# SPDX-License-Identifier: Apache-2.0 # Copyright 2026 Nara contributors """Portable Nara SAR final decoding: original beam rules and bilingual ARPA LM. No Torch, NumPy or KenLM runtime is required. The compact memory-mapped LM is built from the production ARPA and checked against KenLM before packaging. """ import base64 from collections import defaultdict from functools import lru_cache import json import math import mmap from pathlib import Path import struct from sar_syllables import SyllableConstraint class LanguageModel: def __init__(self, path): self.file = open(path, 'rb') self.data = mmap.mmap(self.file.fileno(), 0, access=mmap.ACCESS_READ) if self.data[:8] != b'NARALM01': raise ValueError('Invalid SAR language model') length = struct.unpack_from(''] self.bos = (self.vocab[''],) self.lookup = lru_cache(maxsize=65536)(self._lookup) self.advance = lru_cache(maxsize=32768)(self._advance) def close(self): self.lookup.cache_clear() self.advance.cache_clear() self.data.close() self.file.close() def _lookup(self, ids): if not ids: return None order = len(ids) block = self.header['blocks'][order - 1] width = self.header['keyWidth'] key = b''.join(index.to_bytes(width, 'big') for index in ids) record_size = len(key) + 8 first = self.start + block['offset'] low, high = 0, block['count'] while low < high: middle = (low + high) // 2 at = first + middle * record_size candidate = self.data[at:at + len(key)] if candidate < key: low = middle + 1 elif candidate > key: high = middle else: return struct.unpack_from('