| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import pytest |
| 4 | |
| 5 | from tests.reasoning.utils import run_reasoning_extraction |
| 6 | from vllm.reasoning import ReasoningParser, ReasoningParserManager |
| 7 | from vllm.reasoning.hy_v3_reasoning_parser import HYV3ReasoningParser |
| 8 | from vllm.tokenizers import get_tokenizer |
| 9 | |
| 10 | parser_name = "hy_v3" |
| 11 | MODEL = "tencent/Hy3-preview" |
| 12 | |
| 13 | |
| 14 | @pytest.fixture(scope="module") |
| 15 | def hy_v3_tokenizer(): |
| 16 | return get_tokenizer(tokenizer_name=MODEL) |
| 17 | |
| 18 | |
| 19 | WITH_THINK = { |
| 20 | "output": "This is a reasoning section</think>This is the rest", |
| 21 | "reasoning": "This is a reasoning section", |
| 22 | "content": "This is the rest", |
| 23 | "is_reasoning_end": True, |
| 24 | "reasoning_effort": "high", |
| 25 | } |
| 26 | |
| 27 | WITH_THINK_STREAM = { |
| 28 | "output": "This is a reasoning section</think>This is the rest", |
| 29 | "reasoning": "This is a reasoning section", |
| 30 | "content": "This is the rest", |
| 31 | "is_reasoning_end": True, |
| 32 | "reasoning_effort": "high", |
| 33 | } |
| 34 | |
| 35 | WITHOUT_THINK = { |
| 36 | "output": "This is the rest", |
| 37 | "reasoning": None, |
| 38 | "content": "This is the rest", |
| 39 | "is_reasoning_end": True, |
| 40 | "reasoning_effort": "no_think", |
| 41 | } |
| 42 | |
| 43 | WITHOUT_THINK_STREAM = { |
| 44 | "output": "This is the rest", |
| 45 | "reasoning": None, |
| 46 | "content": "This is the rest", |
| 47 | "is_reasoning_end": True, |
| 48 | "reasoning_effort": "no_think", |
| 49 | } |
| 50 | |
| 51 | WITH_REASONING_EFFORT_NONE = { |
| 52 | "output": "This is the rest", |
| 53 | "reasoning": None, |
| 54 | "content": "This is the rest", |
| 55 | "is_reasoning_end": True, |
| 56 | } |
| 57 | |
| 58 | WITH_REASONING_EFFORT_NONE_STREAM = { |
| 59 | "output": "This is the rest", |
| 60 | "reasoning": None, |
| 61 | "content": "This is the rest", |
| 62 | "is_reasoning_end": True, |
| 63 | } |
| 64 | |
| 65 | COMPLETE_REASONING = { |
| 66 | "output": "This is a reasoning section</think>", |
| 67 | "reasoning": "This is a reasoning section", |
| 68 | "content": None, |
| 69 | "is_reasoning_end": True, |
| 70 | "reasoning_effort": "high", |
| 71 | } |
| 72 | MULTILINE_REASONING = { |
| 73 | "output": "This is a reasoning\nsection</think>This is the rest\nThat", |
| 74 | "reasoning": "This is a reasoning\nsection", |
| 75 | "content": "This is the rest\nThat", |
| 76 | "is_reasoning_end": True, |
| 77 | "reasoning_effort": "high", |
| 78 | } |
| 79 | ONLY_OPEN_TAG = { |
| 80 | "output": "This is a reasoning section", |
| 81 | "reasoning": "This is a reasoning section", |
| 82 | "content": None, |
| 83 | "is_reasoning_end": False, |
| 84 | "reasoning_effort": "high", |
| 85 | } |
| 86 | |
| 87 | ONLY_OPEN_TAG_STREAM = { |
| 88 | "output": "This is a reasoning section", |
| 89 | "reasoning": "This is a reasoning section", |
| 90 | "content": None, |
| 91 | "is_reasoning_end": False, |
| 92 | "reasoning_effort": "high", |
| 93 | } |
| 94 | |
| 95 | TEST_CASES = [ |
| 96 | pytest.param( |
| 97 | False, |
| 98 | WITH_THINK, |
| 99 | id="with_think", |
| 100 | ), |
| 101 | pytest.param( |
| 102 | True, |
