[CI] Fix test_resolution_is_reproducible after cuda_ipc became opt-in

[mmangkad]

Motivation

test/registered/unit/server_args/test_resolution_is_reproducible.py is failing on the GPU
registrations (seen on base-b-test-1-gpu-small, 1-gpu-5090):

FAIL: test_a_resolution_does_not_leak_into_the_next (intermediate='multimodal')
  File "test/registered/unit/server_args/test_resolution_is_reproducible.py", line 343
    self.assertEqual(after.mm_feature_transport, expected)
AssertionError: 'cpu' != 'cuda_ipc'

#34662 (69bf601e3c, "fix: restore VLM nightly regression coverage") made CUDA IPC opt-in: the
nnodes == 1 arm of _handle_multimodal_feature_transport now resolves to cpu instead of
cuda_ipc, because even an idle IPC pool consumes HBM that would otherwise back the KV cache. That
PR updated the CLI help text and the TestMultimodalFeatureTransport cases in test_server_args.py,
but this test still asserted expected = "cuda_ipc" if is_cuda() else "cpu". The CPU and ROCm
registrations kept passing because they already expect cpu, so only the CUDA runners went red.

The policy change itself is intentional, so the fix belongs in the test.

There is a second, quieter consequence worth calling out. With auto-resolution now picking cpu
everywhere, the multimodal intermediate writes exactly what the next resolution would have picked on
its own, so the documented _STICKY_ACROSS_RESOLUTIONS exception no longer pins anything observable.
Flipping the expectation to cpu and stopping there would have left a green but vacuous assertion.

Modifications

Test-only; no runtime behaviour is changed.

1. Add a CUDA-gated multimodal_cuda_ipc shape (_CUDA_IPC_SHAPES) that requests
   mm_feature_transport="cuda_ipc" explicitly. Post-fix: restore VLM nightly regression coverage #34662 this is the only construction under
   which the SGLANG_USE_CUDA_IPC_TRANSPORT carry is observable — the handler writes 1 only for
   cuda_ipc, and the following text-only resolution adopts it via the legacy_ipc_is_set branch —
   and it is also the only remaining coverage of the handler's cuda_ipc arm, which no
   auto-resolution reaches any more. It is wired into _SHAPES (dual-resolve and sibling matrices)
   as well as the leak test's intermediates. Gated on is_cuda() rather than
   torch.cuda.is_available() because the handler raises for cuda_ipc off NVIDIA CUDA, ROCm
   included.

2. Generalize the sticky assertion to run for every intermediate instead of only the
   label == "multimodal" one (every resolution runs the transport handler, so
   after.mm_feature_transport is pinned for all of them), and derive the expected value from what
   the intermediate itself resolved to rather than from its kwargs. What carries is the legacy
   boolean, not the tri-state field, so any non-cuda_ipc selection — including a cuda_vmm shape
   if one is ever added — carries cpu.

Verification

Run	Result
CUDA host, this branch	4 tests / 16 subtests pass
No visible NVIDIA device (both gates drop out)	4 tests / 10 subtests pass
Pre-change, CUDA host	fails at intermediate='multimodal' — reproduces the CI failure
Mutation check: handler's env write stubbed to always set("0")	fails at intermediate='multimodal_cuda_ipc', so the new shape genuinely guards the carry rather than passing trivially
Mutation check: hypothetical explicit cuda_vmm intermediate	passes under the new derivation; would have wrongly demanded cuda_vmm under a kwargs-derived expectation
Sibling suites: test_server_args.py (142), test_mm_process_config.py (42)	pass
pre-commit run --files <file>	clean

Accuracy Tests

Not applicable — this changes a ServerArgs resolution unit test only; no kernel, model forward, or
output path is touched.

Speed Tests and Profiling

Not applicable — no inference path is touched.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): 🚫 Run #31710106430
Latest PR Test (Extra): ❌ Run #31710105883

[mmangkad]

/rerun-test test/registered/unit/server_args/test_resolution_is_reproducible.py

[github-actions]

Results for /rerun-test test/registered/unit/server_args/test_resolution_is_reproducible.py:

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

cd test/ && python3 registered/unit/server_args/test_resolution_is_reproducible.py

[Kangyan-Zhou]

/tag-and-rerun-ci