registry.py 49.8 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_late_interaction,
53
    supports_mamba_prefix_caching,
54
55
56
57
58
59
60
    supports_multimodal,
    supports_multimodal_encoder_tp_data,
    supports_multimodal_raw_input_only,
    supports_pp,
    supports_transcription,
)
from .interfaces_base import (
61
    get_attn_type,
62
63
    get_default_seq_pooling_type,
    get_default_tok_pooling_type,
64
65
66
    is_pooling_model,
    is_text_generation_model,
)
67
68
69

logger = init_logger(__name__)

70
71
_TEXT_GENERATION_MODELS = {
    # [Decoder-only]
72
    "AfmoeForCausalLM": ("afmoe", "AfmoeForCausalLM"),
73
    "ApertusForCausalLM": ("apertus", "ApertusForCausalLM"),
74
75
    "AquilaModel": ("llama", "LlamaForCausalLM"),
    "AquilaForCausalLM": ("llama", "LlamaForCausalLM"),  # AquilaChat2
Raghav Ravishankar's avatar
Raghav Ravishankar committed
76
    "ArceeForCausalLM": ("arcee", "ArceeForCausalLM"),
77
    "ArcticForCausalLM": ("arctic", "ArcticForCausalLM"),
78
79
80
81
    # baichuan-7b, upper case 'C' in the class name
    "BaiChuanForCausalLM": ("baichuan", "BaiChuanForCausalLM"),
    # baichuan-13b, lower case 'c' in the class name
    "BaichuanForCausalLM": ("baichuan", "BaichuanForCausalLM"),
82
    "BailingMoeForCausalLM": ("bailing_moe", "BailingMoeForCausalLM"),
ant-yy's avatar
ant-yy committed
83
    "BailingMoeV2ForCausalLM": ("bailing_moe", "BailingMoeV2ForCausalLM"),
Yu Chin Fabian Lim's avatar
Yu Chin Fabian Lim committed
84
    "BambaForCausalLM": ("bamba", "BambaForCausalLM"),
85
    "BloomForCausalLM": ("bloom", "BloomForCausalLM"),
86
    "ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
87
    "ChatGLMForConditionalGeneration": ("chatglm", "ChatGLMForCausalLM"),
88
    "CohereForCausalLM": ("commandr", "CohereForCausalLM"),
89
    "Cohere2ForCausalLM": ("commandr", "CohereForCausalLM"),
90
    "CwmForCausalLM": ("llama", "LlamaForCausalLM"),
91
    "DbrxForCausalLM": ("dbrx", "DbrxForCausalLM"),
92
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
93
    "DeepseekForCausalLM": ("deepseek_v2", "DeepseekForCausalLM"),
94
    "DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
95
    "DeepseekV3ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
96
    "DeepseekV32ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
97
    "Dots1ForCausalLM": ("dots1", "Dots1ForCausalLM"),
98
    "Ernie4_5ForCausalLM": ("ernie45", "Ernie4_5ForCausalLM"),
99
    "Ernie4_5_MoeForCausalLM": ("ernie45_moe", "Ernie4_5_MoeForCausalLM"),
100
    "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
101
    "Exaone4ForCausalLM": ("exaone4", "Exaone4ForCausalLM"),
Kyungmin Lee's avatar
Kyungmin Lee committed
102
    "ExaoneMoEForCausalLM": ("exaone_moe", "ExaoneMoeForCausalLM"),
103
    "Fairseq2LlamaForCausalLM": ("fairseq2_llama", "Fairseq2LlamaForCausalLM"),
104
105
106
    "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
    "FalconH1ForCausalLM": ("falcon_h1", "FalconH1ForCausalLM"),
107
    "FlexOlmoForCausalLM": ("flex_olmo", "FlexOlmoForCausalLM"),
108
109
    "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
    "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
110
    "Gemma3ForCausalLM": ("gemma3", "Gemma3ForCausalLM"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
111
    "Gemma3nForCausalLM": ("gemma3n", "Gemma3nForCausalLM"),
112
    "Qwen3NextForCausalLM": ("qwen3_next", "Qwen3NextForCausalLM"),
113
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
114
    "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
115
    "Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
116
    "Glm4MoeLiteForCausalLM": ("glm4_moe_lite", "Glm4MoeLiteForCausalLM"),
Jee Jee Li's avatar
Jee Jee Li committed
117
    "GlmMoeDsaForCausalLM": ("deepseek_v2", "GlmMoeDsaForCausalLM"),
118
    "GptOssForCausalLM": ("gpt_oss", "GptOssForCausalLM"),
119
120
121
122
123
124
    "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
    "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
    "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
    "GPTNeoXForCausalLM": ("gpt_neox", "GPTNeoXForCausalLM"),
    "GraniteForCausalLM": ("granite", "GraniteForCausalLM"),
    "GraniteMoeForCausalLM": ("granitemoe", "GraniteMoeForCausalLM"),
125
126
    "GraniteMoeHybridForCausalLM": ("granitemoehybrid", "GraniteMoeHybridForCausalLM"),  # noqa: E501
    "GraniteMoeSharedForCausalLM": ("granitemoeshared", "GraniteMoeSharedForCausalLM"),  # noqa: E501
127
    "GritLM": ("gritlm", "GritLM"),
Bijaya Dangol's avatar
Bijaya Dangol committed
128
129
    "Grok1ModelForCausalLM": ("grok1", "GrokForCausalLM"),
    "Grok1ForCausalLM": ("grok1", "GrokForCausalLM"),
130
131
