registry.py 45.2 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
5
6
"""
Whenever you add an architecture to this page, please also update
`tests/models/registry.py` with example HuggingFace models for it.
"""
7

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

import torch.nn as nn
23
import transformers
24

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

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


44
45
46
47
48
49
from .interfaces import (
    has_inner_state,
    has_noops,
    is_attention_free,
    is_hybrid,
    supports_cross_encoding,
50
    supports_mamba_prefix_caching,
51
52
53
54
55
56
57
    supports_multimodal,
    supports_multimodal_encoder_tp_data,
    supports_multimodal_raw_input_only,
    supports_pp,
    supports_transcription,
)
from .interfaces_base import (
58
    get_attn_type,
59
60
61
62
    get_default_pooling_type,
    is_pooling_model,
    is_text_generation_model,
)
63
64
65

logger = init_logger(__name__)

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

_EMBEDDING_MODELS = {
190
    # [Text-only]
191
    "BertModel": ("bert", "BertEmbeddingModel"),
192
    "BertSpladeSparseEmbeddingModel": ("bert", "BertSpladeSparseEmbeddingModel"),
193
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
194
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
195
    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
196
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
197
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
198
    "GritLM": ("gritlm", "GritLM"),
199
200
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
201
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
202
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
203
    "LlamaModel": ("llama", "LlamaForCausalLM"),
204
205
    **{
        # Multiple models share the same architecture, so we include them all
206
207
        k: (mod, arch)
        for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
208
209
        if arch == "LlamaForCausalLM"
    },
210
    "MistralModel": ("llama", "LlamaForCausalLM"),
211
    "ModernBertModel": ("modernbert", "ModernBertModel"),
212
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
213
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
214
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
215
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
216
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
217
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
218
219
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
220
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
221
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
222
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
223
    # [Multimodal]
224
    "CLIPModel": ("clip", "CLIPEmbeddingModel"),
225
226
227
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
228
    ),
229
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
230
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
231
    "SiglipModel": ("siglip", "SiglipEmbeddingModel"),
232
233
    # Technically Terratorch models work on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
234
    # models for the time being.
235
236
    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
    "Terratorch": ("terratorch", "Terratorch"),
237
238
}

239
240
_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
241
    "BertForTokenClassification": ("bert", "BertForTokenClassification"),
242
243
244
245
246
247
248
249
    "GteNewForSequenceClassification": (
        "bert_with_rope",
        "GteNewForSequenceClassification",
    ),
    "ModernBertForSequenceClassification": (
        "modernbert",
        "ModernBertForSequenceClassification",
    ),
250
251
252
253
    "ModernBertForTokenClassification": (
        "modernbert",
        "ModernBertForTokenClassification",
    ),
254
255
256
257
258
    "RobertaForSequenceClassification": ("roberta", "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": (
        "roberta",
        "RobertaForSequenceClassification",
    ),
259
    # [Auto-converted (see adapters.py)]
260
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"),  # noqa: E501,
261
262
}

