| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import json |
| 5 | |
| 6 | import numpy as np |
| 7 | import pybase64 as base64 |
| 8 | import pytest |
| 9 | import requests |
| 10 | import torch |
| 11 | |
| 12 | from tests.models.utils import check_embeddings_close |
| 13 | from tests.utils import RemoteOpenAIServer |
| 14 | from vllm.entrypoints.chat_utils import ChatCompletionMessageParam |
| 15 | from vllm.entrypoints.pooling.pooling.protocol import PoolingResponse |
| 16 | from vllm.entrypoints.pooling.utils import ( |
| 17 | MetadataItem, |
| 18 | build_metadata_items, |
| 19 | decode_pooling_output, |
| 20 | ) |
| 21 | from vllm.tokenizers import get_tokenizer |
| 22 | from vllm.utils.serial_utils import EMBED_DTYPES, ENDIANNESS, binary2tensor |
| 23 | |
| 24 | MODEL_NAME = "internlm/internlm2-1_8b-reward" |
| 25 | DUMMY_CHAT_TEMPLATE = """{% for message in messages %}{{message['role'] + ': ' + message['content'] + '\\n'}}{% endfor %}""" # noqa: E501 |
| 26 | input_text = "The chef prepared a delicious meal." |
| 27 | input_tokens = [1, 918, 29981, 10166, 395, 18067, 15265, 281] |
| 28 | |
| 29 | |
| 30 | @pytest.fixture(scope="module") |
| 31 | def server(): |
| 32 | args = [ |
| 33 | "--runner", |
| 34 | "pooling", |
| 35 | # use half precision for speed and memory savings in CI environment |
| 36 | "--dtype", |
| 37 | "bfloat16", |
| 38 | "--enforce-eager", |
| 39 | "--max-model-len", |
| 40 | "512", |
| 41 | "--chat-template", |
| 42 | DUMMY_CHAT_TEMPLATE, |
| 43 | "--trust-remote-code", |
| 44 | ] |
| 45 | |
| 46 | with RemoteOpenAIServer(MODEL_NAME, args) as remote_server: |
| 47 | yield remote_server |
| 48 | |
| 49 | |
| 50 | @pytest.mark.parametrize("model_name", [MODEL_NAME]) |
| 51 | def test_basic(server: RemoteOpenAIServer, model_name: str): |
| 52 | # test /v1/models |
| 53 | response = requests.get(server.url_for("/v1/models")) |
| 54 | served_model = response.json()["data"][0]["id"] |
| 55 | assert served_model == MODEL_NAME |
| 56 | |
| 57 | # test /tokenize |
| 58 | response = requests.post( |
| 59 | server.url_for("/tokenize"), |
| 60 | json={"model": model_name, "prompt": input_text}, |
| 61 | ) |
| 62 | assert response.json()["tokens"] == input_tokens |
| 63 | |
| 64 | |
| 65 | @pytest.mark.asyncio |
| 66 | @pytest.mark.parametrize("model_name", [MODEL_NAME]) |
| 67 | def test_completion_request(server: RemoteOpenAIServer, model_name: str): |
| 68 | # test input: str |
| 69 | response = requests.post( |
| 70 | server.url_for("pooling"), |
| 71 | json={"model": model_name, "input": input_text, "encoding_format": "float"}, |
| 72 | ) |
| 73 | response.raise_for_status() |
| 74 | poolings = PoolingResponse.model_validate(response.json()) |
| 75 | |
| 76 | assert poolings.id is not None |
| 77 | assert len(poolings.data) == 1 |
| 78 | assert len(poolings.data[0].data) == len(input_tokens) |
| 79 | assert poolings.usage.completion_tokens == 0 |
| 80 | assert poolings.usage.prompt_tokens == len(input_tokens) |
| 81 | assert poolings.usage.total_tokens == len(input_tokens) |
| 82 | |
| 83 | # test input: list[int] |
| 84 | response = requests.post( |
| 85 | server.url_for("pooling"), |
| 86 | json={"model": model_name, "input": input_tokens, "encoding_format": "float"}, |
| 87 | ) |
| 88 | response.raise_for_status() |
| 89 | poolings = PoolingResponse.model_validate(response.json()) |
| 90 | |
| 91 | assert poolings.id is not None |
