"tests/vscode:/vscode.git/clone" did not exist on "6a895197fafa7069be75ff615709b77546bcec30"
registry.py 48.1 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
5
6
"""
Whenever you add an architecture to this page, please also update
`tests/models/registry.py` with example HuggingFace models for it.
"""
7

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

import torch.nn as nn
23
import transformers
24

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

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


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

logger = init_logger(__name__)

69
70
_TEXT_GENERATION_MODELS = {
    # [Decoder-only]
71
    "AfmoeForCausalLM": ("afmoe", "AfmoeForCausalLM"),
72
    "ApertusForCausalLM": ("apertus", "ApertusForCausalLM"),
73
74
    "AquilaModel": ("llama", "LlamaForCausalLM"),
    "AquilaForCausalLM": ("llama", "LlamaForCausalLM"),  # AquilaChat2
Raghav Ravishankar's avatar
Raghav Ravishankar committed
75
    "ArceeForCausalLM": ("arcee", "ArceeForCausalLM"),
76
    "ArcticForCausalLM": ("arctic", "ArcticForCausalLM"),
77
78
79
80
    # baichuan-7b, upper case 'C' in the class name
    "BaiChuanForCausalLM": ("baichuan", "BaiChuanForCausalLM"),
    # baichuan-13b, lower case 'c' in the class name
    "BaichuanForCausalLM": ("baichuan", "BaichuanForCausalLM"),
81
    "BailingMoeForCausalLM": ("bailing_moe", "BailingMoeForCausalLM"),
ant-yy's avatar
ant-yy committed
82
    "BailingMoeV2ForCausalLM": ("bailing_moe", "BailingMoeV2ForCausalLM"),
Yu Chin Fabian Lim's avatar
Yu Chin Fabian Lim committed
83
    "BambaForCausalLM": ("bamba", "BambaForCausalLM"),
84
    "BloomForCausalLM": ("bloom", "BloomForCausalLM"),
85
    "ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
86
    "ChatGLMForConditionalGeneration": ("chatglm", "ChatGLMForCausalLM"),
87
    "CohereForCausalLM": ("commandr", "CohereForCausalLM"),
88
    "Cohere2ForCausalLM": ("commandr", "CohereForCausalLM"),
89
    "CwmForCausalLM": ("llama", "LlamaForCausalLM"),
90
    "DbrxForCausalLM": ("dbrx", "DbrxForCausalLM"),
91
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
92
    "DeepseekForCausalLM": ("deepseek_v2", "DeepseekForCausalLM"),
93
    "DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
94
    "DeepseekV3ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
95
    "DeepseekV32ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
96
    "Dots1ForCausalLM": ("dots1", "Dots1ForCausalLM"),
97
    "Ernie4_5ForCausalLM": ("ernie45", "Ernie4_5ForCausalLM"),
98
    "Ernie4_5_MoeForCausalLM": ("ernie45_moe", "Ernie4_5_MoeForCausalLM"),
99
    "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
100
    "Exaone4ForCausalLM": ("exaone4", "Exaone4ForCausalLM"),
Kyungmin Lee's avatar
Kyungmin Lee committed
101
    "ExaoneMoEForCausalLM": ("exaone_moe", "ExaoneMoeForCausalLM"),
102
    "Fairseq2LlamaForCausalLM": ("fairseq2_llama", "Fairseq2LlamaForCausalLM"),
103
104
105
    "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
    "FalconH1ForCausalLM": ("falcon_h1", "FalconH1ForCausalLM"),
106
    "FlexOlmoForCausalLM": ("flex_olmo", "FlexOlmoForCausalLM"),
107
108
    "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
    "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
109
    "Gemma3ForCausalLM": ("gemma3", "Gemma3ForCausalLM"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
110
    "Gemma3nForCausalLM": ("gemma3n", "Gemma3nForCausalLM"),
111
    "Qwen3NextForCausalLM": ("qwen3_next", "Qwen3NextForCausalLM"),
112
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
113
    "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
114
    "Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
115
    "Glm4MoeLiteForCausalLM": ("glm4_moe_lite", "Glm4MoeLiteForCausalLM"),
116
    "GptOssForCausalLM": ("gpt_oss", "GptOssForCausalLM"),
117
118
119
120
121
122
    "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
    "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
    "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
    "GPTNeoXForCausalLM": ("gpt_neox", "GPTNeoXForCausalLM"),
    "GraniteForCausalLM": ("granite", "GraniteForCausalLM"),
    "GraniteMoeForCausalLM": ("granitemoe", "GraniteMoeForCausalLM"),
123
124
    "GraniteMoeHybridForCausalLM": ("granitemoehybrid", "GraniteMoeHybridForCausalLM"),  # noqa: E501
    "GraniteMoeSharedForCausalLM": ("granitemoeshared", "GraniteMoeSharedForCausalLM"),  # noqa: E501
125
    "GritLM": ("gritlm", "GritLM"),
Bijaya Dangol's avatar
Bijaya Dangol committed
126
127
    "Grok1ModelForCausalLM": ("grok1", "GrokForCausalLM"),
    "Grok1ForCausalLM": ("grok1", "GrokForCausalLM"),
128
129
