| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Tests for DeepSeek V4-specific parser engine semantics.""" |
| 4 | |
| 5 | import json |
| 6 | |
| 7 | import pytest |
| 8 | |
| 9 | from tests.parser.engine.conftest import make_mock_tokenizer |
| 10 | from tests.parser.engine.replay_harness import ( |
| 11 | DUMMY_TOOLS, |
| 12 | MockTokenizer, |
| 13 | _test_request, |
| 14 | collect_output, |
| 15 | replay_streaming, |
| 16 | ) |
| 17 | from tests.parser.engine.streaming_helpers import ( |
| 18 | collect_content, |
| 19 | collect_function_name, |
| 20 | collect_tool_arguments, |
| 21 | simulate_reasoning_streaming, |
| 22 | simulate_tool_streaming, |
| 23 | ) |
| 24 | from vllm.parser.abstract_parser import DelegatingParser |
| 25 | from vllm.parser.deepseek_v4 import ( |
| 26 | DSML_INVOKE_END, |
| 27 | DSML_INVOKE_NAME_END, |
| 28 | DSML_INVOKE_PREFIX, |
| 29 | DSML_THINK_END, |
| 30 | DSML_THINK_START, |
| 31 | DSML_TOOL_END, |
| 32 | DSML_TOOL_START, |
| 33 | DSML_TOOL_START_VARIANTS, |
| 34 | DeepSeekV4Parser, |
| 35 | _dsml_arg_converter, |
| 36 | _unwrap_wrapper_args, |
| 37 | deepseek_v4_config, |
| 38 | ) |
| 39 | from vllm.parser.engine.registered_adapters import ( |
| 40 | DeepSeekV4ParserReasoningAdapter, |
| 41 | DeepSeekV4ParserToolAdapter, |
| 42 | ) |
| 43 | |
| 44 | _THINK_START_ID = 50 |
| 45 | _THINK_END_ID = 51 |
| 46 | |
| 47 | _PARAM_OPEN = '｜DSML｜parameter name="{name}" string="{is_str}">' |
| 48 | _PARAM_CLOSE = "</｜DSML｜parameter>" |
| 49 | |
| 50 | |
| 51 | def _param(name: str, is_str: str, value: str) -> str: |
| 52 | return f"<{_PARAM_OPEN.format(name=name, is_str=is_str)}{value}{_PARAM_CLOSE}" |
| 53 | |
| 54 | |
| 55 | @pytest.fixture |
| 56 | def mock_tokenizer(): |
| 57 | return make_mock_tokenizer( |
| 58 | { |
| 59 | DSML_THINK_START: _THINK_START_ID, |
| 60 | DSML_THINK_END: _THINK_END_ID, |
| 61 | } |
| 62 | ) |
| 63 | |
| 64 | |
| 65 | # ── Arg converter unit tests ───────────────────────────────────────── |
| 66 | |
| 67 | |
| 68 | class TestArgConverter: |
| 69 | def _raw(self, *params: tuple[str, str, str]) -> str: |
| 70 | lines = [_param(n, s, v) for n, s, v in params] |
| 71 | return "\n" + "\n".join(lines) + "\n" |
| 72 | |
| 73 | def test_string_param(self): |
| 74 | raw = self._raw(("city", "true", "杭州")) |
| 75 | result = json.loads(_dsml_arg_converter(raw, partial=False)) |
| 76 | assert result == {"city": "杭州"} |
| 77 | |
| 78 | def test_string_with_spaces_and_quotes(self): |
| 79 | raw = self._raw(("msg", "true", 'He said "hello world"')) |
| 80 | result = json.loads(_dsml_arg_converter(raw, partial=False)) |
| 81 | assert result["msg"] == 'He said "hello world"' |
| 82 | |
| 83 | def test_integer_param(self): |
| 84 | raw = self._raw(("count", "false", "42")) |
| 85 | result = json.loads(_dsml_arg_converter(raw, partial=False)) |
| 86 | assert result["count"] == 42 |
| 87 | assert isinstance(result["count"], int) |
| 88 | |
| 89 | def test_float_param(self): |
| 90 | raw = self._raw(("ratio", "false", "3.14")) |
| 91 | result = json.loads(_dsml_arg_converter(raw, partial=False)) |
| 92 | assert abs(result["ratio"] - 3.14) < 1e-9 |
| 93 | |
| 94 | def test_bool_param(self): |
| 95 | raw = self._raw(("flag", "false", "true")) |
| 96 | result = json.loads(_dsml_arg_converter(raw, partial=False)) |
| 97 | assert result["flag"] is True |
| 98 | |
| 99 | def test_array_param(self): |
| 100 | raw = self._raw(("items", "false", '["a", "b", "c"]')) |
| 101 | result = json.loads(_dsml_arg_converter(raw, partial=False)) |
| 102 | assert result["items"] == ["a", "b", "c"] |
| 103 | |
| 104 | def test_object_param(self): |
| 105 | raw = self._raw(("opts", "false", '{"key": "val"}')) |
| 106 | result = json.loads(_dsml_arg_converter(raw, partial=False)) |
| 107 | assert result["opts"] == {"key": "val"} |
| 108 | |
| 109 | def test_mixed_types(self): |
| 110 | raw = self._raw( |
| 111 | ("location", "true", "Tokyo"), |
| 112 | ("limit", "false", "10"), |
| 113 | ("active", "false", "false"), |
| 114 | ) |
| 115 | result = json.loads(_dsml_arg_converter(raw, partial=False)) |
| 116 | assert result == {"location": "Tokyo", "limit": 10, "active": False} |
| 117 | |
| 118 | def test_empty_args(self): |
| 119 | result = json.loads(_dsml_arg_converter("", partial=False)) |
| 120 | assert result == {} |
| 121 | |
| 122 | def _bare(self, name: str, value: str) -> str: |
| 123 | return f'<｜DSML｜parameter name="{name}">{value}{_PARAM_CLOSE}' |
| 124 | |
| 125 | def test_missing_string_attribute_keeps_literal(self): |
| 126 | """The model sometimes omits ``string``; the parameter must survive.""" |
| 127 | raw = self._bare("name", "alpha") |
| 128 | assert json.loads(_dsml_arg_converter(raw, partial=False)) == {"name": "alpha"} |
| 129 | |
| 130 | def test_missing_string_attribute_parses_json(self): |
| 131 | raw = self._bare("count", "7") + self._bare("opts", '{"k": [1]}') |
| 132 | assert json.loads(_dsml_arg_converter(raw, partial=False)) == { |
| 133 | "count": 7, |
| 134 | "opts": {"k": [1]}, |
| 135 | } |
| 136 | |
| 137 | def test_missing_string_attribute_alongside_annotated(self): |
| 138 | raw = self._bare("a", "x") + _param("b", "true", "42") |
| 139 | assert json.loads(_dsml_arg_converter(raw, partial=False)) == { |
| 140 | "a": "x", |
| 141 | "b": "42", |
| 142 | } |
| 143 | |
| 144 | def test_missing_string_attribute_partial_value(self): |
| 145 | raw = _param("a", "true", "x") + '<｜DSML｜parameter name="n">12' |
| 146 | assert json.loads(_dsml_arg_converter(raw, partial=True)) == { |
| 147 | "a": "x", |
| 148 | "n": 12, |
| 149 | } |
| 150 | |
| 151 | def test_invalid_json_fallback(self): |
| 152 | raw = self._raw(("data", "false", "[broken")) |
| 153 | result = json.loads(_dsml_arg_converter(raw, partial=False)) |
| 154 | assert result["data"] == "[broken" |
| 155 | |
| 156 | def test_chinese_chars_preserved_in_json(self): |
| 157 | raw = self._raw(("query", "true", "你好世界")) |
| 158 | raw_json = _dsml_arg_converter(raw, partial=False) |
| 159 | assert "你好世界" in raw_json |
| 160 | result = json.loads(raw_json) |
| 161 | assert result["query"] == "你好世界" |
| 162 | |
| 163 | def test_partial_complete_plus_in_progress(self): |
| 164 | raw = self._raw(("city", "true", "Tokyo")) |
| 165 | raw += f"<{_PARAM_OPEN.format(name='unit', is_str='true')}celsi" |
| 166 | result = json.loads(_dsml_arg_converter(raw, partial=True)) |
| 167 | assert result["city"] == "Tokyo" |
| 168 | assert result["unit"] == "celsi" |
| 169 | |
| 170 | def test_partial_no_in_progress(self): |
| 171 | raw = self._raw(("city", "true", "Tokyo")) |
| 172 | result = json.loads(_dsml_arg_converter(raw, partial=True)) |
| 173 | assert result == {"city": "Tokyo"} |
| 174 | |
| 175 | def test_partial_value_with_angle_bracket(self): |
| 176 | raw = f"<{_PARAM_OPEN.format(name='code', is_str='true')}a<b" |
| 177 | result = json.loads(_dsml_arg_converter(raw, partial=True)) |
| 178 | assert result == {"code": "a<b"} |
| 179 | |
| 180 | def test_partial_value_with_angle_bracket_and_complete_param(self): |
| 181 | raw = self._raw(("city", "true", "Tokyo")) |
| 182 | raw += f"<{_PARAM_OPEN.format(name='expr', is_str='true')}x<5" |
