| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Unit tests for EmbedIOProcessor.""" |
| 4 | |
| 5 | import pytest |
| 6 | import torch |
| 7 | from pydantic import TypeAdapter, ValidationError |
| 8 | |
| 9 | from vllm import PoolingParams |
| 10 | from vllm.entrypoints.pooling.embed.io_processor import EmbedIOProcessor |
| 11 | from vllm.entrypoints.pooling.embed.protocol import ( |
| 12 | CohereEmbedRequest, |
| 13 | EmbeddingBatchChatInputRequest, |
| 14 | EmbeddingBatchChatRequest, |
| 15 | EmbeddingChatInputRequest, |
| 16 | EmbeddingChatRequest, |
| 17 | EmbeddingCompletionRequest, |
| 18 | EmbeddingRequest, |
| 19 | ) |
| 20 | from vllm.entrypoints.pooling.typing import ( |
| 21 | PoolingEngineInput, |
| 22 | PoolingServeContext, |
| 23 | ) |
| 24 | from vllm.inputs import tokens_input |
| 25 | from vllm.outputs import PoolingOutput, PoolingRequestOutput |
| 26 | |
| 27 | |
| 28 | class TestEmbeddingRequestParsing: |
| 29 | """Unit tests for OpenAI embedding request parsing.""" |
| 30 | |
| 31 | def test_input_messages_parses_as_chat_request(self): |
| 32 | request: EmbeddingRequest = TypeAdapter(EmbeddingRequest).validate_python( |
| 33 | { |
| 34 | "model": "test", |
| 35 | "input": [{"role": "user", "content": "hello"}], |
| 36 | "chat_template_kwargs": {"instruction": "Represent the query: "}, |
| 37 | } |
| 38 | ) |
| 39 | |
| 40 | assert isinstance(request, EmbeddingChatInputRequest) |
| 41 | assert request.input == [{"role": "user", "content": "hello"}] |
| 42 | assert request.messages == [{"role": "user", "content": "hello"}] |
| 43 | assert request.chat_template_kwargs == {"instruction": "Represent the query: "} |
| 44 | |
| 45 | def test_batched_input_messages_parses_as_batch_chat_input_request(self): |
| 46 | request: EmbeddingRequest = TypeAdapter(EmbeddingRequest).validate_python( |
| 47 | { |
| 48 | "model": "test", |
| 49 | "input": [ |
| 50 | [{"role": "user", "content": "hello"}], |
| 51 | [{"role": "user", "content": "goodbye"}], |
| 52 | ], |
| 53 | "chat_template_kwargs": {"instruction": "Represent the query: "}, |
| 54 | } |
| 55 | ) |
| 56 | |
| 57 | assert isinstance(request, EmbeddingBatchChatInputRequest) |
| 58 | assert request.input == [ |
| 59 | [{"role": "user", "content": "hello"}], |
| 60 | [{"role": "user", "content": "goodbye"}], |
| 61 | ] |
| 62 | assert request.messages == [ |
| 63 | [{"role": "user", "content": "hello"}], |
| 64 | [{"role": "user", "content": "goodbye"}], |
| 65 | ] |
| 66 | assert request.chat_template_kwargs == {"instruction": "Represent the query: "} |
| 67 | |
| 68 | def test_token_ids_still_parse_as_completion_request(self): |
| 69 | request: EmbeddingRequest = TypeAdapter(EmbeddingRequest).validate_python( |
| 70 | { |
| 71 | "model": "test", |
| 72 | "input": [[1, 2, 3], [4, 5]], |
| 73 | } |
| 74 | ) |
| 75 | |
| 76 | assert isinstance(request, EmbeddingCompletionRequest) |
| 77 | assert request.input == [[1, 2, 3], [4, 5]] |
| 78 | |
| 79 | def test_messages_still_parses_as_chat_request(self): |
| 80 | request: EmbeddingRequest = TypeAdapter(EmbeddingRequest).validate_python( |
| 81 | { |
| 82 | "model": "test", |