    "HunYuanMoEV1ForCausalLM": ("hunyuan_v1", "HunYuanMoEV1ForCausalLM"),
    "HunYuanDenseV1ForCausalLM": ("hunyuan_v1", "HunYuanDenseV1ForCausalLM"),
132
    "HCXVisionForCausalLM": ("hyperclovax_vision", "HCXVisionForCausalLM"),
133
134
    "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
    "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"),
135
    "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
136
    "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
137
138
    "IQuestCoderForCausalLM": ("llama", "LlamaForCausalLM"),
    "IQuestLoopCoderForCausalLM": ("iquest_loopcoder", "IQuestLoopCoderForCausalLM"),
139
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
140
    "Jais2ForCausalLM": ("jais2", "Jais2ForCausalLM"),
141
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
142
    "KimiLinearForCausalLM": ("kimi_linear", "KimiLinearForCausalLM"),  # noqa: E501
143
    "Lfm2ForCausalLM": ("lfm2", "Lfm2ForCausalLM"),
Paul Pak's avatar
Paul Pak committed
144
    "Lfm2MoeForCausalLM": ("lfm2_moe", "Lfm2MoeForCausalLM"),
145
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
146
    "Llama4ForCausalLM": ("llama4", "Llama4ForCausalLM"),
147
148
    # For decapoda-research/llama-*
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
XuruiYang's avatar
XuruiYang committed
149
    "LongcatFlashForCausalLM": ("longcat_flash", "LongcatFlashForCausalLM"),
150
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
151
    "Mamba2ForCausalLM": ("mamba2", "Mamba2ForCausalLM"),
152
153
    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
    "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
154
155
156
    "MiniMaxForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxText01ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxM1ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
157
    "MiniMaxM2ForCausalLM": ("minimax_m2", "MiniMaxM2ForCausalLM"),
158
    "MistralForCausalLM": ("mistral", "MistralForCausalLM"),
159
    "MistralLarge3ForCausalLM": ("mistral_large_3", "MistralLarge3ForCausalLM"),
160
161
162
163
    "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
    # transformers's mpt class has lower case
    "MptForCausalLM": ("mpt", "MPTForCausalLM"),
    "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
164
    "MiMoForCausalLM": ("mimo", "MiMoForCausalLM"),
165
    "MiMoV2FlashForCausalLM": ("mimo_v2_flash", "MiMoV2FlashForCausalLM"),
166
    "NemotronForCausalLM": ("nemotron", "NemotronForCausalLM"),
Luis Vega's avatar
Luis Vega committed
167
    "NemotronHForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
168
    "NemotronHPuzzleForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
169
    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
170
    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
171
    "Olmo3ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
172
173
174
    "OlmoeForCausalLM": ("olmoe", "OlmoeForCausalLM"),
    "OPTForCausalLM": ("opt", "OPTForCausalLM"),
    "OrionForCausalLM": ("orion", "OrionForCausalLM"),
175
    "OuroForCausalLM": ("ouro", "OuroForCausalLM"),
176
    "PanguEmbeddedForCausalLM": ("openpangu", "PanguEmbeddedForCausalLM"),
177
    "PanguProMoEV2ForCausalLM": ("openpangu", "PanguProMoEV2ForCausalLM"),
178
    "PanguUltraMoEForCausalLM": ("openpangu", "PanguUltraMoEForCausalLM"),
179
180
181
182
    "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
    "PhiForCausalLM": ("phi", "PhiForCausalLM"),
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
    "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
183
    "Plamo2ForCausalLM": ("plamo2", "Plamo2ForCausalLM"),
184
    "Plamo3ForCausalLM": ("plamo3", "Plamo3ForCausalLM"),
185
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
186
187
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
188
189
    "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),
    "Qwen3MoeForCausalLM": ("qwen3_moe", "Qwen3MoeForCausalLM"),
190
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
191
    "SeedOssForCausalLM": ("seed_oss", "SeedOssForCausalLM"),
Li Xie's avatar
Li Xie committed
192
    "Step1ForCausalLM": ("step1", "Step1ForCausalLM"),
Song's avatar
Song committed
193
    "Step3TextForCausalLM": ("step3_text", "Step3TextForCausalLM"),
csy0225's avatar
csy0225 committed
194
    "Step3p5ForCausalLM": ("step3p5", "Step3p5ForCausalLM"),
195
196
197
198
    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
199
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
200
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
201
    "TeleFLMForCausalLM": ("teleflm", "TeleFLMForCausalLM"),
202
    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
203
    "Zamba2ForCausalLM": ("zamba2", "Zamba2ForCausalLM"),
204
205
206
}

