| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import io |
| 5 | from collections.abc import Sequence |
| 6 | from dataclasses import dataclass |
| 7 | from typing import Any |
| 8 | |
| 9 | import pybase64 |
| 10 | import pytest |
| 11 | import torch |
| 12 | |
| 13 | from vllm.config import ModelConfig |
| 14 | from vllm.exceptions import VLLMValidationError |
| 15 | from vllm.inputs import SingletonPrompt |
| 16 | from vllm.renderers import TokenizeParams |
| 17 | from vllm.renderers.hf import HfRenderer |
| 18 | from vllm.renderers.inputs.preprocess import parse_model_prompt, prompt_to_seq |
| 19 | |
| 20 | MODEL_NAME = "openai-community/gpt2" |
| 21 | |
| 22 | |
| 23 | @dataclass |
| 24 | class MockHFConfig: |
| 25 | model_type: str = "any" |
| 26 | |
| 27 | |
| 28 | @dataclass |
| 29 | class MockModelConfig: |
| 30 | runner_type = "generate" |
| 31 | model: str = MODEL_NAME |
| 32 | tokenizer: str = MODEL_NAME |
| 33 | trust_remote_code: bool = False |
| 34 | tokenizer_revision = None |
| 35 | tokenizer_mode = "auto" |
| 36 | hf_config = MockHFConfig() |
| 37 | encoder_config: dict[str, Any] | None = None |
| 38 | enable_prompt_embeds: bool = True |
| 39 | skip_tokenizer_init: bool = False |
| 40 | is_encoder_decoder: bool = False |
| 41 | is_multimodal_model: bool = False |
| 42 | supports_multimodal_inputs: bool = False |
| 43 | renderer_num_workers: int = 1 |
| 44 | hidden_size: int = 768 |
| 45 | dtype: torch.dtype = torch.float32 |
| 46 | |
| 47 | def get_hidden_size(self) -> int: |
| 48 | return self.hidden_size |
| 49 | |
| 50 | |
| 51 | @dataclass |
| 52 | class MockParallelConfig: |
| 53 | _api_process_rank: int = 0 |
| 54 | |
| 55 | |
| 56 | @dataclass |
| 57 | class MockVllmConfig: |
| 58 | model_config: MockModelConfig |
| 59 | parallel_config: MockParallelConfig |
| 60 | |
| 61 | |
| 62 | @dataclass |
| 63 | class DummyTokenizer: |
| 64 | truncation_side: str = "left" |
| 65 | max_chars_per_token: int = 1 |
| 66 | # Deliberately outside the range of ids `encode` returns, so a test can |
| 67 | # tell a pad token apart from a real one. |
| 68 | pad_token_id: int = 99999 |
| 69 | |
| 70 | def __post_init__(self) -> None: |
| 71 | self._captured_encode_kwargs: dict = {} |
| 72 | self._captured_text_len: int = 0 |
| 73 | |
| 74 | def decode(self, tokens: list[int]): |
| 75 | return str(tokens) |
| 76 | |
| 77 | def encode(self, text: str, **kwargs): |
| 78 | self._captured_encode_kwargs = kwargs |
| 79 | self._captured_text_len = len(text) |
| 80 | |
| 81 | in_length = len(text) |
| 82 | truncation = kwargs.get("truncation") |
| 83 | max_length = kwargs.get("max_length") |
| 84 | if truncation and max_length is not None: |
| 85 | return list(range(min(in_length, max_length))) |
| 86 | |
| 87 | return list(range(in_length)) |
| 88 | |
| 89 | def __call__(self, text: str, **kwargs): |
| 90 | # BaseRenderer._tokenize_prompt calls the tokenizer via __call__ (to |
| 91 | # unify the output type), so mirror a real tokenizer's BatchEncoding. |
| 92 | return {"input_ids": self.encode(text, **kwargs)} |
| 93 | |
| 94 | |
| 95 | def _build_renderer( |
| 96 | model_config: MockModelConfig, |
| 97 | *, |
| 98 | truncation_side: str = "left", |
| 99 | max_chars_per_token: int = 1, |
| 100 | ): |
| 101 | renderer = HfRenderer( |
