Skip to content

1.4.19

Choose a tag to compare

@GBeurier GBeurier released this 18 Nov 08:59
526052d

Changed

  • The SDK depends on the other nirs4all repositories at minor granularity
    (dag-ml>=0.4,<0.5, dag-ml-data>=0.2,<0.3, nirs4all-io>=0.3,<0.4,
    nirs4all-core>=0.5,<0.6, nirs4all-methods/pls4all>=1.3,<2), following the
    ecosystem version-constraint contract. Exact versions stay in the install
    cohorts: docker/native-constraints.txt, the release image check, the
    installed-example runtime and the installed-package smokes now use dag-ml
    0.4.3, nirs4all-io 0.3.6 and nirs4all-core 0.5.1.
    DAG-ML 0.4.3, and Core 0.5.1 whose native package writer is built on it, accept
    the ddof parameter that Methods >= 1.3.15 records in standard_scale
    classifier states. With DAG-ML <= 0.4.2 or Core <= 0.5.0 the portable export
    of a multimodal classifier that scales a source is refused, so install
    DAG-ML 0.4.3 and Core 0.5.1 or later.
  • The Studio Playground callable (nirs4all.api.library_playground.studio_playground_job_v1)
    no longer imposes content-size ceilings. MAX_PLAYGROUND_REQUEST_BYTES (8 MiB)
    and MAX_PLAYGROUND_RESPONSE_BYTES (32 MiB) and their request_too_large and
    response_too_large refusals are removed: request and response byte budgets
    belong to the host transport. The closed request contract, nesting depth, finite
    JSON values, identifier bounds and operator authorization are still checked.
    The fixed limit of 50 Playground steps is removed as well; a host may still bound
    the step count with options.max_steps.
  • read_model_catalogue(workspace_path, max_models=None) reads every chain by
    default. max_models is an optional host bound (any positive integer) and a
    catalogue larger than the bound is still refused, never truncated; the fixed
    10,000-model ceiling is removed.
  • Hosted CI runs pytest again: the tests job of CI.yaml runs every suite
    except tests/integration/{parity,api,pipeline} and tests/e2e on each push
    and pull request, and tests-full.yml runs the whole of tests/ nightly, on
    demand and when it changes, in three shards with the dag-ml-cli of the
    installed DAG-ML release and an installed-wheel interpreter.

Fixed

  • ResidualModel on the legacy engine no longer fails at random with
    Processing 'residual__<base>_<hex>' already exists when it runs more than once
    on a dataset (branches, variants). The residual target was named after 16 bits
    of an object address, which CPython reuses; it is now numbered per dataset.

  • SpectroDataset.content_hash() and nirs4all.utils.hashing.compute_data_hash
    always return the SHA-256 digest (64 hex characters) of the feature bytes. They
    used the 128-bit xxhash digest (32 hex characters) whenever the undeclared
    optional xxhash package happened to be importable, so the dataset identity
    persisted in workspace runs, bundle fingerprints and Studio provenance depended
    on the environment, and validate_studio_provenance refused every Studio run
    in such an environment. Identities recorded by an environment that had xxhash
    (32 characters) are not converted or matched: the same data now has its SHA-256
    identity there, as it always had in a plain installation.

  • Component serialization no longer reads user-chosen mapping keys as fields of
    a step envelope. Branch names in a named branch mapping, the per-branch
    steps and values keys of a separation branch (source names, partition
    values, group names) and tag names in a named tag mapping are names, whatever
    they spell. A source, branch, partition or tag named name, description,
    metadata, _metadata_, _tags_, class, function or instance
    previously had its steps restored as inert strings and dictionaries instead of
    components, so a serialized pipeline using it was refused by engine="dag-ml"
    (run/unsupported_shape, for example a by_source stack over a source named
    metadata) while the same live objects trained. Under names spelling an
    argument field (params, args, *_params) the descriptive fields of the
    nested steps were resolved as references. Only pipelines using such a name
    change: their bodies now deserialize like under any other name, and string
    references under them are canonicalized on serialization, which changes the
    identity hash of those pipelines when they are declared as non-canonical
    strings. Descriptive fields of step and component envelopes, _-prefixed keys
    of a duplication mapping and generator annotations remain literal.

  • Repair the test suite, which no hosted workflow had run since 2026-10-08: tests
    read host bundles and native packages through the Core V1 container, the
    RunResult store double honours fold_id, the Methods binding identity table
    records 1.3.18, the SNV/Savitzky-Golay oracles evaluate in float64 like Methods,
    non-finite training controls are expected to round-trip, and the frozen
    retrain/load_calibrated_result signatures follow the current API. Schema
    validation of runtime envelopes skips with a stated reason when the sibling
    dag-ml ScoreSet/SelectionDecision schemas are not checked out, instead of
    failing on an unresolvable reference.

Qualified wheel sha256 94061f9b5379c428fdf15d8b89b72b2c50171ceb871af88c88a694c55e6691c9 (source 32df220).