registry.py 47.8 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_mamba_prefix_caching,
53
54
55
56
57
58
59
    supports_multimodal,
    supports_multimodal_encoder_tp_data,
    supports_multimodal_raw_input_only,
    supports_pp,
    supports_transcription,
)
from .interfaces_base import (
60
    get_attn_type,
61
62
    get_default_seq_pooling_type,
    get_default_tok_pooling_type,
63
64
65
    is_pooling_model,
    is_text_generation_model,
)
66
67
68

logger = init_logger(__name__)

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

_EMBEDDING_MODELS = {
203
    # [Text-only]
204
    "BertModel": ("bert", "BertEmbeddingModel"),
205
    "BertSpladeSparseEmbeddingModel": ("bert", "BertSpladeSparseEmbeddingModel"),
206
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
207
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
208
    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
209
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
210
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
211
    "GritLM": ("gritlm", "GritLM"),
212
213
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
214
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
215
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
216
    "LlamaBidirectionalModel": ("llama", "LlamaBidirectionalModel"),
217
    "LlamaModel": ("llama", "LlamaForCausalLM"),
218
219
    **{
        # Multiple models share the same architecture, so we include them all
220
221
        k: (mod, arch)
        for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
222
223
        if arch == "LlamaForCausalLM"
    },
224
    "MistralModel": ("llama", "LlamaForCausalLM"),
225
    "ModernBertModel": ("modernbert", "ModernBertModel"),
226
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
227
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
228
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
229
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
230
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
231
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
232
233
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
234
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
235
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
236
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
237
    "BgeM3EmbeddingModel": ("roberta", "BgeM3EmbeddingModel"),
238
    # [Multimodal]
239
    "CLIPModel": ("clip", "CLIPEmbeddingModel"),
240
241
242
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
243
    ),
244
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
245
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
246
    "SiglipModel": ("siglip", "SiglipEmbeddingModel"),
247
248
    # Technically Terratorch models work on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
249
    # models for the time being.
250
251
    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
    "Terratorch": ("terratorch", "Terratorch"),
252
253
}

254
255
_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
256
    "BertForTokenClassification": ("bert", "BertForTokenClassification"),
257
258
259
260
    "GteNewForSequenceClassification": (
        "bert_with_rope",
        "GteNewForSequenceClassification",
    ),
261
262
263
264
265
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"),
    "LlamaBidirectionalForSequenceClassification": (
        "llama",
        "LlamaBidirectionalForSequenceClassification",
    ),
266
267
268
269
    "ModernBertForSequenceClassification": (
        "modernbert",
        "ModernBertForSequenceClassification",
    ),
270
271
272
273
    "ModernBertForTokenClassification": (
        "modernbert",
        "ModernBertForTokenClassification",
    ),
274
275
276
277
278
    "RobertaForSequenceClassification": ("roberta", "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": (
        "roberta",
        "RobertaForSequenceClassification",
    ),
279
280
}

