lvsa-add-model

To add a new video diffusion model to LVSA, implement a dedicated ModelAdapter by copying an existing template like wan.py, wire it into the adapter registry, and provide a minimal example wrapper to plug it into the generation pipeline.

Implementing a video diffusion model adapter requires following an 11-method template to handle geometry, QKV, and RoPE, then wiring the adapter into the example generation script and adding vllm-omni hooks for special pre-attention behavior if needed.

LVSA supports Wan-like single-stream, HV-like dual-stream, and CogVideoX-style joint-attention video diffusion models by choosing the appropriate integration path, such as an ABC adapter for single or joint streams or a processor swap for asymmetric attention.

Validate a new video diffusion model adapter before GPU deployment by running CPU smoke tests to ensure the model's generation script and the 11-method adapter implementation function correctly within the LVSA pipeline.

For video diffusion models with asymmetric attention, use the processor swap integration path instead of the ABC adapter to properly handle the custom pre-attention behavior within the LVSA adapters framework.

You can copy the closest existing LVSA adapter, such as wan.py or hunyuan_video.py, as a template to start implementing your model-specific adapter and required methods for geometry, QKV, and RoPE.