registry.py 46.7 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
38
39
40
41
42
43
if TYPE_CHECKING:
    from vllm.config.model import AttnTypeStr
    from vllm.config.pooler import PoolingTypeStr
else:
    AttnTypeStr = Any
    PoolingTypeStr = Any


44
45
46
47
48
from .interfaces import (
    has_inner_state,
    has_noops,
    is_attention_free,
    is_hybrid,
Patrick von Platen's avatar
Patrick von Platen committed
49
    requires_raw_input_tokens,
50
    supports_cross_encoding,
51
    supports_mamba_prefix_caching,
52
53
54
55
56
57
58
    supports_multimodal,
    supports_multimodal_encoder_tp_data,
    supports_multimodal_raw_input_only,
    supports_pp,
    supports_transcription,
)
from .interfaces_base import (
59
    get_attn_type,
60
61
62
63
    get_default_pooling_type,
    is_pooling_model,
    is_text_generation_model,
)
64
65
66

logger = init_logger(__name__)

67
68
_TEXT_GENERATION_MODELS = {
    # [Decoder-only]
69
    "AfmoeForCausalLM": ("afmoe", "AfmoeForCausalLM"),
70
    "ApertusForCausalLM": ("apertus", "ApertusForCausalLM"),
71
72
    "AquilaModel": ("llama", "LlamaForCausalLM"),
    "AquilaForCausalLM": ("llama", "LlamaForCausalLM"),  # AquilaChat2
Raghav Ravishankar's avatar
Raghav Ravishankar committed
73
    "ArceeForCausalLM": ("arcee", "ArceeForCausalLM"),
74
    "ArcticForCausalLM": ("arctic", "ArcticForCausalLM"),
75
76
77
78
    # baichuan-7b, upper case 'C' in the class name
    "BaiChuanForCausalLM": ("baichuan", "BaiChuanForCausalLM"),
    # baichuan-13b, lower case 'c' in the class name
    "BaichuanForCausalLM": ("baichuan", "BaichuanForCausalLM"),
79
    "BailingMoeForCausalLM": ("bailing_moe", "BailingMoeForCausalLM"),
ant-yy's avatar
ant-yy committed
80
    "BailingMoeV2ForCausalLM": ("bailing_moe", "BailingMoeV2ForCausalLM"),
Yu Chin Fabian Lim's avatar
Yu Chin Fabian Lim committed
81
    "BambaForCausalLM": ("bamba", "BambaForCausalLM"),
82
    "BloomForCausalLM": ("bloom", "BloomForCausalLM"),
83
    "ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
84
    "ChatGLMForConditionalGeneration": ("chatglm", "ChatGLMForCausalLM"),
85
    "CohereForCausalLM": ("commandr", "CohereForCausalLM"),
86
    "Cohere2ForCausalLM": ("commandr", "CohereForCausalLM"),
87
    "CwmForCausalLM": ("llama", "LlamaForCausalLM"),
88
    "DbrxForCausalLM": ("dbrx", "DbrxForCausalLM"),
89
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
90
    "DeepseekForCausalLM": ("deepseek_v2", "DeepseekForCausalLM"),
91
    "DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
92
    "DeepseekV3ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
93
    "DeepseekV32ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
94
    "Dots1ForCausalLM": ("dots1", "Dots1ForCausalLM"),
95
    "Ernie4_5ForCausalLM": ("ernie45", "Ernie4_5ForCausalLM"),
96
    "Ernie4_5_MoeForCausalLM": ("ernie45_moe", "Ernie4_5_MoeForCausalLM"),
97
    "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
98
    "Exaone4ForCausalLM": ("exaone4", "Exaone4ForCausalLM"),
99
    "Fairseq2LlamaForCausalLM": ("fairseq2_llama", "Fairseq2LlamaForCausalLM"),
100
101
102
    "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
    "FalconH1ForCausalLM": ("falcon_h1", "FalconH1ForCausalLM"),
103
    "FlexOlmoForCausalLM": ("flex_olmo", "FlexOlmoForCausalLM"),
104
105
    "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
    "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
106
    "Gemma3ForCausalLM": ("gemma3", "Gemma3ForCausalLM"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
107
    "Gemma3nForCausalLM": ("gemma3n", "Gemma3nForCausalLM"),
108
    "Qwen3NextForCausalLM": ("qwen3_next", "Qwen3NextForCausalLM"),
109
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
110
    "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
111
    "Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
112
    "GptOssForCausalLM": ("gpt_oss", "GptOssForCausalLM"),
113
114
115
116
117
118
    "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
    "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
    "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
    "GPTNeoXForCausalLM": ("gpt_neox", "GPTNeoXForCausalLM"),
    "GraniteForCausalLM": ("granite", "GraniteForCausalLM"),
    "GraniteMoeForCausalLM": ("granitemoe", "GraniteMoeForCausalLM"),
119
120
    "GraniteMoeHybridForCausalLM": ("granitemoehybrid", "GraniteMoeHybridForCausalLM"),  # noqa: E501
    "GraniteMoeSharedForCausalLM": ("granitemoeshared", "GraniteMoeSharedForCausalLM"),  # noqa: E501
121
    "GritLM": ("gritlm", "GritLM"),
Bijaya Dangol's avatar
Bijaya Dangol committed
