DRYML

DRYML provides immutable construction definitions, exact graph references, and local checkpoint publication for Python machine-learning objects.

CDef V2

Definition is a soft declaration. Concretizing it produces a ConcreteDefinition (CDef): a fully bound structural recipe. CDef equality and stable_hash() compare class and semantic parameters; they intentionally do not include graph sharing or object identity. CDef.graph_equal() and graph_hash() additionally preserve shared versus independent equal nodes.

ObjectRef adds durable ObjectId lineage for every owned Serializable node. StateRef adds one immutable local-state hash for each of those paths. Inspecting CDefs, ObjectRefs, StateRefs, and structural queries is import-free. CDef .args and .kwargs are call projections and may import the current class; use .cls and .parameters for inspection.

from dryml import Definition, Object, Repo, Serializable
from dryml.core.store.dir import DirStore


class Counter(Serializable):
    def __init__(self, value=0):
        super().__init__()
        self.value = value


repo = Repo(DirStore("./checkpoint-store"))
counter = Counter(3, repo=repo)
state = counter.save(repo=repo)
restored = repo.load_state_ref(state)
assert restored.object_ref == state.object

Object.save() publishes immutable directory state and returns a StateRef. federated=False is self-contained by default; federated=True may retain immutable dependencies in connected Stores. deep_capture=True captures each owned serializable node once. There is no automatic mutation detection: call save() after relevant mutation.

See Objects and Definitions, Repos and Stores, Formats, and Graph Querying.