registry.py 47.6 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
5
6
"""
Whenever you add an architecture to this page, please also update
`tests/models/registry.py` with example HuggingFace models for it.
"""
7

8
import importlib
9
import json
10
import os
11
import pickle
12
13
import subprocess
import sys
14
import tempfile
15
from abc import ABC, abstractmethod
16
from collections.abc import Callable, Set
17
from dataclasses import asdict, dataclass, field
18
from functools import lru_cache
19
from pathlib import Path
20
from typing import TYPE_CHECKING, Any, TypeVar
21
22

import torch.nn as nn
23
import transformers
24

25
from vllm import envs
26
27
28
29
30
from vllm.config import (
    ModelConfig,
    iter_architecture_defaults,
    try_match_architecture_defaults,
)
31
from vllm.logger import init_logger
32
from vllm.logging_utils import logtime
33
from vllm.transformers_utils.dynamic_module import try_get_class_from_dynamic_module
34
from vllm.utils.hashing import safe_hash
35

36
37
if TYPE_CHECKING:
    from vllm.config.model import AttnTypeStr
38
    from vllm.config.pooler import SequencePoolingType, TokenPoolingType
39
40
else:
    AttnTypeStr = Any
41
42
    SequencePoolingType = Any
    TokenPoolingType = Any
43
44


45
46
47
48
49
from .interfaces import (
    has_inner_state,
    has_noops,
    is_attention_free,
    is_hybrid,
Patrick von Platen's avatar
Patrick von Platen committed
50
    requires_raw_input_tokens,
51
    supports_cross_encoding,
52
    supports_mamba_prefix_caching,
53
54
55
56
57
58
59
    supports_multimodal,
    supports_multimodal_encoder_tp_data,
    supports_multimodal_raw_input_only,
    supports_pp,
    supports_transcription,
)
from .interfaces_base import (
60
    get_attn_type,
61
62
    get_default_seq_pooling_type,
    get_default_tok_pooling_type,
63
64
65
    is_pooling_model,
    is_text_generation_model,
)
66
67
68

logger = init_logger(__name__)

