| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import pytest |
| 4 | from transformers import AutoTokenizer |
| 5 | |
| 6 | from tests.reasoning.utils import DeltaMessage, run_reasoning_extraction |
| 7 | from vllm.reasoning import ReasoningParser, ReasoningParserManager |
| 8 | |
| 9 | parser_name = "granite" |
| 10 | START_REASONING = "Here is my thought process:" |
| 11 | START_RESPONSE = "Here is my response:" |
| 12 | |
| 13 | SIMPLE_REASONING = { |
| 14 | "output": f"{START_REASONING}This is a reasoning section{START_RESPONSE}This is the rest", # noqa: E501 |
| 15 | "reasoning": "This is a reasoning section", |
| 16 | "content": "This is the rest", |
| 17 | } |
| 18 | COMPLETE_REASONING = { |
| 19 | "output": f"{START_REASONING}This is a reasoning section{START_RESPONSE}", |
| 20 | "reasoning": "This is a reasoning section", |
| 21 | "content": None, |
| 22 | } |
| 23 | NO_REASONING = { |
| 24 | "output": "This is content", |
| 25 | "reasoning": None, |
| 26 | "content": "This is content", |
| 27 | } |
| 28 | MULTIPLE_LINES = { |
| 29 | "output": f"{START_REASONING}This\nThat{START_RESPONSE}This is the rest\nThat", |
| 30 | "reasoning": "This\nThat", |
| 31 | "content": "This is the rest\nThat", |
| 32 | } |
| 33 | REASONING_WITH_THINK = { |
| 34 | "output": f"{START_REASONING}This is a reasoning section{START_RESPONSE}This is the rest", # noqa: E501 |
| 35 | "reasoning": "This is a reasoning section", |
| 36 | "content": "This is the rest", |
| 37 | } |
| 38 | COMPLETE_REASONING_WITH_THINK = { |
| 39 | "output": f"{START_REASONING}This is a reasoning section{START_RESPONSE}", |
| 40 | "reasoning": "This is a reasoning section", |
| 41 | "content": None, |
| 42 | } |
| 43 | MULTIPLE_LINES_WITH_THINK = { |
| 44 | "output": f"{START_REASONING}This\nThat{START_RESPONSE}This is the rest\nThat", |
| 45 | "reasoning": "This\nThat", |
| 46 | "content": "This is the rest\nThat", |
| 47 | } |
| 48 | |
| 49 | TEST_CASES = [ |
| 50 | pytest.param( |
| 51 | False, |
| 52 | SIMPLE_REASONING, |
| 53 | id="simple_reasoning", |
| 54 | ), |
| 55 | pytest.param( |
| 56 | False, |
| 57 | COMPLETE_REASONING, |
| 58 | id="complete_reasoning", |
| 59 | ), |
| 60 | pytest.param( |
| 61 | False, |
| 62 | NO_REASONING, |
| 63 | id="no_reasoning", |
| 64 | ), |
| 65 | pytest.param( |
| 66 | False, |
| 67 | MULTIPLE_LINES, |
| 68 | id="multiple_lines", |
| 69 | ), |
| 70 | pytest.param( |
| 71 | False, |
| 72 | REASONING_WITH_THINK, |
| 73 | id="reasoning_with_think", |
| 74 | ), |
| 75 | pytest.param( |
| 76 | False, |
| 77 | COMPLETE_REASONING_WITH_THINK, |
| 78 | id="complete_reasoning_with_think", |
| 79 | ), |
| 80 | pytest.param( |
| 81 | False, |
| 82 | MULTIPLE_LINES_WITH_THINK, |
| 83 | id="multiple_lines_with_think", |
| 84 | ), |
| 85 | pytest.param( |
| 86 | True, |
| 87 | SIMPLE_REASONING, |
| 88 | id="simple_reasoning_streaming", |
| 89 | ), |
| 90 | pytest.param( |
| 91 | True, |
| 92 | COMPLETE_REASONING, |
| 93 | id="complete_reasoning_streaming", |
| 94 | ), |
| 95 | pytest.param( |
| 96 | True, |
| 97 | NO_REASONING, |
| 98 | id="no_reasoning_streaming", |
| 99 | ), |
| 100 | pytest.param( |
| 101 | True, |
| 102 | MULTIPLE_LINES, |
| 103 | id="multiple_lines_streaming", |
| 104 | ), |
| 105 | pytest.param( |
| 106 | True, |
| 107 | REASONING_WITH_THINK, |
| 108 | id="reasoning_with_think_streaming", |
| 109 | ), |
| 110 | pytest.param( |
| 111 | True, |
| 112 | COMPLETE_REASONING_WITH_THINK, |
| 113 | id="complete_reasoning_with_think_streaming", |
| 114 | ), |
| 115 | pytest.param( |
| 116 | True, |
| 117 | MULTIPLE_LINES_WITH_THINK, |
| 118 | id="multiple_lines_with_think_streaming", |
| 119 | ), |
| 120 | ] |
| 121 | |
| 122 | # Global tokenizer initialization to avoid repeated loading |
| 123 | tokenizer = AutoTokenizer.from_pretrained("facebook/opt-125m") |
| 124 | |
| 125 | |
| 126 | @pytest.mark.parametrize("streaming, param_dict", TEST_CASES) |
| 127 | def test_reasoning( |
| 128 | streaming: bool, |
| 129 | param_dict: dict, |
| 130 | ): |
| 131 | output = tokenizer.tokenize(param_dict["output"]) |
