[hnyls2002]

Facts about work in flight now come from what each step recorded when it was built or launched, instead of a per-request counter or a batch snapshot. This fixes a miscount of chunked prefill on main.

The fix

Req.inflight_middle_chunks was raised whenever a chunked request existed after an admission pass that admitted something, even when the chunked request got no chunk in that pass. Under the hybrid-SWA joint prefill budget, a chunk whose remaining prompt does not fit with its new-token headroom is parked, while a smaller waiting request can still be admitted in the same pass. The counter then never came back down, and the request's final chunk was processed as a middle chunk: its first token was not committed (PD prefill: it was requeued as if it had yielded).

The same state made contains_last_prefill_chunk False for a batch whose only row ran a final prefill, so pipeline parallelism with SGLANG_PP_SKIP_PURE_CHUNKED_OUTPUT_COMM skipped its output exchange.

A pause that retracts under pipeline parallelism reset the counter while other micro-batches still carried the request, so their results could be labelled wrongly too; each step's own record labels it by itself.

Changes

• Result processing (prefill, embedding, PD prefill, the pipeline-parallel output check, the grammar skip) asks whether a row is its batch's chunked_req, which the batch already records at build: the request that runs a chunk in this batch and is still chunked after it. A parked request is named there too but is not a row, so it is never asked about. copy() now carries chunked_req, since overlap processes the copy. The counter, its resets and the never-read dLLM increments are removed.
• contains_last_prefill_chunk counts the chunked request only when it is admitted in this pass.
• A PD optimistic request that yields the chunked slot is requeued once no step of it is in flight (Req.inflight_steps): at once at the yield, otherwise when its last step in flight is processed. This replaces the overlap flag and the extend_end inference; both rules agree wherever the yield can happen (optimistic prefill is off with pipeline parallelism, and a request otherwise has at most one step in flight).
• _reqs_finishing_inflight finds running requests whose queued result finishes them by inflight_steps > 0 instead of reading the queued batch copy, and drops the one-batch assert. Same result: only the overlap loop enables it, and there the queued step is the only step in flight.

Behavior changes: only the parked-chunk and pipeline-parallel retract states above.

Test plan

• test_chunked_prefill_does_not_yield_for_decode_admission covers a yield with and without a step in flight.
• The unit tests under test/registered/unit/managers/, test/registered/unit/disaggregation/ and test/registered/unit/dllm/ pass locally with SGLANG_IS_IN_CI=1, apart from failures main has too.

---

CI States

Latest PR Test (Base): 🚫 Run #41560652630
Latest PR Test (Extra): ❌ Run #41560652511
Latest PR Test (AMD ROCm 10): 🚫 Run #41560652473

Tip

Base CI only covers the core cases, and its workload may not exercise your change. Use /rerun-test to run the tests your change touches, or copy this prompt for an agent:

Find the registered tests this PR affects and run them with /rerun-test, following https://github.com/sgl-project/sglang/blob/main/.agents/skills/rerun-test/SKILL.md

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[hnyls2002]

/tag-and-rerun-ci

[hnyls2002]

/rerun-test test_disaggregation_optimistic_prefill.py test_prefill_complete_allocation.py test_scripted_swa_1gpu.py test_scripted_core_1gpu.py test_pp_single_node_extra.py registered/disaggregation/test_disaggregation_pp.py test_retract_decode.py test_dp_spec_prefill_coordination_mixed.py test_disaggregation_hisparse.py test_disaggregation_unified_memory.py

[github-actions]

🚀 2-gpu-h100 (3 tests): ✅ View workflow run

cd test/ && python3 registered/disaggregation/test_disaggregation_optimistic_prefill.py
cd test/ && python3 registered/e2e/disaggregation/test_prefill_complete_allocation.py
cd test/ && python3 registered/disaggregation/test_disaggregation_unified_memory.py

🚀 1-gpu-h100 (1 test): ❌ View workflow run

cd test/ && python3 registered/chunked_prefill/test_scripted_swa_1gpu.py

🚀 1-gpu-5090 (2 tests): ✅ View workflow run

cd test/ && python3 registered/chunked_prefill/test_scripted_core_1gpu.py
cd test/ && python3 registered/scheduler/test_retract_decode.py

🚀 4-gpu-h100 (2 tests): ✅ View workflow run

cd test/ && python3 registered/pp/test_pp_single_node_extra.py
cd test/ && python3 registered/spec/eagle/test_dp_spec_prefill_coordination_mixed.py

🚀 8-gpu-h20 (1 test): ✅ View workflow run

cd test/ && python3 registered/disaggregation/test_disaggregation_pp.py

🚀 8-gpu-h200 (1 test): ✅ View workflow run

cd test/ && python3 registered/disaggregation/test_disaggregation_hisparse.py