69
70
_TEXT_GENERATION_MODELS = {
    # [Decoder-only]
71
    "AfmoeForCausalLM": ("afmoe", "AfmoeForCausalLM"),
72
    "ApertusForCausalLM": ("apertus", "ApertusForCausalLM"),
73
74
    "AquilaModel": ("llama", "LlamaForCausalLM"),
    "AquilaForCausalLM": ("llama", "LlamaForCausalLM"),  # AquilaChat2
Raghav Ravishankar's avatar
Raghav Ravishankar committed
75
    "ArceeForCausalLM": ("arcee", "ArceeForCausalLM"),
76
    "ArcticForCausalLM": ("arctic", "ArcticForCausalLM"),
77
78
79
80
    # baichuan-7b, upper case 'C' in the class name
    "BaiChuanForCausalLM": ("baichuan", "BaiChuanForCausalLM"),
    # baichuan-13b, lower case 'c' in the class name
    "BaichuanForCausalLM": ("baichuan", "BaichuanForCausalLM"),
81
    "BailingMoeForCausalLM": ("bailing_moe", "BailingMoeForCausalLM"),
ant-yy's avatar
ant-yy committed
82
    "BailingMoeV2ForCausalLM": ("bailing_moe", "BailingMoeV2ForCausalLM"),
Yu Chin Fabian Lim's avatar
Yu Chin Fabian Lim committed
83
    "BambaForCausalLM": ("bamba", "BambaForCausalLM"),
84
    "BloomForCausalLM": ("bloom", "BloomForCausalLM"),
85
    "ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
86
    "ChatGLMForConditionalGeneration": ("chatglm", "ChatGLMForCausalLM"),
87
    "CohereForCausalLM": ("commandr", "CohereForCausalLM"),
88
    "Cohere2ForCausalLM": ("commandr", "CohereForCausalLM"),
89
    "CwmForCausalLM": ("llama", "LlamaForCausalLM"),
90
    "DbrxForCausalLM": ("dbrx", "DbrxForCausalLM"),
91
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
92
    "DeepseekForCausalLM": ("deepseek_v2", "DeepseekForCausalLM"),
93
    "DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
94
    "DeepseekV3ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
95
    "DeepseekV32ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
96
    "Dots1ForCausalLM": ("dots1", "Dots1ForCausalLM"),
97
    "Ernie4_5ForCausalLM": ("ernie45", "Ernie4_5ForCausalLM"),
98
    "Ernie4_5_MoeForCausalLM": ("ernie45_moe", "Ernie4_5_MoeForCausalLM"),
99
    "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
100
    "Exaone4ForCausalLM": ("exaone4", "Exaone4ForCausalLM"),
Kyungmin Lee's avatar
Kyungmin Lee committed
101
    "ExaoneMoEForCausalLM": ("exaone_moe", "ExaoneMoeForCausalLM"),
102
    "Fairseq2LlamaForCausalLM": ("fairseq2_llama", "Fairseq2LlamaForCausalLM"),
103
104
105
    "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
    "FalconH1ForCausalLM": ("falcon_h1", "FalconH1ForCausalLM"),
106
    "FlexOlmoForCausalLM": ("flex_olmo", "FlexOlmoForCausalLM"),
107
108
    "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
    "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
109
    "Gemma3ForCausalLM": ("gemma3", "Gemma3ForCausalLM"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
110
    "Gemma3nForCausalLM": ("gemma3n", "Gemma3nForCausalLM"),
111
    "Qwen3NextForCausalLM": ("qwen3_next", "Qwen3NextForCausalLM"),
112
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
113
    "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
114
    "Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
115
    "Glm4MoeLiteForCausalLM": ("glm4_moe_lite", "Glm4MoeLiteForCausalLM"),
116
    "GptOssForCausalLM": ("gpt_oss", "GptOssForCausalLM"),
117
118
119
120
121
122
    "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
    "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
    "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
    "GPTNeoXForCausalLM": ("gpt_neox", "GPTNeoXForCausalLM"),
    "GraniteForCausalLM": ("granite", "GraniteForCausalLM"),
    "GraniteMoeForCausalLM": ("granitemoe", "GraniteMoeForCausalLM"),
123
124
    "GraniteMoeHybridForCausalLM": ("granitemoehybrid", "GraniteMoeHybridForCausalLM"),  # noqa: E501
    "GraniteMoeSharedForCausalLM": ("granitemoeshared", "GraniteMoeSharedForCausalLM"),  # noqa: E501
125
    "GritLM": ("gritlm", "GritLM"),
Bijaya Dangol's avatar
Bijaya Dangol committed
126
127
    "Grok1ModelForCausalLM": ("grok1", "GrokForCausalLM"),
    "Grok1ForCausalLM": ("grok1", "GrokForCausalLM"),
128
129
    "HunYuanMoEV1ForCausalLM": ("hunyuan_v1", "HunYuanMoEV1ForCausalLM"),
    "HunYuanDenseV1ForCausalLM": ("hunyuan_v1", "HunYuanDenseV1ForCausalLM"),
130
    "HCXVisionForCausalLM": ("hyperclovax_vision", "HCXVisionForCausalLM"),
131
132
    "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
    "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"),
133
    "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
134
    "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
135
136
    "IQuestCoderForCausalLM": ("llama", "LlamaForCausalLM"),
    "IQuestLoopCoderForCausalLM": ("iquest_loopcoder", "IQuestLoopCoderForCausalLM"),
137
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
138
    "Jais2ForCausalLM": ("jais2", "Jais2ForCausalLM"),
139
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
140
    "KimiLinearForCausalLM": ("kimi_linear", "KimiLinearForCausalLM"),  # noqa: E501
141
    "Lfm2ForCausalLM": ("lfm2", "Lfm2ForCausalLM"),
Paul Pak's avatar
Paul Pak committed
142
    "Lfm2MoeForCausalLM": ("lfm2_moe", "Lfm2MoeForCausalLM"),
143
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
144
    "Llama4ForCausalLM": ("llama4", "Llama4ForCausalLM"),
145
146
    # For decapoda-research/llama-*
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
XuruiYang's avatar
XuruiYang committed
147
    "LongcatFlashForCausalLM": ("longcat_flash", "LongcatFlashForCausalLM"),
148
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
149
    "Mamba2ForCausalLM": ("mamba2", "Mamba2ForCausalLM"),
150
151
    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
    "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
152
153
154
    "MiniMaxForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxText01ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxM1ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
155
    "MiniMaxM2ForCausalLM": ("minimax_m2", "MiniMaxM2ForCausalLM"),
156
    "MistralForCausalLM": ("llama", "LlamaForCausalLM"),
157
    "MistralLarge3ForCausalLM": ("mistral_large_3", "MistralLarge3ForCausalLM"),
158
159
160
161
    "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
    # transformers's mpt class has lower case
    "MptForCausalLM": ("mpt", "MPTForCausalLM"),
    "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
162
    "MiMoForCausalLM": ("mimo", "MiMoForCausalLM"),
163
    "MiMoV2FlashForCausalLM": ("mimo_v2_flash", "MiMoV2FlashForCausalLM"),
164
    "NemotronForCausalLM": ("nemotron", "NemotronForCausalLM"),
Luis Vega's avatar
Luis Vega committed
165
    "NemotronHForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
166
    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
167
    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
168
    "Olmo3ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
169
170
171
    "OlmoeForCausalLM": ("olmoe", "OlmoeForCausalLM"),
    "OPTForCausalLM": ("opt", "OPTForCausalLM"),
    "OrionForCausalLM": ("orion", "OrionForCausalLM"),
172
    "OuroForCausalLM": ("ouro", "OuroForCausalLM"),
173
    "PanguEmbeddedForCausalLM": ("openpangu", "PanguEmbeddedForCausalLM"),
174
    "PanguProMoEV2ForCausalLM": ("openpangu", "PanguProMoEV2ForCausalLM"),
175
    "PanguUltraMoEForCausalLM": ("openpangu", "PanguUltraMoEForCausalLM"),
176
177
178
179
    "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
    "PhiForCausalLM": ("phi", "PhiForCausalLM"),
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
    "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
180
    "Plamo2ForCausalLM": ("plamo2", "Plamo2ForCausalLM"),
181
    "Plamo3ForCausalLM": ("plamo3", "Plamo3ForCausalLM"),
182
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
183
184
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
185
186
    "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),
    "Qwen3MoeForCausalLM": ("qwen3_moe", "Qwen3MoeForCausalLM"),
187
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
188
    "SeedOssForCausalLM": ("seed_oss", "SeedOssForCausalLM"),
Li Xie's avatar
Li Xie committed
189
    "Step1ForCausalLM": ("step1", "Step1ForCausalLM"),
Song's avatar
Song committed
190
    "Step3TextForCausalLM": ("step3_text", "Step3TextForCausalLM"),
191
192
193
194
    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
195
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
196
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
197
    "TeleFLMForCausalLM": ("teleflm", "TeleFLMForCausalLM"),
198
    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
199
    "Zamba2ForCausalLM": ("zamba2", "Zamba2ForCausalLM"),
200
201
202
}

