"tests/vscode:/vscode.git/clone" did not exist on "067c5d9ad1be71a36f7878631d97d7239dd097ab"
registry.py 52 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.tasks import ScoreType
34
from vllm.transformers_utils.dynamic_module import try_get_class_from_dynamic_module
35
from vllm.utils.hashing import safe_hash
36

37
38
if TYPE_CHECKING:
    from vllm.config.model import AttnTypeStr
39
    from vllm.config.pooler import SequencePoolingType, TokenPoolingType
40
41
else:
    AttnTypeStr = Any
42
43
    SequencePoolingType = Any
    TokenPoolingType = Any
44
45


46
47
48
49
50
from .interfaces import (
    has_inner_state,
    has_noops,
    is_attention_free,
    is_hybrid,
Patrick von Platen's avatar
Patrick von Platen committed
51
    requires_raw_input_tokens,
52
    supports_mamba_prefix_caching,
53
54
55
56
57
58
59
    supports_multimodal,
    supports_multimodal_encoder_tp_data,
    supports_multimodal_raw_input_only,
    supports_pp,
    supports_transcription,
)
from .interfaces_base import (
60
    get_attn_type,
61
62
    get_default_seq_pooling_type,
    get_default_tok_pooling_type,
63
    get_score_type,
64
65
66
    is_pooling_model,
    is_text_generation_model,
)
67
68
69

logger = init_logger(__name__)

