registry.py 48.2 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": ("mistral", "MistralForCausalLM"),
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
    "NemotronHPuzzleForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
167
    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
168
    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
169
    "Olmo3ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
170
171
172
    "OlmoeForCausalLM": ("olmoe", "OlmoeForCausalLM"),
    "OPTForCausalLM": ("opt", "OPTForCausalLM"),
    "OrionForCausalLM": ("orion", "OrionForCausalLM"),
173
    "OuroForCausalLM": ("ouro", "OuroForCausalLM"),
174
    "PanguEmbeddedForCausalLM": ("openpangu", "PanguEmbeddedForCausalLM"),
175
    "PanguProMoEV2ForCausalLM": ("openpangu", "PanguProMoEV2ForCausalLM"),
176
    "PanguUltraMoEForCausalLM": ("openpangu", "PanguUltraMoEForCausalLM"),
177
178
179
180
    "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
    "PhiForCausalLM": ("phi", "PhiForCausalLM"),
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
    "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
181
    "Plamo2ForCausalLM": ("plamo2", "Plamo2ForCausalLM"),
182
    "Plamo3ForCausalLM": ("plamo3", "Plamo3ForCausalLM"),
183
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
184
185
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
186
187
    "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),
    "Qwen3MoeForCausalLM": ("qwen3_moe", "Qwen3MoeForCausalLM"),
188
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
189
    "SeedOssForCausalLM": ("seed_oss", "SeedOssForCausalLM"),
Li Xie's avatar
Li Xie committed
190
    "Step1ForCausalLM": ("step1", "Step1ForCausalLM"),
Song's avatar
Song committed
191
    "Step3TextForCausalLM": ("step3_text", "Step3TextForCausalLM"),
192
193
194
195
    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
196
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
197
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
198
    "TeleFLMForCausalLM": ("teleflm", "TeleFLMForCausalLM"),
199
    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
200
    "Zamba2ForCausalLM": ("zamba2", "Zamba2ForCausalLM"),
201
202
203
}

_EMBEDDING_MODELS = {
204
    # [Text-only]
205
    "BertModel": ("bert", "BertEmbeddingModel"),
206
    "BertSpladeSparseEmbeddingModel": ("bert", "BertSpladeSparseEmbeddingModel"),
207
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
208
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
209
    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
210
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
211
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
212
    "GritLM": ("gritlm", "GritLM"),
213
214
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
215
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
216
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
217
    "LlamaBidirectionalModel": ("llama", "LlamaBidirectionalModel"),
218
    "LlamaModel": ("llama", "LlamaForCausalLM"),
219
220
    **{
        # Multiple models share the same architecture, so we include them all
221
222
        k: (mod, arch)
        for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
223
224
        if arch == "LlamaForCausalLM"
    },
225
    "MistralModel": ("llama", "LlamaForCausalLM"),
226
    "ModernBertModel": ("modernbert", "ModernBertModel"),
227
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
228
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
229
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
230
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
231
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
232
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
233
234
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
235
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
236
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
237
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
238
    "BgeM3EmbeddingModel": ("roberta", "BgeM3EmbeddingModel"),
239
    # [Multimodal]
240
    "CLIPModel": ("clip", "CLIPEmbeddingModel"),
241
242
243
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
244
    ),
245
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
246
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
247
    "SiglipModel": ("siglip", "SiglipEmbeddingModel"),
248
249
    # Technically Terratorch models work on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
250
    # models for the time being.
251
252
    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
    "Terratorch": ("terratorch", "Terratorch"),
253
254
}

255
256
_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
257
    "BertForTokenClassification": ("bert", "BertForTokenClassification"),
258
259
260
261
    "GteNewForSequenceClassification": (
        "bert_with_rope",
        "GteNewForSequenceClassification",
    ),
262
263
264
265
266
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"),
    "LlamaBidirectionalForSequenceClassification": (
        "llama",
        "LlamaBidirectionalForSequenceClassification",
    ),
267
268
269
270
    "ModernBertForSequenceClassification": (
        "modernbert",
        "ModernBertForSequenceClassification",
    ),
271
272
273
274
    "ModernBertForTokenClassification": (
        "modernbert",
        "ModernBertForTokenClassification",
    ),
275
276
277
278
279
    "RobertaForSequenceClassification": ("roberta", "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": (
        "roberta",
        "RobertaForSequenceClassification",
    ),
280
281
}