| 83 | "messages": [{"role": "user", "content": "hello"}], |
| 84 | "chat_template_kwargs": {"instruction": "Represent the query: "}, |
| 85 | } |
| 86 | ) |
| 87 | |
| 88 | assert isinstance(request, EmbeddingChatRequest) |
| 89 | assert request.messages == [{"role": "user", "content": "hello"}] |
| 90 | assert request.chat_template_kwargs == {"instruction": "Represent the query: "} |
| 91 | |
| 92 | def test_batched_messages_parses_as_batch_chat_request(self): |
| 93 | request: EmbeddingRequest = TypeAdapter(EmbeddingRequest).validate_python( |
| 94 | { |
| 95 | "model": "test", |
| 96 | "messages": [ |
| 97 | [{"role": "user", "content": "hello"}], |
| 98 | [{"role": "user", "content": "goodbye"}], |
| 99 | ], |
| 100 | "chat_template_kwargs": {"instruction": "Represent the query: "}, |
| 101 | } |
| 102 | ) |
| 103 | |
| 104 | assert isinstance(request, EmbeddingBatchChatRequest) |
| 105 | assert request.messages == [ |
| 106 | [{"role": "user", "content": "hello"}], |
| 107 | [{"role": "user", "content": "goodbye"}], |
| 108 | ] |
| 109 | assert request.chat_template_kwargs == {"instruction": "Represent the query: "} |
| 110 | |
| 111 | |
| 112 | class TestCohereEmbedRequestParsing: |
| 113 | """Unit tests for Cohere embed request parsing.""" |
| 114 | |
| 115 | @pytest.mark.parametrize( |
| 116 | "request_body", |
| 117 | [ |
| 118 | {"model": "test"}, |
| 119 | {"model": "test", "texts": ["hello"], "images": ["image-uri"]}, |
| 120 | { |
| 121 | "model": "test", |
| 122 | "texts": ["hello"], |
| 123 | "inputs": [ |
| 124 | {"content": [{"type": "text", "text": "hello"}]}, |
| 125 | ], |
| 126 | }, |
| 127 | { |
| 128 | "model": "test", |
| 129 | "images": ["image-uri"], |
| 130 | "inputs": [ |
| 131 | {"content": [{"type": "text", "text": "hello"}]}, |
| 132 | ], |
| 133 | }, |
| 134 | {"model": "test", "texts": []}, |
| 135 | {"model": "test", "images": []}, |
| 136 | {"model": "test", "inputs": []}, |
| 137 | ], |
| 138 | ) |
| 139 | def test_rejects_invalid_input_field_combinations(self, request_body): |
| 140 | with pytest.raises( |
| 141 | ValidationError, |
| 142 | match="Exactly one of texts, images, or inputs must be provided", |
| 143 | ): |
| 144 | CohereEmbedRequest(**request_body) |
| 145 | |
| 146 | @pytest.mark.parametrize( |
| 147 | "request_body", |
| 148 | [ |
| 149 | {"model": "test", "texts": ["hello"]}, |
| 150 | {"model": "test", "images": ["image-uri"]}, |
| 151 | { |
| 152 | "model": "test", |
| 153 | "inputs": [ |
| 154 | {"content": [{"type": "text", "text": "hello"}]}, |
| 155 | ], |
| 156 | }, |
| 157 | { |
| 158 | "model": "test", |
| 159 | "inputs": [ |
| 160 | { |
| 161 | "content": [ |
| 162 | {"type": "image_url", "image_url": {"url": "image-uri"}} |
| 163 | ] |
| 164 | }, |
| 165 | ], |
| 166 | }, |
| 167 | ], |
| 168 | ) |
| 169 | def test_accepts_exactly_one_non_empty_input_field(self, request_body): |
| 170 | request = CohereEmbedRequest(**request_body) |
| 171 | |
| 172 | assert request.model == "test" |
| 173 | |
| 174 | @pytest.mark.parametrize( |
| 175 | ("content", "error"), |
| 176 | [ |
| 177 | ( |
| 178 | {"type": "text"}, |