| 132 | # decode everything to tokens |
| 133 | output_tokens: list[str] = [ |
| 134 | tokenizer.convert_tokens_to_string([token]) for token in output |
| 135 | ] |
| 136 | parser: ReasoningParser = ReasoningParserManager.get_reasoning_parser(parser_name)( |
| 137 | tokenizer |
| 138 | ) |
| 139 | |
| 140 | reasoning, content = run_reasoning_extraction( |
| 141 | parser, output_tokens, streaming=streaming |
| 142 | ) |
| 143 | |
| 144 | assert reasoning == param_dict["reasoning"] |
| 145 | assert content == param_dict["content"] |
| 146 | |
| 147 | |
| 148 | # Additional tests for verifying the correctness of granite streaming; this |
| 149 | # is complicated because granite uses multiple tokens to indicate when thinking |
| 150 | # is starting / when it's starting its response, so skipping special tokens |
| 151 | # is awkward. |
| 152 | |
| 153 | ### Handling the start of reasoning |
| 154 | STREAMING_1 = { |
| 155 | "previous_text": None, |
| 156 | "current_text": "Here", |
| 157 | "delta_text": "Here", |
| 158 | "reasoning": None, |
| 159 | "content": None, |
| 160 | } |
| 161 | # When we fail, we should give what was previously being silenced first |
| 162 | STREAMING_2 = { |
| 163 | "previous_text": "Here is my thought", |
| 164 | "current_text": "Here is my thought failure", |
| 165 | "delta_text": " failure", |
| 166 | "reasoning": None, |
| 167 | "content": "Here is my thought failure", |
| 168 | } |
| 169 | # But then after the first one, we should only add the delta text to content |
| 170 | STREAMING_3 = { |
| 171 | "previous_text": "Here wrong", |
| 172 | "current_text": " words", |
| 173 | "delta_text": " Here wrong words", |
| 174 | "reasoning": None, |
| 175 | "content": " words", |
| 176 | } |
| 177 | # But then after the first one, we should only add the delta text to content |
| 178 | STREAMING_4 = { |
| 179 | "previous_text": "Here is my thought", |
| 180 | "current_text": "Here is my thought process:", |
| 181 | "delta_text": " process:", |
| 182 | "reasoning": None, |
| 183 | "content": None, |
| 184 | } |
| 185 | # Reasoning started successfully; parse reasoning content |
| 186 | STREAMING_5 = { |
| 187 | "previous_text": "Here is my thought process:", |
| 188 | "current_text": "Here is my thought process: foo", |
| 189 | "delta_text": " foo", |
| 190 | "reasoning": " foo", |
| 191 | "content": None, |
| 192 | } |
| 193 | # Response special sequence has started, but not finished. |
| 194 | STREAMING_6 = { |
| 195 | "previous_text": "Here is my thought process: foo", |
| 196 | "current_text": "Here is my thought process: foo Here is", |
| 197 | "delta_text": " Here is", |
| 198 | "reasoning": " ", |
| 199 | "content": None, |
| 200 | } |
| 201 | # Response special sequence started, but was broken; the reasoning |
| 202 | # content should be the content that was previously unused. |
| 203 | STREAMING_7 = { |
| 204 | "previous_text": "Here is my thought process: foo Here is", |
| 205 | "current_text": "Here is my thought process: foo Here is Here", |
| 206 | "delta_text": " Here", |
| 207 | "reasoning": "Here is ", |
| 208 | "content": None, |
| 209 | } |
| 210 | # Response special sequence is ongoing |
| 211 | STREAMING_8 = { |
| 212 | "previous_text": "Here is my thought process: foo Here is my response:", |
| 213 | "current_text": "Here is my thought process: foo Here is my response: bar", |
| 214 | "delta_text": " bar", |
| 215 | "reasoning": None, |
| 216 | "content": " bar", |
| 217 | } |
| 218 | # The delta text has everything; we should be able to correctly parse both |
| 219 | STREAMING_9 = { |
| 220 | "previous_text": None, |
| 221 | "current_text": "Here is my thought process: foo Here is my response: bar", |
| 222 | "delta_text": "Here is my thought process: foo Here is my response: bar", |
| 223 | "reasoning": " foo ", |
| 224 | "content": " bar", |
| 225 | } |
| 226 | ## The Response is ongoing, and the delta mixes reasoning content / content |
| 227 | STREAMING_10 = { |
| 228 | "previous_text": "Here is my thought process: foo", |
| 229 | "current_text": "Here is my thought process: foo bar Here is my response: baz", |
