registry.py 53.4 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.tasks import ScoreType
34
from vllm.transformers_utils.dynamic_module import try_get_class_from_dynamic_module
35
from vllm.utils.hashing import safe_hash
36

37
38
if TYPE_CHECKING:
    from vllm.config.model import AttnTypeStr
39
    from vllm.config.pooler import SequencePoolingType, TokenPoolingType
40
41
else:
    AttnTypeStr = Any
42
43
    SequencePoolingType = Any
    TokenPoolingType = Any
44
45


46
47
48
49
50
from .interfaces import (
    has_inner_state,
    has_noops,
    is_attention_free,
    is_hybrid,
Patrick von Platen's avatar
Patrick von Platen committed
51
    requires_raw_input_tokens,
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
    get_score_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
    "AXK1ForCausalLM": ("AXK1", "AXK1ForCausalLM"),
79
80
81
82
    # baichuan-7b, upper case 'C' in the class name
    "BaiChuanForCausalLM": ("baichuan", "BaiChuanForCausalLM"),
    # baichuan-13b, lower case 'c' in the class name
    "BaichuanForCausalLM": ("baichuan", "BaichuanForCausalLM"),
83
    "BailingMoeForCausalLM": ("bailing_moe", "BailingMoeForCausalLM"),
ant-yy's avatar
ant-yy committed
84
    "BailingMoeV2ForCausalLM": ("bailing_moe", "BailingMoeV2ForCausalLM"),
Jiangyun Zhu's avatar
Jiangyun Zhu committed
85
    "BailingMoeV2_5ForCausalLM": ("bailing_moe_linear", "BailingMoeV25ForCausalLM"),
Yu Chin Fabian Lim's avatar
Yu Chin Fabian Lim committed
86
    "BambaForCausalLM": ("bamba", "BambaForCausalLM"),
87
    "BloomForCausalLM": ("bloom", "BloomForCausalLM"),
88
    "ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
89
    "ChatGLMForConditionalGeneration": ("chatglm", "ChatGLMForCausalLM"),
90
    "CohereForCausalLM": ("commandr", "CohereForCausalLM"),
91
    "Cohere2ForCausalLM": ("commandr", "CohereForCausalLM"),
92
    "CwmForCausalLM": ("llama", "LlamaForCausalLM"),
93
    "DbrxForCausalLM": ("dbrx", "DbrxForCausalLM"),
94
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
95
    "DeepseekForCausalLM": ("deepseek_v2", "DeepseekForCausalLM"),
96
    "DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
97
    "DeepseekV3ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
98
    "DeepseekV32ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
99
    "Dots1ForCausalLM": ("dots1", "Dots1ForCausalLM"),
100
    "Ernie4_5ForCausalLM": ("ernie45", "Ernie4_5ForCausalLM"),
101
    "Ernie4_5_MoeForCausalLM": ("ernie45_moe", "Ernie4_5_MoeForCausalLM"),
102
    "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
103
    "Exaone4ForCausalLM": ("exaone4", "Exaone4ForCausalLM"),
Kyungmin Lee's avatar
Kyungmin Lee committed
104
    "ExaoneMoEForCausalLM": ("exaone_moe", "ExaoneMoeForCausalLM"),
105
    "Fairseq2LlamaForCausalLM": ("fairseq2_llama", "Fairseq2LlamaForCausalLM"),
106
107
108
    "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
    "FalconH1ForCausalLM": ("falcon_h1", "FalconH1ForCausalLM"),
109
    "FlexOlmoForCausalLM": ("flex_olmo", "FlexOlmoForCausalLM"),
110
111
    "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
    "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
112
    "Gemma3ForCausalLM": ("gemma3", "Gemma3ForCausalLM"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
113
    "Gemma3nForCausalLM": ("gemma3n", "Gemma3nForCausalLM"),
114
    "Gemma4ForCausalLM": ("gemma4", "Gemma4ForCausalLM"),
115
    "Qwen3NextForCausalLM": ("qwen3_next", "Qwen3NextForCausalLM"),
116
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
117
    "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
118
    "Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
119
    "Glm4MoeLiteForCausalLM": ("glm4_moe_lite", "Glm4MoeLiteForCausalLM"),
Jee Jee Li's avatar
Jee Jee Li committed
120
    "GlmMoeDsaForCausalLM": ("deepseek_v2", "GlmMoeDsaForCausalLM"),
121
    "GptOssForCausalLM": ("gpt_oss", "GptOssForCausalLM"),
122
123
124
125
126
127
    "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
    "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
    "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
    "GPTNeoXForCausalLM": ("gpt_neox", "GPTNeoXForCausalLM"),
    "GraniteForCausalLM": ("granite", "GraniteForCausalLM"),
    "GraniteMoeForCausalLM": ("granitemoe", "GraniteMoeForCausalLM"),
128
129
    "GraniteMoeHybridForCausalLM": ("granitemoehybrid", "GraniteMoeHybridForCausalLM"),
    "GraniteMoeSharedForCausalLM": ("granitemoeshared", "GraniteMoeSharedForCausalLM"),
130
    "GritLM": ("gritlm", "GritLM"),
Bijaya Dangol's avatar
Bijaya Dangol committed
131
132
    "Grok1ModelForCausalLM": ("grok1", "GrokForCausalLM"),
    "Grok1ForCausalLM": ("grok1", "GrokForCausalLM"),
133
134
    "HunYuanMoEV1ForCausalLM": ("hunyuan_v1", "HunYuanMoEV1ForCausalLM"),
    "HunYuanDenseV1ForCausalLM": ("hunyuan_v1", "HunYuanDenseV1ForCausalLM"),
135
    "HCXVisionForCausalLM": ("hyperclovax_vision", "HCXVisionForCausalLM"),
136
    "HCXVisionV2ForCausalLM": ("hyperclovax_vision_v2", "HCXVisionV2ForCausalLM"),
137
    "HyperCLOVAXForCausalLM": ("hyperclovax", "HyperCLOVAXForCausalLM"),
138
139
    "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
    "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"),
140
    "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
141
    "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
142
143
    "IQuestCoderForCausalLM": ("llama", "LlamaForCausalLM"),
    "IQuestLoopCoderForCausalLM": ("iquest_loopcoder", "IQuestLoopCoderForCausalLM"),
144
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
145
    "Jais2ForCausalLM": ("jais2", "Jais2ForCausalLM"),
146
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
147
    "KimiLinearForCausalLM": ("kimi_linear", "KimiLinearForCausalLM"),
148
    "Lfm2ForCausalLM": ("lfm2", "Lfm2ForCausalLM"),
Paul Pak's avatar
Paul Pak committed
149
    "Lfm2MoeForCausalLM": ("lfm2_moe", "Lfm2MoeForCausalLM"),
150
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
151
    "Llama4ForCausalLM": ("llama4", "Llama4ForCausalLM"),
152
153
    # For decapoda-research/llama-*
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
XuruiYang's avatar
XuruiYang committed
154
    "LongcatFlashForCausalLM": ("longcat_flash", "LongcatFlashForCausalLM"),
155
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
156
    "Mamba2ForCausalLM": ("mamba2", "Mamba2ForCausalLM"),
157
158
    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
    "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
159
160
161
    "MiniMaxForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxText01ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxM1ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
162
    "MiniMaxM2ForCausalLM": ("minimax_m2", "MiniMaxM2ForCausalLM"),
163
    "Ministral3ForCausalLM": ("mistral", "MistralForCausalLM"),
164
    "MistralForCausalLM": ("mistral", "MistralForCausalLM"),
165
    "MistralLarge3ForCausalLM": ("mistral_large_3", "MistralLarge3ForCausalLM"),
166
167
168
169
    "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
    # transformers's mpt class has lower case
    "MptForCausalLM": ("mpt", "MPTForCausalLM"),
    "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
170
    "MiMoForCausalLM": ("mimo", "MiMoForCausalLM"),
171
    "MiMoV2FlashForCausalLM": ("mimo_v2_flash", "MiMoV2FlashForCausalLM"),
172
    "NemotronForCausalLM": ("nemotron", "NemotronForCausalLM"),
Luis Vega's avatar
Luis Vega committed
173
    "NemotronHForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
174
    "NemotronHPuzzleForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
175
    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
176
    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
177
    "Olmo3ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
178
    "OlmoHybridForCausalLM": ("olmo_hybrid", "OlmoHybridForCausalLM"),
179
180
181
    "OlmoeForCausalLM": ("olmoe", "OlmoeForCausalLM"),
    "OPTForCausalLM": ("opt", "OPTForCausalLM"),
    "OrionForCausalLM": ("orion", "OrionForCausalLM"),
182
    "OuroForCausalLM": ("ouro", "OuroForCausalLM"),
183
    "PanguEmbeddedForCausalLM": ("openpangu", "PanguEmbeddedForCausalLM"),
184
    "PanguProMoEV2ForCausalLM": ("openpangu", "PanguProMoEV2ForCausalLM"),
185
    "PanguUltraMoEForCausalLM": ("openpangu", "PanguUltraMoEForCausalLM"),
186
    "Param2MoEForCausalLM": ("param2moe", "Param2MoEForCausalLM"),
187
188
189
190
    "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
    "PhiForCausalLM": ("phi", "PhiForCausalLM"),
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
    "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
191
    "Plamo2ForCausalLM": ("plamo2", "Plamo2ForCausalLM"),
192
    "Plamo3ForCausalLM": ("plamo3", "Plamo3ForCausalLM"),
193
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
194
195
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
196
197
    "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),
    "Qwen3MoeForCausalLM": ("qwen3_moe", "Qwen3MoeForCausalLM"),
198
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
199
200
    "SarvamMoEForCausalLM": ("sarvam", "SarvamMoEForCausalLM"),
    "SarvamMLAForCausalLM": ("sarvam", "SarvamMLAForCausalLM"),
201
    "SeedOssForCausalLM": ("seed_oss", "SeedOssForCausalLM"),
Li Xie's avatar
Li Xie committed
202
    "Step1ForCausalLM": ("step1", "Step1ForCausalLM"),
Song's avatar
Song committed
203
    "Step3TextForCausalLM": ("step3_text", "Step3TextForCausalLM"),
csy0225's avatar
csy0225 committed
204
    "Step3p5ForCausalLM": ("step3p5", "Step3p5ForCausalLM"),
205
206
207
208
    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
209
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
210
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
211
    "TeleChat3ForCausalLM": ("llama", "LlamaForCausalLM"),
212
    "TeleFLMForCausalLM": ("teleflm", "TeleFLMForCausalLM"),
213
    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
214
    "Zamba2ForCausalLM": ("zamba2", "Zamba2ForCausalLM"),
215
216
217
}

