registry.py 24 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
3
4
5
"""
Whenever you add an architecture to this page, please also update
`tests/models/registry.py` with example HuggingFace models for it.
"""
6
import importlib
7
import os
8
import pickle
9
10
import subprocess
import sys
11
import tempfile
12
13
14
from abc import ABC, abstractmethod
from dataclasses import dataclass, field
from functools import lru_cache
15
16
from typing import (AbstractSet, Callable, Dict, List, Optional, Tuple, Type,
                    TypeVar, Union)
17

18
import cloudpickle
19
20
21
import torch.nn as nn

from vllm.logger import init_logger
22
from vllm.utils import is_in_doc_build
23

24
25
26
27
from .interfaces import (has_inner_state, has_noops, is_attention_free,
                         is_hybrid, supports_cross_encoding,
                         supports_multimodal, supports_pp,
                         supports_transcription, supports_v0_only)
28
from .interfaces_base import is_text_generation_model
29
30
31

logger = init_logger(__name__)

32
# yapf: disable
33
34
_TEXT_GENERATION_MODELS = {
    # [Decoder-only]
35
36
37
    "AquilaModel": ("llama", "LlamaForCausalLM"),
    "AquilaForCausalLM": ("llama", "LlamaForCausalLM"),  # AquilaChat2
    "ArcticForCausalLM": ("arctic", "ArcticForCausalLM"),
38
    "MiniMaxText01ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
39
40
41
42
    # baichuan-7b, upper case 'C' in the class name
    "BaiChuanForCausalLM": ("baichuan", "BaiChuanForCausalLM"),
    # baichuan-13b, lower case 'c' in the class name
    "BaichuanForCausalLM": ("baichuan", "BaichuanForCausalLM"),
Yu Chin Fabian Lim's avatar
Yu Chin Fabian Lim committed
43
    "BambaForCausalLM": ("bamba", "BambaForCausalLM"),
44
    "BloomForCausalLM": ("bloom", "BloomForCausalLM"),
45
    "ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
46
    "ChatGLMForConditionalGeneration": ("chatglm", "ChatGLMForCausalLM"),
47
    "CohereForCausalLM": ("commandr", "CohereForCausalLM"),
48
    "Cohere2ForCausalLM": ("commandr", "CohereForCausalLM"),
49
    "DbrxForCausalLM": ("dbrx", "DbrxForCausalLM"),
50
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
51
52
    "DeepseekForCausalLM": ("deepseek", "DeepseekForCausalLM"),
    "DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
53
    "DeepseekV3ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
54
55
    "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
    "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
56
    "Fairseq2LlamaForCausalLM": ("fairseq2_llama", "Fairseq2LlamaForCausalLM"),
57
58
    "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
    "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
59
    "Gemma3ForCausalLM": ("gemma3", "Gemma3ForCausalLM"),
60
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
61
    "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
62
63
64
65
66
67
    "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
    "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
    "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
    "GPTNeoXForCausalLM": ("gpt_neox", "GPTNeoXForCausalLM"),
    "GraniteForCausalLM": ("granite", "GraniteForCausalLM"),
    "GraniteMoeForCausalLM": ("granitemoe", "GraniteMoeForCausalLM"),
68
    "GraniteMoeSharedForCausalLM": ("granitemoeshared", "GraniteMoeSharedForCausalLM"),   # noqa: E501
69
    "GritLM": ("gritlm", "GritLM"),
Michael Goin's avatar
Michael Goin committed
70
    "Grok1ModelForCausalLM": ("grok1", "Grok1ForCausalLM"),
71
72
    "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
    "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"),
73
    "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
74
    "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
75
76
77
78
79
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
    # For decapoda-research/llama-*
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
80
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
81
    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
82
    "Mamba2ForCausalLM": ("mamba2", "Mamba2ForCausalLM"),
83
84
    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
    "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
85
86
87
88
89
90
91
92
    "MistralForCausalLM": ("llama", "LlamaForCausalLM"),
    "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
    "QuantMixtralForCausalLM": ("mixtral_quant", "MixtralForCausalLM"),
    # transformers's mpt class has lower case
    "MptForCausalLM": ("mpt", "MPTForCausalLM"),
    "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
    "NemotronForCausalLM": ("nemotron", "NemotronForCausalLM"),
    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
93
    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
94
95
96
97
98
99
100
101
    "OlmoeForCausalLM": ("olmoe", "OlmoeForCausalLM"),
    "OPTForCausalLM": ("opt", "OPTForCausalLM"),
    "OrionForCausalLM": ("orion", "OrionForCausalLM"),
    "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
    "PhiForCausalLM": ("phi", "PhiForCausalLM"),
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
    "Phi3SmallForCausalLM": ("phi3_small", "Phi3SmallForCausalLM"),
    "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
102
    "Plamo2ForCausalLM": ("plamo2", "Plamo2ForCausalLM"),
103
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
104
105
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
106
107
    "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),
    "Qwen3MoeForCausalLM": ("qwen3_moe", "Qwen3MoeForCausalLM"),
108
109
110
111
112
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
113
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
114
    "TeleFLMForCausalLM": ("teleflm", "TeleFLMForCausalLM"),
115
    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
116
    "Zamba2ForCausalLM": ("zamba2", "Zamba2ForCausalLM"),
117
118
119
    # [Encoder-decoder]
    "BartModel": ("bart", "BartForConditionalGeneration"),
    "BartForConditionalGeneration": ("bart", "BartForConditionalGeneration"),
120
121
122
}

