registry.py 47.4 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
5
6
"""
Whenever you add an architecture to this page, please also update
`tests/models/registry.py` with example HuggingFace models for it.
"""
7

8
import importlib
9
import json
10
import os
11
import pickle
12
13
import subprocess
import sys
14
import tempfile
15
from abc import ABC, abstractmethod
16
from collections.abc import Callable, Set
17
from dataclasses import asdict, dataclass, field
18
from functools import lru_cache
19
from pathlib import Path
20
from typing import TYPE_CHECKING, Any, TypeVar
21
22

import torch.nn as nn
23
import transformers
24

25
from vllm import envs
26
27
28
29
30
from vllm.config import (
    ModelConfig,
    iter_architecture_defaults,
    try_match_architecture_defaults,
)
31
from vllm.logger import init_logger
32
from vllm.logging_utils import logtime
33
from vllm.transformers_utils.dynamic_module import try_get_class_from_dynamic_module
34
from vllm.utils.hashing import safe_hash
35

36
37
if TYPE_CHECKING:
    from vllm.config.model import AttnTypeStr
38
    from vllm.config.pooler import SequencePoolingType, TokenPoolingType
39
40
else:
    AttnTypeStr = Any
41
42
    SequencePoolingType = Any
    TokenPoolingType = Any
43
44


45
46
47
48
49
from .interfaces import (
    has_inner_state,
    has_noops,
    is_attention_free,
    is_hybrid,
Patrick von Platen's avatar
Patrick von Platen committed
50
    requires_raw_input_tokens,
51
    supports_cross_encoding,
52
    supports_mamba_prefix_caching,
53
54
55
56
57
58
59
    supports_multimodal,
    supports_multimodal_encoder_tp_data,
    supports_multimodal_raw_input_only,
    supports_pp,
    supports_transcription,
)
from .interfaces_base import (
60
    get_attn_type,
61
62
    get_default_seq_pooling_type,
    get_default_tok_pooling_type,
63
64
65
    is_pooling_model,
    is_text_generation_model,
)
66
67
68

logger = init_logger(__name__)

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

_EMBEDDING_MODELS = {
201
    # [Text-only]
202
    "BertModel": ("bert", "BertEmbeddingModel"),
203
    "BertSpladeSparseEmbeddingModel": ("bert", "BertSpladeSparseEmbeddingModel"),
204
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
205
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
206
    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
207
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
208
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
209
    "GritLM": ("gritlm", "GritLM"),
210
211
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
212
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
213
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
214
    "LlamaBidirectionalModel": ("llama", "LlamaBidirectionalModel"),
215
    "LlamaModel": ("llama", "LlamaForCausalLM"),
216
217
    **{
        # Multiple models share the same architecture, so we include them all
218
219
        k: (mod, arch)
        for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
220
221
        if arch == "LlamaForCausalLM"
    },
222
    "MistralModel": ("llama", "LlamaForCausalLM"),
223
    "ModernBertModel": ("modernbert", "ModernBertModel"),
224
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
225
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
226
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
227
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
228
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
229
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
230
231
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
232
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
233
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
234
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
235
    # [Multimodal]
236
    "CLIPModel": ("clip", "CLIPEmbeddingModel"),
237
238
239
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
240
    ),
241
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
242
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
243
    "SiglipModel": ("siglip", "SiglipEmbeddingModel"),
244
245
    # Technically Terratorch models work on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
246
    # models for the time being.
247
248
    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
    "Terratorch": ("terratorch", "Terratorch"),
249
250
}

