[Bug][PD][NIXL] always send aux on is_last; only expects_state when truthy

[ishandhanani]

Motivation

Fixes #25698.

Two asymmetries introduced in #24932 ([PD] Refactor hybrid state transfer) hang every NIXL P/D disagg request for dense models (no Mamba / SWA / NSA — i.e. LLaMA, Qwen3, Gemma, GPT-OSS, …) on v0.5.12. Bisects cleanly to commit d7f4761a4.

1. NixlKVManager.transfer_worker gates the aux RDMA write inside if kv_chunk.is_last and kv_chunk.state_indices:. For dense models, state_indices is [] (falsy from the prefill scheduler's empty state_types), so the whole branch short-circuits and send_aux is never called. Decode never receives the {room}_aux notification → hangs.
2. NixlKVReceiver.send_metadata flips expects_state=True whenever state_indices is not None. Decode receives state_indices=[] (non-None, empty) for dense models, so expects_state flips on. The corresponding prefill check at line ~692 is truthy (if kv_chunk.state_indices:) and (correctly) never sends a state notif, so is_done() waits forever for a notif that doesn't come.

Either bug alone hangs the request; both fixes are required for end-to-end disagg on dense models.

Mooncake is unaffected — mooncake/conn.py:1344-1357 already gates state and aux on independent if blocks, and mooncake has no expects_state field (per-chunk return-code polling instead). Mooncake users on v0.5.12 can keep running disagg; NIXL users currently can't.

Modifications

python/sglang/srt/disaggregation/nixl/conn.py:

1. Split the is_last and state_indices gates in transfer_worker so the aux RDMA write always runs on is_last, and state only runs when state_indices is non-empty. Matches the v0.5.11 shape and matches mooncake/conn.py:1344-1357.
2. Switch NixlKVReceiver.send_metadata from if state_indices is not None: to if state_indices: so expects_state only flips on when there's actually state to expect.

Accuracy Tests

End-to-end manual verification on Qwen/Qwen3-0.6B (dense Qwen3) with NIXL UCX backend on 2× L40S, single host, tp1+tp1:

13:53:59.769  prefill bootstrap_thread recv KVArgsRegister (17 frames)
13:53:59.774  decode  send_metadata
13:54:00.464  prefill bootstrap_thread recv TransferInfo (10 frames)
13:54:00.555  prefill add_transfer_request (nkv=1, aux_index=1)
13:54:00.640  prefill transfer_worker DONE n_handles=2 elapsed=66ms   ← KV + aux both sent
13:54:00.640  decode  notif tag=kv  → is_done=False (aux still needed)
13:54:00.641  decode  notif tag=aux → is_done=True                    ← unblocks

Before either fix n_handles=1 (aux skipped) and decode is_done stays False forever. After both fixes the request returns tokens cleanly in <2s.

I do not have access to a multi-node setup to exercise heterogeneous-TP staging here, but neither hunk changes any code path that runs only with enable_staging=True, state_indices truthy, MLA, or decode_tp_size != attn_tp_size — they only restore the dense-LLM path to its v0.5.11 behavior.

Benchmarking and Profiling

No perf change expected — same RDMA ops on the wire; no allocation or branch-count changes for the stateful paths (Mamba/SWA/NSA), and the dense-LLM path now actually completes instead of looping in pop_transferred until SGLANG_DISAGGREGATION_WAITING_TIMEOUT.

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests & Adding to CI.
• Update documentation as needed, including docstrings or example tutorials.

---

CI States

Latest PR Test (Base): ❌ Run #26071445981
Latest PR Test (Extra): ❌ Blocked -- run-ci is required first.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[ispobock]

/tag-and-rerun-ci

[ShangmingCai]

LGTM

[ispobock]

@ishandhanani could you fix the lint?

[ishandhanani]

@ishandhanani could you fix the lint?

Yep. Will do in an hour when I get online again

[ShangmingCai]

It would be better not to include the code line in the comment, let me fix this and fix the lint

[ishandhanani]

Stupid AI :)

[ShangmingCai]

@ishandhanani could you fix the lint?

Yep. Will do in an hour when I get online again

@ishandhanani Let me do it for you. Since nixl is not in the CI, we could just merge. And feel free to add a nixl test in test_disaggregation_basic.py (no HCA, requires TCP transport) or add a different TP test in H20 suite when you get online again.