263
_MULTIMODAL_MODELS = {
264
    # [Decoder-only]
265
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
266
267
268
    "AyaVisionForConditionalGeneration": (
        "aya_vision",
        "AyaVisionForConditionalGeneration",
269
    ),
270
    "BeeForConditionalGeneration": ("bee", "BeeForConditionalGeneration"),
271
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
272
273
274
    "ChameleonForConditionalGeneration": (
        "chameleon",
        "ChameleonForConditionalGeneration",
275
    ),
276
277
278
    "Cohere2VisionForConditionalGeneration": (
        "cohere2_vision",
        "Cohere2VisionForConditionalGeneration",
279
    ),
280
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
281
    "DeepseekOCRForCausalLM": ("deepseek_ocr", "DeepseekOCRForCausalLM"),
Roger Wang's avatar
Roger Wang committed
282
    "DotsOCRForCausalLM": ("dots_ocr", "DotsOCRForCausalLM"),
283
284
285
    "Ernie4_5_VLMoeForConditionalGeneration": (
        "ernie45_vl",
        "Ernie4_5_VLMoeForConditionalGeneration",
286
    ),
287
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
288
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
289
290
291
    "Gemma3nForConditionalGeneration": (
        "gemma3n_mm",
        "Gemma3nForConditionalGeneration",
292
    ),
293
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
294
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),  # noqa: E501
Jee Jee Li's avatar
Jee Jee Li committed
295
    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),  # noqa: E501
296
297
298
    "GraniteSpeechForConditionalGeneration": (
        "granite_speech",
        "GraniteSpeechForConditionalGeneration",
299
    ),
300
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
301
302
303
304
    "HunYuanVLForConditionalGeneration": (
        "hunyuan_vision",
        "HunYuanVLForConditionalGeneration",
    ),
305
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
306
    "NemotronH_Nano_VL_V2": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
Zero's avatar
Zero committed
307
308
309
310
    "OpenCUAForConditionalGeneration": (
        "opencua",
        "OpenCUAForConditionalGeneration",
    ),
311
312
313
    "InternS1ForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
314
    ),
315
316
317
    "InternVLForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
318
    ),
319
320
321
322
323
    "Idefics3ForConditionalGeneration": (
        "idefics3",
        "Idefics3ForConditionalGeneration",
    ),
    "SmolVLMForConditionalGeneration": ("smolvlm", "SmolVLMForConditionalGeneration"),  # noqa: E501
324
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
325
326
327
    "KeyeVL1_5ForConditionalGeneration": (
        "keye_vl1_5",
        "KeyeVL1_5ForConditionalGeneration",
328
    ),
329
    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
330
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
331
332
333
334
    "LightOnOCRForConditionalGeneration": (
        "lightonocr",
        "LightOnOCRForConditionalGeneration",
    ),
335
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
336
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),  # noqa: E501
337
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
338
339
340
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
341
    ),
342
343
344
    "LlavaNextVideoForConditionalGeneration": (
        "llava_next_video",
        "LlavaNextVideoForConditionalGeneration",
345
    ),
346
347
348
    "LlavaOnevisionForConditionalGeneration": (
        "llava_onevision",
        "LlavaOnevisionForConditionalGeneration",
349
    ),
350
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
351
    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
352
353
354
    "MiniMaxVL01ForConditionalGeneration": (
        "minimax_vl_01",
        "MiniMaxVL01ForConditionalGeneration",
355
    ),
356
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
357
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
358
359
360
    "Mistral3ForConditionalGeneration": (
        "mistral3",
        "Mistral3ForConditionalGeneration",
361
    ),
362
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
363
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
364
    "Ovis": ("ovis", "Ovis"),
365
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
366
367
368
369
    "PaddleOCRVLForConditionalGeneration": (
        "paddleocr_vl",
        "PaddleOCRVLForConditionalGeneration",
    ),
370
371
372
373
    "PaliGemmaForConditionalGeneration": (
        "paligemma",
        "PaliGemmaForConditionalGeneration",
    ),
374
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
375
376
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
    "Phi4MultimodalForCausalLM": ("phi4_multimodal", "Phi4MultimodalForCausalLM"),  # noqa: E501
377
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
378
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),  # noqa: E501
379
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
380
381
382
    "Qwen2_5_VLForConditionalGeneration": (
        "qwen2_5_vl",
        "Qwen2_5_VLForConditionalGeneration",
383
    ),
384
385
386
    "Qwen2AudioForConditionalGeneration": (
        "qwen2_audio",
        "Qwen2AudioForConditionalGeneration",
387
    ),
388
389
390
    "Qwen2_5OmniModel": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
391
    ),
392
393
394
    "Qwen2_5OmniForConditionalGeneration": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
395
    ),
396
397
398
399
    "Qwen3OmniMoeForConditionalGeneration": (
        "qwen3_omni_moe_thinker",
        "Qwen3OmniMoeThinkerForConditionalGeneration",
    ),
400
    "Qwen3VLForConditionalGeneration": ("qwen3_vl", "Qwen3VLForConditionalGeneration"),  # noqa: E501
401
402
403
    "Qwen3VLMoeForConditionalGeneration": (
        "qwen3_vl_moe",
        "Qwen3VLMoeForConditionalGeneration",
404
    ),
405
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
Song's avatar
Song committed
406
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),  # noqa: E501
汪志鹏's avatar
汪志鹏 committed
407
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),  # noqa: E501
408
409
410
    "Tarsier2ForConditionalGeneration": (
        "qwen2_vl",
        "Tarsier2ForConditionalGeneration",
411
    ),
412
    "UltravoxModel": ("ultravox", "UltravoxModel"),
Patrick von Platen's avatar
Patrick von Platen committed
413
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),  # noqa: E501
414
    # [Encoder-decoder]
415
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
416
}
417
418

