registry.py 44.4 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 hashlib
9
import importlib
10
import json
11
import os
12
import pickle
13
14
import subprocess
import sys
15
import tempfile
16
from abc import ABC, abstractmethod
17
from collections.abc import Callable, Set
18
from dataclasses import asdict, dataclass, field
19
from functools import lru_cache
20
from pathlib import Path
21
from typing import TypeVar
22
23

import torch.nn as nn
24
import transformers
25

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

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

logger = init_logger(__name__)

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

_EMBEDDING_MODELS = {
178
    # [Text-only]
179
    "BertModel": ("bert", "BertEmbeddingModel"),
180
    "BertSpladeSparseEmbeddingModel": ("bert", "BertSpladeSparseEmbeddingModel"),
181
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
182
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
183
    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
184
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
185
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
186
    "GritLM": ("gritlm", "GritLM"),
187
188
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
189
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
190
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
191
    "LlamaModel": ("llama", "LlamaForCausalLM"),
192
193
    **{
        # Multiple models share the same architecture, so we include them all
194
195
        k: (mod, arch)
        for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
196
197
        if arch == "LlamaForCausalLM"
    },
198
    "MistralModel": ("llama", "LlamaForCausalLM"),
199
    "ModernBertModel": ("modernbert", "ModernBertModel"),
200
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
201
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
202
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
203
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
204
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
205
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
206
207
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
208
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
209
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
210
    # [Multimodal]
211
    "CLIPModel": ("clip", "CLIPEmbeddingModel"),
212
213
214
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
215
    ),
216
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
217
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
218
    "SiglipModel": ("siglip", "SiglipEmbeddingModel"),
219
220
    # Technically Terratorch models work on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
221
    # models for the time being.
222
223
    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
    "Terratorch": ("terratorch", "Terratorch"),
224
225
}

226
227
_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
228
    "BertForTokenClassification": ("bert", "BertForTokenClassification"),
229
230
231
232
233
234
235
236
    "GteNewForSequenceClassification": (
        "bert_with_rope",
        "GteNewForSequenceClassification",
    ),
    "ModernBertForSequenceClassification": (
        "modernbert",
        "ModernBertForSequenceClassification",
    ),
237
238
239
240
    "ModernBertForTokenClassification": (
        "modernbert",
        "ModernBertForTokenClassification",
    ),
241
242
243
244
245
    "RobertaForSequenceClassification": ("roberta", "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": (
        "roberta",
        "RobertaForSequenceClassification",
    ),
246
    # [Auto-converted (see adapters.py)]
247
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"),  # noqa: E501,
248
249
}

