registry.py 16.4 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
50
51
52
53
54
    "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"),
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
    # For decapoda-research/llama-*
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
55
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
56
    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
    "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"),
75
    # QWenLMHeadModel supports multimodal
76
77
78
79
80
81
82
83
    "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"),
84
85
86
    # [Encoder-decoder]
    "BartModel": ("bart", "BartForConditionalGeneration"),
    "BartForConditionalGeneration": ("bart", "BartForConditionalGeneration"),
87
88
89
}

_EMBEDDING_MODELS = {
90
    # [Text-only]
91
    "BertModel": ("bert", "BertEmbeddingModel"),
92
93
    "Gemma2Model": ("gemma2", "Gemma2EmbeddingModel"),
    "MistralModel": ("llama", "LlamaEmbeddingModel"),
94
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
95
96
    # [Multimodal]
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
97
98
99
}

_MULTIMODAL_MODELS = {
100
101
102
    # [Decoder-only]
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
    "ChameleonForConditionalGeneration": ("chameleon", "ChameleonForConditionalGeneration"),  # noqa: E501
103
104
    "ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
    "ChatGLMForConditionalGeneration": ("chatglm", "ChatGLMForCausalLM"),
105
106
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
107
108
109
110
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
    "LlavaNextForConditionalGeneration": ("llava_next", "LlavaNextForConditionalGeneration"),  # noqa: E501
    "LlavaNextVideoForConditionalGeneration": ("llava_next_video", "LlavaNextVideoForConditionalGeneration"),  # noqa: E501
    "LlavaOnevisionForConditionalGeneration": ("llava_onevision", "LlavaOnevisionForConditionalGeneration"),  # noqa: E501
111
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
112
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
113
114
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
    "PaliGemmaForConditionalGeneration": ("paligemma", "PaliGemmaForConditionalGeneration"),  # noqa: E501
115
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
116
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
117
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
118
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
119
    "UltravoxModel": ("ultravox", "UltravoxModel"),
120
121
    # [Encoder-decoder]
    "MllamaForConditionalGeneration": ("mllama", "MllamaForConditionalGeneration"),  # noqa: E501
122
}
123
124
125
126
127

_SPECULATIVE_DECODING_MODELS = {
    "EAGLEModel": ("eagle", "EAGLE"),
    "MedusaModel": ("medusa", "Medusa"),
    "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
128
}
129
# yapf: enable
130

131
_VLLM_MODELS = {
132
    **_TEXT_GENERATION_MODELS,
133
134
    **_EMBEDDING_MODELS,
    **_MULTIMODAL_MODELS,
135
    **_SPECULATIVE_DECODING_MODELS,
136
137
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
}

# 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`")
}


164
165
166
167
168
169
@dataclass(frozen=True)
class _ModelInfo:
    is_text_generation_model: bool
    is_embedding_model: bool
    supports_multimodal: bool
    supports_pp: bool
170
171
    has_inner_state: bool
    is_attention_free: bool
172
173

    @staticmethod
174
175
176
177
178
179
    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),
180
181
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
182
        )
183
184


185
class _BaseRegisteredModel(ABC):
186

187
188
189
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
190

191
192
193
    @abstractmethod
    def load_model_cls(self) -> Type[nn.Module]:
        raise NotImplementedError
194
195


196
197
198
199
200
201
202
203
@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]
204
205

    @staticmethod
206
207
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
    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
259
260


261
262
263
264
265
266
267
268
269
270
271
@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
272
273


274
275
276
277
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
    models: Dict[str, _BaseRegisteredModel] = field(default_factory=dict)
278

279
280
    def get_supported_archs(self) -> List[str]:
        return list(self.models.keys())
281

282
283
284
285
286
    def register_model(
        self,
        model_arch: str,
        model_cls: Union[Type[nn.Module], str],
    ) -> None:
287
288
289
290
291
292
293
294
295
296
297
        """
        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`.
        """
298
        if model_arch in self.models:
299
300
301
            logger.warning(
                "Model architecture %s is already registered, and will be "
                "overwritten by the new model class %s.", model_arch,
302
303
304
305
306
307
308
                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)
309

310
            model = _LazyRegisteredModel(*split_str)
311
        else:
312
            model = _RegisteredModel.from_model_cls(model_cls)
313

314
        self.models[model_arch] = model
315

316
317
    def _raise_for_unsupported(self, architectures: List[str]):
        all_supported_archs = self.get_supported_archs()
318

319
320
321
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
            f"Supported architectures: {all_supported_archs}")
322

323
324
325
326
    def _try_load_model_cls(self,
                            model_arch: str) -> Optional[Type[nn.Module]]:
        if model_arch not in self.models:
            return None
327

328
        return _try_load_model_cls(model_arch, self.models[model_arch])
329

330
331
332
    def _try_inspect_model_cls(self, model_arch: str) -> Optional[_ModelInfo]:
        if model_arch not in self.models:
            return None
333

334
        return _try_inspect_model_cls(model_arch, self.models[model_arch])
335

336
337
338
339
    def _normalize_archs(
        self,
        architectures: Union[str, List[str]],
    ) -> List[str]:
340
341
342
343
344
        if isinstance(architectures, str):
            architectures = [architectures]
        if not architectures:
            logger.warning("No model architectures are specified")

345
        return architectures
346

347
348
349
350
351
    def inspect_model_cls(
        self,
        architectures: Union[str, List[str]],
    ) -> _ModelInfo:
        architectures = self._normalize_archs(architectures)
352

353
354
355
356
        for arch in architectures:
            model_info = self._try_inspect_model_cls(arch)
            if model_info is not None:
                return model_info
357

358
        return self._raise_for_unsupported(architectures)
359

360
361
362
363
364
    def resolve_model_cls(
        self,
        architectures: Union[str, List[str]],
    ) -> Tuple[Type[nn.Module], str]:
        architectures = self._normalize_archs(architectures)
365

366
367
368
369
        for arch in architectures:
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
370

371
        return self._raise_for_unsupported(architectures)
372

373
374
375
376
377
    def is_text_generation_model(
        self,
        architectures: Union[str, List[str]],
    ) -> bool:
        return self.inspect_model_cls(architectures).is_text_generation_model
378

379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
    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

397
398
399
400
401
402
403
404
    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

405
406
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

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()
449
450
451

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
452
453
454
455


if __name__ == "__main__":
    _run()