_SPECULATIVE_DECODING_MODELS = {
419
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
420
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
421
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
422
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
423
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
424
    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
425
    "Eagle3Qwen2_5vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
426
    "Eagle3Qwen3vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
427
    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
428
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
429
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
XuruiYang's avatar
XuruiYang committed
430
    "LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
431
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
432
    "MedusaModel": ("medusa", "Medusa"),
433
    "OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
434
    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
435
436
437
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
438
}
439

440
_TRANSFORMERS_SUPPORTED_MODELS = {
441
442
443
    # Text generation models
    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
    # Multimodal models
444
445
446
447
    "Emu3ForConditionalGeneration": (
        "transformers",
        "TransformersMultiModalForCausalLM",
    ),
448
449
450
}

_TRANSFORMERS_BACKEND_MODELS = {
451
    # Text generation models
452
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
    "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
471
    "TransformersForSequenceClassification": (
472
        "transformers",
473
        "TransformersForSequenceClassification",
474
    ),
475
    "TransformersMoEForSequenceClassification": (
476
        "transformers",
477
        "TransformersMoEForSequenceClassification",
478
    ),
479
480
481
    "TransformersMultiModalForSequenceClassification": (
        "transformers",
        "TransformersMultiModalForSequenceClassification",
482
    ),
483
}
484

485
_VLLM_MODELS = {
486
    **_TEXT_GENERATION_MODELS,
487
    **_EMBEDDING_MODELS,
488
    **_CROSS_ENCODER_MODELS,
489
    **_MULTIMODAL_MODELS,
490
    **_SPECULATIVE_DECODING_MODELS,
491
492
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
493
494
}

495
496
497
498
# 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.
499
_SUBPROCESS_COMMAND = [sys.executable, "-m", "vllm.model_executor.models.registry"]
500

501
_PREVIOUSLY_SUPPORTED_MODELS = {
502
    "MotifForCausalLM": "0.10.2",
503
    "Phi3SmallForCausalLM": "0.9.2",
504
    "Phi4FlashForCausalLM": "0.10.2",
505
506
507
508
509
510
511
512
513
    # 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",
}
514

515

516
517
@dataclass(frozen=True)
class _ModelInfo:
518
    architecture: str
519
    is_text_generation_model: bool
520
    is_pooling_model: bool
521
522
    attn_type: AttnTypeStr
    default_pooling_type: PoolingTypeStr
523
    supports_cross_encoding: bool
524
    supports_multimodal: bool
525
    supports_multimodal_raw_input_only: bool
526
    supports_multimodal_encoder_tp_data: bool
527
    supports_pp: bool
528
529
    has_inner_state: bool
    is_attention_free: bool
530
    is_hybrid: bool
531
    has_noops: bool
532
    supports_mamba_prefix_caching: bool
533
    supports_transcription: bool
534
    supports_transcription_only: bool
535
536

    @staticmethod
537
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
538
        return _ModelInfo(
539
            architecture=model.__name__,
540
            is_text_generation_model=is_text_generation_model(model),
541
            is_pooling_model=is_pooling_model(model),
542
            default_pooling_type=get_default_pooling_type(model),
543
            attn_type=get_attn_type(model),
544
            supports_cross_encoding=supports_cross_encoding(model),
545
            supports_multimodal=supports_multimodal(model),
546
547
548
549
550
551
            supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
                model
            ),
            supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
                model
            ),
552
            supports_pp=supports_pp(model),
553
554
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
555
            is_hybrid=is_hybrid(model),
556
            supports_mamba_prefix_caching=supports_mamba_prefix_caching(model),
557
            supports_transcription=supports_transcription(model),
558
559
560
            supports_transcription_only=(
                supports_transcription(model) and model.supports_transcription_only
            ),
561
            has_noops=has_noops(model),
562
        )
563
564


