registry.py 44.6 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 hashlib
9
import importlib
10
import json
11
import os
12
import pickle
13
14
import subprocess
import sys
15
import tempfile
16
from abc import ABC, abstractmethod
17
from collections.abc import Callable, Set
18
from dataclasses import asdict, dataclass, field
19
from functools import lru_cache
20
from pathlib import Path
21
from typing import TypeVar
22
23

import torch.nn as nn
24
import transformers
25

26
from vllm import envs
27
28
29
30
31
from vllm.config import (
    ModelConfig,
    iter_architecture_defaults,
    try_match_architecture_defaults,
)
32
from vllm.logger import init_logger
33
from vllm.logging_utils import logtime
34
35
36
37
38
39
40
41
from vllm.transformers_utils.dynamic_module import try_get_class_from_dynamic_module

from .interfaces import (
    has_inner_state,
    has_noops,
    is_attention_free,
    is_hybrid,
    supports_cross_encoding,
42
    supports_mamba_prefix_caching,
43
44
45
46
47
48
49
50
51
52
53
    supports_multimodal,
    supports_multimodal_encoder_tp_data,
    supports_multimodal_raw_input_only,
    supports_pp,
    supports_transcription,
)
from .interfaces_base import (
    get_default_pooling_type,
    is_pooling_model,
    is_text_generation_model,
)
54
55
56

logger = init_logger(__name__)