122
123
    "Grok1ModelForCausalLM": ("grok1", "GrokForCausalLM"),
    "Grok1ForCausalLM": ("grok1", "GrokForCausalLM"),
124
125
    "HunYuanMoEV1ForCausalLM": ("hunyuan_v1", "HunYuanMoEV1ForCausalLM"),
    "HunYuanDenseV1ForCausalLM": ("hunyuan_v1", "HunYuanDenseV1ForCausalLM"),
126
    "HCXVisionForCausalLM": ("hyperclovax_vision", "HCXVisionForCausalLM"),
127
128
    "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
    "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"),
129
    "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
130
    "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
131
132
    "IQuestCoderForCausalLM": ("llama", "LlamaForCausalLM"),
    "IQuestLoopCoderForCausalLM": ("iquest_loopcoder", "IQuestLoopCoderForCausalLM"),
133
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
134
    "Jais2ForCausalLM": ("jais2", "Jais2ForCausalLM"),
135
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
136
    "KimiLinearForCausalLM": ("kimi_linear", "KimiLinearForCausalLM"),  # noqa: E501
137
    "Lfm2ForCausalLM": ("lfm2", "Lfm2ForCausalLM"),
Paul Pak's avatar
Paul Pak committed
138
    "Lfm2MoeForCausalLM": ("lfm2_moe", "Lfm2MoeForCausalLM"),
139
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
140
    "Llama4ForCausalLM": ("llama4", "Llama4ForCausalLM"),
141
142
    # For decapoda-research/llama-*
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
XuruiYang's avatar
XuruiYang committed
143
    "LongcatFlashForCausalLM": ("longcat_flash", "LongcatFlashForCausalLM"),
144
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
145
    "Mamba2ForCausalLM": ("mamba2", "Mamba2ForCausalLM"),
146
147
    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
    "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
148
149
150
    "MiniMaxForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxText01ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxM1ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
151
    "MiniMaxM2ForCausalLM": ("minimax_m2", "MiniMaxM2ForCausalLM"),
152
    "MistralForCausalLM": ("llama", "LlamaForCausalLM"),
153
    "MistralLarge3ForCausalLM": ("mistral_large_3", "MistralLarge3ForCausalLM"),
154
155
156
157
    "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
    # transformers's mpt class has lower case
    "MptForCausalLM": ("mpt", "MPTForCausalLM"),
    "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
158
    "MiMoForCausalLM": ("mimo", "MiMoForCausalLM"),
159
    "MiMoV2FlashForCausalLM": ("mimo_v2_flash", "MiMoV2FlashForCausalLM"),
160
    "NemotronForCausalLM": ("nemotron", "NemotronForCausalLM"),
Luis Vega's avatar
Luis Vega committed
161
    "NemotronHForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
162
    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
163
    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
164
    "Olmo3ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
165
166
167
    "OlmoeForCausalLM": ("olmoe", "OlmoeForCausalLM"),
    "OPTForCausalLM": ("opt", "OPTForCausalLM"),
    "OrionForCausalLM": ("orion", "OrionForCausalLM"),
168
    "OuroForCausalLM": ("ouro", "OuroForCausalLM"),
169
    "PanguEmbeddedForCausalLM": ("openpangu", "PanguEmbeddedForCausalLM"),
170
    "PanguProMoEV2ForCausalLM": ("openpangu", "PanguProMoEV2ForCausalLM"),
171
    "PanguUltraMoEForCausalLM": ("openpangu", "PanguUltraMoEForCausalLM"),
172
173
174
175
    "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
    "PhiForCausalLM": ("phi", "PhiForCausalLM"),
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
    "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
176
    "Plamo2ForCausalLM": ("plamo2", "Plamo2ForCausalLM"),
177
    "Plamo3ForCausalLM": ("plamo3", "Plamo3ForCausalLM"),
178
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
179
180
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
181
182
    "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),
    "Qwen3MoeForCausalLM": ("qwen3_moe", "Qwen3MoeForCausalLM"),
183
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
184
    "SeedOssForCausalLM": ("seed_oss", "SeedOssForCausalLM"),
Song's avatar
Song committed
185
    "Step3TextForCausalLM": ("step3_text", "Step3TextForCausalLM"),
186
187
188
189
    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
190
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
191
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
192
    "TeleFLMForCausalLM": ("teleflm", "TeleFLMForCausalLM"),
193
    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
194
    "Zamba2ForCausalLM": ("zamba2", "Zamba2ForCausalLM"),
195
196
197
}

