processor.py 15 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
5
import importlib
import inspect
6
from functools import lru_cache
7
from typing import TYPE_CHECKING, Any, cast, get_args, get_type_hints
8

9
10
11
12
13
14
from transformers import (
    AutoFeatureExtractor,
    AutoImageProcessor,
    AutoProcessor,
    AutoVideoProcessor,
)
15
16
from transformers.feature_extraction_utils import FeatureExtractionMixin
from transformers.image_processing_utils import BaseImageProcessor
17
from transformers.processing_utils import ProcessorMixin
18
from transformers.video_processing_utils import BaseVideoProcessor
19
20
from typing_extensions import TypeVar

21
22
from vllm.transformers_utils.gguf_utils import is_gguf
from vllm.transformers_utils.utils import convert_model_repo_to_path
23
from vllm.utils.func_utils import get_allowed_kwarg_only_overrides
24

25
if TYPE_CHECKING:
26
    from vllm.config import ModelConfig, RendererConfig
27
28

_P = TypeVar("_P", bound=ProcessorMixin, default=ProcessorMixin)
29
_V = TypeVar("_V", bound=BaseVideoProcessor, default=BaseVideoProcessor)
30
31
32
33
34
35
36
37
38
39
40
41
42


class HashableDict(dict):
    """
    A dictionary that can be hashed by lru_cache.
    """

    # NOTE: pythonic dict is not hashable,
    # we override on it directly for simplicity
    def __hash__(self) -> int:  # type: ignore[override]
        return hash(frozenset(self.items()))


43
44
45
46
47
48
49
50
51
class HashableList(list):
    """
    A list that can be hashed by lru_cache.
    """

    def __hash__(self) -> int:  # type: ignore[override]
        return hash(tuple(self))


52
def _get_processor_factory_fn(processor_cls: type | tuple[type, ...]):
53
54
55
56
57
58
59
    if isinstance(processor_cls, tuple) or processor_cls == ProcessorMixin:
        return AutoProcessor.from_pretrained
    if hasattr(processor_cls, "from_pretrained"):
        return processor_cls.from_pretrained

    return processor_cls

60

61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
@lru_cache
def _collect_dynamic_keys_from_processing_kwargs(kwargs_cls: type) -> set[str]:
    dynamic_kwargs: set[str] = set()
    if kwargs_cls is None:
        return dynamic_kwargs
    # get kwargs annotations in processor
    # merge text_kwargs / images_kwargs / videos_kwargs / audio_kwargs
    kwargs_type_annotations = get_type_hints(kwargs_cls)
    for kw_type in ("text_kwargs", "images_kwargs", "videos_kwargs", "audio_kwargs"):
        if kw_type in kwargs_type_annotations:
            kw_annotations = get_type_hints(kwargs_type_annotations[kw_type])
            for kw_name in kw_annotations:
                dynamic_kwargs.add(kw_name)
    dynamic_kwargs |= {"text_kwargs", "images_kwargs", "videos_kwargs", "audio_kwargs"}
    return dynamic_kwargs


78
79
def _merge_mm_kwargs(
    model_config: "ModelConfig",
80
    processor_cls: type | tuple[type, ...],
81
82
83
84
85
86
87
88
89
90
91
92
93
    /,
    **kwargs,
):
    mm_config = model_config.get_multimodal_config()
    merged_kwargs = mm_config.merge_mm_processor_kwargs(kwargs)

    factory = _get_processor_factory_fn(processor_cls)
    allowed_kwargs = get_allowed_kwarg_only_overrides(
        factory,
        merged_kwargs,
        requires_kw_only=False,
        allow_var_kwargs=True,
    )
94
95
96
    # NOTE: Pythonic dict is not hashable and will raise unhashable type
    # error when calling `cached_get_processor`, therefore we need to
    # wrap it to a hashable dict.
97
    for key, value in allowed_kwargs.items():
98
        if isinstance(value, dict):
99
            allowed_kwargs[key] = HashableDict(value)
100
        if isinstance(value, list):
101
102
103
            allowed_kwargs[key] = HashableList(value)