_EMBEDDING_MODELS = {
218
    # [Text-only]
219
    "BertModel": ("bert", "BertEmbeddingModel"),
220
    "BertSpladeSparseEmbeddingModel": ("bert", "BertSpladeSparseEmbeddingModel"),
221
    "ErnieModel": ("ernie", "ErnieEmbeddingModel"),
222
    "BgeM3EmbeddingModel": ("roberta", "BgeM3EmbeddingModel"),
223
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
224
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
225
    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
226
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
227
    "GritLM": ("gritlm", "GritLM"),
228
229
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
230
    "JinaEmbeddingsV5Model": ("jina", "JinaEmbeddingsV5Model"),
231
    "LlamaBidirectionalModel": ("llama", "LlamaBidirectionalModel"),
232
    "LlamaModel": ("llama", "LlamaForCausalLM"),
233
234
    **{
        # Multiple models share the same architecture, so we include them all
235
236
        k: (mod, arch)
        for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
237
238
        if arch == "LlamaForCausalLM"
    },
239
    "MistralModel": ("llama", "LlamaForCausalLM"),
240
    "ModernBertModel": ("modernbert", "ModernBertModel"),
241
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
242
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
243
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
244
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
245
246
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
247
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
248
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
chengchengpei's avatar
chengchengpei committed
249
250
251
252
    "VoyageQwen3BidirectionalEmbedModel": (
        "voyage",
        "VoyageQwen3BidirectionalEmbedModel",
    ),
253
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
254
    # [Multimodal]
255
    "CLIPModel": ("clip", "CLIPEmbeddingModel"),
256
    "ColPaliForRetrieval": ("colpali", "ColPaliModel"),
257
    "LlamaNemotronVLModel": ("nemotron_vl", "LlamaNemotronVLForEmbedding"),
258
259
260
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
261
    ),
262
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
263
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),
264
    "SiglipModel": ("siglip", "SiglipEmbeddingModel"),
265
266
    # Technically Terratorch models work on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
267
    # models for the time being.
268
269
    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
    "Terratorch": ("terratorch", "Terratorch"),
270
271
}