282
_MULTIMODAL_MODELS = {
283
    # [Decoder-only]
284
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
285
286
287
288
    "AudioFlamingo3ForConditionalGeneration": (
        "audioflamingo3",
        "AudioFlamingo3ForConditionalGeneration",
    ),
289
290
291
    "AyaVisionForConditionalGeneration": (
        "aya_vision",
        "AyaVisionForConditionalGeneration",
292
    ),
293
    "BagelForConditionalGeneration": ("bagel", "BagelForConditionalGeneration"),
294
    "BeeForConditionalGeneration": ("bee", "BeeForConditionalGeneration"),
295
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
296
297
298
    "ChameleonForConditionalGeneration": (
        "chameleon",
        "ChameleonForConditionalGeneration",
299
    ),
300
301
302
    "Cohere2VisionForConditionalGeneration": (
        "cohere2_vision",
        "Cohere2VisionForConditionalGeneration",
303
    ),
304
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
305
    "DeepseekOCRForCausalLM": ("deepseek_ocr", "DeepseekOCRForCausalLM"),
Roger Wang's avatar
Roger Wang committed
306
    "DotsOCRForCausalLM": ("dots_ocr", "DotsOCRForCausalLM"),
307
308
309
310
    "Eagle2_5_VLForConditionalGeneration": (
        "eagle2_5_vl",
        "Eagle2_5_VLForConditionalGeneration",
    ),
311
312
313
    "Ernie4_5_VLMoeForConditionalGeneration": (
        "ernie45_vl",
        "Ernie4_5_VLMoeForConditionalGeneration",
314
    ),
315
316
317
318
    "FunAudioChatForConditionalGeneration": (
        "funaudiochat",
        "FunAudioChatForConditionalGeneration",
    ),
319
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
320
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
321
322
323
    "Gemma3nForConditionalGeneration": (
        "gemma3n_mm",
        "Gemma3nForConditionalGeneration",
324
    ),
325
    "GlmAsrForConditionalGeneration": ("glmasr", "GlmAsrForConditionalGeneration"),
326
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
327
328
329
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),
    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),
    "GlmOcrForConditionalGeneration": ("glm_ocr", "GlmOcrForConditionalGeneration"),  # noqa: E501
330
331
332
    "GraniteSpeechForConditionalGeneration": (
        "granite_speech",
        "GraniteSpeechForConditionalGeneration",
333
    ),
334
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
335
336
337
338
    "HunYuanVLForConditionalGeneration": (
        "hunyuan_vision",
        "HunYuanVLForConditionalGeneration",
    ),
ltd0924's avatar
ltd0924 committed
339
    "StepVLForConditionalGeneration": ("step_vl", "StepVLForConditionalGeneration"),
340
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
341
    "NemotronH_Nano_VL_V2": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
Zero's avatar
Zero committed
342
343
344
345
    "OpenCUAForConditionalGeneration": (
        "opencua",
        "OpenCUAForConditionalGeneration",
    ),
346
347
348
    "InternS1ForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
349
    ),
350
351
352
    "InternVLForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
353
    ),
354
355
356
357
    "Idefics3ForConditionalGeneration": (
        "idefics3",
        "Idefics3ForConditionalGeneration",
    ),
oscardev256's avatar
oscardev256 committed
358
    "IsaacForConditionalGeneration": ("isaac", "IsaacForConditionalGeneration"),
359
    "SmolVLMForConditionalGeneration": ("smolvlm", "SmolVLMForConditionalGeneration"),  # noqa: E501
360
    "KananaVForConditionalGeneration": ("kanana_v", "KananaVForConditionalGeneration"),
361
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
362
363
364
    "KeyeVL1_5ForConditionalGeneration": (
        "keye_vl1_5",
        "KeyeVL1_5ForConditionalGeneration",
365
    ),
366
    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
367
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
Roger Wang's avatar
Roger Wang committed
368
    "KimiK25ForConditionalGeneration": ("kimi_k25", "KimiK25ForConditionalGeneration"),  # noqa: E501
369
