registry.py 45.7 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
49
from .interfaces import (
    has_inner_state,
    has_noops,
    is_attention_free,
    is_hybrid,
    supports_cross_encoding,
50
    supports_mamba_prefix_caching,
51
52
53
54
55
56
57
    supports_multimodal,
    supports_multimodal_encoder_tp_data,
    supports_multimodal_raw_input_only,
    supports_pp,
    supports_transcription,
)
from .interfaces_base import (
58
    get_attn_type,
59
60
61
62
    get_default_pooling_type,
    is_pooling_model,
    is_text_generation_model,
)
63
64
65

logger = init_logger(__name__)

66
67
_TEXT_GENERATION_MODELS = {
    # [Decoder-only]
68
    "AfmoeForCausalLM": ("afmoe", "AfmoeForCausalLM"),
69
    "ApertusForCausalLM": ("apertus", "ApertusForCausalLM"),
70
71
    "AquilaModel": ("llama", "LlamaForCausalLM"),
    "AquilaForCausalLM": ("llama", "LlamaForCausalLM"),  # AquilaChat2
Raghav Ravishankar's avatar
Raghav Ravishankar committed
72
    "ArceeForCausalLM": ("arcee", "ArceeForCausalLM"),
73
    "ArcticForCausalLM": ("arctic", "ArcticForCausalLM"),
74
75
76
77
    # baichuan-7b, upper case 'C' in the class name
    "BaiChuanForCausalLM": ("baichuan", "BaiChuanForCausalLM"),
    # baichuan-13b, lower case 'c' in the class name
    "BaichuanForCausalLM": ("baichuan", "BaichuanForCausalLM"),
78
    "BailingMoeForCausalLM": ("bailing_moe", "BailingMoeForCausalLM"),
ant-yy's avatar
ant-yy committed
79
    "BailingMoeV2ForCausalLM": ("bailing_moe", "BailingMoeV2ForCausalLM"),
Yu Chin Fabian Lim's avatar
Yu Chin Fabian Lim committed
80
    "BambaForCausalLM": ("bamba", "BambaForCausalLM"),
81
    "BloomForCausalLM": ("bloom", "BloomForCausalLM"),
82
    "ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
83
    "ChatGLMForConditionalGeneration": ("chatglm", "ChatGLMForCausalLM"),
84
    "CohereForCausalLM": ("commandr", "CohereForCausalLM"),
85
    "Cohere2ForCausalLM": ("commandr", "CohereForCausalLM"),
86
    "CwmForCausalLM": ("llama", "LlamaForCausalLM"),
87
    "DbrxForCausalLM": ("dbrx", "DbrxForCausalLM"),
88
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
89
    "DeepseekForCausalLM": ("deepseek_v2", "DeepseekForCausalLM"),
90
    "DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
91
    "DeepseekV3ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
92
    "DeepseekV32ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
93
    "Dots1ForCausalLM": ("dots1", "Dots1ForCausalLM"),
94
    "Ernie4_5ForCausalLM": ("ernie45", "Ernie4_5ForCausalLM"),
95
    "Ernie4_5_MoeForCausalLM": ("ernie45_moe", "Ernie4_5_MoeForCausalLM"),
96
    "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
97
    "Exaone4ForCausalLM": ("exaone4", "Exaone4ForCausalLM"),
98
    "Fairseq2LlamaForCausalLM": ("fairseq2_llama", "Fairseq2LlamaForCausalLM"),
99
100
101
    "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
    "FalconH1ForCausalLM": ("falcon_h1", "FalconH1ForCausalLM"),
102
    "FlexOlmoForCausalLM": ("flex_olmo", "FlexOlmoForCausalLM"),
103
104
    "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
    "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
105
    "Gemma3ForCausalLM": ("gemma3", "Gemma3ForCausalLM"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
106
    "Gemma3nForCausalLM": ("gemma3n", "Gemma3nForCausalLM"),
107
    "Qwen3NextForCausalLM": ("qwen3_next", "Qwen3NextForCausalLM"),
108
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
109
    "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
110
    "Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
111
    "GptOssForCausalLM": ("gpt_oss", "GptOssForCausalLM"),
112
113
114
115
116
117
    "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
    "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
    "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
    "GPTNeoXForCausalLM": ("gpt_neox", "GPTNeoXForCausalLM"),
    "GraniteForCausalLM": ("granite", "GraniteForCausalLM"),
    "GraniteMoeForCausalLM": ("granitemoe", "GraniteMoeForCausalLM"),
118
119
    "GraniteMoeHybridForCausalLM": ("granitemoehybrid", "GraniteMoeHybridForCausalLM"),  # noqa: E501
    "GraniteMoeSharedForCausalLM": ("granitemoeshared", "GraniteMoeSharedForCausalLM"),  # noqa: E501
120
    "GritLM": ("gritlm", "GritLM"),
Michael Goin's avatar
Michael Goin committed
121
