registry.py 53.1 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
5
6
"""
Whenever you add an architecture to this page, please also update
`tests/models/registry.py` with example HuggingFace models for it.
"""
7

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

import torch.nn as nn
23
import transformers
24

25
from vllm import envs
26
27
28
29
30
from vllm.config import (
    ModelConfig,
    iter_architecture_defaults,
    try_match_architecture_defaults,
)
31
from vllm.logger import init_logger
32
from vllm.logging_utils import logtime
33
from vllm.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
    "Gemma4ForCausalLM": ("gemma4", "Gemma4ForCausalLM"),
115
    "Qwen3NextForCausalLM": ("qwen3_next", "Qwen3NextForCausalLM"),
116
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
117
    "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
118
    "Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
119
    "Glm4MoeLiteForCausalLM": ("glm4_moe_lite", "Glm4MoeLiteForCausalLM"),
Jee Jee Li's avatar
Jee Jee Li committed
120
    "GlmMoeDsaForCausalLM": ("deepseek_v2", "GlmMoeDsaForCausalLM"),
121
    "GptOssForCausalLM": ("gpt_oss", "GptOssForCausalLM"),
122
123
124
125
126
127
    "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
    "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
    "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
    "GPTNeoXForCausalLM": ("gpt_neox", "GPTNeoXForCausalLM"),
    "GraniteForCausalLM": ("granite", "GraniteForCausalLM"),
    "GraniteMoeForCausalLM": ("granitemoe", "GraniteMoeForCausalLM"),
128
129
    "GraniteMoeHybridForCausalLM": ("granitemoehybrid", "GraniteMoeHybridForCausalLM"),
    "GraniteMoeSharedForCausalLM": ("granitemoeshared", "GraniteMoeSharedForCausalLM"),
130
    "GritLM": ("gritlm", "GritLM"),
Bijaya Dangol's avatar
Bijaya Dangol committed
131
132
    "Grok1ModelForCausalLM": ("grok1", "GrokForCausalLM"),
    "Grok1ForCausalLM": ("grok1", "GrokForCausalLM"),
133
134
    "HunYuanMoEV1ForCausalLM": ("hunyuan_v1", "HunYuanMoEV1ForCausalLM"),
    "HunYuanDenseV1ForCausalLM": ("hunyuan_v1", "HunYuanDenseV1ForCausalLM"),
135
    "HCXVisionForCausalLM": ("hyperclovax_vision", "HCXVisionForCausalLM"),
136
    "HCXVisionV2ForCausalLM": ("hyperclovax_vision_v2", "HCXVisionV2ForCausalLM"),
137
    "HyperCLOVAXForCausalLM": ("hyperclovax", "HyperCLOVAXForCausalLM"),
138
139
    "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
    "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"),
140
    "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
141
    "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
142
143
    "IQuestCoderForCausalLM": ("llama", "LlamaForCausalLM"),
    "IQuestLoopCoderForCausalLM": ("iquest_loopcoder", "IQuestLoopCoderForCausalLM"),
144
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
145
    "Jais2ForCausalLM": ("jais2", "Jais2ForCausalLM"),
146
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
147
    "KimiLinearForCausalLM": ("kimi_linear", "KimiLinearForCausalLM"),
148
    "Lfm2ForCausalLM": ("lfm2", "Lfm2ForCausalLM"),
Paul Pak's avatar
Paul Pak committed
149
    "Lfm2MoeForCausalLM": ("lfm2_moe", "Lfm2MoeForCausalLM"),
150
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
151
    "Llama4ForCausalLM": ("llama4", "Llama4ForCausalLM"),
152
153
    # For decapoda-research/llama-*
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
XuruiYang's avatar
XuruiYang committed
154
    "LongcatFlashForCausalLM": ("longcat_flash", "LongcatFlashForCausalLM"),
155
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
156
    "Mamba2ForCausalLM": ("mamba2", "Mamba2ForCausalLM"),
157
158
    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
    "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
159
160
161
    "MiniMaxForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxText01ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
    "MiniMaxM1ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
162
    "MiniMaxM2ForCausalLM": ("minimax_m2", "MiniMaxM2ForCausalLM"),
163
    "MistralForCausalLM": ("mistral", "MistralForCausalLM"),
164
    "MistralLarge3ForCausalLM": ("mistral_large_3", "MistralLarge3ForCausalLM"),
165
166
167
168
    "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
    # transformers's mpt class has lower case
    "MptForCausalLM": ("mpt", "MPTForCausalLM"),
    "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
169
    "MiMoForCausalLM": ("mimo", "MiMoForCausalLM"),
170
    "MiMoV2FlashForCausalLM": ("mimo_v2_flash", "MiMoV2FlashForCausalLM"),
171
    "NemotronForCausalLM": ("nemotron", "NemotronForCausalLM"),
Luis Vega's avatar
Luis Vega committed
172
    "NemotronHForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
173
    "NemotronHPuzzleForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
174
    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
175
    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
176
    "Olmo3ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
177
    "OlmoHybridForCausalLM": ("olmo_hybrid", "OlmoHybridForCausalLM"),
178
179
180
    "OlmoeForCausalLM": ("olmoe", "OlmoeForCausalLM"),
    "OPTForCausalLM": ("opt", "OPTForCausalLM"),
    "OrionForCausalLM": ("orion", "OrionForCausalLM"),
181
    "OuroForCausalLM": ("ouro", "OuroForCausalLM"),
182
    "PanguEmbeddedForCausalLM": ("openpangu", "PanguEmbeddedForCausalLM"),
183
    "PanguProMoEV2ForCausalLM": ("openpangu", "PanguProMoEV2ForCausalLM"),
184
    "PanguUltraMoEForCausalLM": ("openpangu", "PanguUltraMoEForCausalLM"),
185
    "Param2MoEForCausalLM": ("param2moe", "Param2MoEForCausalLM"),
186
187
188
189
    "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
    "PhiForCausalLM": ("phi", "PhiForCausalLM"),
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
    "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
190
    "Plamo2ForCausalLM": ("plamo2", "Plamo2ForCausalLM"),
191
    "Plamo3ForCausalLM": ("plamo3", "Plamo3ForCausalLM"),
192
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
193
194
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
195
196
    "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),
    "Qwen3MoeForCausalLM": ("qwen3_moe", "Qwen3MoeForCausalLM"),
197
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
198
199
    "SarvamMoEForCausalLM": ("sarvam", "SarvamMoEForCausalLM"),
    "SarvamMLAForCausalLM": ("sarvam", "SarvamMLAForCausalLM"),
200
    "SeedOssForCausalLM": ("seed_oss", "SeedOssForCausalLM"),
Li Xie's avatar
Li Xie committed
201
    "Step1ForCausalLM": ("step1", "Step1ForCausalLM"),
Song's avatar
Song committed
202
    "Step3TextForCausalLM": ("step3_text", "Step3TextForCausalLM"),
csy0225's avatar
csy0225 committed
203
    "Step3p5ForCausalLM": ("step3p5", "Step3p5ForCausalLM"),
204
205
206
207
    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
208
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
209
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
210
    "TeleChat3ForCausalLM": ("llama", "LlamaForCausalLM"),
211
    "TeleFLMForCausalLM": ("teleflm", "TeleFLMForCausalLM"),
212
    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
213
    "Zamba2ForCausalLM": ("zamba2", "Zamba2ForCausalLM"),
214
215
216
}