272
273
274
275
276
_LATE_INTERACTION_MODELS = {
    # [Text-only]
    "HF_ColBERT": ("colbert", "ColBERTModel"),
    "ColBERTModernBertModel": ("colbert", "ColBERTModernBertModel"),
    "ColBERTJinaRobertaModel": ("colbert", "ColBERTJinaRobertaModel"),
277
    "ColBERTLfm2Model": ("colbert", "ColBERTLfm2Model"),
278
    "JinaForRanking": ("jina", "JinaForRanking"),
279
280
    # [Multimodal]
    "ColModernVBertForRetrieval": ("colmodernvbert", "ColModernVBertForRetrieval"),
281
    "ColPaliForRetrieval": ("colpali", "ColPaliModel"),
282
283
    "ColQwen3": ("colqwen3", "ColQwen3Model"),
    "OpsColQwen3Model": ("colqwen3", "ColQwen3Model"),
284
    "ColQwen3_5": ("colqwen3_5", "ColQwen3_5Model"),
285
286
287
288
289
290
291
292
293
294
    "Qwen3VLNemotronEmbedModel": ("colqwen3", "ColQwen3Model"),
}

_REWARD_MODELS = {
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
}

_TOKEN_CLASSIFICATION_MODELS = {
295
    "BertForTokenClassification": ("bert", "BertForTokenClassification"),
296
    "ErnieForTokenClassification": ("ernie", "ErnieForTokenClassification"),
297
298
299
300
    "ModernBertForTokenClassification": (
        "modernbert",
        "ModernBertForTokenClassification",
    ),
301
302
303
304
    "Qwen3ASRForcedAlignerForTokenClassification": (
        "qwen3_asr_forced_aligner",
        "Qwen3ASRForcedAlignerForTokenClassification",
    ),
305
306
307
308
309
}

_SEQUENCE_CLASSIFICATION_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
310
    "ErnieForSequenceClassification": ("ernie", "ErnieForSequenceClassification"),
311
312
313
314
    "GteNewForSequenceClassification": (
        "bert_with_rope",
        "GteNewForSequenceClassification",
    ),
315
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),
316
317
318
    "LlamaBidirectionalForSequenceClassification": (
        "llama",
        "LlamaBidirectionalForSequenceClassification",
319
    ),
320
321
322
323
324
325
326
327
328
    "ModernBertForSequenceClassification": (
        "modernbert",
        "ModernBertForSequenceClassification",
    ),
    "RobertaForSequenceClassification": ("roberta", "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": (
        "roberta",
        "RobertaForSequenceClassification",
    ),
329
330
331
332
333
334
    # [Multimodal]
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"),
    "LlamaNemotronVLForSequenceClassification": (
        "nemotron_vl",
        "LlamaNemotronVLForSequenceClassification",
    ),
335
336
}