    "Grok1ModelForCausalLM": ("grok1", "Grok1ForCausalLM"),
122
123
    "HunYuanMoEV1ForCausalLM": ("hunyuan_v1", "HunYuanMoEV1ForCausalLM"),
    "HunYuanDenseV1ForCausalLM": ("hunyuan_v1", "HunYuanDenseV1ForCausalLM"),
124
    "HCXVisionForCausalLM": ("hyperclovax_vision", "HCXVisionForCausalLM"),
125
126
    "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
    "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"),
127
    "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
128
    "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
129
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
130
    "Jais2ForCausalLM": ("jais2", "Jais2ForCausalLM"),
131
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
132
    "KimiLinearForCausalLM": ("kimi_linear", "KimiLinearForCausalLM"),  # noqa: E501
133
    "Lfm2ForCausalLM": ("lfm2", "Lfm2ForCausalLM"),
Paul Pak's avatar
Paul Pak committed
134
    "Lfm2MoeForCausalLM": ("lfm2_moe", "Lfm2MoeForCausalLM"),
135
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
136
    "Llama4ForCausalLM": ("llama4", "Llama4ForCausalLM"),
137
138
    # For decapoda-research/llama-*
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
XuruiYang's avatar
XuruiYang committed
139
    "LongcatFlashForCausalLM": ("longcat_flash", "LongcatFlashForCausalLM"),
140
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
141
    "Mamba2ForCausalLM": ("mamba2", "Mamba2ForCausalLM"),
142
143
    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
    "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
144
145
146
    "MiniMaxForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxText01ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxM1ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
147
    "MiniMaxM2ForCausalLM": ("minimax_m2", "MiniMaxM2ForCausalLM"),
148
    "MistralForCausalLM": ("llama", "LlamaForCausalLM"),
149
    "MistralLarge3ForCausalLM": ("mistral_large_3", "MistralLarge3ForCausalLM"),
150
151
152
153
    "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
    # transformers's mpt class has lower case
    "MptForCausalLM": ("mpt", "MPTForCausalLM"),
    "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
154
    "MiMoForCausalLM": ("mimo", "MiMoForCausalLM"),
155
    "MiMoV2FlashForCausalLM": ("mimo_v2_flash", "MiMoV2FlashForCausalLM"),
156
    "NemotronForCausalLM": ("nemotron", "NemotronForCausalLM"),
Luis Vega's avatar
Luis Vega committed
157
    "NemotronHForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
158
    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
159
    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
160
    "Olmo3ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
161
162
163
    "OlmoeForCausalLM": ("olmoe", "OlmoeForCausalLM"),
    "OPTForCausalLM": ("opt", "OPTForCausalLM"),
    "OrionForCausalLM": ("orion", "OrionForCausalLM"),
164
    "OuroForCausalLM": ("ouro", "OuroForCausalLM"),
165
166
    "PanguEmbeddedForCausalLM": ("openpangu", "PanguEmbeddedForCausalLM"),
    "PanguUltraMoEForCausalLM": ("openpangu", "PanguUltraMoEForCausalLM"),
167
168
169
170
    "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
    "PhiForCausalLM": ("phi", "PhiForCausalLM"),
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
    "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
171
    "Plamo2ForCausalLM": ("plamo2", "Plamo2ForCausalLM"),
172
    "Plamo3ForCausalLM": ("plamo3", "Plamo3ForCausalLM"),
173
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
174
175
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
176
177
    "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),
    "Qwen3MoeForCausalLM": ("qwen3_moe", "Qwen3MoeForCausalLM"),
178
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
179
    "SeedOssForCausalLM": ("seed_oss", "SeedOssForCausalLM"),
Song's avatar
Song committed
180
    "Step3TextForCausalLM": ("step3_text", "Step3TextForCausalLM"),
181
182
183
184
    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
185
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
186
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
187
    "TeleFLMForCausalLM": ("teleflm", "TeleFLMForCausalLM"),
188
    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
189
    "Zamba2ForCausalLM": ("zamba2", "Zamba2ForCausalLM"),
190
191
192
}

