FEAT: Switch default frontend to Next.js app

[OliverBryant]

What & why

Switch the default web UI to the Next.js app under frontend/, shipped as a static export bundled into the xinference package and served by the backend itself — single process, no Node runtime needed at runtime. pip install xinference gives the full product again, same as the old CRA UI. This follows the approach used in xagent#732.

How it works

Frontend (static export)

• next.config.ts: output mode is env-driven — builds default to export; NEXT_OUTPUT=standalone keeps the Node server; next dev sets none (export would break dynamic routes in dev). Rewrites (the dev API proxy) only apply outside export mode; in production the UI is same-origin with the API.
• Dynamic routes (launch-model/[modelType], register-model/[modelType], register-model/[modelType]/[modelName], running-model/[modelUid]) are split into a server wrapper (generateStaticParams emits a __shell__ placeholder page) + a client page that reads real params from the URL (useParams/useSearchParams in Suspense).
• layout.tsx drops request-time cookies()/headers() and the server-side cluster-auth fetch; AppInit now resolves /v1/cluster/auth client-side, and I18nProvider falls back to the browser language on first visit.
• Root / redirects to /launch-model client-side (server redirect()/config redirects are unsupported in export).

Backend (xinference/api/frontend_static.py)

• mount_frontend() serves /_next assets plus a SPA catch-all that reconstructs the route table from the export's __shell__ files, serves RSC .txt prefetch payloads, guards path traversal, returns JSON 404 for unmatched API prefixes (derived from the live route table), and redirects legacy /ui/ to /. No-op when the export is absent (API-only).
• ensure_spa_fallback_last() keeps the catch-all at the end of the route table so runtime Gradio mounts at /{model_uid} are not shadowed.
• Replaces the previous XINFERENCE_FRONTEND_ENDPOINT redirect; override the export location with XINFERENCE_FRONTEND_DIST_DIR if needed.

Packaging

• npm run build stages the export at xinference/ui/web/dist via a postbuild hook; setup.py build_web (hooked on install/develop/sdist, as before) runs npm and asserts the staged output. MANIFEST grafts the staged dist, so wheels/sdists are self-contained. Docker flows are unchanged (python setup.py build_web still works; the staged dist is untracked so git restore . keeps it).

CI & tests

• Lint job: ESLint → next build (static export) → verify out/ shape (index.html, 404.html, __shell__ pages) → serve the real export through FastAPI and run the routing tests.
• xinference/api/tests/test_frontend_static.py: unit tests on a synthetic export (no Node needed) — shell pattern building, backend-prefix 404s, RSC payloads, traversal rejection, Gradio-mount reordering.
• xinference/api/tests/test_frontend_static_real_export.py: integration tests against the real frontend/out (skipped when not built).

Verification

• Static export build: 19/19 routes, all 4 dynamic routes emit __shell__ shells; postbuild staging works.
• 37/37 backend tests pass (also with --noconftest, as the CI lint job runs them).
• Built a wheel locally: all 90 export files included under xinference/ui/web/dist.
• pre-commit (black/flake8/isort/mypy/codespell) passes; ESLint 0 errors.

Notes

• The old CRA app under xinference/ui/web/ui is untouched here; it can be removed in a follow-up.
• next@15.2.3 has a published advisory (CVE-2025-66478); worth bumping in a follow-up.

[gemini-code-assist]

Code Review

This pull request migrates the Xinference Web UI to a separate Next.js application in the frontend/ directory, updating the build process, Dockerfiles, and documentation accordingly. Feedback points out a critical bug where the Switch component's state handler was incorrectly changed to onChange, and suggests several improvements: defaulting the backend redirect to the local frontend port, removing the web build step from standard Python installations to save time, excluding Next.js build artifacts from MANIFEST.in, and stripping trailing slashes from API URLs to prevent routing issues.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[OliverBryant]

/gemini review

[gemini-code-assist]

Warning

Gemini encountered an error creating the review. You can try again by commenting /gemini review.

[OliverBryant]

/gemini review

[gemini-code-assist]

Code Review

This pull request transitions the web UI to a Next.js static export served directly by the FastAPI backend, moving the source code to the frontend/ directory and updating build scripts, Dockerfiles, and documentation. It introduces a backend module to handle SPA routing, dynamic route shells, and static assets. The review feedback identifies three key issues: a bug in launch-dialog.tsx where using onChange instead of onCheckedChange on the Switch component breaks state updates, a routing issue in frontend_static.py where trailing slashes can cause 404 errors, and a potential build-time exception in i18n-context.tsx from accessing navigator during static pre-rendering.

[OliverBryant]

/gemini review

[gemini-code-assist]

Code Review

This pull request restructures the Xinference web UI into a Next.js static export under the frontend/ directory, which is staged at xinference/ui/web/dist and served directly by the FastAPI backend, removing the Node.js runtime dependency in production. It adds SPA routing fallback logic in the backend, updates Dockerfiles to Node.js 20.19.0, and transitions server-side layout logic to client-side initialization. The review feedback correctly identifies a bug in launch-dialog.tsx where changing onCheckedChange to onChange on a Radix UI Switch component will cause TypeScript compilation errors and break functionality.

[qinxuye]

LGTM