337
_MULTIMODAL_MODELS = {
338
    # [Decoder-only]
339
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
340
341
342
343
    "AudioFlamingo3ForConditionalGeneration": (
        "audioflamingo3",
        "AudioFlamingo3ForConditionalGeneration",
    ),
344
345
346
347
    "MusicFlamingoForConditionalGeneration": (
        "musicflamingo",
        "MusicFlamingoForConditionalGeneration",
    ),
348
349
350
    "AyaVisionForConditionalGeneration": (
        "aya_vision",
        "AyaVisionForConditionalGeneration",
351
    ),
352
    "BagelForConditionalGeneration": ("bagel", "BagelForConditionalGeneration"),
353
    "BeeForConditionalGeneration": ("bee", "BeeForConditionalGeneration"),
354
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
355
356
357
    "ChameleonForConditionalGeneration": (
        "chameleon",
        "ChameleonForConditionalGeneration",
358
    ),
359
360
    "Cheers": ("cheers", "CheersForConditionalGeneration"),
    "CheersForConditionalGeneration": ("cheers", "CheersForConditionalGeneration"),
361
362
363
    "Cohere2VisionForConditionalGeneration": (
        "cohere2_vision",
        "Cohere2VisionForConditionalGeneration",
364
    ),
365
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
366
    "DeepseekOCRForCausalLM": ("deepseek_ocr", "DeepseekOCRForCausalLM"),
RED's avatar
RED committed
367
    "DeepseekOCR2ForCausalLM": ("deepseek_ocr2", "DeepseekOCR2ForCausalLM"),
Roger Wang's avatar
Roger Wang committed
368
    "DotsOCRForCausalLM": ("dots_ocr", "DotsOCRForCausalLM"),
369
370
371
372
    "Eagle2_5_VLForConditionalGeneration": (
        "eagle2_5_vl",
        "Eagle2_5_VLForConditionalGeneration",
    ),
373
374
375
    "Ernie4_5_VLMoeForConditionalGeneration": (
        "ernie45_vl",
        "Ernie4_5_VLMoeForConditionalGeneration",
376
    ),
Kyungmin Lee's avatar
Kyungmin Lee committed
377
378
379
380
    "Exaone4_5_ForConditionalGeneration": (
        "exaone4_5",
        "Exaone4_5_ForConditionalGeneration",
    ),  # noqa: E501
381
382
383
384
    "FireRedASR2ForConditionalGeneration": (
        "fireredasr2",
        "FireRedASR2ForConditionalGeneration",
    ),
385
    "FunASRForConditionalGeneration": ("funasr", "FunASRForConditionalGeneration"),
386
387
388
389
    "FireRedLIDForConditionalGeneration": (
        "fireredlid",
        "FireRedLIDForConditionalGeneration",
    ),
390
391
392
393
    "FunAudioChatForConditionalGeneration": (
        "funaudiochat",
        "FunAudioChatForConditionalGeneration",
    ),
394
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
395
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),
396
397
398
    "Gemma3nForConditionalGeneration": (
        "gemma3n_mm",
        "Gemma3nForConditionalGeneration",
399
    ),
400
    "Gemma4ForConditionalGeneration": ("gemma4_mm", "Gemma4ForConditionalGeneration"),
401
    "GlmAsrForConditionalGeneration": ("glmasr", "GlmAsrForConditionalGeneration"),
402
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
403
404
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),
    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),
405
    "GlmOcrForConditionalGeneration": ("glm_ocr", "GlmOcrForConditionalGeneration"),
406
407
408
    "GraniteSpeechForConditionalGeneration": (
        "granite_speech",
        "GraniteSpeechForConditionalGeneration",
409
    ),
410
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
411
412
413
414
    "HunYuanVLForConditionalGeneration": (
        "hunyuan_vision",
        "HunYuanVLForConditionalGeneration",
    ),
415
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
416
417
418
    "InternS1ForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
419
    ),
420
421
422
    "InternVLForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
423
    ),
zxy's avatar
zxy committed
424
425
426
427
    "InternS1ProForConditionalGeneration": (
        "interns1_pro",
        "InternS1ProForConditionalGeneration",
    ),
428
429
430
431
    "Idefics3ForConditionalGeneration": (
        "idefics3",
        "Idefics3ForConditionalGeneration",
    ),
oscardev256's avatar
oscardev256 committed
432
    "IsaacForConditionalGeneration": ("isaac", "IsaacForConditionalGeneration"),
433
    "KananaVForConditionalGeneration": ("kanana_v", "KananaVForConditionalGeneration"),
434
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
435
436
437
    "KeyeVL1_5ForConditionalGeneration": (
        "keye_vl1_5",
        "KeyeVL1_5ForConditionalGeneration",
438
    ),
439
440
441
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),
    "KimiK25ForConditionalGeneration": ("kimi_k25", "KimiK25ForConditionalGeneration"),
    "MoonshotKimiaForCausalLM": ("kimi_audio", "KimiAudioForConditionalGeneration"),
442
443
444
445
    "LightOnOCRForConditionalGeneration": (
        "lightonocr",
        "LightOnOCRForConditionalGeneration",
    ),
446
    "Lfm2VlForConditionalGeneration": ("lfm2_vl", "Lfm2VLForConditionalGeneration"),
447
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),
448
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
449
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
450
451
452
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
453
    ),
454
455
456
    "LlavaNextVideoForConditionalGeneration": (
        "llava_next_video",
        "LlavaNextVideoForConditionalGeneration",
457
    ),
458
459
460
    "LlavaOnevisionForConditionalGeneration": (
        "llava_onevision",
        "LlavaOnevisionForConditionalGeneration",
461
    ),
462
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),
463
    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
464
465
466
    "MiniMaxVL01ForConditionalGeneration": (
        "minimax_vl_01",
        "MiniMaxVL01ForConditionalGeneration",
467
    ),
468
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
469
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
470
471
472
    "Mistral3ForConditionalGeneration": (
        "mistral3",
        "Mistral3ForConditionalGeneration",
473
    ),
474
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
475
    "Molmo2ForConditionalGeneration": ("molmo2", "Molmo2ForConditionalGeneration"),
476
    "NemotronH_Nano_VL_V2": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
477
478
    "NemotronH_Nano_Omni_Reasoning_V3": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
    "NemotronH_Super_Omni_Reasoning_V3": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
479
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
480
    "OpenCUAForConditionalGeneration": ("opencua", "OpenCUAForConditionalGeneration"),
481
482
483
484
    "OpenPanguVLForConditionalGeneration": (
        "openpangu_vl",
        "OpenPanguVLForConditionalGeneration",
    ),
485
    "Ovis": ("ovis", "Ovis"),
486
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
487
488
    "Ovis2_6ForCausalLM": ("ovis2_5", "Ovis2_5"),
    "Ovis2_6_MoeForCausalLM": ("ovis2_5", "Ovis2_5"),
489
490
491
492
    "PaddleOCRVLForConditionalGeneration": (
        "paddleocr_vl",
        "PaddleOCRVLForConditionalGeneration",
    ),
493
494
495
496
    "PaliGemmaForConditionalGeneration": (
        "paligemma",
        "PaliGemmaForConditionalGeneration",
    ),
497
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
498
    "Phi4ForCausalLMV": ("phi4siglip", "Phi4ForCausalLMV"),
499
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
500
501
502
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),
503
504
505
    "Qwen2_5_VLForConditionalGeneration": (
        "qwen2_5_vl",
        "Qwen2_5_VLForConditionalGeneration",
506
    ),
507
508
509
    "Qwen2AudioForConditionalGeneration": (
        "qwen2_audio",
        "Qwen2AudioForConditionalGeneration",
510
    ),
511
512
513
    "Qwen2_5OmniModel": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
514
    ),
515
516
517
    "Qwen2_5OmniForConditionalGeneration": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
518
    ),
519
520
521
522
    "Qwen3OmniMoeForConditionalGeneration": (
        "qwen3_omni_moe_thinker",
        "Qwen3OmniMoeThinkerForConditionalGeneration",
    ),
Roger Wang's avatar
Roger Wang committed
523
524
525
526
    "Qwen3ASRForConditionalGeneration": (
        "qwen3_asr",
        "Qwen3ASRForConditionalGeneration",
    ),
527
528
    "Qwen3ASRRealtimeGeneration": ("qwen3_asr_realtime", "Qwen3ASRRealtimeGeneration"),
    "Qwen3VLForConditionalGeneration": ("qwen3_vl", "Qwen3VLForConditionalGeneration"),
529
530
531
    "Qwen3VLMoeForConditionalGeneration": (
        "qwen3_vl_moe",
        "Qwen3VLMoeForConditionalGeneration",
532
    ),
533
    "Qwen3_5ForConditionalGeneration": ("qwen3_5", "Qwen3_5ForConditionalGeneration"),
534
535
536
537
    "Qwen3_5MoeForConditionalGeneration": (
        "qwen3_5",
        "Qwen3_5MoeForConditionalGeneration",
    ),
538
    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
539
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
540
541
542
543
    "SmolVLMForConditionalGeneration": ("smolvlm", "SmolVLMForConditionalGeneration"),
    "StepVLForConditionalGeneration": ("step_vl", "StepVLForConditionalGeneration"),
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),
544
545
546
    "Tarsier2ForConditionalGeneration": (
        "qwen2_vl",
        "Tarsier2ForConditionalGeneration",
547
    ),
548
    "UltravoxModel": ("ultravox", "UltravoxModel"),
549
550
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),
    "VoxtralRealtimeGeneration": ("voxtral_realtime", "VoxtralRealtimeGeneration"),
551
    # [Encoder-decoder]
552
    "CohereAsrForConditionalGeneration": (
Ekagra Ranjan's avatar
Ekagra Ranjan committed
553
        "cohere_asr",
554
        "CohereAsrForConditionalGeneration",
Ekagra Ranjan's avatar
Ekagra Ranjan committed
555
    ),
556
557
558
559
    "NemotronParseForConditionalGeneration": (
        "nemotron_parse",
        "NemotronParseForConditionalGeneration",
    ),
560
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),
561
}
562
563

