Skip to content

vllm.transformers_utils.processor

_P module-attribute 

_P = TypeVar(
    "_P", bound=ProcessorMixin, default=ProcessorMixin
)

_V module-attribute 

_V = TypeVar(
    "_V",
    bound=BaseVideoProcessor,
    default=BaseVideoProcessor,
)

cached_get_feature_extractor module-attribute 

cached_get_feature_extractor = lru_cache(
    get_feature_extractor
)

cached_get_image_processor module-attribute 

cached_get_image_processor = lru_cache(get_image_processor)

cached_get_processor module-attribute 

cached_get_processor = lru_cache(get_processor)

cached_get_video_processor module-attribute 

cached_get_video_processor = lru_cache(get_video_processor)

HashableDict

Bases: dict

A dictionary that can be hashed by lru_cache.

Source code in vllm/transformers_utils/processor.py 
29
30
31
32
33
34
35
36
37
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()))

__hash__ 

__hash__() -> int
Source code in vllm/transformers_utils/processor.py 
36
37
def __hash__(self) -> int:  # type: ignore[override]
    return hash(frozenset(self.items()))

HashableList

Bases: list

A list that can be hashed by lru_cache.

Source code in vllm/transformers_utils/processor.py 
40
41
42
43
44
45
46
class HashableList(list):
    """
    A list that can be hashed by lru_cache.
    """

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

__hash__ 

__hash__() -> int
Source code in vllm/transformers_utils/processor.py 
45
46
def __hash__(self) -> int:  # type: ignore[override]
    return hash(tuple(self))

_get_processor_factory_fn 

_get_processor_factory_fn(
    processor_cls: type | tuple[type, ...],
)
Source code in vllm/transformers_utils/processor.py 
49
50
51
52
53
54
55
def _get_processor_factory_fn(processor_cls: type | tuple[type, ...]):
    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

_merge_mm_kwargs 

_merge_mm_kwargs(
    model_config: ModelConfig,
    processor_cls: type | tuple[type, ...],
    /,
    **kwargs,
)
Source code in vllm/transformers_utils/processor.py 
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
def _merge_mm_kwargs(
    model_config: "ModelConfig",
    processor_cls: type | tuple[type, ...],
    /,
    **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,
    )

    # 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.
    for key, value in allowed_kwargs.items():
        if isinstance(value, dict):
            allowed_kwargs[key] = HashableDict(value)
        if isinstance(value, list):
            allowed_kwargs[key] = HashableList(value)

    return allowed_kwargs

cached_feature_extractor_from_config 

cached_feature_extractor_from_config(
    model_config: ModelConfig, **kwargs: Any
)
Source code in vllm/transformers_utils/processor.py 
201
202
203
204
205
206
207
208
209
210
def cached_feature_extractor_from_config(
    model_config: "ModelConfig",
    **kwargs: Any,
):
    return cached_get_feature_extractor(
        model_config.model,
        revision=model_config.revision,
        trust_remote_code=model_config.trust_remote_code,
        **_merge_mm_kwargs(model_config, AutoFeatureExtractor, **kwargs),
    )

cached_image_processor_from_config 

cached_image_processor_from_config(
    model_config: ModelConfig, **kwargs: Any
)
Source code in vllm/transformers_utils/processor.py 
252
253
254
255
256
257
258
259
260
261
def cached_image_processor_from_config(
    model_config: "ModelConfig",
    **kwargs: Any,
):
    return cached_get_image_processor(
        model_config.model,
        revision=model_config.revision,
        trust_remote_code=model_config.trust_remote_code,
        **_merge_mm_kwargs(model_config, AutoImageProcessor, **kwargs),
    )

cached_processor_from_config 

cached_processor_from_config(
    model_config: ModelConfig,
    processor_cls: type[_P]
    | tuple[type[_P], ...] = ProcessorMixin,
    **kwargs: Any,
) -> _P
Source code in vllm/transformers_utils/processor.py 
148
149
150
151
152
153
154
155
156
157
158
159
def cached_processor_from_config(
    model_config: "ModelConfig",
    processor_cls: type[_P] | tuple[type[_P], ...] = ProcessorMixin,
    **kwargs: Any,
) -> _P:
    return cached_get_processor(
        model_config.model,
        revision=model_config.revision,
        trust_remote_code=model_config.trust_remote_code,
        processor_cls=processor_cls,  # type: ignore[arg-type]
        **_merge_mm_kwargs(model_config, processor_cls, **kwargs),
    )