_EMBEDDING_MODELS = {
193
    # [Text-only]
194
    "BertModel": ("bert", "BertEmbeddingModel"),
195
    "BertSpladeSparseEmbeddingModel": ("bert", "BertSpladeSparseEmbeddingModel"),
196
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
197
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
198
    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
199
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
200
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
201
    "GritLM": ("gritlm", "GritLM"),
202
203
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
204
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
205
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
206
    "LlamaModel": ("llama", "LlamaForCausalLM"),
207
208
    **{
        # Multiple models share the same architecture, so we include them all
209
210
        k: (mod, arch)
        for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
211
212
        if arch == "LlamaForCausalLM"
    },
213
    "MistralModel": ("llama", "LlamaForCausalLM"),
214
    "ModernBertModel": ("modernbert", "ModernBertModel"),
215
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
216
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
217
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
218
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
219
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
220
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
221
222
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
223
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
224
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
225
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
226
    # [Multimodal]
227
    "CLIPModel": ("clip", "CLIPEmbeddingModel"),
228
229
230
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
231
    ),
232
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
233
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
234
    "SiglipModel": ("siglip", "SiglipEmbeddingModel"),
235
236
    # Technically Terratorch models work on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
237
    # models for the time being.
238
239
    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
    "Terratorch": ("terratorch", "Terratorch"),
240
241
}

242
243
_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
244
    "BertForTokenClassification": ("bert", "BertForTokenClassification"),
245
246
247
248
249
250
251
252
    "GteNewForSequenceClassification": (
        "bert_with_rope",
        "GteNewForSequenceClassification",
    ),
    "ModernBertForSequenceClassification": (
        "modernbert",
        "ModernBertForSequenceClassification",
    ),
253
254
255
256
    "ModernBertForTokenClassification": (
        "modernbert",
        "ModernBertForTokenClassification",
    ),
257
258
259
260
261
    "RobertaForSequenceClassification": ("roberta", "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": (
        "roberta",
        "RobertaForSequenceClassification",
    ),
262
    # [Auto-converted (see adapters.py)]
263
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"),  # noqa: E501,
264
265
}

