registry.py 20.4 KB
Newer Older
1
2
3
4
"""
Whenever you add an architecture to this page, please also update
`tests/models/registry.py` with example HuggingFace models for it.
"""
5
import importlib
6
import os
7
import pickle
8
9
import subprocess
import sys
10
import tempfile
11
12
13
from abc import ABC, abstractmethod
from dataclasses import dataclass, field
from functools import lru_cache
14
15
from typing import (AbstractSet, Callable, Dict, List, Optional, Tuple, Type,
                    TypeVar, Union)
16

17
import cloudpickle
18
19
20
import torch.nn as nn

from vllm.logger import init_logger
21
from vllm.platforms import current_platform
22

23
from .adapters import as_embedding_model
24
from .interfaces import (has_inner_state, is_attention_free, is_hybrid,
25
26
                         supports_cross_encoding, supports_multimodal,
                         supports_pp)
27
from .interfaces_base import is_pooling_model, is_text_generation_model
28
29
30

logger = init_logger(__name__)

31
# yapf: disable
32
33
_TEXT_GENERATION_MODELS = {
    # [Decoder-only]
34
35
36
    "AquilaModel": ("llama", "LlamaForCausalLM"),
    "AquilaForCausalLM": ("llama", "LlamaForCausalLM"),  # AquilaChat2
    "ArcticForCausalLM": ("arctic", "ArcticForCausalLM"),
37
38
39
40
    # baichuan-7b, upper case 'C' in the class name
    "BaiChuanForCausalLM": ("baichuan", "BaiChuanForCausalLM"),
    # baichuan-13b, lower case 'c' in the class name
    "BaichuanForCausalLM": ("baichuan", "BaichuanForCausalLM"),
41
    "BloomForCausalLM": ("bloom", "BloomForCausalLM"),
42
    # ChatGLMModel supports multimodal
43
44
45
46
47
48
49
50
51
    "CohereForCausalLM": ("commandr", "CohereForCausalLM"),
    "DbrxForCausalLM": ("dbrx", "DbrxForCausalLM"),
    "DeciLMForCausalLM": ("decilm", "DeciLMForCausalLM"),
    "DeepseekForCausalLM": ("deepseek", "DeepseekForCausalLM"),
    "DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
    "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
    "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
    "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
    "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
52
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
53
54
55
56
57
58
59
60
    "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
    "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
    "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
    "GPTNeoXForCausalLM": ("gpt_neox", "GPTNeoXForCausalLM"),
    "GraniteForCausalLM": ("granite", "GraniteForCausalLM"),
    "GraniteMoeForCausalLM": ("granitemoe", "GraniteMoeForCausalLM"),
    "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
    "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"),
61
    "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
62
63
64
65
66
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
    # For decapoda-research/llama-*
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
67
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
68
    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
69
70
    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
    "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
71
72
73
74
75
76
77
78
    "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"),
79
    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
80
81
82
83
84
85
86
87
    "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"),
88
    # QWenLMHeadModel supports multimodal
89
90
91
92
93
94
95
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
96
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
97
    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
98
99
100
    # [Encoder-decoder]
    "BartModel": ("bart", "BartForConditionalGeneration"),
    "BartForConditionalGeneration": ("bart", "BartForConditionalGeneration"),
101
    "Florence2ForConditionalGeneration": ("florence2", "Florence2ForConditionalGeneration"),  # noqa: E501
102
103
104
}

_EMBEDDING_MODELS = {
105
    # [Text-only]
106
    "BertModel": ("bert", "BertEmbeddingModel"),
107
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
108
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
109
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
110
    "DeciLMForCausalLM": ("decilm", "DeciLMForCausalLM"),
111
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
112
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
113
    "LlamaModel": ("llama", "LlamaForCausalLM"),
114
115
116
117
118
    **{
        # 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"
    },
119
    "MistralModel": ("llama", "LlamaForCausalLM"),
120
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
121
122
    "Qwen2Model": ("qwen2", "Qwen2EmbeddingModel"),
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
123
124
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
    "Qwen2ForSequenceClassification": ("qwen2_cls", "Qwen2ForSequenceClassification"),  # noqa: E501
125
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
126
    # [Multimodal]
Cyrus Leung's avatar
Cyrus Leung committed
127
    "LlavaNextForConditionalGeneration": ("llava_next", "LlavaNextForConditionalGeneration"),  # noqa: E501
128
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
129
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
130
131
}

132
133
134
135
136
137
138
139
_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
    "RobertaForSequenceClassification": ("roberta",
                                         "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": ("roberta",
                                            "RobertaForSequenceClassification"),
}