| 179 | "CohereEmbedContent with type='text' requires text", |
| 180 | ), |
| 181 | ( |
| 182 | {"type": "image_url"}, |
| 183 | "CohereEmbedContent with type='image_url' requires image_url.url", |
| 184 | ), |
| 185 | ( |
| 186 | {"type": "image_url", "image_url": {}}, |
| 187 | "CohereEmbedContent with type='image_url' requires image_url.url", |
| 188 | ), |
| 189 | ( |
| 190 | {"type": "image_url", "image_url": {"url": ""}}, |
| 191 | "CohereEmbedContent with type='image_url' requires image_url.url", |
| 192 | ), |
| 193 | ], |
| 194 | ) |
| 195 | def test_rejects_invalid_mixed_content_payloads(self, content, error): |
| 196 | with pytest.raises(ValidationError, match=error): |
| 197 | CohereEmbedRequest( |
| 198 | model="test", |
| 199 | inputs=[ |
| 200 | { |
| 201 | "content": [content], |
| 202 | }, |
| 203 | ], |
| 204 | ) |
| 205 | |
| 206 | |
| 207 | class TestResolveTruncation: |
| 208 | """Unit tests for EmbedIOProcessor._resolve_cohere_truncation.""" |
| 209 | |
| 210 | @staticmethod |
| 211 | def _make_request(**kwargs) -> CohereEmbedRequest: |
| 212 | defaults = { |
| 213 | "model": "test", |
| 214 | "input_type": "search_document", |
| 215 | "texts": ["hello"], |
| 216 | } |
| 217 | return CohereEmbedRequest(**(defaults | kwargs)) |
| 218 | |
| 219 | def test_truncate_end_default(self): |
| 220 | req = self._make_request() |
| 221 | tokens, side = EmbedIOProcessor._resolve_cohere_truncation(req) |
| 222 | assert tokens == -1 |
| 223 | assert side is None |
| 224 | |
| 225 | def test_truncate_end_explicit(self): |
| 226 | req = self._make_request(truncate="END") |
| 227 | tokens, side = EmbedIOProcessor._resolve_cohere_truncation(req) |
| 228 | assert tokens == -1 |
| 229 | assert side is None |
| 230 | |
| 231 | def test_truncate_end_with_max_tokens(self): |
| 232 | req = self._make_request(truncate="END", max_tokens=128) |
| 233 | tokens, side = EmbedIOProcessor._resolve_cohere_truncation(req) |
| 234 | assert tokens == 128 |
| 235 | assert side is None |
| 236 | |
| 237 | def test_truncate_none(self): |
| 238 | req = self._make_request(truncate="NONE") |
| 239 | tokens, side = EmbedIOProcessor._resolve_cohere_truncation(req) |
| 240 | assert tokens is None |
| 241 | assert side is None |
| 242 | |
| 243 | def test_truncate_none_with_max_tokens(self): |
| 244 | """truncate=NONE should NOT set truncate_prompt_tokens; the |
| 245 | max_tokens limit is enforced separately via _check_max_tokens.""" |
| 246 | req = self._make_request(truncate="NONE", max_tokens=10) |
| 247 | tokens, side = EmbedIOProcessor._resolve_cohere_truncation(req) |
| 248 | assert tokens is None |
| 249 | assert side is None |
| 250 | |
| 251 | def test_truncate_start(self): |
| 252 | req = self._make_request(truncate="START") |
| 253 | tokens, side = EmbedIOProcessor._resolve_cohere_truncation(req) |
| 254 | assert tokens == -1 |
| 255 | assert side == "left" |
| 256 | |
| 257 | def test_truncate_start_with_max_tokens(self): |
| 258 | req = self._make_request(truncate="START", max_tokens=64) |
| 259 | tokens, side = EmbedIOProcessor._resolve_cohere_truncation(req) |
