spec: DSPARK PD-disaggregation via native draft-KV transfer

[bobcao3]

Turns out the DSPARK draft-KV transfer landed upstream already / is in some other PR that landed recently

Reduced from the original DSPARK PD-disagg commit: the spec_info
DSPARK branch and draft-KV transfer landed upstream via #32541 and the
DP/EP draft-batch fields via #33098, this one-line guard is the only piece
still missing on main.

---

Original PR body

Motivation

We are trying to bring up 4xB200 prefill + 4xB200 decode for Dsv4 flash 0731 (excellent model),
but with vanilla nightly we bumped into incompatibilities between DSPARK and PD-disagg (and file based HiCache).

Originally we went with #31466 but found that has incompatibility with HiCache,
after some inspection I think we can get this done with a much smaller set of changes presented here,
instead of the full hidden-row transfer system.

We have now a running setup on top of 07/31 nightly showing DSPARK with 3-5x improvement in single user case and KV cache seems to be functioning correctly.

Modifications

Two hunks make DSPARK work in PD disaggregation with symmetric TP:

1. spec_info.py: build_disagg_draft_input DSPARK branch
2. overlap_utils.py: FutureMap.resolve_future tolerates future_indices=None.

To get a working setup:

• A prefill engine started with --speculative-algorithm DSPARK runs DSparkWorkerV2._forward_prefill on every extend batch, building draft KV from target hidden states exactly like unified serving.
• For DSV4 / SWA, PD transfer now ships draft KV alongside target KV
• --speculative-algorithm DSPARK must be on both engines

Here is our full launch commands:

Prefill:

SGLANG_HICACHE_FILE_BACKEND_STORAGE_DIR=/mnt/nvme/hicache \
sglang serve \
  --model-path deepseek-ai/DeepSeek-V4-Flash-0731 \
  --trust-remote-code \
  --tp 4 \
  --base-gpu-id 0 \
  --moe-runner-backend flashinfer_mxfp4 \
  --disable-flashinfer-autotune \
  --swa-full-tokens-ratio 0.1 \
  --reasoning-parser deepseek-v4 \
  --tool-call-parser deepseekv4 \
  --speculative-algorithm DSPARK \
  --chunked-prefill-size 16384 \
  --mem-fraction-static 0.90 \
  --enable-hierarchical-cache \
  --hicache-ratio 2 \
  --hicache-size 0 \
  --hicache-mem-layout page_first_direct \
  --hicache-io-backend direct \
  --hicache-write-policy write_through \
  --hicache-storage-backend file \
  --hicache-storage-prefetch-policy wait_complete \
  --disaggregation-transfer-backend nixl \
  --disaggregation-mode prefill \
  --disaggregation-bootstrap-port 8998 \
  --host 0.0.0.0 \
  --port 30000

Decode:

sglang serve \
  --model-path deepseek-ai/DeepSeek-V4-Flash-0731 \
  --trust-remote-code \
  --tp 4 \
  --base-gpu-id 4 \
  --moe-runner-backend flashinfer_mxfp4 \
  --disable-flashinfer-autotune \
  --swa-full-tokens-ratio 0.1 \
  --reasoning-parser deepseek-v4 \
  --tool-call-parser deepseekv4 \
  --speculative-algorithm DSPARK \
  --chunked-prefill-size 4096 \
  --mem-fraction-static 0.87 \
  --cuda-graph-max-bs 64 \
  --disaggregation-transfer-backend nixl \
  --disaggregation-mode decode \
  --host 0.0.0.0 \
  --port 30001

Router:

python3 -m sglang_router.launch_router \
  --pd-disaggregation \
  --prefill http://127.0.0.1:30000 8998 \
  --decode http://127.0.0.1:30001 \
  --host 0.0.0.0 \
  --port 8000

---

CI States

Latest PR Test (Base): ❌ Run #30960127898
Latest PR Test (Extra): ❌ Run #30960127702

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[bobcao3]

/tag-and-rerun-ci

[hnyls2002]

/tag-and-rerun-ci

[bobcao3]

Turns out the DSPARK draft-KV transfer landed upstream already / is in some other PR that landed recently.

[Phoenix3334]

Follow-up regression coverage: #34410 pins the existing DSpark PD-decode draft-input handoff on current main.

The new PR is test-only and does not duplicate the production support discussed here. It asserts the first-decode contract (SpecInput must already exist) and the overlap relay initialization (future_indices / FutureMap) so this integration cannot silently regress.