registry.py 10.8 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
from collections.abc import Mapping
3
from dataclasses import dataclass
4
from typing import TYPE_CHECKING, Generic, Optional, Protocol, TypeVar
5

6
import torch.nn as nn
7
from typing_extensions import deprecated
8

9
from vllm.envs import VLLM_MM_INPUT_CACHE_GIB
10
from vllm.inputs import InputProcessingContext
11
from vllm.logger import init_logger
12
13
from vllm.transformers_utils.tokenizer import (AnyTokenizer,
                                               cached_tokenizer_from_config)
14
from vllm.utils import ClassRegistry
15

16
17
from .processing import (BaseMultiModalProcessor, BaseProcessingInfo,
                         ProcessingCache)
18
19
from .profiling import (BaseDummyInputsBuilder, DummyDecoderData,
                        DummyEncoderData, MultiModalProfiler)
20

21
22
23
if TYPE_CHECKING:
    from vllm.config import ModelConfig

24
25
logger = init_logger(__name__)

26
N = TypeVar("N", bound=type[nn.Module])
27
28
_I = TypeVar("_I", bound=BaseProcessingInfo)
_I_co = TypeVar("_I_co", bound=BaseProcessingInfo, covariant=True)
29
30


31
class ProcessingInfoFactory(Protocol[_I_co]):
32
33
34
35
36
    """Constructs a :class:`MultiModalProcessor` instance from the context."""

    def __call__(
        self,
        ctx: InputProcessingContext,
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
    ) -> _I_co:
        ...


class DummyInputsBuilderFactory(Protocol[_I]):
    """
    Constructs a :class:`BaseDummyInputsBuilder` instance from the context.
    """

    def __call__(self, info: _I) -> BaseDummyInputsBuilder[_I]:
        ...


class MultiModalProcessorFactory(Protocol[_I]):
    """Constructs a :class:`MultiModalProcessor` instance from the context."""

    def __call__(
        self,
        info: _I,
        dummy_inputs: BaseDummyInputsBuilder[_I],
57
58
        *,
        cache: Optional[ProcessingCache] = None,
59
    ) -> BaseMultiModalProcessor[_I]:
60
        ...
61

62

63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
@dataclass(frozen=True)
class _ProcessorFactories(Generic[_I]):
    info: ProcessingInfoFactory[_I]
    processor: MultiModalProcessorFactory[_I]
    dummy_inputs: DummyInputsBuilderFactory[_I]

    def build_processor(
        self,
        ctx: InputProcessingContext,
        *,
        cache: Optional[ProcessingCache] = None,
    ):
        info = self.info(ctx)
        dummy_inputs_builder = self.dummy_inputs(info)
        return self.processor(info, dummy_inputs_builder, cache=cache)


80
81
class MultiModalRegistry:
    """
82
    A registry that dispatches data processing according to the model.
83
84
    """

85
    def __init__(self) -> None:
86
        self._processor_factories = ClassRegistry[nn.Module,
87
                                                  _ProcessorFactories]()
88

89
        self._processing_cache = ProcessingCache(VLLM_MM_INPUT_CACHE_GIB)
90

91
92
93
94
    @deprecated("Legacy input processor/mapper pipeline has been removed. "
                "Please update your model runner to use "
                "`seq_group_metadata.multi_modal_data` directly without "
                "further processing.")
95
    def create_input_mapper(self, model_config: "ModelConfig"):
96
        return lambda data, mm_processor_kwargs: data
97

98
99
100
101
102
    def get_max_tokens_per_item_by_modality(
        self,
        model_config: "ModelConfig",
    ) -> Mapping[str, int]:
        """
103
        Get the maximum number of tokens per data item from each modality based
104
        on underlying model configuration.
105
        """
106
107
        if not model_config.is_multimodal_model:
            return {}
108

