[Serve][LLM][SGLang] Wide / elastic expert parallelism serving pattern

[eicherseiji]

The existing Wide-EP pattern (DPServer + decode-as-orchestrator PD + gang scheduling) passes vLLM-keyed kwargs and sets VLLM_RAY_BUNDLE_INDICES. SGLang consumes different names (dist_init_addr, moe_dp_size, enable_dp_attention, etc.) and has in-engine elastic expert backup. This issue tracks a parallel SGLang Wide-EP server class.

Subitems

• A new SGLang DP / Wide-EP server class.
• SGLang's MoE / EP engine args (moe_dp_size, moe_ep_size, enable_dp_attention, elastic_ep_backend, enable_elastic_expert_backup) are first-class through the new server.
• Layering: engine-level elastic backup handles expert-local failures, gang RESTART_GANG handles control-plane / NCCL-level failures. Documented contract.
• SGLang Wide-EP release test on a real MoE model (DeepSeek-V3 class). Fault-injection variant kills one scheduler actor and asserts engine-level recovery.
• SGLang Wide-EP user guide documenting topology and failure-mode contract.

Open questions

• Failure-mode layering between engine-level elastic expert backup and Ray-side RESTART_GANG. Engine handles expert-local; gang handles control-plane. Explicit contract so the two don't fight.
• Whether expert-placement topology is observable at the Ray layer or stays engine-internal.
• Autoscaling behavior on wide-EP deployments: expert rebalance on scale-up, or opaque DP-group scaling.
• Whether the new server class subclasses DPServer or stands alone.

Upstream coordination

• Expert-placement state exposed over a TokenizerManager RPC for Ray-side observability.
• Clear engine signal when elastic backup is in progress, so gang restart can defer.

[KaisennHu]

I'd love to contribute to this!