| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import copy |
| 5 | |
| 6 | import pytest |
| 7 | from transformers import AutoTokenizer, PreTrainedTokenizerBase |
| 8 | |
| 9 | from tests.reasoning.utils import run_reasoning_extraction |
| 10 | from vllm.parser.engine.adapters import ParserEngineReasoningAdapter |
| 11 | from vllm.parser.glm47_moe import Glm47MoeParser |
| 12 | from vllm.reasoning import ReasoningParser, ReasoningParserManager |
| 13 | |
| 14 | parser_name = "glm45" |
| 15 | start_token = "<think>" |
| 16 | end_token = "</think>" |
| 17 | |
| 18 | REASONING_MODEL_NAME = "zai-org/GLM-4.7" |
| 19 | |
| 20 | |
| 21 | @pytest.fixture(scope="module") |
| 22 | def glm45_tokenizer() -> PreTrainedTokenizerBase: |
| 23 | return AutoTokenizer.from_pretrained(REASONING_MODEL_NAME) |
| 24 | |
| 25 | |
| 26 | WITH_THINK = { |
| 27 | "output": "<think>This is a reasoning section</think>This is the rest", |
| 28 | "reasoning": "This is a reasoning section", |
| 29 | "content": "This is the rest", |
| 30 | "is_reasoning_end": True, |
| 31 | } |
| 32 | |
| 33 | WITH_THINK_STREAM = { |
| 34 | "output": "<think>This is a reasoning section</think>This is the rest", |
| 35 | "reasoning": "This is a reasoning section", |
| 36 | "content": "This is the rest", |
| 37 | "is_reasoning_end": True, |
| 38 | } |
| 39 | |
| 40 | WITHOUT_THINK = { |
| 41 | "output": "This is the rest", |
| 42 | "reasoning": "This is the rest", |
| 43 | "content": None, |
| 44 | "is_reasoning_end": False, |
| 45 | } |
| 46 | |
| 47 | WITHOUT_THINK_STREAM = { |
| 48 | "output": "This is the rest", |
| 49 | "reasoning": "This is the rest", |
| 50 | "content": None, |
| 51 | "is_reasoning_end": False, |
| 52 | } |
| 53 | |
| 54 | WITHOUT_OPEN_THINK = { |
| 55 | "output": "This is a reasoning section</think>This is the rest", |
| 56 | "reasoning": "This is a reasoning section", |
| 57 | "content": "This is the rest", |
| 58 | "is_reasoning_end": True, |
| 59 | } |
| 60 | |
| 61 | WITHOUT_OPEN_THINK_STREAM = { |
| 62 | "output": "This is a reasoning section</think>This is the rest", |
| 63 | "reasoning": "This is a reasoning section", |
| 64 | "content": "This is the rest", |
| 65 | "is_reasoning_end": True, |
| 66 | } |
| 67 | |
| 68 | COMPLETE_REASONING = { |
| 69 | "output": "<think>This is a reasoning section</think>", |
| 70 | "reasoning": "This is a reasoning section", |
| 71 | "content": None, |
| 72 | "is_reasoning_end": True, |
| 73 | } |
| 74 | MULTILINE_REASONING = { |
| 75 | "output": "<think>This is a reasoning\nsection</think>This is the rest\nThat", |
| 76 | "reasoning": "This is a reasoning\nsection", |
| 77 | "content": "This is the rest\nThat", |
| 78 | "is_reasoning_end": True, |
| 79 | } |
| 80 | ONLY_OPEN_TAG = { |
| 81 | "output": "<think>This is a reasoning section", |
| 82 | "reasoning": "This is a reasoning section", |
| 83 | "content": None, |
| 84 | "is_reasoning_end": False, |
| 85 | } |
| 86 | |