57
58
_TEXT_GENERATION_MODELS = {
    # [Decoder-only]
59
    "AfmoeForCausalLM": ("afmoe", "AfmoeForCausalLM"),
60
    "ApertusForCausalLM": ("apertus", "ApertusForCausalLM"),
61
62
    "AquilaModel": ("llama", "LlamaForCausalLM"),
    "AquilaForCausalLM": ("llama", "LlamaForCausalLM"),  # AquilaChat2
Raghav Ravishankar's avatar
Raghav Ravishankar committed
63
    "ArceeForCausalLM": ("arcee", "ArceeForCausalLM"),
64
    "ArcticForCausalLM": ("arctic", "ArcticForCausalLM"),
65
66
67
68
    # baichuan-7b, upper case 'C' in the class name
    "BaiChuanForCausalLM": ("baichuan", "BaiChuanForCausalLM"),
    # baichuan-13b, lower case 'c' in the class name
    "BaichuanForCausalLM": ("baichuan", "BaichuanForCausalLM"),
69
    "BailingMoeForCausalLM": ("bailing_moe", "BailingMoeForCausalLM"),
ant-yy's avatar
ant-yy committed
70
    "BailingMoeV2ForCausalLM": ("bailing_moe", "BailingMoeV2ForCausalLM"),
Yu Chin Fabian Lim's avatar
Yu Chin Fabian Lim committed
71
    "BambaForCausalLM": ("bamba", "BambaForCausalLM"),
72
    "BloomForCausalLM": ("bloom", "BloomForCausalLM"),
73
    "ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
74
    "ChatGLMForConditionalGeneration": ("chatglm", "ChatGLMForCausalLM"),
75
    "CohereForCausalLM": ("commandr", "CohereForCausalLM"),
76
    "Cohere2ForCausalLM": ("commandr", "CohereForCausalLM"),
77
    "CwmForCausalLM": ("llama", "LlamaForCausalLM"),
78
    "DbrxForCausalLM": ("dbrx", "DbrxForCausalLM"),
79
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
80
    "DeepseekForCausalLM": ("deepseek_v2", "DeepseekForCausalLM"),
81
    "DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
82
    "DeepseekV3ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
83
    "DeepseekV32ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
84
    "Dots1ForCausalLM": ("dots1", "Dots1ForCausalLM"),
85
    "Ernie4_5ForCausalLM": ("ernie45", "Ernie4_5ForCausalLM"),
86
    "Ernie4_5_MoeForCausalLM": ("ernie45_moe", "Ernie4_5_MoeForCausalLM"),
87
    "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
88
    "Exaone4ForCausalLM": ("exaone4", "Exaone4ForCausalLM"),
89
    "Fairseq2LlamaForCausalLM": ("fairseq2_llama", "Fairseq2LlamaForCausalLM"),
90
91
92
    "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
    "FalconH1ForCausalLM": ("falcon_h1", "FalconH1ForCausalLM"),
93
    "FlexOlmoForCausalLM": ("flex_olmo", "FlexOlmoForCausalLM"),
94
95
    "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
    "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
96
    "Gemma3ForCausalLM": ("gemma3", "Gemma3ForCausalLM"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
97
    "Gemma3nForCausalLM": ("gemma3n", "Gemma3nForCausalLM"),
98
    "Qwen3NextForCausalLM": ("qwen3_next", "Qwen3NextForCausalLM"),
99
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
100
    "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
101
    "Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
102
    "GptOssForCausalLM": ("gpt_oss", "GptOssForCausalLM"),
103
104
105
106
107
108
    "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
    "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
    "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
    "GPTNeoXForCausalLM": ("gpt_neox", "GPTNeoXForCausalLM"),
    "GraniteForCausalLM": ("granite", "GraniteForCausalLM"),
    "GraniteMoeForCausalLM": ("granitemoe", "GraniteMoeForCausalLM"),
109
110
    "GraniteMoeHybridForCausalLM": ("granitemoehybrid", "GraniteMoeHybridForCausalLM"),  # noqa: E501
    "GraniteMoeSharedForCausalLM": ("granitemoeshared", "GraniteMoeSharedForCausalLM"),  # noqa: E501
111
    "GritLM": ("gritlm", "GritLM"),
Michael Goin's avatar
Michael Goin committed
112
    "Grok1ModelForCausalLM": ("grok1", "Grok1ForCausalLM"),
113
114
    "HunYuanMoEV1ForCausalLM": ("hunyuan_v1", "HunYuanMoEV1ForCausalLM"),
    "HunYuanDenseV1ForCausalLM": ("hunyuan_v1", "HunYuanDenseV1ForCausalLM"),
115
    "HCXVisionForCausalLM": ("hyperclovax_vision", "HCXVisionForCausalLM"),
116
117
    "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
    "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"),
118
    "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
119
    "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
120
121
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
122
    "KimiLinearForCausalLM": ("kimi_linear", "KimiLinearForCausalLM"),  # noqa: E501
123
    "Lfm2ForCausalLM": ("lfm2", "Lfm2ForCausalLM"),
Paul Pak's avatar
Paul Pak committed
124
    "Lfm2MoeForCausalLM": ("lfm2_moe", "Lfm2MoeForCausalLM"),
125
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
126
    "Llama4ForCausalLM": ("llama4", "Llama4ForCausalLM"),
127
128
    # For decapoda-research/llama-*
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
XuruiYang's avatar
XuruiYang committed
129
    "LongcatFlashForCausalLM": ("longcat_flash", "LongcatFlashForCausalLM"),
130
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
131
    "Mamba2ForCausalLM": ("mamba2", "Mamba2ForCausalLM"),
132
133
    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
    "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
134
135
136
    "MiniMaxForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxText01ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxM1ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
137
    "MiniMaxM2ForCausalLM": ("minimax_m2", "MiniMaxM2ForCausalLM"),
138
139
140
141
142
    "MistralForCausalLM": ("llama", "LlamaForCausalLM"),
    "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
    # transformers's mpt class has lower case
    "MptForCausalLM": ("mpt", "MPTForCausalLM"),
    "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
143
    "MiMoForCausalLM": ("mimo", "MiMoForCausalLM"),
144
    "NemotronForCausalLM": ("nemotron", "NemotronForCausalLM"),
Luis Vega's avatar
Luis Vega committed
145
    "NemotronHForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
146
    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
147
    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
148
    "Olmo3ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
149
150
151
    "OlmoeForCausalLM": ("olmoe", "OlmoeForCausalLM"),
    "OPTForCausalLM": ("opt", "OPTForCausalLM"),
    "OrionForCausalLM": ("orion", "OrionForCausalLM"),
152
    "OuroForCausalLM": ("ouro", "OuroForCausalLM"),
153
154
    "PanguEmbeddedForCausalLM": ("openpangu", "PanguEmbeddedForCausalLM"),
    "PanguUltraMoEForCausalLM": ("openpangu", "PanguUltraMoEForCausalLM"),
155
156
157
158
    "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
    "PhiForCausalLM": ("phi", "PhiForCausalLM"),
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
    "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
159
    "Plamo2ForCausalLM": ("plamo2", "Plamo2ForCausalLM"),
160
    "Plamo3ForCausalLM": ("plamo3", "Plamo3ForCausalLM"),
161
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
162
163
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
164
165
    "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),
    "Qwen3MoeForCausalLM": ("qwen3_moe", "Qwen3MoeForCausalLM"),
166
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
167
    "SeedOssForCausalLM": ("seed_oss", "SeedOssForCausalLM"),
Song's avatar
Song committed
168
    "Step3TextForCausalLM": ("step3_text", "Step3TextForCausalLM"),
169
170
171
172
    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
173
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
174
    "TeleFLMForCausalLM": ("teleflm", "TeleFLMForCausalLM"),
175
    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
176
    "Zamba2ForCausalLM": ("zamba2", "Zamba2ForCausalLM"),
177
178
179
}

