Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.
Python calibrationclassificationdecision-modelhuggingfacejevmodernbertmultilingualnlppythonpytorchroutingtyped-decisionszero-shot
docs(revisions): name the blank-revision edge the suite already pins
Aashish8417d964ffe2
diff --git a/laya/revisions.py b/laya/revisions.pyindex 85a31f3d..8a8dc805 100644--- a/laya/revisions.py+++ b/laya/revisions.py@@ -57,8 +57,12 @@ def _reviewed_pin(model_id_or_path: str) -> Optional[str]:def resolve_revision(model_id_or_path: str, revision: Optional[str] = None) -> Optional[str]:"""Pick the revision to download.- An explicit `revision` is returned unchanged. Otherwise ``LAYA_REVISION``, stripped; empty or- unset means "not asked for". The variable holds either a commit SHA/branch/tag, applied to+ Both the argument and ``LAYA_REVISION`` are stripped, and the argument wins when it is truthy.+ That has one sharp edge, pinned by `tests/test_revision_pinning.py`: a truthy-but-blank+ `revision` (" ") wins the `or` and then strips away, so it suppresses the environment and+ returns None -- a blank argument is "no pin from me, and no pin from the deployment either",+ not a way to ask for ``LAYA_REVISION``. An empty or unset one means "not asked for", so the+ variable applies. The variable holds either a commit SHA/branch/tag, applied toevery checkpoint load the way `Router(revision=)` applies one, or the word ``reviewed``, whichlooks `model_id_or_path` up in `PINNED_REVISIONS` -- case-insensitively, the way the Hubitself reads repo ids, so `laya.load("ConvaiInnovations/Laya")` finds its pin.