_EMBEDDING_MODELS = {
217
    # [Text-only]
218
    "BertModel": ("bert", "BertEmbeddingModel"),
219
    "BertSpladeSparseEmbeddingModel": ("bert", "BertSpladeSparseEmbeddingModel"),
220
    "ErnieModel": ("ernie", "ErnieEmbeddingModel"),
221
    "BgeM3EmbeddingModel": ("roberta", "BgeM3EmbeddingModel"),
222
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
223
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
224
    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
225
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
226
    "GritLM": ("gritlm", "GritLM"),
227
228
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
229
    "LlamaBidirectionalModel": ("llama", "LlamaBidirectionalModel"),
230
    "LlamaModel": ("llama", "LlamaForCausalLM"),
231
232
    **{
        # Multiple models share the same architecture, so we include them all
233
234
        k: (mod, arch)
        for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
235
236
        if arch == "LlamaForCausalLM"
    },
237
    "MistralModel": ("llama", "LlamaForCausalLM"),
238
    "ModernBertModel": ("modernbert", "ModernBertModel"),
239
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
240
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
241
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
242
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
243
244
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
245
    "TeleChatForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
246
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
chengchengpei's avatar
chengchengpei committed
247
248
249
250
    "VoyageQwen3BidirectionalEmbedModel": (
        "voyage",
        "VoyageQwen3BidirectionalEmbedModel",
    ),
251
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
252
    # [Multimodal]
253
    "CLIPModel": ("clip", "CLIPEmbeddingModel"),
254
    "ColPaliForRetrieval": ("colpali", "ColPaliModel"),
255
    "LlamaNemotronVLModel": ("nemotron_vl", "LlamaNemotronVLForEmbedding"),
256
257
258
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
259
    ),
260
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
261
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),
262
    "SiglipModel": ("siglip", "SiglipEmbeddingModel"),
263
264
    # Technically Terratorch models work on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
265
    # models for the time being.
266
267
    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
    "Terratorch": ("terratorch", "Terratorch"),
268
269
}