    "HunYuanMoEV1ForCausalLM": ("hunyuan_v1", "HunYuanMoEV1ForCausalLM"),
    "HunYuanDenseV1ForCausalLM": ("hunyuan_v1", "HunYuanDenseV1ForCausalLM"),
130
    "HCXVisionForCausalLM": ("hyperclovax_vision", "HCXVisionForCausalLM"),
131
132
    "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
    "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"),
133
    "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
134
    "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
135
136
    "IQuestCoderForCausalLM": ("llama", "LlamaForCausalLM"),
    "IQuestLoopCoderForCausalLM": ("iquest_loopcoder", "IQuestLoopCoderForCausalLM"),
137
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
138
    "Jais2ForCausalLM": ("jais2", "Jais2ForCausalLM"),
139
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
140
    "KimiLinearForCausalLM": ("kimi_linear", "KimiLinearForCausalLM"),  # noqa: E501
141
    "Lfm2ForCausalLM": ("lfm2", "Lfm2ForCausalLM"),
Paul Pak's avatar
Paul Pak committed
142
    "Lfm2MoeForCausalLM": ("lfm2_moe", "Lfm2MoeForCausalLM"),
143
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
144
    "Llama4ForCausalLM": ("llama4", "Llama4ForCausalLM"),
145
146
    # For decapoda-research/llama-*
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
XuruiYang's avatar
XuruiYang committed
147
    "LongcatFlashForCausalLM": ("longcat_flash", "LongcatFlashForCausalLM"),
148
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
149
    "Mamba2ForCausalLM": ("mamba2", "Mamba2ForCausalLM"),
150
151
    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
    "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
152
153
154
    "MiniMaxForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxText01ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxM1ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
155
    "MiniMaxM2ForCausalLM": ("minimax_m2", "MiniMaxM2ForCausalLM"),
156
    "MistralForCausalLM": ("mistral", "MistralForCausalLM"),
157
    "MistralLarge3ForCausalLM": ("mistral_large_3", "MistralLarge3ForCausalLM"),
158
159
160
161
    "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
    # transformers's mpt class has lower case
    "MptForCausalLM": ("mpt", "MPTForCausalLM"),
    "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
162
    "MiMoForCausalLM": ("mimo", "MiMoForCausalLM"),
163
    "MiMoV2FlashForCausalLM": ("mimo_v2_flash", "MiMoV2FlashForCausalLM"),
164
    "NemotronForCausalLM": ("nemotron", "NemotronForCausalLM"),
Luis Vega's avatar
Luis Vega committed
165
    "NemotronHForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
166
    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
167
    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
168
    "Olmo3ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
169
170
171
    "OlmoeForCausalLM": ("olmoe", "OlmoeForCausalLM"),
    "OPTForCausalLM": ("opt", "OPTForCausalLM"),
    "OrionForCausalLM": ("orion", "OrionForCausalLM"),
172
    "OuroForCausalLM": ("ouro", "OuroForCausalLM"),
173
    "PanguEmbeddedForCausalLM": ("openpangu", "PanguEmbeddedForCausalLM"),
174
    "PanguProMoEV2ForCausalLM": ("openpangu", "PanguProMoEV2ForCausalLM"),
175
    "PanguUltraMoEForCausalLM": ("openpangu", "PanguUltraMoEForCausalLM"),
176
177
178
179
    "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
    "PhiForCausalLM": ("phi", "PhiForCausalLM"),
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
    "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
180
    "Plamo2ForCausalLM": ("plamo2", "Plamo2ForCausalLM"),
181
    "Plamo3ForCausalLM": ("plamo3", "Plamo3ForCausalLM"),
182
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
183
184
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
185
186
    "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),
    "Qwen3MoeForCausalLM": ("qwen3_moe", "Qwen3MoeForCausalLM"),
187
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
188
    "SeedOssForCausalLM": ("seed_oss", "SeedOssForCausalLM"),
Li Xie's avatar
Li Xie committed
189
    "Step1ForCausalLM": ("step1", "Step1ForCausalLM"),
Song's avatar
Song committed
190
    "Step3TextForCausalLM": ("step3_text", "Step3TextForCausalLM"),
csy0225's avatar
csy0225 committed
191
    "Step3p5ForCausalLM": ("step3p5", "Step3p5ForCausalLM"),
192
193
194
195
    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
196
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
197
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
198
    "TeleFLMForCausalLM": ("teleflm", "TeleFLMForCausalLM"),
199
    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
200
    "Zamba2ForCausalLM": ("zamba2", "Zamba2ForCausalLM"),
201
202
203
}

