vllm_omni.diffusion.models.dmd2 ¶

Modules:

Name	Description
`config`
`mixin`

DMD2Config `dataclass` ¶

Inference-time contract for a FastGen DMD2-distilled checkpoint.

denoising_timesteps: list[int] | None = None

guidance_scale: float = 1.0

num_inference_steps: int = 4

solver: Solver = 'ode'

from_model_index(model_index: dict) -> DMD2Config

Read the dmd2_config block from a model_index.json dict. Missing block → defaults.

resolve_timesteps() -> list[int]

Mixin for FastGen DMD2-distilled models. Must appear before the base pipeline in MRO.

forward(
    req: OmniDiffusionRequest, **kwargs
) -> DiffusionOutput