270
271
272
273
274
_LATE_INTERACTION_MODELS = {
    # [Text-only]
    "HF_ColBERT": ("colbert", "ColBERTModel"),
    "ColBERTModernBertModel": ("colbert", "ColBERTModernBertModel"),
    "ColBERTJinaRobertaModel": ("colbert", "ColBERTJinaRobertaModel"),
275
    "ColBERTLfm2Model": ("colbert", "ColBERTLfm2Model"),
276
277
    # [Multimodal]
    "ColModernVBertForRetrieval": ("colmodernvbert", "ColModernVBertForRetrieval"),
278
    "ColPaliForRetrieval": ("colpali", "ColPaliModel"),
279
280
    "ColQwen3": ("colqwen3", "ColQwen3Model"),
    "OpsColQwen3Model": ("colqwen3", "ColQwen3Model"),
281
    "ColQwen3_5": ("colqwen3_5", "ColQwen3_5Model"),
282
283
284
285
286
287
288
289
290
291
    "Qwen3VLNemotronEmbedModel": ("colqwen3", "ColQwen3Model"),
}

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

_TOKEN_CLASSIFICATION_MODELS = {
292
    "BertForTokenClassification": ("bert", "BertForTokenClassification"),
293
    "ErnieForTokenClassification": ("ernie", "ErnieForTokenClassification"),
294
295
296
297
    "ModernBertForTokenClassification": (
        "modernbert",
        "ModernBertForTokenClassification",
    ),
298
299
300
301
    "Qwen3ASRForcedAlignerForTokenClassification": (
        "qwen3_asr_forced_aligner",
        "Qwen3ASRForcedAlignerForTokenClassification",
    ),
302
303
304
305
306
}

_SEQUENCE_CLASSIFICATION_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
307
    "ErnieForSequenceClassification": ("ernie", "ErnieForSequenceClassification"),
308
309
310
311
    "GteNewForSequenceClassification": (
        "bert_with_rope",
        "GteNewForSequenceClassification",
    ),
312
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),
313
314
315
    "LlamaBidirectionalForSequenceClassification": (
        "llama",
        "LlamaBidirectionalForSequenceClassification",
316
    ),
317
318
319
320
321
322
323
324
325
    "ModernBertForSequenceClassification": (
        "modernbert",
        "ModernBertForSequenceClassification",
    ),
    "RobertaForSequenceClassification": ("roberta", "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": (
        "roberta",
        "RobertaForSequenceClassification",
    ),
326
327
328
329
330
331
    # [Multimodal]
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"),
    "LlamaNemotronVLForSequenceClassification": (
        "nemotron_vl",
        "LlamaNemotronVLForSequenceClassification",
    ),
332
333
}

