| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import json |
| 5 | from pathlib import Path |
| 6 | from types import SimpleNamespace |
| 7 | |
| 8 | import pytest |
| 9 | import torch |
| 10 | |
| 11 | from vllm.entrypoints.chat_utils import parse_chat_messages |
| 12 | from vllm.renderers.registry import RENDERER_REGISTRY |
| 13 | from vllm.tokenizers.deepseek_v4 import get_deepseek_v4_tokenizer |
| 14 | from vllm.tokenizers.registry import TokenizerRegistry |
| 15 | |
| 16 | FIXTURES_DIR = Path(__file__).parent / "fixtures" / "deepseek_v4" |
| 17 | |
| 18 | |
| 19 | class FakeHfTokenizer: |
| 20 | vocab_size = 100 |
| 21 | |
| 22 | def get_added_vocab(self) -> dict[str, int]: |
| 23 | return {"</think>": 100} |
| 24 | |
| 25 | def encode( |
| 26 | self, |
| 27 | text: str, |
| 28 | add_special_tokens: bool = False, |
| 29 | **kwargs, |
| 30 | ) -> list[int]: |
| 31 | self.last_encode = (text, add_special_tokens, kwargs) |
| 32 | return [len(text)] |
| 33 | |
| 34 | |
| 35 | def _tokenizer(): |
| 36 | return get_deepseek_v4_tokenizer(FakeHfTokenizer()) |
| 37 | |
| 38 | |
| 39 | def _model_config(): |
| 40 | return SimpleNamespace( |
| 41 | multimodal_config=None, |
| 42 | allowed_local_media_path="", |
| 43 | allowed_media_domains=None, |
| 44 | enable_prompt_embeds=False, |
| 45 | ) |
| 46 | |
| 47 | |
| 48 | def _load_reference_case(case_id: int): |
| 49 | data = json.loads((FIXTURES_DIR / f"test_input_{case_id}.json").read_text()) |
| 50 | if isinstance(data, dict): |
| 51 | return data["messages"], data.get("tools") |
| 52 | return data, None |
| 53 | |
| 54 | |
| 55 | def _render_reference_case(case_id: int, **kwargs): |
| 56 | messages, tools = _load_reference_case(case_id) |
| 57 | # Preserve content blocks for encoding without fetching fixture media. |
| 58 | return _tokenizer().apply_chat_template( |
| 59 | conversation=messages, |
| 60 | messages=messages, |
| 61 | tools=tools, |
| 62 | tokenize=False, |
| 63 | **kwargs, |
| 64 | ) |
| 65 | |
| 66 | |
| 67 | def test_deepseek_v4_tokenizer_registered(): |
| 68 | assert TokenizerRegistry.load_tokenizer_cls("deepseek_v4").__name__ == ( |
| 69 | "DeepseekV4Tokenizer" |
| 70 | ) |
| 71 | assert RENDERER_REGISTRY.load_renderer_cls("deepseek_v4").__name__ == ( |
| 72 | "DeepseekV4Renderer" |
| 73 | ) |
| 74 | |
| 75 | |
| 76 | def test_deepseek_v4_defaults_to_thinking_with_high_effort(): |
| 77 | prompt = _tokenizer().apply_chat_template( |
| 78 | [{"role": "user", "content": "Hello"}], |
| 79 | tokenize=False, |
| 80 | ) |
| 81 | |
| 82 | assert prompt.startswith( |
| 83 | "<｜begin▁of▁sentence｜>Reasoning Effort: Absolute maximum" |
| 84 | ) |
| 85 | assert prompt.endswith("<｜Assistant｜><think>") |
| 86 | |
| 87 | |
| 88 | @pytest.mark.parametrize("kwargs", [{"thinking": True}, {"enable_thinking": True}]) |
| 89 | def test_deepseek_v4_enables_thinking_with_compatible_kwargs(kwargs): |
| 90 | prompt = _tokenizer().apply_chat_template( |
| 91 | [{"role": "user", "content": "Hello"}], |