281
_MULTIMODAL_MODELS = {
282
    # [Decoder-only]
283
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
284
285
286
287
    "AudioFlamingo3ForConditionalGeneration": (
        "audioflamingo3",
        "AudioFlamingo3ForConditionalGeneration",
    ),
288
289
290
    "AyaVisionForConditionalGeneration": (
        "aya_vision",
        "AyaVisionForConditionalGeneration",
291
    ),
292
    "BagelForConditionalGeneration": ("bagel", "BagelForConditionalGeneration"),
293
    "BeeForConditionalGeneration": ("bee", "BeeForConditionalGeneration"),
294
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
295
296
297
    "ChameleonForConditionalGeneration": (
        "chameleon",
        "ChameleonForConditionalGeneration",
298
    ),
299
300
301
    "Cohere2VisionForConditionalGeneration": (
        "cohere2_vision",
        "Cohere2VisionForConditionalGeneration",
302
    ),
303
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
304
    "DeepseekOCRForCausalLM": ("deepseek_ocr", "DeepseekOCRForCausalLM"),
Roger Wang's avatar
Roger Wang committed
305
    "DotsOCRForCausalLM": ("dots_ocr", "DotsOCRForCausalLM"),
306
307
308
309
    "Eagle2_5_VLForConditionalGeneration": (
        "eagle2_5_vl",
        "Eagle2_5_VLForConditionalGeneration",
    ),
310
311
312
    "Ernie4_5_VLMoeForConditionalGeneration": (
        "ernie45_vl",
        "Ernie4_5_VLMoeForConditionalGeneration",
313
    ),
314
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
315
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
316
317
318
    "Gemma3nForConditionalGeneration": (
        "gemma3n_mm",
        "Gemma3nForConditionalGeneration",
319
    ),
320
    "GlmAsrForConditionalGeneration": ("glmasr", "GlmAsrForConditionalGeneration"),
321
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
322
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),  # noqa: E501
Jee Jee Li's avatar
Jee Jee Li committed
323
    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),  # noqa: E501
324
325
326
    "GraniteSpeechForConditionalGeneration": (
        "granite_speech",
        "GraniteSpeechForConditionalGeneration",
327
    ),
328
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
329
330
331
332
    "HunYuanVLForConditionalGeneration": (
        "hunyuan_vision",
        "HunYuanVLForConditionalGeneration",
    ),
ltd0924's avatar
ltd0924 committed
333
    "StepVLForConditionalGeneration": ("step_vl", "StepVLForConditionalGeneration"),
334
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
335
    "NemotronH_Nano_VL_V2": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
Zero's avatar
Zero committed
336
337
338
339
    "OpenCUAForConditionalGeneration": (
        "opencua",
        "OpenCUAForConditionalGeneration",
    ),
340
341
342
    "InternS1ForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
343
    ),
344
345
346
    "InternVLForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
347
    ),
348
349
350
351
    "Idefics3ForConditionalGeneration": (
        "idefics3",
        "Idefics3ForConditionalGeneration",
    ),
oscardev256's avatar
oscardev256 committed
352
    "IsaacForConditionalGeneration": ("isaac", "IsaacForConditionalGeneration"),
353
    "SmolVLMForConditionalGeneration": ("smolvlm", "SmolVLMForConditionalGeneration"),  # noqa: E501
354
    "KananaVForConditionalGeneration": ("kanana_v", "KananaVForConditionalGeneration"),
355
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
356
357
358
    "KeyeVL1_5ForConditionalGeneration": (
        "keye_vl1_5",
        "KeyeVL1_5ForConditionalGeneration",
359
    ),
360
    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
361
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
362
363
364
365
    "LightOnOCRForConditionalGeneration": (
        "lightonocr",
        "LightOnOCRForConditionalGeneration",
    ),
366
    "Lfm2VlForConditionalGeneration": ("lfm2_vl", "Lfm2VLForConditionalGeneration"),
367
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
368
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),  # noqa: E501
369
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
370
371
372
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
373
    ),
374
375
376
    "LlavaNextVideoForConditionalGeneration": (
        "llava_next_video",
        "LlavaNextVideoForConditionalGeneration",
377
    ),
378
379
380
    "LlavaOnevisionForConditionalGeneration": (
        "llava_onevision",
        "LlavaOnevisionForConditionalGeneration",
381
    ),
382
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
383
    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
384
385
386
    "MiniMaxVL01ForConditionalGeneration": (
        "minimax_vl_01",
        "MiniMaxVL01ForConditionalGeneration",
387
    ),
388
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
389
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
390
391
392
    "Mistral3ForConditionalGeneration": (
        "mistral3",
        "Mistral3ForConditionalGeneration",
393
    ),
394
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
395
    "Molmo2ForConditionalGeneration": ("molmo2", "Molmo2ForConditionalGeneration"),
396
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
397
    "Ovis": ("ovis", "Ovis"),
398
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
399
400
401
402
    "PaddleOCRVLForConditionalGeneration": (
        "paddleocr_vl",
        "PaddleOCRVLForConditionalGeneration",
    ),
403
404
405
406
    "PaliGemmaForConditionalGeneration": (
        "paligemma",
        "PaliGemmaForConditionalGeneration",
    ),
407
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
408
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
409
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
410
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),  # noqa: E501
411
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
412
413
414
    "Qwen2_5_VLForConditionalGeneration": (
        "qwen2_5_vl",
        "Qwen2_5_VLForConditionalGeneration",
415
    ),
416
417
418
    "Qwen2AudioForConditionalGeneration": (
        "qwen2_audio",
        "Qwen2AudioForConditionalGeneration",
419
    ),
420
421
422
    "Qwen2_5OmniModel": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
423
    ),
424
425
426
    "Qwen2_5OmniForConditionalGeneration": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
427
    ),
428
429
430
431
    "Qwen3OmniMoeForConditionalGeneration": (
        "qwen3_omni_moe_thinker",
        "Qwen3OmniMoeThinkerForConditionalGeneration",
    ),
432
    "Qwen3VLForConditionalGeneration": ("qwen3_vl", "Qwen3VLForConditionalGeneration"),  # noqa: E501
433
434
435
    "Qwen3VLMoeForConditionalGeneration": (
        "qwen3_vl_moe",
        "Qwen3VLMoeForConditionalGeneration",
436
    ),
437
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
Song's avatar
Song committed
438
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),  # noqa: E501
汪志鹏's avatar
汪志鹏 committed
439
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),  # noqa: E501
440
441
442
    "Tarsier2ForConditionalGeneration": (
        "qwen2_vl",
        "Tarsier2ForConditionalGeneration",
443
    ),
444
    "UltravoxModel": ("ultravox", "UltravoxModel"),
Patrick von Platen's avatar
Patrick von Platen committed
445
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),  # noqa: E501
Patrick von Platen's avatar
Patrick von Platen committed
446
    "VoxtralStreamingGeneration": ("voxtral_streaming", "VoxtralStreamingGeneration"),  # noqa: E501
447
    # [Encoder-decoder]
448
449
450
451
    "NemotronParseForConditionalGeneration": (
        "nemotron_parse",
        "NemotronParseForConditionalGeneration",
    ),
452
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
453
}
454
455

