| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Unit tests for sequence and token pooling methods and their factories.""" |
| 4 | |
| 5 | from dataclasses import dataclass |
| 6 | from types import SimpleNamespace |
| 7 | from unittest.mock import patch |
| 8 | |
| 9 | import numpy as np |
| 10 | import pytest |
| 11 | import torch |
| 12 | from transformers import PreTrainedConfig |
| 13 | |
| 14 | from vllm.config import PoolerConfig, set_current_vllm_config |
| 15 | from vllm.model_executor.layers.pooler import PoolingParamsUpdate |
| 16 | from vllm.model_executor.layers.pooler.seqwise.methods import ( |
| 17 | CLSPool, |
| 18 | LastPool, |
| 19 | MeanPool, |
| 20 | get_seq_pooling_method, |
| 21 | ) |
| 22 | from vllm.model_executor.layers.pooler.special import DispatchPooler |
| 23 | from vllm.model_executor.layers.pooler.tokwise.methods import ( |
| 24 | AllPool, |
| 25 | StepPool, |
| 26 | get_tok_pooling_method, |
| 27 | ) |
| 28 | from vllm.pooling_params import PoolingParams |
| 29 | from vllm.tasks import PoolingTask |
| 30 | from vllm.v1.pool.metadata import PoolingCursor, PoolingMetadata, PoolingStates |
| 31 | |
| 32 | _CPU = torch.device("cpu") |
| 33 | |
| 34 | |
| 35 | def _make_pooling_cursor( |
| 36 | prompt_lens: list[int], |
| 37 | *, |
| 38 | num_scheduled_tokens: list[int] | None = None, |
| 39 | seq_lens: list[int] | None = None, |
| 40 | device: torch.device = _CPU, |
| 41 | ) -> PoolingCursor: |
| 42 | """Build a PoolingCursor from a list of per-sequence prompt lengths.""" |
| 43 | prompt_lens_cpu = torch.tensor(prompt_lens, dtype=torch.long) |
| 44 | if num_scheduled_tokens is None: |
| 45 | num_scheduled_tokens_cpu = prompt_lens_cpu.clone() |
| 46 | else: |
| 47 | num_scheduled_tokens_cpu = torch.tensor(num_scheduled_tokens, dtype=torch.long) |
| 48 | if seq_lens is None: |
| 49 | seq_lens_cpu = prompt_lens_cpu.clone() |
| 50 | else: |
| 51 | seq_lens_cpu = torch.tensor(seq_lens, dtype=torch.long) |
| 52 | |
| 53 | cumsum = torch.zeros(len(prompt_lens) + 1, dtype=torch.long, device=device) |
| 54 | torch.cumsum(num_scheduled_tokens_cpu, dim=0, out=cumsum[1:]) |
| 55 | |
| 56 | return PoolingCursor( |
| 57 | first_token_indices_gpu=cumsum[: len(prompt_lens)].to(device), |
| 58 | last_token_indices_gpu=(cumsum[1:] - 1).to(device), |
| 59 | prompt_lens_cpu=prompt_lens_cpu, |
| 60 | seq_lens_cpu=seq_lens_cpu, |
| 61 | num_scheduled_tokens_cpu=num_scheduled_tokens_cpu, |
| 62 | partial_prefill=not torch.equal(prompt_lens_cpu, num_scheduled_tokens_cpu), |
| 63 | finished_mask=torch.eq(prompt_lens_cpu, seq_lens_cpu).tolist(), |
| 64 | ) |
| 65 | |
| 66 | |
| 67 | def _make_metadata( |
| 68 | prompt_lens: list[int], |
| 69 | *, |
| 70 | tasks: list[PoolingTask] | None = None, |
| 71 | token_ids: list[list[int]] | None = None, |
| 72 | pooling_params: list[PoolingParams] | None = None, |
| 73 | num_scheduled_tokens: list[int] | None = None, |
| 74 | seq_lens: list[int] | None = None, |
| 75 | device: torch.device = _CPU, |
| 76 | ) -> PoolingMetadata: |
| 77 | """Build a minimal PoolingMetadata for testing pooling methods.""" |