| 183 | result = json.loads(_dsml_arg_converter(raw, partial=True)) |
| 184 | assert result["city"] == "Tokyo" |
| 185 | assert result["expr"] == "x<5" |
| 186 | |
| 187 | def test_next_parameter_prefix_implicitly_closes_previous(self): |
| 188 | raw = ( |
| 189 | f"<{_PARAM_OPEN.format(name='city', is_str='true')}Tokyo\n" |
| 190 | "<｜DSML｜parameter name=" |
| 191 | ) |
| 192 | result = json.loads(_dsml_arg_converter(raw, partial=False)) |
| 193 | assert result == {"city": "Tokyo\n"} |
| 194 | |
| 195 | def test_implicit_close_preserves_malformed_text(self): |
| 196 | raw = ( |
| 197 | f"<{_PARAM_OPEN.format(name='city', is_str='true')}" |
| 198 | "Tokyo</｜DSML｜parameter\n" |
| 199 | f"<{_PARAM_OPEN.format(name='unit', is_str='true')}celsius{_PARAM_CLOSE}" |
| 200 | ) |
| 201 | result = json.loads(_dsml_arg_converter(raw, partial=False)) |
| 202 | assert result == { |
| 203 | "city": "Tokyo</｜DSML｜parameter\n", |
| 204 | "unit": "celsius", |
| 205 | } |
| 206 | |
| 207 | def test_null_string_false(self): |
| 208 | raw = self._raw(("val", "false", "null")) |
| 209 | result = json.loads(_dsml_arg_converter(raw, partial=False)) |
| 210 | assert result["val"] is None |
| 211 | |
| 212 | def test_string_true_not_json_parsed(self): |
| 213 | raw = self._raw(("n", "true", "42")) |
| 214 | result = json.loads(_dsml_arg_converter(raw, partial=False)) |
| 215 | assert result["n"] == "42" |
| 216 | assert isinstance(result["n"], str) |
| 217 | |
| 218 | |
| 219 | class TestImplicitParameterClose: |
| 220 | def test_non_streaming_recovers_next_parameter(self, mock_tokenizer, mock_request): |
| 221 | text = ( |
| 222 | f"{DSML_TOOL_START}" |
| 223 | f"{DSML_INVOKE_PREFIX}get_weather{DSML_INVOKE_NAME_END}\n" |
| 224 | f"<{_PARAM_OPEN.format(name='location', is_str='true')}Paris " |
| 225 | f"<{_PARAM_OPEN.format(name='date', is_str='true')}" |
| 226 | f"tomorrow{_PARAM_CLOSE}" |
| 227 | f"{DSML_INVOKE_END}{DSML_TOOL_END}" |
| 228 | ) |
| 229 | |
| 230 | result = DeepSeekV4Parser(mock_tokenizer).extract_tool_calls(text, mock_request) |
| 231 | |
| 232 | assert result.tools_called is True |
| 233 | assert json.loads(result.tool_calls[0].function.arguments) == { |
| 234 | "location": "Paris ", |
| 235 | "date": "tomorrow", |
| 236 | } |
| 237 | |
| 238 | def test_streaming_split_next_parameter_tag_is_buffered( |
| 239 | self, mock_tokenizer, mock_request |
| 240 | ): |
| 241 | chunks = [ |
| 242 | DSML_TOOL_START, |
| 243 | f"{DSML_INVOKE_PREFIX}get_weather{DSML_INVOKE_NAME_END}\n", |
| 244 | f"<{_PARAM_OPEN.format(name='location', is_str='true')}" |
| 245 | "Paris a<b><｜DSML｜parameter", |
| 246 | ' name="date" string="true">tomorrow', |
| 247 | _PARAM_CLOSE, |
| 248 | DSML_INVOKE_END, |
| 249 | DSML_TOOL_END, |
| 250 | ] |
| 251 | |
| 252 | results = simulate_tool_streaming( |
| 253 | DeepSeekV4Parser(mock_tokenizer), mock_request, chunks |
| 254 | ) |
| 255 | |
| 256 | arguments = collect_tool_arguments(results) |
| 257 | assert "<｜DSML｜param" not in arguments |
| 258 | assert json.loads(arguments) == { |
| 259 | "location": "Paris a<b>", |
| 260 | "date": "tomorrow", |
| 261 | } |
| 262 | |
| 263 | |
| 264 | class TestMissingStringAttribute: |
| 265 | def test_non_streaming_keeps_parameter(self, mock_tokenizer, mock_request): |
| 266 | text = ( |
| 267 | f"{DSML_TOOL_START}" |
| 268 | f"{DSML_INVOKE_PREFIX}get_weather{DSML_INVOKE_NAME_END}\n" |
| 269 | f'<｜DSML｜parameter name="location">Paris{_PARAM_CLOSE}\n' |
| 270 | f"{_param('date', 'true', 'tomorrow')}\n" |
| 271 | f"{DSML_INVOKE_END}{DSML_TOOL_END}" |
| 272 | ) |
| 273 | |
| 274 | result = DeepSeekV4Parser(mock_tokenizer).extract_tool_calls(text, mock_request) |
| 275 | |
| 276 | assert result.tools_called is True |
| 277 | assert json.loads(result.tool_calls[0].function.arguments) == { |
| 278 | "location": "Paris", |
| 279 | "date": "tomorrow", |
| 280 | } |
| 281 | |
| 282 | def test_streaming_keeps_parameter(self, mock_tokenizer, mock_request): |
| 283 | chunks = [ |
| 284 | DSML_TOOL_START, |
| 285 | f"{DSML_INVOKE_PREFIX}get_weather{DSML_INVOKE_NAME_END}\n", |
| 286 | '<｜DSML｜parameter name="loc', |
| 287 | 'ation">Par', |
| 288 | "is", |
| 289 | _PARAM_CLOSE, |
| 290 | f"\n{_param('date', 'true', 'tomorrow')}\n", |
| 291 | DSML_INVOKE_END, |
| 292 | DSML_TOOL_END, |
| 293 | ] |
| 294 | |
| 295 | results = simulate_tool_streaming( |
| 296 | DeepSeekV4Parser(mock_tokenizer), mock_request, chunks |
| 297 | ) |
| 298 | |
| 299 | assert json.loads(collect_tool_arguments(results)) == { |
| 300 | "location": "Paris", |
| 301 | "date": "tomorrow", |
| 302 | } |
| 303 | |
| 304 | |
| 305 | # ── Bare </think> absorption and duplicate <think> absorption ───────── |
| 306 | |
| 307 | |
| 308 | class TestThinkTagAbsorption: |
| 309 | def test_bare_think_end_not_leaked(self, mock_tokenizer): |
| 310 | parser = DeepSeekV4Parser(mock_tokenizer) |
| 311 | chunks = ["</think>", "Here is the direct answer."] |
| 312 | reasoning, content = simulate_reasoning_streaming(parser, chunks) |
| 313 | assert reasoning == "" |
| 314 | assert "</think>" not in content |
| 315 | assert "Here is the direct answer" in content |
| 316 | |
| 317 | def test_duplicate_think_start_absorbed(self, mock_tokenizer): |
| 318 | parser = DeepSeekV4Parser( |
| 319 | mock_tokenizer, chat_template_kwargs={"thinking": True} |
| 320 | ) |
| 321 | chunks = [ |
| 322 | "<think>\n", |
| 323 | "Some reasoning.\n", |
| 324 | "</think>\n", |
| 325 | "Answer.", |
| 326 | ] |
| 327 | reasoning, content = simulate_reasoning_streaming(parser, chunks) |
| 328 | assert "Some reasoning" in reasoning |
| 329 | assert "Answer" in content |
| 330 | |
| 331 | |
| 332 | # ── Missing </｜DSML｜invoke> before </｜DSML｜tool_calls> ──────────── |
| 333 | |
| 334 | |
| 335 | class TestMissingInvokeEnd: |
| 336 | def test_non_streaming(self, mock_tokenizer, mock_request): |
| 337 | parser = DeepSeekV4Parser(mock_tokenizer) |
| 338 | text = ( |
| 339 | f"{DSML_TOOL_START}" |
| 340 | f"{DSML_INVOKE_PREFIX}get_weather{DSML_INVOKE_NAME_END}\n" |
| 341 | f"{_param('location', 'true', 'NYC')}\n" |
| 342 | f"{DSML_TOOL_END}" |
| 343 | ) |
| 344 | result = parser.extract_tool_calls(text, mock_request) |
| 345 | |
| 346 | assert result.tools_called is True |
| 347 | assert len(result.tool_calls) == 1 |
| 348 | assert result.tool_calls[0].function.name == "get_weather" |
| 349 | args = json.loads(result.tool_calls[0].function.arguments) |
| 350 | assert args == {"location": "NYC"} |
| 351 | |
| 352 | def test_streaming_with_trailing_content(self, mock_tokenizer, mock_request): |
| 353 | parser = DeepSeekV4Parser(mock_tokenizer) |
| 354 | chunks = [ |
| 355 | DSML_TOOL_START, |
| 356 | f"{DSML_INVOKE_PREFIX}get_weather{DSML_INVOKE_NAME_END}\n" |
| 357 | f"{_param('location', 'true', 'NYC')}\n", |
| 358 | DSML_TOOL_END, |
| 359 | "Done.", |
| 360 | ] |
| 361 | |
| 362 | results = simulate_tool_streaming(parser, mock_request, chunks) |
| 363 | |
| 364 | assert collect_function_name(results) == "get_weather" |
| 365 | args = json.loads(collect_tool_arguments(results)) |
| 366 | assert args == {"location": "NYC"} |
| 367 | # A tool call ends the turn; anything after the block is dropped. |
| 368 | assert "Done." not in collect_content(results) |
