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(onnx): answer single-option questions instead of failing in TopK
hemanthrayuduu3313265690be
diff --git a/laya/onnx_agent.py b/laya/onnx_agent.pyindex 714ddd8c..77fa87e5 100644--- a/laya/onnx_agent.py+++ b/laya/onnx_agent.py@@ -775,12 +775,24 @@ class ONNXAgent(HookRegistry):b = collate_items(per_state_items, self.tok.pad_token_id)+ marker_pos = b["marker_pos"].numpy().astype(np.int64)+ marker_mask = b["marker_mask"].numpy().astype(bool)+ if marker_pos.shape[1] < 2:+ # The exported head takes `topk(2)` of the option probabilities: the eager+ # forward's one-option branch is a Python `if`, which the trace bakes in, so+ # ONNX Runtime refuses a run whose questions all have one option. A masked+ # second slot scores -1e4, is sliced off at decode, and gives the act head the+ # zero second probability the eager branch pads in.+ width = ((0, 0), (0, 2 - marker_pos.shape[1]))+ marker_pos = np.pad(marker_pos, width)+ marker_mask = np.pad(marker_mask, width)+# Prepare ONNX inputs as numpy arraysort_inputs = {"input_ids": b["input_ids"].numpy().astype(np.int64),"attention_mask": b["attention_mask"].numpy().astype(np.int64),- "marker_pos": b["marker_pos"].numpy().astype(np.int64),- "marker_mask": b["marker_mask"].numpy().astype(bool),+ "marker_pos": marker_pos,+ "marker_mask": marker_mask,"qtype": b["qtype"].numpy().astype(np.int64),}diff --git a/tests/test_onnx_batch.py b/tests/test_onnx_batch.pyindex a9f6cedf..97d9f58a 100644--- a/tests/test_onnx_batch.py+++ b/tests/test_onnx_batch.py@@ -308,6 +308,60 @@ check("router/results carry the routing key", [sorted(r) for r in routed],check("router/two states share one session run", router_agent.session.calls, [2 * 2])+# ---------------------------------------------------------------- single-option questions+# The exported graph traces the decision head's `topk(2)` (the eager forward's one-option branch is+# a Python `if` on the marker count, which the trace bakes in), so ONNX Runtime refused a run whose+# questions all had one option -- "k argument [2] should not be greater than specified axis dim+# value [1]" -- while the PyTorch Agent answered it. This stub enforces the same rule, and scores a+# masked slot -1e4 the way the graph's `masked_fill` does.+class _ExportedGraphSession:+ def __init__(self):+ self.marker_widths = []+ self.masks = []++ def run(self, names, inputs):+ mask = inputs["marker_mask"]+ self.marker_widths.append(mask.shape[1])+ self.masks.append(mask.tolist())+ if mask.shape[1] < 2:+ raise RuntimeError("[ONNXRuntimeError] : 1 : FAIL : Non-zero status code returned while "+ "running TopK node. Status Message: k argument [2] should not be "+ "greater than specified axis dim value [%d]" % mask.shape[1])+ logits = np.where(mask, 1.0 + np.arange(mask.shape[1], dtype=np.float32), -1e4).astype(np.float32)+ act = np.tile(np.array([[0.25, 0.75]], dtype=np.float32), (len(mask), 1))+ return [logits, act]+++ONE_OPTION = {"dept": {"type": "choice", "instructions": "Which team?", "criteria": {"billing": "money"}}}+ONE_LEVEL = {"level": {"type": "score", "instructions": "How urgent?", "criteria": ["not urgent"]}}++for label, questions, key, want in (("choice", ONE_OPTION, "dept", {"billing": 1.0}),+ ("score", ONE_LEVEL, "level", {"0": 1.0})):+ single = _bare_onnx()+ single.session = _ExportedGraphSession()+ try:+ got = single.predict("aa", questions)["answers"][key]["probabilities"]+ except RuntimeError as e:+ got = str(e)[:80]+ check("single option/a one-option %s is answered, not refused by the graph" % label, got, want)+ check("single option/%s: the extra slot is padded and masked" % label, single.session.masks, [[[True, False]]])++single = _bare_onnx()+single.session = _ExportedGraphSession()+try:+ got = [r["answers"]["dept"]["probabilities"] for r in single.predict_batch(STATES, ONE_OPTION)]+except RuntimeError as e:+ got = str(e)[:80]+check("single option/the batch path answers too, in one run", (got, single.session.marker_widths),+ ([{"billing": 1.0}] * len(STATES), [2]))++wide = _bare_onnx()+wide.session = _ExportedGraphSession()+wide.predict("aa", QUESTIONS)+check("single option/a run that already has two or more slots is sent as before",+ wide.session.marker_widths, [3])++# --------------------------------------------------------------------------- reportprint("\n%d passed, %d failed" % (len(PASS), len(FAIL)))for f in FAIL: