| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Unit tests for sequence and token pooler head classes.""" |
| 4 | |
| 5 | import torch |
| 6 | import torch.nn as nn |
| 7 | |
| 8 | from vllm.model_executor.layers.pooler.activations import PoolerNormalize |
| 9 | from vllm.model_executor.layers.pooler.seqwise.heads import ( |
| 10 | ClassifierPoolerHead, |
| 11 | EmbeddingPoolerHead, |
| 12 | ) |
| 13 | from vllm.model_executor.layers.pooler.tokwise.heads import ( |
| 14 | TokenClassifierPoolerHead, |
| 15 | TokenEmbeddingPoolerHead, |
| 16 | ) |
| 17 | from vllm.pooling_params import PoolingParams |
| 18 | from vllm.tasks import PoolingTask |
| 19 | from vllm.v1.pool.metadata import PoolingMetadata, PoolingStates |
| 20 | |
| 21 | _HIDDEN = 16 |
| 22 | _BATCH = 3 |
| 23 | |
| 24 | |
| 25 | def _make_params( |
| 26 | n: int, |
| 27 | *, |
| 28 | task: PoolingTask = "embed", |
| 29 | dimensions: int | None = None, |
| 30 | use_activation: bool | None = None, |
| 31 | ) -> list[PoolingParams]: |
| 32 | return [ |
| 33 | PoolingParams(task=task, dimensions=dimensions, use_activation=use_activation) |
| 34 | for _ in range(n) |
| 35 | ] |
| 36 | |
| 37 | |
| 38 | def _make_metadata(pooling_params: list[PoolingParams]) -> PoolingMetadata: |
| 39 | n = len(pooling_params) |
| 40 | return PoolingMetadata( |
| 41 | prompt_lens=torch.ones(n, dtype=torch.long), |
| 42 | prompt_token_ids=None, |
| 43 | prompt_token_ids_cpu=None, |
| 44 | pooling_params=pooling_params, |
| 45 | pooling_states=[PoolingStates() for _ in range(n)], |
| 46 | ) |
| 47 | |
| 48 | |
| 49 | def _linear(in_f: int, out_f: int) -> nn.Linear: |
| 50 | torch.manual_seed(42) |
| 51 | return nn.Linear(in_f, out_f, bias=False) |
| 52 | |
| 53 | |
| 54 | # --------------------------------------------------------------------------- |
| 55 | # EmbeddingPoolerHead |
| 56 | # --------------------------------------------------------------------------- |
| 57 | class TestEmbeddingPoolerHead: |
| 58 | def test_supported_tasks(self): |
| 59 | head = EmbeddingPoolerHead() |
| 60 | assert head.get_supported_tasks() == {"embed"} |
| 61 | |
| 62 | def test_passthrough(self): |
| 63 | head = EmbeddingPoolerHead() |
| 64 | x = torch.randn(_BATCH, _HIDDEN) |
| 65 | meta = _make_metadata(_make_params(_BATCH)) |
| 66 | out = head(x, meta) |
| 67 | assert torch.equal(out, x) |
| 68 | |
| 69 | def test_head_dtype(self): |
| 70 | head = EmbeddingPoolerHead(head_dtype=torch.float16) |
| 71 | x = torch.randn(_BATCH, _HIDDEN) |
| 72 | meta = _make_metadata(_make_params(_BATCH)) |
| 73 | out = head(x, meta) |
| 74 | assert out.dtype == torch.float16 |
| 75 | |
| 76 | def test_projector(self): |
| 77 | proj = _linear(_HIDDEN, 8) |
| 78 | head = EmbeddingPoolerHead(projector=proj) |
| 79 | x = torch.randn(_BATCH, _HIDDEN) |
| 80 | meta = _make_metadata(_make_params(_BATCH)) |
| 81 | out = head(x, meta) |
| 82 | assert out.shape == (_BATCH, 8) |
| 83 | assert torch.allclose(out, proj(x)) |
| 84 | |
| 85 | def test_matryoshka_uniform(self): |
| 86 | head = EmbeddingPoolerHead() |
| 87 | x = torch.randn(_BATCH, _HIDDEN) |
| 88 | params = _make_params(_BATCH, dimensions=4) |
| 89 | meta = _make_metadata(params) |
| 90 | out = head(x, meta) |
| 91 | assert out.shape == (_BATCH, 4) |
| 92 | assert torch.equal(out, x[..., :4]) |
| 93 | |
| 94 | def test_matryoshka_mixed(self): |