| 92 | assert len(poolings.data) == 1 |
| 93 | assert len(poolings.data[0].data) == len(input_tokens) |
| 94 | assert poolings.usage.completion_tokens == 0 |
| 95 | assert poolings.usage.prompt_tokens == len(input_tokens) |
| 96 | assert poolings.usage.total_tokens == len(input_tokens) |
| 97 | |
| 98 | |
| 99 | @pytest.mark.parametrize("model_name", [MODEL_NAME]) |
| 100 | def test_completion_request_batched(server: RemoteOpenAIServer, model_name: str): |
| 101 | N = 10 |
| 102 | input_texts = [input_text] * N |
| 103 | |
| 104 | response = requests.post( |
| 105 | server.url_for("pooling"), |
| 106 | json={"model": model_name, "input": input_texts, "encoding_format": "float"}, |
| 107 | ) |
| 108 | response.raise_for_status() |
| 109 | poolings = PoolingResponse.model_validate(response.json()) |
| 110 | |
| 111 | assert poolings.id is not None |
| 112 | assert len(poolings.data) == N |
| 113 | assert len(poolings.data[0].data) == len(input_tokens) |
| 114 | assert poolings.usage.completion_tokens == 0 |
| 115 | assert poolings.usage.prompt_tokens == len(input_tokens) * N |
| 116 | assert poolings.usage.total_tokens == len(input_tokens) * N |
| 117 | |
| 118 | # test list[list[int]] |
| 119 | response = requests.post( |
| 120 | server.url_for("pooling"), |
| 121 | json={ |
| 122 | "model": model_name, |
| 123 | "input": [input_tokens] * N, |
| 124 | "encoding_format": "float", |
| 125 | }, |
| 126 | ) |
| 127 | response.raise_for_status() |
| 128 | poolings = PoolingResponse.model_validate(response.json()) |
| 129 | |
| 130 | assert poolings.id is not None |
| 131 | assert len(poolings.data) == N |
| 132 | assert len(poolings.data[0].data) == len(input_tokens) |
| 133 | assert poolings.usage.completion_tokens == 0 |
| 134 | assert poolings.usage.prompt_tokens == len(input_tokens) * N |
| 135 | assert poolings.usage.total_tokens == len(input_tokens) * N |
| 136 | |
| 137 | |
| 138 | @pytest.mark.asyncio |
| 139 | @pytest.mark.parametrize("model_name", [MODEL_NAME]) |
| 140 | async def test_chat_request(server: RemoteOpenAIServer, model_name: str): |
| 141 | messages: list[ChatCompletionMessageParam] = [ |
| 142 | { |
| 143 | "role": "user", |
| 144 | "content": "The cat sat on the mat.", |
| 145 | }, |
| 146 | { |
| 147 | "role": "assistant", |
| 148 | "content": "A feline was resting on a rug.", |
| 149 | }, |
| 150 | { |
| 151 | "role": "user", |
| 152 | "content": "Stars twinkle brightly in the night sky.", |
| 153 | }, |
| 154 | ] |
| 155 | |
| 156 | # test chat request basic usage |
| 157 | chat_response = requests.post( |
| 158 | server.url_for("pooling"), |
| 159 | json={ |
| 160 | "model": model_name, |
| 161 | "messages": messages, |
| 162 | "encoding_format": "float", |
| 163 | }, |
| 164 | ) |
| 165 | chat_response.raise_for_status() |
| 166 | chat_poolings = PoolingResponse.model_validate(chat_response.json()) |
| 167 | |
| 168 | tokenizer = get_tokenizer(tokenizer_name=model_name, trust_remote_code=True) |
| 169 | prompt = tokenizer.apply_chat_template( |
| 170 | messages, |
| 171 | chat_template=DUMMY_CHAT_TEMPLATE, |
| 172 | add_generation_prompt=True, |
| 173 | continue_final_message=False, |
| 174 | tokenize=False, |
| 175 | ) |
| 176 | completions_response = requests.post( |
| 177 | server.url_for("pooling"), |
| 178 | json={ |
| 179 | "model": model_name, |
| 180 | "input": prompt, |
| 181 | "encoding_format": "float", |
| 182 | # To be consistent with chat |