_EMBEDDING_MODELS = {
180
    # [Text-only]
181
    "BertModel": ("bert", "BertEmbeddingModel"),
182
    "BertSpladeSparseEmbeddingModel": ("bert", "BertSpladeSparseEmbeddingModel"),
183
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
184
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
185
    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
186
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
187
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
188
    "GritLM": ("gritlm", "GritLM"),
189
190
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
191
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
192
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
193
    "LlamaModel": ("llama", "LlamaForCausalLM"),
194
195
    **{
        # Multiple models share the same architecture, so we include them all
196
197
        k: (mod, arch)
        for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
198
199
        if arch == "LlamaForCausalLM"
    },
200
    "MistralModel": ("llama", "LlamaForCausalLM"),
201
    "ModernBertModel": ("modernbert", "ModernBertModel"),
202
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
203
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
204
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
205
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
206
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
207
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
208
209
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
210
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
211
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
212
    # [Multimodal]
213
    "CLIPModel": ("clip", "CLIPEmbeddingModel"),
214
215
216
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
217
    ),
218
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
219
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
220
    "SiglipModel": ("siglip", "SiglipEmbeddingModel"),
221
222
    # Technically Terratorch models work on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
223
    # models for the time being.
224
225
    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
    "Terratorch": ("terratorch", "Terratorch"),
226
227
}

228
229
_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
230
    "BertForTokenClassification": ("bert", "BertForTokenClassification"),
231
232
233
234
235
236
237
238
    "GteNewForSequenceClassification": (
        "bert_with_rope",
        "GteNewForSequenceClassification",
    ),
    "ModernBertForSequenceClassification": (
        "modernbert",
        "ModernBertForSequenceClassification",
    ),
239
240
241
242
    "ModernBertForTokenClassification": (
        "modernbert",
        "ModernBertForTokenClassification",
    ),
243
244
245
246
247
    "RobertaForSequenceClassification": ("roberta", "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": (
        "roberta",
        "RobertaForSequenceClassification",
    ),
248
    # [Auto-converted (see adapters.py)]
249
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"),  # noqa: E501,
250
251
}