_EMBEDDING_MODELS = {
123
    # [Text-only]
124
    "BertModel": ("bert", "BertEmbeddingModel"),
125
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
126
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
127
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
128
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
129
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
130
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
131
    "GritLM": ("gritlm", "GritLM"),
132
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
133
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
134
    "LlamaModel": ("llama", "LlamaForCausalLM"),
135
136
137
138
139
    **{
        # Multiple models share the same architecture, so we include them all
        k: (mod, arch) for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
        if arch == "LlamaForCausalLM"
    },
140
    "MistralModel": ("llama", "LlamaForCausalLM"),
141
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
142
143
    "Qwen2Model": ("qwen2", "Qwen2EmbeddingModel"),
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
144
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
145
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
146
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
147
    # [Multimodal]
Cyrus Leung's avatar
Cyrus Leung committed
148
    "LlavaNextForConditionalGeneration": ("llava_next", "LlavaNextForConditionalGeneration"),  # noqa: E501
149
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
150
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
151
152
    # [Auto-converted (see adapters.py)]
    "Qwen2ForSequenceClassification": ("qwen2", "Qwen2ForCausalLM"),
153
154
155
156
    # Technically PrithviGeoSpatialMAE is a model that works on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
    # models for the time being.
    "PrithviGeoSpatialMAE": ("prithvi_geospatial_mae", "PrithviGeoSpatialMAE"),
157
158
}

159
160
161
162
163
164
_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
    "RobertaForSequenceClassification": ("roberta",
                                         "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": ("roberta",
                                            "RobertaForSequenceClassification"),
xsank's avatar
xsank committed
165
166
    "ModernBertForSequenceClassification": ("modernbert",
                                            "ModernBertForSequenceClassification"),
167
168
}

