fix(inference): per-precision line labels when multiple precisions selected

[functionstackx]

Problem

When more than one precision is selected (e.g. FP4 and FP8), each precision is rendered as its own curve, but the line labels were deduplicated by hardware key. The result:

• Only one of the two same-hardware curves got a label.
• The label itself omitted the precision, so you couldn't tell which curve it belonged to.

See the reported screenshot: with FP4 + FP8 shown, "B200 (vLLM)" labeled only one of the two B200 lines and gave no precision.

Fix

When selectedPrecisions.length > 1:

• Every curve gets its own line label (dedup key becomes the series key hw_precision instead of just hw).
• The precision is appended to the label text, e.g. B200 (vLLM) FP8 and B200 (vLLM) FP4.

With a single precision selected, behavior is unchanged — one label per hardware, no precision suffix.

Applied consistently across:

• Interactivity charts (greedy collision-avoidance placement) — static render + zoom re-placement.
• TTFT / E2E latency charts (endpoint labels + nudge) — static render + zoom re-placement.

Overlay (unofficial run) support

Per AGENTS.md, the change covers the ?unofficialrun= overlay path too: overlay line labels gain the precision suffix when multiple precisions are shown, so the two precision curves of an overlay stay distinguishable. Overlay labels were already keyed per series, so the only gap was the missing precision text, now fixed in both static and zoom paths.

Tests

Added a case to cypress/e2e/line-labels.cy.ts that loads /inference?i_linelabel=1&i_prec=fp4,fp8 and asserts that at least one line label contains FP8 and one contains FP4.

🤖 Generated with Claude Code

---

Note

Low Risk
Chart labeling-only change in ScatterGraph with single-precision behavior preserved; no auth or data pipeline impact.

Overview
When more than one precision is selected on inference scatter charts, line labels now follow one label per curve (dedup by series key instead of hardware only) and append the precision (e.g. B200 (vLLM) FP8 vs FP4) via lineLabelText and getPrecisionLabel.

Behavior is unchanged with a single precision selected. The same rules apply to interactivity greedy placement, TTFT / E2E endpoint labels, unofficial-run overlay labels, and the matching zoom re-placement paths.

A Cypress case loads i_linelabel=1&i_prec=fp4,fp8 and asserts at least one label contains FP8 and one FP4.

^{Reviewed by Cursor Bugbot for commit 22764d5. 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, Comment	Jun 7, 2026 12:29am