251
252
_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
253
    "BertForTokenClassification": ("bert", "BertForTokenClassification"),
254
255
256
257
    "GteNewForSequenceClassification": (
        "bert_with_rope",
        "GteNewForSequenceClassification",
    ),
258
259
260
261
262
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"),
    "LlamaBidirectionalForSequenceClassification": (
        "llama",
        "LlamaBidirectionalForSequenceClassification",
    ),
263
264
265
266
    "ModernBertForSequenceClassification": (
        "modernbert",
        "ModernBertForSequenceClassification",
    ),
267
268
269
270
    "ModernBertForTokenClassification": (
        "modernbert",
        "ModernBertForTokenClassification",
    ),
271
272
273
274
275
    "RobertaForSequenceClassification": ("roberta", "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": (
        "roberta",
        "RobertaForSequenceClassification",
    ),
276
277
}

278
_MULTIMODAL_MODELS = {
279
    # [Decoder-only]
280
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
281
282
283
284
    "AudioFlamingo3ForConditionalGeneration": (
        "audioflamingo3",
        "AudioFlamingo3ForConditionalGeneration",
    ),
285
286
287
    "AyaVisionForConditionalGeneration": (
        "aya_vision",
        "AyaVisionForConditionalGeneration",
288
    ),
289
    "BagelForConditionalGeneration": ("bagel", "BagelForConditionalGeneration"),
290
    "BeeForConditionalGeneration": ("bee", "BeeForConditionalGeneration"),
291
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
292
293
294
    "ChameleonForConditionalGeneration": (
        "chameleon",
        "ChameleonForConditionalGeneration",
295
    ),
296
297
298
    "Cohere2VisionForConditionalGeneration": (
        "cohere2_vision",
        "Cohere2VisionForConditionalGeneration",
299
    ),
300
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
301
    "DeepseekOCRForCausalLM": ("deepseek_ocr", "DeepseekOCRForCausalLM"),
Roger Wang's avatar
Roger Wang committed
302
    "DotsOCRForCausalLM": ("dots_ocr", "DotsOCRForCausalLM"),
303
304
305
    "Ernie4_5_VLMoeForConditionalGeneration": (
        "ernie45_vl",
        "Ernie4_5_VLMoeForConditionalGeneration",
306
    ),
307
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
308
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
309
310
311
    "Gemma3nForConditionalGeneration": (
        "gemma3n_mm",
        "Gemma3nForConditionalGeneration",
312
    ),
313
    "GlmAsrForConditionalGeneration": ("glmasr", "GlmAsrForConditionalGeneration"),
314
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
315
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),  # noqa: E501
Jee Jee Li's avatar
Jee Jee Li committed
316
    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),  # noqa: E501
317
318
319
    "GraniteSpeechForConditionalGeneration": (
        "granite_speech",
        "GraniteSpeechForConditionalGeneration",
320
    ),
321
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
322
323
324
325
    "HunYuanVLForConditionalGeneration": (
        "hunyuan_vision",
        "HunYuanVLForConditionalGeneration",
    ),
ltd0924's avatar
ltd0924 committed
326
    "StepVLForConditionalGeneration": ("step_vl", "StepVLForConditionalGeneration"),
327
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
328
    "NemotronH_Nano_VL_V2": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
Zero's avatar
Zero committed
329
330
331
332
    "OpenCUAForConditionalGeneration": (
        "opencua",
        "OpenCUAForConditionalGeneration",
    ),
333
334
335
    "InternS1ForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
336
    ),
337
338
339
    "InternVLForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
340
    ),
341
342
343
344
    "Idefics3ForConditionalGeneration": (
        "idefics3",
        "Idefics3ForConditionalGeneration",
    ),
oscardev256's avatar
oscardev256 committed
345
    "IsaacForConditionalGeneration": ("isaac", "IsaacForConditionalGeneration"),
346
    "SmolVLMForConditionalGeneration": ("smolvlm", "SmolVLMForConditionalGeneration"),  # noqa: E501
347
    "KananaVForConditionalGeneration": ("kanana_v", "KananaVForConditionalGeneration"),
348
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
349
350