266
_MULTIMODAL_MODELS = {
267
    # [Decoder-only]
268
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
269
270
271
272
    "AudioFlamingo3ForConditionalGeneration": (
        "audioflamingo3",
        "AudioFlamingo3ForConditionalGeneration",
    ),
273
274
275
    "AyaVisionForConditionalGeneration": (
        "aya_vision",
        "AyaVisionForConditionalGeneration",
276
    ),
277
    "BagelForConditionalGeneration": ("bagel", "BagelForConditionalGeneration"),
278
    "BeeForConditionalGeneration": ("bee", "BeeForConditionalGeneration"),
279
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
280
281
282
    "ChameleonForConditionalGeneration": (
        "chameleon",
        "ChameleonForConditionalGeneration",
283
    ),
284
285
286
    "Cohere2VisionForConditionalGeneration": (
        "cohere2_vision",
        "Cohere2VisionForConditionalGeneration",
287
    ),
288
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
289
    "DeepseekOCRForCausalLM": ("deepseek_ocr", "DeepseekOCRForCausalLM"),
Roger Wang's avatar
Roger Wang committed
290
    "DotsOCRForCausalLM": ("dots_ocr", "DotsOCRForCausalLM"),
291
292
293
    "Ernie4_5_VLMoeForConditionalGeneration": (
        "ernie45_vl",
        "Ernie4_5_VLMoeForConditionalGeneration",
294
    ),
295
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
296
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
297
298
299
    "Gemma3nForConditionalGeneration": (
        "gemma3n_mm",
        "Gemma3nForConditionalGeneration",
300
    ),
301
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
302
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),  # noqa: E501
Jee Jee Li's avatar
Jee Jee Li committed
303
    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),  # noqa: E501
304
305
306
    "GraniteSpeechForConditionalGeneration": (
        "granite_speech",
        "GraniteSpeechForConditionalGeneration",
307
    ),
308
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
309
310
311
312
    "HunYuanVLForConditionalGeneration": (
        "hunyuan_vision",
        "HunYuanVLForConditionalGeneration",
    ),
313
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
314
    "NemotronH_Nano_VL_V2": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
Zero's avatar
Zero committed
315
316
317
318
    "OpenCUAForConditionalGeneration": (
        "opencua",
        "OpenCUAForConditionalGeneration",
    ),
319
320
321
    "InternS1ForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
322
    ),
323
324
325
    "InternVLForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
326
    ),
327
328
329
330
331
    "Idefics3ForConditionalGeneration": (
        "idefics3",
        "Idefics3ForConditionalGeneration",
    ),
    "SmolVLMForConditionalGeneration": ("smolvlm", "SmolVLMForConditionalGeneration"),  # noqa: E501
332
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
333
334
335
    "KeyeVL1_5ForConditionalGeneration": (
        "keye_vl1_5",
        "KeyeVL1_5ForConditionalGeneration",
336
    ),
337
    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
338
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
339
340
341
342
    "LightOnOCRForConditionalGeneration": (
        "lightonocr",
        "LightOnOCRForConditionalGeneration",
    ),
343
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
344
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),  # noqa: E501
345
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
346
347
348
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
349
    ),
350
351
352
    "LlavaNextVideoForConditionalGeneration": (
        "llava_next_video",
        "LlavaNextVideoForConditionalGeneration",
353
    ),
354
355
356
    "LlavaOnevisionForConditionalGeneration": (
        "llava_onevision",
        "LlavaOnevisionForConditionalGeneration",
357
    ),
358
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
359
    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
360
361
362
    "MiniMaxVL01ForConditionalGeneration": (
        "minimax_vl_01",
        "MiniMaxVL01ForConditionalGeneration",
363
    ),
364
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
365
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
366
367
368
    "Mistral3ForConditionalGeneration": (
        "mistral3",
        "Mistral3ForConditionalGeneration",
369
    ),
370
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
371
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
372
    "Ovis": ("ovis", "Ovis"),
373
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
374
375
376
377
    "PaddleOCRVLForConditionalGeneration": (
        "paddleocr_vl",
        "PaddleOCRVLForConditionalGeneration",
    ),
378
379
380
381
    "PaliGemmaForConditionalGeneration": (
        "paligemma",
        "PaliGemmaForConditionalGeneration",
    ),
382
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
383
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
384
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
385
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),  # noqa: E501
386
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
387
388
389
    "Qwen2_5_VLForConditionalGeneration": (
        "qwen2_5_vl",
        "Qwen2_5_VLForConditionalGeneration",
390
    ),
391
392
393
    "Qwen2AudioForConditionalGeneration": (
        "qwen2_audio",
        "Qwen2AudioForConditionalGeneration",
394
    ),
395
396
397
    "Qwen2_5OmniModel": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
398
    ),
399
400
401
    "Qwen2_5OmniForConditionalGeneration": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
402
    ),
403
404
405
406
    "Qwen3OmniMoeForConditionalGeneration": (
        "qwen3_omni_moe_thinker",
        "Qwen3OmniMoeThinkerForConditionalGeneration",
    ),
407
    "Qwen3VLForConditionalGeneration": ("qwen3_vl", "Qwen3VLForConditionalGeneration"),  # noqa: E501
408
409
410
    "Qwen3VLMoeForConditionalGeneration": (
        "qwen3_vl_moe",
        "Qwen3VLMoeForConditionalGeneration",
411
    ),
412
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
Song's avatar
Song committed
413
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),  # noqa: E501
汪志鹏's avatar
汪志鹏 committed
414
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),  # noqa: E501
415
416
417
    "Tarsier2ForConditionalGeneration": (
        "qwen2_vl",
        "Tarsier2ForConditionalGeneration",
418
    ),
419
    "UltravoxModel": ("ultravox", "UltravoxModel"),
Patrick von Platen's avatar
Patrick von Platen committed
420
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),  # noqa: E501
421
    # [Encoder-decoder]
422
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
423
}
424
425