_SPECULATIVE_DECODING_MODELS = {
456
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
457
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
458
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
459
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
460
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
461
    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
462
    "Eagle3Qwen2_5vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
463
    "Eagle3Qwen3vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
464
465
466
467
    "EagleMistralLarge3ForCausalLM": (
        "mistral_large_3_eagle",
        "EagleMistralLarge3ForCausalLM",
    ),
468
    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
469
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
470
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
Kyungmin Lee's avatar
Kyungmin Lee committed
471
    "ExaoneMoeMTP": ("exaone_moe_mtp", "ExaoneMoeMTP"),
XuruiYang's avatar
XuruiYang committed
472
    "LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
473
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
474
    "Glm4MoeLiteMTPModel": ("glm4_moe_lite_mtp", "Glm4MoeLiteMTP"),
475
    "MedusaModel": ("medusa", "Medusa"),
476
    "OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
477
    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
478
479
480
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
481
}
482

483
_TRANSFORMERS_SUPPORTED_MODELS = {
484
485
486
    # Text generation models
    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
    # Multimodal models
487
488
489
490
    "Emu3ForConditionalGeneration": (
        "transformers",
        "TransformersMultiModalForCausalLM",
    ),
491
492
493
}

_TRANSFORMERS_BACKEND_MODELS = {
494
    # Text generation models
495
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
    "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
514
    "TransformersForSequenceClassification": (
515
        "transformers",
516
        "TransformersForSequenceClassification",
517
    ),
518
    "TransformersMoEForSequenceClassification": (
519
        "transformers",
520
        "TransformersMoEForSequenceClassification",
521
    ),
522
523
524
    "TransformersMultiModalForSequenceClassification": (
        "transformers",
        "TransformersMultiModalForSequenceClassification",
525
    ),
526
}
527

528
_VLLM_MODELS = {
529
    **_TEXT_GENERATION_MODELS,
530
    **_EMBEDDING_MODELS,
531
    **_CROSS_ENCODER_MODELS,
532
    **_MULTIMODAL_MODELS,
533
    **_SPECULATIVE_DECODING_MODELS,
534
535
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
536
537
}

538
539
540
541
# 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.
542
_SUBPROCESS_COMMAND = [sys.executable, "-m", "vllm.model_executor.models.registry"]
543

544
_PREVIOUSLY_SUPPORTED_MODELS = {
545
    "MotifForCausalLM": "0.10.2",
546
    "Phi3SmallForCausalLM": "0.9.2",
547
    "Phi4FlashForCausalLM": "0.10.2",
548
    "Phi4MultimodalForCausalLM": "0.12.0",
549
550
551
552
553
554
555
556
557
    # 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",
}
558