351
    "KeyeVL1_5ForConditionalGeneration": (
        "keye_vl1_5",
        "KeyeVL1_5ForConditionalGeneration",
352
    ),
353
    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
354
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
355
356
357
358
    "LightOnOCRForConditionalGeneration": (
        "lightonocr",
        "LightOnOCRForConditionalGeneration",
    ),
359
    "Lfm2VlForConditionalGeneration": ("lfm2_vl", "Lfm2VLForConditionalGeneration"),
360
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
361
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),  # noqa: E501
362
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
363
364
365
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
366
    ),
367
368
369
    "LlavaNextVideoForConditionalGeneration": (
        "llava_next_video",
        "LlavaNextVideoForConditionalGeneration",
370
    ),
371
372
373
    "LlavaOnevisionForConditionalGeneration": (
        "llava_onevision",
        "LlavaOnevisionForConditionalGeneration",
374
    ),
375
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
376
    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
377
378
379
    "MiniMaxVL01ForConditionalGeneration": (
        "minimax_vl_01",
        "MiniMaxVL01ForConditionalGeneration",
380
    ),
381
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
382
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
383
384
385
    "Mistral3ForConditionalGeneration": (
        "mistral3",
        "Mistral3ForConditionalGeneration",
386
    ),
387
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
388
    "Molmo2ForConditionalGeneration": ("molmo2", "Molmo2ForConditionalGeneration"),
389
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
390
    "Ovis": ("ovis", "Ovis"),
391
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
392
393
394
395
    "PaddleOCRVLForConditionalGeneration": (
        "paddleocr_vl",
        "PaddleOCRVLForConditionalGeneration",
    ),
396
397
398
399
    "PaliGemmaForConditionalGeneration": (
        "paligemma",
        "PaliGemmaForConditionalGeneration",
    ),
400
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
401
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
402
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
403
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),  # noqa: E501
404
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
405
406
407
    "Qwen2_5_VLForConditionalGeneration": (
        "qwen2_5_vl",
        "Qwen2_5_VLForConditionalGeneration",
408
    ),
409
410
411
    "Qwen2AudioForConditionalGeneration": (
        "qwen2_audio",
        "Qwen2AudioForConditionalGeneration",
412
    ),
413
414
415
    "Qwen2_5OmniModel": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
416
    ),
417
418
419
    "Qwen2_5OmniForConditionalGeneration": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
420
    ),
421
422
423
424
    "Qwen3OmniMoeForConditionalGeneration": (
        "qwen3_omni_moe_thinker",
        "Qwen3OmniMoeThinkerForConditionalGeneration",
    ),
425
    "Qwen3VLForConditionalGeneration": ("qwen3_vl", "Qwen3VLForConditionalGeneration"),  # noqa: E501
426
427
428
    "Qwen3VLMoeForConditionalGeneration": (
        "qwen3_vl_moe",
        "Qwen3VLMoeForConditionalGeneration",
429
    ),
430
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
Song's avatar
Song committed
431
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),  # noqa: E501
汪志鹏's avatar
汪志鹏 committed
432
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),  # noqa: E501
433
434
435
    "Tarsier2ForConditionalGeneration": (
        "qwen2_vl",
        "Tarsier2ForConditionalGeneration",
436
    ),
437
    "UltravoxModel": ("ultravox", "UltravoxModel"),
Patrick von Platen's avatar
Patrick von Platen committed
438
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),  # noqa: E501
Patrick von Platen's avatar
Patrick von Platen committed
439
    "VoxtralStreamingGeneration": ("voxtral_streaming", "VoxtralStreamingGeneration"),  # noqa: E501
440
    # [Encoder-decoder]
441
442
443
444
    "NemotronParseForConditionalGeneration": (
        "nemotron_parse",
        "NemotronParseForConditionalGeneration",
    ),
445
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
446
}
447
448

