DOC-1581 inference provider production path

[djwfyi]

Content Description

Adds the Inference Provider production path for teams building managed model-serving endpoints on GPU infrastructure. Includes a new production guide, a model-serving runtimes integration page, GPU and inference autoscaling patterns (KEDA and Prometheus), a GPU Operator install guide inside tenant clusters, and a readability pass on the new content.

Preview Link

• (NEW) Inference Provider production path
• (NEW) Model serving runtimes integration page
• (Updated) GPU and inference autoscaling integration page
• (Updated) GPU and accelerator support page

Internal Reference

Partially addresses DOC-1581

AI review: mention @claude in a comment to request a review or changes. See CONTRIBUTING.md for available commands.

FORK LIMITATION: @claude does not work on pull requests opened from forks. GitHub Actions cannot access the required secrets for fork-originated PRs. To use AI review, push your branch directly to this repository.

@netlify /docs

[netlify]

✅ Deploy Preview for vcluster-docs-site ready!

Name	Link
🔨 Latest commit	3b515b8
🔍 Latest deploy log	https://app.netlify.com/projects/vcluster-docs-site/deploys/6a44314d3d3d2c00078cf68e
😎 Deploy Preview	https://deploy-preview-2348--vcluster-docs-site.netlify.app/docs
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.

---

To edit notification comments on pull requests, go to your Netlify project configuration.

[github-actions]

DOC-1581: Docs: make Inference Provider a first-class cross-product docs journey

[djwfyi]

Remaining work and known gaps from DOC-1581

This PR covers Phases 1–3 of the issue. The items below are either out of scope for this PR or deferred.

Done in this PR

• inference-provider.mdx — new production guide (Day 0/1/2)
• model-serving.mdx — tenant-runtime pattern, smoke-test deployment, model storage, Gateway API endpoint exposure
• gpu-hpa-dcgm.mdx — serving-metrics section: KEDA, queue depth, request concurrency, TTFT, latency
• gpu-and-accelerator-support.mdx — GPU Operator install guide inside tenant clusters
• Cross-links from overview.mdx, ai-cloud.mdx, gpu-cloud-platform.mdx
• Architecture diagram (inference-provider.svg)

Not done — cross-repo (separate PRs needed)

• vmetal-docs Phase 4: inference-specific GPU capacity design in docs/operate/gpu-fleet.mdx (hardware SKUs for serving tiers, per-tenant node types, warm pool guidance, OS image/driver rollout for model-serving fleets) and docs/overview.mdx (add inference provider path)
• vnode-docs Phase 5: inference isolation section in docs/concepts/tenancy-models.mdx and docs/concepts/vcluster-integration.mdx (when vNode matters for inference: untrusted adapters, custom containers, privileged workloads)
• Cross-links from vmetal-docs and vnode-docs back to the inference provider path in vcluster-docs

Not done — vcluster-docs (future issues)

• Security/trust model page (Phase 6): dedicated page or section with buyer-facing validation checklist — what an enterprise buyer can verify about isolation boundaries, RBAC, quota enforcement, and audit trail
• Detailed Day 2 operations content: the Day 2 table links out but there is no prose for capacity exhaustion handling, endpoint drain and rollout, stuck model pods, GPU burn-in/validation, node reclaim on customer offboarding, or regional failover
• SEO pass (Phase 7): natural placement of inference-provider search terms and final cross-link sweep across the three repos
• End-to-end "serve a model" walkthrough: model-serving.mdx covers the runtime pattern but not the full step-by-step from vMetal node prep through validated external endpoint; the issue scoped this as a potential child issue

[djwfyi]

Reciprocal dependency: vmetal-docs#29 (DOC-1581 Phase 4)

The vMetal half of this inference provider journey is now up in loft-sh/vmetal-docs#29 (DOC-1581 Phase 4). It adds an "Inference provider capacity" section to the GPU fleet ops page and cross-links the vMetal docs back to the inference provider production path created here.

Merge order: vmetal-docs#29 links to https://www.vcluster.com/docs/vcluster/production-guide/inference-provider, which doesn't exist until this PR merges and deploys. The links are external URLs so no build gate fails either way, but please merge #2348 first (or together) so the vMetal links don't point at a not-yet-published page.

This PR already links into the vMetal pages (operate/gpu-fleet, deploy/gpu-quickstart), which are live, so there's no dependency in this direction.