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
perf(agent): reuse sequence stats option count instead of re-rendering options
Hamidreza0e07dc462005
diff --git a/laya/agent.py b/laya/agent.pyindex 67ea7760..a89fded3 100644--- a/laya/agent.py+++ b/laya/agent.py@@ -40,7 +40,6 @@ from .common import (_disable_question_token_reuse,encode_text,render_criterion,- render_options,resolve_lang_temperatures,serialize_state,temp_bucket,@@ -1154,7 +1153,7 @@ class Agent(HookRegistry):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, return_layout=self.parallel_options)- n_opts = len(render_options(q))+ n_opts = stats["options"]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 --diff --git a/laya/onnx_agent.py b/laya/onnx_agent.pyindex 010190a3..a38f8edf 100644--- a/laya/onnx_agent.py+++ b/laya/onnx_agent.py@@ -22,7 +22,6 @@ from laya.common import (confidence_from_probs,encode_text,render_criterion,- render_options,resolve_lang_temperatures,serialize_state,window_batch_cap,@@ -642,7 +641,7 @@ class ONNXAgent(HookRegistry):truncate_left=truncate_left, state_ids=state_ids, return_stats=True,return_truncation_stats=True,)- n_opts = len(render_options(q))+ n_opts = stats["options"]if len(markers) != n_opts:# 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