cached_video_processor_from_config 

cached_video_processor_from_config(
    model_config: ModelConfig,
    processor_cls: type[_V] | None = None,
    **kwargs: Any,
)
Source code in vllm/transformers_utils/processor.py 
305
306
307
308
309
310
311
312
313
314
315
316
def cached_video_processor_from_config(
    model_config: "ModelConfig",
    processor_cls: type[_V] | None = None,
    **kwargs: Any,
):
    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),
    )

get_feature_extractor 

get_feature_extractor(
    processor_name: str,
    *args: Any,
    revision: str | None = None,
    trust_remote_code: bool = False,
    **kwargs: Any,
)

Load an audio feature extractor for the given model name via HuggingFace.

Source code in vllm/transformers_utils/processor.py 
162
163
164
165
166
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
def get_feature_extractor(
    processor_name: str,
    *args: Any,
    revision: str | None = None,
    trust_remote_code: bool = False,
    **kwargs: Any,
):
    """Load an audio feature extractor for the given model name
    via HuggingFace."""
    try:
        processor_name = convert_model_repo_to_path(processor_name)
        feature_extractor = AutoFeatureExtractor.from_pretrained(
            processor_name,
            *args,
            revision=revision,
            trust_remote_code=trust_remote_code,
            **kwargs,
        )
    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 "
                "`--trust-remote-code` flag in the CLI."
            )
            raise RuntimeError(err_msg) from e
        else:
            raise e
    return cast(FeatureExtractionMixin, feature_extractor)

get_image_processor 

get_image_processor(
    processor_name: str,
    *args: Any,
    revision: str | None = None,
    trust_remote_code: bool = False,
    **kwargs: Any,
)

Load an image processor for the given model name via HuggingFace.

Source code in vllm/transformers_utils/processor.py 
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
def get_image_processor(
    processor_name: str,
    *args: Any,
    revision: str | None = None,
    trust_remote_code: bool = False,
    **kwargs: Any,
):
    """Load an image processor for the given model name via HuggingFace."""
    try:
        processor_name = convert_model_repo_to_path(processor_name)
        processor = AutoImageProcessor.from_pretrained(
            processor_name,
            *args,
            revision=revision,
            trust_remote_code=trust_remote_code,
            **kwargs,
        )
    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 "
                "`--trust-remote-code` flag in the CLI."
            )
            raise RuntimeError(err_msg) from e
        else:
            raise e

    return cast(BaseImageProcessor, processor)

get_processor 

get_processor(
    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

Load a processor for the given model name via HuggingFace.

Source code in vllm/transformers_utils/processor.py 
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
def get_processor(
    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:
    """Load a processor for the given model name via HuggingFace."""
    if revision is None:
        revision = "main"
    try:
        processor_name = convert_model_repo_to_path(processor_name)
        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)
    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 "
                "`--trust-remote-code` flag in the CLI."
            )
            raise RuntimeError(err_msg) from e
        else:
            raise e

    if not isinstance(processor, processor_cls):
        raise TypeError(
            "Invalid type of HuggingFace processor. "
            f"Expected type: {processor_cls}, but "
            f"found type: {type(processor)}"
        )

    return processor

get_video_processor 

get_video_processor(
    processor_name: str,
    *args: Any,
    revision: str | None = None,
    trust_remote_code: bool = False,
    processor_cls_overrides: type[_V] | None = None,
    **kwargs: Any,
)

Load a video processor for the given model name via HuggingFace.

Source code in vllm/transformers_utils/processor.py 
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
def get_video_processor(
    processor_name: str,
    *args: Any,
    revision: str | None = None,
    trust_remote_code: bool = False,
    processor_cls_overrides: type[_V] | None = None,
    **kwargs: Any,
):
    """Load a video processor for the given model name via HuggingFace."""
    try:
        processor_name = convert_model_repo_to_path(processor_name)
        processor_cls = processor_cls_overrides or AutoVideoProcessor
        processor = processor_cls.from_pretrained(
            processor_name,
            *args,
            revision=revision,
            trust_remote_code=trust_remote_code,
            **kwargs,
        )
    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 "
                "`--trust-remote-code` flag in the CLI."
            )
            raise RuntimeError(err_msg) from e
        else:
            raise e

    return cast(BaseVideoProcessor, processor)