| 92 | tokenize=False, |
| 93 | **kwargs, |
| 94 | ) |
| 95 | |
| 96 | assert prompt.startswith( |
| 97 | "<｜begin▁of▁sentence｜>Reasoning Effort: Absolute maximum" |
| 98 | ) |
| 99 | assert prompt.endswith("<｜Assistant｜><think>") |
| 100 | |
| 101 | |
| 102 | @pytest.mark.parametrize("kwargs", [{"thinking": False}, {"enable_thinking": False}]) |
| 103 | def test_deepseek_v4_explicitly_disables_thinking(kwargs): |
| 104 | prompt = _tokenizer().apply_chat_template( |
| 105 | [{"role": "user", "content": "Hello"}], |
| 106 | tokenize=False, |
| 107 | **kwargs, |
| 108 | ) |
| 109 | |
| 110 | assert prompt == ("<｜begin▁of▁sentence｜><｜User｜>Hello<｜Assistant｜></think>") |
| 111 | |
| 112 | |
| 113 | @pytest.mark.parametrize( |
| 114 | ("enable_thinking", "thinking_token"), |
| 115 | [(False, "</think>"), (True, "<think>")], |
| 116 | ) |
| 117 | def test_deepseek_v4_appends_assistant_transition_after_trailing_system( |
| 118 | enable_thinking, thinking_token |
| 119 | ): |
| 120 | prompt = _tokenizer().apply_chat_template( |
| 121 | [ |
| 122 | {"role": "system", "content": "You are helpful."}, |
| 123 | {"role": "user", "content": "What is 2+2?"}, |
| 124 | {"role": "system", "content": "Be concise."}, |
| 125 | ], |
| 126 | tokenize=False, |
| 127 | enable_thinking=enable_thinking, |
| 128 | reasoning_effort="low", |
| 129 | ) |
| 130 | |
| 131 | assert prompt == ( |
| 132 | "<｜begin▁of▁sentence｜>You are helpful." |
| 133 | f"<｜User｜>What is 2+2?Be concise.<｜Assistant｜>{thinking_token}" |
| 134 | ) |
| 135 | |
| 136 | |
| 137 | def test_deepseek_v4_does_not_transition_after_mid_conversation_system(): |
| 138 | prompt = _tokenizer().apply_chat_template( |
| 139 | [ |
| 140 | {"role": "user", "content": "What is 3+3?"}, |
| 141 | {"role": "assistant", "content": "6"}, |
| 142 | {"role": "system", "content": "Answer in French."}, |
| 143 | {"role": "user", "content": "What is 5+5?"}, |
| 144 | ], |
| 145 | tokenize=False, |
| 146 | enable_thinking=False, |
| 147 | ) |
| 148 | |
| 149 | assert prompt == ( |
| 150 | "<｜begin▁of▁sentence｜><｜User｜>What is 3+3?" |
| 151 | "<｜Assistant｜></think>6<｜end▁of▁sentence｜>" |
| 152 | "Answer in French.<｜User｜>What is 5+5?" |
| 153 | "<｜Assistant｜></think>" |
| 154 | ) |
| 155 | |
| 156 | |
| 157 | @pytest.mark.parametrize( |
| 158 | ("enable_thinking", "expected"), |
| 159 | [ |
| 160 | ( |
| 161 | False, |
| 162 | "<｜begin▁of▁sentence｜>rules<｜Assistant｜></think>answer" |
| 163 | "<｜end▁of▁sentence｜>", |
| 164 | ), |
| 165 | ( |
| 166 | True, |
| 167 | "<｜begin▁of▁sentence｜>rules<｜Assistant｜><think>reason</think>answer" |
| 168 | "<｜end▁of▁sentence｜>", |
| 169 | ), |
| 170 | ], |
| 171 | ) |
| 172 | def test_deepseek_v4_transitions_from_system_to_assistant(enable_thinking, expected): |
| 173 | prompt = _tokenizer().apply_chat_template( |
| 174 | [ |
| 175 | {"role": "system", "content": "rules"}, |
| 176 | { |
| 177 | "role": "assistant", |
| 178 | "reasoning": "reason", |
| 179 | "content": "answer", |
| 180 | }, |
