pycodelift

Build-time class-inheritance flattener. Write thin diff-only child classes against a base file; get back a single self-contained module with the inheritance inlined.

Generalizes HuggingFace transformers' modular codegen into a framework-agnostic library. Status: alpha.

When this is worth using

When all three apply:

1. Many sibling files share a base (10+).
2. They are read and forked, not just called.
3. Edits to the base must not silently ripple into already-released siblings.

For five subclasses in a normal app, plain Python inheritance is the right answer.

Install

pip install pycodelift     # or: uv add pycodelift

Quickstart

mypkg/
├── base/modeling_base.py       # parent
└── child/modular_child.py      # diff

mypkg/child/modular_child.py:

from ..base.modeling_base import BaseModel

class ChildModel(BaseModel):
    def __init__(self, config):
        super().__init__(config)
        self.y = 2

__all__ = ["ChildModel"]

Convert:

pycodelift convert mypkg/child/modular_child.py --package mypkg

Or:

from pycodelift import Flattener
files = Flattener(package_name="mypkg").convert("mypkg/child/modular_child.py")
print(files["modeling"])

The output modeling_child.py has BaseModel's body inlined and no inheritance edge.

How it works

Eight CST passes: parse → walk → resolve parent classes → inline-flatten one level → case-preserving rename → pull in transitive helpers → split outputs by class-suffix → render. Each pass with code examples in docs/how-it-works.md.

Policies

Flattener(...) keyword args (all optional):

flatten_levels=2 and non-empty special_bases raise NotImplementedError in v0.1.

transformers compatibility

from pycodelift.adapters.transformers import make_transformers_flattener

flattener = make_transformers_flattener()
files = flattener.convert("src/transformers/models/olmo2/modular_olmo2.py")

The adapter pins every transformers convention so output matches upstream utils/modular_model_converter.py. CLI: --transformers. Combine with --package optimum for transformers forks.

CLI

pycodelift convert <file>...      # one or many
pycodelift convert-tree <root>     # recursive
pycodelift describe                # dump effective config as JSON

Flags: --package, --transformers, --write, --target-dir.

Caveats

• One-level flatten only. Foo2(Foo) works; Foo's own parent is not inlined.
• Not thread-safe (v0.1 mutates module-level vendor state). Use a process pool for parallelism.
• del self.x is line-level — removes the assignment, not later references.

Examples

Three runnable mini-projects under examples/: 01_minimal, 02_split_outputs, 03_transformers_compat.

Lineage

Derived from huggingface/transformers's utils/modular_model_converter.py (Apache-2.0). The CST passes are theirs (battle-tested across 200+ models); pycodelift extracts the policy slots they hard-coded. See NOTICE.

License

Apache-2.0.