_EMBEDDING_MODELS = {
198
    # [Text-only]
199
    "BertModel": ("bert", "BertEmbeddingModel"),
200
    "BertSpladeSparseEmbeddingModel": ("bert", "BertSpladeSparseEmbeddingModel"),
201
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
202
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
203
    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
204
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
205
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
206
    "GritLM": ("gritlm", "GritLM"),
207
208
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
209
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
210
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
211
    "LlamaBidirectionalModel": ("llama", "LlamaBidirectionalModel"),
212
    "LlamaModel": ("llama", "LlamaForCausalLM"),
213
214
    **{
        # Multiple models share the same architecture, so we include them all
215
216
        k: (mod, arch)
        for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
217
218
        if arch == "LlamaForCausalLM"
    },
219
    "MistralModel": ("llama", "LlamaForCausalLM"),
220
    "ModernBertModel": ("modernbert", "ModernBertModel"),
221
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
222
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
223
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
224
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
225
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
226
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
227
228
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
229
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
230
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
231
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
232
    # [Multimodal]
233
    "CLIPModel": ("clip", "CLIPEmbeddingModel"),
234
235
236
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
237
    ),
238
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
239
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
240
    "SiglipModel": ("siglip", "SiglipEmbeddingModel"),
241
242
    # Technically Terratorch models work on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
243
    # models for the time being.
244
245
    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
    "Terratorch": ("terratorch", "Terratorch"),
246
247
}

248
249
_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
250
    "BertForTokenClassification": ("bert", "BertForTokenClassification"),
251
252
253
254
    "GteNewForSequenceClassification": (
        "bert_with_rope",
        "GteNewForSequenceClassification",
    ),
255
256
257
258
259
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"),
    "LlamaBidirectionalForSequenceClassification": (
        "llama",
        "LlamaBidirectionalForSequenceClassification",
    ),
260
261
262
263
    "ModernBertForSequenceClassification": (
        "modernbert",
        "ModernBertForSequenceClassification",
    ),
264
265
266
267
    "ModernBertForTokenClassification": (
        "modernbert",
        "ModernBertForTokenClassification",
    ),
268
269
270
271
272
    "RobertaForSequenceClassification": ("roberta", "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": (
        "roberta",
        "RobertaForSequenceClassification",
    ),
273
274
}