_SPECULATIVE_DECODING_MODELS = {
426
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
427
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
428
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
429
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
430
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
431
    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
432
    "Eagle3Qwen2_5vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
433
    "Eagle3Qwen3vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
434
435
436
437
    "EagleMistralLarge3ForCausalLM": (
        "mistral_large_3_eagle",
        "EagleMistralLarge3ForCausalLM",
    ),
438
    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
439
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
440
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
XuruiYang's avatar
XuruiYang committed
441
    "LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
442
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
443
    "MedusaModel": ("medusa", "Medusa"),
444
    "OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
445
    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
446
447
448
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
449
}
450

451
_TRANSFORMERS_SUPPORTED_MODELS = {
452
453
454
    # Text generation models
    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
    # Multimodal models
455
456
457
458
    "Emu3ForConditionalGeneration": (
        "transformers",
        "TransformersMultiModalForCausalLM",
    ),
459
460
461
}

_TRANSFORMERS_BACKEND_MODELS = {
462
    # Text generation models
463
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
    "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
482
    "TransformersForSequenceClassification": (
483
        "transformers",
484
        "TransformersForSequenceClassification",
485
    ),
486
    "TransformersMoEForSequenceClassification": (
487
        "transformers",
488
        "TransformersMoEForSequenceClassification",
489
    ),
490
491
492
    "TransformersMultiModalForSequenceClassification": (
        "transformers",
        "TransformersMultiModalForSequenceClassification",
493
    ),
494
}
495

496
_VLLM_MODELS = {
497
    **_TEXT_GENERATION_MODELS,
498
    **_EMBEDDING_MODELS,
499
    **_CROSS_ENCODER_MODELS,
500
    **_MULTIMODAL_MODELS,
501
    **_SPECULATIVE_DECODING_MODELS,
502
503
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
504
505
}

506
507
508
509
# 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.
510
_SUBPROCESS_COMMAND = [sys.executable, "-m", "vllm.model_executor.models.registry"]
511

512
_PREVIOUSLY_SUPPORTED_MODELS = {
513
    "MotifForCausalLM": "0.10.2",
514
    "Phi3SmallForCausalLM": "0.9.2",
515
    "Phi4FlashForCausalLM": "0.10.2",
516
    "Phi4MultimodalForCausalLM": "0.12.0",
517
518
519
520
521
522
523
524
525
    # 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",
}
526

527

528
529
@dataclass(frozen=True)
class _ModelInfo:
530
    architecture: str
531
    is_text_generation_model: bool
532
    is_pooling_model: bool
533
534
    attn_type: AttnTypeStr
    default_pooling_type: PoolingTypeStr
535
    supports_cross_encoding: bool
536
    supports_multimodal: bool
537
    supports_multimodal_raw_input_only: bool
538
    supports_multimodal_encoder_tp_data: bool
539
    supports_pp: bool
540
541
    has_inner_state: bool
    is_attention_free: bool
542
    is_hybrid: bool
543
    has_noops: bool
544
    supports_mamba_prefix_caching: bool
545
    supports_transcription: bool
546
    supports_transcription_only: bool
547
548

    @staticmethod
549
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
550
        return _ModelInfo(
551
            architecture=model.__name__,
552
            is_text_generation_model=is_text_generation_model(model),
553
            is_pooling_model=is_pooling_model(model),
554
            default_pooling_type=get_default_pooling_type(model),
555
            attn_type=get_attn_type(model),
556
            supports_cross_encoding=supports_cross_encoding(model),
557
            supports_multimodal=supports_multimodal(model),
558
559
560
561
562
563
            supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
                model
            ),
            supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
                model
            ),
564
            supports_pp=supports_pp(model),
565
566
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
567
            is_hybrid=is_hybrid(model),
568
            supports_mamba_prefix_caching=supports_mamba_prefix_caching(model),
569
            supports_transcription=supports_transcription(model),
570
571
572
            supports_transcription_only=(
                supports_transcription(model) and model.supports_transcription_only
            ),
573
            has_noops=has_noops(model),
574
        )
575
576


577
578
579
580
class _BaseRegisteredModel(ABC):
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
581

582
    @abstractmethod
583
    def load_model_cls(self) -> type[nn.Module]:
584
        raise NotImplementedError
585
586


587
588
589
590
591
592
593
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
594
    model_cls: type[nn.Module]
595
596

    @staticmethod
597
    def from_model_cls(model_cls: type[nn.Module]):
598
599
600
601
602
603
604
605
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

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

606
    def load_model_cls(self) -> type[nn.Module]:
607
608
609
610
611
612
613
614
        return self.model_cls


