Motivation

PR #23882 introduced state_type="dsv4" for the new DeepSeek-V4 flat heterogeneous state pool (SWA + compress + indexer pools) and added a matching branch to NixlKVManager.maybe_send_extra. The mooncake sibling, MooncakeKVManager.maybe_send_extra, was never updated.

DSv4 disaggregated runs over the mooncake transfer backend silently fall through maybe_send_extra's final else: return 0 branch -- the SWA / compress / indexer state pool is never transferred from prefill to decode. The decode-side state buffers keep whatever stale data they were initialized with, producing wrong outputs whenever the model attends to that state.

Repro on GB300 disaggregated DSv4-Pro, /v1/completions with raw prompt_token_ids ending on the literal <think> token (id 128821):

• Monolithic sglang: correct (gsm8k Janet question -> #### 18).
• Disagg + NIXL: correct.
• Disagg + mooncake (unpatched): wrong (model regurgitates an earlier few-shot answer, e.g. Weapon: ... #### 84).

The corruption is most visible when the prompt's last attended position lives in the SWA / indexer state pool. Other endings often look correct because the K/V cache itself does transfer via send_kvcache; the <think> ending is the worst-case noise location for that token's attention pattern.

Modifications

Add the missing dsv4 branch to MooncakeKVManager.maybe_send_extra that delegates to the existing _send_kvcache_generic -- the same helper the swa / nsa branches use. This routes DSv4's flat state pool through get_mla_kv_ptrs_with_pp, which iterates per-pool entry and uses prefill's state_item_lens for offset arithmetic.

Why delegation rather than a fresh per-page flat path:

• The per-page flat path used by NIXL hard-asserts src_state_item_lens[i] == dst_state_item_lens[i] for every entry. With MTP enabled, decode's indexer-pool entry is 2x prefill's (decode carries the EAGLE-3 draft layer), so the assertion fires on the very first transfer and the engine never makes progress.
• Delegating to _send_kvcache_generic matches what the staging DSv4 branch did before this code path was split into a dedicated dsv4 state_type. Prefill writes its half-size at the natural offset on decode; the decode-only MTP half is left untouched, which is correct -- decode populates its own draft state.

The diff is purely additive and bails for asymmetric TP between prefill and decode (matching the constraint of the surrounding swa / nsa path). Existing mamba / swa / nsa / none paths are unchanged.

Validation

GB300 1P+1D DSv4-Pro disagg, mooncake backend, with the patched conn.py hot-mounted into lmsysorg/sglang:nightly-dev-cu13-20260509-9ee83034.

Sample post-patch response for the original <think>-ending repro on mooncake:

We are given: "Janet's ducks lay 16 eggs per day. She eats three for breakfast every
morning and bakes muffins for her friends every day with four. She sells the remainder
at the farmers' market daily for $2 per fresh duck egg..."

Eggs laid: 16 per day - 7 = 9 eggs.
She sells these at $2 each, so daily earnings: 9 * $2 = $18.
Answer: 18.</think>

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings as needed.
• Provide a benchmark / accuracy verification

cc @hnyls2002 (PR #23882 author).