| 369 | assert "DSML" not in collect_content(results) |
| 370 | |
| 371 | |
| 372 | class TestMissingToolCallsWrapper: |
| 373 | """The model sometimes drops ``<｜DSML｜tool_calls>`` at long context and |
| 374 | opens ``<｜DSML｜invoke ...>`` directly (#48931). The invoke itself must |
| 375 | anchor tool-call detection so the block is never leaked as content. |
| 376 | """ |
| 377 | |
| 378 | _ORPHAN = ( |
| 379 | f"{DSML_INVOKE_PREFIX}terminal{DSML_INVOKE_NAME_END}\n" |
| 380 | f"{_param('command', 'true', 'echo hi')}\n" |
| 381 | f"{DSML_INVOKE_END}\n" |
| 382 | ) |
| 383 | |
| 384 | @pytest.mark.parametrize("trailing_end", [True, False]) |
| 385 | def test_non_streaming(self, mock_tokenizer, mock_request, trailing_end): |
| 386 | text = "Let me run it.\n" + self._ORPHAN |
| 387 | if trailing_end: |
| 388 | text += DSML_TOOL_END |
| 389 | parser = DeepSeekV4Parser( |
| 390 | mock_tokenizer, chat_template_kwargs={"thinking": False} |
| 391 | ) |
| 392 | result = parser.extract_tool_calls(text, mock_request) |
| 393 | |
| 394 | assert result.tools_called is True |
| 395 | assert len(result.tool_calls) == 1 |
| 396 | assert result.tool_calls[0].function.name == "terminal" |
| 397 | args = json.loads(result.tool_calls[0].function.arguments) |
| 398 | assert args == {"command": "echo hi"} |
| 399 | assert result.content == "Let me run it.\n" |
| 400 | |
| 401 | @pytest.mark.parametrize("trailing_end", [True, False]) |
| 402 | def test_trailing_content_dropped_consistently( |
| 403 | self, mock_tokenizer, mock_request, trailing_end |
| 404 | ): |
| 405 | """Text after a tool block is dropped whether or not the closing |
| 406 | wrapper is present, so a missing ``</｜DSML｜tool_calls>`` does not |
| 407 | change what the client sees.""" |
| 408 | text = "pre\n" + self._ORPHAN |
| 409 | if trailing_end: |
| 410 | text += DSML_TOOL_END |
| 411 | text += "Done." |
| 412 | parser = DeepSeekV4Parser( |
| 413 | mock_tokenizer, chat_template_kwargs={"thinking": False} |
| 414 | ) |
| 415 | result = parser.extract_tool_calls(text, mock_request) |
| 416 | assert [tc.function.name for tc in result.tool_calls] == ["terminal"] |
| 417 | assert result.content == "pre\n" |
| 418 | |
| 419 | parser = DeepSeekV4Parser( |
| 420 | mock_tokenizer, chat_template_kwargs={"thinking": False} |
| 421 | ) |
| 422 | results = simulate_tool_streaming(parser, mock_request, list(text)) |
| 423 | assert collect_function_name(results) == "terminal" |
| 424 | assert collect_content(results) == "pre\n" |
| 425 | |
| 426 | def test_second_wrapped_block_after_orphan(self, mock_tokenizer, mock_request): |
| 427 | parser = DeepSeekV4Parser( |
| 428 | mock_tokenizer, chat_template_kwargs={"thinking": False} |
| 429 | ) |
| 430 | text = self._ORPHAN + DSML_TOOL_START + self._ORPHAN + DSML_TOOL_END |
| 431 | result = parser.extract_tool_calls(text, mock_request) |
| 432 | assert [tc.function.name for tc in result.tool_calls] == [ |
| 433 | "terminal", |
| 434 | "terminal", |
| 435 | ] |
| 436 | assert result.content is None |
| 437 | |
| 438 | def test_streaming_marker_split_across_deltas(self, mock_tokenizer, mock_request): |
| 439 | text = "Let me run it.\n" + self._ORPHAN + DSML_TOOL_END |
| 440 | parser = DeepSeekV4Parser( |
| 441 | mock_tokenizer, chat_template_kwargs={"thinking": False} |
| 442 | ) |
| 443 | results = simulate_tool_streaming(parser, mock_request, list(text)) |
| 444 | |
| 445 | assert collect_function_name(results) == "terminal" |
| 446 | args = json.loads(collect_tool_arguments(results)) |
| 447 | assert args == {"command": "echo hi"} |
| 448 | content = collect_content(results) |
| 449 | assert "Let me run it." in content |
| 450 | assert "DSML" not in content |
| 451 | |
| 452 | def test_parallel_orphan_invokes(self, mock_tokenizer, mock_request): |
| 453 | parser = DeepSeekV4Parser( |
| 454 | mock_tokenizer, chat_template_kwargs={"thinking": False} |
| 455 | ) |
| 456 | result = parser.extract_tool_calls(self._ORPHAN + self._ORPHAN, mock_request) |
| 457 | |
| 458 | assert [tc.function.name for tc in result.tool_calls] == [ |
| 459 | "terminal", |
| 460 | "terminal", |
| 461 | ] |
| 462 | assert result.content is None |
| 463 | |
| 464 | def test_orphan_invoke_inside_reasoning_is_not_a_tool_call( |
| 465 | self, mock_tokenizer, mock_request |
| 466 | ): |
| 467 | parser = DeepSeekV4Parser( |
| 468 | mock_tokenizer, chat_template_kwargs={"thinking": True} |
| 469 | ) |
| 470 | text = "Let me run it.\n\n" + self._ORPHAN + DSML_TOOL_END |
| 471 | reasoning, content, tool_calls = parser.parse(text, mock_request) |
| 472 | |
| 473 | assert not tool_calls |
| 474 | assert content is None |
| 475 | assert reasoning is not None and reasoning.startswith("Let me run it.") |
| 476 | |
| 477 | def test_drafted_invoke_in_reasoning_does_not_hijack_real_call( |
| 478 | self, mock_tokenizer, mock_request |
| 479 | ): |
| 480 | """A drafted invoke in ``<think>`` must not steal the real call.""" |
| 481 | parser = DeepSeekV4Parser( |
| 482 | mock_tokenizer, chat_template_kwargs={"thinking": True} |
| 483 | ) |
| 484 | text = ( |
| 485 | f"Maybe {DSML_INVOKE_PREFIX}search{DSML_INVOKE_NAME_END} is wrong; " |
| 486 | f"use terminal.{DSML_THINK_END}\n" |
| 487 | f"{DSML_TOOL_START}\n{self._ORPHAN}{DSML_TOOL_END}" |
| 488 | ) |
| 489 | reasoning, content, tool_calls = parser.parse(text, mock_request) |
| 490 | |
| 491 | assert tool_calls is not None |
| 492 | assert [tc.name for tc in tool_calls] == ["terminal"] |
| 493 | assert json.loads(tool_calls[0].arguments) == {"command": "echo hi"} |
| 494 | assert reasoning is not None and "use terminal." in reasoning |
| 495 | |
| 496 | def test_streaming_orphan_invoke_inside_reasoning_is_not_a_tool_call( |
| 497 | self, mock_tokenizer |
| 498 | ): |
| 499 | parser = DeepSeekV4Parser( |
| 500 | mock_tokenizer, chat_template_kwargs={"thinking": True} |
| 501 | ) |
| 502 | chunks = ["Let me run it.\n\n", *list(self._ORPHAN), DSML_TOOL_END] |
| 503 | reasoning, content = simulate_reasoning_streaming(parser, chunks) |
| 504 | |
| 505 | assert reasoning.startswith("Let me run it.") |
| 506 | assert content == "" |
| 507 | |
| 508 | def test_wrapper_present_is_unchanged(self, mock_tokenizer, mock_request): |
| 509 | parser = DeepSeekV4Parser(mock_tokenizer) |
| 510 | text = DSML_TOOL_START + "\n" + self._ORPHAN + DSML_TOOL_END |
| 511 | result = parser.extract_tool_calls(text, mock_request) |
| 512 | |
| 513 | assert len(result.tool_calls) == 1 |
| 514 | assert result.tool_calls[0].function.name == "terminal" |
| 515 | assert result.content is None |
| 516 | |
| 517 | |
| 518 | # ── Thinking mode initial state ────────────────────────────────────── |
| 519 | |
| 520 | |
| 521 | class TestThinkingModeConfig: |
| 522 | def test_thinking_true_starts_in_reasoning(self): |
| 523 | cfg = deepseek_v4_config(thinking=True) |
| 524 | assert cfg.initial_state.name == "REASONING" |
| 525 | |
| 526 | def test_thinking_false_starts_in_content(self): |
| 527 | cfg = deepseek_v4_config(thinking=False) |
| 528 | assert cfg.initial_state.name == "CONTENT" |
| 529 | |
| 530 | @pytest.mark.parametrize( |
| 531 | ("chat_template_kwargs", "expected_state"), |
| 532 | [ |