| 260 | assert tokens == 64 |
| 261 | assert side == "left" |
| 262 | |
| 263 | |
| 264 | class TestApplyStPrompt: |
| 265 | """Unit tests for EmbedIOProcessor._apply_task_instruction.""" |
| 266 | |
| 267 | @staticmethod |
| 268 | def _make_handler(task_instructions: dict[str, str] | None): |
| 269 | handler = object.__new__(EmbedIOProcessor) |
| 270 | handler.task_instructions = task_instructions |
| 271 | return handler |
| 272 | |
| 273 | def test_no_prompts_configured(self): |
| 274 | handler = self._make_handler(None) |
| 275 | texts = ["hello", "world"] |
| 276 | assert handler._apply_task_instruction(texts, "query") is texts |
| 277 | |
| 278 | def test_matching_input_type(self): |
| 279 | handler = self._make_handler({"query": "search_query: "}) |
| 280 | result = handler._apply_task_instruction(["hello"], "query") |
| 281 | assert result == ["search_query: hello"] |
| 282 | |
| 283 | def test_non_matching_input_type(self): |
| 284 | handler = self._make_handler({"query": "search_query: "}) |
| 285 | texts = ["hello"] |
| 286 | assert handler._apply_task_instruction(texts, "document") is texts |
| 287 | |
| 288 | def test_multiple_texts(self): |
| 289 | handler = self._make_handler( |
| 290 | {"query": "Represent this sentence for searching: "} |
| 291 | ) |
| 292 | result = handler._apply_task_instruction(["a", "b", "c"], "query") |
| 293 | assert result == [ |
| 294 | "Represent this sentence for searching: a", |
| 295 | "Represent this sentence for searching: b", |
| 296 | "Represent this sentence for searching: c", |
| 297 | ] |
| 298 | |
| 299 | def test_empty_prefix_returns_unchanged(self): |
| 300 | handler = self._make_handler({"passage": ""}) |
| 301 | texts = ["hello"] |
| 302 | assert handler._apply_task_instruction(texts, "passage") is texts |
| 303 | |
| 304 | |
| 305 | class TestLoadTaskInstructions: |
| 306 | """Unit tests for EmbedIOProcessor._load_task_instructions.""" |
| 307 | |
| 308 | def test_no_attribute(self): |
| 309 | class FakeConfig: |
| 310 | pass |
| 311 | |
| 312 | assert EmbedIOProcessor._load_task_instructions(FakeConfig()) is None |
| 313 | |
| 314 | def test_with_task_instructions(self): |
| 315 | class FakeConfig: |
| 316 | task_instructions = { |
| 317 | "retrieval.query": "Represent the query: ", |
| 318 | "retrieval.passage": "", |
| 319 | } |
| 320 | |
| 321 | result = EmbedIOProcessor._load_task_instructions(FakeConfig()) |
| 322 | assert result == { |
| 323 | "retrieval.query": "Represent the query: ", |
| 324 | "retrieval.passage": "", |
| 325 | } |
| 326 | |
| 327 | def test_empty_dict(self): |
| 328 | class FakeConfig: |
| 329 | task_instructions = {} |
| 330 | |
| 331 | assert EmbedIOProcessor._load_task_instructions(FakeConfig()) is None |
| 332 | |
| 333 | def test_non_dict(self): |
| 334 | class FakeConfig: |
| 335 | task_instructions = "not a dict" |
| 336 | |
| 337 | assert EmbedIOProcessor._load_task_instructions(FakeConfig()) is None |
| 338 | |
| 339 | |
| 340 | class TestCheckMaxTokens: |
| 341 | """Unit tests for EmbedIOProcessor._check_cohere_max_tokens.""" |
| 342 | |
| 343 | @staticmethod |
| 344 | def _fake_output(n_tokens: int): |