370
371
372
    "LightOnOCRForConditionalGeneration": (
        "lightonocr",
        "LightOnOCRForConditionalGeneration",
    ),
373
    "Lfm2VlForConditionalGeneration": ("lfm2_vl", "Lfm2VLForConditionalGeneration"),
374
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
375
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),  # noqa: E501
376
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
377
378
379
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
380
    ),
381
382
383
    "LlavaNextVideoForConditionalGeneration": (
        "llava_next_video",
        "LlavaNextVideoForConditionalGeneration",
384
    ),
385
386
387
    "LlavaOnevisionForConditionalGeneration": (
        "llava_onevision",
        "LlavaOnevisionForConditionalGeneration",
388
    ),
389
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
390
    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
391
392
393
    "MiniMaxVL01ForConditionalGeneration": (
        "minimax_vl_01",
        "MiniMaxVL01ForConditionalGeneration",
394
    ),
395
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
396
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
397
398
399
    "Mistral3ForConditionalGeneration": (
        "mistral3",
        "Mistral3ForConditionalGeneration",
400
    ),
401
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
402
    "Molmo2ForConditionalGeneration": ("molmo2", "Molmo2ForConditionalGeneration"),
403
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
404
    "Ovis": ("ovis", "Ovis"),
405
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
406
407
408
409
    "PaddleOCRVLForConditionalGeneration": (
        "paddleocr_vl",
        "PaddleOCRVLForConditionalGeneration",
    ),
410
411
412
413
    "PaliGemmaForConditionalGeneration": (
        "paligemma",
        "PaliGemmaForConditionalGeneration",
    ),
414
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
415
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
416
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
417
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),  # noqa: E501
418
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
419
420
421
    "Qwen2_5_VLForConditionalGeneration": (
        "qwen2_5_vl",
        "Qwen2_5_VLForConditionalGeneration",
422
    ),
423
424
425
    "Qwen2AudioForConditionalGeneration": (
        "qwen2_audio",
        "Qwen2AudioForConditionalGeneration",
426
    ),
427
428
429
    "Qwen2_5OmniModel": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
430
    ),
431
432
433
    "Qwen2_5OmniForConditionalGeneration": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
434
    ),
435
436
437
438
    "Qwen3OmniMoeForConditionalGeneration": (
        "qwen3_omni_moe_thinker",
        "Qwen3OmniMoeThinkerForConditionalGeneration",
    ),
439
    "Qwen3VLForConditionalGeneration": ("qwen3_vl", "Qwen3VLForConditionalGeneration"),  # noqa: E501
440
441
442
    "Qwen3VLMoeForConditionalGeneration": (
        "qwen3_vl_moe",
        "Qwen3VLMoeForConditionalGeneration",
443
    ),
444
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
Song's avatar
Song committed
445
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),  # noqa: E501
汪志鹏's avatar
汪志鹏 committed
446
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),  # noqa: E501
447
448
449
    "Tarsier2ForConditionalGeneration": (
        "qwen2_vl",
        "Tarsier2ForConditionalGeneration",
450
    ),
451
    "UltravoxModel": ("ultravox", "UltravoxModel"),
Patrick von Platen's avatar
Patrick von Platen committed
452
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),  # noqa: E501
Patrick von Platen's avatar
Patrick von Platen committed
453
    "VoxtralStreamingGeneration": ("voxtral_streaming", "VoxtralStreamingGeneration"),  # noqa: E501
454
    # [Encoder-decoder]
455
456
457
458
    "NemotronParseForConditionalGeneration": (
        "nemotron_parse",
        "NemotronParseForConditionalGeneration",
    ),
459
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
460
}
461
462