334
_MULTIMODAL_MODELS = {
335
    # [Decoder-only]
336
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
337
338
339
340
    "AudioFlamingo3ForConditionalGeneration": (
        "audioflamingo3",
        "AudioFlamingo3ForConditionalGeneration",
    ),
341
342
343
344
    "MusicFlamingoForConditionalGeneration": (
        "musicflamingo",
        "MusicFlamingoForConditionalGeneration",
    ),
345
346
347
    "AyaVisionForConditionalGeneration": (
        "aya_vision",
        "AyaVisionForConditionalGeneration",
348
    ),
349
    "BagelForConditionalGeneration": ("bagel", "BagelForConditionalGeneration"),
350
    "BeeForConditionalGeneration": ("bee", "BeeForConditionalGeneration"),
351
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
352
353
354
    "ChameleonForConditionalGeneration": (
        "chameleon",
        "ChameleonForConditionalGeneration",
355
    ),
356
357
    "Cheers": ("cheers", "CheersForConditionalGeneration"),
    "CheersForConditionalGeneration": ("cheers", "CheersForConditionalGeneration"),
358
359
360
    "Cohere2VisionForConditionalGeneration": (
        "cohere2_vision",
        "Cohere2VisionForConditionalGeneration",
361
    ),
362
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
363
    "DeepseekOCRForCausalLM": ("deepseek_ocr", "DeepseekOCRForCausalLM"),
RED's avatar
RED committed
364
    "DeepseekOCR2ForCausalLM": ("deepseek_ocr2", "DeepseekOCR2ForCausalLM"),
Roger Wang's avatar
Roger Wang committed
365
    "DotsOCRForCausalLM": ("dots_ocr", "DotsOCRForCausalLM"),
366
367
368
369
    "Eagle2_5_VLForConditionalGeneration": (
        "eagle2_5_vl",
        "Eagle2_5_VLForConditionalGeneration",
    ),
370
371
372
    "Ernie4_5_VLMoeForConditionalGeneration": (
        "ernie45_vl",
        "Ernie4_5_VLMoeForConditionalGeneration",
373
    ),
Kyungmin Lee's avatar
Kyungmin Lee committed
374
375
376
377
    "Exaone4_5_ForConditionalGeneration": (
        "exaone4_5",
        "Exaone4_5_ForConditionalGeneration",
    ),  # noqa: E501
378
379
380
381
    "FireRedASR2ForConditionalGeneration": (
        "fireredasr2",
        "FireRedASR2ForConditionalGeneration",
    ),
382
    "FunASRForConditionalGeneration": ("funasr", "FunASRForConditionalGeneration"),
383
384
385
386
    "FireRedLIDForConditionalGeneration": (
        "fireredlid",
        "FireRedLIDForConditionalGeneration",
    ),
387
388
389
390
    "FunAudioChatForConditionalGeneration": (
        "funaudiochat",
        "FunAudioChatForConditionalGeneration",
    ),
391
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
392
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),
393
394
395
    "Gemma3nForConditionalGeneration": (
        "gemma3n_mm",
        "Gemma3nForConditionalGeneration",
396
    ),
397
    "Gemma4ForConditionalGeneration": ("gemma4_mm", "Gemma4ForConditionalGeneration"),
398
    "GlmAsrForConditionalGeneration": ("glmasr", "GlmAsrForConditionalGeneration"),
399
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
400
401
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),
    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),
402
    "GlmOcrForConditionalGeneration": ("glm_ocr", "GlmOcrForConditionalGeneration"),
403
404
405
    "GraniteSpeechForConditionalGeneration": (
        "granite_speech",
        "GraniteSpeechForConditionalGeneration",
406
    ),
407
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
408
409
410
411
    "HunYuanVLForConditionalGeneration": (
        "hunyuan_vision",
        "HunYuanVLForConditionalGeneration",
    ),
412
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
413
414
415
    "InternS1ForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
416
    ),
417
418
419
    "InternVLForConditionalGeneration": (
        "interns1",
        "InternS1ForConditionalGeneration",
420
    ),
zxy's avatar
zxy committed
421
422
423
424
    "InternS1ProForConditionalGeneration": (
        "interns1_pro",
        "InternS1ProForConditionalGeneration",
    ),
425
426
427
428
    "Idefics3ForConditionalGeneration": (
        "idefics3",
        "Idefics3ForConditionalGeneration",
    ),
oscardev256's avatar
oscardev256 committed
429
    "IsaacForConditionalGeneration": ("isaac", "IsaacForConditionalGeneration"),
430
    "KananaVForConditionalGeneration": ("kanana_v", "KananaVForConditionalGeneration"),
431
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
432
433
434
    "KeyeVL1_5ForConditionalGeneration": (
        "keye_vl1_5",
        "KeyeVL1_5ForConditionalGeneration",
435
    ),
436
437
438
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),
    "KimiK25ForConditionalGeneration": ("kimi_k25", "KimiK25ForConditionalGeneration"),
    "MoonshotKimiaForCausalLM": ("kimi_audio", "KimiAudioForConditionalGeneration"),
439
440
441
442
    "LightOnOCRForConditionalGeneration": (
        "lightonocr",
        "LightOnOCRForConditionalGeneration",
    ),
443
    "Lfm2VlForConditionalGeneration": ("lfm2_vl", "Lfm2VLForConditionalGeneration"),
444
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),
445
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
446
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
447
448
449
    "LlavaNextForConditionalGeneration": (
        "llava_next",
        "LlavaNextForConditionalGeneration",
450
    ),
451
452
453
    "LlavaNextVideoForConditionalGeneration": (
        "llava_next_video",
        "LlavaNextVideoForConditionalGeneration",
454
    ),
455
456
457
    "LlavaOnevisionForConditionalGeneration": (
        "llava_onevision",
        "LlavaOnevisionForConditionalGeneration",
458
    ),
459
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),
460
    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
461
462
463
    "MiniMaxVL01ForConditionalGeneration": (
        "minimax_vl_01",
        "MiniMaxVL01ForConditionalGeneration",
464
    ),
465
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
466
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
467
468
469
    "Mistral3ForConditionalGeneration": (
        "mistral3",
        "Mistral3ForConditionalGeneration",
470
    ),
471
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
472
    "Molmo2ForConditionalGeneration": ("molmo2", "Molmo2ForConditionalGeneration"),
473
    "NemotronH_Nano_VL_V2": ("nano_nemotron_vl", "NemotronH_Nano_VL_V2"),
474
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
475
    "OpenCUAForConditionalGeneration": ("opencua", "OpenCUAForConditionalGeneration"),
476
477
478
479
    "OpenPanguVLForConditionalGeneration": (
        "openpangu_vl",
        "OpenPanguVLForConditionalGeneration",
    ),
480
    "Ovis": ("ovis", "Ovis"),
481
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
482
483
    "Ovis2_6ForCausalLM": ("ovis2_5", "Ovis2_5"),
    "Ovis2_6_MoeForCausalLM": ("ovis2_5", "Ovis2_5"),
484
485
486
487
    "PaddleOCRVLForConditionalGeneration": (
        "paddleocr_vl",
        "PaddleOCRVLForConditionalGeneration",
    ),
488
489
490
491
    "PaliGemmaForConditionalGeneration": (
        "paligemma",
        "PaliGemmaForConditionalGeneration",
    ),
492
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
493
    "Phi4ForCausalLMV": ("phi4siglip", "Phi4ForCausalLMV"),
494
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
495
496
497
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),
498
499
500
    "Qwen2_5_VLForConditionalGeneration": (
        "qwen2_5_vl",
        "Qwen2_5_VLForConditionalGeneration",
501
    ),
502
503
504
    "Qwen2AudioForConditionalGeneration": (
        "qwen2_audio",
        "Qwen2AudioForConditionalGeneration",
505
    ),
506
507
508
    "Qwen2_5OmniModel": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
509
    ),
510
511
512
    "Qwen2_5OmniForConditionalGeneration": (
        "qwen2_5_omni_thinker",
        "Qwen2_5OmniThinkerForConditionalGeneration",
513
    ),
514
515
516
517
    "Qwen3OmniMoeForConditionalGeneration": (
        "qwen3_omni_moe_thinker",
        "Qwen3OmniMoeThinkerForConditionalGeneration",
    ),
Roger Wang's avatar
Roger Wang committed
518
519
520
521
    "Qwen3ASRForConditionalGeneration": (
        "qwen3_asr",
        "Qwen3ASRForConditionalGeneration",
    ),
522
523
    "Qwen3ASRRealtimeGeneration": ("qwen3_asr_realtime", "Qwen3ASRRealtimeGeneration"),
    "Qwen3VLForConditionalGeneration": ("qwen3_vl", "Qwen3VLForConditionalGeneration"),
524
525
526
    "Qwen3VLMoeForConditionalGeneration": (
        "qwen3_vl_moe",
        "Qwen3VLMoeForConditionalGeneration",
527
    ),
528
    "Qwen3_5ForConditionalGeneration": ("qwen3_5", "Qwen3_5ForConditionalGeneration"),
529
530
531
532
    "Qwen3_5MoeForConditionalGeneration": (
        "qwen3_5",
        "Qwen3_5MoeForConditionalGeneration",
    ),
533
    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
534
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
535
536
537
538
    "SmolVLMForConditionalGeneration": ("smolvlm", "SmolVLMForConditionalGeneration"),
    "StepVLForConditionalGeneration": ("step_vl", "StepVLForConditionalGeneration"),
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),
539
540
541
    "Tarsier2ForConditionalGeneration": (
        "qwen2_vl",
        "Tarsier2ForConditionalGeneration",
542
    ),
543
    "UltravoxModel": ("ultravox", "UltravoxModel"),
544
545
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),
    "VoxtralRealtimeGeneration": ("voxtral_realtime", "VoxtralRealtimeGeneration"),
546
    # [Encoder-decoder]
547
    "CohereAsrForConditionalGeneration": (
Ekagra Ranjan's avatar
Ekagra Ranjan committed
548
        "cohere_asr",
549
        "CohereAsrForConditionalGeneration",
Ekagra Ranjan's avatar
Ekagra Ranjan committed
550
    ),
551
552
553
554
    "NemotronParseForConditionalGeneration": (
        "nemotron_parse",
        "NemotronParseForConditionalGeneration",
    ),
555
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),
556
}
557
558