169
_MULTIMODAL_MODELS = {
170
    # [Decoder-only]
171
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
Jennifer Zhao's avatar
Jennifer Zhao committed
172
    "AyaVisionForConditionalGeneration": ("aya_vision", "AyaVisionForConditionalGeneration"),  # noqa: E501
173
174
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
    "ChameleonForConditionalGeneration": ("chameleon", "ChameleonForConditionalGeneration"),  # noqa: E501
175
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
176
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
177
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
178
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
179
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
180
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
181
    "Idefics3ForConditionalGeneration":("idefics3","Idefics3ForConditionalGeneration"),
182
    "SmolVLMForConditionalGeneration": ("smolvlm","SmolVLMForConditionalGeneration"),  # noqa: E501
183
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
184
185
186
187
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
    "LlavaNextForConditionalGeneration": ("llava_next", "LlavaNextForConditionalGeneration"),  # noqa: E501
    "LlavaNextVideoForConditionalGeneration": ("llava_next_video", "LlavaNextVideoForConditionalGeneration"),  # noqa: E501
    "LlavaOnevisionForConditionalGeneration": ("llava_onevision", "LlavaOnevisionForConditionalGeneration"),  # noqa: E501
188
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
189
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
190
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
191
    "Mistral3ForConditionalGeneration": ("mistral3", "Mistral3ForConditionalGeneration"),  # noqa: E501
192
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
193
194
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
    "PaliGemmaForConditionalGeneration": ("paligemma", "PaliGemmaForConditionalGeneration"),  # noqa: E501
195
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
196
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
197
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),  # noqa: E501
198
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
Roger Wang's avatar
Roger Wang committed
199
    "Qwen2_5_VLForConditionalGeneration": ("qwen2_5_vl", "Qwen2_5_VLForConditionalGeneration"),  # noqa: E501
200
    "Qwen2AudioForConditionalGeneration": ("qwen2_audio", "Qwen2AudioForConditionalGeneration"),  # noqa: E501
201
    "UltravoxModel": ("ultravox", "UltravoxModel"),
202
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
203
    # [Encoder-decoder]
204
    "Florence2ForConditionalGeneration": ("florence2", "Florence2ForConditionalGeneration"),  # noqa: E501
205
    "MllamaForConditionalGeneration": ("mllama", "MllamaForConditionalGeneration"),  # noqa: E501
206
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),  # noqa: E501
207
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
208
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
209
}
210
211
212

_SPECULATIVE_DECODING_MODELS = {
    "EAGLEModel": ("eagle", "EAGLE"),
213
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
214
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
215
216
    "MedusaModel": ("medusa", "Medusa"),
    "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
217
}
218

219
_TRANSFORMERS_MODELS = {
220
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
221
}
222
# yapf: enable
223

224
_VLLM_MODELS = {
225
    **_TEXT_GENERATION_MODELS,
226
    **_EMBEDDING_MODELS,
227
    **_CROSS_ENCODER_MODELS,
228
    **_MULTIMODAL_MODELS,
229
    **_SPECULATIVE_DECODING_MODELS,
230
    **_TRANSFORMERS_MODELS,
231
232
}

233
234
235
236
237
238
239
240
# This variable is used as the args for subprocess.run(). We
# can modify  this variable to alter the args if needed. e.g.
# when we use par format to pack things together, sys.executable
# might not be the target we want to run.
_SUBPROCESS_COMMAND = [
    sys.executable, "-m", "vllm.model_executor.models.registry"
]

241

242
243
@dataclass(frozen=True)
class _ModelInfo:
244
    architecture: str
245
    is_text_generation_model: bool
246
    is_pooling_model: bool
247
    supports_cross_encoding: bool
248
249
    supports_multimodal: bool
    supports_pp: bool
250
251
    has_inner_state: bool
    is_attention_free: bool
252
    is_hybrid: bool
253
    has_noops: bool
254
    supports_transcription: bool
255
    supports_v0_only: bool
256
257

    @staticmethod
258
259
    def from_model_cls(model: Type[nn.Module]) -> "_ModelInfo":
        return _ModelInfo(
260
            architecture=model.__name__,
261
            is_text_generation_model=is_text_generation_model(model),
262
            is_pooling_model=True,  # Can convert any model into a pooling model
263
            supports_cross_encoding=supports_cross_encoding(model),
264
265
            supports_multimodal=supports_multimodal(model),
            supports_pp=supports_pp(model),
266
267
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
268
            is_hybrid=is_hybrid(model),
269
270
            supports_transcription=supports_transcription(model),
            supports_v0_only=supports_v0_only(model),
271
            has_noops=has_noops(model),
272
        )
273
274


275
class _BaseRegisteredModel(ABC):
276

277
278
279
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
280

281
282
283
    @abstractmethod
    def load_model_cls(self) -> Type[nn.Module]:
        raise NotImplementedError
284
285


286
287
288
289
290
291
292
293
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
    model_cls: Type[nn.Module]
294
295