_EMBEDDING_MODELS = {
203
    # [Text-only]
204
    "BertModel": ("bert", "BertEmbeddingModel"),
205
    "BertSpladeSparseEmbeddingModel": ("bert", "BertSpladeSparseEmbeddingModel"),
206
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
207
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
208
    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
209
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
210
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
211
    "GritLM": ("gritlm", "GritLM"),
212
213
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
214
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
215
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
216
    "LlamaBidirectionalModel": ("llama", "LlamaBidirectionalModel"),
217
    "LlamaModel": ("llama", "LlamaForCausalLM"),
218
219
    **{
        # Multiple models share the same architecture, so we include them all
220
221
        k: (mod, arch)
        for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
222
223
        if arch == "LlamaForCausalLM"
    },
224
    "MistralModel": ("llama", "LlamaForCausalLM"),
225
    "ModernBertModel": ("modernbert", "ModernBertModel"),
226
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
227
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
228
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
229
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
230
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
231
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
232
233
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
234
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
235
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
236
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
237
    # [Multimodal]
238
    "CLIPModel": ("clip", "CLIPEmbeddingModel"),
239
240
241
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
242
    ),
243
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
244
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
245
    "SiglipModel": ("siglip", "SiglipEmbeddingModel"),
246
247
    # Technically Terratorch models work on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
248
    # models for the time being.
249
250
    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
    "Terratorch": ("terratorch", "Terratorch"),
251
252
}

253
254
_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
255
    "BertForTokenClassification": ("bert", "BertForTokenClassification"),
256
257
258
259
    "GteNewForSequenceClassification": (
        "bert_with_rope",
        "GteNewForSequenceClassification",
    ),
260
261
262
263
264
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"),
    "LlamaBidirectionalForSequenceClassification": (
        "llama",
        "LlamaBidirectionalForSequenceClassification",
    ),
265
266
267
268
    "ModernBertForSequenceClassification": (
        "modernbert",
        "ModernBertForSequenceClassification",
    ),
269
270
271
272
    "ModernBertForTokenClassification": (
        "modernbert",
        "ModernBertForTokenClassification",
    ),
273
274
275
276
277
    "RobertaForSequenceClassification": ("roberta", "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": (
        "roberta",
        "RobertaForSequenceClassification",
    ),
278
279
}