@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """
615

616
617
618
    module_name: str
    class_name: str

619
620
621
622
623
624
625
626
    @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"

627
    def _load_modelinfo_from_cache(self, module_hash: str) -> _ModelInfo | None:
628
629
        try:
            try:
630
                modelinfo_path = self._get_cache_dir() / self._get_cache_filename()
631
632
633
                with open(modelinfo_path, encoding="utf-8") as file:
                    mi_dict = json.load(file)
            except FileNotFoundError:
634
                logger.debug(
635
                    "Cached model info file for class %s.%s not found",
636
637
638
                    self.module_name,
                    self.class_name,
                )
639
640
641
                return None

            if mi_dict["hash"] != module_hash:
642
                logger.debug(
643
                    "Cached model info file for class %s.%s is stale",
644
645
646
                    self.module_name,
                    self.class_name,
                )
647
648
649
650
651
                return None

            # file not changed, use cached _ModelInfo properties
            return _ModelInfo(**mi_dict["modelinfo"])
        except Exception:
652
            logger.debug(
653
                "Cached model info for class %s.%s error. ",
654
655
656
                self.module_name,
                self.class_name,
            )
657
658
            return None

659
    def _save_modelinfo_to_cache(self, mi: _ModelInfo, module_hash: str) -> None:
660
661
        """save dictionary json file to cache"""
        from vllm.model_executor.model_loader.weight_utils import atomic_writer
662

663
664
665
666
667
668
669
670
        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()
671
            with atomic_writer(modelinfo_path, encoding="utf-8") as f:
672
673
674
675
676
                json.dump(modelinfo_dict, f, indent=2)
        except Exception:
            logger.exception("Error saving model info cache.")

    @logtime(logger=logger, msg="Registry inspect model class")
677
    def inspect_model_cls(self) -> _ModelInfo:
678
        model_path = Path(__file__).parent / f"{self.module_name.split('.')[-1]}.py"
679
        module_hash = None
680

681
682
        if model_path.exists():
            with open(model_path, "rb") as f:
683
                module_hash = safe_hash(f.read(), usedforsecurity=False).hexdigest()
684
685
686

            mi = self._load_modelinfo_from_cache(module_hash)
            if mi is not None:
687
                logger.debug(
688
                    "Loaded model info for class %s.%s from cache",
689
690
691
                    self.module_name,
                    self.class_name,
                )
692
693
                return mi
            else:
694
                logger.debug(
695
                    "Cache model info for class %s.%s miss. Loading model instead.",
696
697
698
                    self.module_name,
                    self.class_name,
                )
699
700
701

        # Performed in another process to avoid initializing CUDA
        mi = _run_in_subprocess(
702
703
704
705
706
            lambda: _ModelInfo.from_model_cls(self.load_model_cls())
        )
        logger.debug(
            "Loaded model info for class %s.%s", self.module_name, self.class_name
        )
707
708

        # save cache file
709
710
        if module_hash is not None:
            self._save_modelinfo_to_cache(mi, module_hash)
711
712

        return mi
713

714
    def load_model_cls(self) -> type[nn.Module]:
715
716
717
718
719
720
721
722
        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,
723
) -> type[nn.Module] | None:
724
    from vllm.platforms import current_platform
725

726
    current_platform.verify_model_arch(model_arch)
727
728
729
    try:
        return model.load_model_cls()
    except Exception:
730
        logger.exception("Error in loading model architecture '%s'", model_arch)
731
        return None
732
733


734
735
736
737
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
738
) -> _ModelInfo | None:
739
740
741
    try:
        return model.inspect_model_cls()
    except Exception:
742
        logger.exception("Error in inspecting model architecture '%s'", model_arch)
743
        return None
744
745


746
747
748
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
749
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
750

751
    def get_supported_archs(self) -> Set[str]:
752
        return self.models.keys()
753

754
755
756
    def register_model(
        self,
        model_arch: str,
757
        model_cls: type[nn.Module] | str,
758
    ) -> None:
759
760
761
        """
        Register an external model to be used in vLLM.