_SPECULATIVE_DECODING_MODELS = {
449
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
450
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
451
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
452
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
453
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
454
    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
455
    "Eagle3Qwen2_5vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
456
    "Eagle3Qwen3vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
457
458
459
460
    "EagleMistralLarge3ForCausalLM": (
        "mistral_large_3_eagle",
        "EagleMistralLarge3ForCausalLM",
    ),
461
    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
462
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
463
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
Kyungmin Lee's avatar
Kyungmin Lee committed
464
    "ExaoneMoeMTP": ("exaone_moe_mtp", "ExaoneMoeMTP"),
XuruiYang's avatar
XuruiYang committed
465
    "LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
466
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
467
    "MedusaModel": ("medusa", "Medusa"),
468
    "OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
469
    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
470
471
472
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
473
}
474

475
_TRANSFORMERS_SUPPORTED_MODELS = {
476
477
478
    # Text generation models
    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
    # Multimodal models
479
480
481
482
    "Emu3ForConditionalGeneration": (
        "transformers",
        "TransformersMultiModalForCausalLM",
    ),
483
484
485
}

_TRANSFORMERS_BACKEND_MODELS = {
486
    # Text generation models
487
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
    "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
506
    "TransformersForSequenceClassification": (
507
        "transformers",
508
        "TransformersForSequenceClassification",
509
    ),
510
    "TransformersMoEForSequenceClassification": (
511
        "transformers",
512
        "TransformersMoEForSequenceClassification",
513
    ),
514
515
516
    "TransformersMultiModalForSequenceClassification": (
        "transformers",
        "TransformersMultiModalForSequenceClassification",
517
    ),
518
}
519

520
_VLLM_MODELS = {
521
    **_TEXT_GENERATION_MODELS,
522
    **_EMBEDDING_MODELS,
523
    **_CROSS_ENCODER_MODELS,
524
    **_MULTIMODAL_MODELS,
525
    **_SPECULATIVE_DECODING_MODELS,
526
527
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
528
529
}

530
531
532
533
# 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.
534
_SUBPROCESS_COMMAND = [sys.executable, "-m", "vllm.model_executor.models.registry"]
535

536
_PREVIOUSLY_SUPPORTED_MODELS = {
537
    "MotifForCausalLM": "0.10.2",
538
    "Phi3SmallForCausalLM": "0.9.2",
539
    "Phi4FlashForCausalLM": "0.10.2",
540
    "Phi4MultimodalForCausalLM": "0.12.0",
541
542
543
544
545
546
547
548
549
    # 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",
}
550

551

552
553
@dataclass(frozen=True)
class _ModelInfo:
554
    architecture: str
555
    is_text_generation_model: bool
556
    is_pooling_model: bool
557
    attn_type: AttnTypeStr
558
559
    default_seq_pooling_type: SequencePoolingType
    default_tok_pooling_type: TokenPoolingType
560
    supports_cross_encoding: bool
561
    supports_multimodal: bool
562
    supports_multimodal_raw_input_only: bool
Patrick von Platen's avatar
Patrick von Platen committed
563
    requires_raw_input_tokens: bool
564
    supports_multimodal_encoder_tp_data: bool
565
    supports_pp: bool
566
567
    has_inner_state: bool
    is_attention_free: bool
568
    is_hybrid: bool
569
    has_noops: bool
570
    supports_mamba_prefix_caching: bool
571
    supports_transcription: bool
572
    supports_transcription_only: bool
573
574

    @staticmethod
575
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
576
        return _ModelInfo(
577
            architecture=model.__name__,
578
            is_text_generation_model=is_text_generation_model(model),
579
            is_pooling_model=is_pooling_model(model),
580
581
            default_seq_pooling_type=get_default_seq_pooling_type(model),
            default_tok_pooling_type=get_default_tok_pooling_type(model),
582
            attn_type=get_attn_type(model),
583
            supports_cross_encoding=supports_cross_encoding(model),
584
            supports_multimodal=supports_multimodal(model),
585
586
587
            supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
                model
            ),
Patrick von Platen's avatar
Patrick von Platen committed
588
            requires_raw_input_tokens=requires_raw_input_tokens(model),
589
590
591
            supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
                model
            ),
592
            supports_pp=supports_pp(model),
593
594
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
595
            is_hybrid=is_hybrid(model),
596
            supports_mamba_prefix_caching=supports_mamba_prefix_caching(model),
597
            supports_transcription=supports_transcription(model),
598
599
600
            supports_transcription_only=(
                supports_transcription(model) and model.supports_transcription_only
            ),
601
            has_noops=has_noops(model),
602
        )
603
604


