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
Merge pull request #501 from PerryLink/fix/onnx-maxlen-message
Nandakishorab525a8aff34
diff --cc laya/onnx_agent.pyindex f53ad4c5,35c544e0..327dd07a--- a/laya/onnx_agent.py+++ b/laya/onnx_agent.py@@@ -485,18 -250,46 +485,29 @@@ class ONNXAgent(HookRegistry)self.tok, state, q, max_len, head_max_len,truncate_left=truncate_left, state_ids=state_ids,)- if len(markers) != len(render_options(q)):- raise ValueError("question %r options exceed head_max_len=%d" % (qid, head_max_len))+ n_opts = len(render_options(q))+ if len(markers) != n_opts:+ # The markers are placed at absolute positions and `build_sequence` then drops the+ # ones past `max_len`, so this is about the question fitting in the sequence --+ # `head_max_len` is how much of it the options were given, and `max_len` is the+ # ceiling that dropped them. Naming only `head_max_len` pointed at the wrong knob+ # in both directions: lowering it shortens the option block and can make the+ # call succeed, while raising it makes the overflow worse.+ raise ValueError(+ "question %r: its %d options and question text need %d tokens, more than "+ "max_len=%d allows once head_max_len=%d is spent on them; lower head_max_len, "+ "raise max_len, or use fewer options"+ % (qid, n_opts, len(seq), max_len, head_max_len))items.append({"ids": seq, "markers": markers, "qtype": QTYPES[q["t"]]})+ return items- b = collate_items([items], self.tok.pad_token_id)-- # Prepare ONNX inputs as numpy arrays- ort_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),- "qtype": b["qtype"].numpy().astype(np.int64),- }-- # Run ONNX inference- ort_outs = self.session.run(["logits", "act_logits"], ort_inputs)- logits = ort_outs[0]- act_logits = ort_outs[1]-- # Compute softmax for actions manually in numpy- act_exp = np.exp(act_logits - np.max(act_logits, axis=-1, keepdims=True))- act = act_exp / np.sum(act_exp, axis=-1, keepdims=True)-+ def _decode_answers(self, logits, act, items: List[Dict[str, Any]], ids: List[str],+ internal: Dict[str, Any], offset: int,+ lang: Optional[str] = None) -> Dict[str, Any]:+ """Decode the `len(ids)` head rows starting at `offset` into this state's answers."""answers = {}- n_tokens = int(b["attention_mask"].sum())-for r, qid in enumerate(ids):- q = self._to_internal(questions[qid])+ q = internal[qid]k = len(items[r]["markers"])qt = QTYPES[q["t"]]t_scale = self.temperature_by_options.get(temp_bucket(qt, k), self.temperature[qt])