Skip to content

vllm_omni.diffusion.models.cosmos3.mixed_precision.config

Configuration parsing and denoising-step policy for mixed precision.

ReasonerPolicy module-attribute 

ReasonerPolicy = Literal['native', 'a16']

Cosmos3MixedPrecisionConfig dataclass

Validated precision policy shared by every quantization strategy.

first_steps class-attribute instance-attribute 

first_steps: int = 3

last_steps class-attribute instance-attribute 

last_steps: int = 3

reasoner class-attribute instance-attribute 

reasoner: ReasonerPolicy = 'a16'

from_additional_config classmethod 

from_additional_config(
    additional_config: Mapping[str, object] | None,
) -> Cosmos3MixedPrecisionConfig | None

Parse Cosmos3 fields from vLLM-Omni's additional configuration.

resolve_additional_config classmethod 

resolve_additional_config(
    additional_config: Mapping[str, object] | None,
) -> tuple[bool, Cosmos3MixedPrecisionConfig | None]

Return whether an override was supplied and its effective policy.

The separate presence bit distinguishes an absent runtime override from an explicit disable. That distinction matters when a checkpoint owns a default policy.

use_high_precision 

use_high_precision(step_index: int, num_steps: int) -> bool

Select the high path for an actual scheduler-step index.