_EMBEDDING_MODELS = {
207
    # [Text-only]
208
    "BertModel": ("bert", "BertEmbeddingModel"),
209
    "BertSpladeSparseEmbeddingModel": ("bert", "BertSpladeSparseEmbeddingModel"),
210
    "HF_ColBERT": ("colbert", "ColBERTModel"),
211
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
212
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
213
    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
214
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
215
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
216
    "GritLM": ("gritlm", "GritLM"),
217
218
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
219
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
220
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
221
    "LlamaBidirectionalModel": ("llama", "LlamaBidirectionalModel"),
222
    "LlamaModel": ("llama", "LlamaForCausalLM"),
223
224
    **{
        # Multiple models share the same architecture, so we include them all
225
226
        k: (mod, arch)
        for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
227
228
        if arch == "LlamaForCausalLM"
    },
229
    "MistralModel": ("llama", "LlamaForCausalLM"),
230
    "ModernBertModel": ("modernbert", "ModernBertModel"),
231
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
232
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
233
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
234
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
235
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
236
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
237
238
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
239
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
240
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
chengchengpei's avatar
chengchengpei committed
241
242
243
244
    "VoyageQwen3BidirectionalEmbedModel": (
        "voyage",
        "VoyageQwen3BidirectionalEmbedModel",
    ),
245
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
246
    "BgeM3EmbeddingModel": ("roberta", "BgeM3EmbeddingModel"),
247
    # [Multimodal]
248
    "CLIPModel": ("clip", "CLIPEmbeddingModel"),
249
250
251
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
252
    ),
253
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
254
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
255
    "SiglipModel": ("siglip", "SiglipEmbeddingModel"),
256
257
    # Technically Terratorch models work on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
258
    # models for the time being.
259
260
    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
    "Terratorch": ("terratorch", "Terratorch"),
261
262
}

263
264
_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
265
    "BertForTokenClassification": ("bert", "BertForTokenClassification"),
266
267
268
269
    "GteNewForSequenceClassification": (
        "bert_with_rope",
        "GteNewForSequenceClassification",
    ),
270
271
272
273
274
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"),
    "LlamaBidirectionalForSequenceClassification": (
        "llama",
        "LlamaBidirectionalForSequenceClassification",
    ),
275
276
277
278
    "ModernBertForSequenceClassification": (
        "modernbert",
        "ModernBertForSequenceClassification",
    ),
279
280
281
282
    "ModernBertForTokenClassification": (
        "modernbert",
        "ModernBertForTokenClassification",
    ),
283
284
285
286
287
    "RobertaForSequenceClassification": ("roberta", "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": (
        "roberta",
        "RobertaForSequenceClassification",
    ),
288
289
}