| 102 | MockVllmConfig(model_config, parallel_config=MockParallelConfig()), |
| 103 | tokenizer=( |
| 104 | None |
| 105 | if model_config.skip_tokenizer_init |
| 106 | else DummyTokenizer( |
| 107 | truncation_side=truncation_side, |
| 108 | max_chars_per_token=max_chars_per_token, |
| 109 | ) |
| 110 | ), |
| 111 | ) |
| 112 | |
| 113 | return renderer |
| 114 | |
| 115 | |
| 116 | def _preprocess_prompt( |
| 117 | model_config: ModelConfig, |
| 118 | prompt_or_prompts: SingletonPrompt | bytes | Sequence[SingletonPrompt | bytes], |
| 119 | ): |
| 120 | return [ |
| 121 | ( |
| 122 | prompt |
| 123 | if isinstance(prompt, bytes) |
| 124 | else parse_model_prompt(model_config, prompt) |
| 125 | ) |
| 126 | for prompt in prompt_to_seq(prompt_or_prompts) |
| 127 | ] |
| 128 | |
| 129 | |
| 130 | class TestValidatePrompt: |
| 131 | def test_empty_input(self): |
| 132 | renderer = _build_renderer(MockModelConfig()) |
| 133 | |
| 134 | with pytest.raises(ValueError, match="at least one prompt"): |
| 135 | renderer.render_prompts(_preprocess_prompt(renderer.model_config, [])) |
| 136 | |
| 137 | def test_invalid_type(self): |
| 138 | renderer = _build_renderer(MockModelConfig()) |
| 139 | |
| 140 | with pytest.raises(TypeError, match="should be a list of integers"): |
| 141 | renderer.render_prompts( |
| 142 | _preprocess_prompt(renderer.model_config, [[1, 2], ["foo", "bar"]]) |
| 143 | ) |
| 144 | |
| 145 | |
| 146 | class TestRenderPrompt: |
| 147 | def test_tokens_input(self): |
| 148 | renderer = _build_renderer(MockModelConfig()) |
| 149 | |
| 150 | tokens = [101, 7592, 2088] |
| 151 | prompts = renderer.render_prompts( |
| 152 | _preprocess_prompt(renderer.model_config, tokens) |
| 153 | ) |
| 154 | results = renderer.tokenize_prompts( |
| 155 | prompts, |
| 156 | TokenizeParams(max_total_tokens=100), |
| 157 | ) |
| 158 | |
| 159 | assert len(results) == 1 |
| 160 | assert results[0]["prompt_token_ids"] == tokens |
| 161 | |
| 162 | def test_token_list_input(self): |
| 163 | renderer = _build_renderer(MockModelConfig()) |
| 164 | |
| 165 | token_lists = [[101, 7592, 2088], [102, 1234, 5678, 9012], [103, 4567]] |
| 166 | prompts = renderer.render_prompts( |
| 167 | _preprocess_prompt(renderer.model_config, token_lists) |
| 168 | ) |
| 169 | results = renderer.tokenize_prompts( |
| 170 | prompts, |
| 171 | TokenizeParams(max_total_tokens=100), |
| 172 | ) |
| 173 | |
| 174 | assert len(results) == 3 |
| 175 | assert results[0]["prompt_token_ids"] == [101, 7592, 2088] |
| 176 | assert results[1]["prompt_token_ids"] == [102, 1234, 5678, 9012] |
| 177 | assert results[2]["prompt_token_ids"] == [103, 4567] |
| 178 | |
| 179 | def test_text_input(self): |
| 180 | renderer = _build_renderer(MockModelConfig()) |
| 181 | |
| 182 | text_input = "x" * 10 |
| 183 | prompts = renderer.render_prompts( |
| 184 | _preprocess_prompt(renderer.model_config, text_input) |
| 185 | ) |
| 186 | results = renderer.tokenize_prompts( |
| 187 | prompts, |
| 188 | TokenizeParams(max_total_tokens=100), |
| 189 | ) |
| 190 | |
| 191 | assert len(results) == 1 |
| 192 | assert len(results[0]["prompt_token_ids"]) == 10 |
| 193 | |
| 194 | def test_text_list_input(self): |
| 195 | renderer = _build_renderer(MockModelConfig()) |
| 196 | |
| 197 | text_list_input = ["x" * 10, "x" * 12, "x" * 14] |
| 198 | prompts = renderer.render_prompts( |