280
_MULTIMODAL_MODELS = {
281
    # [Decoder-only]
282
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
283
284
285
286
    "AudioFlamingo3ForConditionalGeneration": (
        "audioflamingo3",
        "AudioFlamingo3ForConditionalGeneration",
    ),
287
288
289
    "AyaVisionForConditionalGeneration": (
        "aya_vision",
        "AyaVisionForConditionalGeneration",
290
    ),
291
    "BagelForConditionalGeneration": ("bagel", "BagelForConditionalGeneration"),
292
    "BeeForConditionalGeneration": ("bee", "BeeForConditionalGeneration"),
293
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
294
295
296
    "ChameleonForConditionalGeneration": (
        "chameleon",
        "ChameleonForConditionalGeneration",
297
    ),
298
299
300
    "Cohere2VisionForConditionalGeneration": (
        "cohere2_vision",
        "Cohere2VisionForConditionalGeneration",
301
    ),
302
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
303
    "DeepseekOCRForCausalLM": ("deepseek_ocr", "DeepseekOCRForCausalLM"),
Roger Wang's avatar
Roger Wang committed
304
    "DotsOCRForCausalLM": ("dots_ocr", "DotsOCRForCausalLM"),
305
306
307
    "Ernie4_5_VLMoeForConditionalGeneration": (
        "ernie45_vl",
        "Ernie4_5_VLMoeForConditionalGeneration",
308
    ),
309
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
310
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
311
312
313
    "Gemma3nForConditionalGeneration": (
        "gemma3n_mm",
        "Gemma3nForConditionalGeneration",
314
    ),
315
    "GlmAsrForConditionalGeneration": ("glmasr", "GlmAsrForConditionalGeneration"),
316
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
317
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),  # noqa: E501
Jee Jee Li's avatar
Jee Jee Li committed
318
    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),  # noqa: E501
319
320
321
    "GraniteSpeechForConditionalGeneration": (
        "granite_speech",
        "GraniteSpeechForConditionalGeneration",
322
    ),
323
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
324
325
326
327
    "HunYuanVLForConditionalGeneration": (
        "hunyuan_vision",
        "HunYuanVLForConditionalGeneration",
    ),
ltd0924's avatar
ltd0924 committed
328
    "StepVLForConditionalGeneration": ("step_vl", "StepVLForConditionalGeneration"),
329
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
330
    "NemotronH_Nano_VL_V2": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
Zero's avatar
Zero committed
331
332
333
334
    "OpenCUAForConditionalGeneration": (
        "opencua",
        "OpenCUAForConditionalGeneration",
    ),
335
336
337
    "InternS1ForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
338
    ),
339
340
341
    "InternVLForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
342
    ),
343
344
345
346
    "Idefics3ForConditionalGeneration": (
        "idefics3",
        "Idefics3ForConditionalGeneration",
    ),
oscardev256's avatar
oscardev256 committed
347
    "IsaacForConditionalGeneration": ("isaac", "IsaacForConditionalGeneration"),
348
    "SmolVLMForConditionalGeneration": ("smolvlm", "SmolVLMForConditionalGeneration"),  # noqa: E501
349
    "KananaVForConditionalGeneration": ("kanana_v", "KananaVForConditionalGeneration"),
350
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
351
352
353
    "KeyeVL1_5ForConditionalGeneration": (
        "keye_vl1_5",
        "KeyeVL1_5ForConditionalGeneration",
354
    ),
355
    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
356
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
357
358
359
360
    "LightOnOCRForConditionalGeneration": (
        "lightonocr",
        "LightOnOCRForConditionalGeneration",
    ),
361
    "Lfm2VlForConditionalGeneration": ("lfm2_vl", "Lfm2VLForConditionalGeneration"),
362
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
363
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),  # noqa: E501
364
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
365
366
367
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
368
    ),
369
370
371
    "LlavaNextVideoForConditionalGeneration": (
        "llava_next_video",
        "LlavaNextVideoForConditionalGeneration",
372
    ),
373
374
375
    "LlavaOnevisionForConditionalGeneration": (
        "llava_onevision",
        "LlavaOnevisionForConditionalGeneration",
376
    ),
377
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
378
    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
379
380
381
    "MiniMaxVL01ForConditionalGeneration": (
        "minimax_vl_01",
        "MiniMaxVL01ForConditionalGeneration",
382
    ),
383
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
384
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
385
386
387
    "Mistral3ForConditionalGeneration": (
        "mistral3",
        "Mistral3ForConditionalGeneration",
388
    ),
389
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
390
    "Molmo2ForConditionalGeneration": ("molmo2", "Molmo2ForConditionalGeneration"),
391
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
392
    "Ovis": ("ovis", "Ovis"),
393
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
394
395
396
397
    "PaddleOCRVLForConditionalGeneration": (
        "paddleocr_vl",
        "PaddleOCRVLForConditionalGeneration",
    ),
398
399
400
401
    "PaliGemmaForConditionalGeneration": (
        "paligemma",
        "PaliGemmaForConditionalGeneration",
    ),
402
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
403
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
404
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
405
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),  # noqa: E501
406
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
407
408
409
    "Qwen2_5_VLForConditionalGeneration": (
        "qwen2_5_vl",
        "Qwen2_5_VLForConditionalGeneration",
410
    ),
411
412
413
    "Qwen2AudioForConditionalGeneration": (
        "qwen2_audio",
        "Qwen2AudioForConditionalGeneration",
414
    ),
415
416
417
    "Qwen2_5OmniModel": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
418
    ),
419
420
421
    "Qwen2_5OmniForConditionalGeneration": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
422
    ),
423
424
425
426
    "Qwen3OmniMoeForConditionalGeneration": (
        "qwen3_omni_moe_thinker",
        "Qwen3OmniMoeThinkerForConditionalGeneration",
    ),
427
    "Qwen3VLForConditionalGeneration": ("qwen3_vl", "Qwen3VLForConditionalGeneration"),  # noqa: E501
428
429
430
    "Qwen3VLMoeForConditionalGeneration": (
        "qwen3_vl_moe",
        "Qwen3VLMoeForConditionalGeneration",
431
    ),
432
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
Song's avatar
Song committed
433
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),  # noqa: E501
汪志鹏's avatar
汪志鹏 committed
434
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),  # noqa: E501
435
436
437
    "Tarsier2ForConditionalGeneration": (
        "qwen2_vl",
        "Tarsier2ForConditionalGeneration",
438
    ),
439
    "UltravoxModel": ("ultravox", "UltravoxModel"),
Patrick von Platen's avatar
Patrick von Platen committed
440
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),  # noqa: E501
Patrick von Platen's avatar
Patrick von Platen committed
441
    "VoxtralStreamingGeneration": ("voxtral_streaming", "VoxtralStreamingGeneration"),  # noqa: E501
442
    # [Encoder-decoder]
443
444
445
446
    "NemotronParseForConditionalGeneration": (
        "nemotron_parse",
        "NemotronParseForConditionalGeneration",
    ),
447
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
448
}
449
450

_SPECULATIVE_DECODING_MODELS = {
451
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
452
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
453
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
454
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
455
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
456
    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
457
    "Eagle3Qwen2_5vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
458
    "Eagle3Qwen3vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
459
460
461
462
    "EagleMistralLarge3ForCausalLM": (
        "mistral_large_3_eagle",
        "EagleMistralLarge3ForCausalLM",
    ),
463
    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
464
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
465
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
Kyungmin Lee's avatar
Kyungmin Lee committed
466
    "ExaoneMoeMTP": ("exaone_moe_mtp", "ExaoneMoeMTP"),
XuruiYang's avatar
XuruiYang committed
467
    "LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
468
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
469
    "Glm4MoeLiteMTPModel": ("glm4_moe_lite_mtp", "Glm4MoeLiteMTP"),
470
    "MedusaModel": ("medusa", "Medusa"),
471
    "OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
472
    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
473
474
475
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
476
}
477

478
_TRANSFORMERS_SUPPORTED_MODELS = {
479
480
481
    # Text generation models
    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
    # Multimodal models
482
483
484
485
    "Emu3ForConditionalGeneration": (
        "transformers",
        "TransformersMultiModalForCausalLM",
    ),
486
487
488
}