70
71
_TEXT_GENERATION_MODELS = {
    # [Decoder-only]
72
    "AfmoeForCausalLM": ("afmoe", "AfmoeForCausalLM"),
73
    "ApertusForCausalLM": ("apertus", "ApertusForCausalLM"),
74
75
    "AquilaModel": ("llama", "LlamaForCausalLM"),
    "AquilaForCausalLM": ("llama", "LlamaForCausalLM"),  # AquilaChat2
Raghav Ravishankar's avatar
Raghav Ravishankar committed
76
    "ArceeForCausalLM": ("arcee", "ArceeForCausalLM"),
77
    "ArcticForCausalLM": ("arctic", "ArcticForCausalLM"),
78
    "AXK1ForCausalLM": ("AXK1", "AXK1ForCausalLM"),
79
80
81
82
    # baichuan-7b, upper case 'C' in the class name
    "BaiChuanForCausalLM": ("baichuan", "BaiChuanForCausalLM"),
    # baichuan-13b, lower case 'c' in the class name
    "BaichuanForCausalLM": ("baichuan", "BaichuanForCausalLM"),
83
    "BailingMoeForCausalLM": ("bailing_moe", "BailingMoeForCausalLM"),
ant-yy's avatar
ant-yy committed
84
    "BailingMoeV2ForCausalLM": ("bailing_moe", "BailingMoeV2ForCausalLM"),
Jiangyun Zhu's avatar
Jiangyun Zhu committed
85
    "BailingMoeV2_5ForCausalLM": ("bailing_moe_linear", "BailingMoeV25ForCausalLM"),
Yu Chin Fabian Lim's avatar
Yu Chin Fabian Lim committed
86
    "BambaForCausalLM": ("bamba", "BambaForCausalLM"),
87
    "BloomForCausalLM": ("bloom", "BloomForCausalLM"),
88
    "ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
89
    "ChatGLMForConditionalGeneration": ("chatglm", "ChatGLMForCausalLM"),
90
    "CohereForCausalLM": ("commandr", "CohereForCausalLM"),
91
    "Cohere2ForCausalLM": ("commandr", "CohereForCausalLM"),
92
    "CwmForCausalLM": ("llama", "LlamaForCausalLM"),
93
    "DbrxForCausalLM": ("dbrx", "DbrxForCausalLM"),
94
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
95
    "DeepseekForCausalLM": ("deepseek_v2", "DeepseekForCausalLM"),
96
    "DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
97
    "DeepseekV3ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
98
    "DeepseekV32ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
99
    "Dots1ForCausalLM": ("dots1", "Dots1ForCausalLM"),
100
    "Ernie4_5ForCausalLM": ("ernie45", "Ernie4_5ForCausalLM"),
101
    "Ernie4_5_MoeForCausalLM": ("ernie45_moe", "Ernie4_5_MoeForCausalLM"),
102
    "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
103
    "Exaone4ForCausalLM": ("exaone4", "Exaone4ForCausalLM"),
Kyungmin Lee's avatar
Kyungmin Lee committed
104
    "ExaoneMoEForCausalLM": ("exaone_moe", "ExaoneMoeForCausalLM"),
105
    "Fairseq2LlamaForCausalLM": ("fairseq2_llama", "Fairseq2LlamaForCausalLM"),
106
107
108
    "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
    "FalconH1ForCausalLM": ("falcon_h1", "FalconH1ForCausalLM"),
109
    "FlexOlmoForCausalLM": ("flex_olmo", "FlexOlmoForCausalLM"),
110
111
    "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
    "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
112
    "Gemma3ForCausalLM": ("gemma3", "Gemma3ForCausalLM"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
113
    "Gemma3nForCausalLM": ("gemma3n", "Gemma3nForCausalLM"),
114
    "Qwen3NextForCausalLM": ("qwen3_next", "Qwen3NextForCausalLM"),
115
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
116
    "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
117
    "Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
118
    "Glm4MoeLiteForCausalLM": ("glm4_moe_lite", "Glm4MoeLiteForCausalLM"),
Jee Jee Li's avatar
Jee Jee Li committed
119
    "GlmMoeDsaForCausalLM": ("deepseek_v2", "GlmMoeDsaForCausalLM"),
120
    "GptOssForCausalLM": ("gpt_oss", "GptOssForCausalLM"),
121
122
123
124
125
126
    "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
    "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
    "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
    "GPTNeoXForCausalLM": ("gpt_neox", "GPTNeoXForCausalLM"),
    "GraniteForCausalLM": ("granite", "GraniteForCausalLM"),
    "GraniteMoeForCausalLM": ("granitemoe", "GraniteMoeForCausalLM"),
127
128
    "GraniteMoeHybridForCausalLM": ("granitemoehybrid", "GraniteMoeHybridForCausalLM"),
    "GraniteMoeSharedForCausalLM": ("granitemoeshared", "GraniteMoeSharedForCausalLM"),
129
    "GritLM": ("gritlm", "GritLM"),
Bijaya Dangol's avatar
Bijaya Dangol committed
130
131
    "Grok1ModelForCausalLM": ("grok1", "GrokForCausalLM"),
    "Grok1ForCausalLM": ("grok1", "GrokForCausalLM"),
132
133
    "HunYuanMoEV1ForCausalLM": ("hunyuan_v1", "HunYuanMoEV1ForCausalLM"),
    "HunYuanDenseV1ForCausalLM": ("hunyuan_v1", "HunYuanDenseV1ForCausalLM"),
134
    "HCXVisionForCausalLM": ("hyperclovax_vision", "HCXVisionForCausalLM"),
135
    "HCXVisionV2ForCausalLM": ("hyperclovax_vision_v2", "HCXVisionV2ForCausalLM"),
136
    "HyperCLOVAXForCausalLM": ("hyperclovax", "HyperCLOVAXForCausalLM"),
137
138
    "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
    "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"),
139
    "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
140
    "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
141
142
    "IQuestCoderForCausalLM": ("llama", "LlamaForCausalLM"),
    "IQuestLoopCoderForCausalLM": ("iquest_loopcoder", "IQuestLoopCoderForCausalLM"),
143
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
144
    "Jais2ForCausalLM": ("jais2", "Jais2ForCausalLM"),
145
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
146
    "KimiLinearForCausalLM": ("kimi_linear", "KimiLinearForCausalLM"),
147
    "Lfm2ForCausalLM": ("lfm2", "Lfm2ForCausalLM"),
Paul Pak's avatar
Paul Pak committed
148
    "Lfm2MoeForCausalLM": ("lfm2_moe", "Lfm2MoeForCausalLM"),
149
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
150
    "Llama4ForCausalLM": ("llama4", "Llama4ForCausalLM"),
151
152
    # For decapoda-research/llama-*
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
XuruiYang's avatar
XuruiYang committed
153
    "LongcatFlashForCausalLM": ("longcat_flash", "LongcatFlashForCausalLM"),
154
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
155
    "Mamba2ForCausalLM": ("mamba2", "Mamba2ForCausalLM"),
156
157
    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
    "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
158
159
160
    "MiniMaxForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxText01ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxM1ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
161
    "MiniMaxM2ForCausalLM": ("minimax_m2", "MiniMaxM2ForCausalLM"),
162
    "MistralForCausalLM": ("mistral", "MistralForCausalLM"),
163
    "MistralLarge3ForCausalLM": ("mistral_large_3", "MistralLarge3ForCausalLM"),
164
165
166
167
    "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
    # transformers's mpt class has lower case
    "MptForCausalLM": ("mpt", "MPTForCausalLM"),
    "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
168
    "MiMoForCausalLM": ("mimo", "MiMoForCausalLM"),
169
    "MiMoV2FlashForCausalLM": ("mimo_v2_flash", "MiMoV2FlashForCausalLM"),
170
    "NemotronForCausalLM": ("nemotron", "NemotronForCausalLM"),
Luis Vega's avatar
Luis Vega committed
171
    "NemotronHForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
172
    "NemotronHPuzzleForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
173
    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
174
    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
175
    "Olmo3ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
176
    "OlmoHybridForCausalLM": ("olmo_hybrid", "OlmoHybridForCausalLM"),
177
178
179
    "OlmoeForCausalLM": ("olmoe", "OlmoeForCausalLM"),
    "OPTForCausalLM": ("opt", "OPTForCausalLM"),
    "OrionForCausalLM": ("orion", "OrionForCausalLM"),
180
    "OuroForCausalLM": ("ouro", "OuroForCausalLM"),
181
    "PanguEmbeddedForCausalLM": ("openpangu", "PanguEmbeddedForCausalLM"),
182
    "PanguProMoEV2ForCausalLM": ("openpangu", "PanguProMoEV2ForCausalLM"),
183
    "PanguUltraMoEForCausalLM": ("openpangu", "PanguUltraMoEForCausalLM"),
184
185
186
187
    "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
    "PhiForCausalLM": ("phi", "PhiForCausalLM"),
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
    "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
188
    "Plamo2ForCausalLM": ("plamo2", "Plamo2ForCausalLM"),
189
    "Plamo3ForCausalLM": ("plamo3", "Plamo3ForCausalLM"),
190
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
191
192
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
193
194
    "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),
    "Qwen3MoeForCausalLM": ("qwen3_moe", "Qwen3MoeForCausalLM"),
195
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
196
197
    "SarvamMoEForCausalLM": ("sarvam", "SarvamMoEForCausalLM"),
    "SarvamMLAForCausalLM": ("sarvam", "SarvamMLAForCausalLM"),
198
    "SeedOssForCausalLM": ("seed_oss", "SeedOssForCausalLM"),
Li Xie's avatar
Li Xie committed
199
    "Step1ForCausalLM": ("step1", "Step1ForCausalLM"),
Song's avatar
Song committed
200
    "Step3TextForCausalLM": ("step3_text", "Step3TextForCausalLM"),
csy0225's avatar
csy0225 committed
201
    "Step3p5ForCausalLM": ("step3p5", "Step3p5ForCausalLM"),
202
203
204
205
    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
206
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
207
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
208
    "TeleFLMForCausalLM": ("teleflm", "TeleFLMForCausalLM"),
209
    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
210
    "Zamba2ForCausalLM": ("zamba2", "Zamba2ForCausalLM"),
211
212
213
}