| 199 | _preprocess_prompt(renderer.model_config, text_list_input) |
| 200 | ) |
| 201 | results = renderer.tokenize_prompts( |
| 202 | prompts, |
| 203 | TokenizeParams(max_total_tokens=100), |
| 204 | ) |
| 205 | |
| 206 | assert len(results) == 3 |
| 207 | for text_input, result in zip(text_list_input, results): |
| 208 | assert len(result["prompt_token_ids"]) == len(text_input) |
| 209 | |
| 210 | def test_zero_truncation(self): |
| 211 | renderer = _build_renderer(MockModelConfig()) |
| 212 | |
| 213 | prompts = renderer.render_prompts( |
| 214 | _preprocess_prompt(renderer.model_config, "x" * 200) |
| 215 | ) |
| 216 | results = renderer.tokenize_prompts( |
| 217 | prompts, |
| 218 | TokenizeParams(max_total_tokens=100, truncate_prompt_tokens=0), |
| 219 | ) |
| 220 | |
| 221 | assert len(results) == 1 |
| 222 | assert len(results[0]["prompt_token_ids"]) == 0 |
| 223 | |
| 224 | def test_pos_truncation(self): |
| 225 | renderer = _build_renderer(MockModelConfig()) |
| 226 | |
| 227 | prompts = renderer.render_prompts( |
| 228 | _preprocess_prompt(renderer.model_config, "x" * 200) |
| 229 | ) |
| 230 | results = renderer.tokenize_prompts( |
| 231 | prompts, |
| 232 | TokenizeParams(max_total_tokens=100, truncate_prompt_tokens=50), |
| 233 | ) |
| 234 | |
| 235 | assert len(results) == 1 |
| 236 | assert len(results[0]["prompt_token_ids"]) == 50 |
| 237 | |
| 238 | def test_neg_truncation(self): |
| 239 | renderer = _build_renderer(MockModelConfig()) |
| 240 | |
| 241 | prompts = renderer.render_prompts( |
| 242 | _preprocess_prompt(renderer.model_config, "x" * 200) |
| 243 | ) |
| 244 | results = renderer.tokenize_prompts( |
| 245 | prompts, |
| 246 | TokenizeParams(max_total_tokens=100, truncate_prompt_tokens=-1), |
| 247 | ) |
| 248 | |
| 249 | assert len(results) == 1 |
| 250 | assert len(results[0]["prompt_token_ids"]) == 100 # max_total_tokens |
| 251 | |
| 252 | def test_truncation_left(self): |
| 253 | renderer = _build_renderer(MockModelConfig(), truncation_side="left") |
| 254 | |
| 255 | long_tokens = [100, 101, 102, 103, 104, 105, 106, 107, 108, 109] # 10 tokens |
| 256 | prompts = renderer.render_prompts( |
| 257 | _preprocess_prompt(renderer.model_config, long_tokens) |
| 258 | ) |
| 259 | results = renderer.tokenize_prompts( |
| 260 | prompts, |
| 261 | TokenizeParams(max_total_tokens=100, truncate_prompt_tokens=5), |
| 262 | ) |
| 263 | |
| 264 | assert len(results) == 1 |
| 265 | # Should keep the last 5 tokens: [105, 106, 107, 108, 109] |
| 266 | assert results[0]["prompt_token_ids"] == [105, 106, 107, 108, 109] |
| 267 | |
| 268 | def test_truncation_right(self): |
| 269 | renderer = _build_renderer(MockModelConfig(), truncation_side="right") |
| 270 | |
| 271 | long_tokens = [100, 101, 102, 103, 104, 105, 106, 107, 108, 109] # 10 tokens |
| 272 | prompts = renderer.render_prompts( |
| 273 | _preprocess_prompt(renderer.model_config, long_tokens) |
| 274 | ) |
| 275 | results = renderer.tokenize_prompts( |
| 276 | prompts, |
| 277 | TokenizeParams(max_total_tokens=100, truncate_prompt_tokens=5), |
| 278 | ) |
| 279 | |
| 280 | assert len(results) == 1 |
| 281 | # Should keep the first 5 tokens: [100, 101, 102, 103, 104] |
| 282 | assert results[0]["prompt_token_ids"] == [100, 101, 102, 103, 104] |
| 283 | |