_EMBEDDING_MODELS = {
204
    # [Text-only]
205
    "BertModel": ("bert", "BertEmbeddingModel"),
206
    "BertSpladeSparseEmbeddingModel": ("bert", "BertSpladeSparseEmbeddingModel"),
207
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
208
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
209
    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
210
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
211
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
212
    "GritLM": ("gritlm", "GritLM"),
213
214
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
215
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
216
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
217
    "LlamaBidirectionalModel": ("llama", "LlamaBidirectionalModel"),
218
    "LlamaModel": ("llama", "LlamaForCausalLM"),
219
220
    **{
        # Multiple models share the same architecture, so we include them all
221
222
        k: (mod, arch)
        for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
223
224
        if arch == "LlamaForCausalLM"
    },
225
    "MistralModel": ("llama", "LlamaForCausalLM"),
226
    "ModernBertModel": ("modernbert", "ModernBertModel"),
227
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
228
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
229
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
230
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
231
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
232
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
233
234
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
235
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
236
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
237
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
238
    "BgeM3EmbeddingModel": ("roberta", "BgeM3EmbeddingModel"),
239
    # [Multimodal]
240
    "CLIPModel": ("clip", "CLIPEmbeddingModel"),
241
242
243
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
244
    ),
245
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
246
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
247
    "SiglipModel": ("siglip", "SiglipEmbeddingModel"),
248
249
    # Technically Terratorch models work on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
250
    # models for the time being.
251
252
    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
    "Terratorch": ("terratorch", "Terratorch"),
253
254
}

255
256
_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
257
    "BertForTokenClassification": ("bert", "BertForTokenClassification"),
258
259
260
261
    "GteNewForSequenceClassification": (
        "bert_with_rope",
        "GteNewForSequenceClassification",
    ),
262
263
264
265
266
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"),
    "LlamaBidirectionalForSequenceClassification": (
        "llama",
        "LlamaBidirectionalForSequenceClassification",
    ),
267
268
269
270
    "ModernBertForSequenceClassification": (
        "modernbert",
        "ModernBertForSequenceClassification",
    ),
271
272
273
274
    "ModernBertForTokenClassification": (
        "modernbert",
        "ModernBertForTokenClassification",
    ),
275
276
277
278
279
    "RobertaForSequenceClassification": ("roberta", "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": (
        "roberta",
        "RobertaForSequenceClassification",
    ),
280
281
}

