LTX-2.5: A2VidPipelineTwoStage has no recipe row, so audio-to-video rides the distilled two-stage trajectory

[localai-bot]

A2VidPipelineTwoStage (a2vid_two_stage.py:53 @ fd4ded7f) has no recipe row. Its audio conditioning landed with #922 and is reachable today through the audio_path per-generation extra; its recipe — the two-stage schedule, the guiders and the required audio input — does not exist, so pipeline_kind = a2vid_two_stage gets the generic table refusal at src/vllm/model_executor/models/ltx2_pipeline.cpp:1328-1332 naming the pair rather than the missing machinery. Measured at daeff67f2: git grep -n '"a2vid' over src include tests docs examples returns TWO hits, both upstream anchors inside Fail message assertions (tests/vllm/multimodal/test_ltx2_video.cpp:4363,:4427), against a control of 4 hits for "one_stage" in include/ alone. A closed conditioning issue is not a recipe.

What a supplied take rides today is distilled_two_stage, which is a DIFFERENT trajectory and the a2v spec says so under ## Owed (ltx25-a2v-audio-input.md:438-446). Four differences, each read at the pin:

1. Stage 1 is guided and caller-configured. GuidedDenoiser with MultiModalGuider(params=video_guider_params, negative_context=v_context_n) (a2vid_two_stage.py:230-240), whose six fields come from the params table through the CLI defaults (utils/args.py:947-996, --a2v-guidance-scale defaulting to video_guider.modality_scale at :987-996, which is 3.0 at utils/constants.py:54,:64). distilled_two_stage fixes allow_guidance_override = false and carries the default positive-only guider on both phases.
2. Stage 1's schedule is scheduler-derived, self._scheduler.execute(steps=num_inference_steps) (:225-227), against our fixed DistilledSigmas() (ltx2_pipeline.cpp:1163).
3. Stage 1 is plain Euler. a2vid_two_stage.py:229-258 passes no stepper, so DiffusionStage.__call__'s EulerDiffusionStep() default applies (utils/blocks.py:524-527); our distilled_two_stage selects kEulerAncestral for stage 1 on generation 2.5 (distilled.py:76-84).
4. The audio stream's guider is the DEFAULT one, MultiModalGuiderParams() at :237-239 — cfg 1.0 / stg 0.0 / modality 1.0 (guiders.py:199-211) — and NOT the params table's audio row, which carries cfg 7.0.

Two things that are not schedule and must not be guessed:

• The audio file is REQUIRED. --audio-path is required=True (:312-317). Nothing in this tree makes a pipeline_kind demand a per-generation extra today, so an a2vid load with no take renders unconditioned video and looks like the feature working.
• The distilled LoRA rides stage 2 ALONE. stage_2_loras = (*loras, *distilled_lora) at :114, against loras=tuple(loras) for stage 1 at :107; --distilled-lora is required=True (utils/args.py:1140-1153). This engine fuses at LOAD into ONE weight set (src/vllm/multimodal/ltx2_video.cpp:816-820, the only dit_options.loras.push_back in the tree), so the per-phase adapter cannot be expressed. Same seam blocks #1093 (ti2vid_two_stages.py:140,:151) and #921 (ti2vid_two_stages_hq.py:154,:165, two strengths). distilled.py:131 builds ONE stage set, which is why no shipped recipe has needed it until now.

The blocker this row waited on is gone: Ltx2GuidedDenoise landed at daeff67f2 (#1092/#1102) with four production callers in src/vllm/multimodal/ltx2_video.cpp, and .agents/specs/ltx25-guided-video.md §1 lists a2vid_two_stage under ## Owed as "its own row; none is blocked on this seam any more". ltx-2.5-22b-distilled-lora-450-bf16.safetensors — the artifact #1093's control could not find — is now on the NAS.

Scope of this row: the recipe row, the required-audio refusal, the required-LoRA refusal mirroring --distilled-lora, and a gate that the frozen take is CONSUMED on the a2vid arm rather than merely carried. NOT in scope and owed: the per-phase adapter, and any real-checkpoint render (the GPU is out of bounds for this row).