| 181 | ], |
| 182 | tokenize=False, |
| 183 | enable_thinking=enable_thinking, |
| 184 | reasoning_effort="low", |
| 185 | ) |
| 186 | |
| 187 | assert prompt == expected |
| 188 | |
| 189 | |
| 190 | def test_deepseek_v4_unknown_role_raises_value_error(): |
| 191 | # Invalid roles are client errors: they must surface as ValueError |
| 192 | # (mapped to HTTP 400 by the OpenAI serving layer), not |
| 193 | # NotImplementedError (mapped to HTTP 501). |
| 194 | with pytest.raises(ValueError, match="Invalid role: SYSTEM"): |
| 195 | _tokenizer().apply_chat_template( |
| 196 | [{"role": "SYSTEM", "content": "Hello"}], |
| 197 | tokenize=False, |
| 198 | ) |
| 199 | |
| 200 | |
| 201 | def test_deepseek_v4_uses_v4_tool_prompt_from_request_tools(): |
| 202 | tools = [ |
| 203 | { |
| 204 | "type": "function", |
| 205 | "function": { |
| 206 | "name": "get_weather", |
| 207 | "description": "Get weather for a city", |
| 208 | "parameters": { |
| 209 | "type": "object", |
| 210 | "properties": {"city": {"type": "string"}}, |
| 211 | "required": ["city"], |
| 212 | }, |
| 213 | }, |
| 214 | } |
| 215 | ] |
| 216 | |
| 217 | prompt = _tokenizer().apply_chat_template( |
| 218 | [{"role": "user", "content": "Weather?"}], |
| 219 | tools=tools, |
| 220 | tokenize=False, |
| 221 | ) |
| 222 | |
| 223 | assert "## Tools" in prompt |
| 224 | assert "<｜DSML｜tool_calls>" in prompt |
| 225 | assert "</｜DSML｜tool_calls>" in prompt |
| 226 | assert "function_calls" not in prompt |
| 227 | assert '"name": "get_weather"' in prompt |
| 228 | assert prompt.startswith( |
| 229 | "<｜begin▁of▁sentence｜>Reasoning Effort: Absolute maximum" |
| 230 | ) |
| 231 | assert prompt.endswith("<｜User｜>Weather?<｜Assistant｜><think>") |
| 232 | |
| 233 | |
| 234 | def test_deepseek_v4_renders_parsed_history_tool_arguments(): |
| 235 | messages = [ |
| 236 | {"role": "user", "content": "List the repo"}, |
| 237 | { |
| 238 | "role": "assistant", |
| 239 | "tool_calls": [ |
| 240 | { |
| 241 | "id": "call_1", |
| 242 | "type": "function", |
| 243 | "function": { |
| 244 | "name": "str_replace_editor", |
| 245 | "arguments": '{"command": "view", "path": "/testbed"}', |
| 246 | }, |
| 247 | } |
| 248 | ], |
| 249 | }, |
| 250 | { |
| 251 | "role": "tool", |
| 252 | "tool_call_id": "call_1", |
| 253 | "content": "file list", |
| 254 | }, |
| 255 | ] |
| 256 | tools = [ |
| 257 | { |
| 258 | "type": "function", |
| 259 | "function": { |
| 260 | "name": "str_replace_editor", |
| 261 | "description": "Edit files", |
| 262 | "parameters": { |
| 263 | "type": "object", |
| 264 | "properties": { |
| 265 | "command": {"type": "string"}, |
| 266 | "path": {"type": "string"}, |
| 267 | }, |
| 268 | "required": ["command", "path"], |
| 269 | }, |
| 270 | }, |
| 271 | } |
| 272 | ] |
| 273 | conversation, _, _ = parse_chat_messages( |
| 274 | messages, |
| 275 | _model_config(), |
| 276 | content_format="string", |
| 277 | ) |
| 278 | |
| 279 | prompt = _tokenizer().apply_chat_template( |