| 78 | n_seqs = len(prompt_lens) |
| 79 | if tasks is None: |
| 80 | tasks = ["embed"] * n_seqs |
| 81 | if pooling_params is None: |
| 82 | pooling_params = [PoolingParams(task=t) for t in tasks] |
| 83 | |
| 84 | prompt_lens_tensor = torch.tensor(prompt_lens, dtype=torch.long) |
| 85 | |
| 86 | prompt_token_ids_cpu = None |
| 87 | prompt_token_ids = None |
| 88 | if token_ids is not None: |
| 89 | max_len = max(len(t) for t in token_ids) |
| 90 | padded = [t + [0] * (max_len - len(t)) for t in token_ids] |
| 91 | prompt_token_ids_cpu = torch.tensor(padded, dtype=torch.long) |
| 92 | prompt_token_ids = prompt_token_ids_cpu.to(device) |
| 93 | |
| 94 | cursor = _make_pooling_cursor( |
| 95 | prompt_lens, |
| 96 | num_scheduled_tokens=num_scheduled_tokens, |
| 97 | seq_lens=seq_lens, |
| 98 | device=device, |
| 99 | ) |
| 100 | |
| 101 | pooling_states = [PoolingStates() for _ in range(n_seqs)] |
| 102 | |
| 103 | return PoolingMetadata( |
| 104 | prompt_lens=prompt_lens_tensor, |
| 105 | prompt_token_ids=prompt_token_ids, |
| 106 | prompt_token_ids_cpu=prompt_token_ids_cpu, |
| 107 | pooling_params=pooling_params, |
| 108 | pooling_states=pooling_states, |
| 109 | pooling_cursor=cursor, |
| 110 | ) |
| 111 | |
| 112 | |
| 113 | def test_build_pooling_cursor_caches_cpu_status() -> None: |
| 114 | metadata = PoolingMetadata( |
| 115 | prompt_lens=torch.tensor([3, 4, 2], dtype=torch.long), |
| 116 | prompt_token_ids=None, |
| 117 | prompt_token_ids_cpu=None, |
| 118 | pooling_params=[PoolingParams(task="embed") for _ in range(3)], |
| 119 | pooling_states=[PoolingStates() for _ in range(3)], |
| 120 | ) |
| 121 | metadata.build_pooling_cursor( |
| 122 | np.array([3, 2, 2], dtype=np.int32), |
| 123 | seq_lens_cpu=torch.tensor([3, 2, 1], dtype=torch.long), |
| 124 | device=_CPU, |
| 125 | ) |
| 126 | |
| 127 | cursor = metadata.get_pooling_cursor() |
| 128 | |
| 129 | assert cursor.is_partial_prefill() is True |
| 130 | assert cursor.get_finished_mask() == [True, False, False] |
| 131 | assert cursor[:1].is_partial_prefill() is False |
| 132 | assert cursor[1:2].is_partial_prefill() is True |
| 133 | assert cursor[1:].get_finished_mask() == [False, False] |
| 134 | |
| 135 | |
| 136 | def test_pooling_params_update_combines_token_id_requirements() -> None: |
| 137 | update = PoolingParamsUpdate() | PoolingParamsUpdate(requires_token_ids=True) |
| 138 | params = PoolingParams(task="embed") |
| 139 | |
| 140 | update.apply(params) |
| 141 | |
| 142 | assert params.requires_token_ids is True |
| 143 | |
| 144 | |
| 145 | # --------------------------------------------------------------------------- |
| 146 | # CLSPool |
| 147 | # --------------------------------------------------------------------------- |
| 148 | class TestCLSPool: |
| 149 | def test_extracts_first_token(self): |
| 150 | hidden = torch.tensor( |
| 151 | [[1.0, 2.0], [3.0, 4.0], [5.0, 6.0], [7.0, 8.0], [9.0, 10.0]] |
| 152 | ) |
| 153 | metadata = _make_metadata([2, 3]) |
| 154 | pooler = CLSPool() |
| 155 | out = pooler(hidden, metadata) |
| 156 | expected = torch.tensor([[1.0, 2.0], [5.0, 6.0]]) |
