bun run test:unit — 4,772 tests pass. The relevant guards: glossary.test.ts (unique slugs, 8–40 word plain-English gloss, ≥90 word body, ≥3 resolvable related terms, ≥1 real article, no em/en dashes, complete zh translation with a Han-character floor and distinct measurement labels) and zh-copy.test.ts (mechanical Chinese rules, including the untranslated-chip and duplicated-loanword checks).
• bun run lint, bun run fmt, bun run typecheck clean.

中文说明

我梳理了 2026-08-11 以来合入的 104 个提交，挑出这段时间内新出现在用户界面、但术语表尚未收录的概念，以及释义已经与仪表板当前行为不符的条目，中英文在同一 PR 中同步更新。术语表条目从 54 条增加到 62 条。

新增术语（8 条）

术语	分类	新增理由
端到端归一化交互性	基准指标	已作为 agentic 场景的实验性 X 轴上线，并配有 FAQ 说明与直达链接（#847、#848）
每美元 token 数	基准指标	已成为推理图表的默认 Y 轴，并提供输入/输出与人民币口径（#770、#787、#793）
每 token 能耗	基准指标	新增每请求能耗与 TDP 占比坐标轴，并区分全量供电口径与实测遥测（#734、#735）
上下文并行（PCP / DCP）	并行策略	DCP 与 PCP 并行度已在数据点 tooltip 和并行标签中展示（#745）
KV cache offload	推理服务	图表上的 offload 光环，以及点详情中的 offload 类型、引擎与命中率（#729、#818、#760）
KV cache 管理器	软件栈	kv_offload_backend 已作为运行元数据记录；Mooncake / LMCache / HiCache 位于引擎之下
KV 感知路由	推理服务	路由层是方案标签的一部分（Dynamo、llm-d、Mooncake ATOMesh），也是 AgentX 的瓶颈所在
TileRT	软件栈	已是独立的框架标签，并在每个 SKU 最佳配置视图中特意保留（#843）

每条新增条目都沿用既有结构：通俗解释、定义、机制、意义，以及说明在本仪表板上如何解读的 benchmarkContext，而不是重复定义。

更新释义（9 条）

这些是纠错而非文字润色：

• speculative-decoding 原文称 InferenceX「区分开启和关闭 MTP 的曲线」。自 Merge mixed speculative decoding points on agentic curves #695 起这只对定长场景成立，agentic 曲线已把投机解码作为数据点级元数据并合并这些点。同时补充了固定接受长度的方法说明——之所以需要它，是因为 AgentX 回放内容是合成的。
• kv-cache 与 prefix-caching 原本都称在随机数据上禁用前缀缓存。这对定长场景仍然成立，但 AgentX 是有意为之的例外：复用正是被测量的性质，命中率是逐点上报的指标。
• iso-interactivity 补充了 fix(inference): derive $/M tok and J/token from throughput instead of splining them / 每 token 成本与能耗改为按吞吐量推导而非直接插值 #726 的规则：前沿建立在吞吐量与交互性之上，每百万 token 成本和每 token 焦耳由插值吞吐量推导，而不是各自单独做样条，否则会破坏「每芯片常数除以吞吐量」这一恒等关系。
• pareto-frontier 说明每个 SKU 最佳配置视图会把允许的优化合并成一条曲线，因此同一条线上相邻的点可能在投机解码、分离式部署或 offload 上不同。
• input-output-sequence-length 补充了 agentic 运行的情况：没有单一 ISL/OSL 取值可引用，点详情视图会绘制拟合的对数正态分布并给出分位数（feat(inference): plot agentic ISL/OSL distributions as lognormal / 将智能体 ISL/OSL 分布改为对数正态呈现 #810、feat: show resident ISL and OSL percentiles #811）。
• interactivity、performance-per-dollar、tokens-per-megawatt 增加了与新指标的交叉链接。

评审提示

• 不新增文件或路由。术语表索引页、单条目页和 sitemap 均由 getAllGlossaryEntries() / getAllZhGlossaryEntries() 派生，因此 /glossary/* 与 /zh/glossary/* 会自动收录全部八条术语。
• 新增条目的 articleSlugs 只指向确实讨论该术语的文章，多数引用 AgentX v3 移植文，因为这些概念在那里有充分说明。
• 有意未做：Add normalized interactivity FAQ help link / 为归一化交互性添加 FAQ 帮助链接 #848 的 FAQ 条目和 AgentX 遥测教程都可以链接到新的术语页，但那属于其他界面的交叉链接改动，应当单独提 PR。

验证

• bun run test:unit — 4,772 项测试通过。相关守卫：glossary.test.ts（slug 唯一性、通俗解释 8–40 词、正文 ≥90 词、≥3 个可解析的相关术语、≥1 篇真实文章、禁用长破折号、中文翻译完整且汉字数量达标、measurement 标签需与英文不同）与 zh-copy.test.ts（机械性中文规则，包括未翻译的 chip 和重复借词检查）。
• bun run lint、bun run fmt、bun run typecheck 均通过。

🤖 Generated with Claude Code

---

Note

Low Risk
Documentation-only changes to glossary data in two lib files; no runtime logic, auth, or API behavior changes.

Overview
Expands the InferenceX glossary from 54 to 62 entries by adding eight AgentX-era concepts in glossary.ts with matching glossary-zh.ts translations. New terms cover dashboard-visible metrics and serving features: E2E Normalized Interactivity, tokens per dollar (default chart Y-axis), energy per token, context parallelism (PCP/DCP), KV cache offload, KV cache manager, KV-aware routing, and TileRT.

Nine existing entries get corrected benchmarkContext (not cosmetic copy): fixed-sequence vs AgentX behavior for prefix caching and speculative decoding; iso-interactivity derivation of cost/energy from interpolated throughput; Pareto best-per-SKU merged curves; agentic ISL/OSL distributions; and cross-links from interactivity, perf/$, and tokens/MW to the new metrics. relatedTerms are updated so glossary navigation stays consistent.

No new routes or files—/glossary/* and /zh/glossary/* pick up slugs from getAllGlossaryEntries() / getAllZhGlossaryEntries() automatically.

^{Reviewed by Cursor Bugbot for commit b4d5c6c. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview	Aug 25, 2026 9:32pm