_TRANSFORMERS_BACKEND_MODELS = {
489
    # Text generation models
490
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
    "TransformersMoEForCausalLM": ("transformers", "TransformersMoEForCausalLM"),
    # Multimodal models
    "TransformersMultiModalForCausalLM": (
        "transformers",
        "TransformersMultiModalForCausalLM",
    ),
    "TransformersMultiModalMoEForCausalLM": (
        "transformers",
        "TransformersMultiModalMoEForCausalLM",
    ),
    # Embedding models
    "TransformersEmbeddingModel": ("transformers", "TransformersEmbeddingModel"),
    "TransformersMoEEmbeddingModel": ("transformers", "TransformersMoEEmbeddingModel"),
    "TransformersMultiModalEmbeddingModel": (
        "transformers",
        "TransformersMultiModalEmbeddingModel",
    ),
    # Sequence classification models
509
    "TransformersForSequenceClassification": (
510
        "transformers",
511
        "TransformersForSequenceClassification",
512
    ),
513
    "TransformersMoEForSequenceClassification": (
514
        "transformers",
515
        "TransformersMoEForSequenceClassification",
516
    ),
517
518
519
    "TransformersMultiModalForSequenceClassification": (
        "transformers",
        "TransformersMultiModalForSequenceClassification",
520
    ),
521
}
522

523
_VLLM_MODELS = {
524
    **_TEXT_GENERATION_MODELS,
525
    **_EMBEDDING_MODELS,
526
    **_CROSS_ENCODER_MODELS,
527
    **_MULTIMODAL_MODELS,
528
    **_SPECULATIVE_DECODING_MODELS,
529
530
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
531
532
}

533
534
535
536
# 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.
537
_SUBPROCESS_COMMAND = [sys.executable, "-m", "vllm.model_executor.models.registry"]
538

539
_PREVIOUSLY_SUPPORTED_MODELS = {
540
    "MotifForCausalLM": "0.10.2",
541
    "Phi3SmallForCausalLM": "0.9.2",
542
    "Phi4FlashForCausalLM": "0.10.2",
543
    "Phi4MultimodalForCausalLM": "0.12.0",
544
545
546
547
548
549
550
551
552
    # encoder-decoder models except whisper
    # have been removed for V0 deprecation.
    "BartModel": "0.10.2",
    "BartForConditionalGeneration": "0.10.2",
    "DonutForConditionalGeneration": "0.10.2",
    "Florence2ForConditionalGeneration": "0.10.2",
    "MBartForConditionalGeneration": "0.10.2",
    "MllamaForConditionalGeneration": "0.10.2",
}
553

554

555
556
@dataclass(frozen=True)
class _ModelInfo:
557
    architecture: str
558
    is_text_generation_model: bool
559
    is_pooling_model: bool
560
    attn_type: AttnTypeStr
561
562
    default_seq_pooling_type: SequencePoolingType
    default_tok_pooling_type: TokenPoolingType
563
    supports_cross_encoding: bool
564
    supports_multimodal: bool
565
    supports_multimodal_raw_input_only: bool
Patrick von Platen's avatar
Patrick von Platen committed
566
    requires_raw_input_tokens: bool
567
    supports_multimodal_encoder_tp_data: bool
568
    supports_pp: bool
569
570
    has_inner_state: bool
    is_attention_free: bool
571
    is_hybrid: bool
572
    has_noops: bool
573
    supports_mamba_prefix_caching: bool
574
    supports_transcription: bool
575
    supports_transcription_only: bool
576
577

    @staticmethod
578
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
579
        return _ModelInfo(
580
            architecture=model.__name__,
581
            is_text_generation_model=is_text_generation_model(model),
582
            is_pooling_model=is_pooling_model(model),
583
584
            default_seq_pooling_type=get_default_seq_pooling_type(model),
            default_tok_pooling_type=get_default_tok_pooling_type(model),
585
            attn_type=get_attn_type(model),
586
            supports_cross_encoding=supports_cross_encoding(model),
587
            supports_multimodal=supports_multimodal(model),
588
589
590
            supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
                model
            ),
Patrick von Platen's avatar
Patrick von Platen committed
591
            requires_raw_input_tokens=requires_raw_input_tokens(model),
592
593
594
            supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
                model
            ),
595
            supports_pp=supports_pp(model),
596
597
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
598
            is_hybrid=is_hybrid(model),
599
            supports_mamba_prefix_caching=supports_mamba_prefix_caching(model),
600
            supports_transcription=supports_transcription(model),
601
602
603
            supports_transcription_only=(
                supports_transcription(model) and model.supports_transcription_only
            ),
604
            has_noops=has_noops(model),
605
        )
606
607


608
609
610
611
class _BaseRegisteredModel(ABC):
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
612

613
    @abstractmethod
614
    def load_model_cls(self) -> type[nn.Module]:
615
        raise NotImplementedError
616
617


618
619
620
621
622
623
624
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
625
    model_cls: type[nn.Module]
626
627

    @staticmethod
628
    def from_model_cls(model_cls: type[nn.Module]):
629
630
631
632
633
634
635
636
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

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

637
    def load_model_cls(self) -> type[nn.Module]:
638
639
640
641
642
643
644
645
        return self.model_cls


@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """
646

647
648
649
    module_name: str
    class_name: str

650
651
652
653
654
655
656
657
    @staticmethod
    def _get_cache_dir() -> Path:
        return Path(envs.VLLM_CACHE_ROOT) / "modelinfos"

    def _get_cache_filename(self) -> str:
        cls_name = f"{self.module_name}-{self.class_name}".replace(".", "-")
        return f"{cls_name}.json"

