Filed as the ## Nothing lands dead disclosure owed by the #889 repair, not as a request to wire anything.
Ltx2Schedule(Ltx2SchedulerKind, ...) (include/vllm/model_executor/models/ltx2_pipeline.h, defined at src/vllm/model_executor/models/ltx2_pipeline.cpp:191) is the scheduler-kind dispatcher. It has zero call sites in src/, include/ or examples/. The engine calls Ltx2SigmaSchedule directly at src/vllm/multimodal/ltx2_video.cpp:1665. The only driver is tests/vllm/models/test_ltx2_pipeline.cpp:318-324, which constructs the enumerator by hand — the "test-only driver" shape in .agents/reachability.md.
#889 reclassifies the kBeta arm of that dispatcher from a reachable refusal to a declared marker, which is the honest label for what is there. It deliberately does not decide the dispatcher's own fate, because that is a behavior change with its own red-first evidence, not a review-finding repair.
The decision this owes. Upstream has no scheduler selection to mirror: at Lightricks/LTX-2 fd4ded7f every one of the seven ltx-pipelines entry points hard-codes LTX2Scheduler(), and BetaScheduler (ltx-core components/schedulers.py:91) is constructed nowhere. vLLM-Omni a4ea67a21 has zero hits. So there is no wiring wave coming that would make Ltx2Schedule reachable — which means the two live options are:
- Retire it, the way
kMultishot and kVideoEngineWiring were retired by row LTX25-RETIRE-DEAD-ARMS. Ltx2SigmaSchedule and Ltx2LinearQuadraticSchedule are both separately callable and separately gated, so nothing ported is lost. kBetaScheduler would then be recorded only as an unported upstream component, with no dispatcher arm.
- Keep it as a deliberately unreached mirror of upstream's
SchedulerProtocol, on the argument that the enum is the ported surface of a real upstream abstraction and the kBeta arm is where the unported third scheduler is recorded.
Naming which, with the argument, is what is owed. Listed under ## Owed in specs/ltx25-retire-dead-arms.md.
Filed as the
## Nothing lands deaddisclosure owed by the #889 repair, not as a request to wire anything.Ltx2Schedule(Ltx2SchedulerKind, ...)(include/vllm/model_executor/models/ltx2_pipeline.h, defined atsrc/vllm/model_executor/models/ltx2_pipeline.cpp:191) is the scheduler-kind dispatcher. It has zero call sites insrc/,include/orexamples/. The engine callsLtx2SigmaScheduledirectly atsrc/vllm/multimodal/ltx2_video.cpp:1665. The only driver istests/vllm/models/test_ltx2_pipeline.cpp:318-324, which constructs the enumerator by hand — the "test-only driver" shape in.agents/reachability.md.#889 reclassifies the
kBetaarm of that dispatcher from a reachable refusal to a declared marker, which is the honest label for what is there. It deliberately does not decide the dispatcher's own fate, because that is a behavior change with its own red-first evidence, not a review-finding repair.The decision this owes. Upstream has no scheduler selection to mirror: at Lightricks/LTX-2
fd4ded7fevery one of the sevenltx-pipelinesentry points hard-codesLTX2Scheduler(), andBetaScheduler(ltx-core components/schedulers.py:91) is constructed nowhere. vLLM-Omnia4ea67a21has zero hits. So there is no wiring wave coming that would makeLtx2Schedulereachable — which means the two live options are:kMultishotandkVideoEngineWiringwere retired by rowLTX25-RETIRE-DEAD-ARMS.Ltx2SigmaScheduleandLtx2LinearQuadraticScheduleare both separately callable and separately gated, so nothing ported is lost.kBetaSchedulerwould then be recorded only as an unported upstream component, with no dispatcher arm.SchedulerProtocol, on the argument that the enum is the ported surface of a real upstream abstraction and thekBetaarm is where the unported third scheduler is recorded.Naming which, with the argument, is what is owed. Listed under
## Owedinspecs/ltx25-retire-dead-arms.md.