build()

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

Build an ONNX ModelPackage from a HuggingFace model ID.

from mobius import build

Signature

def build(
    model_id: str,
    task: str | ModelTask | None = None,
    *,
    module_class: type[nn.Module] | None = None,
    dtype: str | ir.DataType | None = None,
    load_weights: bool = True,
    trust_remote_code: bool = False,
) -> ModelPackage:

Parameters

Parameter	Type	Default	Description
model_id	str	(required)	HuggingFace model repository ID (e.g. "meta-llama/Llama-3.2-1B").
task	str | ModelTask | None	None	Model task (e.g. "text-generation"). Auto-detected when None.
module_class	type[nn.Module] | None	None	Custom module class. Auto-detected from registry when None.
dtype	str | ir.DataType | None	None	Target dtype ("f32", "f16", "bf16"). Auto-detected from HF config when None.
load_weights	bool	True	Whether to download and apply weights from HuggingFace.
trust_remote_code	bool	False	Whether to trust remote code when loading the HF config.

Returns

ModelPackage — A dict-like collection of named ir.Model objects.

Examples

from mobius import build

# Auto-detect architecture and task
pkg = build("meta-llama/Llama-3.2-1B")
pkg.save("output/llama/")

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

# Override dtype
pkg = build("meta-llama/Llama-3.2-1B", dtype="f16")

# Custom module class
pkg = build("meta-llama/Llama-3.2-1B", module_class=MyCustomModule)

Behavior

1. Downloads the HuggingFace config via transformers.AutoConfig

2. Detects model_type and looks up the module class in the registry

3. Falls back to diffusers pipeline detection if not a transformer model

4. Builds the ONNX graph via build_from_module()

5. Downloads and applies weights (if load_weights=True)