    return allowed_kwargs
104

105
106
107

def get_processor(
    processor_name: str,
108
    *args: Any,
109
    revision: str | None = None,
110
    trust_remote_code: bool = False,
111
    processor_cls: type[_P] | tuple[type[_P], ...] = ProcessorMixin,
112
    **kwargs: Any,
113
) -> _P:
114
    """Load a processor for the given model name via HuggingFace."""
115
116
    if revision is None:
        revision = "main"
117
    try:
118
        processor_name = convert_model_repo_to_path(processor_name)
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
        if isinstance(processor_cls, tuple) or processor_cls == ProcessorMixin:
            processor = AutoProcessor.from_pretrained(
                processor_name,
                *args,
                revision=revision,
                trust_remote_code=trust_remote_code,
                **kwargs,
            )
        elif issubclass(processor_cls, ProcessorMixin):
            processor = processor_cls.from_pretrained(
                processor_name,
                *args,
                revision=revision,
                trust_remote_code=trust_remote_code,
                **kwargs,
            )
        else:
            # Processors that are standalone classes unrelated to HF
            processor = processor_cls(*args, **kwargs)
138
139
140
141
142
143
144
145
146
147
    except ValueError as e:
        # If the error pertains to the processor class not existing or not
        # currently being imported, suggest using the --trust-remote-code flag.
        # Unlike AutoTokenizer, AutoProcessor does not separate such errors
        if not trust_remote_code:
            err_msg = (
                "Failed to load the processor. If the processor is "
                "a custom processor not yet available in the HuggingFace "
                "transformers library, consider setting "
                "`trust_remote_code=True` in LLM or using the "
148
149
                "`--trust-remote-code` flag in the CLI."
            )
150
151
152
153
            raise RuntimeError(err_msg) from e
        else:
            raise e

154
    if not isinstance(processor, processor_cls):
155
156
157
158
159
        raise TypeError(
            "Invalid type of HuggingFace processor. "
            f"Expected type: {processor_cls}, but "
            f"found type: {type(processor)}"
        )
160
161

    return processor
162
163


164
165
166
cached_get_processor = lru_cache(get_processor)


167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
@lru_cache
def get_processor_kwargs_from_processor(processor: _P) -> set[str]:
    try:
        # get kwargs annotations in processor
        call_kwargs = inspect.signature(type(processor).__call__).parameters.get(
            "kwargs"
        )
        call_kwargs_annotations = call_kwargs.annotation if call_kwargs else None
        # if the processor has explicit kwargs annotation, use it
        if call_kwargs_annotations not in (None, inspect._empty):
            # get_type_hints will parse all type annotations at runtime,
            # and if an annotation refers to a type or
            # name that hasn’t been imported or defined, it will raise an error.
            # So we use __annotations__ to get the raw annotations directly.
            return _collect_dynamic_keys_from_processing_kwargs(
                get_args(call_kwargs_annotations)[0]
            )
        # otherwise, try to get from ProcessingKwargs
        else:
            module_name = type(processor).__module__
            mod = importlib.import_module(module_name)
            # find *ProcessingKwargs in the module
            processor_kwargs: set[str] = set()
            for name, obj in vars(mod).items():
                if name.endswith("ProcessingKwargs"):
                    processor_kwargs = (
                        processor_kwargs
                        | _collect_dynamic_keys_from_processing_kwargs(obj)
                    )
            return processor_kwargs
    except Exception:
        return set()


def cached_get_processor_without_dynamic_kwargs(
    processor_name: str,
    *args: Any,
    revision: str | None = None,
    trust_remote_code: bool = False,
    processor_cls: type[_P] | tuple[type[_P], ...] = ProcessorMixin,
    **kwargs: Any,
) -> _P:
    # Step 1: use default kwargs to get a temporary processor instance
    processor = cached_get_processor(
        processor_name,
        revision=revision,
        trust_remote_code=trust_remote_code,
        processor_cls=processor_cls,  # type: ignore[arg-type]
    )