| 183 | "add_special_tokens": False, |
| 184 | }, |
| 185 | ) |
| 186 | completions_response.raise_for_status() |
| 187 | completion_poolings = PoolingResponse.model_validate(completions_response.json()) |
| 188 | |
| 189 | assert chat_poolings.id is not None |
| 190 | assert completion_poolings.id is not None |
| 191 | assert chat_poolings.created <= completion_poolings.created |
| 192 | assert chat_poolings.model_dump(exclude={"id", "created"}) == ( |
| 193 | completion_poolings.model_dump(exclude={"id", "created"}) |
| 194 | ) |
| 195 | |
| 196 | # test add_generation_prompt |
| 197 | response = requests.post( |
| 198 | server.url_for("pooling"), |
| 199 | json={"model": model_name, "messages": messages, "add_generation_prompt": True}, |
| 200 | ) |
| 201 | |
| 202 | response.raise_for_status() |
| 203 | output = PoolingResponse.model_validate(response.json()) |
| 204 | |
| 205 | assert output.object == "list" |
| 206 | assert len(output.data) == 1 |
| 207 | assert output.model == MODEL_NAME |
| 208 | assert output.usage.prompt_tokens == 33 |
| 209 | |
| 210 | # test continue_final_message |
| 211 | # The continue_final_message parameter doesn't seem to be working with this model. |
| 212 | response = requests.post( |
| 213 | server.url_for("pooling"), |
| 214 | json={ |
| 215 | "model": model_name, |
| 216 | "messages": messages, |
| 217 | "continue_final_message": True, |
| 218 | }, |
| 219 | ) |
| 220 | |
| 221 | response.raise_for_status() |
| 222 | output = PoolingResponse.model_validate(response.json()) |
| 223 | |
| 224 | assert output.object == "list" |
| 225 | assert len(output.data) == 1 |
| 226 | assert output.model == MODEL_NAME |
| 227 | assert output.usage.prompt_tokens == 33 |
| 228 | |
| 229 | # test add_special_tokens |
| 230 | response = requests.post( |
| 231 | server.url_for("pooling"), |
| 232 | json={"model": model_name, "messages": messages, "add_special_tokens": True}, |
| 233 | ) |
| 234 | |
| 235 | response.raise_for_status() |
| 236 | output = PoolingResponse.model_validate(response.json()) |
| 237 | |
| 238 | assert output.object == "list" |
| 239 | assert len(output.data) == 1 |
| 240 | assert output.model == MODEL_NAME |
| 241 | assert output.usage.prompt_tokens == 34 |
| 242 | |
| 243 | # test continue_final_message with add_generation_prompt |
| 244 | response = requests.post( |
| 245 | server.url_for("pooling"), |
| 246 | json={ |
| 247 | "model": model_name, |
| 248 | "messages": messages, |
| 249 | "continue_final_message": True, |
| 250 | "add_generation_prompt": True, |
| 251 | }, |
| 252 | ) |
| 253 | assert ( |
| 254 | "Cannot set both `continue_final_message` and `add_generation_prompt` to True." |
| 255 | in response.json()["error"]["message"] |
| 256 | ) |
| 257 | |
| 258 | |
| 259 | @pytest.mark.asyncio |
| 260 | @pytest.mark.parametrize("model_name", [MODEL_NAME]) |
| 261 | async def test_batch_base64_pooling(server: RemoteOpenAIServer, model_name: str): |
| 262 | input_texts = [ |
| 263 | "Hello my name is", |
| 264 | "The best thing about vLLM is that it supports many different models", |
| 265 | ] |
| 266 | |
| 267 | float_response = requests.post( |
| 268 | server.url_for("pooling"), |
| 269 | json={ |
| 270 | "input": input_texts, |
| 271 | "model": model_name, |
| 272 | "encoding_format": "float", |
| 273 | }, |
| 274 | ) |
| 275 | float_response.raise_for_status() |
| 276 | responses_float = PoolingResponse.model_validate(float_response.json()) |