250
_MULTIMODAL_MODELS = {
251
    # [Decoder-only]
252
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
253
254
255
    "AyaVisionForConditionalGeneration": (
        "aya_vision",
        "AyaVisionForConditionalGeneration",
256
    ),
257
    "BeeForConditionalGeneration": ("bee", "BeeForConditionalGeneration"),
258
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
259
260
261
    "ChameleonForConditionalGeneration": (
        "chameleon",
        "ChameleonForConditionalGeneration",
262
    ),
263
264
265
    "Cohere2VisionForConditionalGeneration": (
        "cohere2_vision",
        "Cohere2VisionForConditionalGeneration",
266
    ),
267
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
268
    "DeepseekOCRForCausalLM": ("deepseek_ocr", "DeepseekOCRForCausalLM"),
Roger Wang's avatar
Roger Wang committed
269
    "DotsOCRForCausalLM": ("dots_ocr", "DotsOCRForCausalLM"),
270
271
272
    "Ernie4_5_VLMoeForConditionalGeneration": (
        "ernie45_vl",
        "Ernie4_5_VLMoeForConditionalGeneration",
273
    ),
274
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
275
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
276
277
278
    "Gemma3nForConditionalGeneration": (
        "gemma3n_mm",
        "Gemma3nForConditionalGeneration",
279
    ),
280
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
281
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),  # noqa: E501
Jee Jee Li's avatar
Jee Jee Li committed
282
    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),  # noqa: E501
283
284
285
    "GraniteSpeechForConditionalGeneration": (
        "granite_speech",
        "GraniteSpeechForConditionalGeneration",
286
    ),
287
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
288
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
289
    "NemotronH_Nano_VL_V2": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
290
291
292
    "InternS1ForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
293
    ),
294
295
296
    "InternVLForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
297
    ),
298
299
300
301
302
    "Idefics3ForConditionalGeneration": (
        "idefics3",
        "Idefics3ForConditionalGeneration",
    ),
    "SmolVLMForConditionalGeneration": ("smolvlm", "SmolVLMForConditionalGeneration"),  # noqa: E501
303
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
304
305
306
    "KeyeVL1_5ForConditionalGeneration": (
        "keye_vl1_5",
        "KeyeVL1_5ForConditionalGeneration",
307
    ),
308
    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
309
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
310
311
312
313
    "LightOnOCRForConditionalGeneration": (
        "lightonocr",
        "LightOnOCRForConditionalGeneration",
    ),
314
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
315
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),  # noqa: E501
316
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
317
318
319
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
320
    ),
321
322
323
    "LlavaNextVideoForConditionalGeneration": (
        "llava_next_video",
        "LlavaNextVideoForConditionalGeneration",
324
    ),
325
326
327
    "LlavaOnevisionForConditionalGeneration": (
        "llava_onevision",
        "LlavaOnevisionForConditionalGeneration",
328
    ),
329
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
330
    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
331
332
333
    "MiniMaxVL01ForConditionalGeneration": (
        "minimax_vl_01",
        "MiniMaxVL01ForConditionalGeneration",
334
    ),
335
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
336
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
337
338
339
    "Mistral3ForConditionalGeneration": (
        "mistral3",
        "Mistral3ForConditionalGeneration",
340
    ),
341
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
342
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
343
    "Ovis": ("ovis", "Ovis"),
344
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
345
346
347
348
    "PaddleOCRVLForConditionalGeneration": (
        "paddleocr_vl",
        "PaddleOCRVLForConditionalGeneration",
    ),
349
350
351
352
    "PaliGemmaForConditionalGeneration": (
        "paligemma",
        "PaliGemmaForConditionalGeneration",
    ),
353
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
354
355
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
    "Phi4MultimodalForCausalLM": ("phi4_multimodal", "Phi4MultimodalForCausalLM"),  # noqa: E501
356
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
357
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),  # noqa: E501
358
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
359
360
361
    "Qwen2_5_VLForConditionalGeneration": (
        "qwen2_5_vl",
        "Qwen2_5_VLForConditionalGeneration",
362
    ),
363
364
365
    "Qwen2AudioForConditionalGeneration": (
        "qwen2_audio",
        "Qwen2AudioForConditionalGeneration",
366
    ),
367
368
369
    "Qwen2_5OmniModel": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
370
    ),
371
372
373
    "Qwen2_5OmniForConditionalGeneration": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
374
    ),
375
376
377
378
    "Qwen3OmniMoeForConditionalGeneration": (
        "qwen3_omni_moe_thinker",
        "Qwen3OmniMoeThinkerForConditionalGeneration",
    ),
379
    "Qwen3VLForConditionalGeneration": ("qwen3_vl", "Qwen3VLForConditionalGeneration"),  # noqa: E501
380
381
382
    "Qwen3VLMoeForConditionalGeneration": (
        "qwen3_vl_moe",
        "Qwen3VLMoeForConditionalGeneration",
383
    ),
384
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
Song's avatar
Song committed
385
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),  # noqa: E501
汪志鹏's avatar
汪志鹏 committed
386
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),  # noqa: E501
387
388
389
    "Tarsier2ForConditionalGeneration": (
        "qwen2_vl",
        "Tarsier2ForConditionalGeneration",
390
    ),
391
    "UltravoxModel": ("ultravox", "UltravoxModel"),
Patrick von Platen's avatar
Patrick von Platen committed
392
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),  # noqa: E501
393
    # [Encoder-decoder]
394
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
395
}
396
397

_SPECULATIVE_DECODING_MODELS = {
398
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
399
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
400
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
401
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
402
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
403
    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
404
    "Eagle3Qwen2_5vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
405
    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
406
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
407
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
XuruiYang's avatar
XuruiYang committed
408
    "LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
409
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
410
    "MedusaModel": ("medusa", "Medusa"),
411
    "OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
412
    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
413
414
415
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
416
}
417

418
_TRANSFORMERS_SUPPORTED_MODELS = {
419
420
421
    # Text generation models
    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
    # Multimodal models
422
423
424
425
    "Emu3ForConditionalGeneration": (
        "transformers",
        "TransformersMultiModalForCausalLM",
    ),
426
427
428
}

