Skip to content

vllm_omni.diffusion.models.minimax_h3.fasth3_checkpoint

Sampling contract for FastVideo FastH3 full checkpoints.

FASTH3_V2_BASE_SCHEDULE module-attribute 

FASTH3_V2_BASE_SCHEDULE = DMD2SigmaSchedule.from_positions(
    (
        0.999,
        0.874,
        0.749,
        0.624,
        0.5,
        0.375,
        0.25,
        0.125,
        0.0,
    )
)

FASTH3_V2_MODEL_ID module-attribute 

FASTH3_V2_MODEL_ID = 'FastVideo/FastVideo-FastH3-8-Step-V2'

FastH3CheckpointSpec dataclass

The full V2 release has its own schedule and trained attention policy.

This is independent of FastH3WeightFusion: the released weights already contain the entire student, including its learned compression gates.

vsa_sparsity class-attribute instance-attribute 

vsa_sparsity: float = 0.8

check_request 

check_request(
    sampling: OmniDiffusionSamplingParams,
    *,
    step_execution: bool = False,
) -> None

check_serving_contract 

check_serving_contract(
    *, partition: str, od_config: OmniDiffusionConfig
) -> None

from_metadata classmethod 

from_metadata(
    metadata: Mapping[str, object],
) -> FastH3CheckpointSpec

Validate the release's own fastvideo_inference.json.

release_metadata 

release_metadata() -> dict[str, Any]

Express the sampling policy in the existing H3 pipeline schema.

resolve_native_vaes 

resolve_native_vaes(model_root: Path) -> Path

Reuse the frozen base VAEs with Omni's native tiled/parallel runtime.

Only VAE components are fetched; the student and text encoder come directly from the FastVideo release. The release pins the base revision.