| 345 | class _Out: |
| 346 | def __init__(self, n: int): |
| 347 | self.prompt_token_ids = list(range(n)) |
| 348 | |
| 349 | return _Out(n_tokens) |
| 350 | |
| 351 | def test_none_check_is_noop(self): |
| 352 | outs = [self._fake_output(100)] |
| 353 | EmbedIOProcessor._check_cohere_max_tokens(outs, None) |
| 354 | |
| 355 | def test_within_limit(self): |
| 356 | outs = [self._fake_output(5), self._fake_output(3)] |
| 357 | EmbedIOProcessor._check_cohere_max_tokens(outs, 5) |
| 358 | |
| 359 | def test_exceeds_limit(self): |
| 360 | outs = [self._fake_output(3), self._fake_output(10)] |
| 361 | with pytest.raises(ValueError, match="exceeds max_tokens=5"): |
| 362 | EmbedIOProcessor._check_cohere_max_tokens(outs, 5) |
| 363 | |
| 364 | def test_exact_limit(self): |
| 365 | outs = [self._fake_output(5)] |
| 366 | EmbedIOProcessor._check_cohere_max_tokens(outs, 5) |
| 367 | |
| 368 | |
| 369 | class TestValidateInputType: |
| 370 | """Unit tests for EmbedIOProcessor._validate_input_type.""" |
| 371 | |
| 372 | @staticmethod |
| 373 | def _make_handler(task_instructions: dict[str, str] | None): |
| 374 | handler = object.__new__(EmbedIOProcessor) |
| 375 | handler.task_instructions = task_instructions |
| 376 | return handler |
| 377 | |
| 378 | def test_none_input_type_always_accepted(self): |
| 379 | handler = self._make_handler(None) |
| 380 | handler._validate_input_type(None) |
| 381 | handler_with = self._make_handler({"query": "q: "}) |
| 382 | handler_with._validate_input_type(None) |
| 383 | |
| 384 | def test_no_prompts_rejects(self): |
| 385 | handler = self._make_handler(None) |
| 386 | with pytest.raises(ValueError, match="does not define any input_type"): |
| 387 | handler._validate_input_type("anything") |
| 388 | |
| 389 | def test_known_type_accepted(self): |
| 390 | handler = self._make_handler({"query": "q: ", "document": "d: "}) |
| 391 | handler._validate_input_type("query") |
| 392 | handler._validate_input_type("document") |
| 393 | |
| 394 | def test_unknown_type_rejected(self): |
| 395 | handler = self._make_handler({"query": "q: ", "document": "d: "}) |
| 396 | with pytest.raises(ValueError, match="Unsupported input_type 'other'"): |
| 397 | handler._validate_input_type("other") |
| 398 | |
| 399 | def test_error_lists_supported(self): |
| 400 | handler = self._make_handler({"a": "", "b": ""}) |
| 401 | with pytest.raises(ValueError, match="Supported values: a, b"): |
| 402 | handler._validate_input_type("z") |
| 403 | |
| 404 | |
| 405 | class TestChunkedEmbeddingProcessing: |
| 406 | """Unit tests for chunked embedding aggregation.""" |
| 407 | |
| 408 | class _FakeModelConfig: |
| 409 | max_model_len = 3 |
| 410 | |
| 411 | @classmethod |
| 412 | def _make_handler(cls): |
| 413 | handler = object.__new__(EmbedIOProcessor) |
| 414 | handler.model_config = cls._FakeModelConfig() |
| 415 | handler.enable_chunked_processing = True |
| 416 | return handler |
| 417 | |
| 418 | @staticmethod |
| 419 | def _make_context() -> PoolingServeContext[EmbeddingCompletionRequest]: |
| 420 | request: EmbeddingRequest = TypeAdapter(EmbeddingRequest).validate_python( |
