Development¶

git clone https://github.com/sirmmo/openai-maple
cd openai-maple
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install -e '.[test]'

Tests¶

Three tiers, selected with pytest markers:

Command	What runs	Needs
`pytest`	offline suite: fake engine behind the HTTP surface, SSE and parser tests, OpenAI SDK round trip, numerical checks of the attention shim and the ternary packing	nothing
`pytest -m network`	builds a two-layer random-init Maple from the HF-hosted custom code and runs real `generate()` through the engine: streaming, stop, cancel, seeds, packing	a few MB from HuggingFace, cached
`OPENAI_MAPLE_URL=http://127.0.0.1:8000 pytest -m live`	end-to-end against a running server with the real weights	a server; minutes per call on CPU

CI runs the first two on Python 3.10 and 3.12, plus ruff check and ruff format --check.

Iterating on the generation path¶

Do not reload 40 GB to debug generate(). The tiny_model_dir fixture in tests/test_engine_tiny.py writes a 2-layer, 8-expert, 64-wide Maple (about 90 MB) that goes through the same custom modeling_maple.py, the shim, the fa3.py patch and the DynamicCache path. Both upstream bugs reproduce on it in seconds.

Releases¶

Pushes to main publish ghcr.io/sirmmo/openai-maple:latest (CPU, amd64 + arm64), :latest-cuda and immutable sha-* tags (.github/workflows/docker.yml).
A vX.Y.Z tag additionally builds the sdist and wheel, publishes to PyPI through trusted publishing, creates a GitHub release, and tags the images X.Y.Z, X.Y and their -cuda twins (.github/workflows/release.yml). workflow_dispatch on that workflow is a dry run: it builds and checks the distributions and stops.
This site is built with MkDocs Material from docs/ on every push that touches it (.github/workflows/docs.yml).

Bump version in pyproject.toml and openai_maple/__init__.py, add a CHANGELOG.md entry, tag, push.