565
566
567
568
class _BaseRegisteredModel(ABC):
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
569

570
    @abstractmethod
571
    def load_model_cls(self) -> type[nn.Module]:
572
        raise NotImplementedError
573
574


575
576
577
578
579
580
581
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
582
    model_cls: type[nn.Module]
583
584

    @staticmethod
585
    def from_model_cls(model_cls: type[nn.Module]):
586
587
588
589
590
591
592
593
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

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

594
    def load_model_cls(self) -> type[nn.Module]:
595
596
597
598
599
600
601
602
        return self.model_cls


@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """
603

604
605
606
    module_name: str
    class_name: str

607
608
609
610
611
612
613
614
    @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"

615
    def _load_modelinfo_from_cache(self, module_hash: str) -> _ModelInfo | None:
616
617
        try:
            try:
618
                modelinfo_path = self._get_cache_dir() / self._get_cache_filename()
619
620
621
                with open(modelinfo_path, encoding="utf-8") as file:
                    mi_dict = json.load(file)
            except FileNotFoundError:
622
                logger.debug(
623
                    "Cached model info file for class %s.%s not found",
624
625
626
                    self.module_name,
                    self.class_name,
                )
627
628
629
                return None

            if mi_dict["hash"] != module_hash:
630
                logger.debug(
631
                    "Cached model info file for class %s.%s is stale",
632
633
634
                    self.module_name,
                    self.class_name,
                )
635
636
637
638
639
                return None

            # file not changed, use cached _ModelInfo properties
            return _ModelInfo(**mi_dict["modelinfo"])
        except Exception:
640
            logger.debug(
641
                "Cached model info for class %s.%s error. ",
642
643
644
                self.module_name,
                self.class_name,
            )
645
646
            return None

647
    def _save_modelinfo_to_cache(self, mi: _ModelInfo, module_hash: str) -> None:
648
649
        """save dictionary json file to cache"""
        from vllm.model_executor.model_loader.weight_utils import atomic_writer
650

651
652
653
654
655
656
657
658
        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()
659
            with atomic_writer(modelinfo_path, encoding="utf-8") as f:
660
661
662
663
664
                json.dump(modelinfo_dict, f, indent=2)
        except Exception:
            logger.exception("Error saving model info cache.")

    @logtime(logger=logger, msg="Registry inspect model class")
665
    def inspect_model_cls(self) -> _ModelInfo:
666
        model_path = Path(__file__).parent / f"{self.module_name.split('.')[-1]}.py"
667
        module_hash = None
668

669
670
        if model_path.exists():
            with open(model_path, "rb") as f:
671
                module_hash = safe_hash(f.read(), usedforsecurity=False).hexdigest()
672
673
674

            mi = self._load_modelinfo_from_cache(module_hash)
            if mi is not None:
675
                logger.debug(
676
                    "Loaded model info for class %s.%s from cache",
677
678
679
                    self.module_name,
                    self.class_name,
                )
680
681
                return mi
            else:
682
                logger.debug(
683
                    "Cache model info for class %s.%s miss. Loading model instead.",
684
685
686
                    self.module_name,
                    self.class_name,
                )
687
688
689

        # Performed in another process to avoid initializing CUDA
        mi = _run_in_subprocess(
690
691
692
693
694
            lambda: _ModelInfo.from_model_cls(self.load_model_cls())
        )
        logger.debug(
            "Loaded model info for class %s.%s", self.module_name, self.class_name
        )
695
696

        # save cache file
697
698
        if module_hash is not None:
            self._save_modelinfo_to_cache(mi, module_hash)
699
700

        return mi
701

702
    def load_model_cls(self) -> type[nn.Module]:
703
704
705
706
707
708
709
710
        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,
711
) -> type[nn.Module] | None:
712
    from vllm.platforms import current_platform
713