    # Step 2: use temporary processor collect dynamic keys
    dynamic_keys = get_processor_kwargs_from_processor(processor)

    # Step 3: use dynamic_keys filter kwargs
    filtered_kwargs = {k: v for k, v in kwargs.items() if k not in dynamic_keys}

    # Step 4: use filtered kwargs to get final processor instance
    final_processor = cached_get_processor(
        processor_name,
        revision=revision,
        trust_remote_code=trust_remote_code,
        processor_cls=processor_cls,  # type: ignore[arg-type]
        **filtered_kwargs,
    )

    return final_processor


235
def cached_processor_from_config(
236
    renderer_config: "RendererConfig",
237
    processor_cls: type[_P] | tuple[type[_P], ...] = ProcessorMixin,
238
239
    **kwargs: Any,
) -> _P:
240
    model_config = renderer_config.model_config
241
    if is_gguf(model_config.model):
242
        assert not is_gguf(renderer_config.tokenizer), (
243
244
245
            "For multimodal GGUF models, the original tokenizer "
            "should be used to correctly load processor."
        )
246
247
        model = renderer_config.tokenizer
        revision = renderer_config.tokenizer_revision
248
249
250
251
    else:
        model = model_config.model
        revision = model_config.revision

252
    return cached_get_processor_without_dynamic_kwargs(
253
254
        model,
        revision=revision,
255
256
        trust_remote_code=model_config.trust_remote_code,
        processor_cls=processor_cls,  # type: ignore[arg-type]
257
        **_merge_mm_kwargs(model_config, processor_cls, **kwargs),
258
259
260
    )


261
262
263
def get_feature_extractor(
    processor_name: str,
    *args: Any,
264
    revision: str | None = None,
265
266
267
    trust_remote_code: bool = False,
    **kwargs: Any,
):
268
    """Load an audio feature extractor for the given model name
269
270
    via HuggingFace."""
    try:
271
        processor_name = convert_model_repo_to_path(processor_name)
272
273
274
        feature_extractor = AutoFeatureExtractor.from_pretrained(
            processor_name,
            *args,
275
            revision=revision,
276
            trust_remote_code=trust_remote_code,
277
278
            **kwargs,
        )
279
280
281
282
283
284
285
286
287
288
    except ValueError as e:
        # If the error pertains to the processor class not existing or not
        # currently being imported, suggest using the --trust-remote-code flag.
        # Unlike AutoTokenizer, AutoImageProcessor does not separate such errors
        if not trust_remote_code:
            err_msg = (
                "Failed to load the feature extractor. If the feature "
                "extractor is a custom extractor not yet available in the "
                "HuggingFace transformers library, consider setting "
                "`trust_remote_code=True` in LLM or using the "
289
290
                "`--trust-remote-code` flag in the CLI."
            )
291
292
293
294
295
296
297
298
299
300
            raise RuntimeError(err_msg) from e
        else:
            raise e
    return cast(FeatureExtractionMixin, feature_extractor)


cached_get_feature_extractor = lru_cache(get_feature_extractor)


def cached_feature_extractor_from_config(
301
    renderer_config: "RendererConfig",
302
303
    **kwargs: Any,
):
304
305
    model_config = renderer_config.model_config

306
307
    return cached_get_feature_extractor(
        model_config.model,
308
        revision=model_config.revision,
309
        trust_remote_code=model_config.trust_remote_code,
310
        **_merge_mm_kwargs(model_config, AutoFeatureExtractor, **kwargs),
311
312
313
    )