605
606
607
608
class _BaseRegisteredModel(ABC):
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
609

610
    @abstractmethod
611
    def load_model_cls(self) -> type[nn.Module]:
612
        raise NotImplementedError
613
614


615
616
617
618
619
620
621
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
622
    model_cls: type[nn.Module]
623
624

    @staticmethod
625
    def from_model_cls(model_cls: type[nn.Module]):
626
627
628
629
630
631
632
633
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

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

634
    def load_model_cls(self) -> type[nn.Module]:
635
636
637
638
639
640
641
642
        return self.model_cls


@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """
643

644
645
646
    module_name: str
    class_name: str

647
648
649
650
651
652
653
654
    @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"

655
    def _load_modelinfo_from_cache(self, module_hash: str) -> _ModelInfo | None:
656
657
        try:
            try:
658
                modelinfo_path = self._get_cache_dir() / self._get_cache_filename()
659
660
661
                with open(modelinfo_path, encoding="utf-8") as file:
                    mi_dict = json.load(file)
            except FileNotFoundError:
662
                logger.debug(
663
                    "Cached model info file for class %s.%s not found",
664
665
666
                    self.module_name,
                    self.class_name,
                )
667
668
669
                return None

            if mi_dict["hash"] != module_hash:
670
                logger.debug(
671
                    "Cached model info file for class %s.%s is stale",
672
673
674
                    self.module_name,
                    self.class_name,
                )
675
676
677
678
679
                return None

            # file not changed, use cached _ModelInfo properties
            return _ModelInfo(**mi_dict["modelinfo"])
        except Exception:
680
            logger.debug(
681
                "Cached model info for class %s.%s error. ",
682
683
684
                self.module_name,
                self.class_name,
            )
685
686
            return None

687
    def _save_modelinfo_to_cache(self, mi: _ModelInfo, module_hash: str) -> None:
688
689
        """save dictionary json file to cache"""
        from vllm.model_executor.model_loader.weight_utils import atomic_writer
690

691
692
693
694
695
696
697
698
        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()
699
            with atomic_writer(modelinfo_path, encoding="utf-8") as f:
700
701
702
703
704
                json.dump(modelinfo_dict, f, indent=2)
        except Exception:
            logger.exception("Error saving model info cache.")

    @logtime(logger=logger, msg="Registry inspect model class")
705
    def inspect_model_cls(self) -> _ModelInfo:
706
        model_path = Path(__file__).parent / f"{self.module_name.split('.')[-1]}.py"
707
        module_hash = None
708

709
710
        if model_path.exists():
            with open(model_path, "rb") as f:
711
                module_hash = safe_hash(f.read(), usedforsecurity=False).hexdigest()
712
713
714

            mi = self._load_modelinfo_from_cache(module_hash)
            if mi is not None:
715
                logger.debug(
716
                    "Loaded model info for class %s.%s from cache",
717
718
719
                    self.module_name,
                    self.class_name,
                )
720
721
                return mi
            else:
722
                logger.debug(
723
                    "Cache model info for class %s.%s miss. Loading model instead.",
724
725
726
                    self.module_name,
                    self.class_name,
                )
727
728
729

        # Performed in another process to avoid initializing CUDA
        mi = _run_in_subprocess(
730
731
732
733
734
            lambda: _ModelInfo.from_model_cls(self.load_model_cls())
        )
        logger.debug(
            "Loaded model info for class %s.%s", self.module_name, self.class_name
        )
735
736