| 421 | { |
| 422 | "model": "test", |
| 423 | "input": [[0, 1, 2, 3, 4], [10, 11]], |
| 424 | } |
| 425 | ) |
| 426 | assert isinstance(request, EmbeddingCompletionRequest) |
| 427 | pooling_params = PoolingParams() |
| 428 | return PoolingServeContext( |
| 429 | request=request, |
| 430 | pooling_params=pooling_params, |
| 431 | model_name="test", |
| 432 | request_id="embd-client-prompt-999-chunk-888", |
| 433 | engine_inputs=[ |
| 434 | PoolingEngineInput( |
| 435 | prompts=tokens_input(prompt_token_ids=[0, 1, 2, 3, 4]), |
| 436 | params=pooling_params, |
| 437 | lora_requests=None, |
| 438 | priorities=0, |
| 439 | ), |
| 440 | PoolingEngineInput( |
| 441 | prompts=tokens_input(prompt_token_ids=[10, 11]), |
| 442 | params=pooling_params, |
| 443 | lora_requests=None, |
| 444 | priorities=0, |
| 445 | ), |
| 446 | ], |
| 447 | lora_request=None, |
| 448 | priorities=0, |
| 449 | prompt_extras=None, |
| 450 | ) |
| 451 | |
| 452 | @staticmethod |
| 453 | def _make_output( |
| 454 | request_id: str, |
| 455 | prompt_token_ids: list[int], |
| 456 | embedding: list[float], |
| 457 | ) -> PoolingRequestOutput: |
| 458 | return PoolingRequestOutput( |
| 459 | request_id=request_id, |
| 460 | outputs=PoolingOutput(data=torch.tensor(embedding)), |
| 461 | prompt_token_ids=prompt_token_ids, |
| 462 | num_cached_tokens=0, |
| 463 | finished=True, |
| 464 | ) |
| 465 | |
| 466 | def test_aggregation_uses_metadata_not_request_id_parsing(self): |
| 467 | handler = self._make_handler() |
| 468 | ctx = self._make_context() |
| 469 | |
| 470 | handler.maybe_pre_process_chunked(ctx) |
| 471 | |
| 472 | assert ctx.prompt_request_ids == [ |
| 473 | "embd-client-prompt-999-chunk-888-prompt-0-chunk-0", |
| 474 | "embd-client-prompt-999-chunk-888-prompt-0-chunk-1", |
| 475 | "embd-client-prompt-999-chunk-888-prompt-1-chunk-0", |
| 476 | ] |
| 477 | assert ctx.chunked_embedding_metadata is not None |
| 478 | assert [ |
| 479 | (item.prompt_index, item.chunk_index) |
| 480 | for item in ctx.chunked_embedding_metadata |
| 481 | ] == [(0, 0), (0, 1), (1, 0)] |
| 482 | |
| 483 | ctx.final_res_batch = [ |
| 484 | self._make_output(ctx.prompt_request_ids[0], [0, 1, 2], [1.0, 1.0]), |
| 485 | self._make_output(ctx.prompt_request_ids[1], [3, 4], [4.0, 7.0]), |
| 486 | self._make_output(ctx.prompt_request_ids[2], [10, 11], [9.0, 9.0]), |
| 487 | ] |
| 488 | |
| 489 | handler._post_process_chunked(ctx) |
| 490 | |
| 491 | assert len(ctx.final_res_batch) == 2 |
| 492 | assert ctx.final_res_batch[0].request_id == ( |
| 493 | "embd-client-prompt-999-chunk-888-prompt-0" |
| 494 | ) |
| 495 | assert ctx.final_res_batch[0].prompt_token_ids == [0, 1, 2, 3, 4] |
| 496 | assert torch.allclose( |
| 497 | ctx.final_res_batch[0].outputs.data, |
| 498 | torch.tensor([2.2, 3.4]), |
| 499 | ) |
| 500 | assert ctx.final_res_batch[1].request_id == ( |
| 501 | "embd-client-prompt-999-chunk-888-prompt-1" |
| 502 | ) |
| 503 | assert ctx.final_res_batch[1].prompt_token_ids == [10, 11] |
| 504 | assert torch.allclose( |
| 505 | ctx.final_res_batch[1].outputs.data, |
| 506 | torch.tensor([9.0, 9.0]), |
| 507 | ) |