| 95 | head = EmbeddingPoolerHead() |
| 96 | x = torch.randn(2, _HIDDEN) |
| 97 | params = [ |
| 98 | PoolingParams(task="embed", dimensions=4), |
| 99 | PoolingParams(task="embed", dimensions=8), |
| 100 | ] |
| 101 | meta = _make_metadata(params) |
| 102 | out = head(x, meta) |
| 103 | assert isinstance(out, list) |
| 104 | assert len(out) == 2 |
| 105 | assert out[0].shape[-1] == 4 |
| 106 | assert out[1].shape[-1] == 8 |
| 107 | |
| 108 | def test_matryoshka_mixed_with_none(self): |
| 109 | head = EmbeddingPoolerHead() |
| 110 | x = torch.randn(2, _HIDDEN) |
| 111 | params = [ |
| 112 | PoolingParams(task="embed", dimensions=4), |
| 113 | PoolingParams(task="embed", dimensions=None), |
| 114 | ] |
| 115 | meta = _make_metadata(params) |
| 116 | out = head(x, meta) |
| 117 | assert isinstance(out, list) |
| 118 | assert out[0].shape[-1] == 4 |
| 119 | assert torch.equal(out[1], x[1]) |
| 120 | |
| 121 | def test_activation_uniform_true(self): |
| 122 | head = EmbeddingPoolerHead(activation=PoolerNormalize()) |
| 123 | x = torch.randn(_BATCH, _HIDDEN) |
| 124 | params = _make_params(_BATCH, use_activation=True) |
| 125 | meta = _make_metadata(params) |
| 126 | out = head(x, meta) |
| 127 | norms = torch.linalg.norm(out, dim=-1) |
| 128 | assert torch.allclose(norms, torch.ones(_BATCH), atol=1e-5) |
| 129 | |
| 130 | def test_activation_uniform_false(self): |
| 131 | head = EmbeddingPoolerHead(activation=PoolerNormalize()) |
| 132 | x = torch.randn(_BATCH, _HIDDEN) |
| 133 | params = _make_params(_BATCH, use_activation=False) |
| 134 | meta = _make_metadata(params) |
| 135 | out = head(x, meta) |
| 136 | assert torch.equal(out, x) |
| 137 | |
| 138 | def test_activation_mixed_flags(self): |
| 139 | head = EmbeddingPoolerHead(activation=PoolerNormalize()) |
| 140 | x = torch.randn(2, _HIDDEN) |
| 141 | params = [ |
| 142 | PoolingParams(task="embed", use_activation=True), |
| 143 | PoolingParams(task="embed", use_activation=False), |
| 144 | ] |
| 145 | meta = _make_metadata(params) |
| 146 | out = head(x, meta) |
| 147 | assert isinstance(out, list) |
| 148 | norm_0 = torch.linalg.norm(out[0], dim=-1) |
| 149 | assert torch.allclose(norm_0, torch.ones(1), atol=1e-5) |
| 150 | assert torch.equal(out[1], x[1]) |
| 151 | |
| 152 | def test_list_input_gets_stacked(self): |
| 153 | head = EmbeddingPoolerHead() |
| 154 | tensors = [torch.randn(_HIDDEN) for _ in range(_BATCH)] |
| 155 | meta = _make_metadata(_make_params(_BATCH)) |
| 156 | out = head(tensors, meta) |
| 157 | assert out.shape == (_BATCH, _HIDDEN) |
| 158 | expected = torch.stack(tensors) |
| 159 | assert torch.equal(out, expected) |
| 160 | |
| 161 | def test_projector_then_matryoshka(self): |
| 162 | proj = _linear(_HIDDEN, 8) |
| 163 | head = EmbeddingPoolerHead(projector=proj) |
| 164 | x = torch.randn(_BATCH, _HIDDEN) |
| 165 | params = _make_params(_BATCH, dimensions=4) |
| 166 | meta = _make_metadata(params) |
| 167 | out = head(x, meta) |
| 168 | assert out.shape == (_BATCH, 4) |
| 169 | assert torch.equal(out, proj(x)[..., :4]) |
| 170 | |
| 171 | def test_matryoshka_then_activation(self): |
| 172 | head = EmbeddingPoolerHead(activation=PoolerNormalize()) |
| 173 | x = torch.randn(_BATCH, _HIDDEN) |
| 174 | params = _make_params(_BATCH, dimensions=4, use_activation=True) |
| 175 | meta = _make_metadata(params) |
| 176 | out = head(x, meta) |