| 157 | assert torch.equal(out, expected) |
| 158 | |
| 159 | def test_rejects_partial_prefill(self): |
| 160 | hidden = torch.tensor([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]]) |
| 161 | metadata = _make_metadata([3], num_scheduled_tokens=[2]) |
| 162 | pooler = CLSPool() |
| 163 | with pytest.raises(RuntimeError, match="partial prefill"): |
| 164 | pooler(hidden, metadata) |
| 165 | |
| 166 | |
| 167 | # --------------------------------------------------------------------------- |
| 168 | # LastPool |
| 169 | # --------------------------------------------------------------------------- |
| 170 | class TestLastPool: |
| 171 | def test_extracts_last_token(self): |
| 172 | hidden = torch.tensor( |
| 173 | [[1.0, 2.0], [3.0, 4.0], [5.0, 6.0], [7.0, 8.0], [9.0, 10.0]] |
| 174 | ) |
| 175 | metadata = _make_metadata([2, 3]) |
| 176 | pooler = LastPool() |
| 177 | out = pooler(hidden, metadata) |
| 178 | expected = torch.tensor([[3.0, 4.0], [9.0, 10.0]]) |
| 179 | assert torch.equal(out, expected) |
| 180 | |
| 181 | def test_partial_prefill_extracts_last_scheduled(self): |
| 182 | hidden = torch.tensor([[1.0, 2.0], [3.0, 4.0]]) |
| 183 | metadata = _make_metadata([4], num_scheduled_tokens=[2]) |
| 184 | pooler = LastPool() |
| 185 | out = pooler(hidden, metadata) |
| 186 | expected = torch.tensor([[3.0, 4.0]]) |
| 187 | assert torch.equal(out, expected) |
| 188 | |
| 189 | |
| 190 | # --------------------------------------------------------------------------- |
| 191 | # MeanPool |
| 192 | # --------------------------------------------------------------------------- |
| 193 | class TestMeanPool: |
| 194 | def test_computes_mean(self): |
| 195 | hidden = torch.tensor( |
| 196 | [[1.0, 2.0], [3.0, 4.0], [10.0, 20.0]], dtype=torch.float32 |
| 197 | ) |
| 198 | metadata = _make_metadata([2, 1]) |
| 199 | pooler = MeanPool() |
| 200 | out = pooler(hidden, metadata) |
| 201 | expected = torch.tensor([[2.0, 3.0], [10.0, 20.0]], dtype=torch.float32) |
| 202 | assert torch.allclose(out, expected, atol=1e-5) |
| 203 | |
| 204 | def test_single_token_is_identity(self): |
| 205 | hidden = torch.tensor([[5.0, 10.0]], dtype=torch.float32) |
| 206 | metadata = _make_metadata([1]) |
| 207 | pooler = MeanPool() |
| 208 | out = pooler(hidden, metadata) |
| 209 | assert torch.allclose(out, hidden, atol=1e-5) |
| 210 | |
| 211 | def test_uniform_values_return_same(self): |
| 212 | hidden = torch.full((4, 3), 7.0, dtype=torch.float32) |
| 213 | metadata = _make_metadata([4]) |
| 214 | pooler = MeanPool() |
| 215 | out = pooler(hidden, metadata) |
| 216 | expected = torch.full((1, 3), 7.0, dtype=torch.float32) |
| 217 | assert torch.allclose(out, expected, atol=1e-5) |
| 218 | |
| 219 | def test_multiple_sequences(self): |
| 220 | hidden = torch.tensor( |
| 221 | [ |
| 222 | [0.0, 0.0], |
| 223 | [2.0, 4.0], |
| 224 | [4.0, 8.0], |
| 225 | [10.0, 10.0], |
| 226 | ], |
| 227 | dtype=torch.float32, |
| 228 | ) |
| 229 | metadata = _make_metadata([3, 1]) |
| 230 | pooler = MeanPool() |
| 231 | out = pooler(hidden, metadata) |
| 232 | expected = torch.tensor([[2.0, 4.0], [10.0, 10.0]], dtype=torch.float32) |