282
_MULTIMODAL_MODELS = {
283
    # [Decoder-only]
284
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
285
286
287
288
    "AudioFlamingo3ForConditionalGeneration": (
        "audioflamingo3",
        "AudioFlamingo3ForConditionalGeneration",
    ),
289
290
291
    "AyaVisionForConditionalGeneration": (
        "aya_vision",
        "AyaVisionForConditionalGeneration",
292
    ),
293
    "BagelForConditionalGeneration": ("bagel", "BagelForConditionalGeneration"),
294
    "BeeForConditionalGeneration": ("bee", "BeeForConditionalGeneration"),
295
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
296
297
298
    "ChameleonForConditionalGeneration": (
        "chameleon",
        "ChameleonForConditionalGeneration",
299
    ),
300
301
302
    "Cohere2VisionForConditionalGeneration": (
        "cohere2_vision",
        "Cohere2VisionForConditionalGeneration",
303
    ),
304
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
305
    "DeepseekOCRForCausalLM": ("deepseek_ocr", "DeepseekOCRForCausalLM"),
Roger Wang's avatar
Roger Wang committed
306
    "DotsOCRForCausalLM": ("dots_ocr", "DotsOCRForCausalLM"),
307
308
309
310
    "Eagle2_5_VLForConditionalGeneration": (
        "eagle2_5_vl",
        "Eagle2_5_VLForConditionalGeneration",
    ),
311
312
313
    "Ernie4_5_VLMoeForConditionalGeneration": (
        "ernie45_vl",
        "Ernie4_5_VLMoeForConditionalGeneration",
314
    ),
315
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
316
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
317
318
319
    "Gemma3nForConditionalGeneration": (
        "gemma3n_mm",
        "Gemma3nForConditionalGeneration",
320
    ),
321
    "GlmAsrForConditionalGeneration": ("glmasr", "GlmAsrForConditionalGeneration"),
322
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
323
324
325
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),
    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),
    "GlmOcrForConditionalGeneration": ("glm_ocr", "GlmOcrForConditionalGeneration"),  # noqa: E501
326
327
328
    "GraniteSpeechForConditionalGeneration": (
        "granite_speech",
        "GraniteSpeechForConditionalGeneration",
329
    ),
330
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
331
332
333
334
    "HunYuanVLForConditionalGeneration": (
        "hunyuan_vision",
        "HunYuanVLForConditionalGeneration",
    ),
ltd0924's avatar
ltd0924 committed
335
    "StepVLForConditionalGeneration": ("step_vl", "StepVLForConditionalGeneration"),
336
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
337
    "NemotronH_Nano_VL_V2": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
Zero's avatar
Zero committed
338
339
340
341
    "OpenCUAForConditionalGeneration": (
        "opencua",
        "OpenCUAForConditionalGeneration",
    ),
342
343
344
    "InternS1ForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
345
    ),
346
347
348
    "InternVLForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
349
    ),
350
351
352
353
    "Idefics3ForConditionalGeneration": (
        "idefics3",
        "Idefics3ForConditionalGeneration",
    ),
oscardev256's avatar
oscardev256 committed
354
    "IsaacForConditionalGeneration": ("isaac", "IsaacForConditionalGeneration"),
355
    "SmolVLMForConditionalGeneration": ("smolvlm", "SmolVLMForConditionalGeneration"),  # noqa: E501
356
    "KananaVForConditionalGeneration": ("kanana_v", "KananaVForConditionalGeneration"),
357
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
358
359
360
    "KeyeVL1_5ForConditionalGeneration": (
        "keye_vl1_5",
        "KeyeVL1_5ForConditionalGeneration",
361
    ),
362
    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
363
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
Roger Wang's avatar
Roger Wang committed
364
    "KimiK25ForConditionalGeneration": ("kimi_k25", "KimiK25ForConditionalGeneration"),  # noqa: E501
365
366
367
368
    "LightOnOCRForConditionalGeneration": (
        "lightonocr",
        "LightOnOCRForConditionalGeneration",
    ),
369
    "Lfm2VlForConditionalGeneration": ("lfm2_vl", "Lfm2VLForConditionalGeneration"),
370
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
371
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),  # noqa: E501
372
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
373
374
375
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
376
    ),
377
378
379
    "LlavaNextVideoForConditionalGeneration": (
        "llava_next_video",
        "LlavaNextVideoForConditionalGeneration",
380
    ),
381
382
383
    "LlavaOnevisionForConditionalGeneration": (
        "llava_onevision",
        "LlavaOnevisionForConditionalGeneration",
384
    ),
385
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
386
    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
387
388
389
    "MiniMaxVL01ForConditionalGeneration": (
        "minimax_vl_01",
        "MiniMaxVL01ForConditionalGeneration",
390
    ),
391
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
392
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
393
394
395
    "Mistral3ForConditionalGeneration": (
        "mistral3",
        "Mistral3ForConditionalGeneration",
396
    ),
397
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
398
    "Molmo2ForConditionalGeneration": ("molmo2", "Molmo2ForConditionalGeneration"),
399
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
400
    "Ovis": ("ovis", "Ovis"),
401
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
402
403
404
405
    "PaddleOCRVLForConditionalGeneration": (
        "paddleocr_vl",
        "PaddleOCRVLForConditionalGeneration",
    ),
406
407
408
409
    "PaliGemmaForConditionalGeneration": (
        "paligemma",
        "PaliGemmaForConditionalGeneration",
    ),
410
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
411
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
412
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
413
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),  # noqa: E501
414
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
415
416
417
    "Qwen2_5_VLForConditionalGeneration": (
        "qwen2_5_vl",
        "Qwen2_5_VLForConditionalGeneration",
418
    ),
419
420
421
    "Qwen2AudioForConditionalGeneration": (
        "qwen2_audio",
        "Qwen2AudioForConditionalGeneration",
422
    ),
423
424
425
    "Qwen2_5OmniModel": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
426
    ),
427
428
429
    "Qwen2_5OmniForConditionalGeneration": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
430
    ),
431
432
433
434
    "Qwen3OmniMoeForConditionalGeneration": (
        "qwen3_omni_moe_thinker",
        "Qwen3OmniMoeThinkerForConditionalGeneration",
    ),
435
    "Qwen3VLForConditionalGeneration": ("qwen3_vl", "Qwen3VLForConditionalGeneration"),  # noqa: E501
436
437
438
    "Qwen3VLMoeForConditionalGeneration": (
        "qwen3_vl_moe",
        "Qwen3VLMoeForConditionalGeneration",
439
    ),
440
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
Song's avatar
Song committed
441
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),  # noqa: E501
汪志鹏's avatar
汪志鹏 committed
442
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),  # noqa: E501
443
444
445
    "Tarsier2ForConditionalGeneration": (
        "qwen2_vl",
        "Tarsier2ForConditionalGeneration",
446
    ),
447
    "UltravoxModel": ("ultravox", "UltravoxModel"),
Patrick von Platen's avatar
Patrick von Platen committed
448
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),  # noqa: E501
Patrick von Platen's avatar
Patrick von Platen committed
449
    "VoxtralStreamingGeneration": ("voxtral_streaming", "VoxtralStreamingGeneration"),  # noqa: E501
450
    # [Encoder-decoder]
451
452
453
454
    "NemotronParseForConditionalGeneration": (
        "nemotron_parse",
        "NemotronParseForConditionalGeneration",
    ),
455
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
456
}
457
458