| 87 | ONLY_OPEN_TAG_STREAM = { |
| 88 | "output": "<think>This is a reasoning section", |
| 89 | "reasoning": "This is a reasoning section", |
| 90 | "content": None, |
| 91 | "is_reasoning_end": False, |
| 92 | } |
| 93 | |
| 94 | TEST_CASES = [ |
| 95 | pytest.param( |
| 96 | False, |
| 97 | WITH_THINK, |
| 98 | id="with_think", |
| 99 | ), |
| 100 | pytest.param( |
| 101 | True, |
| 102 | WITH_THINK_STREAM, |
| 103 | id="with_think_stream", |
| 104 | ), |
| 105 | pytest.param( |
| 106 | False, |
| 107 | WITHOUT_THINK, |
| 108 | id="without_think", |
| 109 | ), |
| 110 | pytest.param( |
| 111 | True, |
| 112 | WITHOUT_THINK_STREAM, |
| 113 | id="without_think_stream", |
| 114 | ), |
| 115 | pytest.param( |
| 116 | False, |
| 117 | WITHOUT_OPEN_THINK, |
| 118 | id="without_open_think", |
| 119 | ), |
| 120 | pytest.param( |
| 121 | True, |
| 122 | WITHOUT_OPEN_THINK_STREAM, |
| 123 | id="without_open_think_stream", |
| 124 | ), |
| 125 | pytest.param( |
| 126 | False, |
| 127 | COMPLETE_REASONING, |
| 128 | id="complete_reasoning", |
| 129 | ), |
| 130 | pytest.param( |
| 131 | True, |
| 132 | COMPLETE_REASONING, |
| 133 | id="complete_reasoning_stream", |
| 134 | ), |
| 135 | pytest.param( |
| 136 | False, |
| 137 | MULTILINE_REASONING, |
| 138 | id="multiline_reasoning", |
| 139 | ), |
| 140 | pytest.param( |
| 141 | True, |
| 142 | MULTILINE_REASONING, |
| 143 | id="multiline_reasoning_stream", |
| 144 | ), |
| 145 | pytest.param( |
| 146 | False, |
| 147 | ONLY_OPEN_TAG, |
| 148 | id="only_open_tag", |
| 149 | ), |
| 150 | pytest.param( |
| 151 | True, |
| 152 | ONLY_OPEN_TAG_STREAM, |
| 153 | id="only_open_tag_stream", |
| 154 | ), |
| 155 | ] |
| 156 | |
| 157 | STILL_REASONING_PROMPT = """[gMASK]<sop><|system|> |
| 158 | You are a helpful assistant.<|user|> |
| 159 | What is the capital of France?<|assistant|> |
| 160 | <think>The user is asking for the capital of""" |
| 161 | |
| 162 | DONE_REASONING_PROMPT = """[gMASK]<sop><|system|> |
| 163 | You are a helpful assistant.<|user|> |
| 164 | What is the capital of France?<|assistant|> |
| 165 | <think>The user is asking for the capital of France.</think> |
| 166 | The capital of France is Paris.""" |
| 167 | |
| 168 | MULTI_TURN_STILL_REASONING_PROMPT = """[gMASK]<sop><|system|> |
| 169 | You are a helpful assistant.<|user|> |
| 170 | What is the capital of France?<|assistant|> |
| 171 | <think></think> |
| 172 | The capital of France is Paris.<|user|> |
| 173 | What about Chile?<|assistant|> |
| 174 | <think>The user is asking for the capital of""" |
| 175 | |
| 176 | MULTI_TURN_DONE_REASONING_PROMPT = """[gMASK]<sop><|system|> |
| 177 | You are a helpful assistant.<|user|> |
| 178 | What is the capital of France?<|assistant|> |
| 179 | <think></think> |
| 180 | The capital of France is Paris.<|user|> |
| 181 | What about Chile?<|assistant|> |
| 182 | <think>The user is asking for the capital of Chile.</think> |
| 183 | The capital of Chile is Santiago.""" |