| 280 | conversation=conversation, |
| 281 | messages=messages, |
| 282 | tools=tools, |
| 283 | tokenize=False, |
| 284 | ) |
| 285 | |
| 286 | assert '<｜DSML｜parameter name="command" string="true">view' in prompt |
| 287 | assert '<｜DSML｜parameter name="path" string="true">/testbed' in prompt |
| 288 | assert 'parameter name="arguments"' not in prompt |
| 289 | |
| 290 | |
| 291 | @pytest.mark.parametrize( |
| 292 | ("reasoning_effort", "expected_prefix"), |
| 293 | [ |
| 294 | ("low", "<｜begin▁of▁sentence｜><｜User｜>Hello"), |
| 295 | ("high", "<｜begin▁of▁sentence｜>Reasoning Effort: Absolute maximum"), |
| 296 | ("max", "<｜begin▁of▁sentence｜>Reasoning Effort: Beyond maximum"), |
| 297 | ], |
| 298 | ) |
| 299 | def test_deepseek_v4_renders_0731_reasoning_effort_prompts( |
| 300 | reasoning_effort, expected_prefix |
| 301 | ): |
| 302 | prompt = _tokenizer().apply_chat_template( |
| 303 | [{"role": "user", "content": "Hello"}], |
| 304 | tokenize=False, |
| 305 | enable_thinking=True, |
| 306 | reasoning_effort=reasoning_effort, |
| 307 | ) |
| 308 | |
| 309 | assert prompt.endswith("<｜Assistant｜><think>") |
| 310 | assert prompt.startswith(expected_prefix) |
| 311 | |
| 312 | |
| 313 | def test_deepseek_v4_none_reasoning_effort_disables_thinking(): |
| 314 | prompt = _tokenizer().apply_chat_template( |
| 315 | [{"role": "user", "content": "Hello"}], |
| 316 | tokenize=False, |
| 317 | enable_thinking=True, |
| 318 | reasoning_effort="none", |
| 319 | ) |
| 320 | |
| 321 | assert prompt == ("<｜begin▁of▁sentence｜><｜User｜>Hello<｜Assistant｜></think>") |
| 322 | |
| 323 | |
| 324 | @pytest.mark.parametrize( |
| 325 | ("reasoning_effort", "expected_mode", "expected_effort"), |
| 326 | [ |
| 327 | ("none", "chat", None), |
| 328 | ("minimal", "thinking", "low"), |
| 329 | ("low", "thinking", "low"), |
| 330 | ("medium", "thinking", "low"), |
| 331 | ("high", "thinking", "high"), |
| 332 | ("xhigh", "thinking", "high"), |
| 333 | ("max", "thinking", "max"), |
| 334 | ("unexpected", "thinking", "high"), |
| 335 | ], |
| 336 | ) |
| 337 | def test_deepseek_v4_maps_compatible_thinking_reasoning_effort_values( |
| 338 | monkeypatch: pytest.MonkeyPatch, |
| 339 | reasoning_effort, |
| 340 | expected_mode, |
| 341 | expected_effort, |
| 342 | ): |
| 343 | captured_kwargs = [] |
| 344 | |
| 345 | def fake_encode_messages(messages, **kwargs): |
| 346 | captured_kwargs.append(kwargs) |
| 347 | return "prompt" |
| 348 | |
| 349 | monkeypatch.setattr( |
| 350 | "vllm.tokenizers.deepseek_v4.encode_messages", |
| 351 | fake_encode_messages, |
| 352 | ) |
| 353 | |
| 354 | _tokenizer().apply_chat_template( |
| 355 | [{"role": "user", "content": "Hello"}], |
| 356 | tokenize=False, |
| 357 | enable_thinking=True, |
| 358 | reasoning_effort=reasoning_effort, |
| 359 | ) |
| 360 | |
| 361 | assert captured_kwargs[-1]["thinking_mode"] == expected_mode |
| 362 | assert captured_kwargs[-1]["reasoning_effort"] == expected_effort |
| 363 | |
| 364 | |