658
    def _load_modelinfo_from_cache(self, module_hash: str) -> _ModelInfo | None:
659
660
        try:
            try:
661
                modelinfo_path = self._get_cache_dir() / self._get_cache_filename()
662
663
664
                with open(modelinfo_path, encoding="utf-8") as file:
                    mi_dict = json.load(file)
            except FileNotFoundError:
665
                logger.debug(
666
                    "Cached model info file for class %s.%s not found",
667
668
669
                    self.module_name,
                    self.class_name,
                )
670
671
672
                return None

            if mi_dict["hash"] != module_hash:
673
                logger.debug(
674
                    "Cached model info file for class %s.%s is stale",
675
676
677
                    self.module_name,
                    self.class_name,
                )
678
679
680
681
682
                return None

            # file not changed, use cached _ModelInfo properties
            return _ModelInfo(**mi_dict["modelinfo"])
        except Exception:
683
            logger.debug(
684
                "Cached model info for class %s.%s error. ",
685
686
687
                self.module_name,
                self.class_name,
            )
688
689
            return None

690
    def _save_modelinfo_to_cache(self, mi: _ModelInfo, module_hash: str) -> None:
691
692
        """save dictionary json file to cache"""
        from vllm.model_executor.model_loader.weight_utils import atomic_writer
693

694
695
696
697
698
699
700
701
        try:
            modelinfo_dict = {
                "hash": module_hash,
                "modelinfo": asdict(mi),
            }
            cache_dir = self._get_cache_dir()
            cache_dir.mkdir(parents=True, exist_ok=True)
            modelinfo_path = cache_dir / self._get_cache_filename()
702
            with atomic_writer(modelinfo_path, encoding="utf-8") as f:
703
704
705
706
707
                json.dump(modelinfo_dict, f, indent=2)
        except Exception:
            logger.exception("Error saving model info cache.")

    @logtime(logger=logger, msg="Registry inspect model class")
708
    def inspect_model_cls(self) -> _ModelInfo:
709
        model_path = Path(__file__).parent / f"{self.module_name.split('.')[-1]}.py"
710
        module_hash = None
711

712
713
        if model_path.exists():
            with open(model_path, "rb") as f:
714
                module_hash = safe_hash(f.read(), usedforsecurity=False).hexdigest()
715
716
717

            mi = self._load_modelinfo_from_cache(module_hash)
            if mi is not None:
718
                logger.debug(
719
                    "Loaded model info for class %s.%s from cache",
720
721
722
                    self.module_name,
                    self.class_name,
                )
723
724
                return mi
            else:
725
                logger.debug(
726
                    "Cache model info for class %s.%s miss. Loading model instead.",
727
728
729
                    self.module_name,
                    self.class_name,
                )
730
731
732

        # Performed in another process to avoid initializing CUDA
        mi = _run_in_subprocess(
733
734
735
736
737
            lambda: _ModelInfo.from_model_cls(self.load_model_cls())
        )
        logger.debug(
            "Loaded model info for class %s.%s", self.module_name, self.class_name
        )
738
739

        # save cache file
740
741
        if module_hash is not None:
            self._save_modelinfo_to_cache(mi, module_hash)
742
743

        return mi
744

745
    def load_model_cls(self) -> type[nn.Module]:
746
747
748
749
750
751
752
753
        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,
754
) -> type[nn.Module] | None:
755
    from vllm.platforms import current_platform
756

757
    current_platform.verify_model_arch(model_arch)
758
759
760
    try:
        return model.load_model_cls()
    except Exception:
761
        logger.exception("Error in loading model architecture '%s'", model_arch)
762
        return None
763
764


765
766
767
768
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
769
) -> _ModelInfo | None:
770
771
772
    try:
        return model.inspect_model_cls()
    except Exception:
773
        logger.exception("Error in inspecting model architecture '%s'", model_arch)
774
        return None
775
776


777
778
779
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
780
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
781

782
    def get_supported_archs(self) -> Set[str]:
783
        return self.models.keys()
784

785
786
787
    def register_model(
        self,
        model_arch: str,
788
        model_cls: type[nn.Module] | str,
789
    ) -> None:
790
791
792
        """
        Register an external model to be used in vLLM.

793
        `model_cls` can be either:
794

795
        - A [`torch.nn.Module`][] class directly referencing the model.
796
        - A string in the format `<module>:<class>` which can be used to
797
798
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
799
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
800
        """
801
802
803
804
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

805
        if model_arch in self.models:
806
807
            logger.warning(
                "Model architecture %s is already registered, and will be "
808
809
810
811
                "overwritten by the new model class %s.",
                model_arch,
                model_cls,
            )
812
813
814
815
816
817

        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)
818

819
            model = _LazyRegisteredModel(*split_str)
820
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
821
            model = _RegisteredModel.from_model_cls(model_cls)
822
        else:
823
824
825
826
            msg = (
                "`model_cls` should be a string or PyTorch model class, "
                f"not a {type(model_arch)}"
            )
827
            raise TypeError(msg)
828

829
        self.models[model_arch] = model
830

831
    def _raise_for_unsupported(self, architectures: list[str]):
832
        all_supported_archs = self.get_supported_archs()
833

834
835
836
        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
837
838
                "to be inspected. Please check the logs for more details."
            )
839

840
841
842
843
844
845
846
847
        for arch in architectures:
            if arch in _PREVIOUSLY_SUPPORTED_MODELS:
                previous_version = _PREVIOUSLY_SUPPORTED_MODELS[arch]

                raise ValueError(
                    f"Model architecture {arch} was supported in vLLM until "
                    f"v{previous_version}, and is not supported anymore. "
                    "Please use an older version of vLLM if you want to "
848
849
                    "use this model architecture."
                )
850