_SPECULATIVE_DECODING_MODELS = {
559
    "ExtractHiddenStatesModel": ("extract_hidden_states", "ExtractHiddenStatesModel"),
560
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
561
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
562
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
563
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
564
    "DFlashDraftModel": ("qwen3_dflash", "DFlashQwen3ForCausalLM"),
565
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
566
    "Eagle3MiniMaxM2ForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
567
    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
568
    "Eagle3Qwen2_5vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
569
    "Eagle3Qwen3vlForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
570
571
572
573
    "EagleMistralLarge3ForCausalLM": (
        "mistral_large_3_eagle",
        "EagleMistralLarge3ForCausalLM",
    ),
574
575
    "Eagle3DeepseekV2ForCausalLM": ("deepseek_eagle3", "Eagle3DeepseekV2ForCausalLM"),
    "Eagle3DeepseekV3ForCausalLM": ("deepseek_eagle3", "Eagle3DeepseekV2ForCausalLM"),
576
    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
577
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
578
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
Kyungmin Lee's avatar
Kyungmin Lee committed
579
    "ExaoneMoeMTP": ("exaone_moe_mtp", "ExaoneMoeMTP"),
Kyungmin Lee's avatar
Kyungmin Lee committed
580
    "Exaone4_5_MTP": ("exaone4_5_mtp", "Exaone4_5_MTP"),
581
    "NemotronHMTPModel": ("nemotron_h_mtp", "NemotronHMTP"),
XuruiYang's avatar
XuruiYang committed
582
    "LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
583
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
584
    "Glm4MoeLiteMTPModel": ("glm4_moe_lite_mtp", "Glm4MoeLiteMTP"),
585
    "GlmOcrMTPModel": ("glm_ocr_mtp", "GlmOcrMTP"),
586
    "MedusaModel": ("medusa", "Medusa"),
587
    "OpenPanguMTPModel": ("openpangu_mtp", "OpenPanguMTP"),
588
    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
csy0225's avatar
csy0225 committed
589
    "Step3p5MTP": ("step3p5_mtp", "Step3p5MTP"),
590
591
    "Qwen3_5MTP": ("qwen3_5_mtp", "Qwen3_5MTP"),
    "Qwen3_5MoeMTP": ("qwen3_5_mtp", "Qwen3_5MoeMTP"),
592
593
594
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
595
}
596

