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

[Oseltamivir]

Summary

$/M tok and spec-sheet J/token are defined as a per-chip constant divided by throughput. Previously, the calculator and historical trends independently splined both throughput and the reciprocal metric, so the two interpolated curves could violate:

metric × throughput = constant

This PR now splines throughput and derives the reciprocal metric from it. Measured knots are unchanged; only values between knots can move.

The size and direction of that movement depend on frontier density and change as new benchmark runs land. Earlier snapshot-specific impact figures have therefore been removed rather than presented as permanent results for the live dataset.

Safety

• recoverReciprocalNumerator applies derivation only when every usable point agrees on the numerator within 0.1%.
• Metrics with a varying numerator, including measured-energy fields, keep the existing direct-spline fallback.
• Total, input, and output token costs are validated against one shared provider rate.
• The Python blog helper now returns null when a reciprocal throughput input is missing, matching TypeScript.
• Existing unit tests execute the bundled Python helper directly, so TypeScript/Python parity works out of the box without a separate analysis workflow.
• Official and ?unofficialrun= calculator paths share the same interpolation implementation.

Validation

• bun run lint
• bun run fmt
• bun run typecheck
• bun run test:unit — 3,821 tests passed across workspaces
• bun run test:e2e — 88 smoke tests passed
• Focused reciprocal interpolation suite — 104 tests passed

中文说明

$/M tok 与基于规格功耗计算的 J/token，其定义都是“单芯片常数 ÷ 吞吐量”。此前 TCO Calculator 与 Historical Trends 会分别对吞吐量和倒数指标做样条插值，导致两条插值曲线之间可能不再满足：

指标 × 吞吐量 = 常数

本 PR 改为只对吞吐量做样条插值，再由插值后的吞吐量推导倒数指标。所有实测节点保持不变，只有节点之间的估算值可能变化。

变化幅度和方向取决于 Pareto frontier 的密度，也会随着新 benchmark 结果持续变化。因此，本 PR 已移除旧快照中的影响百分比，避免将其误写成线上数据集长期不变的结论。

安全措施

• 只有当所有可用数据点的分子在 0.1% 容差内一致时，recoverReciprocalNumerator 才会启用推导逻辑。
• measured energy 等分子会随数据点变化的指标继续使用原有的直接样条插值回退路径。
• total、input 与 output token 成本会共同校验同一个 provider rate。
• Python blog helper 在缺少对应吞吐量输入时会返回 null，与 TypeScript 行为一致。
• 现有单元测试会直接运行随附的 Python helper，无需额外分析流程即可验证 TypeScript/Python 一致性。
• 官方数据与 ?unofficialrun= overlay 共用同一套 calculator 插值实现。

验证

• bun run lint
• bun run fmt
• bun run typecheck
• bun run test:unit — 各 workspace 共 3,821 项测试通过
• bun run test:e2e — 88 项 smoke test 通过
• 倒数指标插值专项测试 — 104 项测试通过

---

Note

High Risk
Changes how cost and energy are interpolated across the calculator, fleet cost-cap lookup, trend charts, and blog helper—published numbers will move, often downward. Guarded by a consistency check with extensive unit coverage, but still user-facing TCO math.

Overview
$/M tok and J/token are now derived from interpolated throughput instead of being splined independently. That preserves metric × throughput = constant between measured knots (previously broken by a median ~72%).

Adds recoverReciprocalNumerator / reciprocalMetricAt and wires them through interpolateForGPU, maxInteractivityAtCost, and interpolateMetricAtInteractivity. A 0.1% consistency guard falls back to the old spline when the numerator is not constant (e.g. measured* energy). Cost rates are recovered across all three token types together so inverse lookups stay on the same curve as the bars.

Syncs iso_interactivity.py with a reciprocal_of option, documents the method in docs/tco-calculator.md, and adds tests that pin the identity, the Python/TS match, and the fallback path.

^{Reviewed by Cursor Bugbot for commit 4138124. 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 12, 2026 11:00am