Operations — install

morel is published as a wheel and source distribution. This page documents the supported install patterns.

Python

Tested against CPython 3.10, 3.11, 3.12, and 3.13. Earlier versions fail at import time because of PEP 604 int | None annotations used throughout the public API.

Dependencies

The runtime dependencies declared in pyproject.toml:

torch>=2.0,<3
numpy>=1.24,<3
scipy>=1.10,<2
sentence-transformers>=2.2,<7
torchvision>=0.15,<1
Pillow>=10.0,<13
pandas>=2.0,<3
pyyaml>=6.0,<7
pydantic>=2.0,<3

Optional extras

Extra	Installs
`[serve]`	FastAPI, uvicorn, prometheus-client, httpx
`[bench]`	`pytest-benchmark`
`[text]`	sentence-transformers (required for text encoders)
`[vision]`	torchvision (required for visual encoders)
`[dev]`	pytest stack, ruff, mypy, pip-audit, build, twine

Install patterns

PyPI

pip install morel

This installs the runtime dependencies plus the package. The [serve], [bench], [text], and [vision] extras are not pulled in automatically — install the one(s) you need:

pip install 'morel[serve]'
pip install 'morel[serve,bench]'

From source

git clone https://github.com/sachncs/morel.git
cd morel
pip install -e '.[dev]'

The -e editable install lets you iterate on the package without re-installing.

Docker

docker build -t morel:latest .
docker compose up --build

The shipped Dockerfile is multi-stage: a build stage that installs .[dev], and a runtime stage that only carries the non-test dependencies and runs as a non-root user.

Reproducibility

requirements.lock is a hash-pinned lockfile generated from pyproject.toml:

make lock

This is the install flavour CI uses; the lock is checked against pyproject.toml on every PR via make lock-check.

Determinism

morel.core.seed.seed(value) configures deterministic seeding for torch, torch CUDA, numpy, Python random, PYTHONHASHSEED, and cuDNN:

from morel import seed_everything

seed_everything(42)

For a scoped variant (restoring the caller's RNG state on exit):

from morel.core.seed import deterministic

with deterministic(42):
    pipeline = build_pipeline(config)

A run that has been seeded this way is bit-identical across reruns on the same torch version. CUDA non-determinism in torch.sparse.mm is documented under Reference: limitations.

Verification

After installing, run the smoke test:

python -m pytest tests/unit -x

The full suite takes a few minutes and asserts the unit, property, and integration paths. CI runs the same suite on every push.