_TRANSFORMERS_BACKEND_MODELS = {
429
    # Text generation models
430
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
    "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
449
    "TransformersForSequenceClassification": (
450
        "transformers",
451
        "TransformersForSequenceClassification",
452
    ),
453
    "TransformersMoEForSequenceClassification": (
454
        "transformers",
455
        "TransformersMoEForSequenceClassification",
456
    ),
457
458
459
    "TransformersMultiModalForSequenceClassification": (
        "transformers",
        "TransformersMultiModalForSequenceClassification",
460
    ),
461
}
462

463
_VLLM_MODELS = {
464
    **_TEXT_GENERATION_MODELS,
465
    **_EMBEDDING_MODELS,
466
    **_CROSS_ENCODER_MODELS,
467
    **_MULTIMODAL_MODELS,
468
    **_SPECULATIVE_DECODING_MODELS,
469
470
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
471
472
}

473
474
475
476
# 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.
477
_SUBPROCESS_COMMAND = [sys.executable, "-m", "vllm.model_executor.models.registry"]
478

479
_PREVIOUSLY_SUPPORTED_MODELS = {
480
    "MotifForCausalLM": "0.10.2",
481
    "Phi3SmallForCausalLM": "0.9.2",
482
    "Phi4FlashForCausalLM": "0.10.2",
483
484
485
486
487
488
489
490
491
    # 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",
}
492

493

494
495
@dataclass(frozen=True)
class _ModelInfo:
496
    architecture: str
497
    is_text_generation_model: bool
498
    is_pooling_model: bool
499
    default_pooling_type: str
500
    supports_cross_encoding: bool
501
    supports_multimodal: bool
502
    supports_multimodal_raw_input_only: bool
503
    supports_multimodal_encoder_tp_data: bool
504
    supports_pp: bool
505
506
    has_inner_state: bool
    is_attention_free: bool
507
    is_hybrid: bool
508
    has_noops: bool
509
    supports_mamba_prefix_caching: bool
510
    supports_transcription: bool
511
    supports_transcription_only: bool
512
513

    @staticmethod
514
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
515
        return _ModelInfo(
516
            architecture=model.__name__,
517
            is_text_generation_model=is_text_generation_model(model),
518
            is_pooling_model=is_pooling_model(model),
519
            default_pooling_type=get_default_pooling_type(model),
520
            supports_cross_encoding=supports_cross_encoding(model),
521
            supports_multimodal=supports_multimodal(model),
522
523
524
525
526
527
            supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
                model
            ),
            supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
                model
            ),
528
            supports_pp=supports_pp(model),
529
530
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
531
            is_hybrid=is_hybrid(model),
532
            supports_mamba_prefix_caching=supports_mamba_prefix_caching(model),
533
            supports_transcription=supports_transcription(model),
534
535
536
            supports_transcription_only=(
                supports_transcription(model) and model.supports_transcription_only
            ),
537
            has_noops=has_noops(model),
538
        )
539
540


541
542
543
544
class _BaseRegisteredModel(ABC):
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
545

546
    @abstractmethod
547
    def load_model_cls(self) -> type[nn.Module]:
548
        raise NotImplementedError
549
550


551
552
553
554
555
556
557
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
558
    model_cls: type[nn.Module]
559
560

    @staticmethod
561
    def from_model_cls(model_cls: type[nn.Module]):
562
563
564
565
566
567
568
569
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

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

570
    def load_model_cls(self) -> type[nn.Module]:
571
572
573
574
575
576
577
578
        return self.model_cls


@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """
579

580
581
582
    module_name: str
    class_name: str

583
584
585
586
587
588
589
590
    @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"

591
    def _load_modelinfo_from_cache(self, module_hash: str) -> _ModelInfo | None:
592
593
        try:
            try:
594
                modelinfo_path = self._get_cache_dir() / self._get_cache_filename()
595
596
597
                with open(modelinfo_path, encoding="utf-8") as file:
                    mi_dict = json.load(file)
            except FileNotFoundError:
598
599
600
601
602
                logger.debug(
                    ("Cached model info file for class %s.%s not found"),
                    self.module_name,
                    self.class_name,
                )
603
604
605
                return None

            if mi_dict["hash"] != module_hash:
606
607
608
609
610
                logger.debug(
                    ("Cached model info file for class %s.%s is stale"),
                    self.module_name,
                    self.class_name,
                )
611
612
613
614
615
                return None