| 177 | assert out.shape == (_BATCH, 4) |
| 178 | norms = torch.linalg.norm(out, dim=-1) |
| 179 | assert torch.allclose(norms, torch.ones(_BATCH), atol=1e-5) |
| 180 | |
| 181 | def test_empty_batch(self): |
| 182 | head = EmbeddingPoolerHead() |
| 183 | x = torch.randn(0, _HIDDEN) |
| 184 | meta = _make_metadata([]) |
| 185 | out = head(x, meta) |
| 186 | assert out.shape == (0, _HIDDEN) |
| 187 | |
| 188 | |
| 189 | # --------------------------------------------------------------------------- |
| 190 | # ClassifierPoolerHead |
| 191 | # --------------------------------------------------------------------------- |
| 192 | class TestClassifierPoolerHead: |
| 193 | def test_supported_tasks(self): |
| 194 | head = ClassifierPoolerHead() |
| 195 | assert head.get_supported_tasks() == {"classify"} |
| 196 | |
| 197 | def test_passthrough(self): |
| 198 | head = ClassifierPoolerHead() |
| 199 | x = torch.randn(_BATCH, _HIDDEN) |
| 200 | meta = _make_metadata(_make_params(_BATCH, task="classify")) |
| 201 | out = head(x, meta) |
| 202 | assert torch.equal(out, x) |
| 203 | |
| 204 | def test_head_dtype(self): |
| 205 | head = ClassifierPoolerHead(head_dtype=torch.float16) |
| 206 | x = torch.randn(_BATCH, _HIDDEN) |
| 207 | meta = _make_metadata(_make_params(_BATCH, task="classify")) |
| 208 | out = head(x, meta) |
| 209 | assert out.dtype == torch.float16 |
| 210 | |
| 211 | def test_classifier(self): |
| 212 | clf = _linear(_HIDDEN, 3) |
| 213 | head = ClassifierPoolerHead(classifier=clf) |
| 214 | x = torch.randn(_BATCH, _HIDDEN) |
| 215 | meta = _make_metadata(_make_params(_BATCH, task="classify")) |
| 216 | out = head(x, meta) |
| 217 | assert out.shape == (_BATCH, 3) |
| 218 | assert torch.allclose(out, clf(x)) |
| 219 | |
| 220 | def test_logit_mean(self): |
| 221 | head = ClassifierPoolerHead(logit_mean=2.0) |
| 222 | x = torch.randn(_BATCH, _HIDDEN) |
| 223 | meta = _make_metadata(_make_params(_BATCH, task="classify")) |
| 224 | out = head(x, meta) |
| 225 | assert torch.allclose(out, x - 2.0) |
| 226 | |
| 227 | def test_logit_sigma(self): |
| 228 | head = ClassifierPoolerHead(logit_sigma=0.5) |
| 229 | x = torch.randn(_BATCH, _HIDDEN) |
| 230 | meta = _make_metadata(_make_params(_BATCH, task="classify")) |
| 231 | out = head(x, meta) |
| 232 | assert torch.allclose(out, x / 0.5) |
| 233 | |
| 234 | def test_platt_scaling_combined(self): |
| 235 | head = ClassifierPoolerHead(logit_mean=1.0, logit_sigma=2.0) |
| 236 | x = torch.randn(_BATCH, _HIDDEN) |
| 237 | meta = _make_metadata(_make_params(_BATCH, task="classify")) |
| 238 | out = head(x, meta) |
| 239 | assert torch.allclose(out, (x - 1.0) / 2.0) |
| 240 | |
| 241 | def test_activation_uniform_true(self): |
| 242 | head = ClassifierPoolerHead(activation=PoolerNormalize()) |
| 243 | x = torch.randn(_BATCH, _HIDDEN) |
| 244 | params = _make_params(_BATCH, task="classify", use_activation=True) |
| 245 | meta = _make_metadata(params) |
| 246 | out = head(x, meta) |
| 247 | norms = torch.linalg.norm(out, dim=-1) |
| 248 | assert torch.allclose(norms, torch.ones(_BATCH), atol=1e-5) |
| 249 | |
| 250 | def test_activation_uniform_false(self): |
| 251 | head = ClassifierPoolerHead(activation=PoolerNormalize()) |
| 252 | x = torch.randn(_BATCH, _HIDDEN) |
| 253 | params = _make_params(_BATCH, task="classify", use_activation=False) |