559

560
561
@dataclass(frozen=True)
class _ModelInfo:
562
    architecture: str
563
    is_text_generation_model: bool
564
    is_pooling_model: bool
565
    attn_type: AttnTypeStr
566
567
    default_seq_pooling_type: SequencePoolingType
    default_tok_pooling_type: TokenPoolingType
568
    supports_cross_encoding: bool
569
    supports_multimodal: bool
570
    supports_multimodal_raw_input_only: bool
Patrick von Platen's avatar
Patrick von Platen committed
571
    requires_raw_input_tokens: bool
572
    supports_multimodal_encoder_tp_data: bool
573
    supports_pp: bool
574
575
    has_inner_state: bool
    is_attention_free: bool
576
    is_hybrid: bool
577
    has_noops: bool
578
    supports_mamba_prefix_caching: bool
579
    supports_transcription: bool
580
    supports_transcription_only: bool
581
582

    @staticmethod
583
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
584
        return _ModelInfo(
585
            architecture=model.__name__,
586
            is_text_generation_model=is_text_generation_model(model),
587
            is_pooling_model=is_pooling_model(model),
588
589
            default_seq_pooling_type=get_default_seq_pooling_type(model),
            default_tok_pooling_type=get_default_tok_pooling_type(model),
590
            attn_type=get_attn_type(model),
591
            supports_cross_encoding=supports_cross_encoding(model),
592
            supports_multimodal=supports_multimodal(model),
593
594
595
            supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
                model
            ),
Patrick von Platen's avatar
Patrick von Platen committed
596
            requires_raw_input_tokens=requires_raw_input_tokens(model),
597
598
599
            supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
                model
            ),
600
            supports_pp=supports_pp(model),
601
602
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
603
            is_hybrid=is_hybrid(model),
604
            supports_mamba_prefix_caching=supports_mamba_prefix_caching(model),
605
            supports_transcription=supports_transcription(model),
606
607
608
            supports_transcription_only=(
                supports_transcription(model) and model.supports_transcription_only
            ),
609
            has_noops=has_noops(model),
610
        )
611
612


613
614
615
616
class _BaseRegisteredModel(ABC):
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
617

618
    @abstractmethod
619
    def load_model_cls(self) -> type[nn.Module]:
620
        raise NotImplementedError
621
622


623
624
625
626
627
628
629
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
630
    model_cls: type[nn.Module]
631
632

    @staticmethod
633
    def from_model_cls(model_cls: type[nn.Module]):
634
635
636
637
638
639
640
641
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

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

642
    def load_model_cls(self) -> type[nn.Module]:
643
644
645
646
647
648
649
650
        return self.model_cls


@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """
651

652
653
654
    module_name: str
    class_name: str

655
656
657
658
659
660
661
662
    @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"

663
    def _load_modelinfo_from_cache(self, module_hash: str) -> _ModelInfo | None:
664
665
        try:
            try:
666
                modelinfo_path = self._get_cache_dir() / self._get_cache_filename()
667
668
669
                with open(modelinfo_path, encoding="utf-8") as file:
                    mi_dict = json.load(file)
            except FileNotFoundError:
670
                logger.debug(
671
                    "Cached model info file for class %s.%s not found",
672
673
674
                    self.module_name,
                    self.class_name,
                )
675
676
677
                return None

            if mi_dict["hash"] != module_hash:
678
                logger.debug(
679
                    "Cached model info file for class %s.%s is stale",
680
681
682
                    self.module_name,
                    self.class_name,
                )
683
684
685
686
687
                return None

            # file not changed, use cached _ModelInfo properties
            return _ModelInfo(**mi_dict["modelinfo"])
        except Exception:
688
            logger.debug(
689
                "Cached model info for class %s.%s error. ",
690
691
692
                self.module_name,
                self.class_name,
            )
693
694
            return None

695
    def _save_modelinfo_to_cache(self, mi: _ModelInfo, module_hash: str) -> None:
696
697
        """save dictionary json file to cache"""
        from vllm.model_executor.model_loader.weight_utils import atomic_writer
698

699
700
701
702
703
704
705
706
        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()
707
            with atomic_writer(modelinfo_path, encoding="utf-8") as f:
708
709
710
711
712
                json.dump(modelinfo_dict, f, indent=2)
        except Exception:
            logger.exception("Error saving model info cache.")