290
_MULTIMODAL_MODELS = {
291
    # [Decoder-only]
292
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
293
294
295
296
    "AudioFlamingo3ForConditionalGeneration": (
        "audioflamingo3",
        "AudioFlamingo3ForConditionalGeneration",
    ),
297
298
299
300
    "MusicFlamingoForConditionalGeneration": (
        "musicflamingo",
        "MusicFlamingoForConditionalGeneration",
    ),
301
302
303
    "AyaVisionForConditionalGeneration": (
        "aya_vision",
        "AyaVisionForConditionalGeneration",
304
    ),
305
    "BagelForConditionalGeneration": ("bagel", "BagelForConditionalGeneration"),
306
    "BeeForConditionalGeneration": ("bee", "BeeForConditionalGeneration"),
307
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
308
309
310
    "ChameleonForConditionalGeneration": (
        "chameleon",
        "ChameleonForConditionalGeneration",
311
    ),
312
313
314
    "Cohere2VisionForConditionalGeneration": (
        "cohere2_vision",
        "Cohere2VisionForConditionalGeneration",
315
    ),
316
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
317
    "DeepseekOCRForCausalLM": ("deepseek_ocr", "DeepseekOCRForCausalLM"),
RED's avatar
RED committed
318
    "DeepseekOCR2ForCausalLM": ("deepseek_ocr2", "DeepseekOCR2ForCausalLM"),
Roger Wang's avatar
Roger Wang committed
319
    "DotsOCRForCausalLM": ("dots_ocr", "DotsOCRForCausalLM"),
320
321
322
323
    "Eagle2_5_VLForConditionalGeneration": (
        "eagle2_5_vl",
        "Eagle2_5_VLForConditionalGeneration",
    ),
324
325
326
    "Ernie4_5_VLMoeForConditionalGeneration": (
        "ernie45_vl",
        "Ernie4_5_VLMoeForConditionalGeneration",
327
    ),
328
    "FunASRForConditionalGeneration": ("funasr", "FunASRForConditionalGeneration"),  # noqa: E501
329
330
331
332
    "FunAudioChatForConditionalGeneration": (
        "funaudiochat",
        "FunAudioChatForConditionalGeneration",
    ),
333
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
334
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
335
336
337
    "Gemma3nForConditionalGeneration": (
        "gemma3n_mm",
        "Gemma3nForConditionalGeneration",
338
    ),
339
    "GlmAsrForConditionalGeneration": ("glmasr", "GlmAsrForConditionalGeneration"),
340
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
341
342
343
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),
    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),
    "GlmOcrForConditionalGeneration": ("glm_ocr", "GlmOcrForConditionalGeneration"),  # noqa: E501
344
345
346
    "GraniteSpeechForConditionalGeneration": (
        "granite_speech",
        "GraniteSpeechForConditionalGeneration",
347
    ),
348
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
349
350
351
352
    "HunYuanVLForConditionalGeneration": (
        "hunyuan_vision",
        "HunYuanVLForConditionalGeneration",
    ),
ltd0924's avatar
ltd0924 committed
353
    "StepVLForConditionalGeneration": ("step_vl", "StepVLForConditionalGeneration"),
354
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
355
    "NemotronH_Nano_VL_V2": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
Zero's avatar
Zero committed
356
357
358
359
    "OpenCUAForConditionalGeneration": (
        "opencua",
        "OpenCUAForConditionalGeneration",
    ),
360
361
362
    "InternS1ForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
363
    ),
364
365
366
    "InternVLForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
367
    ),
zxy's avatar
zxy committed
368
369
370
371
    "InternS1ProForConditionalGeneration": (
        "interns1_pro",
        "InternS1ProForConditionalGeneration",
    ),
372
373
374
375
    "Idefics3ForConditionalGeneration": (
        "idefics3",
        "Idefics3ForConditionalGeneration",
    ),
oscardev256's avatar
oscardev256 committed
376
    "IsaacForConditionalGeneration": ("isaac", "IsaacForConditionalGeneration"),
377
    "SmolVLMForConditionalGeneration": ("smolvlm", "SmolVLMForConditionalGeneration"),  # noqa: E501
378
    "KananaVForConditionalGeneration": ("kanana_v", "KananaVForConditionalGeneration"),
379
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
380
381
382
    "KeyeVL1_5ForConditionalGeneration": (
        "keye_vl1_5",
        "KeyeVL1_5ForConditionalGeneration",
383
    ),
384
    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
385
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
Roger Wang's avatar
Roger Wang committed
386
    "KimiK25ForConditionalGeneration": ("kimi_k25", "KimiK25ForConditionalGeneration"),  # noqa: E501
387
388
389
390
    "LightOnOCRForConditionalGeneration": (
        "lightonocr",
        "LightOnOCRForConditionalGeneration",
    ),
391
    "Lfm2VlForConditionalGeneration": ("lfm2_vl", "Lfm2VLForConditionalGeneration"),
392
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
393
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),  # noqa: E501
394
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
395
396
397
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
398
    ),
399
400
401
    "LlavaNextVideoForConditionalGeneration": (
        "llava_next_video",
        "LlavaNextVideoForConditionalGeneration",
402
    ),
403
404
405
    "LlavaOnevisionForConditionalGeneration": (
        "llava_onevision",
        "LlavaOnevisionForConditionalGeneration",
406
    ),
407
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
408
    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
409
410
411
    "MiniMaxVL01ForConditionalGeneration": (
        "minimax_vl_01",
        "MiniMaxVL01ForConditionalGeneration",
412
    ),
413
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
414
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
415
416
417
    "Mistral3ForConditionalGeneration": (
        "mistral3",
        "Mistral3ForConditionalGeneration",
418
    ),
419
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
420
    "Molmo2ForConditionalGeneration": ("molmo2", "Molmo2ForConditionalGeneration"),
421
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
422
423
424
425
    "OpenPanguVLForConditionalGeneration": (
        "openpangu_vl",
        "OpenPanguVLForConditionalGeneration",
    ),
426
    "Ovis": ("ovis", "Ovis"),
427
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
428
429
    "Ovis2_6ForCausalLM": ("ovis2_5", "Ovis2_5"),
    "Ovis2_6_MoeForCausalLM": ("ovis2_5", "Ovis2_5"),
430
431
432
433
    "PaddleOCRVLForConditionalGeneration": (
        "paddleocr_vl",
        "PaddleOCRVLForConditionalGeneration",
    ),
434
435
436
437
    "PaliGemmaForConditionalGeneration": (
        "paligemma",
        "PaliGemmaForConditionalGeneration",
    ),
438
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
439
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
440
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
441
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),  # noqa: E501
442
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
443
444
445
    "Qwen2_5_VLForConditionalGeneration": (
        "qwen2_5_vl",
        "Qwen2_5_VLForConditionalGeneration",
446
    ),
447
448
449
    "Qwen2AudioForConditionalGeneration": (
        "qwen2_audio",
        "Qwen2AudioForConditionalGeneration",
450
    ),
451
452
453
    "Qwen2_5OmniModel": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
454
    ),
455
456
457
    "Qwen2_5OmniForConditionalGeneration": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
458
    ),
459
460
461
462
    "Qwen3OmniMoeForConditionalGeneration": (
        "qwen3_omni_moe_thinker",
        "Qwen3OmniMoeThinkerForConditionalGeneration",
    ),
Roger Wang's avatar
Roger Wang committed
463
464
465
466
    "Qwen3ASRForConditionalGeneration": (
        "qwen3_asr",
        "Qwen3ASRForConditionalGeneration",
    ),
467
    "Qwen3VLForConditionalGeneration": ("qwen3_vl", "Qwen3VLForConditionalGeneration"),  # noqa: E501
468
469
470
    "Qwen3VLMoeForConditionalGeneration": (
        "qwen3_vl_moe",
        "Qwen3VLMoeForConditionalGeneration",
471
    ),
472
473
474
475
476
477
478
479
    "Qwen3_5ForConditionalGeneration": (
        "qwen3_5",
        "Qwen3_5ForConditionalGeneration",
    ),
    "Qwen3_5MoeForConditionalGeneration": (
        "qwen3_5",
        "Qwen3_5MoeForConditionalGeneration",
    ),
480
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
Song's avatar
Song committed
481
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),  # noqa: E501
汪志鹏's avatar
汪志鹏 committed
482
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),  # noqa: E501
483
484
485
    "Tarsier2ForConditionalGeneration": (
        "qwen2_vl",
        "Tarsier2ForConditionalGeneration",
486
    ),
487
    "UltravoxModel": ("ultravox", "UltravoxModel"),
Patrick von Platen's avatar
Patrick von Platen committed
488
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),  # noqa: E501
489
    "VoxtralRealtimeGeneration": ("voxtral_realtime", "VoxtralRealtimeGeneration"),  # noqa: E501
490
    # [Encoder-decoder]
491
492
493
494
    "NemotronParseForConditionalGeneration": (
        "nemotron_parse",
        "NemotronParseForConditionalGeneration",
    ),
495
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
496
}
497
498

