mcm-mycelium 0.0.0a133

mycelium — experiment contracts, statistical optimisation, and portable executors for small-model training

mycelium

Experiment management for statistical optimisation, with integrations for portable executors and architectures — built for small-model fine-tuning.

Alpha. Versioned as 0.0.0aN prereleases; every release may break the API. Published on PyPI as mcm-mycelium (a PEP 541 request for the bare mycelium name is pending).

Install

uv add mcm-mycelium               # contract layer (reishi) only
uv add 'mcm-mycelium[physarum]'   # + Optuna sweep tooling
uv add 'mcm-mycelium[oyster]'     # + MLX mesh-executor deps (Apple Silicon)
uv add 'mcm-mycelium[enoki]'      # KubeRay executor (deps are image-managed)
uv add 'mcm-mycelium[all]'

One wheel ships all four modules; extras only add each module's third-party dependencies. Importing a module whose extra is missing fails with a plain ModuleNotFoundError naming the dependency.

Modules

Module	What it is
reishi	The experiment contract layer: Task, Dataset, Recipe, Trial, and Board records, storage backends, and the plugin CLI.
physarum	Statistical optimisation: Optuna-backed sweeps over recipe search spaces, with epsilon-constraint objectives and pruning.
enoki	KubeRay (cloud) execution — renders recipes into RayJobs. Its torch/ray stack is installed by the training image, not from PyPI.
oyster	Self-hosted mesh execution on Apple Silicon (MLX LoRA training, GitHub Actions transport).

enoki and physarum register mcm.plugins entry points; the reishi plugin CLI discovers them automatically and warns (rather than fails) when an extra's dependencies are absent.

License

Apache-2.0 — see LICENSE and NOTICE.