597
_TRANSFORMERS_SUPPORTED_MODELS = {
598
599
600
    # Text generation models
    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
    # Multimodal models
601
602
603
604
    "Emu3ForConditionalGeneration": (
        "transformers",
        "TransformersMultiModalForCausalLM",
    ),
605
606
607
}

_TRANSFORMERS_BACKEND_MODELS = {
608
    # Text generation models
609
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
    "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
628
    "TransformersForSequenceClassification": (
629
        "transformers",
630
        "TransformersForSequenceClassification",
631
    ),
632
    "TransformersMoEForSequenceClassification": (
633
        "transformers",
634
        "TransformersMoEForSequenceClassification",
635
    ),
636
637
638
    "TransformersMultiModalForSequenceClassification": (
        "transformers",
        "TransformersMultiModalForSequenceClassification",
639
    ),
640
}
641

642
_VLLM_MODELS = {
643
    **_TEXT_GENERATION_MODELS,
644
    **_EMBEDDING_MODELS,
645
646
647
648
    **_LATE_INTERACTION_MODELS,
    **_REWARD_MODELS,
    **_TOKEN_CLASSIFICATION_MODELS,
    **_SEQUENCE_CLASSIFICATION_MODELS,
649
    **_MULTIMODAL_MODELS,
650
    **_SPECULATIVE_DECODING_MODELS,
651
652
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
653
654
}

655
656
657
658
# 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.
659
_SUBPROCESS_COMMAND = [sys.executable, "-m", "vllm.model_executor.models.registry"]
660

661
_PREVIOUSLY_SUPPORTED_MODELS = {
662
    "MotifForCausalLM": "0.10.2",
663
    "Phi3SmallForCausalLM": "0.9.2",
664
    "Phi4FlashForCausalLM": "0.10.2",
665
    "Phi4MultimodalForCausalLM": "0.12.0",
666
667
668
669
670
    # encoder-decoder models except whisper
    # have been removed for V0 deprecation.
    "DonutForConditionalGeneration": "0.10.2",
    "MllamaForConditionalGeneration": "0.10.2",
}
671