| 365 | def test_deepseek_v4_renders_0731_max_reasoning_effort(): |
| 366 | prompt = _tokenizer().apply_chat_template( |
| 367 | [{"role": "user", "content": "Hello"}], |
| 368 | tokenize=False, |
| 369 | enable_thinking=True, |
| 370 | reasoning_effort="max", |
| 371 | ) |
| 372 | |
| 373 | assert prompt.startswith("<｜begin▁of▁sentence｜>Reasoning Effort: Beyond maximum") |
| 374 | |
| 375 | |
| 376 | def test_deepseek_v4_maps_xhigh_to_high_reasoning_effort(): |
| 377 | prompt = _tokenizer().apply_chat_template( |
| 378 | [{"role": "user", "content": "Hello"}], |
| 379 | tokenize=False, |
| 380 | enable_thinking=True, |
| 381 | reasoning_effort="xhigh", |
| 382 | ) |
| 383 | |
| 384 | assert prompt.startswith( |
| 385 | "<｜begin▁of▁sentence｜>Reasoning Effort: Absolute maximum" |
| 386 | ) |
| 387 | |
| 388 | |
| 389 | @pytest.mark.parametrize( |
| 390 | ("case_id", "kwargs"), |
| 391 | [ |
| 392 | (1, {"thinking": True, "reasoning_effort": "low"}), |
| 393 | (2, {"thinking": True, "reasoning_effort": "low"}), |
| 394 | (3, {"thinking": True, "reasoning_effort": "low"}), |
| 395 | (4, {"thinking": False}), |
| 396 | (5, {"thinking": False}), |
| 397 | ], |
| 398 | ) |
| 399 | def test_deepseek_v4_matches_reference_golden_fixtures(case_id, kwargs): |
| 400 | prompt = _render_reference_case(case_id, **kwargs) |
| 401 | |
| 402 | expected = (FIXTURES_DIR / f"test_output_{case_id}.txt").read_text() |
| 403 | assert prompt == expected |
| 404 | |
| 405 | |
| 406 | def test_deepseek_v4_rejects_empty_developer_content(): |
| 407 | with pytest.raises(ValueError): |
| 408 | _tokenizer().apply_chat_template( |
| 409 | [{"role": "developer", "content": ""}], |
| 410 | tokenize=False, |
| 411 | enable_thinking=True, |
| 412 | reasoning_effort="low", |
| 413 | ) |
| 414 | |
| 415 | |
| 416 | @pytest.mark.parametrize( |
| 417 | "kwargs", |
| 418 | [ |
| 419 | {"thinking_mode": "bogus"}, |
| 420 | {"thinking_mode": "thinking", "reasoning_effort": "turbo"}, |
| 421 | ], |
| 422 | ) |
| 423 | def test_deepseek_v4_encode_messages_rejects_invalid_arguments(kwargs): |
| 424 | from vllm.tokenizers.deepseek_v4_encoding import encode_messages |
| 425 | |
| 426 | with pytest.raises(ValueError): |
| 427 | encode_messages([{"role": "user", "content": "Hello"}], **kwargs) |
| 428 | |
| 429 | |
| 430 | def test_deepseek_v4_image_blocks_become_placeholders(): |
| 431 | prompt = _tokenizer().apply_chat_template( |
| 432 | [ |
| 433 | { |
| 434 | "role": "user", |
| 435 | "content": [ |
| 436 | {"type": "text", "text": "first:"}, |
| 437 | {"type": "image_url", "image_url": {"url": "file:///a.png"}}, |
| 438 | {"type": "text", "text": "second:"}, |
| 439 | {"type": "image", "source": {"data": "AAAA"}}, |
| 440 | ], |
| 441 | } |
| 442 | ], |
| 443 | tokenize=False, |
| 444 | thinking=False, |
| 445 | ) |
| 446 | |
| 447 | assert ( |
| 448 | "<｜User｜>first:\n\n<｜deepseek_image｜>\n\n" |
| 449 | "second:\n\n<｜deepseek_image｜>" in prompt |
| 450 | ) |
| 451 | |
| 452 | |
| 453 | def test_deepseek_v4_image_sentinel_ids_match_tokenizer(): |