| 230 | "delta_text": " bar Here is my response: baz", |
| 231 | "reasoning": " bar ", |
| 232 | "content": " baz", |
| 233 | } |
| 234 | # The delta text starts a new substring that might be a response special seq |
| 235 | STREAMING_11 = { |
| 236 | "previous_text": "Here is my thought process: This is a reasoning section ", |
| 237 | "current_text": "Here is my thought process: This is a reasoning section Here", |
| 238 | "delta_text": "Here", |
| 239 | "reasoning": None, |
| 240 | "content": None, |
| 241 | } |
| 242 | # The delta text is finishing the response special seq |
| 243 | STREAMING_12 = { |
| 244 | "previous_text": "Here is my thought process: foo Here is my response", |
| 245 | "current_text": "Here is my thought process: foo Here is my response:", |
| 246 | "delta_text": ":", |
| 247 | "reasoning": None, |
| 248 | "content": None, |
| 249 | } |
| 250 | STREAMING_13 = { |
| 251 | "previous_text": "Here is my thought process: foo Here", |
| 252 | "current_text": "Here is my thought process: foo Here was", |
| 253 | "delta_text": " was", |
| 254 | "reasoning": "Here was", |
| 255 | "content": None, |
| 256 | } |
| 257 | |
| 258 | STREAMING_SUBCASES = [ |
| 259 | pytest.param( |
| 260 | STREAMING_1, |
| 261 | id="Starting reasoning special sequence", |
| 262 | ), |
| 263 | pytest.param( |
| 264 | STREAMING_2, |
| 265 | id="Unexpected start reasoning sequence", |
| 266 | ), |
| 267 | pytest.param( |
| 268 | STREAMING_3, |
| 269 | id="Continuing unexpected start reasoning sequence", |
| 270 | ), |
| 271 | pytest.param( |
| 272 | STREAMING_4, |
| 273 | id="Only start reasoning sequence and nothing else", |
| 274 | ), |
| 275 | pytest.param( |
| 276 | STREAMING_5, |
| 277 | id="Reasoning content has started", |
| 278 | ), |
| 279 | pytest.param( |
| 280 | STREAMING_6, |
| 281 | id="Response special sequence has started", |
| 282 | ), |
| 283 | pytest.param( |
| 284 | STREAMING_7, |
| 285 | id="Response special sequence reset", |
| 286 | ), |
| 287 | pytest.param( |
| 288 | STREAMING_8, |
| 289 | id="Response text has started", |
| 290 | ), |
| 291 | pytest.param( |
| 292 | STREAMING_9, |
| 293 | id="Delta contains everything", |
| 294 | ), |
| 295 | pytest.param( |
| 296 | STREAMING_10, |
| 297 | id="Delta contains some reasoning and response", |
| 298 | ), |
| 299 | pytest.param( |
| 300 | STREAMING_11, |
| 301 | id="Delta starts response sequence", |
| 302 | ), |
| 303 | pytest.param( |
| 304 | STREAMING_12, |
| 305 | id="Delta finishes response sequence", |
| 306 | ), |
| 307 | pytest.param( |
| 308 | STREAMING_13, |
| 309 | id="Delta breaks potential responise sequence", |
| 310 | ), |
| 311 | ] |
| 312 | |
| 313 | |
| 314 | @pytest.mark.parametrize("param_dict", STREAMING_SUBCASES) |
| 315 | def test_streaming_subcases(param_dict): |
| 316 | # Get all of the token IDs |
| 317 | previous_token_ids = ( |
| 318 | tokenizer.encode(param_dict["previous_text"]) |
| 319 | if param_dict["previous_text"] is not None |
| 320 | else [] |
| 321 | ) |
| 322 | current_token_ids = tokenizer.encode(param_dict["current_text"]) |
| 323 | delta_token_ids = tokenizer.encode(param_dict["delta_text"]) |
| 324 | |
| 325 | parser: ReasoningParser = ReasoningParserManager.get_reasoning_parser(parser_name)( |
| 326 | tokenizer |
| 327 | ) |
| 328 | |
| 329 | response = parser.extract_reasoning_streaming( |
| 330 | previous_text=param_dict["previous_text"], |
| 331 | current_text=param_dict["current_text"], |
| 332 | delta_text=param_dict["delta_text"], |
| 333 | previous_token_ids=previous_token_ids, |
| 334 | current_token_ids=current_token_ids, |
| 335 | delta_token_ids=delta_token_ids, |
| 336 | ) |
| 337 | # Streaming currently expects at least one of reasoning content / content, |
| 338 | # so the response should return None in that case. |
| 339 | if param_dict["reasoning"] is None and param_dict["content"] is None: |
| 340 | assert response is None |
| 341 | else: |
| 342 | assert isinstance(response, DeltaMessage) |
| 343 | assert param_dict["reasoning"] == response.reasoning |
| 344 | assert param_dict["content"] == response.content |