672
673
674
675
676
677
678
_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",
}

679

680
681
@dataclass(frozen=True)
class _ModelInfo:
682
    architecture: str
683
    is_text_generation_model: bool
684
    is_pooling_model: bool
685
    attn_type: AttnTypeStr
686
687
    default_seq_pooling_type: SequencePoolingType
    default_tok_pooling_type: TokenPoolingType
688
    score_type: ScoreType
689
    supports_multimodal: bool
690
    supports_multimodal_raw_input_only: bool
Patrick von Platen's avatar
Patrick von Platen committed
691
    requires_raw_input_tokens: bool
692
    supports_multimodal_encoder_tp_data: bool
693
    supports_pp: bool
694
695
    has_inner_state: bool
    is_attention_free: bool
696
    is_hybrid: bool
697
    has_noops: bool
698
    supports_mamba_prefix_caching: bool
699
    supports_transcription: bool
700
    supports_transcription_only: bool
701
702

    @staticmethod
703
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
704
        return _ModelInfo(
705
            architecture=model.__name__,
706
            is_text_generation_model=is_text_generation_model(model),
707
            is_pooling_model=is_pooling_model(model),
708
709
            default_seq_pooling_type=get_default_seq_pooling_type(model),
            default_tok_pooling_type=get_default_tok_pooling_type(model),
710
            attn_type=get_attn_type(model),
711
            score_type=get_score_type(model),
712
            supports_multimodal=supports_multimodal(model),
713
714
715
            supports_multimodal_raw_input_only=supports_multimodal_raw_input_only(
                model
            ),
Patrick von Platen's avatar
Patrick von Platen committed
716
            requires_raw_input_tokens=requires_raw_input_tokens(model),
717
718
719
            supports_multimodal_encoder_tp_data=supports_multimodal_encoder_tp_data(
                model
            ),
720
            supports_pp=supports_pp(model),
721
722
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
723
            is_hybrid=is_hybrid(model),
724
            supports_mamba_prefix_caching=supports_mamba_prefix_caching(model),
725
            supports_transcription=supports_transcription(model),
726
727
728
            supports_transcription_only=(
                supports_transcription(model) and model.supports_transcription_only
            ),
729
            has_noops=has_noops(model),
730
        )
731
732


733
734
735
736
class _BaseRegisteredModel(ABC):
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
737

738
    @abstractmethod
739
    def load_model_cls(self) -> type[nn.Module]:
740
        raise NotImplementedError
741
742


743
744
745
746
747
748
749
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
750
    model_cls: type[nn.Module]
751
752

    @staticmethod
753
    def from_model_cls(model_cls: type[nn.Module]):
754
755
756
757
758
759
760
761
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

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

762
    def load_model_cls(self) -> type[nn.Module]:
763
764
765
766
767
768
769
770
        return self.model_cls


@dataclass(frozen=True)
class _LazyRegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has not been imported in the main process.
    """
771

772
773
774
    module_name: str
    class_name: str

775
776
777
778
779
780
781
782
    @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"

783
    def _load_modelinfo_from_cache(self, module_hash: str) -> _ModelInfo | None:
784
785
        try:
            try:
786
                modelinfo_path = self._get_cache_dir() / self._get_cache_filename()
787
788
789
                with open(modelinfo_path, encoding="utf-8") as file:
                    mi_dict = json.load(file)
            except FileNotFoundError:
790
                logger.debug(
791
                    "Cached model info file for class %s.%s not found",
792
793
794
                    self.module_name,
                    self.class_name,
                )
795
796
797
                return None

            if mi_dict["hash"] != module_hash:
798
                logger.debug(
799
                    "Cached model info file for class %s.%s is stale",
800
801
802
                    self.module_name,
                    self.class_name,
                )
803
804
805
806
807
                return None

            # file not changed, use cached _ModelInfo properties
            return _ModelInfo(**mi_dict["modelinfo"])
        except Exception:
808
            logger.debug(
809
                "Cached model info for class %s.%s error. ",
810
811
812
                self.module_name,
                self.class_name,
            )
813
814
            return None

815
    def _save_modelinfo_to_cache(self, mi: _ModelInfo, module_hash: str) -> None:
816
817
        """save dictionary json file to cache"""
        from vllm.model_executor.model_loader.weight_utils import atomic_writer
818

819
820
821
822
823
824
825
826
        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()
827
            with atomic_writer(modelinfo_path, encoding="utf-8") as f:
828
829
830
831
832
                json.dump(modelinfo_dict, f, indent=2)
        except Exception:
            logger.exception("Error saving model info cache.")