| 233 | assert torch.allclose(out, expected, atol=1e-5) |
| 234 | |
| 235 | def test_empty_batch(self): |
| 236 | hidden = torch.empty((0, 8), dtype=torch.float32) |
| 237 | metadata = _make_metadata([]) |
| 238 | pooler = MeanPool() |
| 239 | out = pooler(hidden, metadata) |
| 240 | assert out.shape == (0, 8) |
| 241 | |
| 242 | def test_rejects_partial_prefill(self): |
| 243 | hidden = torch.tensor([[1.0, 2.0], [3.0, 4.0]], dtype=torch.float32) |
| 244 | metadata = _make_metadata([3], num_scheduled_tokens=[2]) |
| 245 | pooler = MeanPool() |
| 246 | with pytest.raises(RuntimeError, match="partial prefill"): |
| 247 | pooler(hidden, metadata) |
| 248 | |
| 249 | def test_chunked_accumulation(self): |
| 250 | hidden = torch.arange(20, dtype=torch.float32).reshape(5, 4) |
| 251 | metadata = _make_metadata([3, 2]) |
| 252 | pooler = MeanPool() |
| 253 | with patch( |
| 254 | "vllm.model_executor.layers.pooler.seqwise.methods" |
| 255 | "._MEAN_POOL_ACCUMULATION_CHUNK_BYTES", |
| 256 | 16, |
| 257 | ): |
| 258 | out = pooler(hidden, metadata) |
| 259 | expected_seq0 = hidden[:3].float().mean(dim=0, keepdim=True) |
| 260 | expected_seq1 = hidden[3:].float().mean(dim=0, keepdim=True) |
| 261 | expected = torch.cat([expected_seq0, expected_seq1], dim=0) |
| 262 | assert torch.allclose(out, expected, atol=1e-5) |
| 263 | |
| 264 | def test_upcasts_to_float32(self): |
| 265 | hidden = torch.tensor([[1.0, 2.0], [3.0, 4.0]], dtype=torch.float16) |
| 266 | metadata = _make_metadata([2]) |
| 267 | pooler = MeanPool() |
| 268 | out = pooler(hidden, metadata) |
| 269 | assert out.dtype == torch.float32 |
| 270 | expected = torch.tensor([[2.0, 3.0]], dtype=torch.float32) |
| 271 | assert torch.allclose(out, expected, atol=1e-2) |
| 272 | |
| 273 | |
| 274 | # --------------------------------------------------------------------------- |
| 275 | # get_seq_pooling_method factory |
| 276 | # --------------------------------------------------------------------------- |
| 277 | class TestGetSeqPoolingMethod: |
| 278 | def test_cls(self): |
| 279 | assert isinstance(get_seq_pooling_method("CLS"), CLSPool) |
| 280 | |
| 281 | def test_last(self): |
| 282 | assert isinstance(get_seq_pooling_method("LAST"), LastPool) |
| 283 | |
| 284 | def test_mean(self): |
| 285 | assert isinstance(get_seq_pooling_method("MEAN"), MeanPool) |
| 286 | |
| 287 | def test_unknown_raises(self): |
| 288 | with pytest.raises(NotImplementedError, match="UNKNOWN"): |
| 289 | get_seq_pooling_method("UNKNOWN") |
| 290 | |
| 291 | |
| 292 | # --------------------------------------------------------------------------- |
| 293 | # AllPool |
| 294 | # --------------------------------------------------------------------------- |
| 295 | |
| 296 | |
| 297 | @dataclass |
| 298 | class _FakeSchedulerConfig: |
| 299 | enable_chunked_prefill: bool = False |
| 300 | async_scheduling: bool = False |
| 301 | |
| 302 | |
| 303 | @dataclass |
| 304 | class _FakeVllmConfig: |
| 305 | scheduler_config: _FakeSchedulerConfig |
| 306 | |
| 307 | |
| 308 | @pytest.mark.parametrize("tok_pooling_type", ["ALL", "STEP"]) |