_SPECULATIVE_DECODING_MODELS = {
499
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
500
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
501
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
502
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
503
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
504
    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
505
    "Eagle3Qwen2_5vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
506
    "Eagle3Qwen3vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
507
508
509
510
    "EagleMistralLarge3ForCausalLM": (
        "mistral_large_3_eagle",
        "EagleMistralLarge3ForCausalLM",
    ),
511
    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
512
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
513
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
Kyungmin Lee's avatar
Kyungmin Lee committed
514
    "ExaoneMoeMTP": ("exaone_moe_mtp", "ExaoneMoeMTP"),
XuruiYang's avatar
XuruiYang committed
515
    "LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
516
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
517
    "Glm4MoeLiteMTPModel": ("glm4_moe_lite_mtp", "Glm4MoeLiteMTP"),
518
    "GlmOcrMTPModel": ("glm_ocr_mtp", "GlmOcrMTP"),
519
    "MedusaModel": ("medusa", "Medusa"),
520
    "OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
521
    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
csy0225's avatar
csy0225 committed
522
    "Step3p5MTP": ("step3p5_mtp", "Step3p5MTP"),
523
524
    "Qwen3_5MTP": ("qwen3_5_mtp", "Qwen3_5MTP"),
    "Qwen3_5MoeMTP": ("qwen3_5_mtp", "Qwen3_5MoeMTP"),
525
526
527
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
528
}
529

530
_TRANSFORMERS_SUPPORTED_MODELS = {
531
532
533
    # Text generation models
    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
    # Multimodal models
534
535
536
537
    "Emu3ForConditionalGeneration": (
        "transformers",
        "TransformersMultiModalForCausalLM",
    ),
538
539
540
}

_TRANSFORMERS_BACKEND_MODELS = {
541
    # Text generation models
542
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
    "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
561
    "TransformersForSequenceClassification": (
562
        "transformers",
563
        "TransformersForSequenceClassification",
564
    ),
565
    "TransformersMoEForSequenceClassification": (
566
        "transformers",
567
        "TransformersMoEForSequenceClassification",
568
    ),
569
570
571
    "TransformersMultiModalForSequenceClassification": (
        "transformers",
        "TransformersMultiModalForSequenceClassification",
572
    ),
573
}
574

575
_VLLM_MODELS = {
576
    **_TEXT_GENERATION_MODELS,
577
    **_EMBEDDING_MODELS,
578
    **_CROSS_ENCODER_MODELS,
579
    **_MULTIMODAL_MODELS,
580
    **_SPECULATIVE_DECODING_MODELS,
581
582
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
583
584
}

585
586
587
588
# 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.
589
_SUBPROCESS_COMMAND = [sys.executable, "-m", "vllm.model_executor.models.registry"]
590

591
_PREVIOUSLY_SUPPORTED_MODELS = {
592
    "MotifForCausalLM": "0.10.2",
593
    "Phi3SmallForCausalLM": "0.9.2",
594
    "Phi4FlashForCausalLM": "0.10.2",
595
    "Phi4MultimodalForCausalLM": "0.12.0",
596
597
598
599
600
601
602
603
604
    # 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",
}
605