    @staticmethod
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
    def from_model_cls(model_cls: Type[nn.Module]):
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

    def inspect_model_cls(self) -> _ModelInfo:
        return self.interfaces

    def load_model_cls(self) -> Type[nn.Module]:
        return self.model_cls


@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """
    module_name: str
    class_name: str

    # Performed in another process to avoid initializing CUDA
    def inspect_model_cls(self) -> _ModelInfo:
        return _run_in_subprocess(
            lambda: _ModelInfo.from_model_cls(self.load_model_cls()))

    def load_model_cls(self) -> Type[nn.Module]:
        mod = importlib.import_module(self.module_name)
        return getattr(mod, self.class_name)


@lru_cache(maxsize=128)
def _try_load_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
) -> Optional[Type[nn.Module]]:
332
    from vllm.platforms import current_platform
333
    current_platform.verify_model_arch(model_arch)
334
335
336
337
338
339
    try:
        return model.load_model_cls()
    except Exception:
        logger.exception("Error in loading model architecture '%s'",
                         model_arch)
        return None
340
341


342
343
344
345
346
347
348
349
350
351
352
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
) -> Optional[_ModelInfo]:
    try:
        return model.inspect_model_cls()
    except Exception:
        logger.exception("Error in inspecting model architecture '%s'",
                         model_arch)
        return None
353
354


355
356
357
358
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
    models: Dict[str, _BaseRegisteredModel] = field(default_factory=dict)
359

360
361
    def get_supported_archs(self) -> AbstractSet[str]:
        return self.models.keys()
362

363
364
365
366
367
    def register_model(
        self,
        model_arch: str,
        model_cls: Union[Type[nn.Module], str],
    ) -> None:
368
369
370
371
372
373
374
375
376
377
378
        """
        Register an external model to be used in vLLM.

        :code:`model_cls` can be either:

        - A :class:`torch.nn.Module` class directly referencing the model.
        - A string in the format :code:`<module>:<class>` which can be used to
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
          :code:`RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
        """
379
380
381
382
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

383
        if model_arch in self.models:
384
385
386
            logger.warning(
                "Model architecture %s is already registered, and will be "
                "overwritten by the new model class %s.", model_arch,
387
388
389
390
391
392
393
                model_cls)

        if isinstance(model_cls, str):
            split_str = model_cls.split(":")
            if len(split_str) != 2:
                msg = "Expected a string in the format `<module>:<class>`"
                raise ValueError(msg)
394

395
            model = _LazyRegisteredModel(*split_str)
396
397
        elif isinstance(model_cls, type) and (is_in_doc_build() or issubclass(
                model_cls, nn.Module)):
398
            model = _RegisteredModel.from_model_cls(model_cls)
399
400
401
402
        else:
            msg = ("`model_cls` should be a string or PyTorch model class, "
                   f"not a {type(model_arch)}")
            raise TypeError(msg)
403

404
        self.models[model_arch] = model
405

406
407
    def _raise_for_unsupported(self, architectures: List[str]):
        all_supported_archs = self.get_supported_archs()
408

409
410
411
412
413
        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
                "to be inspected. Please check the logs for more details.")

414
415
416
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
            f"Supported architectures: {all_supported_archs}")
417

418
419
420
421
    def _try_load_model_cls(self,
                            model_arch: str) -> Optional[Type[nn.Module]]:
        if model_arch not in self.models:
            return None
422

423
        return _try_load_model_cls(model_arch, self.models[model_arch])
424

425
426
427
    def _try_inspect_model_cls(self, model_arch: str) -> Optional[_ModelInfo]:
        if model_arch not in self.models:
            return None
428

429
        return _try_inspect_model_cls(model_arch, self.models[model_arch])
430

431
432
433
434
    def _normalize_archs(
        self,
        architectures: Union[str, List[str]],
    ) -> List[str]:
435
436
437
438
439
        if isinstance(architectures, str):
            architectures = [architectures]
        if not architectures:
            logger.warning("No model architectures are specified")

440
441
442
443
        # filter out support architectures
        normalized_arch = list(
            filter(lambda model: model in self.models, architectures))