| 277 | float_data = [np.array(d.data).squeeze(-1).tolist() for d in responses_float.data] |
| 278 | |
| 279 | base64_response = requests.post( |
| 280 | server.url_for("pooling"), |
| 281 | json={ |
| 282 | "input": input_texts, |
| 283 | "model": model_name, |
| 284 | "encoding_format": "base64", |
| 285 | }, |
| 286 | ) |
| 287 | base64_response.raise_for_status() |
| 288 | responses_base64 = PoolingResponse.model_validate(base64_response.json()) |
| 289 | |
| 290 | decoded_responses_base64_data = [] |
| 291 | for data in responses_base64.data: |
| 292 | decoded_responses_base64_data.append( |
| 293 | np.frombuffer(base64.b64decode(data.data), dtype="float32").tolist() |
| 294 | ) |
| 295 | |
| 296 | check_embeddings_close( |
| 297 | embeddings_0_lst=float_data, |
| 298 | embeddings_1_lst=decoded_responses_base64_data, |
| 299 | name_0="float32", |
| 300 | name_1="base64", |
| 301 | ) |
| 302 | |
| 303 | # Default response is float32 decoded from base64 by OpenAI Client |
| 304 | default_response = requests.post( |
| 305 | server.url_for("pooling"), |
| 306 | json={ |
| 307 | "input": input_texts, |
| 308 | "model": model_name, |
| 309 | }, |
| 310 | ) |
| 311 | default_response.raise_for_status() |
| 312 | responses_default = PoolingResponse.model_validate(default_response.json()) |
| 313 | default_data = [ |
| 314 | np.array(d.data).squeeze(-1).tolist() for d in responses_default.data |
| 315 | ] |
| 316 | |
| 317 | check_embeddings_close( |
| 318 | embeddings_0_lst=float_data, |
| 319 | embeddings_1_lst=default_data, |
| 320 | name_0="float32", |
| 321 | name_1="default", |
| 322 | ) |
| 323 | |
| 324 | |
| 325 | @pytest.mark.asyncio |
| 326 | @pytest.mark.parametrize("model_name", [MODEL_NAME]) |
| 327 | async def test_base64_embed_dtype_and_endianness( |
| 328 | server: RemoteOpenAIServer, model_name: str |
| 329 | ): |
| 330 | input_texts = [input_text] * 3 |
| 331 | |
| 332 | url = server.url_for("pooling") |
| 333 | float_response = requests.post( |
| 334 | url, |
| 335 | json={ |
| 336 | "model": model_name, |
| 337 | "input": input_texts, |
| 338 | "encoding_format": "float", |
| 339 | }, |
| 340 | ) |
| 341 | responses_float = PoolingResponse.model_validate(float_response.json()) |
| 342 | float_data = [np.array(d.data).squeeze(-1).tolist() for d in responses_float.data] |
| 343 | |
| 344 | for embed_dtype in EMBED_DTYPES: |
| 345 | for endianness in ENDIANNESS: |
| 346 | responses_base64 = requests.post( |
| 347 | url, |
| 348 | json={ |
| 349 | "model": model_name, |
| 350 | "input": input_texts, |
| 351 | "encoding_format": "base64", |
| 352 | "embed_dtype": embed_dtype, |
| 353 | "endianness": endianness, |
| 354 | }, |
| 355 | ) |
| 356 | |
| 357 | base64_data = [] |
| 358 | for data in responses_base64.json()["data"]: |
| 359 | binary = base64.b64decode(data["data"]) |
| 360 | tensor = binary2tensor(binary, (-1,), embed_dtype, endianness) |
| 361 | base64_data.append(tensor.to(torch.float32).tolist()) |
| 362 | |
| 363 | check_embeddings_close( |
| 364 | embeddings_0_lst=float_data, |
| 365 | embeddings_1_lst=base64_data, |
| 366 | name_0="float_data", |
| 367 | name_1="base64_data", |
| 368 | tol=1e-2, |
| 369 | ) |
| 370 | |
| 371 | |
| 372 | @pytest.mark.asyncio |
| 373 | @pytest.mark.parametrize("model_name", [MODEL_NAME]) |
| 374 | async def test_bytes_embed_dtype_and_endianness( |
| 375 | server: RemoteOpenAIServer, model_name: str |
