vllm_omni.diffusion.models.cosmos3.mixed_precision.strategy ¶

Format-specific live-weight validation and reference A16 materialization.

Cosmos3PrecisionStrategy ¶

Quantization-format boundary behind the shared precision schedule.

accepts(method: object | None) -> bool

Return whether this strategy owns a native linear method.

apply_high(
    layer: Module, x: Tensor, bias: Tensor | None
) -> Tensor

Apply the reference dense A16 path.

materialize(layer: Module) -> Tensor

Materialize the live quantized weight as a dense BF16 matrix.

validate_after_processing(
    layer: Module, module_name: str
) -> None

Validate the live native weight representation.

validate_before_processing(
    method: object, layer: Module, module_name: str
) -> None

Reject native transforms that cannot support live dequantization.

Use native ModelOpt W8A8 or reference dense W8A16.

accepts(method: object | None) -> bool

materialize(layer: Module) -> Tensor

validate_after_processing(
    layer: Module, module_name: str
) -> None

validate_before_processing(
    method: object, layer: Module, module_name: str
) -> None

Use native ModelOpt W4A4 or reference dense W4A16.

accepts(method: object | None) -> bool

materialize(layer: Module) -> Tensor

validate_after_processing(
    layer: Module, module_name: str
) -> None

validate_before_processing(
    method: object, layer: Module, module_name: str
) -> None