_SPECULATIVE_DECODING_MODELS = {
459
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
460
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
461
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
462
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
463
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
464
    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
465
    "Eagle3Qwen2_5vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
466
    "Eagle3Qwen3vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
467
468
469
470
    "EagleMistralLarge3ForCausalLM": (
        "mistral_large_3_eagle",
        "EagleMistralLarge3ForCausalLM",
    ),
471
    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
472
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
473
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
Kyungmin Lee's avatar
Kyungmin Lee committed
474
    "ExaoneMoeMTP": ("exaone_moe_mtp", "ExaoneMoeMTP"),
XuruiYang's avatar
XuruiYang committed
475
    "LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
476
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
477
    "Glm4MoeLiteMTPModel": ("glm4_moe_lite_mtp", "Glm4MoeLiteMTP"),
478
    "GlmOcrMTPModel": ("glm_ocr_mtp", "GlmOcrMTP"),
479
    "MedusaModel": ("medusa", "Medusa"),
480
    "OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
481
    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
csy0225's avatar
csy0225 committed
482
    "Step3p5MTP": ("step3p5_mtp", "Step3p5MTP"),
483
484
485
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
486
}
487

488
_TRANSFORMERS_SUPPORTED_MODELS = {
489
490
491
    # Text generation models
    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
    # Multimodal models
492
493
494
495
    "Emu3ForConditionalGeneration": (
        "transformers",
        "TransformersMultiModalForCausalLM",
    ),
496
497
498
}

_TRANSFORMERS_BACKEND_MODELS = {
499
    # Text generation models
500
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
    "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
519
    "TransformersForSequenceClassification": (
520
        "transformers",
521
        "TransformersForSequenceClassification",
522
    ),
523
    "TransformersMoEForSequenceClassification": (
524
        "transformers",
525
        "TransformersMoEForSequenceClassification",
526
    ),
527
528
529
    "TransformersMultiModalForSequenceClassification": (
        "transformers",
        "TransformersMultiModalForSequenceClassification",
530
    ),
531
}
532

533
_VLLM_MODELS = {
534
    **_TEXT_GENERATION_MODELS,
535
    **_EMBEDDING_MODELS,
536
    **_CROSS_ENCODER_MODELS,
537
    **_MULTIMODAL_MODELS,
538
    **_SPECULATIVE_DECODING_MODELS,
539
540
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
541
542
}

543
544
545
546
# 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.
547
_SUBPROCESS_COMMAND = [sys.executable, "-m", "vllm.model_executor.models.registry"]
548

549
_PREVIOUSLY_SUPPORTED_MODELS = {
550
    "MotifForCausalLM": "0.10.2",
551
    "Phi3SmallForCausalLM": "0.9.2",
552
    "Phi4FlashForCausalLM": "0.10.2",
553
    "Phi4MultimodalForCausalLM": "0.12.0",
554
555
556
557
558
559
560
561
562
    # 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",
}
563