| 284 | def test_text_max_length_exceeded_obvious(self): |
| 285 | renderer = _build_renderer(MockModelConfig(), max_chars_per_token=1) |
| 286 | |
| 287 | # Exceeds max_total_tokens and max_total_tokens * VLLM_MAX_CHARS_PER_TOKEN |
| 288 | long_tokens = "x" * 150 |
| 289 | prompts = renderer.render_prompts( |
| 290 | _preprocess_prompt(renderer.model_config, long_tokens) |
| 291 | ) |
| 292 | |
| 293 | with pytest.raises( |
| 294 | VLLMValidationError, |
| 295 | match="maximum context length is", |
| 296 | ): |
| 297 | renderer.tokenize_prompts( |
| 298 | prompts, |
| 299 | TokenizeParams(max_total_tokens=100), |
| 300 | ) |
| 301 | |
| 302 | # Should not even attempt tokenization |
| 303 | assert renderer.tokenizer._captured_encode_kwargs == {} |
| 304 | |
| 305 | def test_text_max_length_exceeded_nonobvious(self): |
| 306 | renderer = _build_renderer(MockModelConfig(), max_chars_per_token=2) |
| 307 | |
| 308 | # Exceeds max_total_tokens but not max_total_tokens * VLLM_MAX_CHARS_PER_TOKEN |
| 309 | long_tokens = "x" * 150 |
| 310 | prompts = renderer.render_prompts( |
| 311 | _preprocess_prompt(renderer.model_config, long_tokens) |
| 312 | ) |
| 313 | |
| 314 | with pytest.raises( |
| 315 | VLLMValidationError, |
| 316 | match="maximum context length is", |
| 317 | ): |
| 318 | renderer.tokenize_prompts( |
| 319 | prompts, |
| 320 | TokenizeParams(max_total_tokens=100), |
| 321 | ) |
| 322 | |
| 323 | # Should only tokenize the first max_total_tokens + 1 tokens |
| 324 | assert renderer.tokenizer._captured_encode_kwargs["truncation"] is True |
| 325 | assert renderer.tokenizer._captured_encode_kwargs["max_length"] == 101 |
| 326 | |
| 327 | def test_token_max_length_exceeded(self): |
| 328 | renderer = _build_renderer(MockModelConfig()) |
| 329 | |
| 330 | long_tokens = list(range(150)) # Exceeds max_total_tokens=100 |
| 331 | prompts = renderer.render_prompts( |
| 332 | _preprocess_prompt(renderer.model_config, long_tokens) |
| 333 | ) |
| 334 | |
| 335 | with pytest.raises( |
| 336 | VLLMValidationError, |
| 337 | match="maximum context length is", |
| 338 | ): |
| 339 | renderer.tokenize_prompts( |
| 340 | prompts, |
| 341 | TokenizeParams(max_total_tokens=100, truncate_prompt_tokens=None), |
| 342 | ) |
| 343 | |
| 344 | def test_no_tokenizer_for_text(self): |
| 345 | renderer = _build_renderer(MockModelConfig(skip_tokenizer_init=True)) |
| 346 | |
| 347 | prompts = renderer.render_prompts( |
| 348 | _preprocess_prompt(renderer.model_config, "Hello world") |
| 349 | ) |
| 350 | |
| 351 | with pytest.raises(ValueError, match="`skip_tokenizer_init=True`"): |
| 352 | renderer.tokenize_prompts( |
| 353 | prompts, |
| 354 | TokenizeParams(max_total_tokens=100), |
| 355 | ) |
| 356 | |
| 357 | def test_tokens_input_with_needs_detokenization(self): |
| 358 | renderer = _build_renderer(MockModelConfig()) |
| 359 | |
| 360 | tokens = [1, 2, 3, 4] |
| 361 | prompts = renderer.render_prompts( |
| 362 | _preprocess_prompt(renderer.model_config, tokens) |
| 363 | ) |
| 364 | results = renderer.tokenize_prompts( |
| 365 | prompts, |
| 366 | TokenizeParams( |
| 367 | max_total_tokens=100, |
| 368 | needs_detokenization=True, |
| 369 | ), |
| 370 | ) |
| 371 | |
| 372 | assert len(results) == 1 |
| 373 | assert results[0]["prompt_token_ids"] == tokens |
| 374 | assert results[0]["prompt"] == "[1, 2, 3, 4]" |