314
315
316
def get_image_processor(
    processor_name: str,
    *args: Any,
317
    revision: str | None = None,
318
319
320
321
322
    trust_remote_code: bool = False,
    **kwargs: Any,
):
    """Load an image processor for the given model name via HuggingFace."""
    try:
323
        processor_name = convert_model_repo_to_path(processor_name)
324
325
326
        processor = AutoImageProcessor.from_pretrained(
            processor_name,
            *args,
327
            revision=revision,
328
            trust_remote_code=trust_remote_code,
329
330
            **kwargs,
        )
331
332
333
334
335
336
337
338
339
340
    except ValueError as e:
        # If the error pertains to the processor class not existing or not
        # currently being imported, suggest using the --trust-remote-code flag.
        # Unlike AutoTokenizer, AutoImageProcessor does not separate such errors
        if not trust_remote_code:
            err_msg = (
                "Failed to load the image processor. If the image processor is "
                "a custom processor not yet available in the HuggingFace "
                "transformers library, consider setting "
                "`trust_remote_code=True` in LLM or using the "
341
342
                "`--trust-remote-code` flag in the CLI."
            )
343
344
345
346
347
348
349
            raise RuntimeError(err_msg) from e
        else:
            raise e

    return cast(BaseImageProcessor, processor)


350
351
352
353
cached_get_image_processor = lru_cache(get_image_processor)


def cached_image_processor_from_config(
354
    renderer_config: "RendererConfig",
355
356
    **kwargs: Any,
):
357
    model_config = renderer_config.model_config
358
    if is_gguf(model_config.model):
359
        assert not is_gguf(renderer_config.tokenizer), (
360
361
362
            "For multimodal GGUF models, the original tokenizer "
            "should be used to correctly load image processor."
        )
363
364
        model = renderer_config.tokenizer
        revision = renderer_config.tokenizer_revision
365
366
367
    else:
        model = model_config.model
        revision = model_config.revision
368
    return cached_get_image_processor(
369
370
        model,
        revision=revision,
371
        trust_remote_code=model_config.trust_remote_code,
372
        **_merge_mm_kwargs(model_config, AutoImageProcessor, **kwargs),
373
    )
374
375
376
377
378


def get_video_processor(
    processor_name: str,
    *args: Any,
379
    revision: str | None = None,
380
    trust_remote_code: bool = False,
381
    processor_cls_overrides: type[_V] | None = None,
382
383
384
385
    **kwargs: Any,
):
    """Load a video processor for the given model name via HuggingFace."""
    try:
386
        processor_name = convert_model_repo_to_path(processor_name)
387
388
389
390
391
392
        processor_cls = processor_cls_overrides or AutoVideoProcessor
        processor = processor_cls.from_pretrained(
            processor_name,
            *args,
            revision=revision,
            trust_remote_code=trust_remote_code,
393
394
            **kwargs,
        )
395
396
397
398
399
400
401
402
403
404
    except ValueError as e:
        # If the error pertains to the processor class not existing or not
        # currently being imported, suggest using the --trust-remote-code flag.
        # Unlike AutoTokenizer, AutoVideoProcessor does not separate such errors
        if not trust_remote_code:
            err_msg = (
                "Failed to load the video processor. If the video processor is "
                "a custom processor not yet available in the HuggingFace "
                "transformers library, consider setting "
                "`trust_remote_code=True` in LLM or using the "
405
406
                "`--trust-remote-code` flag in the CLI."
            )
407
408
409
410
411
412
413
414
415
416
417
            raise RuntimeError(err_msg) from e
        else:
            raise e

    return cast(BaseVideoProcessor, processor)


cached_get_video_processor = lru_cache(get_video_processor)


def cached_video_processor_from_config(
418
    renderer_config: "RendererConfig",
419
    processor_cls: type[_V] | None = None,
420
421
    **kwargs: Any,
):
422
423
    model_config = renderer_config.model_config

424
425
426
427
428
429
430
    return cached_get_video_processor(
        model_config.model,
        revision=model_config.revision,
        trust_remote_code=model_config.trust_remote_code,
        processor_cls_overrides=processor_cls,  # type: ignore[arg-type]
        **_merge_mm_kwargs(model_config, AutoVideoProcessor, **kwargs),
    )