| 184 | |
| 185 | REASONING_END_TEST_CASES = [ |
| 186 | pytest.param(STILL_REASONING_PROMPT, False, id="still_reasoning"), |
| 187 | pytest.param(DONE_REASONING_PROMPT, True, id="done_reasoning"), |
| 188 | pytest.param( |
| 189 | MULTI_TURN_STILL_REASONING_PROMPT, False, id="multi_turn_still_reasoning" |
| 190 | ), |
| 191 | pytest.param( |
| 192 | MULTI_TURN_DONE_REASONING_PROMPT, True, id="multi_turn_done_reasoning" |
| 193 | ), |
| 194 | ] |
| 195 | |
| 196 | |
| 197 | @pytest.mark.parametrize("streaming, param_dict", TEST_CASES) |
| 198 | def test_reasoning( |
| 199 | streaming: bool, |
| 200 | param_dict: dict, |
| 201 | glm45_tokenizer, |
| 202 | ): |
| 203 | output = glm45_tokenizer.tokenize(param_dict["output"]) |
| 204 | output_tokens: list[str] = [ |
| 205 | glm45_tokenizer.convert_tokens_to_string([token]) for token in output |
| 206 | ] |
| 207 | parser: ReasoningParser = ReasoningParserManager.get_reasoning_parser(parser_name)( |
| 208 | glm45_tokenizer |
| 209 | ) |
| 210 | |
| 211 | reasoning, content = run_reasoning_extraction( |
| 212 | parser, output_tokens, streaming=streaming |
| 213 | ) |
| 214 | |
| 215 | assert reasoning == param_dict["reasoning"] |
| 216 | assert content == param_dict["content"] |
| 217 | |
| 218 | output_ids = glm45_tokenizer.convert_tokens_to_ids(output) |
| 219 | is_reasoning_end = parser.is_reasoning_end(output_ids) |
| 220 | assert is_reasoning_end == param_dict["is_reasoning_end"] |
| 221 | |
| 222 | |
| 223 | @pytest.mark.parametrize("prompt, is_reasoning_end", REASONING_END_TEST_CASES) |
| 224 | def test_is_reasoning_end_full_prompt( |
| 225 | prompt: str, is_reasoning_end: bool, glm45_tokenizer |
| 226 | ): |
| 227 | parser: ReasoningParser = ReasoningParserManager.get_reasoning_parser(parser_name)( |
| 228 | glm45_tokenizer |
| 229 | ) |
| 230 | tokens = glm45_tokenizer.tokenize(prompt) |
| 231 | token_ids = glm45_tokenizer.convert_tokens_to_ids(tokens) |
| 232 | check_is_reasoning_end = parser.is_reasoning_end(token_ids) |
| 233 | assert check_is_reasoning_end == is_reasoning_end |
| 234 | |
| 235 | |
| 236 | GLM53_TEMPLATE = ( |
| 237 | "[gMASK]<sop>\n" |
| 238 | "{%- set effective_reasoning_effort = reasoning_effort if reasoning_effort is" |
| 239 | " defined and reasoning_effort in ['low', 'high'] else 'max' -%}\n" |
| 240 | "<|system|>Reasoning Effort: {{ effective_reasoning_effort | capitalize }}\n" |
| 241 | "{% for tc in m.tool_calls %}\n" |
| 242 | "{{- '<tool_call>' + tc.name -}}\n" |
| 243 | "{% set _args = tc.arguments %}" |
| 244 | "{% for k, v in _args.items() %}" |
| 245 | "<arg_key>{{ k }}</arg_key><arg_value>{{ v }}</arg_value>" |
| 246 | "{% endfor %}</tool_call>\n" |
| 247 | "{% endfor %}\n" |
| 248 | "<|assistant|>{{- '<think>' -}}" |
| 249 | ) |
| 250 | |
| 251 | GLM53_LEAK = { |
| 252 | "output": "Simple question.</think>2 + 2 = **4**", |
| 253 | "reasoning": "Simple question.", |
| 254 | "content": "2 + 2 = **4**", |