252
_MULTIMODAL_MODELS = {
253
    # [Decoder-only]
254
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
255
256
257
    "AyaVisionForConditionalGeneration": (
        "aya_vision",
        "AyaVisionForConditionalGeneration",
258
    ),
259
    "BeeForConditionalGeneration": ("bee", "BeeForConditionalGeneration"),
260
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
261
262
263
    "ChameleonForConditionalGeneration": (
        "chameleon",
        "ChameleonForConditionalGeneration",
264
    ),
265
266
267
    "Cohere2VisionForConditionalGeneration": (
        "cohere2_vision",
        "Cohere2VisionForConditionalGeneration",
268
    ),
269
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
270
    "DeepseekOCRForCausalLM": ("deepseek_ocr", "DeepseekOCRForCausalLM"),
Roger Wang's avatar
Roger Wang committed
271
    "DotsOCRForCausalLM": ("dots_ocr", "DotsOCRForCausalLM"),
272
273
274
    "Ernie4_5_VLMoeForConditionalGeneration": (
        "ernie45_vl",
        "Ernie4_5_VLMoeForConditionalGeneration",
275
    ),
276
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
277
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
278
279
280
    "Gemma3nForConditionalGeneration": (
        "gemma3n_mm",
        "Gemma3nForConditionalGeneration",
281
    ),
282
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
283
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),  # noqa: E501
Jee Jee Li's avatar
Jee Jee Li committed
284
    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),  # noqa: E501
285
286
287
    "GraniteSpeechForConditionalGeneration": (
        "granite_speech",
        "GraniteSpeechForConditionalGeneration",
288
    ),
289
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
290
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
291
    "NemotronH_Nano_VL_V2": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
Zero's avatar
Zero committed
292
293
294
295
    "OpenCUAForConditionalGeneration": (
        "opencua",
        "OpenCUAForConditionalGeneration",
    ),
296
297
298
    "InternS1ForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
299
    ),
300
301
302
    "InternVLForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
303
    ),
304
305
306
307
308
    "Idefics3ForConditionalGeneration": (
        "idefics3",
        "Idefics3ForConditionalGeneration",
    ),
    "SmolVLMForConditionalGeneration": ("smolvlm", "SmolVLMForConditionalGeneration"),  # noqa: E501
309
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
310
311
312
    "KeyeVL1_5ForConditionalGeneration": (
        "keye_vl1_5",
        "KeyeVL1_5ForConditionalGeneration",
313
    ),
314
    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
315
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
316
317
318
319
    "LightOnOCRForConditionalGeneration": (
        "lightonocr",
        "LightOnOCRForConditionalGeneration",
    ),
320
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
321
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),  # noqa: E501
322
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
323
324
325
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
326
    ),
327
328
329
    "LlavaNextVideoForConditionalGeneration": (
        "llava_next_video",
        "LlavaNextVideoForConditionalGeneration",
330
    ),
331
332
333
    "LlavaOnevisionForConditionalGeneration": (
        "llava_onevision",
        "LlavaOnevisionForConditionalGeneration",
334
    ),
335
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
336
    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
337
338
339
    "MiniMaxVL01ForConditionalGeneration": (
        "minimax_vl_01",
        "MiniMaxVL01ForConditionalGeneration",
340
    ),
341
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
342
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
343
344
345
    "Mistral3ForConditionalGeneration": (
        "mistral3",
        "Mistral3ForConditionalGeneration",
346
    ),
347
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
348
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
349
    "Ovis": ("ovis", "Ovis"),
350
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
351
352
353
354
    "PaddleOCRVLForConditionalGeneration": (
        "paddleocr_vl",
        "PaddleOCRVLForConditionalGeneration",
    ),
355
356
357
358
    "PaliGemmaForConditionalGeneration": (
        "paligemma",
        "PaliGemmaForConditionalGeneration",
    ),
359
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
360
361
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
    "Phi4MultimodalForCausalLM": ("phi4_multimodal", "Phi4MultimodalForCausalLM"),  # noqa: E501
362
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
363
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),  # noqa: E501
364
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
365
366
367
    "Qwen2_5_VLForConditionalGeneration": (
        "qwen2_5_vl",
        "Qwen2_5_VLForConditionalGeneration",
368
    ),
369
370
371
    "Qwen2AudioForConditionalGeneration": (
        "qwen2_audio",
        "Qwen2AudioForConditionalGeneration",
372
    ),
373
374
375
    "Qwen2_5OmniModel": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
376
    ),
377
378
379
    "Qwen2_5OmniForConditionalGeneration": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
380
    ),
381
382
383
384
    "Qwen3OmniMoeForConditionalGeneration": (
        "qwen3_omni_moe_thinker",
        "Qwen3OmniMoeThinkerForConditionalGeneration",
    ),
385
    "Qwen3VLForConditionalGeneration": ("qwen3_vl", "Qwen3VLForConditionalGeneration"),  # noqa: E501
386
387
388
    "Qwen3VLMoeForConditionalGeneration": (
        "qwen3_vl_moe",
        "Qwen3VLMoeForConditionalGeneration",
389
    ),
390
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
Song's avatar
Song committed
391
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),  # noqa: E501
汪志鹏's avatar
汪志鹏 committed
392
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),  # noqa: E501
393
394
395
    "Tarsier2ForConditionalGeneration": (
        "qwen2_vl",
        "Tarsier2ForConditionalGeneration",
396
    ),
397
    "UltravoxModel": ("ultravox", "UltravoxModel"),
Patrick von Platen's avatar
Patrick von Platen committed
398
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),  # noqa: E501
399
    # [Encoder-decoder]
400
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
401
}
402
403