| 309 | def test_dispatch_seq_cls_honors_token_pooling_type(tok_pooling_type): |
| 310 | """Filter STEP rows before classification; preserve all rows for resolved ALL.""" |
| 311 | pooler_config = PoolerConfig( |
| 312 | seq_pooling_type="LAST", tok_pooling_type=tok_pooling_type |
| 313 | ) |
| 314 | config = SimpleNamespace( |
| 315 | scheduler_config=_FakeSchedulerConfig(), |
| 316 | model_config=SimpleNamespace( |
| 317 | pooler_config=pooler_config, |
| 318 | head_dtype=None, |
| 319 | hf_config=PreTrainedConfig(num_labels=3), |
| 320 | ), |
| 321 | ) |
| 322 | classifier = torch.nn.Linear(4, 3) |
| 323 | with set_current_vllm_config(config): |
| 324 | pooler = DispatchPooler.for_seq_cls(pooler_config, classifier=classifier) |
| 325 | |
| 326 | token_ids = [[2, 99, 3, 99], [4, 99], [5, 6]] |
| 327 | params = [ |
| 328 | PoolingParams( |
| 329 | task="token_classify", |
| 330 | use_activation=False, |
| 331 | step_tag_id=99 if tok_pooling_type == "STEP" else None, |
| 332 | ) |
| 333 | for _ in token_ids |
| 334 | ] |
| 335 | metadata = _make_metadata([4, 2, 2], token_ids=token_ids, pooling_params=params) |
| 336 | hidden = torch.arange(32, dtype=torch.float32).reshape(8, 4) |
| 337 | outputs = pooler(hidden, metadata) |
| 338 | |
| 339 | expected_rows = ( |
| 340 | [[1, 3], [5], []] |
| 341 | if tok_pooling_type == "STEP" |
| 342 | else [[0, 1, 2, 3], [4, 5], [6, 7]] |
| 343 | ) |
| 344 | assert len(outputs) == len(expected_rows) |
| 345 | for output, rows in zip(outputs, expected_rows): |
| 346 | torch.testing.assert_close(output, classifier(hidden[rows])) |
| 347 | assert pooler.get_pooling_updates("token_classify").requires_token_ids == ( |
| 348 | tok_pooling_type == "STEP" |
| 349 | ) |
| 350 | |
| 351 | |
| 352 | class TestAllPool: |
| 353 | @staticmethod |
| 354 | def _make_all_pool( |
| 355 | *, chunked: bool = False, async_scheduling: bool = False |
| 356 | ) -> AllPool: |
| 357 | fake_config = _FakeVllmConfig( |
| 358 | scheduler_config=_FakeSchedulerConfig( |
| 359 | enable_chunked_prefill=chunked, |
| 360 | async_scheduling=async_scheduling, |
| 361 | ), |
| 362 | ) |
| 363 | with patch( |
| 364 | "vllm.model_executor.layers.pooler.tokwise.methods.get_current_vllm_config", |
| 365 | return_value=fake_config, |
| 366 | ): |
| 367 | return AllPool() |
| 368 | |
| 369 | def test_splits_by_sequence(self): |
| 370 | pooler = self._make_all_pool() |
| 371 | hidden = torch.tensor( |
| 372 | [[1.0, 2.0], [3.0, 4.0], [5.0, 6.0], [7.0, 8.0], [9.0, 10.0]] |
| 373 | ) |
| 374 | metadata = _make_metadata([2, 3]) |
| 375 | out = pooler(hidden, metadata) |
| 376 | assert len(out) == 2 |
| 377 | assert torch.equal(out[0], hidden[:2]) |
| 378 | assert torch.equal(out[1], hidden[2:]) |
| 379 | |
| 380 | def test_single_sequence(self): |
| 381 | pooler = self._make_all_pool() |
| 382 | hidden = torch.tensor([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]]) |
| 383 | metadata = _make_metadata([3]) |
| 384 | out = pooler(hidden, metadata) |
| 385 | assert len(out) == 1 |
| 386 | assert torch.equal(out[0], hidden) |
| 387 | |
