| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import json |
| 4 | from pathlib import Path |
| 5 | from types import SimpleNamespace |
| 6 | from unittest.mock import patch |
| 7 | |
| 8 | import pytest |
| 9 | from transformers import AutoConfig |
| 10 | from transformers.models.auto.configuration_auto import CONFIG_MAPPING |
| 11 | |
| 12 | from vllm.tokenizers import TokenizerLike |
| 13 | from vllm.tokenizers.registry import ( |
| 14 | TokenizerRegistry, |
| 15 | cached_get_tokenizer, |
| 16 | cached_resolve_tokenizer_args, |
| 17 | cached_tokenizer_from_config, |
| 18 | get_tokenizer, |
| 19 | resolve_tokenizer_args, |
| 20 | ) |
| 21 | from vllm.transformers_utils.configs.qwen3_5_moe import Qwen3_5MoeConfig |
| 22 | |
| 23 | |
| 24 | class TestTokenizer(TokenizerLike): |
| 25 | @classmethod |
| 26 | def from_pretrained( |
| 27 | cls, |
| 28 | path_or_repo_id: str | Path, |
| 29 | *args, |
| 30 | trust_remote_code: bool = False, |
| 31 | revision: str | None = None, |
| 32 | download_dir: str | None = None, |
| 33 | **kwargs, |
| 34 | ) -> "TestTokenizer": |
| 35 | return TestTokenizer(path_or_repo_id) # type: ignore |
| 36 | |
| 37 | def __init__(self, path_or_repo_id: str | Path) -> None: |
| 38 | super().__init__() |
| 39 | |
| 40 | self.path_or_repo_id = path_or_repo_id |
| 41 | |
| 42 | @property |
| 43 | def bos_token_id(self) -> int: |
| 44 | return 0 |
| 45 | |
| 46 | @property |
| 47 | def eos_token_id(self) -> int: |
| 48 | return 1 |
| 49 | |
| 50 | @property |
| 51 | def pad_token_id(self) -> int: |
| 52 | return 2 |
| 53 | |
| 54 | @property |
| 55 | def is_fast(self) -> bool: |
| 56 | return True |
| 57 | |
| 58 | |
| 59 | @pytest.mark.parametrize("runner_type", ["generate", "pooling"]) |
| 60 | def test_resolve_tokenizer_args_idempotent(runner_type): |
| 61 | tokenizer_mode, tokenizer_name, args, kwargs = resolve_tokenizer_args( |
| 62 | "facebook/opt-125m", |
| 63 | runner_type=runner_type, |
| 64 | ) |
| 65 | |
| 66 | assert (tokenizer_mode, tokenizer_name, args, kwargs) == resolve_tokenizer_args( |
| 67 | tokenizer_name, *args, **kwargs |
| 68 | ) |
| 69 | |
| 70 | |
| 71 | @pytest.mark.parametrize("input_kwargs", [{}, {"mistral_format": False}]) |
| 72 | def test_resolve_tokenizer_args_forces_hf_mistral_format_false(input_kwargs): |
| 73 | resolved_mode, _, _, kwargs = resolve_tokenizer_args( |
| 74 | "mistralai/Mistral-Nemo-Instruct-2407", |
| 75 | tokenizer_mode="hf", |
| 76 | **input_kwargs, |
| 77 | ) |
| 78 | |
| 79 | assert resolved_mode == "hf" |
| 80 | assert kwargs["mistral_format"] is False |
| 81 | |
| 82 | |
| 83 | def test_resolve_tokenizer_args_rejects_hf_mistral_format_true(): |
| 84 | with pytest.raises( |
| 85 | ValueError, |
| 86 | match="mistral_format=True is not supported with tokenizer_mode='hf'", |
| 87 | ): |
| 88 | resolve_tokenizer_args( |
| 89 | "mistralai/Mistral-Nemo-Instruct-2407", |
| 90 | tokenizer_mode="hf", |