606

607
608
@dataclass(frozen=True)
class _ModelInfo:
609
    architecture: str
610
    is_text_generation_model: bool
611
    is_pooling_model: bool
612
    attn_type: AttnTypeStr
613
614
    default_seq_pooling_type: SequencePoolingType
    default_tok_pooling_type: TokenPoolingType
615
    supports_cross_encoding: bool
616
    supports_late_interaction: bool
617
    supports_multimodal: bool
618
    supports_multimodal_raw_input_only: bool
Patrick von Platen's avatar
Patrick von Platen committed
619
    requires_raw_input_tokens: bool
620
    supports_multimodal_encoder_tp_data: bool
621
    supports_pp: bool
622
623
    has_inner_state: bool
    is_attention_free: bool
624
    is_hybrid: bool
625
    has_noops: bool
626
    supports_mamba_prefix_caching: bool
627
    supports_transcription: bool
628
    supports_transcription_only: bool
629
630

    @staticmethod
631
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
632
        return _ModelInfo(
633
            architecture=model.__name__,
634
            is_text_generation_model=is_text_generation_model(model),
635
            is_pooling_model=is_pooling_model(model),
636
637
            default_seq_pooling_type=get_default_seq_pooling_type(model),
            default_tok_pooling_type=get_default_tok_pooling_type(model),
638
            attn_type=get_attn_type(model),
639
            supports_cross_encoding=supports_cross_encoding(model),
640
            supports_late_interaction=supports_late_interaction(model),
641
            supports_multimodal=supports_multimodal(model),
642
643
644
            supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
                model
            ),
Patrick von Platen's avatar
Patrick von Platen committed
645
            requires_raw_input_tokens=requires_raw_input_tokens(model),
646
647
648
            supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
                model
            ),
649
            supports_pp=supports_pp(model),
650
651
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
652
            is_hybrid=is_hybrid(model),
653
            supports_mamba_prefix_caching=supports_mamba_prefix_caching(model),
654
            supports_transcription=supports_transcription(model),
655
656
657
            supports_transcription_only=(
                supports_transcription(model) and model.supports_transcription_only
            ),
658
            has_noops=has_noops(model),
659
        )
660
661


662
663
664
665
class _BaseRegisteredModel(ABC):
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
666

667
    @abstractmethod
668
    def load_model_cls(self) -> type[nn.Module]:
669
        raise NotImplementedError
670
671


672
673
674
675
676
677
678
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
679
    model_cls: type[nn.Module]
680
681

    @staticmethod
682
    def from_model_cls(model_cls: type[nn.Module]):
683
684
685
686
687
688
689
690
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

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

691
    def load_model_cls(self) -> type[nn.Module]:
692
693
694
695
696
697
698
699
        return self.model_cls


@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """
700

701
702
703
    module_name: str
    class_name: str

704
705
706
707
708
709
710
711
    @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"

712
    def _load_modelinfo_from_cache(self, module_hash: str) -> _ModelInfo | None:
713
714
        try:
            try:
715
                modelinfo_path = self._get_cache_dir() / self._get_cache_filename()
716
717
718
                with open(modelinfo_path, encoding="utf-8") as file:
                    mi_dict = json.load(file)
            except FileNotFoundError:
719
                logger.debug(
720
                    "Cached model info file for class %s.%s not found",
721
722
723
                    self.module_name,
                    self.class_name,
                )
724
725
726
                return None

            if mi_dict["hash"] != module_hash:
727
                logger.debug(
728
                    "Cached model info file for class %s.%s is stale",
729
730
731
                    self.module_name,
                    self.class_name,
                )
732
733
734
735
736
                return None

            # file not changed, use cached _ModelInfo properties
            return _ModelInfo(**mi_dict["modelinfo"])
        except Exception:
737
            logger.debug(
738
                "Cached model info for class %s.%s error. ",
739
740
741
                self.module_name,
                self.class_name,
            )
742
743
            return None

744
    def _save_modelinfo_to_cache(self, mi: _ModelInfo, module_hash: str) -> None:
745
746
        """save dictionary json file to cache"""
        from vllm.model_executor.model_loader.weight_utils import atomic_writer
747

748
749
750
751
752
753
754
755
        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()
756
            with atomic_writer(modelinfo_path, encoding="utf-8") as f:
757
758
759
760
761
                json.dump(modelinfo_dict, f, indent=2)
        except Exception:
            logger.exception("Error saving model info cache.")

    @logtime(logger=logger, msg="Registry inspect model class")
762
    def inspect_model_cls(self) -> _ModelInfo:
763
        model_path = Path(__file__).parent / f"{self.module_name.split('.')[-1]}.py"
764
        module_hash = None
765

766
767
        if model_path.exists():
            with open(model_path, "rb") as f:
768
                module_hash = safe_hash(f.read(), usedforsecurity=False).hexdigest()
769
770
771

            mi = self._load_modelinfo_from_cache(module_hash)
            if mi is not None:
772
                logger.debug(
773
                    "Loaded model info for class %s.%s from cache",
774
775
776
                    self.module_name,
                    self.class_name,
                )
777
778
                return mi
            else:
779
                logger.debug(
780
                    "Cache model info for class %s.%s miss. Loading model instead.",
781
782
783
                    self.module_name,
                    self.class_name,
                )
784
785
786

        # Performed in another process to avoid initializing CUDA
        mi = _run_in_subprocess(
787
788
789
790
791
            lambda: _ModelInfo.from_model_cls(self.load_model_cls())
        )
        logger.debug(
            "Loaded model info for class %s.%s", self.module_name, self.class_name
        )
792
793