| 375 | |
| 376 | def test_explicit_side_tokenizer_unbounded(self): |
| 377 | renderer = _build_renderer(MockModelConfig()) |
| 378 | |
| 379 | prompts = renderer.render_prompts( |
| 380 | _preprocess_prompt(renderer.model_config, "x" * 500) |
| 381 | ) |
| 382 | results = renderer.tokenize_prompts( |
| 383 | prompts, |
| 384 | TokenizeParams( |
| 385 | max_total_tokens=100, |
| 386 | truncate_prompt_tokens=4, |
| 387 | truncation_side="left", |
| 388 | ), |
| 389 | ) |
| 390 | |
| 391 | assert len(results) == 1 |
| 392 | assert len(results[0]["prompt_token_ids"]) == 4 |
| 393 | |
| 394 | kwargs = renderer.tokenizer._captured_encode_kwargs |
| 395 | assert kwargs["truncation"] is False |
| 396 | |
| 397 | def test_explicit_side_left_text(self): |
| 398 | renderer = _build_renderer(MockModelConfig()) |
| 399 | |
| 400 | prompts = renderer.render_prompts( |
| 401 | _preprocess_prompt(renderer.model_config, "x" * 50) |
| 402 | ) |
| 403 | results = renderer.tokenize_prompts( |
| 404 | prompts, |
| 405 | TokenizeParams( |
| 406 | max_total_tokens=100, |
| 407 | truncate_prompt_tokens=5, |
| 408 | truncation_side="left", |
| 409 | ), |
| 410 | ) |
| 411 | |
| 412 | assert len(results) == 1 |
| 413 | assert len(results[0]["prompt_token_ids"]) == 5 |
| 414 | assert results[0]["prompt_token_ids"] == list(range(45, 50)) |
| 415 | |
| 416 | def test_explicit_side_right_text(self): |
| 417 | renderer = _build_renderer(MockModelConfig()) |
| 418 | |
| 419 | prompts = renderer.render_prompts( |
| 420 | _preprocess_prompt(renderer.model_config, "x" * 50) |
| 421 | ) |
| 422 | results = renderer.tokenize_prompts( |
| 423 | prompts, |
| 424 | TokenizeParams( |
| 425 | max_total_tokens=100, |
| 426 | truncate_prompt_tokens=5, |
| 427 | truncation_side="right", |
| 428 | ), |
| 429 | ) |
| 430 | |
| 431 | assert len(results) == 1 |
| 432 | assert len(results[0]["prompt_token_ids"]) == 5 |
| 433 | assert results[0]["prompt_token_ids"] == list(range(5)) |
| 434 | |
| 435 | def test_padding_with_left_truncation_keeps_the_prompt(self): |
| 436 | """Padding must not be truncated away. |
| 437 | |
| 438 | `padding` and `truncate_prompt_tokens`/`truncation_side` are all |
| 439 | settable on one pooling request. Padding to the full input length |
| 440 | before truncating from the left leaves a prompt made entirely of pad |
| 441 | tokens, and the request still succeeds -- so the model embeds nothing |
| 442 | but padding. |
| 443 | """ |
| 444 | renderer = _build_renderer(MockModelConfig()) |
| 445 | pad_id = renderer.tokenizer.pad_token_id |
| 446 | |
| 447 | prompts = renderer.render_prompts( |
| 448 | _preprocess_prompt(renderer.model_config, "x" * 50) |
| 449 | ) |
| 450 | results = renderer.tokenize_prompts( |
| 451 | prompts, |
| 452 | TokenizeParams( |
| 453 | max_total_tokens=100, |
| 454 | pad_prompt_tokens=-1, |
| 455 | truncate_prompt_tokens=5, |
| 456 | truncation_side="left", |
| 457 | ), |
| 458 | ) |
| 459 | |
| 460 | assert len(results) == 1 |
| 461 | token_ids = results[0]["prompt_token_ids"] |
| 462 | |
| 463 | # The sentinel: on a padding-first pipeline every surviving id is a |
| 464 | # pad token, so the prompt reaches the model with no content at all. |
| 465 | assert set(token_ids) != {pad_id} |
| 466 | |
| 467 | # Transformers semantics: truncate to 5, then pad out to the full |
