Relax routed_experts capture KV-connector check to a warning for P/D

[S1ro1]

Relax routed_experts capture KV-connector check to a warning for P/D

Background

--enable-return-routed-experts records, per token, which MoE experts it routed
to. Today VllmConfig.__post_init__ hard-rejects it whenever a KV connector is
configured:

if self.kv_transfer_config is not None and self.kv_transfer_config.is_kv_transfer_instance:
    raise ValueError("--enable-return-routed-experts is incompatible with KV "
                     "connectors (PD disaggregation, KV cache offload).")

Why relax it

For P/D disaggregation this is not fundamentally incompatible. The decode
replica pulls the prompt KV from prefill and never forwards the prompt, so its
prompt-region routing rows are invalid — but the prefill replica returns the
correct prompt-region routing, and a P/D-aware router/proxy can splice it back in:

merged = concat( prefill_rows[:Lp], decode_rows[Lp:] )

So whether routed-experts capture works under disaggregation is a property of the
router/proxy, not something vLLM can decide at config time.

Change

Replace the ValueError for KV-transfer instances with a warning:

You are using P/D disaggregation with routed_experts capture, for this to work
your router/proxy needs to support it

The PP>1 incompatibility is unchanged.

Note: kv_role alone cannot distinguish P/D transfer from single-instance KV
offload — NixlConnector P/D itself runs as kv_both (see
docs/features/disagg_prefill.md, docs/serving/expert_parallel_deployment.md) —
which is why this is a warning rather than a role-gated error.

Router/proxy support

Companion PRs add the prefill→decode routed_experts merge to two routers, each
verified end-to-end against this change on a 2-node Qwen3-30B-A3B P/D deployment
(NIXL + Mooncake), checked against a non-disaggregated oracle under greedy
decoding:

• vllm-project/router
• llm-d/llm-d-router

Router PRs: vllm-project/router#184 · llm-d/llm-d-router#1627