Runner interface contract

Recipes use the vocabulary from the shared controls. Public LTX image/video and Qwen entries expose the controls implemented by each route. The two native entries' complete shared DDO vocabulary is in the native image interface. Workload-profile, repeat and native allocation experiments use dedicated diagnostic entries. Public LTX image entries share the custom-block loader and unified LoRA interface with public LTX VMM; SDNQ preserves its packed forward and exposes no unimplemented dynamic encoder switch.

An explicit transformer selector takes precedence over an environment preset. For Windows purge, explicit positive/negative CLI wins, then the corresponding DDO_PURGE_WINDOWS_STANDBY_<PHASE> variable, then the preset default. Unsupported phase switches must be rejected, not accepted as no-ops.

Image --ddo-vmm-total-staging-gb specifies total reusable host staging in GiB. The runner converts that capacity to the slots required by the selected route. Video uses the same total-capacity vocabulary and delegates staging conversion to the public DDO backend configuration. Public native video uses the installed allocator bootstrap; isolated DLL selection belongs to diagnostics. Video source roles and supported subsets are in the video interface matrix.

Preserve official Diffusers and quantized backend forwards when selecting compatibility runtime routes. Parser tests must verify supported controls, unsupported-option errors and unchanged runtime destinations.

Custom code identity

A Hub revision selects downloaded executable code; it does not switch an existing local clone. Reports distinguish the requested revision, resolved snapshot and local clone HEAD. Public LTX video loaders use Hub main when the revision is omitted, and pass an explicit branch, tag or commit through unchanged. Benchmarks explicitly select immutable code revisions.

Video decode defaults

Video runners currently default to CPU decode without explicit temporal streaming. Select --decode-device cuda:0 --stream-video-decode for the GPU temporal decoder. This is independent of transformer precision and offload preset. Refer to the video guide for model-specific controls.