714
    current_platform.verify_model_arch(model_arch)
715
716
717
    try:
        return model.load_model_cls()
    except Exception:
718
        logger.exception("Error in loading model architecture '%s'", model_arch)
719
        return None
720
721


722
723
724
725
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
726
) -> _ModelInfo | None:
727
728
729
    try:
        return model.inspect_model_cls()
    except Exception:
730
        logger.exception("Error in inspecting model architecture '%s'", model_arch)
731
        return None
732
733


734
735
736
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
737
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
738

739
    def get_supported_archs(self) -> Set[str]:
740
        return self.models.keys()
741

742
743
744
    def register_model(
        self,
        model_arch: str,
745
        model_cls: type[nn.Module] | str,
746
    ) -> None:
747
748
749
        """
        Register an external model to be used in vLLM.

750
        `model_cls` can be either:
751

752
        - A [`torch.nn.Module`][] class directly referencing the model.
753
        - A string in the format `<module>:<class>` which can be used to
754
755
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
756
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
757
        """
758
759
760
761
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

762
        if model_arch in self.models:
763
764
            logger.warning(
                "Model architecture %s is already registered, and will be "
765
766
767
768
                "overwritten by the new model class %s.",
                model_arch,
                model_cls,
            )
769
770
771
772
773
774

        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)
775

776
            model = _LazyRegisteredModel(*split_str)
777
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
778
            model = _RegisteredModel.from_model_cls(model_cls)
779
        else:
780
781
782
783
            msg = (
                "`model_cls` should be a string or PyTorch model class, "
                f"not a {type(model_arch)}"
            )
784
            raise TypeError(msg)
785

786
        self.models[model_arch] = model
787

788
    def _raise_for_unsupported(self, architectures: list[str]):
789
        all_supported_archs = self.get_supported_archs()
790

791
792
793
        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
794
795
                "to be inspected. Please check the logs for more details."
            )
796

797
798
799
800
801
802
803
804
        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 "
805
806
                    "use this model architecture."
                )
807

808
809
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
810
811
            f"Supported architectures: {all_supported_archs}"
        )
812

813
    def _try_load_model_cls(self, model_arch: str) -> type[nn.Module] | None:
814
815
        if model_arch not in self.models:
            return None
816

817
        return _try_load_model_cls(model_arch, self.models[model_arch])
818

819
    def _try_inspect_model_cls(self, model_arch: str) -> _ModelInfo | None:
820
821
        if model_arch not in self.models:
            return None
822

823
824
825
826
827
828
        return _try_inspect_model_cls(model_arch, self.models[model_arch])

    def _try_resolve_transformers(
        self,
        architecture: str,
        model_config: ModelConfig,
829
    ) -> str | None:
830
831
832
        if architecture in _TRANSFORMERS_BACKEND_MODELS:
            return architecture

833
834
835
        auto_map: dict[str, str] = (
            getattr(model_config.hf_config, "auto_map", None) or dict()
        )
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868

        # 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:
869
                if model_config.model_impl != "transformers":
870
871
872
873
874
875
876
                    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 "
877
878
                    "'auto_map' (relevant if the model is custom)."
                )
879
880

        if not model_module.is_backend_compatible():
881
            if model_config.model_impl != "transformers":
882
                return None
883

884
885
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
886
887
                "is not compatible with vLLM."
            )
888

889
        return model_config._get_transformers_backend_cls()
890

891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
    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
916

917
918
    def inspect_model_cls(
        self,
919
        architectures: str | list[str],
920
        model_config: ModelConfig,
921
    ) -> tuple[_ModelInfo, str]:
922
923
        if isinstance(architectures, str):
            architectures = [architectures]
924
925
        if not architectures:
            raise ValueError("No model architectures are specified")
926
927

        # Require transformers impl
928
        if model_config.model_impl == "transformers":
929
            arch = self._try_resolve_transformers(architectures[0], model_config)
930
931
932
933
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)
934
        elif model_config.model_impl == "terratorch":
935
936
            model_info = self._try_inspect_model_cls("Terratorch")
            return (model_info, "Terratorch")
937

