registry.py 16.5 KB
Newer Older
1
import importlib
2
import pickle
3
4
import subprocess
import sys
5
import tempfile
6
7
8
9
from abc import ABC, abstractmethod
from dataclasses import dataclass, field
from functools import lru_cache
from typing import Callable, Dict, List, Optional, Tuple, Type, TypeVar, Union
10

11
import cloudpickle
12
13
14
15
16
import torch.nn as nn

from vllm.logger import init_logger
from vllm.utils import is_hip

17
18
from .interfaces import (has_inner_state, is_attention_free,
                         supports_multimodal, supports_pp)
19
from .interfaces_base import is_embedding_model, is_text_generation_model
20
21
22

logger = init_logger(__name__)

23
# yapf: disable
24
25
_TEXT_GENERATION_MODELS = {
    # [Decoder-only]
26
27
28
29
30
31
    "AquilaModel": ("llama", "LlamaForCausalLM"),
    "AquilaForCausalLM": ("llama", "LlamaForCausalLM"),  # AquilaChat2
    "ArcticForCausalLM": ("arctic", "ArcticForCausalLM"),
    "BaiChuanForCausalLM": ("baichuan", "BaiChuanForCausalLM"),  # baichuan-7b
    "BaichuanForCausalLM": ("baichuan", "BaichuanForCausalLM"),  # baichuan-13b
    "BloomForCausalLM": ("bloom", "BloomForCausalLM"),
32
    # ChatGLMModel supports multimodal
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
    "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"),
    "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"),
50
    "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
51
52
53
54
55
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
    # For decapoda-research/llama-*
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
56
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
57
    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
    "MistralForCausalLM": ("llama", "LlamaForCausalLM"),
    "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
    "QuantMixtralForCausalLM": ("mixtral_quant", "MixtralForCausalLM"),
    # transformers's mpt class has lower case
    "MptForCausalLM": ("mpt", "MPTForCausalLM"),
    "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
    "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
    "NemotronForCausalLM": ("nemotron", "NemotronForCausalLM"),
    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
    "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"),
76
    # QWenLMHeadModel supports multimodal
77
78
79
80
81
82
83
84
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
    "XverseForCausalLM": ("xverse", "XverseForCausalLM"),
85
86
87
    # [Encoder-decoder]
    "BartModel": ("bart", "BartForConditionalGeneration"),
    "BartForConditionalGeneration": ("bart", "BartForConditionalGeneration"),
88
89
90
}

_EMBEDDING_MODELS = {
91
    # [Text-only]
92
    "BertModel": ("bert", "BertEmbeddingModel"),
93
94
    "Gemma2Model": ("gemma2", "Gemma2EmbeddingModel"),
    "MistralModel": ("llama", "LlamaEmbeddingModel"),
95
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
96
    # [Multimodal]
Cyrus Leung's avatar
Cyrus Leung committed
97
    "LlavaNextForConditionalGeneration": ("llava_next", "LlavaNextForConditionalGeneration"),  # noqa: E501
98
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
99
100
101
}

_MULTIMODAL_MODELS = {
102
103
104
    # [Decoder-only]
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
    "ChameleonForConditionalGeneration": ("chameleon", "ChameleonForConditionalGeneration"),  # noqa: E501
105
106
    "ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
    "ChatGLMForConditionalGeneration": ("chatglm", "ChatGLMForCausalLM"),
107
108
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
109
110
111
112
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
    "LlavaNextForConditionalGeneration": ("llava_next", "LlavaNextForConditionalGeneration"),  # noqa: E501
    "LlavaNextVideoForConditionalGeneration": ("llava_next_video", "LlavaNextVideoForConditionalGeneration"),  # noqa: E501
    "LlavaOnevisionForConditionalGeneration": ("llava_onevision", "LlavaOnevisionForConditionalGeneration"),  # noqa: E501
113
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
114
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
115
116
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
    "PaliGemmaForConditionalGeneration": ("paligemma", "PaliGemmaForConditionalGeneration"),  # noqa: E501
117
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
118
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
119
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
120
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
121
    "UltravoxModel": ("ultravox", "UltravoxModel"),
122
123
    # [Encoder-decoder]
    "MllamaForConditionalGeneration": ("mllama", "MllamaForConditionalGeneration"),  # noqa: E501
124
}
125
126
127
128
129

_SPECULATIVE_DECODING_MODELS = {
    "EAGLEModel": ("eagle", "EAGLE"),
    "MedusaModel": ("medusa", "Medusa"),
    "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
130
}
131
# yapf: enable
132

133
_VLLM_MODELS = {
134
    **_TEXT_GENERATION_MODELS,
135
136
    **_EMBEDDING_MODELS,
    **_MULTIMODAL_MODELS,
137
    **_SPECULATIVE_DECODING_MODELS,
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
}