| 254 | meta = _make_metadata(params) |
| 255 | out = head(x, meta) |
| 256 | assert torch.equal(out, x) |
| 257 | |
| 258 | def test_activation_mixed_flags(self): |
| 259 | head = ClassifierPoolerHead(activation=PoolerNormalize()) |
| 260 | x = torch.randn(2, _HIDDEN) |
| 261 | params = [ |
| 262 | PoolingParams(task="classify", use_activation=True), |
| 263 | PoolingParams(task="classify", use_activation=False), |
| 264 | ] |
| 265 | meta = _make_metadata(params) |
| 266 | out = head(x, meta) |
| 267 | assert isinstance(out, list) |
| 268 | norm_0 = torch.linalg.norm(out[0], dim=-1) |
| 269 | assert torch.allclose(norm_0, torch.ones(1), atol=1e-5) |
| 270 | assert torch.equal(out[1], x[1]) |
| 271 | |
| 272 | def test_list_input_gets_stacked(self): |
| 273 | head = ClassifierPoolerHead() |
| 274 | tensors = [torch.randn(_HIDDEN) for _ in range(_BATCH)] |
| 275 | meta = _make_metadata(_make_params(_BATCH, task="classify")) |
| 276 | out = head(tensors, meta) |
| 277 | assert out.shape == (_BATCH, _HIDDEN) |
| 278 | expected = torch.stack(tensors) |
| 279 | assert torch.equal(out, expected) |
| 280 | |
| 281 | def test_classifier_then_platt_scaling(self): |
| 282 | clf = _linear(_HIDDEN, 3) |
| 283 | head = ClassifierPoolerHead(classifier=clf, logit_mean=1.0, logit_sigma=2.0) |
| 284 | x = torch.randn(_BATCH, _HIDDEN) |
| 285 | meta = _make_metadata(_make_params(_BATCH, task="classify")) |
| 286 | out = head(x, meta) |
| 287 | expected = (clf(x) - 1.0) / 2.0 |
| 288 | assert torch.allclose(out, expected) |
| 289 | |
| 290 | def test_empty_batch(self): |
| 291 | head = ClassifierPoolerHead() |
| 292 | x = torch.randn(0, _HIDDEN) |
| 293 | meta = _make_metadata([]) |
| 294 | out = head(x, meta) |
| 295 | assert out.shape == (0, _HIDDEN) |
| 296 | |
| 297 | |
| 298 | # --------------------------------------------------------------------------- |
| 299 | # TokenEmbeddingPoolerHead |
| 300 | # --------------------------------------------------------------------------- |
| 301 | class TestTokenEmbeddingPoolerHead: |
| 302 | def test_supported_tasks(self): |
| 303 | head = TokenEmbeddingPoolerHead() |
| 304 | assert head.get_supported_tasks() == {"token_embed"} |
| 305 | |
| 306 | def test_passthrough(self): |
| 307 | head = TokenEmbeddingPoolerHead() |
| 308 | x = torch.randn(5, _HIDDEN) |
| 309 | param = PoolingParams(task="token_embed") |
| 310 | out = head.forward_chunk(x, param) |
| 311 | assert torch.equal(out, x) |
| 312 | |
| 313 | def test_none_chunked_prefill(self): |
| 314 | head = TokenEmbeddingPoolerHead() |
| 315 | param = PoolingParams(task="token_embed") |
| 316 | out = head.forward_chunk(None, param) |
| 317 | assert out is None |
| 318 | |
| 319 | def test_head_dtype(self): |
| 320 | head = TokenEmbeddingPoolerHead(head_dtype=torch.float16) |
| 321 | x = torch.randn(5, _HIDDEN) |
| 322 | param = PoolingParams(task="token_embed") |
| 323 | out = head.forward_chunk(x, param) |
| 324 | assert out is not None |
| 325 | assert out.dtype == torch.float16 |
| 326 | |
| 327 | def test_projector(self): |
| 328 | proj = _linear(_HIDDEN, 8) |
| 329 | head = TokenEmbeddingPoolerHead(projector=proj) |
| 330 | x = torch.randn(5, _HIDDEN) |
| 331 | param = PoolingParams(task="token_embed") |
| 332 | out = head.forward_chunk(x, param) |
| 333 | assert out is not None |
| 334 | assert out.shape == (5, 8) |