938
        # Fallback to transformers impl (after resolving convert_type)
939
940
941
942
943
944
        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)
945
946
947
948
949
950
951
            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)
952
            model_info = self._try_inspect_model_cls(normalized_arch)
953
            if model_info is not None:
954
                return (model_info, arch)
955

956
        # Fallback to transformers impl (before resolving runner_type)
957
958
959
960
961
        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)
962
963
964
965
966
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

967
        return self._raise_for_unsupported(architectures)
968

969
970
    def resolve_model_cls(
        self,
971
        architectures: str | list[str],
972
        model_config: ModelConfig,
973
    ) -> tuple[type[nn.Module], str]:
974
975
        if isinstance(architectures, str):
            architectures = [architectures]
976
977
        if not architectures:
            raise ValueError("No model architectures are specified")
978
979

        # Require transformers impl
980
        if model_config.model_impl == "transformers":
981
            arch = self._try_resolve_transformers(architectures[0], model_config)
982
983
984
985
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)
986
        elif model_config.model_impl == "terratorch":
987
988
989
990
            arch = "Terratorch"
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
991

992
        # Fallback to transformers impl (after resolving convert_type)
993
994
995
996
997
998
        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)
999
1000
1001
1002
1003
1004
1005
            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)
1006
            model_cls = self._try_load_model_cls(normalized_arch)
1007
1008
            if model_cls is not None:
                return (model_cls, arch)
1009

1010
        # Fallback to transformers impl (before resolving runner_type)
1011
1012
1013
1014
1015
        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)
1016
1017
1018
1019
1020
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

1021
        return self._raise_for_unsupported(architectures)
1022

1023
1024
    def is_text_generation_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.is_text_generation_model
1030

1031
    def is_pooling_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.is_pooling_model
1038

1039
1040
    def is_cross_encoder_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_cross_encoding
1046

1047
1048
    def is_multimodal_model(
        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.supports_multimodal
1054

1055
    def is_multimodal_raw_input_only_model(
1056
        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.supports_multimodal_raw_input_only
1062

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

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

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

1087
1088
    def is_hybrid_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.is_hybrid

1095
1096
    def is_noops_model(
        self,
1097
        architectures: str | list[str],
1098
        model_config: ModelConfig,
1099
    ) -> bool:
1100
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1101
1102
        return model_cls.has_noops

1103
1104
    def is_transcription_model(
        self,
1105
        architectures: str | list[str],
1106
        model_config: ModelConfig,
1107
    ) -> bool:
1108
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1109
1110
        return model_cls.supports_transcription

1111
1112
    def is_transcription_only_model(
        self,
1113
        architectures: str | list[str],
1114
        model_config: ModelConfig,
1115
    ) -> bool:
1116
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1117
1118
        return model_cls.supports_transcription_only

1119

1120
1121
1122
1123
1124
1125
1126
1127
1128
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()
    }
)
1129
1130
1131
1132
1133

_T = TypeVar("_T")


def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
1134
1135
1136
1137
1138
    # 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")

1139
        # `cloudpickle` allows pickling lambda functions directly
1140
        import cloudpickle
1141

1142
        input_bytes = cloudpickle.dumps((fn, output_filepath))
1143
1144
1145

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
1146
1147
1148
        returned = subprocess.run(
            _SUBPROCESS_COMMAND, input=input_bytes, capture_output=True
        )
1149
1150
1151
1152
1153
1154

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
1155
1156
1157
            raise RuntimeError(
                f"Error raised in subprocess:\n{returned.stderr.decode()}"
            ) from e
1158

1159
        with open(output_filepath, "rb") as f:
1160
1161
1162
1163
1164
1165
            return pickle.load(f)


def _run() -> None:
    # Setup plugins
    from vllm.plugins import load_general_plugins
1166

1167
1168
1169
1170
1171
    load_general_plugins()

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

    result = fn()
1172
1173
1174

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
1175
1176
1177


if __name__ == "__main__":
1178
    _run()