_SPECULATIVE_DECODING_MODELS = {
463
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
464
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
465
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
466
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
467
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
468
    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
469
    "Eagle3Qwen2_5vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
470
    "Eagle3Qwen3vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
471
472
473
474
    "EagleMistralLarge3ForCausalLM": (
        "mistral_large_3_eagle",
        "EagleMistralLarge3ForCausalLM",
    ),
475
    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
476
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
477
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
Kyungmin Lee's avatar
Kyungmin Lee committed
478
    "ExaoneMoeMTP": ("exaone_moe_mtp", "ExaoneMoeMTP"),
XuruiYang's avatar
XuruiYang committed
479
    "LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
480
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
481
    "Glm4MoeLiteMTPModel": ("glm4_moe_lite_mtp", "Glm4MoeLiteMTP"),
482
    "GlmOcrMTPModel": ("glm_ocr_mtp", "GlmOcrMTP"),
483
    "MedusaModel": ("medusa", "Medusa"),
484
    "OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
485
    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
486
487
488
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
489
}
490

491
_TRANSFORMERS_SUPPORTED_MODELS = {
492
493
494
    # Text generation models
    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
    # Multimodal models
495
496
497
498
    "Emu3ForConditionalGeneration": (
        "transformers",
        "TransformersMultiModalForCausalLM",
    ),
499
500
501
}

_TRANSFORMERS_BACKEND_MODELS = {
502
    # Text generation models
503
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
    "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
522
    "TransformersForSequenceClassification": (
523
        "transformers",
524
        "TransformersForSequenceClassification",
525
    ),
526
    "TransformersMoEForSequenceClassification": (
527
        "transformers",
528
        "TransformersMoEForSequenceClassification",
529
    ),
530
531
532
    "TransformersMultiModalForSequenceClassification": (
        "transformers",
        "TransformersMultiModalForSequenceClassification",
533
    ),
534
}
535

536
_VLLM_MODELS = {
537
    **_TEXT_GENERATION_MODELS,
538
    **_EMBEDDING_MODELS,
539
    **_CROSS_ENCODER_MODELS,
540
    **_MULTIMODAL_MODELS,
541
    **_SPECULATIVE_DECODING_MODELS,
542
543
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
544
545
}

546
547
548
549
# 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.
550
_SUBPROCESS_COMMAND = [sys.executable, "-m", "vllm.model_executor.models.registry"]
551

552
_PREVIOUSLY_SUPPORTED_MODELS = {
553
    "MotifForCausalLM": "0.10.2",
554
    "Phi3SmallForCausalLM": "0.9.2",
555
    "Phi4FlashForCausalLM": "0.10.2",
556
    "Phi4MultimodalForCausalLM": "0.12.0",
557
558
559
560
561
562
563
564
565
    # 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",
}
566

567

568
569
@dataclass(frozen=True)
class _ModelInfo:
570
    architecture: str
571
    is_text_generation_model: bool
572
    is_pooling_model: bool
573
    attn_type: AttnTypeStr
574
575
    default_seq_pooling_type: SequencePoolingType
    default_tok_pooling_type: TokenPoolingType
576
    supports_cross_encoding: bool
577
    supports_multimodal: bool
578
    supports_multimodal_raw_input_only: bool
Patrick von Platen's avatar
Patrick von Platen committed
579
    requires_raw_input_tokens: bool
580
    supports_multimodal_encoder_tp_data: bool
581
    supports_pp: bool
582
583
    has_inner_state: bool
    is_attention_free: bool
584
    is_hybrid: bool
585
    has_noops: bool
586
    supports_mamba_prefix_caching: bool
587
    supports_transcription: bool
588
    supports_transcription_only: bool
589
590

    @staticmethod
591
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
592
        return _ModelInfo(
593
            architecture=model.__name__,
594
            is_text_generation_model=is_text_generation_model(model),
595
            is_pooling_model=is_pooling_model(model),
596
597
            default_seq_pooling_type=get_default_seq_pooling_type(model),
            default_tok_pooling_type=get_default_tok_pooling_type(model),
598
            attn_type=get_attn_type(model),
599
            supports_cross_encoding=supports_cross_encoding(model),
600
            supports_multimodal=supports_multimodal(model),
601
602
603
            supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
                model
            ),
Patrick von Platen's avatar
Patrick von Platen committed
604
            requires_raw_input_tokens=requires_raw_input_tokens(model),
605
606
607
            supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
                model
            ),
608
            supports_pp=supports_pp(model),
609
610
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
611
            is_hybrid=is_hybrid(model),
612
            supports_mamba_prefix_caching=supports_mamba_prefix_caching(model),
613
            supports_transcription=supports_transcription(model),
614
615
616
            supports_transcription_only=(
                supports_transcription(model) and model.supports_transcription_only
            ),
617
            has_noops=has_noops(model),
618
        )
619
620


621
622
623
624
class _BaseRegisteredModel(ABC):
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
625

626
    @abstractmethod
627
    def load_model_cls(self) -> type[nn.Module]:
628
        raise NotImplementedError
629
630


631
632
633
634
635
636
637
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
638
    model_cls: type[nn.Module]
639
640

    @staticmethod
641
    def from_model_cls(model_cls: type[nn.Module]):
642
643
644
645
646
647
648
649
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

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

650
    def load_model_cls(self) -> type[nn.Module]:
651
652
653
654
655
656
657
658
        return self.model_cls


@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """
659

660
661
662
    module_name: str
    class_name: str

663
664
665
666
667
668
669
670
    @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"

