docs(pd): cover vLLM HTTP disaggregation, KV-protocol pairing, and EPD

[slin1237]

PD disaggregation changed a lot in v1.10 and v1.11, and the docs still described the pre-1.10 behavior:

• vLLM PD now works over plain HTTP. The gateway dispatches sequentially and relays kv_transfer_params (feat(pd): sequential vLLM disaggregation with kv_transfer_params relay over HTTP smg#2255, with MultiConnector resolution in #2425).
• Anthropic Messages, Responses, and count_tokens route through PD (#2250, #2638).
• Prefill and decode are paired by KV-transfer protocol, controlled by --pd-pairing-mode (#2484, #2486, #2504).
• n>1 fans out into one pair per sample (#2482).
• A failed leg fails fast (#2466), and every router uses the same 503 for an unavailable leg (#2479).
• Kubernetes PD KV metadata comes from pod annotations (#2358, #2359).

The old pages also showed wrong SGLang flags, invented pairing strategies and sizing tables, and no EPD at all.

Changes

• concepts/routing/pd-disaggregation.md (rewritten)
  • A supported-engines table covering SGLang, vLLM, and TokenSpeed; TRT-LLM/MLX (400 runtime_pd_not_supported); and ZMQ (rejected).
  • Parallel vs sequential dispatch: bootstrap fields, the 8998 default, NIXL/Mooncake/passthrough connector modes and recompute fallbacks, and where the connector comes from over gRPC vs HTTP.
  • Failures, cancellation, and streaming: fail-fast, deferred decode abort, and early stream commit with prefill drain.
  • n>1 fan-out.
  • A multimodal note with the three pd_decode_language_model_only_* codes.
  • A supported-API matrix.
  • A new Pairing section: what is compared, off/lenient/strict, pairing_protocol / SMG_PAIRING_PROTOCOL, and 503 no_compatible_pd_pair.
  • Admission and errors: the gRPC decode window, context length, and an error-code table.
  • A new EPD section.
  • Monitoring rebuilt on the real smg_pd_* series.
• getting-started/pd-disaggregation.md (rewritten)
  • A setup chooser.
  • SGLang HTTP and gRPC tabs with correct flags.
  • vLLM over gRPC with NIXL and Mooncake.
  • A new vLLM-over-HTTP walkthrough that registers kv_connector through POST /workers.
  • TokenSpeed PD and EPD, Kubernetes, and pairing and readiness checks.
• getting-started/index.md, PD Disaggregation Workers
  • Uses smg launch (the pip CLI requires a subcommand).
  • Adds a vLLM PD (HTTP + NIXL) tab.
  • Moves the SGLang gRPC decode worker to port 50061, because SGLang's gRPC mode opens an HTTP sidecar on --port + 1, which collided with a decode worker on 50052.
• Both service-discovery pages, PD sections only: a pod-annotation table (sglang.ai/bootstrap-port, smg.ai/kv-connector, smg.ai/kv-engine-id), a vLLM Mooncake pod example, and notes on role precedence and readiness.
• concepts/architecture/overview.md, PD Mode: one accurate paragraph replaces a stale four-step list.

Verification

• Checked against smg v1.11.0 (3be823a7): routers/http/pd_router.rs, routers/grpc/common/stages/{worker_selection,request_execution,context_length}.rs, worker/{pd_pairing,pd_pair_index,kv_transfer}.rs, service_discovery/kubernetes.rs, main.rs, and config/validation.rs. Launch commands were cross-checked with the e2e harness. The sidecar port was checked in the SGLang source.
• pnpm lint, pnpm check (0 errors), and pnpm build pass.
• The link check reports only pages added by sibling PRs: zmq-workers.md (docs: add a guide for direct ZMQ engine connections #45), multimodal.md (docs: add a multimodal pipeline concept page #44), and overload-protection.md (docs(reliability): document worker overload protection and load monitoring #37).
• All six pages render with status 200 under wrangler pages dev.

This is one of a series of docs PRs that bring the site up to smg v1.11.0, one per area.

🤖 Generated with Claude Code

[coderabbitai]

Warning

Review limit reached

Next included review available in 43 minutes.

Check out review usage here.

View limit details

Limit details: You’ve used all 4 included reviews currently available. Your 62 included PR review attempts over the past 7 days set your current allowance at 4 reviews per hour.

Your organization has reached its usage spending cap. Adjust your spending cap in the billing tab.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Team

Run ID: 69e0bd83-2aca-4031-a898-f06feadb5821

📥 Commits

Reviewing files that changed from the base of the PR and between 7362b40 and eb5279a.

📒 Files selected for processing (6)

• src/lib/content/concepts/architecture/overview.md
• src/lib/content/concepts/architecture/service-discovery.md
• src/lib/content/concepts/routing/pd-disaggregation.md
• src/lib/content/getting-started/index.md
• src/lib/content/getting-started/pd-disaggregation.md
• src/lib/content/getting-started/service-discovery.md

---

_{Comment @coderabbitai help to get the list of available commands.}