_SPECULATIVE_DECODING_MODELS = {
564
    "ExtractHiddenStatesModel": ("extract_hidden_states", "ExtractHiddenStatesModel"),
565
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
566
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
567
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
568
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
569
    "DFlashDraftModel": ("qwen3_dflash", "DFlashQwen3ForCausalLM"),
570
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
571
    "Eagle3MiniMaxM2ForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
572
    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
573
    "Eagle3Qwen2_5vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
574
    "Eagle3Qwen3vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
575
576
577
578
    "EagleMistralLarge3ForCausalLM": (
        "mistral_large_3_eagle",
        "EagleMistralLarge3ForCausalLM",
    ),
579
580
    "Eagle3DeepseekV2ForCausalLM": ("deepseek_eagle3", "Eagle3DeepseekV2ForCausalLM"),
    "Eagle3DeepseekV3ForCausalLM": ("deepseek_eagle3", "Eagle3DeepseekV2ForCausalLM"),
581
    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
582
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
583
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
Kyungmin Lee's avatar
Kyungmin Lee committed
584
    "ExaoneMoeMTP": ("exaone_moe_mtp", "ExaoneMoeMTP"),
Kyungmin Lee's avatar
Kyungmin Lee committed
585
    "Exaone4_5_MTP": ("exaone4_5_mtp", "Exaone4_5_MTP"),
586
    "NemotronHMTPModel": ("nemotron_h_mtp", "NemotronHMTP"),
XuruiYang's avatar
XuruiYang committed
587
    "LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
588
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
589
    "Glm4MoeLiteMTPModel": ("glm4_moe_lite_mtp", "Glm4MoeLiteMTP"),
590
    "GlmOcrMTPModel": ("glm_ocr_mtp", "GlmOcrMTP"),
591
    "MedusaModel": ("medusa", "Medusa"),
592
    "OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
593
    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
csy0225's avatar
csy0225 committed
594
    "Step3p5MTP": ("step3p5_mtp", "Step3p5MTP"),
595
596
    "Qwen3_5MTP": ("qwen3_5_mtp", "Qwen3_5MTP"),
    "Qwen3_5MoeMTP": ("qwen3_5_mtp", "Qwen3_5MoeMTP"),
597
598
599
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
600
}
601

602
_TRANSFORMERS_SUPPORTED_MODELS = {
603
604
605
    # Text generation models
    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
    # Multimodal models
606
607
608
609
    "Emu3ForConditionalGeneration": (
        "transformers",
        "TransformersMultiModalForCausalLM",
    ),
610
611
612
}

_TRANSFORMERS_BACKEND_MODELS = {
613
    # Text generation models
614
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
    "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
633
    "TransformersForSequenceClassification": (
634
        "transformers",
635
        "TransformersForSequenceClassification",
636
    ),
637
    "TransformersMoEForSequenceClassification": (
638
        "transformers",
639
        "TransformersMoEForSequenceClassification",
640
    ),
641
642
643
    "TransformersMultiModalForSequenceClassification": (
        "transformers",
        "TransformersMultiModalForSequenceClassification",
644
    ),
645
}
646