            # file not changed, use cached _ModelInfo properties
            return _ModelInfo(**mi_dict["modelinfo"])
        except Exception:
616
            logger.debug(
617
618
619
620
                ("Cached model info for class %s.%s error. "),
                self.module_name,
                self.class_name,
            )
621
622
            return None

623
    def _save_modelinfo_to_cache(self, mi: _ModelInfo, module_hash: str) -> None:
624
625
        """save dictionary json file to cache"""
        from vllm.model_executor.model_loader.weight_utils import atomic_writer
626

627
628
629
630
631
632
633
634
        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()
635
            with atomic_writer(modelinfo_path, encoding="utf-8") as f:
636
637
638
639
640
                json.dump(modelinfo_dict, f, indent=2)
        except Exception:
            logger.exception("Error saving model info cache.")

    @logtime(logger=logger, msg="Registry inspect model class")
641
    def inspect_model_cls(self) -> _ModelInfo:
642
        model_path = Path(__file__).parent / f"{self.module_name.split('.')[-1]}.py"
643
        module_hash = None
644

645
646
        if model_path.exists():
            with open(model_path, "rb") as f:
647
                module_hash = hashlib.md5(f.read(), usedforsecurity=False).hexdigest()
648
649
650

            mi = self._load_modelinfo_from_cache(module_hash)
            if mi is not None:
651
652
653
654
655
                logger.debug(
                    ("Loaded model info for class %s.%s from cache"),
                    self.module_name,
                    self.class_name,
                )
656
657
                return mi
            else:
658
659
660
661
662
                logger.debug(
                    ("Cache model info for class %s.%s miss. Loading model instead."),
                    self.module_name,
                    self.class_name,
                )
663
664
665

        # Performed in another process to avoid initializing CUDA
        mi = _run_in_subprocess(
666
667
668
669
670
            lambda: _ModelInfo.from_model_cls(self.load_model_cls())
        )
        logger.debug(
            "Loaded model info for class %s.%s", self.module_name, self.class_name
        )
671
672

        # save cache file
673
674
        if module_hash is not None:
            self._save_modelinfo_to_cache(mi, module_hash)
675
676

        return mi
677

678
    def load_model_cls(self) -> type[nn.Module]:
679
680
681
682
683
684
685
686
        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,
687
) -> type[nn.Module] | None:
688
    from vllm.platforms import current_platform
689

690
    current_platform.verify_model_arch(model_arch)
691
692
693
    try:
        return model.load_model_cls()
    except Exception:
694
        logger.exception("Error in loading model architecture '%s'", model_arch)
695
        return None
696
697


698
699
700
701
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
702
) -> _ModelInfo | None:
703
704
705
    try:
        return model.inspect_model_cls()
    except Exception:
706
        logger.exception("Error in inspecting model architecture '%s'", model_arch)
707
        return None
708
709


710
711
712
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
713
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
714

715
    def get_supported_archs(self) -> Set[str]:
716
        return self.models.keys()
717

718
719
720
    def register_model(
        self,
        model_arch: str,
721
        model_cls: type[nn.Module] | str,
722
    ) -> None:
723
724
725
        """
        Register an external model to be used in vLLM.

726
        `model_cls` can be either:
727

728
        - A [`torch.nn.Module`][] class directly referencing the model.
729
        - A string in the format `<module>:<class>` which can be used to
730
731
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
732
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
733
        """
734
735
736
737
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

738
        if model_arch in self.models:
739
740
            logger.warning(
                "Model architecture %s is already registered, and will be "
741
742
743
744
                "overwritten by the new model class %s.",
                model_arch,
                model_cls,
            )
745
746
747
748
749
750

        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)
751

752
            model = _LazyRegisteredModel(*split_str)
753
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
754
            model = _RegisteredModel.from_model_cls(model_cls)
755
        else:
756
757
758
759
            msg = (
                "`model_cls` should be a string or PyTorch model class, "
                f"not a {type(model_arch)}"
            )
760
            raise TypeError(msg)
761

762
        self.models[model_arch] = model
763

764
    def _raise_for_unsupported(self, architectures: list[str]):
765
        all_supported_archs = self.get_supported_archs()
766

767
768
769
        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
770
771
                "to be inspected. Please check the logs for more details."
            )
772