109
110
111
112
113
114
115
116
117
118
119
120
121
        processor = self.create_processor(model_config, disable_cache=True)
        profiler = MultiModalProfiler(processor)

        seq_len = model_config.max_model_len
        mm_limits = self.get_mm_limits_per_prompt(model_config)

        return profiler.get_mm_max_tokens(
            seq_len,
            {
                modality: 1
                for modality, limit in mm_limits.items() if limit > 0
            },
        )
122

123
124
125
126
127
128
    def get_max_tokens_per_item_by_nonzero_modality(
        self,
        model_config: "ModelConfig",
    ) -> Mapping[str, int]:
        """
        Get the maximum number of tokens per data item from each modality based
129
        on underlying model configuration, excluding modalities that user
130
131
132
        explicitly disabled via `limit_mm_per_prompt`.

        Note:
133
            This is currently directly used only in V1 for profiling the memory
134
135
            usage of a model.
        """
136
        mm_limits = self.get_mm_limits_per_prompt(model_config)
137
138
139
140
141

        return {
            key: max_tokens_per_mm_item
            for key, max_tokens_per_mm_item in
            self.get_max_tokens_per_item_by_modality(model_config).items()
142
            if mm_limits[key] > 0
143
144
        }

145
146
147
148
    def get_max_tokens_by_modality(
        self,
        model_config: "ModelConfig",
    ) -> Mapping[str, int]:
149
        """
150
        Get the maximum number of tokens from each modality
151
        for profiling the memory usage of a model.
152

153
        See :meth:`MultiModalPlugin.get_max_multimodal_tokens` for more details.
154
        """
155
        mm_limits = self.get_mm_limits_per_prompt(model_config)
156

157
        return {
158
            key: mm_limits[key] * max_tokens_per_mm_item
159
160
            for key, max_tokens_per_mm_item in
            self.get_max_tokens_per_item_by_modality(model_config).items()
161
162
163
164
165
166
167
168
169
170
        }

    def get_max_multimodal_tokens(self, model_config: "ModelConfig") -> int:
        """
        Get the maximum number of multi-modal tokens
        for profiling the memory usage of a model.

        See :meth:`MultiModalPlugin.get_max_multimodal_tokens` for more details.
        """
        return sum(self.get_max_tokens_by_modality(model_config).values())
171

172
173
174
175
    @deprecated("Legacy input processor/mapper pipeline has been removed. "
                "Please update your model runner to use "
                "`seq_group_metadata.multi_modal_data` directly without "
                "further processing.")
176
177
    def init_mm_limits_per_prompt(
        self,
178
        model_config: "ModelConfig",
179
    ) -> None:
180
        pass
181
182
183

    def get_mm_limits_per_prompt(
        self,
184
        model_config: "ModelConfig",
185
186
187
188
    ) -> Mapping[str, int]:
        """
        Get the maximum number of multi-modal input instances for each modality
        that are allowed per prompt for a model class.
189
        """
190
191
        if not model_config.is_multimodal_model:
            return {}
192

193
194
195
        processor = self.create_processor(model_config, disable_cache=True)
        profiler = MultiModalProfiler(processor)
        return profiler.get_mm_limits()
196
197
198

    def register_processor(
        self,
199
200
201
202
        processor: MultiModalProcessorFactory[_I],
        *,
        info: ProcessingInfoFactory[_I],
        dummy_inputs: DummyInputsBuilderFactory[_I],
203
204
    ):
        """
205
206
        Register a multi-modal processor to a model class. The processor
        is constructed lazily, hence a factory method should be passed.
207
208
209
210
211

        When the model receives multi-modal data, the provided function is
        invoked to transform the data into a dictionary of model inputs.