| 335 | assert torch.allclose(out, proj(x)) |
| 336 | |
| 337 | def test_matryoshka_truncation(self): |
| 338 | head = TokenEmbeddingPoolerHead() |
| 339 | x = torch.randn(5, _HIDDEN) |
| 340 | param = PoolingParams(task="token_embed", dimensions=4) |
| 341 | out = head.forward_chunk(x, param) |
| 342 | assert out is not None |
| 343 | assert out.shape == (5, 4) |
| 344 | assert torch.equal(out, x[..., :4]) |
| 345 | |
| 346 | def test_activation_true(self): |
| 347 | head = TokenEmbeddingPoolerHead(activation=PoolerNormalize()) |
| 348 | x = torch.randn(5, _HIDDEN) |
| 349 | param = PoolingParams(task="token_embed", use_activation=True) |
| 350 | out = head.forward_chunk(x, param) |
| 351 | norms = torch.linalg.norm(out, dim=-1) |
| 352 | assert torch.allclose(norms, torch.ones(5), atol=1e-5) |
| 353 | |
| 354 | def test_activation_false(self): |
| 355 | head = TokenEmbeddingPoolerHead(activation=PoolerNormalize()) |
| 356 | x = torch.randn(5, _HIDDEN) |
| 357 | param = PoolingParams(task="token_embed", use_activation=False) |
| 358 | out = head.forward_chunk(x, param) |
| 359 | assert torch.equal(out, x) |
| 360 | |
| 361 | def test_projector_then_matryoshka(self): |
| 362 | proj = _linear(_HIDDEN, 8) |
| 363 | head = TokenEmbeddingPoolerHead(projector=proj) |
| 364 | x = torch.randn(5, _HIDDEN) |
| 365 | param = PoolingParams(task="token_embed", dimensions=4) |
| 366 | out = head.forward_chunk(x, param) |
| 367 | assert out is not None |
| 368 | assert out.shape == (5, 4) |
| 369 | assert torch.equal(out, proj(x)[..., :4]) |
| 370 | |
| 371 | def test_matryoshka_then_activation(self): |
| 372 | head = TokenEmbeddingPoolerHead(activation=PoolerNormalize()) |
| 373 | x = torch.randn(5, _HIDDEN) |
| 374 | param = PoolingParams(task="token_embed", dimensions=4, use_activation=True) |
| 375 | out = head.forward_chunk(x, param) |
| 376 | assert out is not None |
| 377 | assert out.shape == (5, 4) |
| 378 | norms = torch.linalg.norm(out, dim=-1) |
| 379 | assert torch.allclose(norms, torch.ones(5), atol=1e-5) |
| 380 | |
| 381 | def test_forward_mixed_batch_chunked_prefill(self): |
| 382 | head = TokenEmbeddingPoolerHead() |
| 383 | pooled_data = [torch.randn(5, _HIDDEN), None, torch.randn(3, _HIDDEN)] |
| 384 | params = _make_params(3, task="token_embed") |
| 385 | meta = _make_metadata(params) |
| 386 | out = head(pooled_data, meta) |
| 387 | assert len(out) == 3 |
| 388 | assert torch.equal(out[0], pooled_data[0]) |
| 389 | assert out[1] is None |
| 390 | assert torch.equal(out[2], pooled_data[2]) |
| 391 | |
| 392 | def test_forward_empty_batch(self): |
| 393 | head = TokenEmbeddingPoolerHead() |
| 394 | meta = _make_metadata([]) |
| 395 | out = head([], meta) |
| 396 | assert out == [] |
| 397 | |
| 398 | |
| 399 | # --------------------------------------------------------------------------- |
| 400 | # TokenClassifierPoolerHead |
| 401 | # --------------------------------------------------------------------------- |
| 402 | class TestTokenClassifierPoolerHead: |
| 403 | def test_supported_tasks(self): |
| 404 | head = TokenClassifierPoolerHead() |
| 405 | assert head.get_supported_tasks() == {"token_classify"} |
| 406 | |
| 407 | def test_passthrough(self): |
| 408 | head = TokenClassifierPoolerHead() |
| 409 | x = torch.randn(5, _HIDDEN) |
| 410 | param = PoolingParams(task="token_classify") |