_EMBEDDING_MODELS = {
214
    # [Text-only]
215
    "BertModel": ("bert", "BertEmbeddingModel"),
216
    "BertSpladeSparseEmbeddingModel": ("bert", "BertSpladeSparseEmbeddingModel"),
217
    "ErnieModel": ("ernie", "ErnieEmbeddingModel"),
218
    "BgeM3EmbeddingModel": ("roberta", "BgeM3EmbeddingModel"),
219
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
220
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
221
    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
222
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
223
    "GritLM": ("gritlm", "GritLM"),
224
225
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
226
    "LlamaBidirectionalModel": ("llama", "LlamaBidirectionalModel"),
227
    "LlamaModel": ("llama", "LlamaForCausalLM"),
228
229
    **{
        # Multiple models share the same architecture, so we include them all
230
231
        k: (mod, arch)
        for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
232
233
        if arch == "LlamaForCausalLM"
    },
234
    "MistralModel": ("llama", "LlamaForCausalLM"),
235
    "ModernBertModel": ("modernbert", "ModernBertModel"),
236
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
237
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
238
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
239
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
240
241
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
242
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
243
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
chengchengpei's avatar
chengchengpei committed
244
245
246
247
    "VoyageQwen3BidirectionalEmbedModel": (
        "voyage",
        "VoyageQwen3BidirectionalEmbedModel",
    ),
248
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
249
    # [Multimodal]
250
    "CLIPModel": ("clip", "CLIPEmbeddingModel"),
251
    "ColPaliForRetrieval": ("colpali", "ColPaliModel"),
252
    "LlamaNemotronVLModel": ("nemotron_vl", "LlamaNemotronVLForEmbedding"),
253
254
255
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
256
    ),
257
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
258
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),
259
    "SiglipModel": ("siglip", "SiglipEmbeddingModel"),
260
261
    # Technically Terratorch models work on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
262
    # models for the time being.
263
264
    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
    "Terratorch": ("terratorch", "Terratorch"),
265
266
}

267
268
269
270
271
272
273
_LATE_INTERACTION_MODELS = {
    # [Text-only]
    "HF_ColBERT": ("colbert", "ColBERTModel"),
    "ColBERTModernBertModel": ("colbert", "ColBERTModernBertModel"),
    "ColBERTJinaRobertaModel": ("colbert", "ColBERTJinaRobertaModel"),
    # [Multimodal]
    "ColModernVBertForRetrieval": ("colmodernvbert", "ColModernVBertForRetrieval"),
274
    "ColPaliForRetrieval": ("colpali", "ColPaliModel"),
275
276
    "ColQwen3": ("colqwen3", "ColQwen3Model"),
    "OpsColQwen3Model": ("colqwen3", "ColQwen3Model"),
277
    "ColQwen3_5": ("colqwen3_5", "ColQwen3_5Model"),
278
279
280
281
282
283
284
285
286
287
    "Qwen3VLNemotronEmbedModel": ("colqwen3", "ColQwen3Model"),
}

_REWARD_MODELS = {
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
}

_TOKEN_CLASSIFICATION_MODELS = {
288
    "BertForTokenClassification": ("bert", "BertForTokenClassification"),
289
    "ErnieForTokenClassification": ("ernie", "ErnieForTokenClassification"),
290
291
292
293
294
295
296
297
298
    "ModernBertForTokenClassification": (
        "modernbert",
        "ModernBertForTokenClassification",
    ),
}

_SEQUENCE_CLASSIFICATION_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
299
    "ErnieForSequenceClassification": ("ernie", "ErnieForSequenceClassification"),
300
301
302
303
    "GteNewForSequenceClassification": (
        "bert_with_rope",
        "GteNewForSequenceClassification",
    ),
304
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),
305
306
307
    "LlamaBidirectionalForSequenceClassification": (
        "llama",
        "LlamaBidirectionalForSequenceClassification",
308
    ),
309
310
311
312
313
314
315
316
317
    "ModernBertForSequenceClassification": (
        "modernbert",
        "ModernBertForSequenceClassification",
    ),
    "RobertaForSequenceClassification": ("roberta", "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": (
        "roberta",
        "RobertaForSequenceClassification",
    ),
318
319
320
321
322
323
    # [Multimodal]
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"),
    "LlamaNemotronVLForSequenceClassification": (
        "nemotron_vl",
        "LlamaNemotronVLForSequenceClassification",
    ),
324
325
}