762
        `model_cls` can be either:
763

764
        - A [`torch.nn.Module`][] class directly referencing the model.
765
        - A string in the format `<module>:<class>` which can be used to
766
767
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
768
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
769
        """
770
771
772
773
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

774
        if model_arch in self.models:
775
776
            logger.warning(
                "Model architecture %s is already registered, and will be "
777
778
779
780
                "overwritten by the new model class %s.",
                model_arch,
                model_cls,
            )
781
782
783
784
785
786

        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)
787

788
            model = _LazyRegisteredModel(*split_str)
789
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
790
            model = _RegisteredModel.from_model_cls(model_cls)
791
        else:
792
793
794
795
            msg = (
                "`model_cls` should be a string or PyTorch model class, "
                f"not a {type(model_arch)}"
            )
796
            raise TypeError(msg)
797

798
        self.models[model_arch] = model
799

800
    def _raise_for_unsupported(self, architectures: list[str]):
801
        all_supported_archs = self.get_supported_archs()
802

803
804
805
        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
806
807
                "to be inspected. Please check the logs for more details."
            )
808

809
810
811
812
813
814
815
816
        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 "
817
818
                    "use this model architecture."
                )
819

820
821
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
822
823
            f"Supported architectures: {all_supported_archs}"
        )
824

825
    def _try_load_model_cls(self, model_arch: str) -> type[nn.Module] | None:
826
827
        if model_arch not in self.models:
            return None
828

829
        return _try_load_model_cls(model_arch, self.models[model_arch])
830

831
    def _try_inspect_model_cls(self, model_arch: str) -> _ModelInfo | None:
832
833
        if model_arch not in self.models:
            return None
834

835
836
837
838
839
840
        return _try_inspect_model_cls(model_arch, self.models[model_arch])

    def _try_resolve_transformers(
        self,
        architecture: str,
        model_config: ModelConfig,
841
    ) -> str | None:
842
843
844
        if architecture in _TRANSFORMERS_BACKEND_MODELS:
            return architecture

845
846
847
        auto_map: dict[str, str] = (
            getattr(model_config.hf_config, "auto_map", None) or dict()
        )
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880

        # 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:
881
                if model_config.model_impl != "transformers":
882
883
884
885
886
887
888
                    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 "
889
890
                    "'auto_map' (relevant if the model is custom)."
                )
891
892

        if not model_module.is_backend_compatible():
893
            if model_config.model_impl != "transformers":
894
                return None
895

896
897
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
898
899
                "is not compatible with vLLM."
            )
900

901
        return model_config._get_transformers_backend_cls()
902

903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
    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
928

929
930
    def inspect_model_cls(
        self,
931
        architectures: str | list[str],
932
        model_config: ModelConfig,
933
    ) -> tuple[_ModelInfo, str]:
934
935
        if isinstance(architectures, str):
            architectures = [architectures]
936
937
        if not architectures:
            raise ValueError("No model architectures are specified")
938
939

        # Require transformers impl
940
        if model_config.model_impl == "transformers":
941
            arch = self._try_resolve_transformers(architectures[0], model_config)
942
943
944
945
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)
946
        elif model_config.model_impl == "terratorch":
947
948
            model_info = self._try_inspect_model_cls("Terratorch")
            return (model_info, "Terratorch")
949

950
        # Fallback to transformers impl (after resolving convert_type)
951
952
953
954
955
956
        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)
957
958
959
960
961
962
963
            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)
964
            model_info = self._try_inspect_model_cls(normalized_arch)
965
            if model_info is not None:
966
                return (model_info, arch)
967

968
        # Fallback to transformers impl (before resolving runner_type)
969
970
971
972
973
        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)
974
975
976
977
978
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

979
        return self._raise_for_unsupported(architectures)
980

981
982
    def resolve_model_cls(
        self,
983
        architectures: str | list[str],
984
        model_config: ModelConfig,
985
    ) -> tuple[type[nn.Module], str]:
986
987
        if isinstance(architectures, str):
            architectures = [architectures]
988
989
        if not architectures:
            raise ValueError("No model architectures are specified")
990
991

        # Require transformers impl
992
        if model_config.model_impl == "transformers":
993
            arch = self._try_resolve_transformers(architectures[0], model_config)
994
995
996
997
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)
998
        elif model_config.model_impl == "terratorch":
999
1000
1001
1002
            arch = "Terratorch"
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
1003