| 533 | ({}, "REASONING"), |
| 534 | ({"thinking": True}, "REASONING"), |
| 535 | ({"enable_thinking": True}, "REASONING"), |
| 536 | ({"reasoning_effort": "high"}, "REASONING"), |
| 537 | ({"thinking": False}, "CONTENT"), |
| 538 | ({"enable_thinking": False}, "CONTENT"), |
| 539 | ( |
| 540 | {"enable_thinking": True, "reasoning_effort": "none"}, |
| 541 | "CONTENT", |
| 542 | ), |
| 543 | ], |
| 544 | ) |
| 545 | def test_parser_thinking_mode_matches_tokenizer_default( |
| 546 | self, mock_tokenizer, chat_template_kwargs, expected_state |
| 547 | ): |
| 548 | parser = DeepSeekV4Parser( |
| 549 | mock_tokenizer, |
| 550 | chat_template_kwargs=chat_template_kwargs, |
| 551 | ) |
| 552 | assert parser.parser_engine_config.initial_state.name == expected_state |
| 553 | |
| 554 | def test_thinking_mode_reasoning_without_tags(self, mock_tokenizer): |
| 555 | parser = DeepSeekV4Parser( |
| 556 | mock_tokenizer, chat_template_kwargs={"thinking": True} |
| 557 | ) |
| 558 | chunks = [ |
| 559 | "\n\nLet me consider ", |
| 560 | "this carefully.\n", |
| 561 | "</think>\n", |
| 562 | "Here is the result.", |
| 563 | ] |
| 564 | reasoning, content = simulate_reasoning_streaming(parser, chunks) |
| 565 | assert "Let me consider" in reasoning |
| 566 | assert "Here is the result" in content |
| 567 | |
| 568 | def test_thinking_mode_all_reasoning_no_end_tag(self, mock_tokenizer): |
| 569 | parser = DeepSeekV4Parser( |
| 570 | mock_tokenizer, chat_template_kwargs={"thinking": True} |
| 571 | ) |
| 572 | chunks = ["I'll review ", "the PR."] |
| 573 | reasoning, content = simulate_reasoning_streaming(parser, chunks) |
| 574 | assert "review" in reasoning |
| 575 | assert "the PR" in reasoning |
| 576 | assert content == "" |
| 577 | |
| 578 | def test_reasoning_effort_none_overrides_enable_thinking(self, mock_tokenizer): |
| 579 | p = DeepSeekV4Parser( |
| 580 | mock_tokenizer, |
| 581 | chat_template_kwargs={ |
| 582 | "enable_thinking": True, |
| 583 | "reasoning_effort": "none", |
| 584 | }, |
| 585 | ) |
| 586 | assert p.parser_engine_config.initial_state.name == "CONTENT" |
| 587 | |
| 588 | |
| 589 | # ── Implicit reasoning end (missing </think> before tool calls) ───── |
| 590 | |
| 591 | |
| 592 | class TestImplicitReasoningEnd: |
| 593 | """Tool call markers end reasoning implicitly when </think> is missing. |
| 594 | |
| 595 | DeepSeek V4 models occasionally omit </think> before emitting tool calls. |
| 596 | The (REASONING, TOOL_START) transition handles this gracefully. |
| 597 | """ |
| 598 | |
| 599 | @pytest.fixture |
| 600 | def thinking_parser(self, mock_tokenizer): |
| 601 | return DeepSeekV4Parser(mock_tokenizer, chat_template_kwargs={"thinking": True}) |
| 602 | |
| 603 | def _reasoning_then_tool(self, reasoning_text: str) -> str: |
| 604 | return reasoning_text + _tool_calls( |
| 605 | _invoke("get_weather", ("location", "true", "NYC")), |
| 606 | ) |
| 607 | |
| 608 | def test_non_streaming_extract_reasoning_implicit_end(self, thinking_parser): |
| 609 | text = self._reasoning_then_tool("Let me look up the weather.\n\n") |
| 610 | reasoning, content = thinking_parser.extract_reasoning(text, None) |
| 611 | assert reasoning == "Let me look up the weather." |
| 612 | assert DSML_TOOL_START not in reasoning |
| 613 | assert DSML_INVOKE_PREFIX not in reasoning |
| 614 | assert content is None |
| 615 | |
| 616 | def test_non_streaming_extract_tool_calls_implicit_end( |
| 617 | self, thinking_parser, mock_request |
| 618 | ): |
| 619 | text = self._reasoning_then_tool("Let me look up the weather.\n\n") |
| 620 | result = thinking_parser.extract_tool_calls(text, mock_request) |
| 621 | assert result.tools_called is True |
| 622 | assert len(result.tool_calls) == 1 |
| 623 | assert result.tool_calls[0].function.name == "get_weather" |
| 624 | args = json.loads(result.tool_calls[0].function.arguments) |
| 625 | assert args == {"location": "NYC"} |
| 626 | |
| 627 | def test_non_streaming_parse_implicit_end(self, thinking_parser, mock_request): |
| 628 | text = self._reasoning_then_tool("Let me look up the weather.\n\n") |
| 629 | reasoning, content, tool_calls = thinking_parser.parse(text, mock_request) |
| 630 | assert reasoning == "Let me look up the weather." |
| 631 | assert content is None |
| 632 | assert tool_calls is not None |
| 633 | assert len(tool_calls) == 1 |
| 634 | assert tool_calls[0].name == "get_weather" |
| 635 | args = json.loads(tool_calls[0].arguments) |
| 636 | assert args == {"location": "NYC"} |
| 637 | |
| 638 | def test_reasoning_end_token_ids_drop_multi_token_tool_opener( |
| 639 | self, thinking_parser |
| 640 | ): |
| 641 | # Like the real DeepSeek-V4 tokenizer, the mock vocab has no single |
| 642 | # token for the DSML tool opener, so only </think> is tracked on the |
| 643 | # token-ID fast path; the implicit tool-call exit still goes through |
| 644 | # the text lexer. |
| 645 | assert thinking_parser.reasoning_end_token_ids == {_THINK_END_ID} |
| 646 | assert thinking_parser.find_reasoning_end_offset([7, _THINK_END_ID, 8]) == 1 |
| 647 | assert thinking_parser.find_reasoning_end_offset([7, 8]) == 2 |
| 648 | |
| 649 | def test_streaming_reasoning_implicit_end(self, thinking_parser): |
| 650 | chunks = [ |
| 651 | "Let me look up the weather.\n\n", |
| 652 | DSML_TOOL_START, |
| 653 | DSML_INVOKE_PREFIX + "get_weather" + DSML_INVOKE_NAME_END, |
| 654 | ] |
| 655 | reasoning, content = simulate_reasoning_streaming(thinking_parser, chunks) |
| 656 | assert reasoning == "Let me look up the weather." |
| 657 | assert DSML_TOOL_START not in reasoning |
| 658 | assert DSML_INVOKE_PREFIX not in reasoning |
| 659 | |
| 660 | def test_streaming_tool_extraction_implicit_end( |
| 661 | self, thinking_parser, mock_request |
| 662 | ): |
| 663 | chunks = [ |
| 664 | "Let me check.\n\n", |
| 665 | DSML_TOOL_START, |
| 666 | DSML_INVOKE_PREFIX |
| 667 | + "get_weather" |
| 668 | + DSML_INVOKE_NAME_END |
| 669 | + "\n" |
| 670 | + _param("location", "true", "NYC") |
| 671 | + "\n" |
| 672 | + DSML_INVOKE_END, |
| 673 | DSML_TOOL_END, |
| 674 | ] |
| 675 | results = simulate_tool_streaming(thinking_parser, mock_request, chunks) |
| 676 | assert collect_function_name(results) == "get_weather" |
| 677 | args = json.loads(collect_tool_arguments(results)) |
| 678 | assert args == {"location": "NYC"} |
| 679 | |
| 680 | def test_thinking_false_explicit_think_then_tool_call(self, mock_tokenizer): |
| 681 | parser = DeepSeekV4Parser(mock_tokenizer) |
| 682 | chunks = [ |
| 683 | DSML_THINK_START, |
| 684 | "Let me check the weather.", |
| 685 | DSML_TOOL_START, |
| 686 | DSML_INVOKE_PREFIX + "get_weather" + DSML_INVOKE_NAME_END, |
| 687 | ] |
| 688 | reasoning, content = simulate_reasoning_streaming(parser, chunks) |
| 689 | assert "Let me check the weather" in reasoning |
| 690 | assert DSML_TOOL_START not in reasoning |
| 691 | assert DSML_THINK_START not in reasoning |
| 692 | |
| 693 | def test_non_streaming_parallel_tools_after_implicit_end( |
| 694 | self, thinking_parser, mock_request |
| 695 | ): |
| 696 | text = "I need both.\n\n" + _tool_calls( |
| 697 | _invoke("get_weather", ("location", "true", "NYC")), |
| 698 | _invoke("get_time", ("timezone", "true", "EST")), |