        # save cache file
794
795
        if module_hash is not None:
            self._save_modelinfo_to_cache(mi, module_hash)
796
797

        return mi
798

799
    def load_model_cls(self) -> type[nn.Module]:
800
801
802
803
804
805
806
807
        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,
808
) -> type[nn.Module] | None:
809
    from vllm.platforms import current_platform
810

811
    current_platform.verify_model_arch(model_arch)
812
813
814
    try:
        return model.load_model_cls()
    except Exception:
815
        logger.exception("Error in loading model architecture '%s'", model_arch)
816
        return None
817
818


819
820
821
822
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
823
) -> _ModelInfo | None:
824
825
826
    try:
        return model.inspect_model_cls()
    except Exception:
827
        logger.exception("Error in inspecting model architecture '%s'", model_arch)
828
        return None
829
830


831
832
833
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
834
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
835

836
    def get_supported_archs(self) -> Set[str]:
837
        return self.models.keys()
838

839
840
841
    def register_model(
        self,
        model_arch: str,
842
        model_cls: type[nn.Module] | str,
843
    ) -> None:
844
845
846
        """
        Register an external model to be used in vLLM.

847
        `model_cls` can be either:
848

849
        - A [`torch.nn.Module`][] class directly referencing the model.
850
        - A string in the format `<module>:<class>` which can be used to
851
852
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
853
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
854
        """
855
856
857
858
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

859
        if model_arch in self.models:
860
861
            logger.warning(
                "Model architecture %s is already registered, and will be "
862
863
864
865
                "overwritten by the new model class %s.",
                model_arch,
                model_cls,
            )
866
867
868
869
870
871

        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)
872

873
            model = _LazyRegisteredModel(*split_str)
874
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
875
            model = _RegisteredModel.from_model_cls(model_cls)
876
        else:
877
878
879
880
            msg = (
                "`model_cls` should be a string or PyTorch model class, "
                f"not a {type(model_arch)}"
            )
881
            raise TypeError(msg)
882

883
        self.models[model_arch] = model
884

885
    def _raise_for_unsupported(self, architectures: list[str]):
886
        all_supported_archs = self.get_supported_archs()
887

888
889
890
        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
891
892
                "to be inspected. Please check the logs for more details."
            )
893

894
895
896
897
898
899
900
901
        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 "
902
903
                    "use this model architecture."
                )
904

905
906
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
907
908
            f"Supported architectures: {all_supported_archs}"
        )
909

910
    def _try_load_model_cls(self, model_arch: str) -> type[nn.Module] | None:
911
912
        if model_arch not in self.models:
            return None
913

914
        return _try_load_model_cls(model_arch, self.models[model_arch])
915

916
    def _try_inspect_model_cls(self, model_arch: str) -> _ModelInfo | None:
917
918
        if model_arch not in self.models:
            return None
919

920
921
922
923
924
925
        return _try_inspect_model_cls(model_arch, self.models[model_arch])

    def _try_resolve_transformers(
        self,
        architecture: str,
        model_config: ModelConfig,
926
    ) -> str | None:
927
928
929
        if architecture in _TRANSFORMERS_BACKEND_MODELS:
            return architecture

930
931
932
        auto_map: dict[str, str] = (
            getattr(model_config.hf_config, "auto_map", None) or dict()
        )
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948

        # 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,
949
                        trust_remote_code=model_config.trust_remote_code,
950
951
952
953
954
955
956
957
958
959
960
961
                        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,
962
                        trust_remote_code=model_config.trust_remote_code,
963
964
965
966
967
                        warn_on_fail=True,
                    )
                    if model_module is not None:
                        break
            else:
968
                if model_config.model_impl != "transformers":
969
970
971
972
973
974
975
                    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 "
976
977
                    "'auto_map' (relevant if the model is custom)."
                )
978
979

        if not model_module.is_backend_compatible():
980
            if model_config.model_impl != "transformers":
981
                return None
982

983
984
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
985
986
                "is not compatible with vLLM."
            )
987

988
        return model_config._get_transformers_backend_cls()
989

990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
    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
1015

1016
1017
    def inspect_model_cls(
        self,
1018
        architectures: str | list[str],
1019
        model_config: ModelConfig,
1020
    ) -> tuple[_ModelInfo, str]:
1021
1022
        if isinstance(architectures, str):
            architectures = [architectures]
1023
1024
        if not architectures:
            raise ValueError("No model architectures are specified")
1025
1026

        # Require transformers impl
1027
        if model_config.model_impl == "transformers":
1028
            arch = self._try_resolve_transformers(architectures[0], model_config)
1029
1030
1031
1032
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)
1033
        elif model_config.model_impl == "terratorch":
1034
1035
            model_info = self._try_inspect_model_cls("Terratorch")
            return (model_info, "Terratorch")
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_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)
1051
            model_info = self._try_inspect_model_cls(normalized_arch)
1052
            if model_info is not None:
1053
                return (model_info, 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_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