275
_MULTIMODAL_MODELS = {
276
    # [Decoder-only]
277
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
278
279
280
281
    "AudioFlamingo3ForConditionalGeneration": (
        "audioflamingo3",
        "AudioFlamingo3ForConditionalGeneration",
    ),
282
283
284
    "AyaVisionForConditionalGeneration": (
        "aya_vision",
        "AyaVisionForConditionalGeneration",
285
    ),
286
    "BagelForConditionalGeneration": ("bagel", "BagelForConditionalGeneration"),
287
    "BeeForConditionalGeneration": ("bee", "BeeForConditionalGeneration"),
288
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
289
290
291
    "ChameleonForConditionalGeneration": (
        "chameleon",
        "ChameleonForConditionalGeneration",
292
    ),
293
294
295
    "Cohere2VisionForConditionalGeneration": (
        "cohere2_vision",
        "Cohere2VisionForConditionalGeneration",
296
    ),
297
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
298
    "DeepseekOCRForCausalLM": ("deepseek_ocr", "DeepseekOCRForCausalLM"),
Roger Wang's avatar
Roger Wang committed
299
    "DotsOCRForCausalLM": ("dots_ocr", "DotsOCRForCausalLM"),
300
301
302
    "Ernie4_5_VLMoeForConditionalGeneration": (
        "ernie45_vl",
        "Ernie4_5_VLMoeForConditionalGeneration",
303
    ),
304
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
305
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
306
307
308
    "Gemma3nForConditionalGeneration": (
        "gemma3n_mm",
        "Gemma3nForConditionalGeneration",
309
    ),
310
    "GlmAsrForConditionalGeneration": ("glmasr", "GlmAsrForConditionalGeneration"),
311
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
312
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),  # noqa: E501
Jee Jee Li's avatar
Jee Jee Li committed
313
    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),  # noqa: E501
314
315
316
    "GraniteSpeechForConditionalGeneration": (
        "granite_speech",
        "GraniteSpeechForConditionalGeneration",
317
    ),
318
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
319
320
321
322
    "HunYuanVLForConditionalGeneration": (
        "hunyuan_vision",
        "HunYuanVLForConditionalGeneration",
    ),
323
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
324
    "NemotronH_Nano_VL_V2": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
Zero's avatar
Zero committed
325
326
327
328
    "OpenCUAForConditionalGeneration": (
        "opencua",
        "OpenCUAForConditionalGeneration",
    ),
329
330
331
    "InternS1ForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
332
    ),
333
334
335
    "InternVLForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
336
    ),
337
338
339
340
    "Idefics3ForConditionalGeneration": (
        "idefics3",
        "Idefics3ForConditionalGeneration",
    ),
oscardev256's avatar
oscardev256 committed
341
    "IsaacForConditionalGeneration": ("isaac", "IsaacForConditionalGeneration"),
342
    "SmolVLMForConditionalGeneration": ("smolvlm", "SmolVLMForConditionalGeneration"),  # noqa: E501
343
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
344
345
346
    "KeyeVL1_5ForConditionalGeneration": (
        "keye_vl1_5",
        "KeyeVL1_5ForConditionalGeneration",
347
    ),
348
    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
349
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
350
351
352
353
    "LightOnOCRForConditionalGeneration": (
        "lightonocr",
        "LightOnOCRForConditionalGeneration",
    ),
354
    "Lfm2VlForConditionalGeneration": ("lfm2_vl", "Lfm2VLForConditionalGeneration"),
355
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
356
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),  # noqa: E501
357
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
358
359
360
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
361
    ),
362
363
364
    "LlavaNextVideoForConditionalGeneration": (
        "llava_next_video",
        "LlavaNextVideoForConditionalGeneration",
365
    ),
366
367
368
    "LlavaOnevisionForConditionalGeneration": (
        "llava_onevision",
        "LlavaOnevisionForConditionalGeneration",
369
    ),
370
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
371
    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
372
373
374
    "MiniMaxVL01ForConditionalGeneration": (
        "minimax_vl_01",
        "MiniMaxVL01ForConditionalGeneration",
375
    ),
376
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
377
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
378
379
380
    "Mistral3ForConditionalGeneration": (
        "mistral3",
        "Mistral3ForConditionalGeneration",
381
    ),
382
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
383
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
384
    "Ovis": ("ovis", "Ovis"),
385
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
386
387
388
389
    "PaddleOCRVLForConditionalGeneration": (
        "paddleocr_vl",
        "PaddleOCRVLForConditionalGeneration",
    ),
390
391
392
393
    "PaliGemmaForConditionalGeneration": (
        "paligemma",
        "PaliGemmaForConditionalGeneration",
    ),
394
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
395
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
396
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
397
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),  # noqa: E501
398
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
399
400
401
    "Qwen2_5_VLForConditionalGeneration": (
        "qwen2_5_vl",
        "Qwen2_5_VLForConditionalGeneration",
402
    ),
403
404
405
    "Qwen2AudioForConditionalGeneration": (
        "qwen2_audio",
        "Qwen2AudioForConditionalGeneration",
406
    ),
407
408
409
    "Qwen2_5OmniModel": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
410
    ),
411
412
413
    "Qwen2_5OmniForConditionalGeneration": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
414
    ),
415
416
417
418
    "Qwen3OmniMoeForConditionalGeneration": (
        "qwen3_omni_moe_thinker",
        "Qwen3OmniMoeThinkerForConditionalGeneration",
    ),
419
    "Qwen3VLForConditionalGeneration": ("qwen3_vl", "Qwen3VLForConditionalGeneration"),  # noqa: E501
420
421
422
    "Qwen3VLMoeForConditionalGeneration": (
        "qwen3_vl_moe",
        "Qwen3VLMoeForConditionalGeneration",
423
    ),
424
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
Song's avatar
Song committed
425
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),  # noqa: E501
汪志鹏's avatar
汪志鹏 committed
426
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),  # noqa: E501
427
428
429
    "Tarsier2ForConditionalGeneration": (
        "qwen2_vl",
        "Tarsier2ForConditionalGeneration",
430
    ),
431
    "UltravoxModel": ("ultravox", "UltravoxModel"),
Patrick von Platen's avatar
Patrick von Platen committed
432
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),  # noqa: E501
Patrick von Platen's avatar
Patrick von Platen committed
433
    "VoxtralStreamingGeneration": ("voxtral_streaming", "VoxtralStreamingGeneration"),  # noqa: E501
434
    # [Encoder-decoder]
435
436
437
438
    "NemotronParseForConditionalGeneration": (
        "nemotron_parse",
        "NemotronParseForConditionalGeneration",
    ),
439
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
440
}
441
442