671
    def _load_modelinfo_from_cache(self, module_hash: str) -> _ModelInfo | None:
672
673
        try:
            try:
674
                modelinfo_path = self._get_cache_dir() / self._get_cache_filename()
675
676
677
                with open(modelinfo_path, encoding="utf-8") as file:
                    mi_dict = json.load(file)
            except FileNotFoundError:
678
                logger.debug(
679
                    "Cached model info file for class %s.%s not found",
680
681
682
                    self.module_name,
                    self.class_name,
                )
683
684
685
                return None

            if mi_dict["hash"] != module_hash:
686
                logger.debug(
687
                    "Cached model info file for class %s.%s is stale",
688
689
690
                    self.module_name,
                    self.class_name,
                )
691
692
693
694
695
                return None

            # file not changed, use cached _ModelInfo properties
            return _ModelInfo(**mi_dict["modelinfo"])
        except Exception:
696
            logger.debug(
697
                "Cached model info for class %s.%s error. ",
698
699
700
                self.module_name,
                self.class_name,
            )
701
702
            return None

703
    def _save_modelinfo_to_cache(self, mi: _ModelInfo, module_hash: str) -> None:
704
705
        """save dictionary json file to cache"""
        from vllm.model_executor.model_loader.weight_utils import atomic_writer
706

707
708
709
710
711
712
713
714
        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()
715
            with atomic_writer(modelinfo_path, encoding="utf-8") as f:
716
717
718
719
720
                json.dump(modelinfo_dict, f, indent=2)
        except Exception:
            logger.exception("Error saving model info cache.")

    @logtime(logger=logger, msg="Registry inspect model class")
721
    def inspect_model_cls(self) -> _ModelInfo:
722
        model_path = Path(__file__).parent / f"{self.module_name.split('.')[-1]}.py"
723
        module_hash = None
724

725
726
        if model_path.exists():
            with open(model_path, "rb") as f:
727
                module_hash = safe_hash(f.read(), usedforsecurity=False).hexdigest()
728
729
730

            mi = self._load_modelinfo_from_cache(module_hash)
            if mi is not None:
731
                logger.debug(
732
                    "Loaded model info for class %s.%s from cache",
733
734
735
                    self.module_name,
                    self.class_name,
                )
736
737
                return mi
            else:
738
                logger.debug(
739
                    "Cache model info for class %s.%s miss. Loading model instead.",
740
741
742
                    self.module_name,
                    self.class_name,
                )
743
744
745

        # Performed in another process to avoid initializing CUDA
        mi = _run_in_subprocess(
746
747
748
749
750
            lambda: _ModelInfo.from_model_cls(self.load_model_cls())
        )
        logger.debug(
            "Loaded model info for class %s.%s", self.module_name, self.class_name
        )
751
752

        # save cache file
753
754
        if module_hash is not None:
            self._save_modelinfo_to_cache(mi, module_hash)
755
756

        return mi
757

758
    def load_model_cls(self) -> type[nn.Module]:
759
760
761
762
763
764
765
766
        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,
767
) -> type[nn.Module] | None:
768
    from vllm.platforms import current_platform
769

770
    current_platform.verify_model_arch(model_arch)
771
772
773
    try:
        return model.load_model_cls()
    except Exception:
774
        logger.exception("Error in loading model architecture '%s'", model_arch)
775
        return None
776
777


778
779
780
781
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
782
) -> _ModelInfo | None:
783
784
785
    try:
        return model.inspect_model_cls()
    except Exception:
786
        logger.exception("Error in inspecting model architecture '%s'", model_arch)
787
        return None
788
789


790
791
792
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
793
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
794

795
    def get_supported_archs(self) -> Set[str]:
796
        return self.models.keys()
797

798
799
800
    def register_model(
        self,
        model_arch: str,
801
        model_cls: type[nn.Module] | str,
802
    ) -> None:
803
804
805
        """
        Register an external model to be used in vLLM.

