feat(inference): surface DCP and PCP across point views

[cquil11]

Summary

• Preserve aggregate dcp_size/pcp_size and role-specific prefill/decode DCP/PCP fields when benchmark rows become chart points.
• Use the shared parallelism formatter for chart labels, hover content, and agentic detail sibling chips. For example, TP8 + DCP8 renders consistently as TP8/DCP8; PCP is appended when greater than one.
• Show DCP and PCP widths in official, comparison, and unofficial-run point hovers; disaggregated labels and tooltips retain per-role values.
• Normalize aggregate DCP/PCP values in the benchmark-siblings API so agentic detail pages use the same topology label source of truth.
• Add unit and browser regression coverage.

Context

Fixes the frontend follow-up identified in the InferenceX #2616 review.

Run 31893747354 attempt 2 emits aggregate dcp_size: 8 and pcp_size: 1; its points and agentic detail chip therefore render as TP8/DCP8, while the default PCP1 remains omitted.

Validation

• bun run typecheck
• bun run lint
• bun run fmt
• bun run test:unit (3,919 tests passed)
• Cypress: unofficial-watermark.cy.ts (rendered SVG point label)
• Cypress: agentic-point-time-series.cy.ts (11 tests passed, including the active TP8/DCP8 detail chip)

Unofficial-run overlays and agentic detail pages now consume the same benchmark normalization and shared parallelism formatter as official chart points.

---

Note

Low Risk
Display and data-mapping changes across inference UI and read APIs; no auth or persistence logic. Main risk is inconsistent labels if metrics JSONB is missing or malformed, mitigated by tests and cache namespace rollover.

Overview
Extends the shared parallelism label formatter so decode-context (DCP) and prefill-context (PCP) widths appear in chart point labels when values are greater than one (e.g. TP8/DCP8, TP8/DCP8/PCP4), while default widths of 1 stay omitted to preserve legacy strings.

Benchmark row → chart point mapping now keeps role-specific and aggregate dcp_size/pcp_size metrics from JSONB. Tooltips (official, GPU comparison, unofficial overlays) and agentic sibling chips use the same formatter and show aggregate or per-role DCP/PCP in English and Chinese.

The benchmark-siblings query/API adds normalized DCP/PCP on sibling rows (coalescing role fields with aggregate metrics) and bumps the cached response namespace to v3 so agentic detail topology chips match scatter labels. Unit and Cypress tests cover labels, tooltips, unofficial scatter overlays, and the active agentic point chip.

^{Reviewed by Cursor Bugbot for commit a823492. 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 16, 2026 7:45pm

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

There are 2 total unresolved issues (including 1 from previous review).

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit a823492. Configure here.}