| 376 | ): |
| 377 | input_texts = [input_text] * 3 |
| 378 | |
| 379 | url = server.url_for("pooling") |
| 380 | float_response = requests.post( |
| 381 | url, |
| 382 | json={ |
| 383 | "model": model_name, |
| 384 | "input": input_texts, |
| 385 | "encoding_format": "float", |
| 386 | }, |
| 387 | ) |
| 388 | responses_float = PoolingResponse.model_validate(float_response.json()) |
| 389 | float_data = [np.array(d.data).squeeze(-1).tolist() for d in responses_float.data] |
| 390 | |
| 391 | for embed_dtype in EMBED_DTYPES: |
| 392 | for endianness in ENDIANNESS: |
| 393 | responses_bytes = requests.post( |
| 394 | url, |
| 395 | json={ |
| 396 | "model": model_name, |
| 397 | "input": input_texts, |
| 398 | "encoding_format": "bytes", |
| 399 | "embed_dtype": embed_dtype, |
| 400 | "endianness": endianness, |
| 401 | }, |
| 402 | ) |
| 403 | |
| 404 | metadata = json.loads(responses_bytes.headers["metadata"]) |
| 405 | body = responses_bytes.content |
| 406 | items = [MetadataItem(**x) for x in metadata["data"]] |
| 407 | |
| 408 | bytes_data = decode_pooling_output(items=items, body=body) |
| 409 | bytes_data = [x.to(torch.float32).view(-1).tolist() for x in bytes_data] |
| 410 | |
| 411 | check_embeddings_close( |
| 412 | embeddings_0_lst=float_data, |
| 413 | embeddings_1_lst=bytes_data, |
| 414 | name_0="float_data", |
| 415 | name_1="bytes_data", |
| 416 | tol=1e-2, |
| 417 | ) |
| 418 | |
| 419 | |
| 420 | @pytest.mark.asyncio |
| 421 | @pytest.mark.parametrize("model_name", [MODEL_NAME]) |
| 422 | async def test_bytes_only_embed_dtype_and_endianness( |
| 423 | server: RemoteOpenAIServer, model_name: str |
| 424 | ): |
| 425 | input_texts = [input_text] * 3 |
| 426 | |
| 427 | url = server.url_for("pooling") |
| 428 | float_response = requests.post( |
| 429 | url, |
| 430 | json={ |
| 431 | "model": model_name, |
| 432 | "input": input_texts, |
| 433 | "encoding_format": "float", |
| 434 | }, |
| 435 | ) |
| 436 | responses_float = PoolingResponse.model_validate(float_response.json()) |
| 437 | float_data = [np.array(d.data).squeeze(-1).tolist() for d in responses_float.data] |
| 438 | n_tokens = responses_float.usage.prompt_tokens // len(input_texts) |
| 439 | |
| 440 | for embed_dtype in EMBED_DTYPES: |
| 441 | for endianness in ENDIANNESS: |
| 442 | responses_bytes = requests.post( |
| 443 | url, |
| 444 | json={ |
| 445 | "model": model_name, |
| 446 | "input": input_texts, |
| 447 | "encoding_format": "bytes_only", |
| 448 | "embed_dtype": embed_dtype, |
| 449 | "endianness": endianness, |
| 450 | }, |
| 451 | ) |
| 452 | |
| 453 | assert "metadata" not in responses_bytes.headers |
| 454 | body = responses_bytes.content |
| 455 | items = build_metadata_items( |
| 456 | embed_dtype=embed_dtype, |
| 457 | endianness=endianness, |
| 458 | shape=(n_tokens, 1), |
| 459 | n_request=len(input_texts), |
| 460 | ) |
| 461 | bytes_data = decode_pooling_output(items=items, body=body) |
| 462 | bytes_data = [x.to(torch.float32).view(-1).tolist() for x in bytes_data] |
| 463 | |
| 464 | check_embeddings_close( |
| 465 | embeddings_0_lst=float_data, |
| 466 | embeddings_1_lst=bytes_data, |
| 467 | name_0="float_data", |
| 468 | name_1="bytes_data", |
| 469 | tol=1e-2, |
| 470 | ) |
| 471 | |
| 472 | |
| 473 | @pytest.mark.asyncio |
| 474 | @pytest.mark.parametrize("model_name", [MODEL_NAME]) |
| 475 | @pytest.mark.parametrize("param_name", ["encoding_format", "embed_dtype", "endianness"]) |
