fix(configs): remove vestigial client dp_rank_count / X-data-parallel-rank

[S1ro1]

Summary

Removes the vestigial client-side dp_rank_count / X-data-parallel-rank mechanism.

Background

dp_rank_count + the X-data-parallel-rank header come from #1940 ("add dp-rank-routing"): with a single hybrid-LB server fronting multiple DP engines, the client pinned each rollout to a DP shard for KV reuse. #2696 ("use external-LB everywhere, drop hybrid-LB") replaced that with one endpoint per DP rank, router-balanced — so the header no longer has a job.

It isn't just dead, it's harmful. The regular vllm-router forwards client headers to the worker, and vLLM validates the rank:

data_parallel_rank N is out of range [0, num_ranks).

Single-DP backends (--data-parallel-size 1, how dense external-LB launches) have num_ranks = 1, so any request carrying rank > 0 is rejected. Affinity / KV reuse is now the router's job (prefix-cache scoring), not the client's.

Change

• Remove client.dp_rank_count from ClientConfig.
• setup_clients: one client per base_url, no X-data-parallel-rank header.
• Drop the dp_rank_count auto-set in the RL resolver and the field pass-through in the elastic pool.
• client_identity keys on api_base_url alone (each DP rank is its own endpoint).
• CHANGELOG + tests updated.

No checked-in config sets dp_rank_count, and the server-side header reader already handles its absence. Breaking only for external configs that set it (extra="forbid" rejects it); nothing to migrate.

Related: #2707 (NCCL world-size) — together they fully fix dense multi-node external-LB.

🤖 Generated with Claude Code

---

Note

Medium Risk
Breaking config removal (extra="forbid") and changed inference routing identity; fixes harmful rank headers on single-DP vLLM but requires one base URL per DP rank in user configs.

Overview
Removes client-side dp_rank_count and the X-data-parallel-rank header path that expanded each base_url into multiple pinned DP clients. With external-LB, each DP rank is already a separate endpoint in client.base_url, so the orchestrator now builds one inference client per URL and load-balances on api_base_url only.

ClientConfig drops the field (breaking for configs that still set it). setup_clients no longer loops ranks or sets rank headers. The RL auto_setup_inference_client hook stops copying inference DP size into the client. Elastic client rebuilds no longer forward dp_rank_count. CHANGELOG documents the migration (delete the key; no replacement).

^{Reviewed by Cursor Bugbot for commit 55d3fa8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[samsja]

ah wait actually do we have router on single node ? if not we cant remove