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 #668 from somtri/fix/truncation-from-applied-budget
Nandakishorf3f1ea94d04a
diff --cc laya/agent.pyindex 8a801b98,7cfc61fd..c98e6859--- a/laya/agent.py+++ b/laya/agent.py@@@ -881,28 -774,13 +881,29 @@@ class Agent(HookRegistry)items = []for qid in ids:q = internal[qid]- seq, markers, stats = build_sequence(self.tok, state, q, max_len, head_max_len,- option_order=q.get("option_order"),- truncate_left=truncate_left, state_ids=state_ids,- return_stats=True)+ seq, markers, stats, state_stats = build_sequence(self.tok, state, q, max_len, head_max_len,++ option_order=q.get("option_order"),+ truncate_left=truncate_left, state_ids=state_ids,+ return_stats=True, return_truncation_stats=True)- 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.+ #+ # The count reported is the markers that SURVIVED, not `len(seq)`: `build_sequence`+ # truncates to `max_len` first, so `len(seq)` is always exactly `max_len` here and+ # would state the ceiling as though it were the requirement.+ raise ValueError(+ "question %r: only %d of its %d option markers fit in max_len=%d with "+ "head_max_len=%d spent on the question; lower head_max_len, raise max_len, "+ "or use fewer options"+ % (qid, len(markers), n_opts, max_len, head_max_len))- items.append({"ids": seq, "markers": markers, "qtype": QTYPES[q["t"]], "options": stats})+ items.append({"ids": seq, "markers": markers, "qtype": QTYPES[q["t"]], "options": stats,+ "state_stats": state_stats})return itemsdef _amp_enabled_for(self, rows: int) -> bool:@@@ -1334,12 -1172,16 +1349,21 @@@the `N` a start hook rewrote them to), and `0` when a start hook answered the document, orleft no states to score, before any window was read -- on either path, so a cached answernever reads as a window the model read.++ Across several windows the truncation keys are combined like every other `usage` field:+ `truncated`, `state_tokens` and `state_tokens_dropped` are summed (so `truncated` is the+ number of windows that were cut, and the token counts include the overlap), and+ `truncated_questions` is the last window's list. The two can disagree: when only an+ earlier window was cut, `truncated` is above 0 and `truncated_questions` is empty. A+ window is cut when it is larger than the room a question's head leaves, from a `window`+ above the default or a start hook that narrows `max_len` / `head_max_len`. Test+ `usage["truncated"] > 0` here, not `is True`."""+ if state is None:+ raise TypeError("state must not be None; pass a string, dict, or list")+ if not isinstance(questions, dict):+ raise TypeError("questions must be a dict of question id -> definition, got %s"+ % type(questions).__name__)if aggregate != "auto":raise ValueError("predict_long: only aggregate='auto' is supported")hook_kwargs = {"hooks": hooks, "on_predict_start": on_predict_start,diff --cc laya/onnx_agent.pyindex c49bea5b,fa3bf502..560ca6bf--- a/laya/onnx_agent.py+++ b/laya/onnx_agent.py@@@ -567,25 -542,26 +569,27 @@@ class ONNXAgent(HookRegistry)items = []for qid in ids:q = internal[qid]- seq, markers, stats = build_sequence(+ seq, markers, stats, state_stats = build_sequence(self.tok, state, q, max_len, head_max_len,+ option_order=q.get("option_order"),truncate_left=truncate_left, state_ids=state_ids, return_stats=True,+ return_truncation_stats=True,)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.+ # Same diagnosis as `Agent._encode_state`, so both backends report the same thing.+ # The markers are placed at absolute positions and `build_sequence` drops the ones+ # past `max_len`, so `head_max_len` is how much of the sequence 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. The count is the+ # markers that survived, not `len(seq)`, which is always exactly `max_len` here.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))+ "question %r: only %d of its %d option markers fit in max_len=%d with "+ "head_max_len=%d spent on the question; lower head_max_len, raise max_len, "+ "or use fewer options"+ % (qid, len(markers), n_opts, max_len, head_max_len))- items.append({"ids": seq, "markers": markers, "qtype": QTYPES[q["t"]], "options": stats})+ items.append({"ids": seq, "markers": markers, "qtype": QTYPES[q["t"]], "options": stats,+ "state_stats": state_stats})return itemsdef _decode_answers(self, logits, act, items: List[Dict[str, Any]], ids: List[str],diff --cc laya/structured.pyindex db4d2942,b152613f..9e75dcec--- a/laya/structured.py+++ b/laya/structured.py@@@ -59,11 -43,8 +59,11 @@@ class DecisionResultconfidence: Dict[str, float]probabilities: Dict[str, Dict[str, Any]]answers: Dict[str, Any]- usage: Optional[Dict[str, int]] = None+ usage: Optional[Dict[str, Any]] = Nonerouting: Optional[Dict[str, Any]] = None+ # Appended with a default so every existing construction of this dataclass keeps working, and+ # the first six fields keep their positions.+ answer_confidence: Dict[str, Optional[float]] = field(default_factory=dict)@dataclass