        # save cache file
737
738
        if module_hash is not None:
            self._save_modelinfo_to_cache(mi, module_hash)
739
740

        return mi
741

742
    def load_model_cls(self) -> type[nn.Module]:
743
744
745
746
747
748
749
750
        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,
751
) -> type[nn.Module] | None:
752
    from vllm.platforms import current_platform
753

754
    current_platform.verify_model_arch(model_arch)
755
756
757
    try:
        return model.load_model_cls()
    except Exception:
758
        logger.exception("Error in loading model architecture '%s'", model_arch)
759
        return None
760
761


762
763
764
765
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
766
) -> _ModelInfo | None:
767
768
769
    try:
        return model.inspect_model_cls()
    except Exception:
770
        logger.exception("Error in inspecting model architecture '%s'", model_arch)
771
        return None
772
773


774
775
776
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
777
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
778

779
    def get_supported_archs(self) -> Set[str]:
780
        return self.models.keys()
781

782
783
784
    def register_model(
        self,
        model_arch: str,
785
        model_cls: type[nn.Module] | str,
786
    ) -> None:
787
788
789
        """
        Register an external model to be used in vLLM.

790
        `model_cls` can be either:
791

792
        - A [`torch.nn.Module`][] class directly referencing the model.
793
        - A string in the format `<module>:<class>` which can be used to
794
795
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
796
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
797
        """
798
799
800
801
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

802
        if model_arch in self.models:
803
804
            logger.warning(
                "Model architecture %s is already registered, and will be "
805
806
807
808
                "overwritten by the new model class %s.",
                model_arch,
                model_cls,
            )
809
810
811
812
813
814

        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)
815

816
            model = _LazyRegisteredModel(*split_str)
817
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
818
            model = _RegisteredModel.from_model_cls(model_cls)
819
        else:
820
821
822
823
            msg = (
                "`model_cls` should be a string or PyTorch model class, "
                f"not a {type(model_arch)}"
            )
824
            raise TypeError(msg)
825

826
        self.models[model_arch] = model
827

828
    def _raise_for_unsupported(self, architectures: list[str]):
829
        all_supported_archs = self.get_supported_archs()
830

831
832
833
        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
834
835
                "to be inspected. Please check the logs for more details."
            )
836

837
838
839
840
841
842
843
844
        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 "
845
846
                    "use this model architecture."
                )
847

848
849
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
850
851
            f"Supported architectures: {all_supported_archs}"
        )
852

853
    def _try_load_model_cls(self, model_arch: str) -> type[nn.Module] | None:
854
855
        if model_arch not in self.models:
            return None
856

857
        return _try_load_model_cls(model_arch, self.models[model_arch])
858

859
    def _try_inspect_model_cls(self, model_arch: str) -> _ModelInfo | None:
860
861
        if model_arch not in self.models:
            return None
862

863
864
865
866
867
868
        return _try_inspect_model_cls(model_arch, self.models[model_arch])

    def _try_resolve_transformers(
        self,
        architecture: str,
        model_config: ModelConfig,
869
    ) -> str | None:
870
871
872
        if architecture in _TRANSFORMERS_BACKEND_MODELS:
            return architecture

873
874
875
        auto_map: dict[str, str] = (
            getattr(model_config.hf_config, "auto_map", None) or dict()
        )
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891

        # 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,
892
                        trust_remote_code=model_config.trust_remote_code,
893
894
895
896
897
898
899
900
901
902
903
904
                        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,
905
                        trust_remote_code=model_config.trust_remote_code,
906
907
908
909
910
                        warn_on_fail=True,
                    )
                    if model_module is not None:
                        break
            else:
911
                if model_config.model_impl != "transformers":
912
913
914
915
916
917
918
                    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 "
919
920
                    "'auto_map' (relevant if the model is custom)."
                )
921
922

        if not model_module.is_backend_compatible():
923
            if model_config.model_impl != "transformers":
924
                return None
925

926
927
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
928
929
                "is not compatible with vLLM."
            )
930

