The generator integration relied on the Transformers model-loading and batched-input conventions to prepare a pinned Parler-TTS model for offline inference. The code path was reviewed and compiled, but the model dependencies and weights were unavailable locally, so inference was not executed.
- What worked
- The API shape supported immutable revision selection and moving encoded inputs to the selected device.
- What got in the way
- Device placement and real model loading could not be validated without the internal GPU image and mirrored weights.