    @logtime(logger=logger, msg="Registry inspect model class")
713
    def inspect_model_cls(self) -> _ModelInfo:
714
        model_path = Path(__file__).parent / f"{self.module_name.split('.')[-1]}.py"
715
        module_hash = None
716

717
718
        if model_path.exists():
            with open(model_path, "rb") as f:
719
                module_hash = safe_hash(f.read(), usedforsecurity=False).hexdigest()
720
721
722

            mi = self._load_modelinfo_from_cache(module_hash)
            if mi is not None:
723
                logger.debug(
724
                    "Loaded model info for class %s.%s from cache",
725
726
727
                    self.module_name,
                    self.class_name,
                )
728
729
                return mi
            else:
730
                logger.debug(
731
                    "Cache model info for class %s.%s miss. Loading model instead.",
732
733
734
                    self.module_name,
                    self.class_name,
                )
735
736
737

        # Performed in another process to avoid initializing CUDA
        mi = _run_in_subprocess(
738
739
740
741
742
            lambda: _ModelInfo.from_model_cls(self.load_model_cls())
        )
        logger.debug(
            "Loaded model info for class %s.%s", self.module_name, self.class_name
        )
743
744

        # save cache file
745
746
        if module_hash is not None:
            self._save_modelinfo_to_cache(mi, module_hash)
747
748

        return mi
749

750
    def load_model_cls(self) -> type[nn.Module]:
751
752
753
754
755
756
757
758
        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,
759
) -> type[nn.Module] | None:
760
    from vllm.platforms import current_platform
761

762
    current_platform.verify_model_arch(model_arch)
763
764
765
    try:
        return model.load_model_cls()
    except Exception:
766
        logger.exception("Error in loading model architecture '%s'", model_arch)
767
        return None
768
769


770
771
772
773
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
774
) -> _ModelInfo | None:
775
776
777
    try:
        return model.inspect_model_cls()
    except Exception:
778
        logger.exception("Error in inspecting model architecture '%s'", model_arch)
779
        return None
780
781


782
783
784
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
785
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
786

787
    def get_supported_archs(self) -> Set[str]:
788
        return self.models.keys()
789

790
791
792
    def register_model(
        self,
        model_arch: str,
793
        model_cls: type[nn.Module] | str,
794
    ) -> None:
795
796
797
        """
        Register an external model to be used in vLLM.

798
        `model_cls` can be either:
799

800
        - A [`torch.nn.Module`][] class directly referencing the model.
801
        - A string in the format `<module>:<class>` which can be used to
802
803
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
804
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
805
        """
806
807
808
809
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

810
        if model_arch in self.models:
811
812
            logger.warning(
                "Model architecture %s is already registered, and will be "
813
814
815
816
                "overwritten by the new model class %s.",
                model_arch,
                model_cls,
            )
817
818
819
820
821
822

        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)
823

824
            model = _LazyRegisteredModel(*split_str)
825
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
826
            model = _RegisteredModel.from_model_cls(model_cls)
827
        else:
828
829
830
831
            msg = (
                "`model_cls` should be a string or PyTorch model class, "
                f"not a {type(model_arch)}"
            )
832
            raise TypeError(msg)
833

834
        self.models[model_arch] = model
835

836
    def _raise_for_unsupported(self, architectures: list[str]):
837
        all_supported_archs = self.get_supported_archs()
838

839
840
841
        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
842
843
                "to be inspected. Please check the logs for more details."
            )
844

845
846
847
848
849
850
851
852
        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 "
853
854
                    "use this model architecture."
                )
855

856
857
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
858
859
            f"Supported architectures: {all_supported_archs}"
        )
860

861
    def _try_load_model_cls(self, model_arch: str) -> type[nn.Module] | None:
862
863
        if model_arch not in self.models:
            return None
864

865
        return _try_load_model_cls(model_arch, self.models[model_arch])
866

867
    def _try_inspect_model_cls(self, model_arch: str) -> _ModelInfo | None:
868
869
        if model_arch not in self.models:
            return None
870

871
872
873
874
875
876
        return _try_inspect_model_cls(model_arch, self.models[model_arch])

    def _try_resolve_transformers(
        self,
        architecture: str,
        model_config: ModelConfig,
877
    ) -> str | None:
878
879
880
        if architecture in _TRANSFORMERS_BACKEND_MODELS:
            return architecture

881
882
883
        auto_map: dict[str, str] = (
            getattr(model_config.hf_config, "auto_map", None) or dict()
        )
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899