| 699 | ) |
| 700 | result = thinking_parser.extract_tool_calls(text, mock_request) |
| 701 | assert result.tools_called is True |
| 702 | assert len(result.tool_calls) == 2 |
| 703 | assert result.tool_calls[0].function.name == "get_weather" |
| 704 | assert result.tool_calls[1].function.name == "get_time" |
| 705 | |
| 706 | def test_streaming_implicit_end_trailing_whitespace_stripped(self, thinking_parser): |
| 707 | chunks = [ |
| 708 | "Reasoning.\n\n\n", |
| 709 | DSML_TOOL_START, |
| 710 | DSML_INVOKE_PREFIX + "func" + DSML_INVOKE_NAME_END, |
| 711 | ] |
| 712 | reasoning, content = simulate_reasoning_streaming(thinking_parser, chunks) |
| 713 | assert reasoning == "Reasoning." |
| 714 | |
| 715 | |
| 716 | # ── Wrapper argument unwrapping ────────────────────────────────────── |
| 717 | |
| 718 | |
| 719 | class TestWrapperUnwrapping: |
| 720 | def test_unwrap_arguments_wrapper(self): |
| 721 | from vllm.entrypoints.openai.chat_completion.protocol import ( |
| 722 | ChatCompletionToolsParam, |
| 723 | ) |
| 724 | |
| 725 | tool = ChatCompletionToolsParam( |
| 726 | type="function", |
| 727 | function={ |
| 728 | "name": "get_weather", |
| 729 | "parameters": { |
| 730 | "type": "object", |
| 731 | "properties": {"location": {"type": "string"}}, |
| 732 | }, |
| 733 | }, |
| 734 | ) |
| 735 | |
| 736 | result = _unwrap_wrapper_args( |
| 737 | '{"arguments": {"location": "Beijing"}}', |
| 738 | [tool], |
| 739 | "get_weather", |
| 740 | ) |
| 741 | assert json.loads(result) == {"location": "Beijing"} |
| 742 | |
| 743 | def test_unwrap_input_wrapper(self): |
| 744 | from vllm.entrypoints.openai.chat_completion.protocol import ( |
| 745 | ChatCompletionToolsParam, |
| 746 | ) |
| 747 | |
| 748 | tool = ChatCompletionToolsParam( |
| 749 | type="function", |
| 750 | function={ |
| 751 | "name": "get_weather", |
| 752 | "parameters": { |
| 753 | "type": "object", |
| 754 | "properties": {"location": {"type": "string"}}, |
| 755 | }, |
| 756 | }, |
| 757 | ) |
| 758 | |
| 759 | result = _unwrap_wrapper_args( |
| 760 | '{"input": {"location": "Beijing"}}', |
| 761 | [tool], |
| 762 | "get_weather", |
| 763 | ) |
| 764 | assert json.loads(result) == {"location": "Beijing"} |
| 765 | |
| 766 | def test_no_unwrap_when_key_in_schema(self): |
| 767 | from vllm.entrypoints.openai.chat_completion.protocol import ( |
| 768 | ChatCompletionToolsParam, |
| 769 | ) |
| 770 | |
| 771 | tool = ChatCompletionToolsParam( |
| 772 | type="function", |
| 773 | function={ |
| 774 | "name": "func", |
| 775 | "parameters": { |
| 776 | "type": "object", |
| 777 | "properties": {"arguments": {"type": "string"}}, |
| 778 | }, |
| 779 | }, |
| 780 | ) |
| 781 | |
| 782 | result = _unwrap_wrapper_args( |
| 783 | '{"arguments": "some value"}', |
| 784 | [tool], |
| 785 | "func", |
| 786 | ) |
| 787 | assert json.loads(result) == {"arguments": "some value"} |
| 788 | |
| 789 | def test_no_unwrap_when_no_tools(self): |
| 790 | result = _unwrap_wrapper_args( |
| 791 | '{"arguments": {"location": "Beijing"}}', |
| 792 | None, |
| 793 | "get_weather", |
| 794 | ) |
| 795 | assert json.loads(result) == {"arguments": {"location": "Beijing"}} |
| 796 | |
| 797 | def test_unwrap_json_string_inner(self): |
| 798 | from vllm.entrypoints.openai.chat_completion.protocol import ( |
| 799 | ChatCompletionToolsParam, |
| 800 | ) |
| 801 | |
| 802 | tool = ChatCompletionToolsParam( |
| 803 | type="function", |
| 804 | function={ |
| 805 | "name": "get_weather", |
| 806 | "parameters": { |
| 807 | "type": "object", |
| 808 | "properties": {"location": {"type": "string"}}, |
| 809 | }, |
| 810 | }, |
| 811 | ) |
| 812 | |
| 813 | result = _unwrap_wrapper_args( |
| 814 | '{"arguments": "{\\"location\\": \\"Beijing\\"}"}', |
| 815 | [tool], |
| 816 | "get_weather", |
| 817 | ) |
| 818 | assert json.loads(result) == {"location": "Beijing"} |
| 819 | |
| 820 | |
| 821 | # ── Parallel tool call wrapper unwrapping ─────────────────────────── |
| 822 | |
| 823 | |
| 824 | def _make_tool(name, properties): |
| 825 | from vllm.entrypoints.openai.chat_completion.protocol import ( # noqa: E501 |
| 826 | ChatCompletionToolsParam, |
| 827 | ) |
| 828 | |
| 829 | return ChatCompletionToolsParam( |
| 830 | type="function", |
| 831 | function={ |
| 832 | "name": name, |
| 833 | "parameters": { |
| 834 | "type": "object", |
| 835 | "properties": properties, |
| 836 | }, |
| 837 | }, |
| 838 | ) |
| 839 | |
| 840 | |
| 841 | def _invoke(name, *params): |
| 842 | body = "\n".join(_param(n, s, v) for n, s, v in params) |
| 843 | return ( |
| 844 | f"{DSML_INVOKE_PREFIX}{name}{DSML_INVOKE_NAME_END}\n{body}\n{DSML_INVOKE_END}" |
| 845 | ) |
| 846 | |
| 847 | |
| 848 | def _tool_calls(*invokes): |
| 849 | return DSML_TOOL_START + "\n".join(invokes) + DSML_TOOL_END |
| 850 | |
| 851 | |
| 852 | class TestParallelUnwrapping: |
| 853 | @pytest.fixture |
| 854 | def weather_tool(self): |
| 855 | return _make_tool( |
| 856 | "get_weather", |
| 857 | { |
| 858 | "location": {"type": "string"}, |
| 859 | "unit": {"type": "string"}, |
| 860 | }, |
| 861 | ) |
| 862 | |
| 863 | @pytest.fixture |
| 864 | def time_tool(self): |
| 865 | return _make_tool( |
| 866 | "get_time", |
| 867 | {"timezone": {"type": "string"}}, |
| 868 | ) |
| 869 | |
| 870 | @pytest.mark.parametrize( |
| 871 | "weather_args, expected", |
| 872 | [ |
| 873 | ( |
| 874 | '{"location": "NYC", "unit": "celsius"}', |
| 875 | {"location": "NYC", "unit": "celsius"}, |
| 876 | ), |
| 877 | ('{"location": "NYC"}', {"location": "NYC"}), |
| 878 | ], |
| 879 | ids=["all_props", "subset_props"], |
| 880 | ) |
| 881 | def test_unwrap_parallel_uses_correct_schema( |
| 882 | self, |
| 883 | mock_tokenizer, |
| 884 | mock_request, |
| 885 | weather_tool, |
| 886 | time_tool, |
| 887 | weather_args, |
| 888 | expected, |
| 889 | ): |
| 890 | tools = [weather_tool, time_tool] |
| 891 | parser = DeepSeekV4Parser(mock_tokenizer, tools=tools) |
| 892 | mock_request.tools = tools |
| 893 | |
| 894 | text = _tool_calls( |
| 895 | _invoke("get_weather", ("arguments", "false", weather_args)), |
| 896 | _invoke("get_time", ("timezone", "true", "EST")), |
| 897 | ) |
| 898 | |
| 899 | result = parser.extract_tool_calls(text, mock_request) |
| 900 | |
| 901 | assert result.tools_called is True |
| 902 | assert len(result.tool_calls) == 2 |
| 903 | assert result.tool_calls[0].function.name == "get_weather" |
| 904 | args0 = json.loads(result.tool_calls[0].function.arguments) |
| 905 | assert args0 == expected |
| 906 | assert result.tool_calls[1].function.name == "get_time" |
| 907 | args1 = json.loads(result.tool_calls[1].function.arguments) |
| 908 | assert args1 == {"timezone": "EST"} |
| 909 | |
| 910 | def test_unwrap_parallel_streaming( |
| 911 | self, mock_tokenizer, mock_request, weather_tool, time_tool |
| 912 | ): |
| 913 | tools = [weather_tool, time_tool] |
| 914 | parser = DeepSeekV4Parser(mock_tokenizer, tools=tools) |
| 915 | mock_request.tools = tools |
| 916 | |
| 917 | chunks = [ |
| 918 | DSML_TOOL_START, |
| 919 | _invoke( |
| 920 | "get_weather", |
| 921 | ("arguments", "false", '{"location": "NYC"}'), |
| 922 | ), |
| 923 | _invoke("get_time", ("timezone", "true", "EST")), |