| 388 | def test_chunked_prefill_returns_none_for_unfinished(self): |
| 389 | pooler = self._make_all_pool(chunked=True) |
| 390 | hidden = torch.tensor([[1.0, 2.0], [3.0, 4.0]]) |
| 391 | metadata = _make_metadata( |
| 392 | [4], |
| 393 | num_scheduled_tokens=[2], |
| 394 | seq_lens=[2], |
| 395 | ) |
| 396 | out = pooler(hidden, metadata) |
| 397 | assert len(out) == 1 |
| 398 | assert out[0] is None |
| 399 | |
| 400 | def test_chunked_prefill_returns_concat_when_finished(self): |
| 401 | pooler = self._make_all_pool(chunked=True) |
| 402 | |
| 403 | chunk1 = torch.tensor([[1.0, 2.0], [3.0, 4.0]]) |
| 404 | metadata1 = _make_metadata( |
| 405 | [4], |
| 406 | num_scheduled_tokens=[2], |
| 407 | seq_lens=[2], |
| 408 | ) |
| 409 | out1 = pooler(chunk1, metadata1) |
| 410 | assert out1[0] is None |
| 411 | expected_chunk1 = chunk1.clone() |
| 412 | chunk1.zero_() |
| 413 | |
| 414 | chunk2 = torch.tensor([[5.0, 6.0], [7.0, 8.0]]) |
| 415 | metadata2 = _make_metadata( |
| 416 | [4], |
| 417 | num_scheduled_tokens=[2], |
| 418 | seq_lens=[4], |
| 419 | ) |
| 420 | metadata2.pooling_states = metadata1.pooling_states |
| 421 | out2 = pooler(chunk2, metadata2) |
| 422 | assert out2[0] is not None |
| 423 | expected = torch.cat([expected_chunk1, chunk2], dim=0) |
| 424 | assert torch.equal(out2[0], expected) |
| 425 | |
| 426 | @pytest.mark.parametrize("async_scheduling", [False, True]) |
| 427 | def test_chunked_prefill_single_shot_matches_non_chunked( |
| 428 | self, async_scheduling: bool |
| 429 | ): |
| 430 | pooler = self._make_all_pool(chunked=True, async_scheduling=async_scheduling) |
| 431 | hidden = torch.tensor( |
| 432 | [[1.0, 2.0], [3.0, 4.0], [5.0, 6.0], [7.0, 8.0], [9.0, 10.0]] |
| 433 | ) |
| 434 | expected = [hidden[:2].clone(), hidden[2:].clone()] |
| 435 | metadata = _make_metadata([2, 3]) |
| 436 | out = pooler(hidden, metadata) |
| 437 | if async_scheduling: |
| 438 | # Finished outputs must own their storage under async scheduling. |
| 439 | hidden.zero_() |
| 440 | assert len(out) == 2 |
| 441 | assert torch.equal(out[0], expected[0]) |
| 442 | assert torch.equal(out[1], expected[1]) |
| 443 | |
| 444 | def test_non_chunked_owns_output_under_async_scheduling(self): |
| 445 | pooler = self._make_all_pool(async_scheduling=True) |
| 446 | hidden = torch.tensor([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]]) |
| 447 | expected = [hidden[:2].clone(), hidden[2:].clone()] |
| 448 | metadata = _make_metadata([2, 1]) |
| 449 | out = pooler(hidden, metadata) |
| 450 | hidden.zero_() |
| 451 | assert len(out) == 2 |
| 452 | assert torch.equal(out[0], expected[0]) |
| 453 | assert torch.equal(out[1], expected[1]) |
| 454 | |
| 455 | def test_chunked_prefill_mixed_finished_unfinished(self): |
| 456 | pooler = self._make_all_pool(chunked=True) |
| 457 | hidden = torch.tensor([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]]) |
| 458 | metadata = _make_metadata( |
| 459 | [2, 4], |
| 460 | num_scheduled_tokens=[2, 1], |
| 461 | seq_lens=[2, 1], |
| 462 | ) |
| 463 | out = pooler(hidden, metadata) |
| 464 | assert len(out) == 2 |