        See also:
212
            :ref:`mm-processing`
213
214
215
        """

        def wrapper(model_cls: N) -> N:
216
            if self._processor_factories.contains(model_cls, strict=True):
217
                logger.warning(
218
                    "Model class %s already has a multi-modal processor "
219
220
221
                    "registered to %s. It is overwritten by the new one.",
                    model_cls, self)

222
223
224
225
226
            self._processor_factories[model_cls] = _ProcessorFactories(
                info=info,
                dummy_inputs=dummy_inputs,
                processor=processor,
            )
227
228
229
230
231

            return model_cls

        return wrapper

232
    def _get_model_cls(self, model_config: "ModelConfig"):
233
234
235
236
        # Avoid circular import
        from vllm.model_executor.model_loader import get_model_architecture

        model_cls, _ = get_model_architecture(model_config)
237
238
        return model_cls

239
240
241
242
    @deprecated("Legacy input processor/mapper pipeline has been removed. "
                "Please update your model runner to use "
                "`seq_group_metadata.multi_modal_data` directly without "
                "further processing.")
243
    def has_processor(self, model_config: "ModelConfig") -> bool:
244
        return True
245
246
247
248

    def create_processor(
        self,
        model_config: "ModelConfig",
249
        *,
250
        tokenizer: Optional[AnyTokenizer] = None,
251
        disable_cache: Optional[bool] = None,
252
    ) -> BaseMultiModalProcessor[BaseProcessingInfo]:
253
254
        """
        Create a multi-modal processor for a specific model and tokenizer.
255
256
257

        See also:
            :ref:`mm-processing`
258
        """
259
260
261
        if not model_config.is_multimodal_model:
            raise ValueError(f"{model_config.model} is not a multimodal model")

262
263
        if tokenizer is None:
            tokenizer = cached_tokenizer_from_config(model_config)
264
        if disable_cache is None:
265
266
            mm_config = model_config.get_multimodal_config()
            disable_cache = mm_config.disable_mm_preprocessor_cache
267

268
        model_cls = self._get_model_cls(model_config)
269
        factories = self._processor_factories[model_cls]
270
271

        ctx = InputProcessingContext(model_config, tokenizer)
272
        cache = None if disable_cache else self._processing_cache
273

274
        return factories.build_processor(ctx, cache=cache)
275
276
277
278
279

    def get_decoder_dummy_data(
        self,
        model_config: "ModelConfig",
        seq_len: int,
280
        mm_counts: Optional[Mapping[str, int]] = None,
281
282
283
284
285
286
287
288
    ) -> DummyDecoderData:
        """
        Create dummy data for profiling the memory usage of a model.

        The model is identified by ``model_config``.
        """
        processor = self.create_processor(model_config, disable_cache=True)
        profiler = MultiModalProfiler(processor)
289
        dummy_data = profiler.get_decoder_dummy_data(seq_len, mm_counts)
290
291
292
293
294
295
296
297
298
299
300
301
302
303

        # Having more tokens is over-conservative but otherwise fine
        token_ids = dummy_data.prompt_token_ids
        if len(token_ids) < seq_len:
            raise AssertionError(
                f"Expected at least {seq_len} dummy tokens for profiling, "
                f"but found {len(token_ids)} tokens instead.")

        return dummy_data

    def get_encoder_dummy_data(
        self,
        model_config: "ModelConfig",
        seq_len: int,
304
        mm_counts: Optional[Mapping[str, int]] = None,
305
306
307
308
309
310
311
312
    ) -> DummyEncoderData:
        """
        Create dummy data for profiling the memory usage of a model.

        The model is identified by ``model_config``.
        """
        processor = self.create_processor(model_config, disable_cache=True)
        profiler = MultiModalProfiler(processor)
313
        dummy_data = profiler.get_encoder_dummy_data(seq_len, mm_counts)
314
315
316
317
318

        # Having more tokens is over-conservative but otherwise fine
        token_ids = dummy_data.prompt_token_ids
        if len(token_ids) < seq_len:
            logger.warning_once(
319
320
321
322
                "Expected at least %d dummy encoder tokens for profiling, but found %d tokens instead.",  # noqa: E501
                seq_len,
                len(token_ids),
            )
323
324

        return dummy_data