_SPECULATIVE_DECODING_MODELS = {
443
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
444
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
445
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
446
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
447
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
448
    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
449
    "Eagle3Qwen2_5vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
450
    "Eagle3Qwen3vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
451
452
453
454
    "EagleMistralLarge3ForCausalLM": (
        "mistral_large_3_eagle",
        "EagleMistralLarge3ForCausalLM",
    ),
455
    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
456
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
457
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
XuruiYang's avatar
XuruiYang committed
458
    "LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
459
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
460
    "MedusaModel": ("medusa", "Medusa"),
461
    "OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
462
    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
463
464
465
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
466
}
467

468
_TRANSFORMERS_SUPPORTED_MODELS = {
469
470
471
    # Text generation models
    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
    # Multimodal models
472
473
474
475
    "Emu3ForConditionalGeneration": (
        "transformers",
        "TransformersMultiModalForCausalLM",
    ),
476
477
478
}

_TRANSFORMERS_BACKEND_MODELS = {
479
    # Text generation models
480
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
    "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
499
    "TransformersForSequenceClassification": (
500
        "transformers",
501
        "TransformersForSequenceClassification",
502
    ),
503
    "TransformersMoEForSequenceClassification": (
504
        "transformers",
505
        "TransformersMoEForSequenceClassification",
506
    ),
507
508
509
    "TransformersMultiModalForSequenceClassification": (
        "transformers",
        "TransformersMultiModalForSequenceClassification",
510
    ),
511
}
512

513
_VLLM_MODELS = {
514
    **_TEXT_GENERATION_MODELS,
515
    **_EMBEDDING_MODELS,
516
    **_CROSS_ENCODER_MODELS,
517
    **_MULTIMODAL_MODELS,
518
    **_SPECULATIVE_DECODING_MODELS,
519
520
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
521
522
}

523
524
525
526
# 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.
527
_SUBPROCESS_COMMAND = [sys.executable, "-m", "vllm.model_executor.models.registry"]
528

529
_PREVIOUSLY_SUPPORTED_MODELS = {
530
    "MotifForCausalLM": "0.10.2",
531
    "Phi3SmallForCausalLM": "0.9.2",
532
    "Phi4FlashForCausalLM": "0.10.2",
533
    "Phi4MultimodalForCausalLM": "0.12.0",
534
535
536
537
538
539
540
541
542
    # 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",
}
543

544

545
546
@dataclass(frozen=True)
class _ModelInfo:
547
    architecture: str
548
    is_text_generation_model: bool
549
    is_pooling_model: bool
550
551
    attn_type: AttnTypeStr
    default_pooling_type: PoolingTypeStr
552
    supports_cross_encoding: bool
553
    supports_multimodal: bool
554
    supports_multimodal_raw_input_only: bool
Patrick von Platen's avatar
Patrick von Platen committed
555
    requires_raw_input_tokens: bool
556
    supports_multimodal_encoder_tp_data: bool
557
    supports_pp: bool
558
559
    has_inner_state: bool
    is_attention_free: bool
560
    is_hybrid: bool
561
    has_noops: bool
562
    supports_mamba_prefix_caching: bool
563
    supports_transcription: bool
564
    supports_transcription_only: bool
565
566

    @staticmethod
567
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
568
        return _ModelInfo(
569
            architecture=model.__name__,
570
            is_text_generation_model=is_text_generation_model(model),
571
            is_pooling_model=is_pooling_model(model),
572
            default_pooling_type=get_default_pooling_type(model),
573
            attn_type=get_attn_type(model),
574
            supports_cross_encoding=supports_cross_encoding(model),
575
            supports_multimodal=supports_multimodal(model),
576
577
578
            supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
                model
            ),
Patrick von Platen's avatar
Patrick von Platen committed
579
            requires_raw_input_tokens=requires_raw_input_tokens(model),
580
581
582
            supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
                model
            ),
583
            supports_pp=supports_pp(model),
584
585
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
586
            is_hybrid=is_hybrid(model),
587
            supports_mamba_prefix_caching=supports_mamba_prefix_caching(model),
588
            supports_transcription=supports_transcription(model),
589
590
591
            supports_transcription_only=(
                supports_transcription(model) and model.supports_transcription_only
            ),
592
            has_noops=has_noops(model),
593
        )
594
595