806
        `model_cls` can be either:
807

808
        - A [`torch.nn.Module`][] class directly referencing the model.
809
        - A string in the format `<module>:<class>` which can be used to
810
811
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
812
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
813
        """
814
815
816
817
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

818
        if model_arch in self.models:
819
820
            logger.warning(
                "Model architecture %s is already registered, and will be "
821
822
823
824
                "overwritten by the new model class %s.",
                model_arch,
                model_cls,
            )
825
826
827
828
829
830

        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)
831

832
            model = _LazyRegisteredModel(*split_str)
833
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
834
            model = _RegisteredModel.from_model_cls(model_cls)
835
        else:
836
837
838
839
            msg = (
                "`model_cls` should be a string or PyTorch model class, "
                f"not a {type(model_arch)}"
            )
840
            raise TypeError(msg)
841

842
        self.models[model_arch] = model
843

844
    def _raise_for_unsupported(self, architectures: list[str]):
845
        all_supported_archs = self.get_supported_archs()
846

847
848
849
        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
850
851
                "to be inspected. Please check the logs for more details."
            )
852

853
854
855
856
857
858
859
860
        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 "
861
862
                    "use this model architecture."
                )
863

864
865
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
866
867
            f"Supported architectures: {all_supported_archs}"
        )
868

869
    def _try_load_model_cls(self, model_arch: str) -> type[nn.Module] | None:
870
871
        if model_arch not in self.models:
            return None
872

873
        return _try_load_model_cls(model_arch, self.models[model_arch])
874

875
    def _try_inspect_model_cls(self, model_arch: str) -> _ModelInfo | None:
876
877
        if model_arch not in self.models:
            return None
878

879
880
881
882
883
884
        return _try_inspect_model_cls(model_arch, self.models[model_arch])

    def _try_resolve_transformers(
        self,
        architecture: str,
        model_config: ModelConfig,
885
    ) -> str | None:
886
887
888
        if architecture in _TRANSFORMERS_BACKEND_MODELS:
            return architecture

889
890
891
        auto_map: dict[str, str] = (
            getattr(model_config.hf_config, "auto_map", None) or dict()
        )
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907

        # 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,
908
                        trust_remote_code=model_config.trust_remote_code,
909
910
911
912
913
914
915
916
917
918
919
920
                        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,
921
                        trust_remote_code=model_config.trust_remote_code,
922
923
924
925
926
                        warn_on_fail=True,
                    )
                    if model_module is not None:
                        break
            else:
927
                if model_config.model_impl != "transformers":
928
929
930
931
932
933
934
                    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 "
935
936
                    "'auto_map' (relevant if the model is custom)."
                )
937
938

        if not model_module.is_backend_compatible():
939
            if model_config.model_impl != "transformers":
940
                return None
941

942
943
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
944
945
                "is not compatible with vLLM."
            )
946

947
        return model_config._get_transformers_backend_cls()
948

949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
    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
974

975
976
    def inspect_model_cls(
        self,
977
        architectures: str | list[str],
978
        model_config: ModelConfig,
979
    ) -> tuple[_ModelInfo, str]:
980
981
        if isinstance(architectures, str):
            architectures = [architectures]
982
983
        if not architectures:
            raise ValueError("No model architectures are specified")
984
985

        # Require transformers impl
986
        if model_config.model_impl == "transformers":
987
            arch = self._try_resolve_transformers(architectures[0], model_config)
988
989
990
991
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)
992
        elif model_config.model_impl == "terratorch":
993
994
            model_info = self._try_inspect_model_cls("Terratorch")
            return (model_info, "Terratorch")
995

996
        # Fallback to transformers impl (after resolving convert_type)
997
998
999
1000
1001
1002
        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)
1003
1004
1005
1006
1007
1008
1009
            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)
1010
            model_info = self._try_inspect_model_cls(normalized_arch)
1011
            if model_info is not None:
1012
                return (model_info, arch)
1013

1014
        # Fallback to transformers impl (before resolving runner_type)
1015
1016
1017
1018
1019
        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)
1020
1021
1022
1023
1024
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

1025
        return self._raise_for_unsupported(architectures)
1026

1027
1028
    def resolve_model_cls(
        self,
1029
        architectures: str | list[str],
1030
        model_config: ModelConfig,
1031
    ) -> tuple[type[nn.Module], str]:
1032
1033
        if isinstance(architectures, str):
            architectures = [architectures]
1034
1035
        if not architectures:
            raise ValueError("No model architectures are specified")
1036
1037

        # Require transformers impl
1038
        if model_config.model_impl == "transformers":
1039
            arch = self._try_resolve_transformers(architectures[0], model_config)
1040
1041
1042
1043
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)
1044
        elif model_config.model_impl == "terratorch":
1045
1046
1047
1048
            arch = "Terratorch"
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
1049