647
_VLLM_MODELS = {
648
    **_TEXT_GENERATION_MODELS,
649
    **_EMBEDDING_MODELS,
650
651
652
653
    **_LATE_INTERACTION_MODELS,
    **_REWARD_MODELS,
    **_TOKEN_CLASSIFICATION_MODELS,
    **_SEQUENCE_CLASSIFICATION_MODELS,
654
    **_MULTIMODAL_MODELS,
655
    **_SPECULATIVE_DECODING_MODELS,
656
657
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
658
659
}

660
661
662
663
# 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.
664
_SUBPROCESS_COMMAND = [sys.executable, "-m", "vllm.model_executor.models.registry"]
665

666
_PREVIOUSLY_SUPPORTED_MODELS = {
667
    "MotifForCausalLM": "0.10.2",
668
    "Phi3SmallForCausalLM": "0.9.2",
669
    "Phi4FlashForCausalLM": "0.10.2",
670
    "Phi4MultimodalForCausalLM": "0.12.0",
671
672
673
674
675
    # encoder-decoder models except whisper
    # have been removed for V0 deprecation.
    "DonutForConditionalGeneration": "0.10.2",
    "MllamaForConditionalGeneration": "0.10.2",
}
676

677
678
679
680
681
682
683
_OOT_SUPPORTED_MODELS = {
    "BartModel": "https://github.com/vllm-project/bart-plugin",
    "BartForConditionalGeneration": "https://github.com/vllm-project/bart-plugin",
    "Florence2ForConditionalGeneration": "https://github.com/vllm-project/bart-plugin",
    "MBartForConditionalGeneration": "https://github.com/vllm-project/bart-plugin",
}

684

685
686
@dataclass(frozen=True)
class _ModelInfo:
687
    architecture: str
688
    is_text_generation_model: bool
689
    is_pooling_model: bool
690
    attn_type: AttnTypeStr
691
692
    default_seq_pooling_type: SequencePoolingType
    default_tok_pooling_type: TokenPoolingType
693
    score_type: ScoreType
694
    supports_multimodal: bool
695
    supports_multimodal_raw_input_only: bool
Patrick von Platen's avatar
Patrick von Platen committed
696
    requires_raw_input_tokens: bool
697
    supports_multimodal_encoder_tp_data: bool
698
    supports_pp: bool
699
700
    has_inner_state: bool
    is_attention_free: bool
701
    is_hybrid: bool
702
    has_noops: bool
703
    supports_mamba_prefix_caching: bool
704
    supports_transcription: bool
705
    supports_transcription_only: bool
706
707

    @staticmethod
708
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
709
        return _ModelInfo(
710
            architecture=model.__name__,
711
            is_text_generation_model=is_text_generation_model(model),
712
            is_pooling_model=is_pooling_model(model),
713
714
            default_seq_pooling_type=get_default_seq_pooling_type(model),
            default_tok_pooling_type=get_default_tok_pooling_type(model),
715
            attn_type=get_attn_type(model),
716
            score_type=get_score_type(model),
717
            supports_multimodal=supports_multimodal(model),
718
719
720
            supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
                model
            ),
Patrick von Platen's avatar
Patrick von Platen committed
721
            requires_raw_input_tokens=requires_raw_input_tokens(model),
722
723
724
            supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
                model
            ),
725
            supports_pp=supports_pp(model),
726
727
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
728
            is_hybrid=is_hybrid(model),
729
            supports_mamba_prefix_caching=supports_mamba_prefix_caching(model),
730
            supports_transcription=supports_transcription(model),
731
732
733
            supports_transcription_only=(
                supports_transcription(model) and model.supports_transcription_only
            ),
734
            has_noops=has_noops(model),
735
        )
736
737


738
739
740
741
class _BaseRegisteredModel(ABC):
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
742

743
    @abstractmethod
744
    def load_model_cls(self) -> type[nn.Module]:
745
        raise NotImplementedError
746
747


748
749
750
751
752
753
754
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
755
    model_cls: type[nn.Module]
756
757

    @staticmethod
758
    def from_model_cls(model_cls: type[nn.Module]):
759
760
761
762
763
764
765
766
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

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

767
    def load_model_cls(self) -> type[nn.Module]:
768
769
770
771
772
773
774
775
        return self.model_cls


@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """
776

777
778
779
    module_name: str
    class_name: str

780
781
782
783
784
785
786
787
    @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"

788
    def _load_modelinfo_from_cache(self, module_hash: str) -> _ModelInfo | None:
789
790
        try:
            try:
791
                modelinfo_path = self._get_cache_dir() / self._get_cache_filename()
792
793
794
                with open(modelinfo_path, encoding="utf-8") as file:
                    mi_dict = json.load(file)
            except FileNotFoundError:
795
                logger.debug(
796
                    "Cached model info file for class %s.%s not found",
797
798
799
                    self.module_name,
                    self.class_name,
                )
800
801
802
                return None

            if mi_dict["hash"] != module_hash:
803
                logger.debug(
804
                    "Cached model info file for class %s.%s is stale",
805
806
807
                    self.module_name,
                    self.class_name,
                )
808
809
810
811
812
                return None

            # file not changed, use cached _ModelInfo properties
            return _ModelInfo(**mi_dict["modelinfo"])
        except Exception:
813
            logger.debug(
814
                "Cached model info for class %s.%s error. ",
815
816
817
                self.module_name,
                self.class_name,
            )
818
819
            return None

820
    def _save_modelinfo_to_cache(self, mi: _ModelInfo, module_hash: str) -> None:
821
822
        """save dictionary json file to cache"""
        from vllm.model_executor.model_loader.weight_utils import atomic_writer
823

824
825
826
827
828
829
830
831
        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()
832
            with atomic_writer(modelinfo_path, encoding="utf-8") as f:
833
834
835
836
837
                json.dump(modelinfo_dict, f, indent=2)
        except Exception:
            logger.exception("Error saving model info cache.")