| 465 | assert torch.equal(out[0], hidden[:2]) |
| 466 | assert out[1] is None |
| 467 | |
| 468 | |
| 469 | # --------------------------------------------------------------------------- |
| 470 | # StepPool |
| 471 | # --------------------------------------------------------------------------- |
| 472 | class TestStepPool: |
| 473 | @staticmethod |
| 474 | def _make_step_pool(*, chunked: bool = False) -> StepPool: |
| 475 | fake_config = _FakeVllmConfig( |
| 476 | scheduler_config=_FakeSchedulerConfig( |
| 477 | enable_chunked_prefill=chunked, |
| 478 | ), |
| 479 | ) |
| 480 | with patch( |
| 481 | "vllm.model_executor.layers.pooler.tokwise.methods.get_current_vllm_config", |
| 482 | return_value=fake_config, |
| 483 | ): |
| 484 | return StepPool() |
| 485 | |
| 486 | def test_filters_by_step_tag_id(self): |
| 487 | pooler = self._make_step_pool() |
| 488 | hidden = torch.tensor([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0], [7.0, 8.0]]) |
| 489 | token_ids = [[10, 99, 10, 20]] |
| 490 | params = [PoolingParams(task="token_classify", step_tag_id=10)] |
| 491 | metadata = _make_metadata([4], token_ids=token_ids, pooling_params=params) |
| 492 | out = pooler(hidden, metadata) |
| 493 | assert len(out) == 1 |
| 494 | expected = torch.tensor([[1.0, 2.0], [5.0, 6.0]]) |
| 495 | assert torch.equal(out[0], expected) |
| 496 | |
| 497 | def test_filters_by_returned_token_ids(self): |
| 498 | pooler = self._make_step_pool() |
| 499 | hidden = torch.tensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]) |
| 500 | token_ids = [[10, 20]] |
| 501 | params = [PoolingParams(task="token_classify", returned_token_ids=[0, 2])] |
| 502 | metadata = _make_metadata([2], token_ids=token_ids, pooling_params=params) |
| 503 | out = pooler(hidden, metadata) |
| 504 | assert len(out) == 1 |
| 505 | expected = torch.tensor([[1.0, 3.0], [4.0, 6.0]]) |
| 506 | assert torch.equal(out[0], expected) |
| 507 | |
| 508 | def test_no_filtering_without_params(self): |
| 509 | pooler = self._make_step_pool() |
| 510 | hidden = torch.tensor([[1.0, 2.0], [3.0, 4.0]]) |
| 511 | token_ids = [[10, 20]] |
| 512 | params = [PoolingParams(task="token_classify")] |
| 513 | metadata = _make_metadata([2], token_ids=token_ids, pooling_params=params) |
| 514 | out = pooler(hidden, metadata) |
| 515 | assert len(out) == 1 |
| 516 | assert torch.equal(out[0], hidden) |
| 517 | |
| 518 | def test_combined_step_tag_and_returned_token_ids(self): |
| 519 | pooler = self._make_step_pool() |
| 520 | hidden = torch.tensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 9.0]]) |
| 521 | token_ids = [[99, 10, 99]] |
| 522 | params = [ |
| 523 | PoolingParams( |
| 524 | task="token_classify", |
| 525 | step_tag_id=10, |
| 526 | returned_token_ids=[0, 2], |
| 527 | ) |
| 528 | ] |
| 529 | metadata = _make_metadata([3], token_ids=token_ids, pooling_params=params) |
| 530 | out = pooler(hidden, metadata) |
| 531 | assert len(out) == 1 |
| 532 | expected = torch.tensor([[4.0, 6.0]]) |
| 533 | assert torch.equal(out[0], expected) |
| 534 | |
| 535 | def test_step_tag_id_no_match_returns_empty(self): |