596
597
598
599
class _BaseRegisteredModel(ABC):
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
600

601
    @abstractmethod
602
    def load_model_cls(self) -> type[nn.Module]:
603
        raise NotImplementedError
604
605


606
607
608
609
610
611
612
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
613
    model_cls: type[nn.Module]
614
615

    @staticmethod
616
    def from_model_cls(model_cls: type[nn.Module]):
617
618
619
620
621
622
623
624
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

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

625
    def load_model_cls(self) -> type[nn.Module]:
626
627
628
629
630
631
632
633
        return self.model_cls


@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """
634

635
636
637
    module_name: str
    class_name: str

638
639
640
641
642
643
644
645
    @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"

646
    def _load_modelinfo_from_cache(self, module_hash: str) -> _ModelInfo | None:
647
648
        try:
            try:
649
                modelinfo_path = self._get_cache_dir() / self._get_cache_filename()
650
651
652
                with open(modelinfo_path, encoding="utf-8") as file:
                    mi_dict = json.load(file)
            except FileNotFoundError:
653
                logger.debug(
654
                    "Cached model info file for class %s.%s not found",
655
656
657
                    self.module_name,
                    self.class_name,
                )
658
659
660
                return None

            if mi_dict["hash"] != module_hash:
661
                logger.debug(
662
                    "Cached model info file for class %s.%s is stale",
663
664
665
                    self.module_name,
                    self.class_name,
                )
666
667
668
669
670
                return None

            # file not changed, use cached _ModelInfo properties
            return _ModelInfo(**mi_dict["modelinfo"])
        except Exception:
671
            logger.debug(
672
                "Cached model info for class %s.%s error. ",
673
674
675
                self.module_name,
                self.class_name,
            )
676
677
            return None

678
    def _save_modelinfo_to_cache(self, mi: _ModelInfo, module_hash: str) -> None:
679
680
        """save dictionary json file to cache"""
        from vllm.model_executor.model_loader.weight_utils import atomic_writer
681

682
683
684
685
686
687
688
689
        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()
690
            with atomic_writer(modelinfo_path, encoding="utf-8") as f:
691
692
693
694
695
                json.dump(modelinfo_dict, f, indent=2)
        except Exception:
            logger.exception("Error saving model info cache.")

    @logtime(logger=logger, msg="Registry inspect model class")
696
    def inspect_model_cls(self) -> _ModelInfo:
697
        model_path = Path(__file__).parent / f"{self.module_name.split('.')[-1]}.py"
698
        module_hash = None
699

700
701
        if model_path.exists():
            with open(model_path, "rb") as f:
702
                module_hash = safe_hash(f.read(), usedforsecurity=False).hexdigest()
703
704
705

            mi = self._load_modelinfo_from_cache(module_hash)
            if mi is not None:
706
                logger.debug(
707
                    "Loaded model info for class %s.%s from cache",
708
709
710
                    self.module_name,
                    self.class_name,
                )
711
712
                return mi
            else:
713
                logger.debug(
714
                    "Cache model info for class %s.%s miss. Loading model instead.",
715
716
717
                    self.module_name,
                    self.class_name,
                )
718
719
720

        # Performed in another process to avoid initializing CUDA
        mi = _run_in_subprocess(
721
722
723
724
725
            lambda: _ModelInfo.from_model_cls(self.load_model_cls())
        )
        logger.debug(
            "Loaded model info for class %s.%s", self.module_name, self.class_name
        )
726
727

        # save cache file
728
729
        if module_hash is not None:
            self._save_modelinfo_to_cache(mi, module_hash)
730
731

        return mi
732

733
    def load_model_cls(self) -> type[nn.Module]:
734
735
736
737
738
739
740
741
        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,
742
) -> type[nn.Module] | None:
743
    from vllm.platforms import current_platform
744

745
    current_platform.verify_model_arch(model_arch)
746
747
748
    try:
        return model.load_model_cls()
    except Exception:
749
        logger.exception("Error in loading model architecture '%s'", model_arch)
750
        return None
751
752


753
754
755
756
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
757
) -> _ModelInfo | None:
758
759
760
    try:
        return model.inspect_model_cls()
    except Exception:
761
        logger.exception("Error in inspecting model architecture '%s'", model_arch)
762
        return None
763
764


765
766
767
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
768
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
769

770
    def get_supported_archs(self) -> Set[str]:
771
        return self.models.keys()
772

773
774
775
    def register_model(
        self,
        model_arch: str,
776
        model_cls: type[nn.Module] | str,
777
    ) -> None:
778
779
780
        """
        Register an external model to be used in vLLM.

