docs: The Rust vLLM servicer switch needs no vLLM change: self-installing plugin and vllm serve --servicer-impl

[github-actions]

What this PR does

Does selecting the Rust vLLM gRPC servicer still require a vLLM whose gRPC entrypoint carries the smg hook, and can the implementation be chosen on vLLM's own command line (vllm serve --grpc --servicer-impl rust) instead of only through SMG_VLLM_SERVICER_IMPL?

Why we need it

Source change: smg-project/smg@50ba16f

Commit 26 (#2755, 50ba16f7, 2026-10-04) removed the need for a vLLM-side hook, one day after merged docs PR #105 (which documented the hook as a requirement). Verified on current main: grpc_servicer/smg_grpc_servicer/vllm/launcher_switch.py installs the servicer switch over vLLM's serve_grpc from the launcher's own module-level import of this package's servicer classes (install_launcher_switch is called from health_servicer.py at import, plus a meta-path finder for late-loading launchers); grpc_servicer/pyproject.toml registers a vllm.general_plugins entry point (smg-servicer = smg_grpc_servicer.vllm.plugin:register), and plugin.py adds --servicer-impl {python,rust} to every FlexibleArgumentParser that defines --grpc (the vllm serve parser), with resolve_servicer_impl (rust.py:343-355) reading args.servicer_impl ahead of SMG_VLLM_SERVICER_IMPL and writing the decision back to the environment so require_python_impl agrees; plugin.py's module doc notes a VLLM_PLUGINS setting that leaves the plugin out only removes the flag — the env var works regardless. rust.py:upstream_hook_installed (def at line 380) now returns true when the launcher imports this package's servicer modules at module level OR carries the check in its own serve_grpc, and require_upstream_hook (lines 470-479) fails only when neither holds; require_python_impl (lines 364-377) names the one unswitchable form — "a launcher executed directly as main, or one that did not import smg_grpc_servicer.vllm.servicer before defining serve_grpc". The commit's grpc_servicer/README.md hunk states "it needs no vLLM change" and shows vllm serve Qwen/Qwen3-0.6B --grpc --port 50051 --servicer-impl rust. Stale docs, all in src/lib/content/getting-started/grpc-workers.md's Rust vLLM Servicer section (written by merged PR #105 on 2026-10-03, before this commit): line 107 "The switch is the SMG_VLLM_SERVICER_IMPL environment variable, read by upstream vLLM's gRPC entrypoint before it builds an engine" (the entrypoint no longer consults anything; the package installs the switch itself, and a --servicer-impl flag now also exists on vllm serve --grpc); line 133's requirement "a vLLM whose gRPC entrypoint consults the switch (smg_grpc_servicer.vllm.resolve_servicer_impl)" (false on main — any vLLM whose gRPC launcher imports the package's servicer classes, the stock shape, is switched unchanged); line 135 "the installed vLLM lacks the hook" (the fail-fast check now verifies the launcher will be switched, and the Python servicer's refusal covers the launcher-run-as-main case). getting-started/index.md:170 and :298 were read and stay accurate (the smg serve flag and env-var selection are unchanged), so the edit is scoped to grpc-workers.md. Version caveat: the commit landed after the v1.11.0 bump (commit 69, #2670, 2026-09-24); the section already frames the whole feature as newer than v1.11.0, so describe the self-installing switch as current-main behavior without naming a release. This is an engine-compatibility concern (which vLLM installations can serve the Rust servicer and what fails loudly), distinct from merged PR #105's smg serve selection surface, which it corrects from a newer primary commit; one independent concern: one commit changed one delivery mechanism for one switch, corrected in the one section that states the old requirement. No open or closed PR covers #2755, and no open PR touches grpc-workers.md (open PRs #110 and #107 hold concepts/architecture/grpc-pipeline.md, reference/api/messages.md, and getting-started/pd-disaggregation.md). Note: the cli-configuration scan's pending sglang-zmq-direct-backend concern also lists grpc-workers.md among its doc_paths; if that PR opens first, this concern must wait for it to merge.

Scope: grpc-workers / rust-servicer-switch-no-vllm-hook. Other concerns are deferred.

How to test

• Passed the documentation path and size guard (under 1000 added plus deleted lines; no file-count limit).
• Passed an independent accuracy and single-concern review.
• Passed git diff --check and pnpm check and the production pnpm build.

Checklist

• Tests added/updated (if applicable)
• Docs updated (if applicable)
• Commit includes a DCO sign-off

[coderabbitai]

Important

Review skipped

Bot user detected.

To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

• Configuration used: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: bf0e8deb-418f-4217-b5f5-f06ccb648953

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

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