fix: dp=1 EP replicas launch dense-style (vLLM rejects external-LB args at dp=1)

[mikasenghaas]

Problem

Multi-node inference with enable_expert_parallel and per-node data-parallel size 1 (e.g. tp=8 on 8-GPU nodes → dp_local=1; the "etp8" shape: experts EP-sharded across the node's 8 GPUs, attention at full tp=8) fails to launch. The EP branch in multi_node_rl.sbatch.j2 unconditionally passes vLLM's external-LB DP args, which vLLM rejects when data_parallel_size == 1:

pydantic_core.ValidationError: data_parallel_external_lb can only be set when data_parallel_size > 1

Fix

Guard the external-LB EP launch on INFER_GROUP_DP > 1; for dp=1 EP replicas, launch dense-style (independent single-engine server). enable_expert_parallel in the engine config alone still shards experts across the TP group — no external-LB coordination needed within a single DP rank.

Split out of #3067 as an independent bug fix (the global-router change there is separate and still under evaluation). Validated: etp8 launches and serves cleanly on 8 nodes with this fix; without it, launch crashes at engine init.

🤖 Generated with Claude Code

---

Note

Low Risk
Narrow SLURM launcher branch change for dp=1 EP; multi-DP EP behavior unchanged when INFER_GROUP_DP > 1.

Overview
Fixes multi-node RL inference launch when enable_expert_parallel is on but the replica’s effective data-parallel size is 1 (e.g. one node with tp=8 → INFER_GROUP_DP=1).

The EP branch in multi_node_rl.sbatch.j2 now requires INFERENCE_ENABLE_EXPERT_PARALLEL and INFER_GROUP_DP > 1 before passing external load-balancer DP fields (data_parallel_rank, data_parallel_address, RPC port, INFER_GROUP_DP). Otherwise it uses the dense-style launch_inference_rank path (dp=1, no DP coordinator args), because vLLM rejects data_parallel_external_lb when data_parallel_size == 1.

Expert sharding within a single TP group still comes from enable_expert_parallel in the inference config; only cross-rank external-LB coordination is skipped for dp=1 EP.

^{Reviewed by Cursor Bugbot for commit c0aa31d. Bugbot is set up for automated code reviews on this repo. Configure here.}

[S1ro1]

lgtm