| 454 | """The borrowed sentinel ids must line up with the reserved |
| 455 | ``<|place_holder_mm_span_XXXX|>`` tokens in the tokenizer.""" |
| 456 | from vllm.models.deepseek_v4.common.mm_preprocess import ( |
| 457 | IMAGE_SENTINEL_BASE_ID, |
| 458 | IMAGE_SENTINEL_TOKEN_NAMES, |
| 459 | image_sentinel_mask, |
| 460 | validate_image_sentinel_ids, |
| 461 | ) |
| 462 | |
| 463 | class FakeTokenizer: |
| 464 | def __init__(self, offset: int = 0) -> None: |
| 465 | self.offset = offset |
| 466 | |
| 467 | def convert_tokens_to_ids(self, token: str) -> int: |
| 468 | return ( |
| 469 | IMAGE_SENTINEL_BASE_ID |
| 470 | + self.offset |
| 471 | + (IMAGE_SENTINEL_TOKEN_NAMES.index(token)) |
| 472 | ) |
| 473 | |
| 474 | validate_image_sentinel_ids(FakeTokenizer()) # no raise |
| 475 | with pytest.raises(ValueError, match="sentinel"): |
| 476 | validate_image_sentinel_ids(FakeTokenizer(offset=1)) |
| 477 | |
| 478 | ids = torch.tensor( |
| 479 | [ |
| 480 | 1, |
| 481 | IMAGE_SENTINEL_BASE_ID, |
| 482 | IMAGE_SENTINEL_BASE_ID + 4, |
| 483 | IMAGE_SENTINEL_BASE_ID + 5, |
| 484 | 129264, |
| 485 | ] |
| 486 | ) |
| 487 | assert image_sentinel_mask(ids).tolist() == [False, True, True, False, False] |
| 488 | |
| 489 | |
| 490 | def _request_tools(): |
| 491 | return [ |
| 492 | { |
| 493 | "type": "function", |
| 494 | "function": { |
| 495 | "name": "get_weather", |
| 496 | "description": "Get weather for a city", |
| 497 | "parameters": { |
| 498 | "type": "object", |
| 499 | "properties": {"city": {"type": "string"}}, |
| 500 | "required": ["city"], |
| 501 | }, |
| 502 | }, |
| 503 | } |
| 504 | ] |
| 505 | |
| 506 | |
| 507 | def _encode_reference(messages): |
| 508 | from vllm.tokenizers.deepseek_v4_encoding import encode_messages |
| 509 | |
| 510 | return encode_messages(messages, thinking_mode="thinking", reasoning_effort="low") |
| 511 | |
| 512 | |
| 513 | def test_deepseek_v4_attaches_request_tools_to_existing_system_message(): |
| 514 | messages = [ |
| 515 | {"role": "system", "content": "You are helpful."}, |
| 516 | {"role": "user", "content": "Weather in Paris?"}, |
| 517 | ] |
| 518 | |
| 519 | prompt = _tokenizer().apply_chat_template( |
| 520 | messages, |
| 521 | tools=_request_tools(), |
| 522 | tokenize=False, |
| 523 | thinking=True, |
| 524 | reasoning_effort="low", |
| 525 | ) |
| 526 | |
| 527 | expected = _encode_reference( |
| 528 | [{**messages[0], "tools": _request_tools()}, messages[1]] |
| 529 | ) |
| 530 | assert prompt == expected |
| 531 | assert prompt.startswith("<｜begin▁of▁sentence｜>You are helpful.\n\n## Tools") |
| 532 | assert prompt.count("## Tools") == 1 |
| 533 | |
| 534 | |
| 535 | def test_deepseek_v4_synthetic_system_only_when_no_system_message(): |
| 536 | """Without a system message, request tools still get a synthetic leading |
| 537 | system entry.""" |
| 538 | messages = [{"role": "user", "content": "Weather in Paris?"}] |
| 539 | |
| 540 | prompt = _tokenizer().apply_chat_template( |