140
_MULTIMODAL_MODELS = {
141
    # [Decoder-only]
142
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
143
144
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
    "ChameleonForConditionalGeneration": ("chameleon", "ChameleonForConditionalGeneration"),  # noqa: E501
145
146
    "ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
    "ChatGLMForConditionalGeneration": ("chatglm", "ChatGLMForCausalLM"),
147
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
148
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
149
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
150
    "Idefics3ForConditionalGeneration":("idefics3","Idefics3ForConditionalGeneration"),
151
152
153
154
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
    "LlavaNextForConditionalGeneration": ("llava_next", "LlavaNextForConditionalGeneration"),  # noqa: E501
    "LlavaNextVideoForConditionalGeneration": ("llava_next_video", "LlavaNextVideoForConditionalGeneration"),  # noqa: E501
    "LlavaOnevisionForConditionalGeneration": ("llava_onevision", "LlavaOnevisionForConditionalGeneration"),  # noqa: E501
155
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
156
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
157
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
158
159
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
    "PaliGemmaForConditionalGeneration": ("paligemma", "PaliGemmaForConditionalGeneration"),  # noqa: E501
160
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
161
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
162
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
163
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
164
    "Qwen2AudioForConditionalGeneration": ("qwen2_audio", "Qwen2AudioForConditionalGeneration"),  # noqa: E501
165
    "UltravoxModel": ("ultravox", "UltravoxModel"),
166
167
    # [Encoder-decoder]
    "MllamaForConditionalGeneration": ("mllama", "MllamaForConditionalGeneration"),  # noqa: E501
168
}
169
170
171
172
173

_SPECULATIVE_DECODING_MODELS = {
    "EAGLEModel": ("eagle", "EAGLE"),
    "MedusaModel": ("medusa", "Medusa"),
    "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
174
}
175
# yapf: enable
176

177
_VLLM_MODELS = {
178
    **_TEXT_GENERATION_MODELS,
179
    **_EMBEDDING_MODELS,
180
    **_CROSS_ENCODER_MODELS,
181
    **_MULTIMODAL_MODELS,
182
    **_SPECULATIVE_DECODING_MODELS,
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
}

# Models not supported by ROCm.
_ROCM_UNSUPPORTED_MODELS: List[str] = []

# Models partially supported by ROCm.
# Architecture -> Reason.
_ROCM_SWA_REASON = ("Sliding window attention (SWA) is not yet supported in "
                    "Triton flash attention. For half-precision SWA support, "
                    "please use CK flash attention by setting "
                    "`VLLM_USE_TRITON_FLASH_ATTN=0`")
_ROCM_PARTIALLY_SUPPORTED_MODELS: Dict[str, str] = {
    "Qwen2ForCausalLM":
    _ROCM_SWA_REASON,
    "MistralForCausalLM":
    _ROCM_SWA_REASON,
    "MixtralForCausalLM":
    _ROCM_SWA_REASON,
    "PaliGemmaForConditionalGeneration":
    ("ROCm flash attention does not yet "
     "fully support 32-bit precision on PaliGemma"),
    "Phi3VForCausalLM":
    ("ROCm Triton flash attention may run into compilation errors due to "
     "excessive use of shared memory. If this happens, disable Triton FA "
     "by setting `VLLM_USE_TRITON_FLASH_ATTN=0`")
}


211
212
@dataclass(frozen=True)
class _ModelInfo:
213
    architecture: str
214
    is_text_generation_model: bool
215
    is_pooling_model: bool
216
    supports_cross_encoding: bool
217
218
    supports_multimodal: bool
    supports_pp: bool
219
220
    has_inner_state: bool
    is_attention_free: bool
221
    is_hybrid: bool
222
223

    @staticmethod
224
    def from_model_cls(model: Type[nn.Module]) -> "_ModelInfo":
225
226
        is_pooling_model_ = is_pooling_model(model)
        if not is_pooling_model_:
227
228
229
230
231
            try:
                as_embedding_model(model)
            except Exception:
                pass
            else:
232
                is_pooling_model_ = True
233

234
        return _ModelInfo(
235
            architecture=model.__name__,
236
            is_text_generation_model=is_text_generation_model(model),
237
            is_pooling_model=is_pooling_model_,
238
            supports_cross_encoding=supports_cross_encoding(model),
239
240
            supports_multimodal=supports_multimodal(model),
            supports_pp=supports_pp(model),
241
242
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
243
            is_hybrid=is_hybrid(model),
244
        )
245
246


247
class _BaseRegisteredModel(ABC):
248

249
250
251
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
252

253
254
255
    @abstractmethod
    def load_model_cls(self) -> Type[nn.Module]:
        raise NotImplementedError
256
257


258
259
260
261
262
263
264
265
@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]
266
267

    @staticmethod
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
300
301
302
303
    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]]:
304
    if current_platform.is_rocm():
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
        if model_arch in _ROCM_UNSUPPORTED_MODELS:
            raise ValueError(f"Model architecture '{model_arch}' is not "
                             "supported by ROCm for now.")