781
        `model_cls` can be either:
782

783
        - A [`torch.nn.Module`][] class directly referencing the model.
784
        - A string in the format `<module>:<class>` which can be used to
785
786
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
787
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
788
        """
789
790
791
792
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

793
        if model_arch in self.models:
794
795
            logger.warning(
                "Model architecture %s is already registered, and will be "
796
797
798
799
                "overwritten by the new model class %s.",
                model_arch,
                model_cls,
            )
800
801
802
803
804
805

        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)
806

807
            model = _LazyRegisteredModel(*split_str)
808
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
809
            model = _RegisteredModel.from_model_cls(model_cls)
810
        else:
811
812
813
814
            msg = (
                "`model_cls` should be a string or PyTorch model class, "
                f"not a {type(model_arch)}"
            )
815
            raise TypeError(msg)
816

817
        self.models[model_arch] = model
818

819
    def _raise_for_unsupported(self, architectures: list[str]):
820
        all_supported_archs = self.get_supported_archs()
821

822
823
824
        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
825
826
                "to be inspected. Please check the logs for more details."
            )
827

828
829
830
831
832
833
834
835
        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 "
836
837
                    "use this model architecture."
                )
838

839
840
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
841
842
            f"Supported architectures: {all_supported_archs}"
        )
843

844
    def _try_load_model_cls(self, model_arch: str) -> type[nn.Module] | None:
845
846
        if model_arch not in self.models:
            return None
847

848
        return _try_load_model_cls(model_arch, self.models[model_arch])
849

850
    def _try_inspect_model_cls(self, model_arch: str) -> _ModelInfo | None:
851
852
        if model_arch not in self.models:
            return None
853

854
855
856
857
858
859
        return _try_inspect_model_cls(model_arch, self.models[model_arch])

    def _try_resolve_transformers(
        self,
        architecture: str,
        model_config: ModelConfig,
860
    ) -> str | None:
861
862
863
        if architecture in _TRANSFORMERS_BACKEND_MODELS:
            return architecture

864
865
866
        auto_map: dict[str, str] = (
            getattr(model_config.hf_config, "auto_map", None) or dict()
        )
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
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,
                        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,
                        warn_on_fail=True,
                    )
                    if model_module is not None:
                        break
            else:
900
                if model_config.model_impl != "transformers":
901
902
903
904
905
906
907
                    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 "
908
909
                    "'auto_map' (relevant if the model is custom)."
                )
910
911

        if not model_module.is_backend_compatible():
912
            if model_config.model_impl != "transformers":
913
                return None
914

915
916
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
917
918
                "is not compatible with vLLM."
            )
919

920
        return model_config._get_transformers_backend_cls()
921

922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
    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
947

948
949
    def inspect_model_cls(
        self,
950
        architectures: str | list[str],
951
        model_config: ModelConfig,
952
    ) -> tuple[_ModelInfo, str]:
953
954
        if isinstance(architectures, str):
            architectures = [architectures]
955
956
        if not architectures:
            raise ValueError("No model architectures are specified")
957
958

        # Require transformers impl
959
        if model_config.model_impl == "transformers":
960
            arch = self._try_resolve_transformers(architectures[0], model_config)
961
962
963
964
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)
965
        elif model_config.model_impl == "terratorch":
966
967
            model_info = self._try_inspect_model_cls("Terratorch")
            return (model_info, "Terratorch")
968

969
        # Fallback to transformers impl (after resolving convert_type)
970
971
972
973
974
975
        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)
976
977
978
979
980
981
982
            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)
983
            model_info = self._try_inspect_model_cls(normalized_arch)
984
            if model_info is not None:
985
                return (model_info, arch)
986

987
        # Fallback to transformers impl (before resolving runner_type)
988
989
990
991
992
        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)
993
994
995
996
997
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

998
        return self._raise_for_unsupported(architectures)
999

1000
1001
    def resolve_model_cls(
        self,
1002
        architectures: str | list[str],
1003
        model_config: ModelConfig,
1004
    ) -> tuple[type[nn.Module], str]:
1005
1006
        if isinstance(architectures, str):
            architectures = [architectures]
1007
1008
        if not architectures:
            raise ValueError("No model architectures are specified")
1009
1010

        # Require transformers impl
1011
        if model_config.model_impl == "transformers":
1012
            arch = self._try_resolve_transformers(architectures[0], model_config)
1013
1014
1015
1016
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)
1017
        elif model_config.model_impl == "terratorch":
1018
1019
1020
1021
            arch = "Terratorch"
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
1022