| 924 | DSML_TOOL_END, |
| 925 | ] |
| 926 | |
| 927 | results = simulate_tool_streaming(parser, mock_request, chunks) |
| 928 | final_delta, _ = results[-1] |
| 929 | finish_delta = parser.finish_streaming() |
| 930 | extracted = parser._build_extracted_result(final_delta, finish_delta) |
| 931 | |
| 932 | assert extracted.tools_called is True |
| 933 | assert len(extracted.tool_calls) == 2 |
| 934 | args0 = json.loads(extracted.tool_calls[0].function.arguments) |
| 935 | assert args0 == {"location": "NYC"} |
| 936 | args1 = json.loads(extracted.tool_calls[1].function.arguments) |
| 937 | assert args1 == {"timezone": "EST"} |
| 938 | |
| 939 | def test_no_unwrap_parallel_when_no_match( |
| 940 | self, mock_tokenizer, mock_request, weather_tool, time_tool |
| 941 | ): |
| 942 | tools = [weather_tool, time_tool] |
| 943 | parser = DeepSeekV4Parser(mock_tokenizer, tools=tools) |
| 944 | mock_request.tools = tools |
| 945 | |
| 946 | text = _tool_calls( |
| 947 | _invoke( |
| 948 | "get_weather", |
| 949 | ("arguments", "false", '{"unknown_key": "val"}'), |
| 950 | ), |
| 951 | _invoke("get_time", ("timezone", "true", "EST")), |
| 952 | ) |
| 953 | |
| 954 | result = parser.extract_tool_calls(text, mock_request) |
| 955 | |
| 956 | assert len(result.tool_calls) == 2 |
| 957 | args0 = json.loads(result.tool_calls[0].function.arguments) |
| 958 | assert args0 == {"arguments": {"unknown_key": "val"}} |
| 959 | args1 = json.loads(result.tool_calls[1].function.arguments) |
| 960 | assert args1 == {"timezone": "EST"} |
| 961 | |
| 962 | def test_unwrap_single_tool_still_works(self, mock_tokenizer, mock_request): |
| 963 | tool = _make_tool("get_weather", {"location": {"type": "string"}}) |
| 964 | tools = [tool] |
| 965 | parser = DeepSeekV4Parser(mock_tokenizer, tools=tools) |
| 966 | mock_request.tools = tools |
| 967 | |
| 968 | text = _tool_calls( |
| 969 | _invoke( |
| 970 | "get_weather", |
| 971 | ("arguments", "false", '{"location": "Beijing"}'), |
| 972 | ), |
| 973 | ) |
| 974 | |
| 975 | result = parser.extract_tool_calls(text, mock_request) |
| 976 | |
| 977 | assert result.tools_called is True |
| 978 | assert len(result.tool_calls) == 1 |
| 979 | args = json.loads(result.tool_calls[0].function.arguments) |
| 980 | assert args == {"location": "Beijing"} |
| 981 | |
| 982 | |
| 983 | # ── Streaming wrapper consistency ───────────────────────────────────── |
| 984 | |
| 985 | |
| 986 | class TestStreamingWrapperConsistency: |
| 987 | """Streamed arg deltas must stay consistent with final extraction |
| 988 | when wrapper params like 'arguments' are unwrapped.""" |
| 989 | |
| 990 | def test_streaming_wrapper_unwrap_consistency(self, mock_tokenizer, mock_request): |
| 991 | tool = _make_tool("get_weather", {"location": {"type": "string"}}) |
| 992 | tools = [tool] |
| 993 | parser = DeepSeekV4Parser(mock_tokenizer, tools=tools) |
| 994 | mock_request.tools = tools |
| 995 | |
| 996 | chunks = [ |
| 997 | DSML_TOOL_START, |
| 998 | _invoke( |
| 999 | "get_weather", |
| 1000 | ("arguments", "false", '{"location": "NYC"}'), |
| 1001 | ), |
| 1002 | DSML_TOOL_END, |
| 1003 | ] |
| 1004 | |
| 1005 | results = simulate_tool_streaming(parser, mock_request, chunks) |
| 1006 | streamed_args = collect_tool_arguments(results) |
| 1007 | |
| 1008 | final_delta, _ = results[-1] |
| 1009 | finish_delta = parser.finish_streaming() |
| 1010 | extracted = parser._build_extracted_result(final_delta, finish_delta) |
| 1011 | |
| 1012 | assert extracted.tools_called is True |
| 1013 | assert len(extracted.tool_calls) == 1 |
| 1014 | |
| 1015 | final_args = extracted.tool_calls[0].function.arguments |
| 1016 | assert json.loads(final_args) == {"location": "NYC"} |
| 1017 | |
| 1018 | assert '"arguments"' not in streamed_args, ( |
| 1019 | f"Streamed args should not contain wrapper key, got: {streamed_args!r}" |
| 1020 | ) |
| 1021 | |
| 1022 | assert final_args.startswith(streamed_args), ( |
| 1023 | f"Extracted args {final_args!r} " |
| 1024 | f"should start with streamed args {streamed_args!r}" |
| 1025 | ) |
| 1026 | |
| 1027 | |
| 1028 | # ── DelegatingParser: large delta with </think> + tool calls ───────── |
| 1029 | |
| 1030 | _DSV4_FULL_VOCAB = { |
| 1031 | DSML_THINK_START: 128821, |
| 1032 | DSML_THINK_END: 128822, |
| 1033 | DSML_TOOL_START: 128823, |
| 1034 | DSML_TOOL_END: 128824, |
| 1035 | } |
| 1036 | |
| 1037 | |
| 1038 | class _DeepSeekV4Delegating(DelegatingParser): |
| 1039 | reasoning_parser_cls = DeepSeekV4ParserReasoningAdapter |
| 1040 | tool_parser_cls = DeepSeekV4ParserToolAdapter |
| 1041 | |
| 1042 | |
| 1043 | def _dsv4_tokens( |
| 1044 | reasoning: str, |
| 1045 | tool_name: str, |
| 1046 | params: list[tuple[str, str, str]], |
| 1047 | ) -> list[tuple[int, str]]: |
| 1048 | """Build a token sequence: reasoning + </think> + DSML tool block.""" |
| 1049 | tokens: list[tuple[int, str]] = [] |
| 1050 | tid = 100 |
| 1051 | |
| 1052 | for word in reasoning.split(" "): |
| 1053 | prefix = " " if tokens else "" |
| 1054 | tokens.append((tid, prefix + word)) |
| 1055 | tid += 1 |
| 1056 | |
| 1057 | tokens.append((_DSV4_FULL_VOCAB[DSML_THINK_END], DSML_THINK_END)) |
| 1058 | |
| 1059 | tokens.append((tid, "\n\n")) |
| 1060 | tid += 1 |
| 1061 | |
| 1062 | tokens.append((_DSV4_FULL_VOCAB[DSML_TOOL_START], DSML_TOOL_START)) |
| 1063 | |
| 1064 | tokens.append((tid, "\n")) |
| 1065 | tid += 1 |
| 1066 | |
| 1067 | invoke_prefix_text = f"{DSML_INVOKE_PREFIX}{tool_name}{DSML_INVOKE_NAME_END}" |
| 1068 | tokens.append((tid, invoke_prefix_text)) |
| 1069 | tid += 1 |
| 1070 | |
| 1071 | tokens.append((tid, "\n")) |
| 1072 | tid += 1 |
| 1073 | |
| 1074 | for name, is_str, value in params: |
| 1075 | param_text = _param(name, is_str, value) |
| 1076 | tokens.append((tid, param_text)) |
| 1077 | tid += 1 |
| 1078 | tokens.append((tid, "\n")) |
| 1079 | tid += 1 |
| 1080 | |
| 1081 | tokens.append((tid, DSML_INVOKE_END)) |
| 1082 | tid += 1 |
| 1083 | |
| 1084 | tokens.append((tid, "\n")) |
| 1085 | tid += 1 |
| 1086 | |
| 1087 | tokens.append((_DSV4_FULL_VOCAB[DSML_TOOL_END], DSML_TOOL_END)) |
| 1088 | |
| 1089 | return tokens |
| 1090 | |
| 1091 | |
| 1092 | class TestDelegatingParserLargeDelta: |
| 1093 | """Regression: tool calls lost when </think> + DSML arrive in same delta. |
| 1094 | |
| 1095 | The DelegatingParser used by the serving layer splits reasoning and |
| 1096 | tool parsing across two separate engine instances. When </think> and |
| 1097 | the entire DSML tool block arrive in a single large streaming delta, |
| 1098 | the content transfer from reasoning adapter to tool adapter must |
| 1099 | preserve the tool call text. |
| 1100 | """ |
| 1101 | |
| 1102 | @pytest.fixture |
| 1103 | def dsv4_tokens(self): |
| 1104 | return _dsv4_tokens( |
| 1105 | reasoning="The user wants the current weather in Berlin.", |
| 1106 | tool_name="get_weather", |
| 1107 | params=[ |
| 1108 | ("location", "true", "Berlin"), |
| 1109 | ("units", "true", "celsius"), |
| 1110 | ], |
| 1111 | ) |
| 1112 | |
| 1113 | @pytest.fixture |
| 1114 | def dsv4_tokenizer(self, dsv4_tokens): |
| 1115 | return MockTokenizer( |
| 1116 | vocab=dict(_DSV4_FULL_VOCAB), |