564

565
566
@dataclass(frozen=True)
class _ModelInfo:
567
    architecture: str
568
    is_text_generation_model: bool
569
    is_pooling_model: bool
570
    attn_type: AttnTypeStr
571
572
    default_seq_pooling_type: SequencePoolingType
    default_tok_pooling_type: TokenPoolingType
573
    supports_cross_encoding: bool
574
    supports_multimodal: bool
575
    supports_multimodal_raw_input_only: bool
Patrick von Platen's avatar
Patrick von Platen committed
576
    requires_raw_input_tokens: bool
577
    supports_multimodal_encoder_tp_data: bool
578
    supports_pp: bool
579
580
    has_inner_state: bool
    is_attention_free: bool
581
    is_hybrid: bool
582
    has_noops: bool
583
    supports_mamba_prefix_caching: bool
584
    supports_transcription: bool
585
    supports_transcription_only: bool
586
587

    @staticmethod
588
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
589
        return _ModelInfo(
590
            architecture=model.__name__,
591
            is_text_generation_model=is_text_generation_model(model),
592
            is_pooling_model=is_pooling_model(model),
593
594
            default_seq_pooling_type=get_default_seq_pooling_type(model),
            default_tok_pooling_type=get_default_tok_pooling_type(model),
595
            attn_type=get_attn_type(model),
596
            supports_cross_encoding=supports_cross_encoding(model),
597
            supports_multimodal=supports_multimodal(model),
598
599
600
            supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
                model
            ),
Patrick von Platen's avatar
Patrick von Platen committed
601
            requires_raw_input_tokens=requires_raw_input_tokens(model),
602
603
604
            supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
                model
            ),
605
            supports_pp=supports_pp(model),
606
607
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
608
            is_hybrid=is_hybrid(model),
609
            supports_mamba_prefix_caching=supports_mamba_prefix_caching(model),
610
            supports_transcription=supports_transcription(model),
611
612
613
            supports_transcription_only=(
                supports_transcription(model) and model.supports_transcription_only
            ),
614
            has_noops=has_noops(model),
615
        )
616
617


618
619
620
621
class _BaseRegisteredModel(ABC):
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
622

623
    @abstractmethod
624
    def load_model_cls(self) -> type[nn.Module]:
625
        raise NotImplementedError
626
627


628
629
630
631
632
633
634
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
635
    model_cls: type[nn.Module]
636
637

    @staticmethod
638
    def from_model_cls(model_cls: type[nn.Module]):
639
640
641
642
643
644
645
646
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

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

647
    def load_model_cls(self) -> type[nn.Module]:
648
649
650
651
652
653
654
655
        return self.model_cls


@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """
656

657
658
659
    module_name: str
    class_name: str

660
661
662
663
664
665
666
667
    @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"

668
    def _load_modelinfo_from_cache(self, module_hash: str) -> _ModelInfo | None:
669
670
        try:
            try:
671
                modelinfo_path = self._get_cache_dir() / self._get_cache_filename()
672
673
674
                with open(modelinfo_path, encoding="utf-8") as file:
                    mi_dict = json.load(file)
            except FileNotFoundError:
675
                logger.debug(
676
                    "Cached model info file for class %s.%s not found",
677
678
679
                    self.module_name,
                    self.class_name,
                )
680
681
682
                return None

            if mi_dict["hash"] != module_hash:
683
                logger.debug(
684
                    "Cached model info file for class %s.%s is stale",
685
686
687
                    self.module_name,
                    self.class_name,
                )
688
689
690
691
692
                return None

            # file not changed, use cached _ModelInfo properties
            return _ModelInfo(**mi_dict["modelinfo"])
        except Exception:
693
            logger.debug(
694
                "Cached model info for class %s.%s error. ",
695
696
697
                self.module_name,
                self.class_name,
            )
698
699
            return None

700
    def _save_modelinfo_to_cache(self, mi: _ModelInfo, module_hash: str) -> None:
701
702
        """save dictionary json file to cache"""
        from vllm.model_executor.model_loader.weight_utils import atomic_writer
703

704
705
706
707
708
709
710
711
        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()
712
            with atomic_writer(modelinfo_path, encoding="utf-8") as f:
713
714
715
716
717
                json.dump(modelinfo_dict, f, indent=2)
        except Exception:
            logger.exception("Error saving model info cache.")

