docs: SGLang gRPC workers can serve from Rust: SMG_SGLANG_SERVICER_IMPL=rust on the same entrypoint

[github-actions]

What this PR does

How do I run an SGLang gRPC worker on the Rust servicer implementation — which environment variable selects it, what is the default, what must be installed, and which modes refuse it?

Why we need it

Source change: smg-project/smg@065807e

Commit 88 (#2750, 065807e1, 2026-10-04) added a Rust implementation of the SGLang gRPC servicer behind one environment variable, verified on current main. Selection: grpc_servicer/smg_grpc_servicer/sglang/rust.py:45 defines SERVICER_IMPL_ENV = "SMG_SGLANG_SERVICER_IMPL"; resolve_servicer_impl (rust.py:52-59) defaults to python and raises on values outside {python, rust}; smg_grpc_servicer/sglang/server.py:112-121 (serve_grpc, the entry behind python -m sglang.launch_server --smg-grpc-mode) hands the process to serve_rust before any Python gRPC machinery exists and warns that SGLang's HTTP sidecar (metrics/profiling endpoints) stays off on this path. serve_rust (rust.py:272-316) starts the Rust smg.servicer.SglangGrpcServer (binding listed in bindings/python/src/smg/servicer.py:26/:32; implementation crates/engine_servicer/src/sglang/) on the worker's host:port, launches the scheduler headless against a same-host ZMQ handshake, and supervises both; per the in-code comment at server.py:112-113 "the entrypoint and every ServerArgs flag stay the same, and the Router cannot tell them apart" — the gateway still connects over grpc://, so this is NOT the ZMQ direct backend of zmq-workers.md (grpc-workers.md:125 draws the same distinction for vLLM). Refusals: refuse_disaggregation (rust.py:252-269) fails startup under PD (disaggregation_mode set) or EPD (language_only/encoder_only); LoRA loading answers UNIMPLEMENTED (crates/engine_servicer/src/sglang/service.rs:17-30) and SubscribeKvEvents is relayed when the engine publishes KV events, UNIMPLEMENTED when they are off (crates/engine_servicer/src/sglang/mod.rs:140); grpc_servicer/README.md:295-313 states the user contract (needs the smg wheel; multimodal inputs and hidden states refused; Embed, FlushCache and profiling answer as on the Python servicer) — note rust.py's module docstring (lines 14-18) lags the README on those capabilities, so the writer must follow the README and the Rust code. No smg serve flag: bindings/python/src/smg/serve.py:826-837 scopes --servicer-impl to the vLLM gRPC servicer, parse_serve_args rejects it for other backends (serve.py:937-941), and only VllmWorkerLauncher.gpu_env manages a servicer env var (serve.py:321-334), so for SGLang the selection surface is the worker's environment. Missing docs: a repo-wide grep for SMG_SGLANG_SERVICER_IMPL and SglangGrpcServer finds nothing in src/lib/content; getting-started/grpc-workers.md's SGLang tab (:55-65) shows only the Python servicer start and the "Rust vLLM Servicer" section (:105-144) is explicitly vLLM-only; getting-started/index.md's SGLang worker recipe (:300-319) lacks the one-line Rust-servicer note its vLLM recipe carries at :298. Version framing: the commit postdates the v1.11.0 bump (commit 130, #2670, 2026-09-24), so present it as current-main behavior newer than v1.11.0, the style grpc-workers.md:107 already uses for the vLLM Rust servicer. One independent concern: one commit added one selection surface for one engine's gRPC servicer, documented where SGLang gRPC workers are started; the Rust path's full runtime capability matrix is the grpc-multimodal scan's charter and the edit should state limits only at the README's level. Dedup: no open or closed PR mentions SMG_SGLANG_SERVICER_IMPL or #2750 — merged #122 covers the different SGLang ZMQ direct backend (#2749), merged #105/#115 cover the vLLM servicer switch, closed-unmerged #87 covered SGLang gRPC worker shutdown (#2428, not recreated). No open PR touches grpc-workers.md or getting-started/index.md, and my other selected item edits only kv-events-cache-aware.md, so no files are shared. The TokenSpeed Rust servicer (commit 92, #2747, SMG_TOKENSPEED_SERVICER_IMPL) is a separate commit and concern, deferred because its canonical home is the same grpc-workers.md.

Scope: cli-configuration / sglang-rust-servicer-selection. 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: 592935b2-0997-4489-82b3-6f1a80407c80

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.}