| 91 | mistral_format=True, |
| 92 | ) |
| 93 | |
| 94 | |
| 95 | def test_customized_tokenizer(): |
| 96 | TokenizerRegistry.register("test_tokenizer", __name__, TestTokenizer.__name__) |
| 97 | |
| 98 | tokenizer = TokenizerRegistry.load_tokenizer("test_tokenizer", "abc") |
| 99 | assert isinstance(tokenizer, TestTokenizer) |
| 100 | assert tokenizer.path_or_repo_id == "abc" |
| 101 | assert tokenizer.bos_token_id == 0 |
| 102 | assert tokenizer.eos_token_id == 1 |
| 103 | assert tokenizer.pad_token_id == 2 |
| 104 | |
| 105 | tokenizer = get_tokenizer("abc", tokenizer_mode="test_tokenizer") |
| 106 | assert isinstance(tokenizer, TestTokenizer) |
| 107 | assert tokenizer.path_or_repo_id == "abc" |
| 108 | assert tokenizer.bos_token_id == 0 |
| 109 | assert tokenizer.eos_token_id == 1 |
| 110 | assert tokenizer.pad_token_id == 2 |
| 111 | |
| 112 | |
| 113 | def test_cached_tokenizer_from_config_registers_local_config(tmp_path: Path): |
| 114 | (tmp_path / "config.json").write_text( |
| 115 | json.dumps({"model_type": "qwen3_5_moe"}), |
| 116 | encoding="utf-8", |
| 117 | ) |
| 118 | |
| 119 | model_config = SimpleNamespace( |
| 120 | skip_tokenizer_init=False, |
| 121 | tokenizer=str(tmp_path), |
| 122 | runner_type="generate", |
| 123 | tokenizer_mode="hf", |
| 124 | tokenizer_revision=None, |
| 125 | trust_remote_code=True, |
| 126 | hf_config=Qwen3_5MoeConfig(), |
| 127 | ) |
| 128 | |
| 129 | registered_config = CONFIG_MAPPING._extra_content.pop("qwen3_5_moe", None) |
| 130 | cached_get_tokenizer.cache_clear() |
| 131 | cached_resolve_tokenizer_args.cache_clear() |
| 132 | |
| 133 | try: |
| 134 | |
| 135 | def fake_from_pretrained(path_or_repo_id: str, *args, **kwargs): |
| 136 | passed_config = kwargs.pop("config") |
| 137 | assert isinstance(passed_config, Qwen3_5MoeConfig) |
| 138 | loaded_config = AutoConfig.from_pretrained( |
| 139 | path_or_repo_id, |
| 140 | trust_remote_code=False, |
| 141 | ) |
| 142 | assert isinstance(loaded_config, Qwen3_5MoeConfig) |
| 143 | return SimpleNamespace(is_fast=True) |
| 144 | |
| 145 | with ( |
| 146 | patch( |
| 147 | "vllm.tokenizers.registry.logger.debug_once", |
| 148 | lambda *args, **kwargs: None, |
| 149 | ), |
| 150 | patch( |
| 151 | "vllm.tokenizers.hf.AutoTokenizer.from_pretrained", |
| 152 | side_effect=fake_from_pretrained, |
| 153 | ), |
| 154 | patch( |
| 155 | "vllm.tokenizers.hf.get_cached_tokenizer", |
| 156 | side_effect=lambda tokenizer: tokenizer, |
| 157 | ), |
| 158 | ): |
| 159 | tokenizer = cached_tokenizer_from_config(model_config) |
| 160 | |
| 161 | assert tokenizer.is_fast is True |
| 162 | finally: |
| 163 | cached_get_tokenizer.cache_clear() |
| 164 | cached_resolve_tokenizer_args.cache_clear() |
| 165 | CONFIG_MAPPING._extra_content.pop("qwen3_5_moe", None) |
| 166 | if registered_config is not None: |
| 167 | CONFIG_MAPPING._extra_content["qwen3_5_moe"] = registered_config |