326
_MULTIMODAL_MODELS = {
327
    # [Decoder-only]
328
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
329
330
331
332
    "AudioFlamingo3ForConditionalGeneration": (
        "audioflamingo3",
        "AudioFlamingo3ForConditionalGeneration",
    ),
333
334
335
336
    "MusicFlamingoForConditionalGeneration": (
        "musicflamingo",
        "MusicFlamingoForConditionalGeneration",
    ),
337
338
339
    "AyaVisionForConditionalGeneration": (
        "aya_vision",
        "AyaVisionForConditionalGeneration",
340
    ),
341
    "BagelForConditionalGeneration": ("bagel", "BagelForConditionalGeneration"),
342
    "BeeForConditionalGeneration": ("bee", "BeeForConditionalGeneration"),
343
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
344
345
346
    "ChameleonForConditionalGeneration": (
        "chameleon",
        "ChameleonForConditionalGeneration",
347
    ),
348
349
350
    "Cohere2VisionForConditionalGeneration": (
        "cohere2_vision",
        "Cohere2VisionForConditionalGeneration",
351
    ),
352
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
353
    "DeepseekOCRForCausalLM": ("deepseek_ocr", "DeepseekOCRForCausalLM"),
RED's avatar
RED committed
354
    "DeepseekOCR2ForCausalLM": ("deepseek_ocr2", "DeepseekOCR2ForCausalLM"),
Roger Wang's avatar
Roger Wang committed
355
    "DotsOCRForCausalLM": ("dots_ocr", "DotsOCRForCausalLM"),
356
357
358
359
    "Eagle2_5_VLForConditionalGeneration": (
        "eagle2_5_vl",
        "Eagle2_5_VLForConditionalGeneration",
    ),
360
361
362
    "Ernie4_5_VLMoeForConditionalGeneration": (
        "ernie45_vl",
        "Ernie4_5_VLMoeForConditionalGeneration",
363
    ),
364
365
366
367
    "FireRedASR2ForConditionalGeneration": (
        "fireredasr2",
        "FireRedASR2ForConditionalGeneration",
    ),
368
    "FunASRForConditionalGeneration": ("funasr", "FunASRForConditionalGeneration"),
369
370
371
372
    "FunAudioChatForConditionalGeneration": (
        "funaudiochat",
        "FunAudioChatForConditionalGeneration",
    ),
373
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
374
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),
375
376
377
    "Gemma3nForConditionalGeneration": (
        "gemma3n_mm",
        "Gemma3nForConditionalGeneration",
378
    ),
379
    "GlmAsrForConditionalGeneration": ("glmasr", "GlmAsrForConditionalGeneration"),
380
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
381
382
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),
    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),
383
    "GlmOcrForConditionalGeneration": ("glm_ocr", "GlmOcrForConditionalGeneration"),
384
385
386
    "GraniteSpeechForConditionalGeneration": (
        "granite_speech",
        "GraniteSpeechForConditionalGeneration",
387
    ),
388
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
389
390
391
392
    "HunYuanVLForConditionalGeneration": (
        "hunyuan_vision",
        "HunYuanVLForConditionalGeneration",
    ),
393
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
394
395
396
    "InternS1ForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
397
    ),
398
399
400
    "InternVLForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
401
    ),
zxy's avatar
zxy committed
402
403
404
405
    "InternS1ProForConditionalGeneration": (
        "interns1_pro",
        "InternS1ProForConditionalGeneration",
    ),
406
407
408
409
    "Idefics3ForConditionalGeneration": (
        "idefics3",
        "Idefics3ForConditionalGeneration",
    ),
oscardev256's avatar
oscardev256 committed
410
    "IsaacForConditionalGeneration": ("isaac", "IsaacForConditionalGeneration"),
411
    "KananaVForConditionalGeneration": ("kanana_v", "KananaVForConditionalGeneration"),
412
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
413
414
415
    "KeyeVL1_5ForConditionalGeneration": (
        "keye_vl1_5",
        "KeyeVL1_5ForConditionalGeneration",
416
    ),
417
418
419
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),
    "KimiK25ForConditionalGeneration": ("kimi_k25", "KimiK25ForConditionalGeneration"),
    "MoonshotKimiaForCausalLM": ("kimi_audio", "KimiAudioForConditionalGeneration"),
420
421
422
423
    "LightOnOCRForConditionalGeneration": (
        "lightonocr",
        "LightOnOCRForConditionalGeneration",
    ),
424
    "Lfm2VlForConditionalGeneration": ("lfm2_vl", "Lfm2VLForConditionalGeneration"),
425
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),
426
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
427
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
428
429
430
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
431
    ),
432
433
434
    "LlavaNextVideoForConditionalGeneration": (
        "llava_next_video",
        "LlavaNextVideoForConditionalGeneration",
435
    ),
436
437
438
    "LlavaOnevisionForConditionalGeneration": (
        "llava_onevision",
        "LlavaOnevisionForConditionalGeneration",
439
    ),
440
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),
441
    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
442
443
444
    "MiniMaxVL01ForConditionalGeneration": (
        "minimax_vl_01",
        "MiniMaxVL01ForConditionalGeneration",
445
    ),
446
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
447
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
448
449
450
    "Mistral3ForConditionalGeneration": (
        "mistral3",
        "Mistral3ForConditionalGeneration",
451
    ),
452
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
453
    "Molmo2ForConditionalGeneration": ("molmo2", "Molmo2ForConditionalGeneration"),
454
    "NemotronH_Nano_VL_V2": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
455
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
456
    "OpenCUAForConditionalGeneration": ("opencua", "OpenCUAForConditionalGeneration"),
457
458
459
460
    "OpenPanguVLForConditionalGeneration": (
        "openpangu_vl",
        "OpenPanguVLForConditionalGeneration",
    ),
461
    "Ovis": ("ovis", "Ovis"),
462
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
463
464
    "Ovis2_6ForCausalLM": ("ovis2_5", "Ovis2_5"),
    "Ovis2_6_MoeForCausalLM": ("ovis2_5", "Ovis2_5"),
465
466
467
468
    "PaddleOCRVLForConditionalGeneration": (
        "paddleocr_vl",
        "PaddleOCRVLForConditionalGeneration",
    ),
469
470
471
472
    "PaliGemmaForConditionalGeneration": (
        "paligemma",
        "PaliGemmaForConditionalGeneration",
    ),
473
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
474
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
475
476
477
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),
478
479
480
    "Qwen2_5_VLForConditionalGeneration": (
        "qwen2_5_vl",
        "Qwen2_5_VLForConditionalGeneration",
481
    ),
482
483
484
    "Qwen2AudioForConditionalGeneration": (
        "qwen2_audio",
        "Qwen2AudioForConditionalGeneration",
485
    ),
486
487
488
    "Qwen2_5OmniModel": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
489
    ),
490
491
492
    "Qwen2_5OmniForConditionalGeneration": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
493
    ),
494
495
496
497
    "Qwen3OmniMoeForConditionalGeneration": (
        "qwen3_omni_moe_thinker",
        "Qwen3OmniMoeThinkerForConditionalGeneration",
    ),
Roger Wang's avatar
Roger Wang committed
498
499
500
501
    "Qwen3ASRForConditionalGeneration": (
        "qwen3_asr",
        "Qwen3ASRForConditionalGeneration",
    ),
502
503
    "Qwen3ASRRealtimeGeneration": ("qwen3_asr_realtime", "Qwen3ASRRealtimeGeneration"),
    "Qwen3VLForConditionalGeneration": ("qwen3_vl", "Qwen3VLForConditionalGeneration"),
504
505
506
    "Qwen3VLMoeForConditionalGeneration": (
        "qwen3_vl_moe",
        "Qwen3VLMoeForConditionalGeneration",
507
    ),
508
    "Qwen3_5ForConditionalGeneration": ("qwen3_5", "Qwen3_5ForConditionalGeneration"),
509
510
511
512
    "Qwen3_5MoeForConditionalGeneration": (
        "qwen3_5",
        "Qwen3_5MoeForConditionalGeneration",
    ),
513
    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
514
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
515
516
517
518
    "SmolVLMForConditionalGeneration": ("smolvlm", "SmolVLMForConditionalGeneration"),
    "StepVLForConditionalGeneration": ("step_vl", "StepVLForConditionalGeneration"),
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),
519
520
521
    "Tarsier2ForConditionalGeneration": (
        "qwen2_vl",
        "Tarsier2ForConditionalGeneration",
522
    ),
523
    "UltravoxModel": ("ultravox", "UltravoxModel"),
524
525
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),
    "VoxtralRealtimeGeneration": ("voxtral_realtime", "VoxtralRealtimeGeneration"),
526
    # [Encoder-decoder]
Ekagra Ranjan's avatar
Ekagra Ranjan committed
527
528
529
530
    "CohereASRForConditionalGeneration": (
        "cohere_asr",
        "CohereASRForConditionalGeneration",
    ),
531
532
533
534
    "NemotronParseForConditionalGeneration": (
        "nemotron_parse",
        "NemotronParseForConditionalGeneration",
    ),
535
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),
536
}
537
538

