"""Public API and README examples, exercised against a real local checkpoint.""" import os import unicodedata import pytest import torch from transformers import AutoModelForCausalLM, AutoTokenizer from manta import MantaKeyboard @pytest.fixture(scope='module') def keyboard(): torch.set_num_threads(2) return MantaKeyboard.from_pretrained(os.environ['MANTA_MODEL_DIR']) def test_standard_transformers_loading(keyboard): folder = os.environ['MANTA_MODEL_DIR'] tokenizer = AutoTokenizer.from_pretrained(folder, trust_remote_code=False) model = AutoModelForCausalLM.from_pretrained(folder, trust_remote_code=False) assert sum(p.numel() for p in model.parameters()) == 20_084_064 for text in ['я щас домой ', 'док', 'ёлка е\u0308лка', 'docker перезапущу', ' \n']: assert tokenizer.encode(text, add_special_tokens=False) == keyboard.tokenizer.encode( text, add_special_tokens=False).ids assert not model.training def test_hub_loader_uses_pinned_release_and_only_model_files(monkeypatch): calls = [] def download(model_id, **kwargs): calls.append((model_id, kwargs)) return os.environ['MANTA_MODEL_DIR'] monkeypatch.setattr('manta.snapshot_download', download) model = MantaKeyboard.from_pretrained('average23/MANTA-20M', local_files_only=True) assert model.model.config.model_type == 'qwen2' assert calls == [('average23/MANTA-20M', { 'revision': 'v0.1.0', 'local_files_only': True, 'allow_patterns': ['config.json', 'model.safetensors', 'tokenizer.json', 'tokenizer_config.json', 'generation_config.json'], })] @pytest.mark.parametrize(('context', 'prefix'), [ ('я щас домой ', ''), ('надо перезапустить ', 'док'), ('', ''), ('ща docker ', 'перезапу'), ('ну это ', 'пиз'), (' \n', 'ё'), ('я ', 'е\u0308'), ]) def test_words_prefixes_and_repeatability(keyboard, context, prefix): keyboard.reset_session() first = keyboard.complete(context, prefix=prefix) if prefix else keyboard.next_word(context) assert first assert 0 < len(first) <= 3 assert len({r['text'] for r in first}) == len(first) for row in first: assert row['text'].startswith(prefix) assert all(c == '_' or unicodedata.category(c)[0] in 'LMN' for c in row['text']) assert not keyboard.last_search['timed_out'] keyboard.reset_session() second = keyboard.complete(context, prefix=prefix) if prefix else keyboard.next_word(context) assert first == second def test_whitespace_and_invalid_input(keyboard): assert keyboard.next_word('я щас домой') == keyboard.next_word('я щас домой ') with pytest.raises(ValueError, match='whitespace'): keyboard.complete('я домой', prefix='до') with pytest.raises(ValueError, match='single Unicode word'): keyboard.complete('я ', prefix='два слова') with pytest.raises(ValueError, match='beam budget'): keyboard.next_word('я ', top_k=11)