| 536 | pooler = self._make_step_pool() |
| 537 | hidden = torch.tensor([[1.0, 2.0], [3.0, 4.0]]) |
| 538 | token_ids = [[10, 20]] |
| 539 | params = [PoolingParams(task="token_classify", step_tag_id=999)] |
| 540 | metadata = _make_metadata([2], token_ids=token_ids, pooling_params=params) |
| 541 | out = pooler(hidden, metadata) |
| 542 | assert len(out) == 1 |
| 543 | assert out[0].shape == (0, 2) |
| 544 | |
| 545 | def test_chunked_prefill_propagates_none_for_unfinished(self): |
| 546 | pooler = self._make_step_pool(chunked=True) |
| 547 | hidden = torch.tensor([[1.0, 2.0], [3.0, 4.0]]) |
| 548 | token_ids = [[10, 20, 30, 40]] |
| 549 | params = [PoolingParams(task="token_classify", step_tag_id=10)] |
| 550 | metadata = _make_metadata( |
| 551 | [4], |
| 552 | token_ids=token_ids, |
| 553 | pooling_params=params, |
| 554 | num_scheduled_tokens=[2], |
| 555 | seq_lens=[2], |
| 556 | ) |
| 557 | out = pooler(hidden, metadata) |
| 558 | assert len(out) == 1 |
| 559 | assert out[0] is None |
| 560 | |
| 561 | def test_chunked_prefill_filters_when_finished(self): |
| 562 | pooler = self._make_step_pool(chunked=True) |
| 563 | hidden = torch.tensor([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0], [7.0, 8.0]]) |
| 564 | token_ids = [[10, 99, 10, 20]] |
| 565 | params = [PoolingParams(task="token_classify", step_tag_id=10)] |
| 566 | metadata = _make_metadata([4], token_ids=token_ids, pooling_params=params) |
| 567 | out = pooler(hidden, metadata) |
| 568 | assert len(out) == 1 |
| 569 | expected = torch.tensor([[1.0, 2.0], [5.0, 6.0]]) |
| 570 | assert torch.equal(out[0], expected) |
| 571 | |
| 572 | def test_requires_token_ids_update(self): |
| 573 | pooler = self._make_step_pool() |
| 574 | update = pooler.get_pooling_updates("token_classify") |
| 575 | assert update.requires_token_ids is True |
| 576 | |
| 577 | |
| 578 | # --------------------------------------------------------------------------- |
| 579 | # get_tok_pooling_method factory |
| 580 | # --------------------------------------------------------------------------- |
| 581 | class TestGetTokPoolingMethod: |
| 582 | def test_all(self): |
| 583 | fake_config = _FakeVllmConfig( |
| 584 | scheduler_config=_FakeSchedulerConfig( |
| 585 | enable_chunked_prefill=False, |
| 586 | ), |
| 587 | ) |
| 588 | with patch( |
| 589 | "vllm.model_executor.layers.pooler.tokwise.methods.get_current_vllm_config", |
| 590 | return_value=fake_config, |
| 591 | ): |
| 592 | assert isinstance(get_tok_pooling_method("ALL"), AllPool) |
| 593 | |
| 594 | def test_step(self): |
| 595 | fake_config = _FakeVllmConfig( |
| 596 | scheduler_config=_FakeSchedulerConfig( |
| 597 | enable_chunked_prefill=False, |
| 598 | ), |
| 599 | ) |
| 600 | with patch( |
| 601 | "vllm.model_executor.layers.pooler.tokwise.methods.get_current_vllm_config", |
| 602 | return_value=fake_config, |
| 603 | ): |
| 604 | assert isinstance(get_tok_pooling_method("STEP"), StepPool) |
| 605 | |
| 606 | def test_unknown_raises(self): |
| 607 | with pytest.raises(NotImplementedError, match="UNKNOWN"): |
| 608 | get_tok_pooling_method("UNKNOWN") |