openbim-ifc (Python)

Read, edit and write IFC STEP files from Python. A thin binding over the openbim-ifc Rust crates: one abi3 wheel for CPython 3.9 and later.

pip install openbim-ifc

Documentation: Python guide · source and issues

from openbim_ifc import IfcModel, Text

with open("model.ifc", "rb") as f:
    model = IfcModel.parse(f.read())

print(model.schema, len(model))
for id in model.ids_of_type_including_subtypes("IfcWall"):
    print(id, model.attribute(id, 2))      # Name, e.g. Text('Wall')

model.set_attribute(id, 2, Text("Renamed"))
with open("out.ifc", "wb") as f:
    f.write(model.write())

Values

Attribute values are frozen dataclasses, one per STEP form: Null ($), Derived (*), Bool, Unknown (.U.), Integer, Real, Text, Binary, Enum, Ref (#42), List, and Typed (IFCLENGTHMEASURE(2.5)). They keep every distinction IFC makes, so a file read and written back through Python is unchanged.

Bare Python values are refused with TypeError: 3 could be an Integer or a Real, "x" a Text or an Enum, and the binding does not guess.

Errors

Every failure raises IfcError with a stable code: parse, write, missing-entity, invalid-value, out-of-range, unsupported-schema, io, unsupported-profile or feature-disabled. The codes are shared with the JavaScript and C bindings; a code is never renamed or reused.

Threads

parse releases the GIL while it reads. A model can be passed between threads; the GIL serialises calls on it.

Scope

The record model -- entities, attributes, the STEP codec, exact and subtype queries -- plus lenient reads (ParseOptions), the STEP header, validation, ifcXML and the reachability lint, each as frozen dataclasses. Domain views (properties, quantities, geometry) and checked multi-edit transactions are not bound yet.

Build from source

uv venv && . .venv/bin/activate
uv pip install maturin
maturin develop --release     # or: scripts/check-python.sh to build + test

Faster reading (opt-in)

maturin build --release --features rusty_alloc

Uses the pure-Rust rusty_alloc allocator for the extension's own memory: reading STEP into a model takes 18-35% less CPU time on seven real IFC files, at 1-7% less peak memory (#49). Python's own allocator is not replaced. Off by default because a global allocator is a build-time choice; the version is pinned exactly.