    @logtime(logger=logger, msg="Registry inspect model class")
718
    def inspect_model_cls(self) -> _ModelInfo:
719
        model_path = Path(__file__).parent / f"{self.module_name.split('.')[-1]}.py"
720
        module_hash = None
721

722
723
        if model_path.exists():
            with open(model_path, "rb") as f:
724
                module_hash = safe_hash(f.read(), usedforsecurity=False).hexdigest()
725
726
727

            mi = self._load_modelinfo_from_cache(module_hash)
            if mi is not None:
728
                logger.debug(
729
                    "Loaded model info for class %s.%s from cache",
730
731
732
                    self.module_name,
                    self.class_name,
                )
733
734
                return mi
            else:
735
                logger.debug(
736
                    "Cache model info for class %s.%s miss. Loading model instead.",
737
738
739
                    self.module_name,
                    self.class_name,
                )
740
741
742

        # Performed in another process to avoid initializing CUDA
        mi = _run_in_subprocess(
743
744
745
746
747
            lambda: _ModelInfo.from_model_cls(self.load_model_cls())
        )
        logger.debug(
            "Loaded model info for class %s.%s", self.module_name, self.class_name
        )
748
749

        # save cache file
750
751
        if module_hash is not None:
            self._save_modelinfo_to_cache(mi, module_hash)
752
753

        return mi
754

755
    def load_model_cls(self) -> type[nn.Module]:
756
757
758
759
760
761
762
763
        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,
764
) -> type[nn.Module] | None:
765
    from vllm.platforms import current_platform
766

767
    current_platform.verify_model_arch(model_arch)
768
769
770
    try:
        return model.load_model_cls()
    except Exception:
771
        logger.exception("Error in loading model architecture '%s'", model_arch)
772
        return None
773
774


775
776
777
778
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
779
) -> _ModelInfo | None:
780
781
782
    try:
        return model.inspect_model_cls()
    except Exception:
783
        logger.exception("Error in inspecting model architecture '%s'", model_arch)
784
        return None
785
786


787
788
789
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
790
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
791

792
    def get_supported_archs(self) -> Set[str]:
793
        return self.models.keys()
794

795
796
797
    def register_model(
        self,
        model_arch: str,
798
        model_cls: type[nn.Module] | str,
799
    ) -> None:
800
801
802
        """
        Register an external model to be used in vLLM.

803
        `model_cls` can be either:
804

805
        - A [`torch.nn.Module`][] class directly referencing the model.
806
        - A string in the format `<module>:<class>` which can be used to
807
808
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
809
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
810
        """
811
812
813
814
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

815
        if model_arch in self.models:
816
817
            logger.warning(
                "Model architecture %s is already registered, and will be "
818
819
820
821
                "overwritten by the new model class %s.",
                model_arch,
                model_cls,
            )
822
823
824
825
826
827

        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)
828

829
            model = _LazyRegisteredModel(*split_str)
830
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
831
            model = _RegisteredModel.from_model_cls(model_cls)
832
        else:
833
834
835
836
            msg = (
                "`model_cls` should be a string or PyTorch model class, "
                f"not a {type(model_arch)}"
            )
837
            raise TypeError(msg)
838

839
        self.models[model_arch] = model
840

841
    def _raise_for_unsupported(self, architectures: list[str]):
842
        all_supported_archs = self.get_supported_archs()
843

844
845
846
        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
847
848
                "to be inspected. Please check the logs for more details."
            )
849

850
851
852
853
854
855
856
857
        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 "
858
859
                    "use this model architecture."
                )
860

861
862
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
863
864
            f"Supported architectures: {all_supported_archs}"
        )
865

866
    def _try_load_model_cls(self, model_arch: str) -> type[nn.Module] | None:
867
868
        if model_arch not in self.models:
            return None
869

870
        return _try_load_model_cls(model_arch, self.models[model_arch])
871

872
    def _try_inspect_model_cls(self, model_arch: str) -> _ModelInfo | None:
873
874
        if model_arch not in self.models:
            return None
875

876
877
878
879
880
881
        return _try_inspect_model_cls(model_arch, self.models[model_arch])

    def _try_resolve_transformers(
        self,
        architecture: str,
        model_config: ModelConfig,
882
    ) -> str | None:
883
884
885
        if architecture in _TRANSFORMERS_BACKEND_MODELS:
            return architecture

886
887
888
        auto_map: dict[str, str] = (
            getattr(model_config.hf_config, "auto_map", None) or dict()
        )
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904

        # 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,
905
                        trust_remote_code=model_config.trust_remote_code,
906
907
908
909
910
911
912
913
914
915
916
917
                        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,
918
                        trust_remote_code=model_config.trust_remote_code,
919
920
921
922
923
                        warn_on_fail=True,
                    )
                    if model_module is not None:
                        break
            else:
924
                if model_config.model_impl != "transformers":
925
926
927
928
929
930
931
                    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 "
932
933
                    "'auto_map' (relevant if the model is custom)."
                )
934
935

        if not model_module.is_backend_compatible():
936
            if model_config.model_impl != "transformers":
937
                return None
938