| 103 | WITH_THINK_STREAM, |
| 104 | id="with_think_stream", |
| 105 | ), |
| 106 | pytest.param( |
| 107 | False, |
| 108 | WITHOUT_THINK, |
| 109 | id="without_think", |
| 110 | ), |
| 111 | pytest.param( |
| 112 | True, |
| 113 | WITHOUT_THINK_STREAM, |
| 114 | id="without_think_stream", |
| 115 | ), |
| 116 | pytest.param( |
| 117 | False, |
| 118 | WITH_REASONING_EFFORT_NONE, |
| 119 | id="with_reasoning_effort_none", |
| 120 | ), |
| 121 | pytest.param( |
| 122 | True, |
| 123 | WITH_REASONING_EFFORT_NONE_STREAM, |
| 124 | id="with_reasoning_effort_none_stream", |
| 125 | ), |
| 126 | pytest.param( |
| 127 | False, |
| 128 | COMPLETE_REASONING, |
| 129 | id="complete_reasoning", |
| 130 | ), |
| 131 | pytest.param( |
| 132 | True, |
| 133 | COMPLETE_REASONING, |
| 134 | id="complete_reasoning_stream", |
| 135 | ), |
| 136 | pytest.param( |
| 137 | False, |
| 138 | MULTILINE_REASONING, |
| 139 | id="multiline_reasoning", |
| 140 | ), |
| 141 | pytest.param( |
| 142 | True, |
| 143 | MULTILINE_REASONING, |
| 144 | id="multiline_reasoning_stream", |
| 145 | ), |
| 146 | pytest.param( |
| 147 | False, |
| 148 | ONLY_OPEN_TAG, |
| 149 | id="only_open_tag", |
| 150 | ), |
| 151 | pytest.param( |
| 152 | True, |
| 153 | ONLY_OPEN_TAG_STREAM, |
| 154 | id="only_open_tag_stream", |
| 155 | ), |
| 156 | ] |
| 157 | |
| 158 | STILL_REASONING_PROMPT = """<｜hy_begin▁of▁sentence｜> |
| 159 | You are a helpful assistant. |
| 160 | <｜reasoning_mode｜>reasoning_effort:high<｜hy_User｜> |
| 161 | What is the capital of France?<｜hy_Assistant｜> |
| 162 | <think>The user is asking for the capital of""" |
| 163 | |
| 164 | DONE_REASONING_PROMPT = """<｜hy_begin▁of▁sentence｜> |
| 165 | You are a helpful assistant. |
| 166 | <｜reasoning_mode｜>reasoning_effort:high<｜hy_User｜> |
| 167 | What is the capital of France?<｜hy_Assistant｜> |
| 168 | <think>The user is asking for the capital of France.</think> |
| 169 | The capital of France is Paris.""" |
| 170 | |
| 171 | MULTI_TURN_STILL_REASONING_PROMPT = """<｜hy_begin▁of▁sentence｜> |
| 172 | You are a helpful assistant. |
| 173 | <｜reasoning_mode｜>reasoning_effort:high<｜hy_User｜> |
| 174 | What is the capital of France?<｜hy_Assistant｜ |
| 175 | ><think></think>The capital of France is Paris.<eos:6124c78e> |
| 176 | <｜hy_User｜>What about Chile?<｜hy_Assistant｜> |
| 177 | <think>The user is asking for the capital of""" |
| 178 | |
| 179 | MULTI_TURN_DONE_REASONING_PROMPT = """<｜hy_begin▁of▁sentence｜> |
| 180 | You are a helpful assistant. |
| 181 | <｜reasoning_mode｜>reasoning_effort:high<｜hy_User｜> |
| 182 | What is the capital of France?<｜hy_Assistant｜ |
| 183 | ><think></think>The capital of France is Paris.<eos:6124c78e> |
| 184 | <｜hy_User｜>What about Chile?<｜hy_Assistant｜> |
| 185 | <think>The user is asking for the capital of Chile.</think> |
| 186 | The capital of Chile is Santiago.""" |
| 187 | |
| 188 | REASONING_END_TEST_CASES = [ |
| 189 | pytest.param(STILL_REASONING_PROMPT, False, id="still_reasoning"), |
| 190 | pytest.param(DONE_REASONING_PROMPT, True, id="done_reasoning"), |
| 191 | pytest.param( |
| 192 | MULTI_TURN_STILL_REASONING_PROMPT, False, id="multi_turn_still_reasoning" |
| 193 | ), |
| 194 | pytest.param( |