        # 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,
900
                        trust_remote_code=model_config.trust_remote_code,
901
902
903
904
905
906
907
908
909
910
911
912
                        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,
913
                        trust_remote_code=model_config.trust_remote_code,
914
915
916
917
918
                        warn_on_fail=True,
                    )
                    if model_module is not None:
                        break
            else:
919
                if model_config.model_impl != "transformers":
920
921
922
923
924
925
926
                    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 "
927
928
                    "'auto_map' (relevant if the model is custom)."
                )
929
930

        if not model_module.is_backend_compatible():
931
            if model_config.model_impl != "transformers":
932
                return None
933

934
935
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
936
937
                "is not compatible with vLLM."
            )
938

939
        return model_config._get_transformers_backend_cls()
940

941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
    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
966

967
968
    def inspect_model_cls(
        self,
969
        architectures: str | list[str],
970
        model_config: ModelConfig,
971
    ) -> tuple[_ModelInfo, str]:
972
973
        if isinstance(architectures, str):
            architectures = [architectures]
974
975
        if not architectures:
            raise ValueError("No model architectures are specified")
976
977

        # Require transformers impl
978
        if model_config.model_impl == "transformers":
979
            arch = self._try_resolve_transformers(architectures[0], model_config)
980
981
982
983
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)
984
        elif model_config.model_impl == "terratorch":
985
986
            model_info = self._try_inspect_model_cls("Terratorch")
            return (model_info, "Terratorch")
987

988
        # Fallback to transformers impl (after resolving convert_type)
989
990
991
992
993
994
        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)
995
996
997
998
999
1000
1001
            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)
1002
            model_info = self._try_inspect_model_cls(normalized_arch)
1003
            if model_info is not None:
1004
                return (model_info, arch)
1005

1006
        # Fallback to transformers impl (before resolving runner_type)
1007
1008
1009
1010
1011
        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)
1012
1013
1014
1015
1016
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

1017
        return self._raise_for_unsupported(architectures)
1018

1019
1020
    def resolve_model_cls(
        self,
1021
        architectures: str | list[str],
1022
        model_config: ModelConfig,
1023
    ) -> tuple[type[nn.Module], str]:
1024
1025
        if isinstance(architectures, str):
            architectures = [architectures]
1026
1027
        if not architectures:
            raise ValueError("No model architectures are specified")
1028
1029

        # Require transformers impl
1030
        if model_config.model_impl == "transformers":
1031
            arch = self._try_resolve_transformers(architectures[0], model_config)
1032
1033
1034
1035
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)
1036
        elif model_config.model_impl == "terratorch":
1037
1038
1039
1040
            arch = "Terratorch"
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
1041

1042
        # Fallback to transformers impl (after resolving convert_type)
1043
1044
1045
1046
1047
1048
        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)
1049
1050
1051
1052
1053
1054
1055
            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)
1056
            model_cls = self._try_load_model_cls(normalized_arch)
1057
1058
            if model_cls is not None:
                return (model_cls, arch)
1059

1060
        # Fallback to transformers impl (before resolving runner_type)
1061
1062
1063
1064
1065
        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)
1066
1067
1068
1069
1070
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

1071
        return self._raise_for_unsupported(architectures)
1072

1073
1074
    def is_text_generation_model(
        self,
1075
        architectures: str | list[str],
1076
        model_config: ModelConfig,
1077
    ) -> bool:
1078
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1079
        return model_cls.is_text_generation_model
1080

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

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

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

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

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

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

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

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

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

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

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

1169

1170
1171
1172
1173
1174
1175
1176
1177
1178
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()
    }
)
1179
1180
1181
1182
1183

_T = TypeVar("_T")


def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
1184
1185
1186
1187
1188
    # 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")

1189
        # `cloudpickle` allows pickling lambda functions directly
1190
        import cloudpickle
1191

1192
        input_bytes = cloudpickle.dumps((fn, output_filepath))
1193
1194
1195

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
1196
1197
1198
        returned = subprocess.run(
            _SUBPROCESS_COMMAND, input=input_bytes, capture_output=True
        )
1199
1200
1201
1202
1203
1204

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
1205
1206
1207
            raise RuntimeError(
                f"Error raised in subprocess:\n{returned.stderr.decode()}"
            ) from e
1208

1209
        with open(output_filepath, "rb") as f:
1210
1211
1212
1213
1214
1215
            return pickle.load(f)


def _run() -> None:
    # Setup plugins
    from vllm.plugins import load_general_plugins
1216

1217
1218
1219
1220
1221
    load_general_plugins()

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

    result = fn()
1222
1223
1224

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
1225
1226
1227


if __name__ == "__main__":
1228
    _run()