    @logtime(logger=logger, msg="Registry inspect model class")
838
    def inspect_model_cls(self) -> _ModelInfo:
839
        model_path = Path(__file__).parent / f"{self.module_name.split('.')[-1]}.py"
840
        module_hash = None
841

842
843
        if model_path.exists():
            with open(model_path, "rb") as f:
844
                module_hash = safe_hash(f.read(), usedforsecurity=False).hexdigest()
845
846
847

            mi = self._load_modelinfo_from_cache(module_hash)
            if mi is not None:
848
                logger.debug(
849
                    "Loaded model info for class %s.%s from cache",
850
851
852
                    self.module_name,
                    self.class_name,
                )
853
854
                return mi
            else:
855
                logger.debug(
856
                    "Cache model info for class %s.%s miss. Loading model instead.",
857
858
859
                    self.module_name,
                    self.class_name,
                )
860
861
862

        # Performed in another process to avoid initializing CUDA
        mi = _run_in_subprocess(
863
864
865
866
867
            lambda: _ModelInfo.from_model_cls(self.load_model_cls())
        )
        logger.debug(
            "Loaded model info for class %s.%s", self.module_name, self.class_name
        )
868
869

        # save cache file
870
871
        if module_hash is not None:
            self._save_modelinfo_to_cache(mi, module_hash)
872
873

        return mi
874

875
    def load_model_cls(self) -> type[nn.Module]:
876
877
878
879
880
881
882
883
        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,
884
) -> type[nn.Module] | None:
885
    from vllm.platforms import current_platform
886

887
    current_platform.verify_model_arch(model_arch)
888
889
890
    try:
        return model.load_model_cls()
    except Exception:
891
        logger.exception("Error in loading model architecture '%s'", model_arch)
892
        return None
893
894


895
896
897
898
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
899
) -> _ModelInfo | None:
900
901
902
    try:
        return model.inspect_model_cls()
    except Exception:
903
        logger.exception("Error in inspecting model architecture '%s'", model_arch)
904
        return None
905
906


907
908
909
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
910
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
911

912
    def get_supported_archs(self) -> Set[str]:
913
        return self.models.keys()
914

915
916
917
    def register_model(
        self,
        model_arch: str,
918
        model_cls: type[nn.Module] | str,
919
    ) -> None:
920
921
922
        """
        Register an external model to be used in vLLM.

923
        `model_cls` can be either:
924

925
        - A [`torch.nn.Module`][] class directly referencing the model.
926
        - A string in the format `<module>:<class>` which can be used to
927
928
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
929
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
930
        """
931
932
933
934
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

935
        if model_arch in self.models:
936
937
            logger.warning(
                "Model architecture %s is already registered, and will be "
938
939
940
941
                "overwritten by the new model class %s.",
                model_arch,
                model_cls,
            )
942
943
944
945
946
947

        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)
948

949
            model = _LazyRegisteredModel(*split_str)
950
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
951
            model = _RegisteredModel.from_model_cls(model_cls)
952
        else:
953
954
955
956
            msg = (
                "`model_cls` should be a string or PyTorch model class, "
                f"not a {type(model_arch)}"
            )
957
            raise TypeError(msg)
958

959
        self.models[model_arch] = model
960

961
    def _raise_for_unsupported(self, architectures: list[str]):
962
        all_supported_archs = self.get_supported_archs()
963

964
965
966
        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
967
968
                "to be inspected. Please check the logs for more details."
            )
969

970
971
972
973
974
975
976
977
        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 "
978
979
                    "use this model architecture."
                )
980
981
982
983
984
985
986
987
            if arch in _OOT_SUPPORTED_MODELS:
                plugin_url = _OOT_SUPPORTED_MODELS[arch]

                raise ValueError(
                    f"Model architecture {arch} is not supported in-tree anymore. "
                    f"Please install the plugin at {plugin_url} if you want to "
                    "use this model architecture."
                )
988

989
990
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
991
992
            f"Supported architectures: {all_supported_archs}"
        )
993

994
    def _try_load_model_cls(self, model_arch: str) -> type[nn.Module] | None:
995
996
        if model_arch not in self.models:
            return None
997

998
        return _try_load_model_cls(model_arch, self.models[model_arch])
999

1000
    def _try_inspect_model_cls(self, model_arch: str) -> _ModelInfo | None:
1001
1002
        if model_arch not in self.models:
            return None
1003

1004
1005
1006
1007
1008
1009
        return _try_inspect_model_cls(model_arch, self.models[model_arch])

    def _try_resolve_transformers(
        self,
        architecture: str,
        model_config: ModelConfig,
1010
    ) -> str | None:
1011
1012
1013
        if architecture in _TRANSFORMERS_BACKEND_MODELS:
            return architecture

1014
1015
1016
        auto_map: dict[str, str] = (
            getattr(model_config.hf_config, "auto_map", None) or dict()
        )
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032

        # 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,
1033
                        trust_remote_code=model_config.trust_remote_code,
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
                        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,
1046
                        trust_remote_code=model_config.trust_remote_code,
1047
1048
1049
1050
1051
                        warn_on_fail=True,
                    )
                    if model_module is not None:
                        break
            else:
1052
                if model_config.model_impl != "transformers":
1053
1054
1055
1056
1057
1058
1059
                    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 "
1060
1061
                    "'auto_map' (relevant if the model is custom)."
                )
1062
1063

        if not model_module.is_backend_compatible():
1064
            if model_config.model_impl != "transformers":
1065
                return None
1066

1067
1068
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
1069
1070
                "is not compatible with vLLM."
            )
1071

1072
        return model_config._get_transformers_backend_cls()
1073

1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
    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
1099

1100
1101
    def inspect_model_cls(
        self,
1102
        architectures: str | list[str],
1103
        model_config: ModelConfig,
1104
    ) -> tuple[_ModelInfo, str]:
1105
1106
        if isinstance(architectures, str):
            architectures = [architectures]
1107
1108
        if not architectures:
            raise ValueError("No model architectures are specified")
1109
1110