_SPECULATIVE_DECODING_MODELS = {
539
    "ExtractHiddenStatesModel": ("extract_hidden_states", "ExtractHiddenStatesModel"),
540
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
541
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
542
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
543
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
544
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
545
    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
546
    "Eagle3Qwen2_5vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
547
    "Eagle3Qwen3vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
548
549
550
551
    "EagleMistralLarge3ForCausalLM": (
        "mistral_large_3_eagle",
        "EagleMistralLarge3ForCausalLM",
    ),
552
553
    "Eagle3DeepseekV2ForCausalLM": ("deepseek_eagle3", "Eagle3DeepseekV2ForCausalLM"),
    "Eagle3DeepseekV3ForCausalLM": ("deepseek_eagle3", "Eagle3DeepseekV2ForCausalLM"),
554
    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
555
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
556
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
Kyungmin Lee's avatar
Kyungmin Lee committed
557
    "ExaoneMoeMTP": ("exaone_moe_mtp", "ExaoneMoeMTP"),
558
    "NemotronHMTPModel": ("nemotron_h_mtp", "NemotronHMTP"),
XuruiYang's avatar
XuruiYang committed
559
    "LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
560
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
561
    "Glm4MoeLiteMTPModel": ("glm4_moe_lite_mtp", "Glm4MoeLiteMTP"),
562
    "GlmOcrMTPModel": ("glm_ocr_mtp", "GlmOcrMTP"),
563
    "MedusaModel": ("medusa", "Medusa"),
564
    "OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
565
    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
csy0225's avatar
csy0225 committed
566
    "Step3p5MTP": ("step3p5_mtp", "Step3p5MTP"),
567
568
    "Qwen3_5MTP": ("qwen3_5_mtp", "Qwen3_5MTP"),
    "Qwen3_5MoeMTP": ("qwen3_5_mtp", "Qwen3_5MoeMTP"),
569
570
571
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
572
}
573

574
_TRANSFORMERS_SUPPORTED_MODELS = {
575
576
577
    # Text generation models
    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
    # Multimodal models
578
579
580
581
    "Emu3ForConditionalGeneration": (
        "transformers",
        "TransformersMultiModalForCausalLM",
    ),
582
583
584
}

_TRANSFORMERS_BACKEND_MODELS = {
585
    # Text generation models
586
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
    "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
605
    "TransformersForSequenceClassification": (
606
        "transformers",
607
        "TransformersForSequenceClassification",
608
    ),
609
    "TransformersMoEForSequenceClassification": (
610
        "transformers",
611
        "TransformersMoEForSequenceClassification",
612
    ),
613
614
615
    "TransformersMultiModalForSequenceClassification": (
        "transformers",
        "TransformersMultiModalForSequenceClassification",
616
    ),
617
}
618

619
_VLLM_MODELS = {
620
    **_TEXT_GENERATION_MODELS,
621
    **_EMBEDDING_MODELS,
622
623
624
625
    **_LATE_INTERACTION_MODELS,
    **_REWARD_MODELS,
    **_TOKEN_CLASSIFICATION_MODELS,
    **_SEQUENCE_CLASSIFICATION_MODELS,
626
    **_MULTIMODAL_MODELS,
627
    **_SPECULATIVE_DECODING_MODELS,
628
629
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
630
631
}