_SPECULATIVE_DECODING_MODELS = {
404
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
405
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
406
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
407
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
408
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
409
    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
410
    "Eagle3Qwen2_5vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
411
    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
412
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
413
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
XuruiYang's avatar
XuruiYang committed
414
    "LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
415
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
416
    "MedusaModel": ("medusa", "Medusa"),
417
    "OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
418
    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
419
420
421
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
422
}
423

424
_TRANSFORMERS_SUPPORTED_MODELS = {
425
426
427
    # Text generation models
    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
    # Multimodal models
428
429
430
431
    "Emu3ForConditionalGeneration": (
        "transformers",
        "TransformersMultiModalForCausalLM",
    ),
432
433
434
}

_TRANSFORMERS_BACKEND_MODELS = {
435
    # Text generation models
436
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
    "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
455
    "TransformersForSequenceClassification": (
456
        "transformers",
457
        "TransformersForSequenceClassification",
458
    ),
459
    "TransformersMoEForSequenceClassification": (
460
        "transformers",
461
        "TransformersMoEForSequenceClassification",
462
    ),
463
464
465
    "TransformersMultiModalForSequenceClassification": (
        "transformers",
        "TransformersMultiModalForSequenceClassification",
466
    ),
467
}
468

469
_VLLM_MODELS = {
470
    **_TEXT_GENERATION_MODELS,
471
    **_EMBEDDING_MODELS,
472
    **_CROSS_ENCODER_MODELS,
473
    **_MULTIMODAL_MODELS,
474
    **_SPECULATIVE_DECODING_MODELS,
475
476
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
477
478
}

479
480
481
482
# 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.
483
_SUBPROCESS_COMMAND = [sys.executable, "-m", "vllm.model_executor.models.registry"]
484

485
_PREVIOUSLY_SUPPORTED_MODELS = {
486
    "MotifForCausalLM": "0.10.2",
487
    "Phi3SmallForCausalLM": "0.9.2",
488
    "Phi4FlashForCausalLM": "0.10.2",
489
490
491
492
493
494
495
496
497
    # 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",
}
498

499

500
501
@dataclass(frozen=True)
class _ModelInfo:
502
    architecture: str
503
    is_text_generation_model: bool
504
    is_pooling_model: bool
505
    default_pooling_type: str
506
    supports_cross_encoding: bool
507
    supports_multimodal: bool
508
    supports_multimodal_raw_input_only: bool
509
    supports_multimodal_encoder_tp_data: bool
510
    supports_pp: bool
511
512
    has_inner_state: bool
    is_attention_free: bool
513
    is_hybrid: bool
514
    has_noops: bool
515
    supports_mamba_prefix_caching: bool
516
    supports_transcription: bool
517
    supports_transcription_only: bool
518
519

    @staticmethod
520
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
521
        return _ModelInfo(
522
            architecture=model.__name__,
523
            is_text_generation_model=is_text_generation_model(model),
524
            is_pooling_model=is_pooling_model(model),
525
            default_pooling_type=get_default_pooling_type(model),
526
            supports_cross_encoding=supports_cross_encoding(model),
527
            supports_multimodal=supports_multimodal(model),
528
529
530
531
532
533
            supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
                model
            ),
            supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
                model
            ),
534
            supports_pp=supports_pp(model),
535
536
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
537
            is_hybrid=is_hybrid(model),
538
            supports_mamba_prefix_caching=supports_mamba_prefix_caching(model),
539
            supports_transcription=supports_transcription(model),
540
541
542
            supports_transcription_only=(
                supports_transcription(model) and model.supports_transcription_only
            ),
543
            has_noops=has_noops(model),
544
        )
545
546


547
548
549
550
class _BaseRegisteredModel(ABC):
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
551

552
    @abstractmethod
553
    def load_model_cls(self) -> type[nn.Module]:
554
        raise NotImplementedError
555
556


557
558
559
560
561
562
563
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
564
    model_cls: type[nn.Module]
565
566

    @staticmethod
567
    def from_model_cls(model_cls: type[nn.Module]):
568
569
570
571
572
573
574
575
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

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

576
    def load_model_cls(self) -> type[nn.Module]:
577
578
579
580
581
582
583
584
        return self.model_cls