| 411 | out = head.forward_chunk(x, param) |
| 412 | assert torch.equal(out, x) |
| 413 | |
| 414 | def test_none_chunked_prefill(self): |
| 415 | head = TokenClassifierPoolerHead() |
| 416 | param = PoolingParams(task="token_classify") |
| 417 | out = head.forward_chunk(None, param) |
| 418 | assert out is None |
| 419 | |
| 420 | def test_head_dtype(self): |
| 421 | head = TokenClassifierPoolerHead(head_dtype=torch.float16) |
| 422 | x = torch.randn(5, _HIDDEN) |
| 423 | param = PoolingParams(task="token_classify") |
| 424 | out = head.forward_chunk(x, param) |
| 425 | assert out is not None |
| 426 | assert out.dtype == torch.float16 |
| 427 | |
| 428 | def test_classifier(self): |
| 429 | clf = _linear(_HIDDEN, 3) |
| 430 | head = TokenClassifierPoolerHead(classifier=clf) |
| 431 | x = torch.randn(5, _HIDDEN) |
| 432 | param = PoolingParams(task="token_classify") |
| 433 | out = head.forward_chunk(x, param) |
| 434 | assert out is not None |
| 435 | assert out.shape == (5, 3) |
| 436 | assert torch.allclose(out, clf(x)) |
| 437 | |
| 438 | def test_logit_mean(self): |
| 439 | head = TokenClassifierPoolerHead(logit_mean=2.0) |
| 440 | x = torch.randn(5, _HIDDEN) |
| 441 | param = PoolingParams(task="token_classify") |
| 442 | out = head.forward_chunk(x, param) |
| 443 | assert torch.allclose(out, x - 2.0) |
| 444 | |
| 445 | def test_logit_sigma(self): |
| 446 | head = TokenClassifierPoolerHead(logit_sigma=0.5) |
| 447 | x = torch.randn(5, _HIDDEN) |
| 448 | param = PoolingParams(task="token_classify") |
| 449 | out = head.forward_chunk(x, param) |
| 450 | assert torch.allclose(out, x / 0.5) |
| 451 | |
| 452 | def test_platt_scaling_combined(self): |
| 453 | head = TokenClassifierPoolerHead(logit_mean=1.0, logit_sigma=2.0) |
| 454 | x = torch.randn(5, _HIDDEN) |
| 455 | param = PoolingParams(task="token_classify") |
| 456 | out = head.forward_chunk(x, param) |
| 457 | assert torch.allclose(out, (x - 1.0) / 2.0) |
| 458 | |
| 459 | def test_activation_true(self): |
| 460 | head = TokenClassifierPoolerHead(activation=PoolerNormalize()) |
| 461 | x = torch.randn(5, _HIDDEN) |
| 462 | param = PoolingParams(task="token_classify", use_activation=True) |
| 463 | out = head.forward_chunk(x, param) |
| 464 | norms = torch.linalg.norm(out, dim=-1) |
| 465 | assert torch.allclose(norms, torch.ones(5), atol=1e-5) |
| 466 | |
| 467 | def test_activation_false(self): |
| 468 | head = TokenClassifierPoolerHead(activation=PoolerNormalize()) |
| 469 | x = torch.randn(5, _HIDDEN) |
| 470 | param = PoolingParams(task="token_classify", use_activation=False) |
| 471 | out = head.forward_chunk(x, param) |
| 472 | assert torch.equal(out, x) |
| 473 | |
| 474 | def test_forward_mixed_batch_chunked_prefill(self): |
| 475 | head = TokenClassifierPoolerHead() |
| 476 | pooled_data = [torch.randn(5, _HIDDEN), None, torch.randn(3, _HIDDEN)] |
| 477 | params = _make_params(3, task="token_classify") |
| 478 | meta = _make_metadata(params) |
| 479 | out = head(pooled_data, meta) |
| 480 | assert len(out) == 3 |
| 481 | assert torch.equal(out[0], pooled_data[0]) |
| 482 | assert out[1] is None |
| 483 | assert torch.equal(out[2], pooled_data[2]) |
| 484 | |
| 485 | def test_forward_empty_batch(self): |
| 486 | head = TokenClassifierPoolerHead() |
| 487 | meta = _make_metadata([]) |
| 488 | out = head([], meta) |
| 489 | assert out == [] |