931
        return model_config._get_transformers_backend_cls()
932

933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
    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
958

959
960
    def inspect_model_cls(
        self,
961
        architectures: str | list[str],
962
        model_config: ModelConfig,
963
    ) -> tuple[_ModelInfo, str]:
964
965
        if isinstance(architectures, str):
            architectures = [architectures]
966
967
        if not architectures:
            raise ValueError("No model architectures are specified")
968
969

        # Require transformers impl
970
        if model_config.model_impl == "transformers":
971
            arch = self._try_resolve_transformers(architectures[0], model_config)
972
973
974
975
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)
976
        elif model_config.model_impl == "terratorch":
977
978
            model_info = self._try_inspect_model_cls("Terratorch")
            return (model_info, "Terratorch")
979

980
        # Fallback to transformers impl (after resolving convert_type)
981
982
983
984
985
986
        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)
987
988
989
990
991
992
993
            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)
994
            model_info = self._try_inspect_model_cls(normalized_arch)
995
            if model_info is not None:
996
                return (model_info, arch)
997

998
        # Fallback to transformers impl (before resolving runner_type)
999
1000
1001
1002
1003
        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)
1004
1005
1006
1007
1008
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

1009
        return self._raise_for_unsupported(architectures)
1010

1011
1012
    def resolve_model_cls(
        self,
1013
        architectures: str | list[str],
1014
        model_config: ModelConfig,
1015
    ) -> tuple[type[nn.Module], str]:
1016
1017
        if isinstance(architectures, str):
            architectures = [architectures]
1018
1019
        if not architectures:
            raise ValueError("No model architectures are specified")
1020
1021

        # Require transformers impl
1022
        if model_config.model_impl == "transformers":
1023
            arch = self._try_resolve_transformers(architectures[0], model_config)
1024
1025
1026
1027
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)
1028
        elif model_config.model_impl == "terratorch":
1029
1030
1031
1032
            arch = "Terratorch"
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
1033

1034
        # Fallback to transformers impl (after resolving convert_type)
1035
1036
1037
1038
1039
1040
        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)
1041
1042
1043
1044
1045
1046
1047
            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)
1048
            model_cls = self._try_load_model_cls(normalized_arch)
1049
1050
            if model_cls is not None:
                return (model_cls, arch)
1051

1052
        # Fallback to transformers impl (before resolving runner_type)
1053
1054
1055
1056
1057
        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)
1058
1059
1060
1061
1062
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

1063
        return self._raise_for_unsupported(architectures)
1064

1065
1066
    def is_text_generation_model(
        self,
1067
        architectures: str | list[str],
1068
        model_config: ModelConfig,
1069
    ) -> bool:
1070
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1071
        return model_cls.is_text_generation_model
1072

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

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

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

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

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

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

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

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

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

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

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

1161

1162
1163
1164
1165
1166
1167
1168
1169
1170
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()
    }
)
1171
1172
1173
1174
1175

_T = TypeVar("_T")


def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
1176
1177
1178
1179
1180
    # 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")

1181
        # `cloudpickle` allows pickling lambda functions directly
1182
        import cloudpickle
1183

1184
        input_bytes = cloudpickle.dumps((fn, output_filepath))
1185
1186
1187

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
1188
1189
1190
        returned = subprocess.run(
            _SUBPROCESS_COMMAND, input=input_bytes, capture_output=True
        )
1191
1192
1193
1194
1195
1196

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
1197
1198
1199
            raise RuntimeError(
                f"Error raised in subprocess:\n{returned.stderr.decode()}"
            ) from e
1200

1201
        with open(output_filepath, "rb") as f:
1202
1203
1204
1205
1206
1207
            return pickle.load(f)


def _run() -> None:
    # Setup plugins
    from vllm.plugins import load_general_plugins
1208

1209
1210
1211
1212
1213
    load_general_plugins()

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

    result = fn()
1214
1215
1216

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
1217
1218
1219


if __name__ == "__main__":
1220
    _run()