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
fix(shortlist): refuse an agent that has no tok or model.encoder
Aashish4985bee8c826
diff --git a/laya/shortlist.py b/laya/shortlist.pyindex 24614d12..2d984aa6 100644--- a/laya/shortlist.py+++ b/laya/shortlist.py@@ -248,10 +248,16 @@ def embed_fn_from_agent(if isinstance(batch_size, bool) or not isinstance(batch_size, int) or batch_size < 1:raise ValueError("batch_size must be a positive integer, got %r" % (batch_size,))- import torch+ # The two attributes the docstring names are the only ones here that were never checked: an+ # agent that is not a loaded checkpoint answered with the interpreter's own AttributeError, the+ # one leak in this function, while `_call_predict` below refuses the same mistake for the agent+ # it does use. Checked before `import torch` so a wrong argument does not pay for the import.+ tok = getattr(agent, "tok", None)+ encoder = getattr(getattr(agent, "model", None), "encoder", None)+ if tok is None or encoder is None:+ raise TypeError("agent must provide tok and model.encoder")- tok = agent.tok- encoder = agent.model.encoder+ import torchdef embed_fn(texts: Sequence[str]) -> np.ndarray:rows = ["" if text is None else str(text) for text in texts]diff --git a/tests/test_shortlist.py b/tests/test_shortlist.pyindex d06f2c3f..db64588a 100644--- a/tests/test_shortlist.py+++ b/tests/test_shortlist.py@@ -496,6 +496,39 @@ check_raises("encoder/bad max_length", lambda: embed_fn_from_agent(tiny, max_lencheck_raises("encoder/bad batch_size", lambda: embed_fn_from_agent(tiny, batch_size=True))+# The two attributes the docstring names -- `agent.tok`, `agent.model.encoder` -- were read with no+# check, so an object that is not a loaded checkpoint answered with the interpreter's own+# AttributeError instead of naming what the helper needs. `_call_predict` refuses the same mistake+# for the agent *it* uses, which is the sibling this now matches.+class NoAttrs:+ pass+++class TokOnly:+ def __init__(self):+ self.tok = TinyTok()+++class ModelWithoutEncoder:+ def __init__(self):+ self.tok = TinyTok()+ self.model = type("M", (), {})()+++for _bad_agent in (NoAttrs(), TokOnly(), ModelWithoutEncoder()):+ check_raises("encoder/agent/%s refused" % type(_bad_agent).__name__,+ lambda _a=_bad_agent: embed_fn_from_agent(_a), TypeError)++try:+ embed_fn_from_agent(NoAttrs())+ FAIL.append("encoder/agent refusal message: no TypeError")+except TypeError as _exc:+ check_true("encoder/agent refusal names both attributes",+ str(_exc) == "agent must provide tok and model.encoder", repr(str(_exc)))+except AttributeError as _exc:+ FAIL.append("encoder/agent refusal message: leaked AttributeError (%s)" % _exc)++# ---------------------------------------------------------------- device changes, simulated entirely on CPUclass DeviceRecordingTensor:def __init__(self, tensor, requests):