        # Require transformers impl
1111
        if model_config.model_impl == "transformers":
1112
            arch = self._try_resolve_transformers(architectures[0], model_config)
1113
1114
1115
1116
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)
1117
        elif model_config.model_impl == "terratorch":
1118
1119
            model_info = self._try_inspect_model_cls("Terratorch")
            return (model_info, "Terratorch")
1120

1121
        # Fallback to transformers impl (after resolving convert_type)
1122
1123
1124
1125
1126
1127
        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)
1128
1129
1130
1131
1132
1133
1134
            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)
1135
            model_info = self._try_inspect_model_cls(normalized_arch)
1136
            if model_info is not None:
1137
                return (model_info, arch)
1138

1139
        # Fallback to transformers impl (before resolving runner_type)
1140
1141
1142
1143
1144
        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)
1145
1146
1147
1148
1149
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

1150
        return self._raise_for_unsupported(architectures)
1151

1152
1153
    def resolve_model_cls(
        self,
1154
        architectures: str | list[str],
1155
        model_config: ModelConfig,
1156
    ) -> tuple[type[nn.Module], str]:
1157
1158
        if isinstance(architectures, str):
            architectures = [architectures]
1159
1160
        if not architectures:
            raise ValueError("No model architectures are specified")
1161
1162

        # Require transformers impl
1163
        if model_config.model_impl == "transformers":
1164
            arch = self._try_resolve_transformers(architectures[0], model_config)
1165
1166
1167
1168
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)
1169
        elif model_config.model_impl == "terratorch":
1170
1171
1172
1173
            arch = "Terratorch"
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
1174

1175
        # Fallback to transformers impl (after resolving convert_type)
1176
1177
1178
1179
1180
1181
        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)
1182
1183
1184
1185
1186
1187
1188
            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)
1189
            model_cls = self._try_load_model_cls(normalized_arch)
1190
1191
            if model_cls is not None:
                return (model_cls, arch)
1192

1193
        # Fallback to transformers impl (before resolving runner_type)
1194
1195
1196
1197
1198
        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)
1199
1200
1201
1202
1203
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

1204
        return self._raise_for_unsupported(architectures)
1205

1206
1207
    def is_text_generation_model(
        self,
1208
        architectures: str | list[str],
1209
        model_config: ModelConfig,
1210
    ) -> bool:
1211
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1212
        return model_cls.is_text_generation_model
1213

1214
    def is_pooling_model(
1215
        self,
1216
        architectures: str | list[str],
1217
        model_config: ModelConfig,
1218
    ) -> bool:
1219
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1220
        return model_cls.is_pooling_model
1221
1222
1223

    def is_multimodal_model(
        self,
1224
        architectures: str | list[str],
1225
        model_config: ModelConfig,
1226
    ) -> bool:
1227
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1228
        return model_cls.supports_multimodal
1229

1230
    def is_multimodal_raw_input_only_model(
1231
        self,
1232
        architectures: str | list[str],
1233
        model_config: ModelConfig,
1234
    ) -> bool:
1235
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1236
        return model_cls.supports_multimodal_raw_input_only
1237

1238
1239
    def is_pp_supported_model(
        self,
1240
        architectures: str | list[str],
1241
        model_config: ModelConfig,
1242
    ) -> bool:
1243
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1244
        return model_cls.supports_pp
1245

1246
1247
    def model_has_inner_state(
        self,
1248
        architectures: str | list[str],
1249
        model_config: ModelConfig,
1250
    ) -> bool:
1251
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1252
        return model_cls.has_inner_state
1253

1254
1255
    def is_attention_free_model(
        self,
1256
        architectures: str | list[str],
1257
        model_config: ModelConfig,
1258
    ) -> bool:
1259
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1260
        return model_cls.is_attention_free
1261

1262
1263
    def is_hybrid_model(
        self,
1264
        architectures: str | list[str],
1265
        model_config: ModelConfig,
1266
    ) -> bool:
1267
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1268
1269
        return model_cls.is_hybrid

1270
1271
    def is_noops_model(
        self,
1272
        architectures: str | list[str],
1273
        model_config: ModelConfig,
1274
    ) -> bool:
1275
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1276
1277
        return model_cls.has_noops

1278
1279
    def is_transcription_model(
        self,
1280
        architectures: str | list[str],
1281
        model_config: ModelConfig,
1282
    ) -> bool:
1283
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1284
1285
        return model_cls.supports_transcription

1286
1287
    def is_transcription_only_model(
        self,
1288
        architectures: str | list[str],
1289
        model_config: ModelConfig,
1290
    ) -> bool:
1291
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1292
1293
        return model_cls.supports_transcription_only

1294

1295
1296
1297
1298
1299
1300
1301
1302
1303
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()
    }
)
1304
1305
1306
1307
1308

_T = TypeVar("_T")


def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
1309
1310
1311
1312
1313
    # 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")

1314
        # `cloudpickle` allows pickling lambda functions directly
1315
        import cloudpickle
1316

1317
        input_bytes = cloudpickle.dumps((fn, output_filepath))
1318
1319
1320

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
1321
1322
1323
        returned = subprocess.run(
            _SUBPROCESS_COMMAND, input=input_bytes, capture_output=True
        )
1324
1325
1326
1327
1328
1329

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
1330
1331
1332
            raise RuntimeError(
                f"Error raised in subprocess:\n{returned.stderr.decode()}"
            ) from e
1333

1334
        with open(output_filepath, "rb") as f:
1335
1336
1337
1338
1339
1340
            return pickle.load(f)


def _run() -> None:
    # Setup plugins
    from vllm.plugins import load_general_plugins
1341

1342
1343
1344
1345
1346
    load_general_plugins()

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

    result = fn()
1347
1348
1349

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
1350
1351
1352


if __name__ == "__main__":
1353
    _run()