444
        # make sure Transformers backend is put at the last as a fallback
445
        if len(normalized_arch) != len(architectures):
446
            normalized_arch.append("TransformersForCausalLM")
447
        return normalized_arch
448

449
450
451
    def inspect_model_cls(
        self,
        architectures: Union[str, List[str]],
452
    ) -> Tuple[_ModelInfo, str]:
453
        architectures = self._normalize_archs(architectures)
454

455
456
457
        for arch in architectures:
            model_info = self._try_inspect_model_cls(arch)
            if model_info is not None:
458
                return (model_info, arch)
459

460
        return self._raise_for_unsupported(architectures)
461

462
463
464
465
466
    def resolve_model_cls(
        self,
        architectures: Union[str, List[str]],
    ) -> Tuple[Type[nn.Module], str]:
        architectures = self._normalize_archs(architectures)
467

468
469
470
471
        for arch in architectures:
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
472

473
        return self._raise_for_unsupported(architectures)
474

475
476
477
478
    def is_text_generation_model(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
479
480
        model_cls, _ = self.inspect_model_cls(architectures)
        return model_cls.is_text_generation_model
481

482
    def is_pooling_model(
483
484
485
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
486
        model_cls, _ = self.inspect_model_cls(architectures)
487
        return model_cls.is_pooling_model
488

489
490
491
492
    def is_cross_encoder_model(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
493
494
        model_cls, _ = self.inspect_model_cls(architectures)
        return model_cls.supports_cross_encoding
495

496
497
498
499
    def is_multimodal_model(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
500
501
        model_cls, _ = self.inspect_model_cls(architectures)
        return model_cls.supports_multimodal
502
503
504
505
506

    def is_pp_supported_model(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
507
508
        model_cls, _ = self.inspect_model_cls(architectures)
        return model_cls.supports_pp
509

510
511
512
513
514
515
    def model_has_inner_state(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures)
        return model_cls.has_inner_state
516

517
518
519
520
521
522
    def is_attention_free_model(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures)
        return model_cls.is_attention_free
523

524
525
526
527
528
529
530
    def is_hybrid_model(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures)
        return model_cls.is_hybrid

531
532
533
534
535
536
537
    def is_noops_model(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures)
        return model_cls.has_noops

538
539
540
541
542
543
544
    def is_transcription_model(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures)
        return model_cls.supports_transcription

545
546
547
548
549
550
551
    def is_v1_compatible(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
        model_cls, _ = self.inspect_model_cls(architectures)
        return not model_cls.supports_v0_only

552
553

ModelRegistry = _ModelRegistry({
554
555
    model_arch:
    _LazyRegisteredModel(
556
557
558
559
560
561
562
563
564
565
        module_name=f"vllm.model_executor.models.{mod_relname}",
        class_name=cls_name,
    )
    for model_arch, (mod_relname, cls_name) in _VLLM_MODELS.items()
})

_T = TypeVar("_T")


def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
566
567
568
569
570
    # NOTE: We use a temporary directory instead of a temporary file to avoid
    # issues like https://stackoverflow.com/questions/23212435/permission-denied-to-write-to-my-temporary-file
    with tempfile.TemporaryDirectory() as tempdir:
        output_filepath = os.path.join(tempdir, "registry_output.tmp")

571
        # `cloudpickle` allows pickling lambda functions directly
572
        input_bytes = cloudpickle.dumps((fn, output_filepath))
573
574
575

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
576
577
578
        returned = subprocess.run(_SUBPROCESS_COMMAND,
                                  input=input_bytes,
                                  capture_output=True)
579
580
581
582
583
584
585
586
587

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
            raise RuntimeError(f"Error raised in subprocess:\n"
                               f"{returned.stderr.decode()}") from e

588
        with open(output_filepath, "rb") as f:
589
590
591
592
593
594
595
596
597
598
599
            return pickle.load(f)


def _run() -> None:
    # Setup plugins
    from vllm.plugins import load_general_plugins
    load_general_plugins()

    fn, output_file = pickle.loads(sys.stdin.buffer.read())

    result = fn()
600
601
602

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
603
604
605


if __name__ == "__main__":
606
    _run()