939
940
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
941
942
                "is not compatible with vLLM."
            )
943

944
        return model_config._get_transformers_backend_cls()
945

946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
    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
971

972
973
    def inspect_model_cls(
        self,
974
        architectures: str | list[str],
975
        model_config: ModelConfig,
976
    ) -> tuple[_ModelInfo, str]:
977
978
        if isinstance(architectures, str):
            architectures = [architectures]
979
980
        if not architectures:
            raise ValueError("No model architectures are specified")
981
982

        # Require transformers impl
983
        if model_config.model_impl == "transformers":
984
            arch = self._try_resolve_transformers(architectures[0], model_config)
985
986
987
988
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)
989
        elif model_config.model_impl == "terratorch":
990
991
            model_info = self._try_inspect_model_cls("Terratorch")
            return (model_info, "Terratorch")
992

993
        # Fallback to transformers impl (after resolving convert_type)
994
995
996
997
998
999
        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)
1000
1001
1002
1003
1004
1005
1006
            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)
1007
            model_info = self._try_inspect_model_cls(normalized_arch)
1008
            if model_info is not None:
1009
                return (model_info, arch)
1010

1011
        # Fallback to transformers impl (before resolving runner_type)
1012
1013
1014
1015
1016
        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)
1017
1018
1019
1020
1021
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

1022
        return self._raise_for_unsupported(architectures)
1023

1024
1025
    def resolve_model_cls(
        self,
1026
        architectures: str | list[str],
1027
        model_config: ModelConfig,
1028
    ) -> tuple[type[nn.Module], str]:
1029
1030
        if isinstance(architectures, str):
            architectures = [architectures]
1031
1032
        if not architectures:
            raise ValueError("No model architectures are specified")
1033
1034

        # Require transformers impl
1035
        if model_config.model_impl == "transformers":
1036
            arch = self._try_resolve_transformers(architectures[0], model_config)
1037
1038
1039
1040
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)
1041
        elif model_config.model_impl == "terratorch":
1042
1043
1044
1045
            arch = "Terratorch"
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
1046

1047
        # Fallback to transformers impl (after resolving convert_type)
1048
1049
1050
1051
1052
1053
        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)
1054
1055
1056
1057
1058
1059
1060
            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)
1061
            model_cls = self._try_load_model_cls(normalized_arch)
1062
1063
            if model_cls is not None:
                return (model_cls, arch)
1064

1065
        # Fallback to transformers impl (before resolving runner_type)
1066
1067
1068
1069
1070
        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)
1071
1072
1073
1074
1075
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

1076
        return self._raise_for_unsupported(architectures)
1077

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

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

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

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

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

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

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

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

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

1150
1151
    def is_noops_model(
        self,
1152
        architectures: str | list[str],
1153
        model_config: ModelConfig,
1154
    ) -> bool:
1155
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1156
1157
        return model_cls.has_noops

1158
1159
    def is_transcription_model(
        self,
1160
        architectures: str | list[str],
1161
        model_config: ModelConfig,
1162
    ) -> bool:
1163
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1164
1165
        return model_cls.supports_transcription

1166
1167
    def is_transcription_only_model(
        self,
1168
        architectures: str | list[str],
1169
        model_config: ModelConfig,
1170
    ) -> bool:
1171
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1172
1173
        return model_cls.supports_transcription_only

1174

1175
1176
1177
1178
1179
1180
1181
1182
1183
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()
    }
)
1184
1185
1186
1187
1188

_T = TypeVar("_T")


def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
1189
1190
1191
1192
1193
    # 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")

1194
        # `cloudpickle` allows pickling lambda functions directly
1195
        import cloudpickle
1196

1197
        input_bytes = cloudpickle.dumps((fn, output_filepath))
1198
1199
1200

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
1201
1202
1203
        returned = subprocess.run(
            _SUBPROCESS_COMMAND, input=input_bytes, capture_output=True
        )
1204
1205
1206
1207
1208
1209

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
1210
1211
1212
            raise RuntimeError(
                f"Error raised in subprocess:\n{returned.stderr.decode()}"
            ) from e
1213

1214
        with open(output_filepath, "rb") as f:
1215
1216
1217
1218
1219
1220
            return pickle.load(f)


def _run() -> None:
    # Setup plugins
    from vllm.plugins import load_general_plugins
1221

1222
1223
1224
1225
1226
    load_general_plugins()

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

    result = fn()
1227
1228
1229

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
1230
1231
1232


if __name__ == "__main__":
1233
    _run()