| 1117 | tokens=dsv4_tokens, |
| 1118 | ) |
| 1119 | |
| 1120 | @pytest.mark.parametrize( |
| 1121 | "chunk_size", |
| 1122 | [1, 2, 3, 5, None], |
| 1123 | ids=lambda c: f"chunk={c}", |
| 1124 | ) |
| 1125 | def test_tool_calls_extracted_at_all_chunk_sizes( |
| 1126 | self, dsv4_tokenizer, dsv4_tokens, chunk_size |
| 1127 | ): |
| 1128 | parser = _DeepSeekV4Delegating( |
| 1129 | dsv4_tokenizer, |
| 1130 | chat_template_kwargs={"thinking": True}, |
| 1131 | ) |
| 1132 | deltas = replay_streaming( |
| 1133 | parser, |
| 1134 | dsv4_tokens, |
| 1135 | chunk_size=chunk_size, |
| 1136 | finished_on_last=True, |
| 1137 | tools=DUMMY_TOOLS, |
| 1138 | ) |
| 1139 | output = collect_output(deltas) |
| 1140 | |
| 1141 | assert "The user wants" in output.reasoning |
| 1142 | assert len(output.tool_calls) == 1, ( |
| 1143 | f"Expected 1 tool call but got {len(output.tool_calls)}; " |
| 1144 | f"reasoning={output.reasoning!r}, content={output.content!r}" |
| 1145 | ) |
| 1146 | assert output.tool_calls[0]["name"] == "get_weather" |
| 1147 | args = json.loads(output.tool_calls[0]["arguments"]) |
| 1148 | assert args == {"location": "Berlin", "units": "celsius"} |
| 1149 | |
| 1150 | def test_default_thinking_extracts_tool_call_without_think_end(self, dsv4_tokens): |
| 1151 | tokens = [ |
| 1152 | token |
| 1153 | for token in dsv4_tokens |
| 1154 | if token[0] != _DSV4_FULL_VOCAB[DSML_THINK_END] |
| 1155 | ] |
| 1156 | tokenizer = MockTokenizer( |
| 1157 | vocab=dict(_DSV4_FULL_VOCAB), |
| 1158 | tokens=tokens, |
| 1159 | ) |
| 1160 | parser = _DeepSeekV4Delegating(tokenizer) |
| 1161 | |
| 1162 | deltas = replay_streaming( |
| 1163 | parser, |
| 1164 | tokens, |
| 1165 | chunk_size=1, |
| 1166 | finished_on_last=True, |
| 1167 | tools=DUMMY_TOOLS, |
| 1168 | prompt_token_ids=[_DSV4_FULL_VOCAB[DSML_THINK_START]], |
| 1169 | ) |
| 1170 | output = collect_output(deltas) |
| 1171 | |
| 1172 | assert "The user wants" in output.reasoning |
| 1173 | assert output.content == "" |
| 1174 | assert len(output.tool_calls) == 1 |
| 1175 | assert output.tool_calls[0]["name"] == "get_weather" |
| 1176 | args = json.loads(output.tool_calls[0]["arguments"]) |
| 1177 | assert args == {"location": "Berlin", "units": "celsius"} |
| 1178 | |
| 1179 | def test_eos_drop_token_does_not_swallow_tool_calls(self): |
| 1180 | """Tool calls must survive when an EOS DROP token's ID is in |
| 1181 | delta_token_ids but its text is absent from delta_text. |
| 1182 | |
| 1183 | At large stream_interval the EOS token ID arrives in the same |
| 1184 | delta as </think> + tool calls but the detokenizer strips the |
| 1185 | EOS text. The scanner's _rebuild_from_anchors defers all text |
| 1186 | after </think> when it can't find the EOS anchor text. The |
| 1187 | reasoning adapter's finish_streaming must flush deferred text |
| 1188 | as content (with skip_tool_parsing), not as tool calls. |
| 1189 | """ |
| 1190 | eos_text = "<｜end▁of▁sentence｜>" |
| 1191 | eos_id = 128801 |
| 1192 | vocab = { |
| 1193 | DSML_THINK_START: 128821, |
| 1194 | DSML_THINK_END: 128822, |
| 1195 | eos_text: eos_id, |
| 1196 | } |
| 1197 | |
| 1198 | reasoning = "The user wants weather." |
| 1199 | tool_block = ( |
| 1200 | "\n\n" |
| 1201 | + DSML_TOOL_START |
| 1202 | + "\n" |
| 1203 | + DSML_INVOKE_PREFIX |
| 1204 | + "get_weather" |
| 1205 | + DSML_INVOKE_NAME_END |
| 1206 | + "\n" |
| 1207 | + _param("location", "true", "Berlin") |
| 1208 | + "\n" |
| 1209 | + DSML_INVOKE_END |
| 1210 | + "\n" |
| 1211 | + DSML_TOOL_END |
| 1212 | ) |
| 1213 | # delta_text does NOT include EOS text (detokenizer strips it) |
| 1214 | full_text = reasoning + DSML_THINK_END + tool_block |
| 1215 | # Build token list: word-split reasoning, then special tokens, |
| 1216 | # then word-split tool block content, then EOS. |
| 1217 | # EOS ID is present but its text is NOT in delta_text. |
| 1218 | tokens: list[tuple[int, str]] = [] |
| 1219 | tid = 100 |
| 1220 | for word in reasoning.split(" "): |
| 1221 | pfx = " " if tokens else "" |
| 1222 | tokens.append((tid, pfx + word)) |
| 1223 | tid += 1 |
| 1224 | tokens.append((128822, DSML_THINK_END)) |
| 1225 | for ch in tool_block: |
| 1226 | tokens.append((tid, ch)) |
| 1227 | tid += 1 |
| 1228 | tokens.append((eos_id, eos_text)) |
| 1229 | |
| 1230 | all_ids = [t[0] for t in tokens] |
| 1231 | tokenizer = MockTokenizer(vocab=vocab, tokens=tokens) |
| 1232 | request = _test_request(tools=DUMMY_TOOLS) |
| 1233 | |
| 1234 | # All-in-one delta: EOS ID in token_ids but text NOT in |
| 1235 | # delta_text (detokenizer strips EOS). This is the scenario |
| 1236 | # at large stream_interval. |
| 1237 | parser = _DeepSeekV4Delegating( |
| 1238 | tokenizer, |
| 1239 | chat_template_kwargs={"thinking": True}, |
| 1240 | ) |
| 1241 | deltas = [ |
| 1242 | parser.parse_delta( |
| 1243 | full_text, |
| 1244 | all_ids, |
| 1245 | request, |
| 1246 | prompt_token_ids=[], |
| 1247 | finished=True, |
| 1248 | ) |
| 1249 | ] |
| 1250 | |
| 1251 | output = collect_output(deltas) |
| 1252 | |
| 1253 | assert "The user wants" in output.reasoning |
| 1254 | assert len(output.tool_calls) == 1, ( |
| 1255 | f"Expected 1 tool call but got {len(output.tool_calls)}; " |
| 1256 | f"reasoning={output.reasoning!r}, content={output.content!r}" |
| 1257 | ) |
| 1258 | assert output.tool_calls[0]["name"] == "get_weather" |
| 1259 | args = json.loads(output.tool_calls[0]["arguments"]) |
| 1260 | assert args == {"location": "Berlin"} |
| 1261 | |
| 1262 | @pytest.mark.parametrize( |
| 1263 | "chunk_size", |
| 1264 | [1, 2, 3, 5, None], |
| 1265 | ids=lambda c: f"chunk={c}", |
| 1266 | ) |
| 1267 | def test_eos_not_leaked_when_reasoning_never_ends(self, chunk_size): |
| 1268 | """EOS must not leak into reasoning_content when the model never |
| 1269 | emits </think> (generation ends while still in REASONING state).""" |
| 1270 | eos_text = "<｜end▁of▁sentence｜>" |
| 1271 | eos_id = 128801 |
| 1272 | vocab = { |
| 1273 | **_DSV4_FULL_VOCAB, |
| 1274 | eos_text: eos_id, |
| 1275 | } |
| 1276 | |
| 1277 | reasoning_text = "Good morning! How can I help you today?" |
| 1278 | tokens: list[tuple[int, str]] = [] |
| 1279 | tid = 100 |
| 1280 | for word in reasoning_text.split(" "): |
| 1281 | prefix = " " if tokens else "" |
| 1282 | tokens.append((tid, prefix + word)) |
| 1283 | tid += 1 |
| 1284 | tokens.append((eos_id, eos_text)) |
| 1285 | |
| 1286 | tokenizer = MockTokenizer(vocab=vocab, tokens=tokens) |
| 1287 | parser = _DeepSeekV4Delegating( |
| 1288 | tokenizer, |
| 1289 | chat_template_kwargs={"thinking": True}, |
| 1290 | ) |
| 1291 | deltas = replay_streaming( |
| 1292 | parser, |
| 1293 | tokens, |
| 1294 | chunk_size=chunk_size, |
| 1295 | finished_on_last=True, |
| 1296 | ) |
| 1297 | output = collect_output(deltas) |
| 1298 | |
| 1299 | assert reasoning_text in output.reasoning |
| 1300 | assert eos_text not in output.reasoning |
| 1301 | assert output.content == "" |
| 1302 | assert output.tool_calls == [] |
| 1303 | |
| 1304 | |
| 1305 | # ── Malformed DSML markup must not leak into content (#51914) ───────── |
| 1306 | |
| 1307 | _WRAPPER_VARIANTS = list(DSML_TOOL_START_VARIANTS) |
| 1308 | |
| 1309 | |