| 195 | MULTI_TURN_DONE_REASONING_PROMPT, True, id="multi_turn_done_reasoning" |
| 196 | ), |
| 197 | ] |
| 198 | |
| 199 | |
| 200 | @pytest.mark.parametrize("streaming, param_dict", TEST_CASES) |
| 201 | def test_reasoning( |
| 202 | streaming: bool, |
| 203 | param_dict: dict, |
| 204 | hy_v3_tokenizer, |
| 205 | ): |
| 206 | output = hy_v3_tokenizer.tokenize(param_dict["output"]) |
| 207 | output_tokens: list[str] = [ |
| 208 | hy_v3_tokenizer.convert_tokens_to_string([token]) for token in output |
| 209 | ] |
| 210 | |
| 211 | parser_kwargs = {} |
| 212 | if "reasoning_effort" in param_dict: |
| 213 | parser_kwargs["chat_template_kwargs"] = { |
| 214 | "reasoning_effort": param_dict["reasoning_effort"] |
| 215 | } |
| 216 | parser: ReasoningParser = ReasoningParserManager.get_reasoning_parser(parser_name)( |
| 217 | hy_v3_tokenizer, |
| 218 | **parser_kwargs, |
| 219 | ) |
| 220 | |
| 221 | reasoning, content = run_reasoning_extraction( |
| 222 | parser, output_tokens, streaming=streaming |
| 223 | ) |
| 224 | |
| 225 | assert reasoning == param_dict["reasoning"] |
| 226 | assert content == param_dict["content"] |
| 227 | |
| 228 | output_ids = hy_v3_tokenizer.convert_tokens_to_ids(output) |
| 229 | is_reasoning_end = parser.is_reasoning_end(output_ids) |
| 230 | assert is_reasoning_end == param_dict["is_reasoning_end"] |
| 231 | |
| 232 | |
| 233 | @pytest.mark.parametrize("prompt, is_reasoning_end", REASONING_END_TEST_CASES) |
| 234 | def test_is_reasoning_end_full_prompt( |
| 235 | prompt: str, is_reasoning_end: bool, hy_v3_tokenizer |
| 236 | ): |
| 237 | parser: ReasoningParser = ReasoningParserManager.get_reasoning_parser(parser_name)( |
| 238 | hy_v3_tokenizer, |
| 239 | chat_template_kwargs={"reasoning_effort": "high"}, |
| 240 | ) |
| 241 | tokens = hy_v3_tokenizer.tokenize(prompt) |
| 242 | token_ids = hy_v3_tokenizer.convert_tokens_to_ids(tokens) |
| 243 | check_is_reasoning_end = parser.is_reasoning_end(token_ids) |
| 244 | assert check_is_reasoning_end == is_reasoning_end |
| 245 | |
| 246 | |
| 247 | def test_constructor_does_not_mutate_shared_chat_template_kwargs(hy_v3_tokenizer): |
| 248 | parser_cls = ReasoningParserManager.get_reasoning_parser(parser_name) |
| 249 | chat_template_kwargs = {"reasoning_effort": "low"} |
| 250 | |
| 251 | first_parser: ReasoningParser = parser_cls( |
| 252 | hy_v3_tokenizer, |
| 253 | chat_template_kwargs=chat_template_kwargs, |
| 254 | ) |
| 255 | second_parser: ReasoningParser = parser_cls( |
| 256 | hy_v3_tokenizer, |
| 257 | chat_template_kwargs=chat_template_kwargs, |
| 258 | ) |
| 259 | |
| 260 | assert chat_template_kwargs == {"reasoning_effort": "low"} |
| 261 | assert isinstance(first_parser, HYV3ReasoningParser) |
| 262 | assert isinstance(second_parser, HYV3ReasoningParser) |
| 263 | assert first_parser._identity_parser is None |
| 264 | assert second_parser._identity_parser is None |
| 265 | |
| 266 | |
| 267 | def test_constructor_falls_back_to_outer_reasoning_effort(hy_v3_tokenizer): |
| 268 | parser: ReasoningParser = ReasoningParserManager.get_reasoning_parser(parser_name)( |
| 269 | hy_v3_tokenizer, |
| 270 | reasoning_effort="low", |
| 271 | ) |
| 272 | |
| 273 | assert isinstance(parser, HYV3ReasoningParser) |
| 274 | assert parser._identity_parser is None |