@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """
585

586
587
588
    module_name: str
    class_name: str

589
590
591
592
593
594
595
596
    @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"

597
    def _load_modelinfo_from_cache(self, module_hash: str) -> _ModelInfo | None:
598
599
        try:
            try:
600
                modelinfo_path = self._get_cache_dir() / self._get_cache_filename()
601
602
603
                with open(modelinfo_path, encoding="utf-8") as file:
                    mi_dict = json.load(file)
            except FileNotFoundError:
604
                logger.debug(
605
                    "Cached model info file for class %s.%s not found",
606
607
608
                    self.module_name,
                    self.class_name,
                )
609
610
611
                return None

            if mi_dict["hash"] != module_hash:
612
                logger.debug(
613
                    "Cached model info file for class %s.%s is stale",
614
615
616
                    self.module_name,
                    self.class_name,
                )
617
618
619
620
621
                return None

            # file not changed, use cached _ModelInfo properties
            return _ModelInfo(**mi_dict["modelinfo"])
        except Exception:
622
            logger.debug(
623
                "Cached model info for class %s.%s error. ",
624
625
626
                self.module_name,
                self.class_name,
            )
627
628
            return None

629
    def _save_modelinfo_to_cache(self, mi: _ModelInfo, module_hash: str) -> None:
630
631
        """save dictionary json file to cache"""
        from vllm.model_executor.model_loader.weight_utils import atomic_writer
632

633
634
635
636
637
638
639
640
        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()
641
            with atomic_writer(modelinfo_path, encoding="utf-8") as f:
642
643
644
645
646
                json.dump(modelinfo_dict, f, indent=2)
        except Exception:
            logger.exception("Error saving model info cache.")

    @logtime(logger=logger, msg="Registry inspect model class")
647
    def inspect_model_cls(self) -> _ModelInfo:
648
        model_path = Path(__file__).parent / f"{self.module_name.split('.')[-1]}.py"
649
        module_hash = None
650

651
652
        if model_path.exists():
            with open(model_path, "rb") as f:
653
                module_hash = hashlib.md5(f.read(), usedforsecurity=False).hexdigest()
654
655
656

            mi = self._load_modelinfo_from_cache(module_hash)
            if mi is not None:
657
                logger.debug(
658
                    "Loaded model info for class %s.%s from cache",
659
660
661
                    self.module_name,
                    self.class_name,
                )
662
663
                return mi
            else:
664
                logger.debug(
665
                    "Cache model info for class %s.%s miss. Loading model instead.",
666
667
668
                    self.module_name,
                    self.class_name,
                )
669
670
671

        # Performed in another process to avoid initializing CUDA
        mi = _run_in_subprocess(
672
673
674
675
676
            lambda: _ModelInfo.from_model_cls(self.load_model_cls())
        )
        logger.debug(
            "Loaded model info for class %s.%s", self.module_name, self.class_name
        )
677
678

        # save cache file
679
680
        if module_hash is not None:
            self._save_modelinfo_to_cache(mi, module_hash)
681
682

        return mi
683

684
    def load_model_cls(self) -> type[nn.Module]:
685
686
687
688
689
690
691
692
        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,
693
) -> type[nn.Module] | None:
694
    from vllm.platforms import current_platform
695

696
    current_platform.verify_model_arch(model_arch)
697
698
699
    try:
        return model.load_model_cls()
    except Exception:
700
        logger.exception("Error in loading model architecture '%s'", model_arch)
701
        return None
702
703


704
705
706
707
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
708
) -> _ModelInfo | None:
709
710
711
    try:
        return model.inspect_model_cls()
    except Exception:
712
        logger.exception("Error in inspecting model architecture '%s'", model_arch)
713
        return None
714
715


716
717
718
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
719
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
720

721
    def get_supported_archs(self) -> Set[str]:
722
        return self.models.keys()
723

724
725
726
    def register_model(
        self,
        model_arch: str,
727
        model_cls: type[nn.Module] | str,
728
    ) -> None:
729
730
731
        """
        Register an external model to be used in vLLM.

