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
test(model): skip the bf16 legs on a Windows CPU, which SIGILLs
thread3d62ac667b6d61
diff --git a/tests/test_decision_model.py b/tests/test_decision_model.pyindex 720c3499..d50a932f 100644--- a/tests/test_decision_model.py+++ b/tests/test_decision_model.py@@ -101,6 +101,13 @@ def test_action_head_explicit_dtype_and_autocast():for amp in (None, torch.bfloat16, torch.float16):if device == "cpu" and amp == torch.float16:continue+ # A bf16 forward on a Windows CPU can die with an uncatchable SIGILL on recent+ # torch builds (pytorch/pytorch#152385); sentence-transformers and TabPFN skip+ # the same combination. It takes the whole `windows-latest` lane down, and+ # Linux/macOS CPU and CUDA still run the bf16 legs.+ if os.name == "nt" and device == "cpu" and (+ weights == torch.bfloat16 or amp == torch.bfloat16):+ continuemodel = _tiny_model().to(device=device, dtype=weights)inputs = tuple(t.to(device) for t in _inputs(2, 10, 4))captured = []