851
852
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
853
854
            f"Supported architectures: {all_supported_archs}"
        )
855

856
    def _try_load_model_cls(self, model_arch: str) -> type[nn.Module] | None:
857
858
        if model_arch not in self.models:
            return None
859

860
        return _try_load_model_cls(model_arch, self.models[model_arch])
861

862
    def _try_inspect_model_cls(self, model_arch: str) -> _ModelInfo | None:
863
864
        if model_arch not in self.models:
            return None
865

866
867
868
869
870
871
        return _try_inspect_model_cls(model_arch, self.models[model_arch])

    def _try_resolve_transformers(
        self,
        architecture: str,
        model_config: ModelConfig,
872
    ) -> str | None:
873
874
875
        if architecture in _TRANSFORMERS_BACKEND_MODELS:
            return architecture

876
877
878
        auto_map: dict[str, str] = (
            getattr(model_config.hf_config, "auto_map", None) or dict()
        )
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894

        # Make sure that config class is always initialized before model class,
        # otherwise the model class won't be able to access the config class,
        # the expected auto_map should have correct order like:
        # "auto_map": {
        #     "AutoConfig": "<your-repo-name>--<config-name>",
        #     "AutoModel": "<your-repo-name>--<config-name>",
        #     "AutoModelFor<Task>": "<your-repo-name>--<config-name>",
        # },
        for prefix in ("AutoConfig", "AutoModel"):
            for name, module in auto_map.items():
                if name.startswith(prefix):
                    try_get_class_from_dynamic_module(
                        module,
                        model_config.model,
                        revision=model_config.revision,
895
                        trust_remote_code=model_config.trust_remote_code,
896
897
898
899
900
901
902
903
904
905
906
907
                        warn_on_fail=False,
                    )

        model_module = getattr(transformers, architecture, None)

        if model_module is None:
            for name, module in auto_map.items():
                if name.startswith("AutoModel"):
                    model_module = try_get_class_from_dynamic_module(
                        module,
                        model_config.model,
                        revision=model_config.revision,
908
                        trust_remote_code=model_config.trust_remote_code,
909
910
911
912
913
                        warn_on_fail=True,
                    )
                    if model_module is not None:
                        break
            else:
914
                if model_config.model_impl != "transformers":
915
916
917
918
919
920
921
                    return None

                raise ValueError(
                    f"Cannot find model module. {architecture!r} is not a "
                    "registered model in the Transformers library (only "
                    "relevant if the model is meant to be in Transformers) "
                    "and 'AutoModel' is not present in the model config's "
922
923
                    "'auto_map' (relevant if the model is custom)."
                )
924
925

        if not model_module.is_backend_compatible():
926
            if model_config.model_impl != "transformers":
927
                return None
928

929
930
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
931
932
                "is not compatible with vLLM."
            )
933

934
        return model_config._get_transformers_backend_cls()
935

936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
    def _normalize_arch(
        self,
        architecture: str,
        model_config: ModelConfig,
    ) -> str:
        if architecture in self.models:
            return architecture

        # This may be called in order to resolve runner_type and convert_type
        # in the first place, in which case we consider the default match
        match = try_match_architecture_defaults(
            architecture,
            runner_type=getattr(model_config, "runner_type", None),
            convert_type=getattr(model_config, "convert_type", None),
        )
        if match:
            suffix, _ = match

            # Get the name of the base model to convert
            for repl_suffix, _ in iter_architecture_defaults():
                base_arch = architecture.replace(suffix, repl_suffix)
                if base_arch in self.models:
                    return base_arch

        return architecture
961

962
963
    def inspect_model_cls(
        self,
964
        architectures: str | list[str],
965
        model_config: ModelConfig,
966
    ) -> tuple[_ModelInfo, str]:
967
968
        if isinstance(architectures, str):
            architectures = [architectures]
969
970
        if not architectures:
            raise ValueError("No model architectures are specified")
971
972

        # Require transformers impl
973
        if model_config.model_impl == "transformers":
974
            arch = self._try_resolve_transformers(architectures[0], model_config)
975
976
977
978
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)
979
        elif model_config.model_impl == "terratorch":
980
981
            model_info = self._try_inspect_model_cls("Terratorch")
            return (model_info, "Terratorch")
982

983
        # Fallback to transformers impl (after resolving convert_type)
984
985
986
987
988
989
        if (
            all(arch not in self.models for arch in architectures)
            and model_config.model_impl == "auto"
            and getattr(model_config, "convert_type", "none") == "none"
        ):
            arch = self._try_resolve_transformers(architectures[0], model_config)
990
991
992
993
994
995
996
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

        for arch in architectures:
            normalized_arch = self._normalize_arch(arch, model_config)
997
            model_info = self._try_inspect_model_cls(normalized_arch)
998
            if model_info is not None:
999
                return (model_info, arch)
1000

1001
        # Fallback to transformers impl (before resolving runner_type)
1002
1003
1004
1005
1006
        if (
            all(arch not in self.models for arch in architectures)
            and model_config.model_impl == "auto"
        ):
            arch = self._try_resolve_transformers(architectures[0], model_config)
1007
1008
1009
1010
1011
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