        if model_arch in _ROCM_PARTIALLY_SUPPORTED_MODELS:
            msg = _ROCM_PARTIALLY_SUPPORTED_MODELS[model_arch]
            logger.warning(
                "Model architecture '%s' is partially "
                "supported by ROCm: %s", model_arch, msg)

    try:
        return model.load_model_cls()
    except Exception:
        logger.exception("Error in loading model architecture '%s'",
                         model_arch)
        return None
321
322


323
324
325
326
327
328
329
330
331
332
333
@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
334
335


336
337
338
339
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
    models: Dict[str, _BaseRegisteredModel] = field(default_factory=dict)
340

341
342
    def get_supported_archs(self) -> AbstractSet[str]:
        return self.models.keys()
343

344
345
346
347
348
    def register_model(
        self,
        model_arch: str,
        model_cls: Union[Type[nn.Module], str],
    ) -> None:
349
350
351
352
353
354
355
356
357
358
359
        """
        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`.
        """
360
        if model_arch in self.models:
361
362
363
            logger.warning(
                "Model architecture %s is already registered, and will be "
                "overwritten by the new model class %s.", model_arch,
364
365
366
367
368
369
370
                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)
371

372
            model = _LazyRegisteredModel(*split_str)
373
        else:
374
            model = _RegisteredModel.from_model_cls(model_cls)
375

376
        self.models[model_arch] = model
377

378
379
    def _raise_for_unsupported(self, architectures: List[str]):
        all_supported_archs = self.get_supported_archs()
380

381
382
383
384
385
        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.")

386
387
388
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
            f"Supported architectures: {all_supported_archs}")
389

390
391
392
393
    def _try_load_model_cls(self,
                            model_arch: str) -> Optional[Type[nn.Module]]:
        if model_arch not in self.models:
            return None
394

395
        return _try_load_model_cls(model_arch, self.models[model_arch])
396

397
398
399
    def _try_inspect_model_cls(self, model_arch: str) -> Optional[_ModelInfo]:
        if model_arch not in self.models:
            return None
400

401
        return _try_inspect_model_cls(model_arch, self.models[model_arch])
402

403
404
405
406
    def _normalize_archs(
        self,
        architectures: Union[str, List[str]],
    ) -> List[str]:
407
408
409
410
411
        if isinstance(architectures, str):
            architectures = [architectures]
        if not architectures:
            logger.warning("No model architectures are specified")

412
        return architectures
413

414
415
416
    def inspect_model_cls(
        self,
        architectures: Union[str, List[str]],
417
    ) -> Tuple[_ModelInfo, str]:
418
        architectures = self._normalize_archs(architectures)
419

420
421
422
        for arch in architectures:
            model_info = self._try_inspect_model_cls(arch)
            if model_info is not None:
423
                return (model_info, arch)
424

425
        return self._raise_for_unsupported(architectures)
426

427
428
429
430
431
    def resolve_model_cls(
        self,
        architectures: Union[str, List[str]],
    ) -> Tuple[Type[nn.Module], str]:
        architectures = self._normalize_archs(architectures)
432

433
434
435
436
        for arch in architectures:
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
437

438
        return self._raise_for_unsupported(architectures)
439

440
441
442
443
    def is_text_generation_model(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
444
445
        model_cls, _ = self.inspect_model_cls(architectures)
        return model_cls.is_text_generation_model
446

447
    def is_pooling_model(
448
449
450
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
451
        model_cls, _ = self.inspect_model_cls(architectures)
452
        return model_cls.is_pooling_model
453

454
455
456
457
    def is_cross_encoder_model(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
458
459
        model_cls, _ = self.inspect_model_cls(architectures)
        return model_cls.supports_cross_encoding
460

461
462
463
464
    def is_multimodal_model(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
465
466
        model_cls, _ = self.inspect_model_cls(architectures)
        return model_cls.supports_multimodal
467
468
469
470
471

    def is_pp_supported_model(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
472
473
        model_cls, _ = self.inspect_model_cls(architectures)
        return model_cls.supports_pp
474

475
476
477
478
479
480
    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
481

482
483
484
485
486
487
    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
488

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

496
497
498
499
500
501
502
503
504
505
506
507
508

ModelRegistry = _ModelRegistry({
    model_arch: _LazyRegisteredModel(
        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:
509
510
511
512
513
    # 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")

514
        # `cloudpickle` allows pickling lambda functions directly
515
        input_bytes = cloudpickle.dumps((fn, output_filepath))
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
        returned = subprocess.run(
            [sys.executable, "-m", "vllm.model_executor.models.registry"],
            input=input_bytes,
            capture_output=True)

        # 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

532
        with open(output_filepath, "rb") as f:
533
534
535
536
537
538
539
540
541
542
543
            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()
544
545
546

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
547
548
549


if __name__ == "__main__":
550
    _run()