773
774
775
776
777
778
779
780
        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 "
781
782
                    "use this model architecture."
                )
783

784
785
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
786
787
            f"Supported architectures: {all_supported_archs}"
        )
788

789
    def _try_load_model_cls(self, model_arch: str) -> type[nn.Module] | None:
790
791
        if model_arch not in self.models:
            return None
792

793
        return _try_load_model_cls(model_arch, self.models[model_arch])
794

795
    def _try_inspect_model_cls(self, model_arch: str) -> _ModelInfo | None:
796
797
        if model_arch not in self.models:
            return None
798

799
800
801
802
803
804
        return _try_inspect_model_cls(model_arch, self.models[model_arch])

    def _try_resolve_transformers(
        self,
        architecture: str,
        model_config: ModelConfig,
805
    ) -> str | None:
806
807
808
        if architecture in _TRANSFORMERS_BACKEND_MODELS:
            return architecture

809
810
811
        auto_map: dict[str, str] = (
            getattr(model_config.hf_config, "auto_map", None) or dict()
        )
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844

        # Make sure that config class is always initialized before model class,
        # otherwise the model class won't be able to access the config class,
        # the expected auto_map should have correct order like:
        # "auto_map": {
        #     "AutoConfig": "<your-repo-name>--<config-name>",
        #     "AutoModel": "<your-repo-name>--<config-name>",
        #     "AutoModelFor<Task>": "<your-repo-name>--<config-name>",
        # },
        for prefix in ("AutoConfig", "AutoModel"):
            for name, module in auto_map.items():
                if name.startswith(prefix):
                    try_get_class_from_dynamic_module(
                        module,
                        model_config.model,
                        revision=model_config.revision,
                        warn_on_fail=False,
                    )

        model_module = getattr(transformers, architecture, None)

        if model_module is None:
            for name, module in auto_map.items():
                if name.startswith("AutoModel"):
                    model_module = try_get_class_from_dynamic_module(
                        module,
                        model_config.model,
                        revision=model_config.revision,
                        warn_on_fail=True,
                    )
                    if model_module is not None:
                        break
            else:
845
                if model_config.model_impl != "transformers":
846
847
848
849
850
851
852
                    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 "
853
854
                    "'auto_map' (relevant if the model is custom)."
                )
855
856

        if not model_module.is_backend_compatible():
857
            if model_config.model_impl != "transformers":
858
                return None
859

860
861
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
862
863
                "is not compatible with vLLM."
            )
864

865
        return model_config._get_transformers_backend_cls()
866

867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
    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
892

893
894
    def inspect_model_cls(
        self,
895
        architectures: str | list[str],
896
        model_config: ModelConfig,
897
    ) -> tuple[_ModelInfo, str]:
898
899
        if isinstance(architectures, str):
            architectures = [architectures]
900
901
        if not architectures:
            raise ValueError("No model architectures are specified")
902
903

        # Require transformers impl
904
        if model_config.model_impl == "transformers":
905
            arch = self._try_resolve_transformers(architectures[0], model_config)
906
907
908
909
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)
910
        elif model_config.model_impl == "terratorch":
911
912
            model_info = self._try_inspect_model_cls("Terratorch")
            return (model_info, "Terratorch")
913

914
        # Fallback to transformers impl (after resolving convert_type)
915
916
917
918
919
920
        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)
921
922
923
924
925
926
927
            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)
928
            model_info = self._try_inspect_model_cls(normalized_arch)
929
            if model_info is not None:
930
                return (model_info, arch)
931

932
        # Fallback to transformers impl (before resolving runner_type)
933
934
935
936
937
        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)
938
939
940
941
942
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

943
        return self._raise_for_unsupported(architectures)
944

945
946
    def resolve_model_cls(
        self,
947
        architectures: str | list[str],
948
        model_config: ModelConfig,
949
    ) -> tuple[type[nn.Module], str]:
950
951
        if isinstance(architectures, str):
            architectures = [architectures]
952
953
        if not architectures:
            raise ValueError("No model architectures are specified")
954
955

        # Require transformers impl
956
        if model_config.model_impl == "transformers":
957
            arch = self._try_resolve_transformers(architectures[0], model_config)
958
959
960
961
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)
962
        elif model_config.model_impl == "terratorch":
963
964
965
966
            arch = "Terratorch"
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
967