1012
        return self._raise_for_unsupported(architectures)
1013

1014
1015
    def resolve_model_cls(
        self,
1016
        architectures: str | list[str],
1017
        model_config: ModelConfig,
1018
    ) -> tuple[type[nn.Module], str]:
1019
1020
        if isinstance(architectures, str):
            architectures = [architectures]
1021
1022
        if not architectures:
            raise ValueError("No model architectures are specified")
1023
1024

        # Require transformers impl
1025
        if model_config.model_impl == "transformers":
1026
            arch = self._try_resolve_transformers(architectures[0], model_config)
1027
1028
1029
1030
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)
1031
        elif model_config.model_impl == "terratorch":
1032
1033
1034
1035
            arch = "Terratorch"
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
1036

1037
        # Fallback to transformers impl (after resolving convert_type)
1038
1039
1040
1041
1042
1043
        if (
            all(arch not in self.models for arch in architectures)
            and model_config.model_impl == "auto"
            and getattr(model_config, "convert_type", "none") == "none"
        ):
            arch = self._try_resolve_transformers(architectures[0], model_config)
1044
1045
1046
1047
1048
1049
1050
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

        for arch in architectures:
            normalized_arch = self._normalize_arch(arch, model_config)
1051
            model_cls = self._try_load_model_cls(normalized_arch)
1052
1053
            if model_cls is not None:
                return (model_cls, arch)
1054

1055
        # Fallback to transformers impl (before resolving runner_type)
1056
1057
1058
1059
1060
        if (
            all(arch not in self.models for arch in architectures)
            and model_config.model_impl == "auto"
        ):
            arch = self._try_resolve_transformers(architectures[0], model_config)
1061
1062
1063
1064
1065
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

1066
        return self._raise_for_unsupported(architectures)
1067

1068
1069
    def is_text_generation_model(
        self,
1070
        architectures: str | list[str],
1071
        model_config: ModelConfig,
1072
    ) -> bool:
1073
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1074
        return model_cls.is_text_generation_model
1075

1076
    def is_pooling_model(
1077
        self,
1078
        architectures: str | list[str],
1079
        model_config: ModelConfig,
1080
    ) -> bool:
1081
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1082
        return model_cls.is_pooling_model
1083

1084
1085
    def is_cross_encoder_model(
        self,
1086
        architectures: str | list[str],
1087
        model_config: ModelConfig,
1088
    ) -> bool:
1089
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1090
        return model_cls.supports_cross_encoding
1091

1092
1093
    def is_multimodal_model(
        self,
1094
        architectures: str | list[str],
1095
        model_config: ModelConfig,
1096
    ) -> bool:
1097
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1098
        return model_cls.supports_multimodal
1099

1100
    def is_multimodal_raw_input_only_model(
1101
        self,
1102
        architectures: str | list[str],
1103
        model_config: ModelConfig,
1104
    ) -> bool:
1105
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1106
        return model_cls.supports_multimodal_raw_input_only
1107

1108
1109
    def is_pp_supported_model(
        self,
1110
        architectures: str | list[str],
1111
        model_config: ModelConfig,
1112
    ) -> bool:
1113
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1114
        return model_cls.supports_pp
1115

1116
1117
    def model_has_inner_state(
        self,
1118
        architectures: str | list[str],
1119
        model_config: ModelConfig,
1120
    ) -> bool:
1121
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1122
        return model_cls.has_inner_state
1123

1124
1125
    def is_attention_free_model(
        self,
1126
        architectures: str | list[str],
1127
        model_config: ModelConfig,
1128
    ) -> bool:
1129
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1130
        return model_cls.is_attention_free
1131

1132
1133
    def is_hybrid_model(
        self,
1134
        architectures: str | list[str],
1135
        model_config: ModelConfig,
1136
    ) -> bool:
1137
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1138
1139
        return model_cls.is_hybrid

1140
1141
    def is_noops_model(
        self,
1142
        architectures: str | list[str],
1143
        model_config: ModelConfig,
1144
    ) -> bool:
1145
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1146
1147
        return model_cls.has_noops

1148
1149
    def is_transcription_model(
        self,
1150
        architectures: str | list[str],
1151
        model_config: ModelConfig,
1152
    ) -> bool:
1153
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1154
1155
        return model_cls.supports_transcription

1156
1157
    def is_transcription_only_model(
        self,
1158
        architectures: str | list[str],
1159
        model_config: ModelConfig,
1160
    ) -> bool:
1161
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1162
1163
        return model_cls.supports_transcription_only

1164

1165
1166
1167
1168
1169
1170
1171
1172
1173
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()
    }
)
1174
1175
1176
1177
1178

_T = TypeVar("_T")


def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
1179
1180
1181
1182
1183
    # 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")

1184
        # `cloudpickle` allows pickling lambda functions directly
1185
        import cloudpickle
1186

1187
        input_bytes = cloudpickle.dumps((fn, output_filepath))
1188
1189
1190

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
1191
1192
1193
        returned = subprocess.run(
            _SUBPROCESS_COMMAND, input=input_bytes, capture_output=True
        )
1194
1195
1196
1197
1198
1199

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
1200
1201
1202
            raise RuntimeError(
                f"Error raised in subprocess:\n{returned.stderr.decode()}"
            ) from e
1203

1204
        with open(output_filepath, "rb") as f:
1205
1206
1207
1208
1209
1210
            return pickle.load(f)


def _run() -> None:
    # Setup plugins
    from vllm.plugins import load_general_plugins
1211

1212
1213
1214
1215
1216
    load_general_plugins()

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

    result = fn()
1217
1218
1219

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
1220
1221
1222


if __name__ == "__main__":
1223
    _run()