`apply_weights()`¶

Created: 2026-03-18 · Last updated: 2026-03-18

Apply weights from a state dict to a built ModelPackage.

from mobius import apply_weights

Signature¶

def apply_weights(
    state_dict: dict[str, torch.Tensor],
    pkg: ModelPackage,
    prefix_map: dict[str, str] | None = None,
) -> None:

This is also available as ModelPackage.apply_weights() (the preferred interface):

pkg.apply_weights(state_dict, prefix_map=prefix_map)

Parameters¶

Parameter	Type	Default	Description
`state_dict`	`dict[str, torch.Tensor]`	(required)	Mapping of parameter names to tensors.
`prefix_map`	`dict[str, str] \| None`	`None`	Mapping from weight-name prefix to component name for multi-model packages.

Examples¶

import safetensors.torch
from mobius import build

# Build without weights, then apply manually
pkg = build("meta-llama/Llama-3.2-1B", load_weights=False)

state_dict = safetensors.torch.load_file("model.safetensors")
pkg.apply_weights(state_dict)

Weight Routing¶

For multi-component packages (e.g. vision-language models), use prefix_map to route weights to the correct component:

pkg.apply_weights(state_dict, prefix_map={
    "model.vision": "vision",
    "model.language": "model",
})

Unmatched weights are tried against all components.