| 468 | # input length. |
| 469 | assert token_ids[:5] == list(range(45, 50)) |
| 470 | assert token_ids[5:] == [pad_id] * 95 |
| 471 | |
| 472 | def test_padding_without_truncation_is_unchanged(self): |
| 473 | renderer = _build_renderer(MockModelConfig()) |
| 474 | pad_id = renderer.tokenizer.pad_token_id |
| 475 | |
| 476 | prompts = renderer.render_prompts( |
| 477 | _preprocess_prompt(renderer.model_config, "x" * 50) |
| 478 | ) |
| 479 | results = renderer.tokenize_prompts( |
| 480 | prompts, |
| 481 | TokenizeParams(max_total_tokens=100, pad_prompt_tokens=-1), |
| 482 | ) |
| 483 | |
| 484 | assert len(results) == 1 |
| 485 | assert results[0]["prompt_token_ids"] == list(range(50)) + [pad_id] * 50 |
| 486 | |
| 487 | def test_explicit_side_text_pretokenization_guard(self): |
| 488 | renderer = _build_renderer(MockModelConfig(), max_chars_per_token=1) |
| 489 | |
| 490 | prompts = renderer.render_prompts( |
| 491 | _preprocess_prompt(renderer.model_config, "x" * 500) |
| 492 | ) |
| 493 | results = renderer.tokenize_prompts( |
| 494 | prompts, |
| 495 | TokenizeParams( |
| 496 | max_total_tokens=100, |
| 497 | truncate_prompt_tokens=4, |
| 498 | truncation_side="left", |
| 499 | ), |
| 500 | ) |
| 501 | |
| 502 | assert len(results) == 1 |
| 503 | assert len(results[0]["prompt_token_ids"]) == 4 |
| 504 | |
| 505 | assert renderer.tokenizer._captured_text_len <= 100 |
| 506 | |
| 507 | |
| 508 | class TestRenderEmbedPrompt: |
| 509 | def _create_test_embed_bytes(self, tensor: torch.Tensor) -> bytes: |
| 510 | """Helper to create base64-encoded tensor bytes.""" |
| 511 | buffer = io.BytesIO() |
| 512 | torch.save(tensor, buffer) |
| 513 | buffer.seek(0) |
| 514 | return pybase64.b64encode(buffer.read()) |
| 515 | |
| 516 | def test_single_prompt_embed(self): |
| 517 | renderer = _build_renderer(MockModelConfig()) |
| 518 | |
| 519 | # Create a test tensor |
| 520 | tensor_input = torch.randn(10, 768, dtype=torch.float32) |
| 521 | embed_bytes = self._create_test_embed_bytes(tensor_input) |
| 522 | |
| 523 | prompts = renderer.render_prompts( |
| 524 | _preprocess_prompt(renderer.model_config, embed_bytes) |
| 525 | ) |
| 526 | results = renderer.tokenize_prompts( |
| 527 | prompts, |
| 528 | TokenizeParams(max_total_tokens=100), |
| 529 | ) |
| 530 | |
| 531 | assert len(results) == 1 |
| 532 | assert torch.equal(results[0]["prompt_embeds"], tensor_input) |
| 533 | |
| 534 | def test_multiple_prompt_embeds(self): |
| 535 | hidden_size = 512 |
| 536 | renderer = _build_renderer(MockModelConfig(hidden_size=hidden_size)) |
| 537 | |
| 538 | # Create multiple test tensors |
| 539 | tensor_inputs = [ |
| 540 | torch.randn(8, hidden_size, dtype=torch.float32), |
| 541 | torch.randn(12, hidden_size, dtype=torch.float32), |
| 542 | ] |
| 543 | |
| 544 | prompts = renderer.render_prompts( |
| 545 | _preprocess_prompt( |
| 546 | renderer.model_config, |
| 547 | [self._create_test_embed_bytes(t) for t in tensor_inputs], |
| 548 | ) |
| 549 | ) |
| 550 | results = renderer.tokenize_prompts( |
| 551 | prompts, |
| 552 | TokenizeParams(max_total_tokens=100), |
| 553 | ) |
| 554 | |
| 555 | assert len(results) == 2 |
| 556 | for i, result in enumerate(results): |
| 557 | assert torch.allclose(result["prompt_embeds"], tensor_inputs[i]) |
| 558 | |