1066
        return self._raise_for_unsupported(architectures)
1067

1068
1069
    def resolve_model_cls(
        self,
1070
        architectures: str | list[str],
1071
        model_config: ModelConfig,
1072
    ) -> tuple[type[nn.Module], str]:
1073
1074
        if isinstance(architectures, str):
            architectures = [architectures]
1075
1076
        if not architectures:
            raise ValueError("No model architectures are specified")
1077
1078

        # Require transformers impl
1079
        if model_config.model_impl == "transformers":
1080
            arch = self._try_resolve_transformers(architectures[0], model_config)
1081
1082
1083
1084
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)
1085
        elif model_config.model_impl == "terratorch":
1086
1087
1088
1089
            arch = "Terratorch"
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
1090

1091
        # Fallback to transformers impl (after resolving convert_type)
1092
1093
1094
1095
1096
1097
        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)
1098
1099
1100
1101
1102
1103
1104
            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)
1105
            model_cls = self._try_load_model_cls(normalized_arch)
1106
1107
            if model_cls is not None:
                return (model_cls, arch)
1108

1109
        # Fallback to transformers impl (before resolving runner_type)
1110
1111
1112
1113
1114
        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)
1115
1116
1117
1118
1119
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

1120
        return self._raise_for_unsupported(architectures)
1121

1122
1123
    def is_text_generation_model(
        self,
1124
        architectures: str | list[str],
1125
        model_config: ModelConfig,
1126
    ) -> bool:
1127
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1128
        return model_cls.is_text_generation_model
1129

1130
    def is_pooling_model(
1131
        self,
1132
        architectures: str | list[str],
1133
        model_config: ModelConfig,
1134
    ) -> bool:
1135
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1136
        return model_cls.is_pooling_model
1137

1138
1139
    def is_cross_encoder_model(
        self,
1140
        architectures: str | list[str],
1141
        model_config: ModelConfig,
1142
    ) -> bool:
1143
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1144
        return model_cls.supports_cross_encoding
1145

1146
1147
    def is_multimodal_model(
        self,
1148
        architectures: str | list[str],
1149
        model_config: ModelConfig,
1150
    ) -> bool:
1151
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1152
        return model_cls.supports_multimodal
1153

1154
    def is_multimodal_raw_input_only_model(
1155
        self,
1156
        architectures: str | list[str],
1157
        model_config: ModelConfig,
1158
    ) -> bool:
1159
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1160
        return model_cls.supports_multimodal_raw_input_only
1161

1162
1163
    def is_pp_supported_model(
        self,
1164
        architectures: str | list[str],
1165
        model_config: ModelConfig,
1166
    ) -> bool:
1167
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1168
        return model_cls.supports_pp
1169

1170
1171
    def model_has_inner_state(
        self,
1172
        architectures: str | list[str],
1173
        model_config: ModelConfig,
1174
    ) -> bool:
1175
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1176
        return model_cls.has_inner_state
1177

1178
1179
    def is_attention_free_model(
        self,
1180
        architectures: str | list[str],
1181
        model_config: ModelConfig,
1182
    ) -> bool:
1183
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1184
        return model_cls.is_attention_free
1185

1186
1187
    def is_hybrid_model(
        self,
1188
        architectures: str | list[str],
1189
        model_config: ModelConfig,
1190
    ) -> bool:
1191
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1192
1193
        return model_cls.is_hybrid

1194
1195
    def is_noops_model(
        self,
1196
        architectures: str | list[str],
1197
        model_config: ModelConfig,
1198
    ) -> bool:
1199
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1200
1201
        return model_cls.has_noops

1202
1203
    def is_transcription_model(
        self,
1204
        architectures: str | list[str],
1205
        model_config: ModelConfig,
1206
    ) -> bool:
1207
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1208
1209
        return model_cls.supports_transcription

1210
1211
    def is_transcription_only_model(
        self,
1212
        architectures: str | list[str],
1213
        model_config: ModelConfig,
1214
    ) -> bool:
1215
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1216
1217
        return model_cls.supports_transcription_only

1218

1219
1220
1221
1222
1223
1224
1225
1226
1227
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()
    }
)
1228
1229
1230
1231
1232

_T = TypeVar("_T")


def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
1233
1234
1235
1236
1237
    # 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")

1238
        # `cloudpickle` allows pickling lambda functions directly
1239
        import cloudpickle
1240

1241
        input_bytes = cloudpickle.dumps((fn, output_filepath))
1242
1243
1244

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
1245
1246
1247
        returned = subprocess.run(
            _SUBPROCESS_COMMAND, input=input_bytes, capture_output=True
        )
1248
1249
1250
1251
1252
1253

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
1254
1255
1256
            raise RuntimeError(
                f"Error raised in subprocess:\n{returned.stderr.decode()}"
            ) from e
1257

1258
        with open(output_filepath, "rb") as f:
1259
1260
1261
1262
1263
1264
            return pickle.load(f)


def _run() -> None:
    # Setup plugins
    from vllm.plugins import load_general_plugins
1265

1266
1267
1268
1269
1270
    load_general_plugins()

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

    result = fn()
1271
1272
1273

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
1274
1275
1276


if __name__ == "__main__":
1277
    _run()