1004
        # Fallback to transformers impl (after resolving convert_type)
1005
1006
1007
1008
1009
1010
        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)
1011
1012
1013
1014
1015
1016
1017
            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)
1018
            model_cls = self._try_load_model_cls(normalized_arch)
1019
1020
            if model_cls is not None:
                return (model_cls, arch)
1021

1022
        # Fallback to transformers impl (before resolving runner_type)
1023
1024
1025
1026
1027
        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)
1028
1029
1030
1031
1032
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

1033
        return self._raise_for_unsupported(architectures)
1034

1035
1036
    def is_text_generation_model(
        self,
1037
        architectures: str | list[str],
1038
        model_config: ModelConfig,
1039
    ) -> bool:
1040
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1041
        return model_cls.is_text_generation_model
1042

1043
    def is_pooling_model(
1044
        self,
1045
        architectures: str | list[str],
1046
        model_config: ModelConfig,
1047
    ) -> bool:
1048
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1049
        return model_cls.is_pooling_model
1050

1051
1052
    def is_cross_encoder_model(
        self,
1053
        architectures: str | list[str],
1054
        model_config: ModelConfig,
1055
    ) -> bool:
1056
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1057
        return model_cls.supports_cross_encoding
1058

1059
1060
    def is_multimodal_model(
        self,
1061
        architectures: str | list[str],
1062
        model_config: ModelConfig,
1063
    ) -> bool:
1064
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1065
        return model_cls.supports_multimodal
1066

1067
    def is_multimodal_raw_input_only_model(
1068
        self,
1069
        architectures: str | list[str],
1070
        model_config: ModelConfig,
1071
    ) -> bool:
1072
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1073
        return model_cls.supports_multimodal_raw_input_only
1074

1075
1076
    def is_pp_supported_model(
        self,
1077
        architectures: str | list[str],
1078
        model_config: ModelConfig,
1079
    ) -> bool:
1080
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1081
        return model_cls.supports_pp
1082

1083
1084
    def model_has_inner_state(
        self,
1085
        architectures: str | list[str],
1086
        model_config: ModelConfig,
1087
    ) -> bool:
1088
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1089
        return model_cls.has_inner_state
1090

1091
1092
    def is_attention_free_model(
        self,
1093
        architectures: str | list[str],
1094
        model_config: ModelConfig,
1095
    ) -> bool:
1096
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1097
        return model_cls.is_attention_free
1098

1099
1100
    def is_hybrid_model(
        self,
1101
        architectures: str | list[str],
1102
        model_config: ModelConfig,
1103
    ) -> bool:
1104
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1105
1106
        return model_cls.is_hybrid

1107
1108
    def is_noops_model(
        self,
1109
        architectures: str | list[str],
1110
        model_config: ModelConfig,
1111
    ) -> bool:
1112
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1113
1114
        return model_cls.has_noops

1115
1116
    def is_transcription_model(
        self,
1117
        architectures: str | list[str],
1118
        model_config: ModelConfig,
1119
    ) -> bool:
1120
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1121
1122
        return model_cls.supports_transcription

1123
1124
    def is_transcription_only_model(
        self,
1125
        architectures: str | list[str],
1126
        model_config: ModelConfig,
1127
    ) -> bool:
1128
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1129
1130
        return model_cls.supports_transcription_only

1131

1132
1133
1134
1135
1136
1137
1138
1139
1140
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()
    }
)
1141
1142
1143
1144
1145

_T = TypeVar("_T")


def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
1146
1147
1148
1149
1150
    # 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")

1151
        # `cloudpickle` allows pickling lambda functions directly
1152
        import cloudpickle
1153

1154
        input_bytes = cloudpickle.dumps((fn, output_filepath))
1155
1156
1157

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
1158
1159
1160
        returned = subprocess.run(
            _SUBPROCESS_COMMAND, input=input_bytes, capture_output=True
        )
1161
1162
1163
1164
1165
1166

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
1167
1168
1169
            raise RuntimeError(
                f"Error raised in subprocess:\n{returned.stderr.decode()}"
            ) from e
1170

1171
        with open(output_filepath, "rb") as f:
1172
1173
1174
1175
1176
1177
            return pickle.load(f)


def _run() -> None:
    # Setup plugins
    from vllm.plugins import load_general_plugins
1178

1179
1180
1181
1182
1183
    load_general_plugins()

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

    result = fn()
1184
1185
1186

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
1187
1188
1189


if __name__ == "__main__":
1190
    _run()