632
633
634
635
# 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.
636
_SUBPROCESS_COMMAND = [sys.executable, "-m", "vllm.model_executor.models.registry"]
637

638
_PREVIOUSLY_SUPPORTED_MODELS = {
639
    "MotifForCausalLM": "0.10.2",
640
    "Phi3SmallForCausalLM": "0.9.2",
641
    "Phi4FlashForCausalLM": "0.10.2",
642
    "Phi4MultimodalForCausalLM": "0.12.0",
643
644
645
646
647
    # encoder-decoder models except whisper
    # have been removed for V0 deprecation.
    "DonutForConditionalGeneration": "0.10.2",
    "MllamaForConditionalGeneration": "0.10.2",
}
648

649
650
651
652
653
654
655
_OOT_SUPPORTED_MODELS = {
    "BartModel": "https://github.com/vllm-project/bart-plugin",
    "BartForConditionalGeneration": "https://github.com/vllm-project/bart-plugin",
    "Florence2ForConditionalGeneration": "https://github.com/vllm-project/bart-plugin",
    "MBartForConditionalGeneration": "https://github.com/vllm-project/bart-plugin",
}

656

657
658
@dataclass(frozen=True)
class _ModelInfo:
659
    architecture: str
660
    is_text_generation_model: bool
661
    is_pooling_model: bool
662
    attn_type: AttnTypeStr
663
664
    default_seq_pooling_type: SequencePoolingType
    default_tok_pooling_type: TokenPoolingType
665
    score_type: ScoreType
666
    supports_multimodal: bool
667
    supports_multimodal_raw_input_only: bool
Patrick von Platen's avatar
Patrick von Platen committed
668
    requires_raw_input_tokens: bool
669
    supports_multimodal_encoder_tp_data: bool
670
    supports_pp: bool
671
672
    has_inner_state: bool
    is_attention_free: bool
673
    is_hybrid: bool
674
    has_noops: bool
675
    supports_mamba_prefix_caching: bool
676
    supports_transcription: bool
677
    supports_transcription_only: bool
678
679

    @staticmethod
680
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
681
        return _ModelInfo(
682
            architecture=model.__name__,
683
            is_text_generation_model=is_text_generation_model(model),
684
            is_pooling_model=is_pooling_model(model),
685
686
            default_seq_pooling_type=get_default_seq_pooling_type(model),
            default_tok_pooling_type=get_default_tok_pooling_type(model),
687
            attn_type=get_attn_type(model),
688
            score_type=get_score_type(model),
689
            supports_multimodal=supports_multimodal(model),
690
691
692
            supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
                model
            ),
Patrick von Platen's avatar
Patrick von Platen committed
693
            requires_raw_input_tokens=requires_raw_input_tokens(model),
694
695
696
            supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
                model
            ),
697
            supports_pp=supports_pp(model),
698
699
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
700
            is_hybrid=is_hybrid(model),
701
            supports_mamba_prefix_caching=supports_mamba_prefix_caching(model),
702
            supports_transcription=supports_transcription(model),
703
704
705
            supports_transcription_only=(
                supports_transcription(model) and model.supports_transcription_only
            ),
706
            has_noops=has_noops(model),
707
        )
708
709


710
711
712
713
class _BaseRegisteredModel(ABC):
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
714

715
    @abstractmethod
716
    def load_model_cls(self) -> type[nn.Module]:
717
        raise NotImplementedError
718
719


720
721
722
723
724
725
726
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
727
    model_cls: type[nn.Module]
728
729

    @staticmethod
730
    def from_model_cls(model_cls: type[nn.Module]):
731
732
733
734
735
736
737
738
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

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

739
    def load_model_cls(self) -> type[nn.Module]:
740
741
742
743
744
745
746
747
        return self.model_cls


@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """
748

749
750
751
    module_name: str
    class_name: str

752
753
754
755
756
757
758
759
    @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"

760
    def _load_modelinfo_from_cache(self, module_hash: str) -> _ModelInfo | None:
761
762
        try:
            try:
763
                modelinfo_path = self._get_cache_dir() / self._get_cache_filename()
764
765
766
                with open(modelinfo_path, encoding="utf-8") as file:
                    mi_dict = json.load(file)
            except FileNotFoundError:
767
                logger.debug(
768
                    "Cached model info file for class %s.%s not found",
769
770
771
                    self.module_name,
                    self.class_name,
                )
772
773
774
                return None

            if mi_dict["hash"] != module_hash:
775
                logger.debug(
776
                    "Cached model info file for class %s.%s is stale",
777
778
779
                    self.module_name,
                    self.class_name,
                )
780
781
782
783
784
                return None

            # file not changed, use cached _ModelInfo properties
            return _ModelInfo(**mi_dict["modelinfo"])
        except Exception:
785
            logger.debug(
786
                "Cached model info for class %s.%s error. ",
787
788
789
                self.module_name,
                self.class_name,
            )
790
791
            return None

792
    def _save_modelinfo_to_cache(self, mi: _ModelInfo, module_hash: str) -> None:
793
794
        """save dictionary json file to cache"""
        from vllm.model_executor.model_loader.weight_utils import atomic_writer
795

796
797
798
799
800
801
802
803
        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()
804
            with atomic_writer(modelinfo_path, encoding="utf-8") as f:
805
806
807
808
809
                json.dump(modelinfo_dict, f, indent=2)
        except Exception:
            logger.exception("Error saving model info cache.")