968
        # Fallback to transformers impl (after resolving convert_type)
969
970
971
972
973
974
        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)
975
976
977
978
979
980
981
            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)
982
            model_cls = self._try_load_model_cls(normalized_arch)
983
984
            if model_cls is not None:
                return (model_cls, arch)
985

986
        # Fallback to transformers impl (before resolving runner_type)
987
988
989
990
991
        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)
992
993
994
995
996
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

997
        return self._raise_for_unsupported(architectures)
998

999
1000
    def is_text_generation_model(
        self,
1001
        architectures: str | list[str],
1002
        model_config: ModelConfig,
1003
    ) -> bool:
1004
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1005
        return model_cls.is_text_generation_model
1006

1007
    def is_pooling_model(
1008
        self,
1009
        architectures: str | list[str],
1010
        model_config: ModelConfig,
1011
    ) -> bool:
1012
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1013
        return model_cls.is_pooling_model
1014

1015
1016
    def is_cross_encoder_model(
        self,
1017
        architectures: str | list[str],
1018
        model_config: ModelConfig,
1019
    ) -> bool:
1020
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1021
        return model_cls.supports_cross_encoding
1022

1023
1024
    def is_multimodal_model(
        self,
1025
        architectures: str | list[str],
1026
        model_config: ModelConfig,
1027
    ) -> bool:
1028
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1029
        return model_cls.supports_multimodal
1030

1031
    def is_multimodal_raw_input_only_model(
1032
        self,
1033
        architectures: str | list[str],
1034
        model_config: ModelConfig,
1035
    ) -> bool:
1036
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1037
        return model_cls.supports_multimodal_raw_input_only
1038

1039
1040
    def is_pp_supported_model(
        self,
1041
        architectures: str | list[str],
1042
        model_config: ModelConfig,
1043
    ) -> bool:
1044
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1045
        return model_cls.supports_pp
1046

1047
1048
    def model_has_inner_state(
        self,
1049
        architectures: str | list[str],
1050
        model_config: ModelConfig,
1051
    ) -> bool:
1052
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1053
        return model_cls.has_inner_state
1054

1055
1056
    def is_attention_free_model(
        self,
1057
        architectures: str | list[str],
1058
        model_config: ModelConfig,
1059
    ) -> bool:
1060
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1061
        return model_cls.is_attention_free
1062

1063
1064
    def is_hybrid_model(
        self,
1065
        architectures: str | list[str],
1066
        model_config: ModelConfig,
1067
    ) -> bool:
1068
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1069
1070
        return model_cls.is_hybrid

1071
1072
    def is_noops_model(
        self,
1073
        architectures: str | list[str],
1074
        model_config: ModelConfig,
1075
    ) -> bool:
1076
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1077
1078
        return model_cls.has_noops

1079
1080
    def is_transcription_model(
        self,
1081
        architectures: str | list[str],
1082
        model_config: ModelConfig,
1083
    ) -> bool:
1084
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1085
1086
        return model_cls.supports_transcription

1087
1088
    def is_transcription_only_model(
        self,
1089
        architectures: str | list[str],
1090
        model_config: ModelConfig,
1091
    ) -> bool:
1092
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1093
1094
        return model_cls.supports_transcription_only

1095

1096
1097
1098
1099
1100
1101
1102
1103
1104
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()
    }
)
1105
1106
1107
1108
1109

_T = TypeVar("_T")


def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
1110
1111
1112
1113
1114
    # 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")

1115
        # `cloudpickle` allows pickling lambda functions directly
1116
        import cloudpickle
1117

1118
        input_bytes = cloudpickle.dumps((fn, output_filepath))
1119
1120
1121

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
1122
1123
1124
        returned = subprocess.run(
            _SUBPROCESS_COMMAND, input=input_bytes, capture_output=True
        )
1125
1126
1127
1128
1129
1130

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
1131
1132
1133
            raise RuntimeError(
                f"Error raised in subprocess:\n{returned.stderr.decode()}"
            ) from e
1134

1135
        with open(output_filepath, "rb") as f:
1136
1137
1138
1139
1140
1141
            return pickle.load(f)


def _run() -> None:
    # Setup plugins
    from vllm.plugins import load_general_plugins
1142

1143
1144
1145
1146
1147
    load_general_plugins()

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

    result = fn()
1148
1149
1150

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
1151
1152
1153


if __name__ == "__main__":
1154
    _run()