1023
        # Fallback to transformers impl (after resolving convert_type)
1024
1025
1026
1027
1028
1029
        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)
1030
1031
1032
1033
1034
1035
1036
            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)
1037
            model_cls = self._try_load_model_cls(normalized_arch)
1038
1039
            if model_cls is not None:
                return (model_cls, arch)
1040

1041
        # Fallback to transformers impl (before resolving runner_type)
1042
1043
1044
1045
1046
        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)
1047
1048
1049
1050
1051
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

1052
        return self._raise_for_unsupported(architectures)
1053

1054
1055
    def is_text_generation_model(
        self,
1056
        architectures: str | list[str],
1057
        model_config: ModelConfig,
1058
    ) -> bool:
1059
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1060
        return model_cls.is_text_generation_model
1061

1062
    def is_pooling_model(
1063
        self,
1064
        architectures: str | list[str],
1065
        model_config: ModelConfig,
1066
    ) -> bool:
1067
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1068
        return model_cls.is_pooling_model
1069

1070
1071
    def is_cross_encoder_model(
        self,
1072
        architectures: str | list[str],
1073
        model_config: ModelConfig,
1074
    ) -> bool:
1075
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1076
        return model_cls.supports_cross_encoding
1077

1078
1079
    def is_multimodal_model(
        self,
1080
        architectures: str | list[str],
1081
        model_config: ModelConfig,
1082
    ) -> bool:
1083
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1084
        return model_cls.supports_multimodal
1085

1086
    def is_multimodal_raw_input_only_model(
1087
        self,
1088
        architectures: str | list[str],
1089
        model_config: ModelConfig,
1090
    ) -> bool:
1091
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1092
        return model_cls.supports_multimodal_raw_input_only
1093

1094
1095
    def is_pp_supported_model(
        self,
1096
        architectures: str | list[str],
1097
        model_config: ModelConfig,
1098
    ) -> bool:
1099
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1100
        return model_cls.supports_pp
1101

1102
1103
    def model_has_inner_state(
        self,
1104
        architectures: str | list[str],
1105
        model_config: ModelConfig,
1106
    ) -> bool:
1107
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1108
        return model_cls.has_inner_state
1109

1110
1111
    def is_attention_free_model(
        self,
1112
        architectures: str | list[str],
1113
        model_config: ModelConfig,
1114
    ) -> bool:
1115
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1116
        return model_cls.is_attention_free
1117

1118
1119
    def is_hybrid_model(
        self,
1120
        architectures: str | list[str],
1121
        model_config: ModelConfig,
1122
    ) -> bool:
1123
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1124
1125
        return model_cls.is_hybrid

1126
1127
    def is_noops_model(
        self,
1128
        architectures: str | list[str],
1129
        model_config: ModelConfig,
1130
    ) -> bool:
1131
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1132
1133
        return model_cls.has_noops

1134
1135
    def is_transcription_model(
        self,
1136
        architectures: str | list[str],
1137
        model_config: ModelConfig,
1138
    ) -> bool:
1139
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1140
1141
        return model_cls.supports_transcription

1142
1143
    def is_transcription_only_model(
        self,
1144
        architectures: str | list[str],
1145
        model_config: ModelConfig,
1146
    ) -> bool:
1147
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1148
1149
        return model_cls.supports_transcription_only

1150

1151
1152
1153
1154
1155
1156
1157
1158
1159
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()
    }
)
1160
1161
1162
1163
1164

_T = TypeVar("_T")


def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
1165
1166
1167
1168
1169
    # 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")

1170
        # `cloudpickle` allows pickling lambda functions directly
1171
        import cloudpickle
1172

1173
        input_bytes = cloudpickle.dumps((fn, output_filepath))
1174
1175
1176

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
1177
1178
1179
        returned = subprocess.run(
            _SUBPROCESS_COMMAND, input=input_bytes, capture_output=True
        )
1180
1181
1182
1183
1184
1185

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
1186
1187
1188
            raise RuntimeError(
                f"Error raised in subprocess:\n{returned.stderr.decode()}"
            ) from e
1189

1190
        with open(output_filepath, "rb") as f:
1191
1192
1193
1194
1195
1196
            return pickle.load(f)


def _run() -> None:
    # Setup plugins
    from vllm.plugins import load_general_plugins
1197

1198
1199
1200
1201
1202
    load_general_plugins()

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

    result = fn()
1203
1204
1205

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
1206
1207
1208


if __name__ == "__main__":
1209
    _run()