732
        `model_cls` can be either:
733

734
        - A [`torch.nn.Module`][] class directly referencing the model.
735
        - A string in the format `<module>:<class>` which can be used to
736
737
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
738
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
739
        """
740
741
742
743
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

744
        if model_arch in self.models:
745
746
            logger.warning(
                "Model architecture %s is already registered, and will be "
747
748
749
750
                "overwritten by the new model class %s.",
                model_arch,
                model_cls,
            )
751
752
753
754
755
756

        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)
757

758
            model = _LazyRegisteredModel(*split_str)
759
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
760
            model = _RegisteredModel.from_model_cls(model_cls)
761
        else:
762
763
764
765
            msg = (
                "`model_cls` should be a string or PyTorch model class, "
                f"not a {type(model_arch)}"
            )
766
            raise TypeError(msg)
767

768
        self.models[model_arch] = model
769

770
    def _raise_for_unsupported(self, architectures: list[str]):
771
        all_supported_archs = self.get_supported_archs()
772

773
774
775
        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
776
777
                "to be inspected. Please check the logs for more details."
            )
778

779
780
781
782
783
784
785
786
        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 "
787
788
                    "use this model architecture."
                )
789

790
791
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
792
793
            f"Supported architectures: {all_supported_archs}"
        )
794

795
    def _try_load_model_cls(self, model_arch: str) -> type[nn.Module] | None:
796
797
        if model_arch not in self.models:
            return None
798

799
        return _try_load_model_cls(model_arch, self.models[model_arch])
800

801
    def _try_inspect_model_cls(self, model_arch: str) -> _ModelInfo | None:
802
803
        if model_arch not in self.models:
            return None
804

805
806
807
808
809
810
        return _try_inspect_model_cls(model_arch, self.models[model_arch])

    def _try_resolve_transformers(
        self,
        architecture: str,
        model_config: ModelConfig,
811
    ) -> str | None:
812
813
814
        if architecture in _TRANSFORMERS_BACKEND_MODELS:
            return architecture

815
816
817
        auto_map: dict[str, str] = (
            getattr(model_config.hf_config, "auto_map", None) or dict()
        )
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850

        # 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:
851
                if model_config.model_impl != "transformers":
852
853
854
855
856
857
858
                    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 "
859
860
                    "'auto_map' (relevant if the model is custom)."
                )
861
862

        if not model_module.is_backend_compatible():
863
            if model_config.model_impl != "transformers":
864
                return None
865

866
867
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
868
869
                "is not compatible with vLLM."
            )
870

871
        return model_config._get_transformers_backend_cls()
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
    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
898

899
900
    def inspect_model_cls(
        self,
901
        architectures: str | list[str],
902
        model_config: ModelConfig,
903
    ) -> tuple[_ModelInfo, str]:
904
905
        if isinstance(architectures, str):
            architectures = [architectures]
906
907
        if not architectures:
            raise ValueError("No model architectures are specified")
908
909

        # Require transformers impl
910
        if model_config.model_impl == "transformers":
911
            arch = self._try_resolve_transformers(architectures[0], model_config)
912
913
914
915
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)
916
        elif model_config.model_impl == "terratorch":
917
918
            model_info = self._try_inspect_model_cls("Terratorch")
            return (model_info, "Terratorch")
919

920
        # Fallback to transformers impl (after resolving convert_type)
921
922
923
924
925
926
        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)
927
928
929
930
931
932
933
            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)
934
            model_info = self._try_inspect_model_cls(normalized_arch)
935
            if model_info is not None:
936
                return (model_info, arch)
937

938
        # Fallback to transformers impl (before resolving runner_type)
939
940
941
942
943
        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)
944
945
946
947
948
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

949
        return self._raise_for_unsupported(architectures)
950

951
952
    def resolve_model_cls(
        self,
953
        architectures: str | list[str],
954
        model_config: ModelConfig,
955
    ) -> tuple[type[nn.Module], str]:
956
957
        if isinstance(architectures, str):
            architectures = [architectures]
958
959
        if not architectures:
            raise ValueError("No model architectures are specified")
960
961

        # Require transformers impl
962
        if model_config.model_impl == "transformers":
963
            arch = self._try_resolve_transformers(architectures[0], model_config)
964
965
966
967
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)
968
        elif model_config.model_impl == "terratorch":
969
970
971
972
            arch = "Terratorch"
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
973