    @logtime(logger=logger, msg="Registry inspect model class")
810
    def inspect_model_cls(self) -> _ModelInfo:
811
        model_path = Path(__file__).parent / f"{self.module_name.split('.')[-1]}.py"
812
        module_hash = None
813

814
815
        if model_path.exists():
            with open(model_path, "rb") as f:
816
                module_hash = safe_hash(f.read(), usedforsecurity=False).hexdigest()
817
818
819

            mi = self._load_modelinfo_from_cache(module_hash)
            if mi is not None:
820
                logger.debug(
821
                    "Loaded model info for class %s.%s from cache",
822
823
824
                    self.module_name,
                    self.class_name,
                )
825
826
                return mi
            else:
827
                logger.debug(
828
                    "Cache model info for class %s.%s miss. Loading model instead.",
829
830
831
                    self.module_name,
                    self.class_name,
                )
832
833
834

        # Performed in another process to avoid initializing CUDA
        mi = _run_in_subprocess(
835
836
837
838
839
            lambda: _ModelInfo.from_model_cls(self.load_model_cls())
        )
        logger.debug(
            "Loaded model info for class %s.%s", self.module_name, self.class_name
        )
840
841

        # save cache file
842
843
        if module_hash is not None:
            self._save_modelinfo_to_cache(mi, module_hash)
844
845

        return mi
846

847
    def load_model_cls(self) -> type[nn.Module]:
848
849
850
851
852
853
854
855
        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,
856
) -> type[nn.Module] | None:
857
    from vllm.platforms import current_platform
858

859
    current_platform.verify_model_arch(model_arch)
860
861
862
    try:
        return model.load_model_cls()
    except Exception:
863
        logger.exception("Error in loading model architecture '%s'", model_arch)
864
        return None
865
866


867
868
869
870
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
871
) -> _ModelInfo | None:
872
873
874
    try:
        return model.inspect_model_cls()
    except Exception:
875
        logger.exception("Error in inspecting model architecture '%s'", model_arch)
876
        return None
877
878


879
880
881
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
882
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
883

884
    def get_supported_archs(self) -> Set[str]:
885
        return self.models.keys()
886

887
888
889
    def register_model(
        self,
        model_arch: str,
890
        model_cls: type[nn.Module] | str,
891
    ) -> None:
892
893
894
        """
        Register an external model to be used in vLLM.

895
        `model_cls` can be either:
896

897
        - A [`torch.nn.Module`][] class directly referencing the model.
898
        - A string in the format `<module>:<class>` which can be used to
899
900
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
901
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
902
        """
903
904
905
906
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

907
        if model_arch in self.models:
908
909
            logger.warning(
                "Model architecture %s is already registered, and will be "
910
911
912
913
                "overwritten by the new model class %s.",
                model_arch,
                model_cls,
            )
914
915
916
917
918
919

        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)
920

921
            model = _LazyRegisteredModel(*split_str)
922
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
923
            model = _RegisteredModel.from_model_cls(model_cls)
924
        else:
925
926
927
928
            msg = (
                "`model_cls` should be a string or PyTorch model class, "
                f"not a {type(model_arch)}"
            )
929
            raise TypeError(msg)
930

931
        self.models[model_arch] = model
932

933
    def _raise_for_unsupported(self, architectures: list[str]):
934
        all_supported_archs = self.get_supported_archs()
935

936
937
938
        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
939
940
                "to be inspected. Please check the logs for more details."
            )
941

942
943
944
945
946
947
948
949
        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 "
950
951
                    "use this model architecture."
                )
952
953
954
955
956
957
958
959
            if arch in _OOT_SUPPORTED_MODELS:
                plugin_url = _OOT_SUPPORTED_MODELS[arch]

                raise ValueError(
                    f"Model architecture {arch} is not supported in-tree anymore. "
                    f"Please install the plugin at {plugin_url} if you want to "
                    "use this model architecture."
                )
960

961
962
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
963
964
            f"Supported architectures: {all_supported_archs}"
        )
965

966
    def _try_load_model_cls(self, model_arch: str) -> type[nn.Module] | None:
967
968
        if model_arch not in self.models:
            return None
969

970
        return _try_load_model_cls(model_arch, self.models[model_arch])
971

972
    def _try_inspect_model_cls(self, model_arch: str) -> _ModelInfo | None:
973
974
        if model_arch not in self.models:
            return None
975

976
977
978
979
980
981
        return _try_inspect_model_cls(model_arch, self.models[model_arch])

    def _try_resolve_transformers(
        self,
        architecture: str,
        model_config: ModelConfig,
982
    ) -> str | None:
983
984
985
        if architecture in _TRANSFORMERS_BACKEND_MODELS:
            return architecture

986
987
988
        auto_map: dict[str, str] = (
            getattr(model_config.hf_config, "auto_map", None) or dict()
        )
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004

        # 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,
1005
                        trust_remote_code=model_config.trust_remote_code,
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
                        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,
1018
                        trust_remote_code=model_config.trust_remote_code,
1019
1020
1021
1022
1023
                        warn_on_fail=True,
                    )
                    if model_module is not None:
                        break
            else:
1024
                if model_config.model_impl != "transformers":
1025
1026
1027
1028
1029
1030
1031
                    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 "
1032
1033
                    "'auto_map' (relevant if the model is custom)."
                )
1034
1035

        if not model_module.is_backend_compatible():
1036
            if model_config.model_impl != "transformers":
1037
                return None
1038

1039
1040
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
1041
1042
                "is not compatible with vLLM."
            )
1043

1044
        return model_config._get_transformers_backend_cls()
1045

1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
    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
1071

1072
1073
    def inspect_model_cls(
        self,
1074
        architectures: str | list[str],
1075
        model_config: ModelConfig,
1076
    ) -> tuple[_ModelInfo, str]:
1077
1078
        if isinstance(architectures, str):
            architectures = [architectures]
1079
1080
        if not architectures:
            raise ValueError("No model architectures are specified")
1081
1082