| 559 | def test_prompt_embed_truncation(self): |
| 560 | renderer = _build_renderer(MockModelConfig()) |
| 561 | |
| 562 | # Create tensor with more tokens than truncation limit |
| 563 | tensor_input = torch.randn(20, 768, dtype=torch.float32) |
| 564 | |
| 565 | prompts = renderer.render_prompts( |
| 566 | _preprocess_prompt( |
| 567 | renderer.model_config, self._create_test_embed_bytes(tensor_input) |
| 568 | ) |
| 569 | ) |
| 570 | results = renderer.tokenize_prompts( |
| 571 | prompts, |
| 572 | TokenizeParams( |
| 573 | max_total_tokens=100, |
| 574 | truncate_prompt_tokens=10, |
| 575 | ), |
| 576 | ) |
| 577 | |
| 578 | assert len(results) == 1 |
| 579 | # Should keep last 10 tokens |
| 580 | expected = tensor_input[-10:] |
| 581 | assert torch.equal(results[0]["prompt_embeds"], expected) |
| 582 | |
| 583 | def test_prompt_embed_different_dtypes(self): |
| 584 | hidden_size = 256 |
| 585 | # Test different supported dtypes |
| 586 | dtypes = [torch.float32, torch.float16, torch.bfloat16] |
| 587 | |
| 588 | for dtype in dtypes: |
| 589 | renderer = _build_renderer( |
| 590 | MockModelConfig(hidden_size=hidden_size, dtype=dtype) |
| 591 | ) |
| 592 | tensor_input = torch.randn(5, hidden_size, dtype=dtype) |
| 593 | |
| 594 | prompts = renderer.render_prompts( |
| 595 | _preprocess_prompt( |
| 596 | renderer.model_config, self._create_test_embed_bytes(tensor_input) |
| 597 | ) |
| 598 | ) |
| 599 | results = renderer.tokenize_prompts( |
| 600 | prompts, |
| 601 | TokenizeParams(max_total_tokens=100), |
| 602 | ) |
| 603 | |
| 604 | assert len(results) == 1 |
| 605 | assert results[0]["prompt_embeds"].dtype == dtype |
| 606 | |
| 607 | def test_prompt_embed_squeeze_batch_dim(self): |
| 608 | renderer = _build_renderer(MockModelConfig()) |
| 609 | |
| 610 | # Test tensor with batch dimension gets squeezed |
| 611 | tensor_input = torch.randn(1, 10, 768, dtype=torch.float32) |
| 612 | |
| 613 | prompts = renderer.render_prompts( |
| 614 | _preprocess_prompt( |
| 615 | renderer.model_config, self._create_test_embed_bytes(tensor_input) |
| 616 | ) |
| 617 | ) |
| 618 | results = renderer.tokenize_prompts( |
| 619 | prompts, |
| 620 | TokenizeParams(max_total_tokens=100), |
| 621 | ) |
| 622 | |
| 623 | assert len(results) == 1 |
| 624 | # Should be squeezed to 2D |
| 625 | assert results[0]["prompt_embeds"].shape == (10, 768) |
| 626 | |
| 627 | def test_both_prompts_and_embeds(self): |
| 628 | hidden_size = 256 |
| 629 | renderer = _build_renderer(MockModelConfig(hidden_size=hidden_size)) |
| 630 | |
| 631 | text_input = "Hello world" |
| 632 | tensor_input = torch.randn(5, hidden_size, dtype=torch.float32) |
| 633 | |
| 634 | prompts = renderer.render_prompts( |
| 635 | _preprocess_prompt( |
| 636 | renderer.model_config, |
| 637 | [text_input, self._create_test_embed_bytes(tensor_input)], |
| 638 | ) |
| 639 | ) |
| 640 | results = renderer.tokenize_prompts( |
| 641 | prompts, |
| 642 | TokenizeParams(max_total_tokens=100), |
| 643 | ) |
| 644 | |
| 645 | assert len(results) == 2 |
| 646 | # First should be tokens prompt |
| 647 | assert "prompt_token_ids" in results[0] |
| 648 | assert len(results[0]["prompt_token_ids"]) == len(text_input) |
| 649 | # Second should be embed prompt |
| 650 | assert torch.equal(results[1]["prompt_embeds"], tensor_input) |