[shiyu7]

Motivation

In the DSPARK/DFLASH spec-v2 decode path, the common SWA maintenance call and decode_batch_idx tick were missing. As a result, out-of-window SWA pages accumulated until the generic decode OOM path retracted a request. The active DSV4 SWA KV pool does not implement CPU-copy backup, so retraction raised NotImplementedError and terminated the scheduler.

This complements sgl-project#36185: host-pool-backed resumable retraction remains useful when available, while this change reclaims SWA before retraction and keeps unsupported backup from crashing the whole decode worker.

Modifications

• Move speculative decode SWA/clock bookkeeping into spec_prepare_for_decode(), so EAGLE, DFLASH, and DSPARK share one owner.
• Force an SWA reclamation pass in check_decode_mem() before decode retraction.
• If CPU backup is unsupported, abort only the affected retracted request instead of terminating the scheduler.
• Skip the unusable CPU backup when the final request is aborted for OOM.
• Add CPU unit tests for speculative bookkeeping ownership, forced SWA reclamation, and retraction fail-safes.

Accuracy Tests

Validated on a 2-node DSV4 DSPARK disaggregated-decode setup based on bytedance-iaas/sglang:ep_main:

python -m sglang.test.run_eval \  --port 8080 \  --eval-name gpqa \  --num-examples 192 \  --max-tokens 64000 \  --repeat 3 \  --top-p 0.95 \  --temperature 1.0 \  --thinking-mode deepseek-v3

• Completed all 3 × 192 examples without scheduler exit or retraction.
• Scores: 0.870, 0.885, 0.880; mean 0.878.
• Output throughput: 317.624 token/s.

Unit/static validation:

• test_decode_bookkeeping_ownership.py: 2 passed
• test_spec_prepare_for_decode.py: 2 passed
• test_schedule_batch_swa_retraction.py: 5 passed
• black --check, isort --check-only, py_compile, and git diff --check passed

Speed Tests and Profiling

No isolated speed comparison; the change only adds forced SWA maintenance after the normal memory check has already failed.

Checklist

• Format code according to the contribution guide.
• Add unit tests for the changed scheduler/speculative paths.
• Provide end-to-end GPQA accuracy and throughput results.
• Follow the SGLang code style guidance.
• Documentation update (not applicable; no user-facing API change).

---

CI States

Latest PR Test (Base): ❌ Run #32817721599
Latest PR Test (Extra): ❌ Run #32817721495