| 541 | messages, |
| 542 | tools=_request_tools(), |
| 543 | tokenize=False, |
| 544 | thinking=True, |
| 545 | reasoning_effort="low", |
| 546 | ) |
| 547 | |
| 548 | expected = _encode_reference( |
| 549 | [{"role": "system", "tools": _request_tools()}, messages[0]] |
| 550 | ) |
| 551 | assert prompt == expected |
| 552 | assert prompt.startswith("<｜begin▁of▁sentence｜>\n\n## Tools") |
| 553 | |
| 554 | |
| 555 | def test_deepseek_v4_request_tools_override_system_message_tools(): |
| 556 | message_tools = [ |
| 557 | { |
| 558 | "type": "function", |
| 559 | "function": { |
| 560 | "name": "message_level_tool", |
| 561 | "description": "Should not be rendered", |
| 562 | "parameters": {"type": "object", "properties": {}}, |
| 563 | }, |
| 564 | } |
| 565 | ] |
| 566 | system_message = { |
| 567 | "role": "system", |
| 568 | "content": "You are helpful.", |
| 569 | "tools": message_tools, |
| 570 | } |
| 571 | user_message = {"role": "user", "content": "Weather in Paris?"} |
| 572 | |
| 573 | prompt = _tokenizer().apply_chat_template( |
| 574 | [system_message, user_message], |
| 575 | tools=_request_tools(), |
| 576 | tokenize=False, |
| 577 | thinking=True, |
| 578 | reasoning_effort="low", |
| 579 | ) |
| 580 | |
| 581 | expected = _encode_reference( |
| 582 | [{**system_message, "tools": _request_tools()}, user_message] |
| 583 | ) |
| 584 | assert prompt == expected |
| 585 | assert '"name": "get_weather"' in prompt |
| 586 | assert "message_level_tool" not in prompt |
| 587 | |
| 588 | |
| 589 | def test_deepseek_v4_attaches_request_tools_to_first_system_message_only(): |
| 590 | """Only the first system message, at any index, gets the request tools.""" |
| 591 | messages = [ |
| 592 | {"role": "user", "content": "Hi"}, |
| 593 | {"role": "system", "content": "First system."}, |
| 594 | {"role": "system", "content": "Second system."}, |
| 595 | {"role": "user", "content": "Weather in Paris?"}, |
| 596 | ] |
| 597 | |
| 598 | prompt = _tokenizer().apply_chat_template( |
| 599 | messages, |
| 600 | tools=_request_tools(), |
| 601 | tokenize=False, |
| 602 | thinking=True, |
| 603 | reasoning_effort="low", |
| 604 | ) |
| 605 | |
| 606 | expected = _encode_reference( |
| 607 | [ |
| 608 | messages[0], |
| 609 | {**messages[1], "tools": _request_tools()}, |
| 610 | messages[2], |
| 611 | messages[3], |
| 612 | ] |
| 613 | ) |
| 614 | assert prompt == expected |
| 615 | assert "First system.\n\n## Tools" in prompt |
| 616 | assert prompt.count("## Tools") == 1 |
| 617 | |
| 618 | |
| 619 | def test_deepseek_v4_request_tools_do_not_mutate_caller_messages(): |
| 620 | import copy |
| 621 | |
| 622 | messages = [ |
| 623 | {"role": "system", "content": "You are helpful."}, |
| 624 | {"role": "user", "content": "Weather in Paris?"}, |
| 625 | ] |
| 626 | snapshot = copy.deepcopy(messages) |
| 627 | |
| 628 | _tokenizer().apply_chat_template(messages, tools=_request_tools(), tokenize=False) |
| 629 | |
| 630 | assert messages == snapshot |