docs(blog): Agentic AI part 4 — InferenceService v0.8.0, single-source thesis (FR)

[Smana]

Part 4 of the Agentic AI series, in French. Follows Self-hosted LLM stack: laying the foundations (part 3).

How the InferenceService abstraction in cloud-native-ref goes from v0.6.0 to v0.8.0 — technical source: Smana/cloud-native-ref#1559. Everything is backed by an e2e run on a from-scratch EKS cluster.

The thesis

The deployment unit is no longer "a pod serving a model": it's a single declarative claim that owns, reconciles and garbage-collects the pod and its route as one object — ending the second source of truth (a model with no route is a silent 404; an orphaned route is a phantom backend). On that basis, the cheapest traffic policy in the ecosystem — a zero-GPU weighted split across LoRA adapters — becomes just a field of that unit.

What's covered

• The claim owns its route (gateway.enabled) — Backend / AIServiceBackend / AIGatewayRoute generated and GC'd with the claim; the readiness latch (route withheld until the Deployment is Available, observed on-cluster).
• Weighted LoRA split at zero GPU cost — two backendRefs to the same Backend, modelNameOverride; measured 8.2% for a configured 10%, corroborated by vllm:lora_requests_info. Stated honestly as a frozen weighted split, not a canary (no bake window, no auto-rollback; the missing link is Flagger/Argo Rollouts).
• engineArgs, the governed escape hatch — the curated-plus-verbatim middle that none of the 8 compared projects occupy — with the leak owned: --max_num_seqs (underscore) slips past the denylist. Lesson: a single enforcement point is only safe if it's leak-proof.
• status.servedModels — one pod, three names, readable in kubectl get.
• Semantic routing (model: MoM) — a distinct platform capability that lives in the Gateway, not the claim; composes with the split because they decide at two different points in the chain.

Deliberately honest

• Cut to "what's next": the Endpoint Picker (wired, opt-in, not yet activated — no measured gain, GPU-capped) and the Run:ai streamer (~4% cold-start gain; real lever is the 124s engine init).
• llm-d's 57× TTFT figure is cited as external and not reproduced.
• Every terminal output was actually executed.

Architecture audit

The rewrite is grounded in a multi-agent architecture audit (8 subsystems mapped, 8 ecosystem projects compared — KServe, llm-d, GAIE, Modelplane, production-stack, Envoy AI Gateway, Crossplane patterns, managed offerings). Notably: the zero-GPU LoRA split is unique across all 8; the governed escape hatch is a genuine differentiator. Kept local (not in this PR).

⚠️ Draft — not mergeable as-is

• architecture.png — global architecture diagram (source: docs/architecture/llm-platform-overview.drawio), to export from drawio (marked DIAGRAM-EN-ATTENTE)
• thumbnail.png — missing; the post is featured = true
• Grafana screenshot of the llm-gateway dashboard (canary vs base attribution) — marked SCREENSHOT-EN-ATTENTE
• draft = true to remove once the above are done
• English translation (out of scope)

Tightened from an earlier 832-line draft to ~450 lines around the single-source thesis.