somatize-runtime
Execution: GraphSession, the node catalog, caching, streaming, search and training strategies.
Where things actually run. GraphSession is the primary orchestrator —
graph plus catalog plus cache plus events.
NodeCatalog is the registry: every node, filter or step, and the
trained states. Filters and steps used to live in two registries joined by
an adapter a caller had to remember to build, which is how .compile()
came to skip every step's schema while .run() checked them.
run_node is the one execution site — input resolution, the output cache,
catch_unwind and the start/complete/fail events happen once for filters
and steps alike. Its three primitives are what StreamRun composes per
chunk, so streaming is the same execution site rather than a parallel one.
Also here: LRU/local/tiered caches, the Grid/Random/Bayesian samplers,
Median/Percentile pruners, StudyRunner, PbtRunner, the effect driver
and journal, and the training strategies (local, data-parallel,
model-parallel and federated all run).
Part of Soma, a computational
graph runtime for research pipelines, agent orchestration and data
virtualization. Most users want the somatize
facade or the Python package rather
than this crate on its own.
- Guides and design notes: https://manucouto1.github.io/soma/
- API documentation: https://docs.rs/somatize-runtime
Licensed under the Elastic License 2.0.