1050
        # Fallback to transformers impl (after resolving convert_type)
1051
1052
1053
1054
1055
1056
        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)
1057
1058
1059
1060
1061
1062
1063
            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)
1064
            model_cls = self._try_load_model_cls(normalized_arch)
1065
1066
            if model_cls is not None:
                return (model_cls, arch)
1067

1068
        # Fallback to transformers impl (before resolving runner_type)
1069
1070
1071
1072
1073
        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)
1074
1075
1076
1077
1078
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

1079
        return self._raise_for_unsupported(architectures)
1080

1081
1082
    def is_text_generation_model(
        self,
1083
        architectures: str | list[str],
1084
        model_config: ModelConfig,
1085
    ) -> bool:
1086
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1087
        return model_cls.is_text_generation_model
1088

1089
    def is_pooling_model(
1090
        self,
1091
        architectures: str | list[str],
1092
        model_config: ModelConfig,
1093
    ) -> bool:
1094
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1095
        return model_cls.is_pooling_model
1096

1097
1098
    def is_cross_encoder_model(
        self,
1099
        architectures: str | list[str],
1100
        model_config: ModelConfig,
1101
    ) -> bool:
1102
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1103
        return model_cls.supports_cross_encoding
1104

1105
1106
    def is_multimodal_model(
        self,
1107
        architectures: str | list[str],
1108
        model_config: ModelConfig,
1109
    ) -> bool:
1110
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1111
        return model_cls.supports_multimodal
1112

1113
    def is_multimodal_raw_input_only_model(
1114
        self,
1115
        architectures: str | list[str],
1116
        model_config: ModelConfig,
1117
    ) -> bool:
1118
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1119
        return model_cls.supports_multimodal_raw_input_only
1120

1121
1122
    def is_pp_supported_model(
        self,
1123
        architectures: str | list[str],
1124
        model_config: ModelConfig,
1125
    ) -> bool:
1126
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1127
        return model_cls.supports_pp
1128

1129
1130
    def model_has_inner_state(
        self,
1131
        architectures: str | list[str],
1132
        model_config: ModelConfig,
1133
    ) -> bool:
1134
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1135
        return model_cls.has_inner_state
1136

1137
1138
    def is_attention_free_model(
        self,
1139
        architectures: str | list[str],
1140
        model_config: ModelConfig,
1141
    ) -> bool:
1142
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1143
        return model_cls.is_attention_free
1144

1145
1146
    def is_hybrid_model(
        self,
1147
        architectures: str | list[str],
1148
        model_config: ModelConfig,
1149
    ) -> bool:
1150
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1151
1152
        return model_cls.is_hybrid

1153
1154
    def is_noops_model(
        self,
1155
        architectures: str | list[str],
1156
        model_config: ModelConfig,
1157
    ) -> bool:
1158
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1159
1160
        return model_cls.has_noops

1161
1162
    def is_transcription_model(
        self,
1163
        architectures: str | list[str],
1164
        model_config: ModelConfig,
1165
    ) -> bool:
1166
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1167
1168
        return model_cls.supports_transcription

1169
1170
    def is_transcription_only_model(
        self,
1171
        architectures: str | list[str],
1172
        model_config: ModelConfig,
1173
    ) -> bool:
1174
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1175
1176
        return model_cls.supports_transcription_only

1177

1178
1179
1180
1181
1182
1183
1184
1185
1186
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()
    }
)
1187
1188
1189
1190
1191

_T = TypeVar("_T")


def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
1192
1193
1194
1195
1196
    # 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")

1197
        # `cloudpickle` allows pickling lambda functions directly
1198
        import cloudpickle
1199

1200
        input_bytes = cloudpickle.dumps((fn, output_filepath))
1201
1202
1203

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
1204
1205
1206
        returned = subprocess.run(
            _SUBPROCESS_COMMAND, input=input_bytes, capture_output=True
        )
1207
1208
1209
1210
1211
1212

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
1213
1214
1215
            raise RuntimeError(
                f"Error raised in subprocess:\n{returned.stderr.decode()}"
            ) from e
1216

1217
        with open(output_filepath, "rb") as f:
1218
1219
1220
1221
1222
1223
            return pickle.load(f)


def _run() -> None:
    # Setup plugins
    from vllm.plugins import load_general_plugins
1224

1225
1226
1227
1228
1229
    load_general_plugins()

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

    result = fn()
1230
1231
1232

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
1233
1234
1235


if __name__ == "__main__":
1236
    _run()