| 476 | async def test_params_not_supported( |
| 477 | server: RemoteOpenAIServer, model_name: str, param_name: str |
| 478 | ): |
| 479 | responses_base64 = requests.post( |
| 480 | server.url_for("pooling"), |
| 481 | json={ |
| 482 | "model": model_name, |
| 483 | "input": input_text, |
| 484 | "encoding_format": "base64", |
| 485 | param_name: f"bad_{param_name}", |
| 486 | }, |
| 487 | ) |
| 488 | |
| 489 | assert responses_base64.status_code == 400 |
| 490 | assert "literal_error" in responses_base64.json()["error"]["message"] |
| 491 | assert f"bad_{param_name}" in responses_base64.json()["error"]["message"] |
| 492 | |
| 493 | |
| 494 | @pytest.mark.asyncio |
| 495 | async def test_invocations_chat_request(server: RemoteOpenAIServer): |
| 496 | request_args = { |
| 497 | "model": MODEL_NAME, |
| 498 | "input": input_text, |
| 499 | "encoding_format": "float", |
| 500 | } |
| 501 | |
| 502 | completion_response = requests.post(server.url_for("pooling"), json=request_args) |
| 503 | completion_response.raise_for_status() |
| 504 | |
| 505 | invocation_response = requests.post( |
| 506 | server.url_for("invocations"), json=request_args |
| 507 | ) |
| 508 | invocation_response.raise_for_status() |
| 509 | |
| 510 | completion_output = completion_response.json() |
| 511 | invocation_output = invocation_response.json() |
| 512 | |
| 513 | assert completion_output.keys() == invocation_output.keys() |
| 514 | for completion_data, invocation_data in zip( |
| 515 | completion_output["data"], invocation_output["data"] |
| 516 | ): |
| 517 | assert completion_data.keys() == invocation_data.keys() |
| 518 | check_embeddings_close( |
| 519 | embeddings_0_lst=completion_data["data"], |
| 520 | embeddings_1_lst=invocation_data["data"], |
| 521 | name_0="completion", |
| 522 | name_1="invocation", |
| 523 | ) |
| 524 | |
| 525 | |
| 526 | @pytest.mark.asyncio |
| 527 | async def test_invocations_conversation_chat_request(server: RemoteOpenAIServer): |
| 528 | messages: list[ChatCompletionMessageParam] = [ |
| 529 | { |
| 530 | "role": "user", |
| 531 | "content": "The cat sat on the mat.", |
| 532 | }, |
| 533 | { |
| 534 | "role": "assistant", |
| 535 | "content": "A feline was resting on a rug.", |
| 536 | }, |
| 537 | { |
| 538 | "role": "user", |
| 539 | "content": "Stars twinkle brightly in the night sky.", |
| 540 | }, |
| 541 | ] |
| 542 | |
| 543 | request_args = { |
| 544 | "model": MODEL_NAME, |
| 545 | "messages": messages, |
| 546 | "encoding_format": "float", |
| 547 | } |
| 548 | |
| 549 | chat_response = requests.post(server.url_for("pooling"), json=request_args) |
| 550 | chat_response.raise_for_status() |
| 551 | |
| 552 | invocation_response = requests.post( |
| 553 | server.url_for("invocations"), json=request_args |
| 554 | ) |
| 555 | invocation_response.raise_for_status() |
| 556 | |
| 557 | chat_output = chat_response.json() |
| 558 | invocation_output = invocation_response.json() |
| 559 | |
| 560 | assert chat_output.keys() == invocation_output.keys() |
| 561 | for chat_data, invocation_data in zip( |
| 562 | chat_output["data"], invocation_output["data"] |
| 563 | ): |
| 564 | assert chat_data.keys() == invocation_data.keys() |
| 565 | check_embeddings_close( |
| 566 | embeddings_0_lst=chat_data["data"], |
| 567 | embeddings_1_lst=invocation_data["data"], |
| 568 | name_0="chat", |
| 569 | name_1="invocation", |
| 570 | ) |