974
        # Fallback to transformers impl (after resolving convert_type)
975
976
977
978
979
980
        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)
981
982
983
984
985
986
987
            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)
988
            model_cls = self._try_load_model_cls(normalized_arch)
989
990
            if model_cls is not None:
                return (model_cls, arch)
991

992
        # Fallback to transformers impl (before resolving runner_type)
993
994
995
996
997
        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)
998
999
1000
1001
1002
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

1003
        return self._raise_for_unsupported(architectures)
1004

1005
1006
    def is_text_generation_model(
        self,
1007
        architectures: str | list[str],
1008
        model_config: ModelConfig,
1009
    ) -> bool:
1010
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1011
        return model_cls.is_text_generation_model
1012

1013
    def is_pooling_model(
1014
        self,
1015
        architectures: str | list[str],
1016
        model_config: ModelConfig,
1017
    ) -> bool:
1018
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1019
        return model_cls.is_pooling_model
1020

1021
1022
    def is_cross_encoder_model(
        self,
1023
        architectures: str | list[str],
1024
        model_config: ModelConfig,
1025
    ) -> bool:
1026
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1027
        return model_cls.supports_cross_encoding
1028

1029
1030
    def is_multimodal_model(
        self,
1031
        architectures: str | list[str],
1032
        model_config: ModelConfig,
1033
    ) -> bool:
1034
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1035
        return model_cls.supports_multimodal
1036

1037
    def is_multimodal_raw_input_only_model(
1038
        self,
1039
        architectures: str | list[str],
1040
        model_config: ModelConfig,
1041
    ) -> bool:
1042
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1043
        return model_cls.supports_multimodal_raw_input_only
1044

1045
1046
    def is_pp_supported_model(
        self,
1047
        architectures: str | list[str],
1048
        model_config: ModelConfig,
1049
    ) -> bool:
1050
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1051
        return model_cls.supports_pp
1052

1053
1054
    def model_has_inner_state(
        self,
1055
        architectures: str | list[str],
1056
        model_config: ModelConfig,
1057
    ) -> bool:
1058
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1059
        return model_cls.has_inner_state
1060

1061
1062
    def is_attention_free_model(
        self,
1063
        architectures: str | list[str],
1064
        model_config: ModelConfig,
1065
    ) -> bool:
1066
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1067
        return model_cls.is_attention_free
1068

1069
1070
    def is_hybrid_model(
        self,
1071
        architectures: str | list[str],
1072
        model_config: ModelConfig,
1073
    ) -> bool:
1074
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1075
1076
        return model_cls.is_hybrid

1077
1078
    def is_noops_model(
        self,
1079
        architectures: str | list[str],
1080
        model_config: ModelConfig,
1081
    ) -> bool:
1082
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1083
1084
        return model_cls.has_noops

1085
1086
    def is_transcription_model(
        self,
1087
        architectures: str | list[str],
1088
        model_config: ModelConfig,
1089
    ) -> bool:
1090
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1091
1092
        return model_cls.supports_transcription

1093
1094
    def is_transcription_only_model(
        self,
1095
        architectures: str | list[str],
1096
        model_config: ModelConfig,
1097
    ) -> bool:
1098
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1099
1100
        return model_cls.supports_transcription_only

1101

1102
1103
1104
1105
1106
1107
1108
1109
1110
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()
    }
)
1111
1112
1113
1114
1115

_T = TypeVar("_T")


def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
1116
1117
1118
1119
1120
    # 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")

1121
        # `cloudpickle` allows pickling lambda functions directly
1122
        import cloudpickle
1123

1124
        input_bytes = cloudpickle.dumps((fn, output_filepath))
1125
1126
1127

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
1128
1129
1130
        returned = subprocess.run(
            _SUBPROCESS_COMMAND, input=input_bytes, capture_output=True
        )
1131
1132
1133
1134
1135
1136

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
1137
1138
1139
            raise RuntimeError(
                f"Error raised in subprocess:\n{returned.stderr.decode()}"
            ) from e
1140

1141
        with open(output_filepath, "rb") as f:
1142
1143
1144
1145
1146
1147
            return pickle.load(f)


def _run() -> None:
    # Setup plugins
    from vllm.plugins import load_general_plugins
1148

1149
1150
1151
1152
1153
    load_general_plugins()

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

    result = fn()
1154
1155
1156

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
1157
1158
1159


if __name__ == "__main__":
1160
    _run()