| 255 | "is_reasoning_end": True, |
| 256 | } |
| 257 | |
| 258 | |
| 259 | @pytest.fixture() |
| 260 | def glm53_style_tokenizer(glm45_tokenizer): |
| 261 | tokenizer = copy.copy(glm45_tokenizer) |
| 262 | tokenizer.chat_template = GLM53_TEMPLATE |
| 263 | return tokenizer |
| 264 | |
| 265 | |
| 266 | def _glm_engine(parser: ReasoningParser) -> Glm47MoeParser: |
| 267 | assert isinstance(parser, ParserEngineReasoningAdapter) |
| 268 | engine = parser._parser_engine |
| 269 | assert isinstance(engine, Glm47MoeParser) |
| 270 | return engine |
| 271 | |
| 272 | |
| 273 | @pytest.mark.parametrize( |
| 274 | "disable_kwargs", [{"enable_thinking": False}, {"thinking": False}] |
| 275 | ) |
| 276 | def test_glm53_template_forces_reasoning(disable_kwargs: dict, glm53_style_tokenizer): |
| 277 | parser_cls = ReasoningParserManager.get_reasoning_parser(parser_name) |
| 278 | parser = parser_cls(glm53_style_tokenizer, chat_template_kwargs=disable_kwargs) |
| 279 | assert _glm_engine(parser).thinking_enabled |
| 280 | |
| 281 | output = glm53_style_tokenizer.tokenize(GLM53_LEAK["output"]) |
| 282 | output_tokens: list[str] = [ |
| 283 | glm53_style_tokenizer.convert_tokens_to_string([token]) for token in output |
| 284 | ] |
| 285 | reasoning, content = run_reasoning_extraction(parser, output_tokens) |
| 286 | assert reasoning == GLM53_LEAK["reasoning"] |
| 287 | assert content == GLM53_LEAK["content"] |
| 288 | |
| 289 | output_ids = glm53_style_tokenizer.convert_tokens_to_ids(output) |
| 290 | assert parser.is_reasoning_end(output_ids) == GLM53_LEAK["is_reasoning_end"] |
| 291 | |
| 292 | |
| 293 | def test_glm53_template_forces_reasoning_streaming(glm53_style_tokenizer): |
| 294 | parser = ReasoningParserManager.get_reasoning_parser(parser_name)( |
| 295 | glm53_style_tokenizer, chat_template_kwargs={"enable_thinking": False} |
| 296 | ) |
| 297 | output = glm53_style_tokenizer.tokenize(GLM53_LEAK["output"]) |
| 298 | output_tokens: list[str] = [ |
| 299 | glm53_style_tokenizer.convert_tokens_to_string([token]) for token in output |
| 300 | ] |
| 301 | reasoning, content = run_reasoning_extraction(parser, output_tokens, streaming=True) |
| 302 | assert reasoning == GLM53_LEAK["reasoning"] |
| 303 | assert content == GLM53_LEAK["content"] |
| 304 | |
| 305 | |
| 306 | def test_glm47_template_honors_thinking_disable(glm45_tokenizer): |
| 307 | parser = ReasoningParserManager.get_reasoning_parser(parser_name)( |
| 308 | glm45_tokenizer, chat_template_kwargs={"enable_thinking": False} |
| 309 | ) |
| 310 | assert not _glm_engine(parser).thinking_enabled |
| 311 | |
| 312 | output = glm45_tokenizer.tokenize(GLM53_LEAK["output"]) |
| 313 | output_tokens: list[str] = [ |
| 314 | glm45_tokenizer.convert_tokens_to_string([token]) for token in output |
| 315 | ] |
| 316 | reasoning, content = run_reasoning_extraction(parser, output_tokens) |
| 317 | assert reasoning is None |
| 318 | assert content == GLM53_LEAK["output"] |