        # Require transformers impl
1083
        if model_config.model_impl == "transformers":
1084
            arch = self._try_resolve_transformers(architectures[0], model_config)
1085
1086
1087
1088
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)
1089
        elif model_config.model_impl == "terratorch":
1090
1091
            model_info = self._try_inspect_model_cls("Terratorch")
            return (model_info, "Terratorch")
1092

1093
        # Fallback to transformers impl (after resolving convert_type)
1094
1095
1096
1097
1098
1099
        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)
1100
1101
1102
1103
1104
1105
1106
            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)
1107
            model_info = self._try_inspect_model_cls(normalized_arch)
1108
            if model_info is not None:
1109
                return (model_info, arch)
1110

1111
        # Fallback to transformers impl (before resolving runner_type)
1112
1113
1114
1115
1116
        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)
1117
1118
1119
1120
1121
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

1122
        return self._raise_for_unsupported(architectures)
1123

1124
1125
    def resolve_model_cls(
        self,
1126
        architectures: str | list[str],
1127
        model_config: ModelConfig,
1128
    ) -> tuple[type[nn.Module], str]:
1129
1130
        if isinstance(architectures, str):
            architectures = [architectures]
1131
1132
        if not architectures:
            raise ValueError("No model architectures are specified")
1133
1134

        # Require transformers impl
1135
        if model_config.model_impl == "transformers":
1136
            arch = self._try_resolve_transformers(architectures[0], model_config)
1137
1138
1139
1140
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)
1141
        elif model_config.model_impl == "terratorch":
1142
1143
1144
1145
            arch = "Terratorch"
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
1146

1147
        # Fallback to transformers impl (after resolving convert_type)
1148
1149
1150
1151
1152
1153
        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)
1154
1155
1156
1157
1158
1159
1160
            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)
1161
            model_cls = self._try_load_model_cls(normalized_arch)
1162
1163
            if model_cls is not None:
                return (model_cls, arch)
1164

1165
        # Fallback to transformers impl (before resolving runner_type)
1166
1167
1168
1169
1170
        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)
1171
1172
1173
1174
1175
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

1176
        return self._raise_for_unsupported(architectures)
1177

1178
1179
    def is_text_generation_model(
        self,
1180
        architectures: str | list[str],
1181
        model_config: ModelConfig,
1182
    ) -> bool:
1183
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1184
        return model_cls.is_text_generation_model
1185

1186
    def is_pooling_model(
1187
        self,
1188
        architectures: str | list[str],
1189
        model_config: ModelConfig,
1190
    ) -> bool:
1191
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1192
        return model_cls.is_pooling_model
1193
1194
1195

    def is_multimodal_model(
        self,
1196
        architectures: str | list[str],
1197
        model_config: ModelConfig,
1198
    ) -> bool:
1199
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1200
        return model_cls.supports_multimodal
1201

1202
    def is_multimodal_raw_input_only_model(
1203
        self,
1204
        architectures: str | list[str],
1205
        model_config: ModelConfig,
1206
    ) -> bool:
1207
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1208
        return model_cls.supports_multimodal_raw_input_only
1209

1210
1211
    def is_pp_supported_model(
        self,
1212
        architectures: str | list[str],
1213
        model_config: ModelConfig,
1214
    ) -> bool:
1215
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1216
        return model_cls.supports_pp
1217

1218
1219
    def model_has_inner_state(
        self,
1220
        architectures: str | list[str],
1221
        model_config: ModelConfig,
1222
    ) -> bool:
1223
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1224
        return model_cls.has_inner_state
1225

1226
1227
    def is_attention_free_model(
        self,
1228
        architectures: str | list[str],
1229
        model_config: ModelConfig,
1230
    ) -> bool:
1231
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1232
        return model_cls.is_attention_free
1233

1234
1235
    def is_hybrid_model(
        self,
1236
        architectures: str | list[str],
1237
        model_config: ModelConfig,
1238
    ) -> bool:
1239
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1240
1241
        return model_cls.is_hybrid

1242
1243
    def is_noops_model(
        self,
1244
        architectures: str | list[str],
1245
        model_config: ModelConfig,
1246
    ) -> bool:
1247
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1248
1249
        return model_cls.has_noops

1250
1251
    def is_transcription_model(
        self,
1252
        architectures: str | list[str],
1253
        model_config: ModelConfig,
1254
    ) -> bool:
1255
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1256
1257
        return model_cls.supports_transcription

1258
1259
    def is_transcription_only_model(
        self,
1260
        architectures: str | list[str],
1261
        model_config: ModelConfig,
1262
    ) -> bool:
1263
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1264
1265
        return model_cls.supports_transcription_only

1266

1267
1268
1269
1270
1271
1272
1273
1274
1275
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()
    }
)
1276
1277
1278
1279
1280

_T = TypeVar("_T")


def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
1281
1282
1283
1284
1285
    # 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")

1286
        # `cloudpickle` allows pickling lambda functions directly
1287
        import cloudpickle
1288

1289
        input_bytes = cloudpickle.dumps((fn, output_filepath))
1290
1291
1292

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
1293
1294
1295
        returned = subprocess.run(
            _SUBPROCESS_COMMAND, input=input_bytes, capture_output=True
        )
1296
1297
1298
1299
1300
1301

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
1302
1303
1304
            raise RuntimeError(
                f"Error raised in subprocess:\n{returned.stderr.decode()}"
            ) from e
1305

1306
        with open(output_filepath, "rb") as f:
1307
1308
1309
1310
1311
1312
            return pickle.load(f)


def _run() -> None:
    # Setup plugins
    from vllm.plugins import load_general_plugins
1313

1314
1315
1316
1317
1318
    load_general_plugins()

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

    result = fn()
1319
1320
1321

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
1322
1323
1324


if __name__ == "__main__":
1325
    _run()