[ycjcx123]

Required human contributor answers

1. What outcome are you trying to achieve?

Update the stale WCPB prelude-frontier assertion in the existing split-prefill
CUDA Graph test so it validates the intended contract instead of failing before
the Graph branch.

2. What is the highest-leverage variable for that outcome?

The correctness of the test expectation for the WCPB prelude frontier relative
to the scheduler contract.

3. How will this work move that variable rather than create busywork?

A one-line test-only assertion update, followed by the full eager and CUDA Graph
capture/replay run. No production behavior changes.

4. Who is the customer or user, and what do they actually need?

Maintainers and future developers who rely on this regression test need it to
exercise the real Graph path rather than fail on a stale expectation.

PR necessity and bugfix evidence

A separate test-only PR is needed because the existing test assertion predates
WCPB. @LyxWxj confirmed that num_preluded_tokens == 4 is intended here: with
total_ut_steps=4, prefill_chunk_size=2 and a 4-token budget, one WCPB prelude
batch stages both chunks. num_prefilled_tokens remains 0 because the final
recurrent depth has not completed.

The previous assertion == 2 is stale for WCPB. This PR does not change
vllm_rlt/ and is independent of M7 #153.

Test Plan

• Source: 28085c6 on fix/wcpb-split-prefill-graph-assertion
• Base: 8fbf4cd
• Hardware: 1×RTX 5090, 32 GB, SM120
• Environment: Python 3.12.3, PyTorch 2.8.0+cu128, Triton 3.4.0, pytest 8.4.2,
  ruff 0.13.2

CUDA_VISIBLE_DEVICES=<reserved-device> PYTHONPATH=. \
  python -m pytest -q --run-gpu \
  tests/test_split_prefill.py::test_cuda_split_prefill_survives_decode_graph_capture_and_replay

PYTHONPATH=. python -m pytest -q -m "not gpu" tests/test_split_prefill.py
python -m ruff check tests/test_split_prefill.py
python -m ruff format --check tests/test_split_prefill.py

Test Result

Selection	Result
Full eager + CUDA Graph capture/replay test	1 passed, 1.84 s
tests/test_split_prefill.py -m "not gpu"	47 passed, 32 deselected
Ruff lint / format	Passed

The test now reaches and completes its CUDA Graph capture/replay portion.