| 1310 | class TestMalformedDsmlNoise: |
| 1311 | """DeepSeek-V4-Flash intermittently misspells the tool-call opener, e.g. |
| 1312 | ``<｜DSML｜toolcalls>``. Known corrupted openers open a tool block like |
| 1313 | the real wrapper so the markup never surfaces as assistant content. |
| 1314 | """ |
| 1315 | |
| 1316 | _INNER = ( |
| 1317 | f"{DSML_INVOKE_PREFIX}get_weather{DSML_INVOKE_NAME_END}\n" |
| 1318 | f"{_param('city', 'true', 'Seoul')}\n" |
| 1319 | f"{DSML_INVOKE_END}\n" |
| 1320 | f"{DSML_TOOL_END}" |
| 1321 | ) |
| 1322 | |
| 1323 | @staticmethod |
| 1324 | def _parser(mock_tokenizer, thinking: bool = False): |
| 1325 | return DeepSeekV4Parser( |
| 1326 | mock_tokenizer, chat_template_kwargs={"thinking": thinking} |
| 1327 | ) |
| 1328 | |
| 1329 | @pytest.mark.parametrize("wrapper", _WRAPPER_VARIANTS) |
| 1330 | @pytest.mark.parametrize("preamble", ["", "Let me check.\n"]) |
| 1331 | def test_corrupted_wrapper_behaves_like_real_one( |
| 1332 | self, mock_tokenizer, mock_request, wrapper, preamble |
| 1333 | ): |
| 1334 | text = f"{preamble}{wrapper}\n{self._INNER}" |
| 1335 | reference = f"{preamble}{DSML_TOOL_START}\n{self._INNER}" |
| 1336 | |
| 1337 | result = self._parser(mock_tokenizer).extract_tool_calls(text, mock_request) |
| 1338 | expected = self._parser(mock_tokenizer).extract_tool_calls( |
| 1339 | reference, mock_request |
| 1340 | ) |
| 1341 | assert [tc.function.name for tc in result.tool_calls] == ["get_weather"] |
| 1342 | assert json.loads(result.tool_calls[0].function.arguments) == {"city": "Seoul"} |
| 1343 | assert result.content == expected.content |
| 1344 | assert "DSML" not in (result.content or "") |
| 1345 | |
| 1346 | results = simulate_tool_streaming( |
| 1347 | self._parser(mock_tokenizer), mock_request, list(text) |
| 1348 | ) |
| 1349 | assert collect_function_name(results) == "get_weather" |
| 1350 | assert json.loads(collect_tool_arguments(results)) == {"city": "Seoul"} |
| 1351 | assert collect_content(results) == (expected.content or "") |
| 1352 | |
| 1353 | @pytest.mark.parametrize("wrapper", _WRAPPER_VARIANTS) |
| 1354 | def test_corrupted_wrapper_ends_reasoning( |
| 1355 | self, mock_tokenizer, mock_request, wrapper |
| 1356 | ): |
| 1357 | """Like the real wrapper, a corrupted one inside ``<think>`` closes |
| 1358 | the reasoning block and starts the tool call.""" |
| 1359 | text = f"thinking{wrapper}\n{self._INNER}" |
| 1360 | parser = self._parser(mock_tokenizer, thinking=True) |
| 1361 | reasoning, content = parser.extract_reasoning(text, mock_request) |
| 1362 | assert reasoning == "thinking" |
| 1363 | assert "DSML" not in (content or "") |
| 1364 | |
| 1365 | result = self._parser(mock_tokenizer, thinking=True).extract_tool_calls( |
| 1366 | text, mock_request |
| 1367 | ) |
| 1368 | assert [tc.function.name for tc in result.tool_calls] == ["get_weather"] |
| 1369 | |
| 1370 | @pytest.mark.parametrize( |
| 1371 | "text", |
| 1372 | [ |
| 1373 | # Fullwidth bars, DSML namespace, but not a registered spelling. |
| 1374 | "see <｜DSML｜foo> and <｜DSML｜tool_call> here", |
| 1375 | # Unclosed prefix. |
| 1376 | "see <｜DSML｜ here\nnext line", |
| 1377 | # ASCII pipes are not the DSML sigil. |
| 1378 | "see <|DSML|tool> and <|DSML|toolcalls> there", |
| 1379 | ], |
| 1380 | ) |
| 1381 | def test_unregistered_markup_is_ordinary_content( |
| 1382 | self, mock_tokenizer, mock_request, text |
| 1383 | ): |
| 1384 | """Only the registered spellings are treated as wrappers.""" |
| 1385 | result = self._parser(mock_tokenizer).extract_tool_calls(text, mock_request) |
| 1386 | assert result.tools_called is False |
| 1387 | assert result.content == text |
| 1388 | |
| 1389 | results = simulate_tool_streaming( |
| 1390 | self._parser(mock_tokenizer), mock_request, list(text) |
| 1391 | ) |
| 1392 | assert collect_content(results) == text |
| 1393 | |
| 1394 | @pytest.mark.parametrize("wrapper", _WRAPPER_VARIANTS) |
| 1395 | def test_prose_mention_matches_canonical_wrapper( |
| 1396 | self, mock_tokenizer, mock_request, wrapper |
| 1397 | ): |
| 1398 | """A registered spelling mentioned in prose behaves exactly like a |
| 1399 | prose mention of the real wrapper: the opener starts a tool block and |
| 1400 | the rest of the message is dropped. That is pre-existing behavior of |
| 1401 | ``<｜DSML｜tool_calls>``; the variants must not differ from it either |
| 1402 | way.""" |
| 1403 | template = "The opener looks like {} and then params follow. Done." |
| 1404 | text = template.format(wrapper) |
| 1405 | reference = template.format(DSML_TOOL_START) |
| 1406 | |
| 1407 | result = self._parser(mock_tokenizer).extract_tool_calls(text, mock_request) |
| 1408 | expected = self._parser(mock_tokenizer).extract_tool_calls( |
| 1409 | reference, mock_request |
| 1410 | ) |
| 1411 | assert result.tools_called == expected.tools_called |
| 1412 | assert result.content == expected.content |
| 1413 | |
| 1414 | results = simulate_tool_streaming( |
| 1415 | self._parser(mock_tokenizer), mock_request, list(text) |
| 1416 | ) |
| 1417 | expected_results = simulate_tool_streaming( |
| 1418 | self._parser(mock_tokenizer), mock_request, list(reference) |
| 1419 | ) |
| 1420 | assert collect_content(results) == collect_content(expected_results) |
| 1421 | |
| 1422 | @pytest.mark.parametrize("wrapper", _WRAPPER_VARIANTS) |
| 1423 | @pytest.mark.parametrize("chunk_size", [1, 3, None], ids=lambda c: f"chunk={c}") |
| 1424 | def test_delegating_parser_corrupted_wrapper(self, wrapper, chunk_size): |
| 1425 | """The serving-layer shape: reasoning and tool adapters on separate |
| 1426 | engines, with the corrupted opener arriving as plain text.""" |
| 1427 | tokens = _dsv4_tokens( |
| 1428 | reasoning="Checking the weather.", |
| 1429 | tool_name="get_weather", |
| 1430 | params=[("city", "true", "Seoul")], |
| 1431 | ) |
| 1432 | start_id = _DSV4_FULL_VOCAB[DSML_TOOL_START] |
| 1433 | tokens = [ |
| 1434 | (900, wrapper) if tid == start_id else (tid, text) for tid, text in tokens |
| 1435 | ] |
| 1436 | # Only the think markers are single tokens in the real |
| 1437 | # DeepSeek-V4-Flash vocab; the DSML wrappers are ordinary text and |
| 1438 | # must be matched by the lexer. Mirror that here. |
| 1439 | vocab = { |
| 1440 | DSML_THINK_START: _DSV4_FULL_VOCAB[DSML_THINK_START], |
| 1441 | DSML_THINK_END: _DSV4_FULL_VOCAB[DSML_THINK_END], |
| 1442 | } |
| 1443 | tokenizer = MockTokenizer(vocab=vocab, tokens=tokens) |
| 1444 | parser = _DeepSeekV4Delegating( |
| 1445 | tokenizer, chat_template_kwargs={"thinking": True} |
| 1446 | ) |
| 1447 | |
| 1448 | deltas = replay_streaming( |
| 1449 | parser, |
| 1450 | tokens, |
| 1451 | chunk_size=chunk_size, |
| 1452 | finished_on_last=True, |
| 1453 | tools=DUMMY_TOOLS, |
| 1454 | ) |
| 1455 | output = collect_output(deltas) |
| 1456 | |
| 1457 | assert "Checking" in output.reasoning |
| 1458 | assert len(output.tool_calls) == 1, ( |
| 1459 | f"reasoning={output.reasoning!r}, content={output.content!r}" |
| 1460 | ) |
| 1461 | assert output.tool_calls[0]["name"] == "get_weather" |
| 1462 | assert json.loads(output.tool_calls[0]["arguments"]) == {"city": "Seoul"} |
| 1463 | assert "DSML" not in (output.content or "") |