# 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`")
}


166
167
168
169
170
171
@dataclass(frozen=True)
class _ModelInfo:
    is_text_generation_model: bool
    is_embedding_model: bool
    supports_multimodal: bool
    supports_pp: bool
172
173
    has_inner_state: bool
    is_attention_free: bool
174
175

    @staticmethod
176
177
178
179
180
181
    def from_model_cls(model: Type[nn.Module]) -> "_ModelInfo":
        return _ModelInfo(
            is_text_generation_model=is_text_generation_model(model),
            is_embedding_model=is_embedding_model(model),
            supports_multimodal=supports_multimodal(model),
            supports_pp=supports_pp(model),
182
183
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
184
        )
185
186


187
class _BaseRegisteredModel(ABC):
188

189
190
191
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
192

193
194
195
    @abstractmethod
    def load_model_cls(self) -> Type[nn.Module]:
        raise NotImplementedError
196
197


198
199
200
201
202
203
204
205
@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]
206
207

    @staticmethod
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
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
    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]]:
    if is_hip():
        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
261
262


263
264
265
266
267
268
269
270
271
272
273
@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
274
275


276
277
278
279
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
    models: Dict[str, _BaseRegisteredModel] = field(default_factory=dict)
280

281
282
    def get_supported_archs(self) -> List[str]:
        return list(self.models.keys())
283

284
285
286
287
288
    def register_model(
        self,
        model_arch: str,
        model_cls: Union[Type[nn.Module], str],
    ) -> None:
289
290
291
292
293
294
295
296
297
298
299
        """
        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`.
        """
300
        if model_arch in self.models:
301
302
303
            logger.warning(
                "Model architecture %s is already registered, and will be "
                "overwritten by the new model class %s.", model_arch,
304
305
306
307
308
309
310
                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)
311

312
            model = _LazyRegisteredModel(*split_str)
313
        else:
314
            model = _RegisteredModel.from_model_cls(model_cls)
315

316
        self.models[model_arch] = model
317

318
319
    def _raise_for_unsupported(self, architectures: List[str]):
        all_supported_archs = self.get_supported_archs()
320

321
322
323
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
            f"Supported architectures: {all_supported_archs}")
324

325
326
327
328
    def _try_load_model_cls(self,
                            model_arch: str) -> Optional[Type[nn.Module]]:
        if model_arch not in self.models:
            return None
329

330
        return _try_load_model_cls(model_arch, self.models[model_arch])
331

332
333
334
    def _try_inspect_model_cls(self, model_arch: str) -> Optional[_ModelInfo]:
        if model_arch not in self.models:
            return None
335

336
        return _try_inspect_model_cls(model_arch, self.models[model_arch])
337

338
339
340
341
    def _normalize_archs(
        self,
        architectures: Union[str, List[str]],
    ) -> List[str]:
342
343
344
345
346
        if isinstance(architectures, str):
            architectures = [architectures]
        if not architectures:
            logger.warning("No model architectures are specified")

347
        return architectures
348

349
350
351
352
353
    def inspect_model_cls(
        self,
        architectures: Union[str, List[str]],
    ) -> _ModelInfo:
        architectures = self._normalize_archs(architectures)
354

355
356
357
358
        for arch in architectures:
            model_info = self._try_inspect_model_cls(arch)
            if model_info is not None:
                return model_info
359

360
        return self._raise_for_unsupported(architectures)
361

362
363
364
365
366
    def resolve_model_cls(
        self,
        architectures: Union[str, List[str]],
    ) -> Tuple[Type[nn.Module], str]:
        architectures = self._normalize_archs(architectures)
367

368
369
370
371
        for arch in architectures:
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
372

373
        return self._raise_for_unsupported(architectures)
374

375
376
377
378
379
    def is_text_generation_model(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
        return self.inspect_model_cls(architectures).is_text_generation_model
380

381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
    def is_embedding_model(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
        return self.inspect_model_cls(architectures).is_embedding_model

    def is_multimodal_model(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
        return self.inspect_model_cls(architectures).supports_multimodal

    def is_pp_supported_model(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
        return self.inspect_model_cls(architectures).supports_pp

399
400
401
402
403
404
405
406
    def model_has_inner_state(self, architectures: Union[str,
                                                         List[str]]) -> bool:
        return self.inspect_model_cls(architectures).has_inner_state

    def is_attention_free_model(self, architectures: Union[str,
                                                           List[str]]) -> bool:
        return self.inspect_model_cls(architectures).is_attention_free

407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450

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:
    with tempfile.NamedTemporaryFile() as output_file:
        # `cloudpickle` allows pickling lambda functions directly
        input_bytes = cloudpickle.dumps((fn, output_file.name))

        # 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

        with open(output_file.name, "rb") as f:
            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()
451
452
453

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
454
455
456
457


if __name__ == "__main__":
    _run()