    @logtime(logger=logger, msg="Registry inspect model class")
833
    def inspect_model_cls(self) -> _ModelInfo:
834
        model_path = Path(__file__).parent / f"{self.module_name.split('.')[-1]}.py"
835
        module_hash = None
836

837
838
        if model_path.exists():
            with open(model_path, "rb") as f:
839
                module_hash = safe_hash(f.read(), usedforsecurity=False).hexdigest()
840
841
842

            mi = self._load_modelinfo_from_cache(module_hash)
            if mi is not None:
843
                logger.debug(
844
                    "Loaded model info for class %s.%s from cache",
845
846
847
                    self.module_name,
                    self.class_name,
                )
848
849
                return mi
            else:
850
                logger.debug(
851
                    "Cache model info for class %s.%s miss. Loading model instead.",
852
853
854
                    self.module_name,
                    self.class_name,
                )
855
856
857

        # Performed in another process to avoid initializing CUDA
        mi = _run_in_subprocess(
858
859
860
861
862
            lambda: _ModelInfo.from_model_cls(self.load_model_cls())
        )
        logger.debug(
            "Loaded model info for class %s.%s", self.module_name, self.class_name
        )
863
864

        # save cache file
865
866
        if module_hash is not None:
            self._save_modelinfo_to_cache(mi, module_hash)
867
868

        return mi
869

870
    def load_model_cls(self) -> type[nn.Module]:
871
872
873
874
875
876
877
878
        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,
879
) -> type[nn.Module] | None:
880
    from vllm.platforms import current_platform
881

882
    current_platform.verify_model_arch(model_arch)
883
884
885
    try:
        return model.load_model_cls()
    except Exception:
886
        logger.exception("Error in loading model architecture '%s'", model_arch)
887
        return None
888
889


890
891
892
893
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
894
) -> _ModelInfo | None:
895
896
897
    try:
        return model.inspect_model_cls()
    except Exception:
898
        logger.exception("Error in inspecting model architecture '%s'", model_arch)
899
        return None
900
901


902
903
904
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
905
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
906

907
    def get_supported_archs(self) -> Set[str]:
908
        return self.models.keys()
909

910
911
912
    def register_model(
        self,
        model_arch: str,
913
        model_cls: type[nn.Module] | str,
914
    ) -> None:
915
916
917
        """
        Register an external model to be used in vLLM.

918
        `model_cls` can be either:
919

920
        - A [`torch.nn.Module`][] class directly referencing the model.
921
        - A string in the format `<module>:<class>` which can be used to
922
923
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
924
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
925
        """
926
927
928
929
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

930
        if model_arch in self.models:
931
932
            logger.warning(
                "Model architecture %s is already registered, and will be "
933
934
935
936
                "overwritten by the new model class %s.",
                model_arch,
                model_cls,
            )
937
938
939
940
941
942

        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)
943

944
            model = _LazyRegisteredModel(*split_str)
945
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
946
            model = _RegisteredModel.from_model_cls(model_cls)
947
        else:
948
949
950
951
            msg = (
                "`model_cls` should be a string or PyTorch model class, "
                f"not a {type(model_arch)}"
            )
952
            raise TypeError(msg)
953

954
        self.models[model_arch] = model
955

956
    def _raise_for_unsupported(self, architectures: list[str]):
957
        all_supported_archs = self.get_supported_archs()
958

959
960
961
        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
962
963
                "to be inspected. Please check the logs for more details."
            )
964

965
966
967
968
969
970
971
972
        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 "
973
974
                    "use this model architecture."
                )
975
976
977
978
979
980
981
982
            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."
                )
983

984
985
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
986
987
            f"Supported architectures: {all_supported_archs}"
        )
988

989
    def _try_load_model_cls(self, model_arch: str) -> type[nn.Module] | None:
990
991
        if model_arch not in self.models:
            return None
992

993
        return _try_load_model_cls(model_arch, self.models[model_arch])
994

995
    def _try_inspect_model_cls(self, model_arch: str) -> _ModelInfo | None:
996
997
        if model_arch not in self.models:
            return None
998

999
1000
1001
1002
1003
1004
        return _try_inspect_model_cls(model_arch, self.models[model_arch])

    def _try_resolve_transformers(
        self,
        architecture: str,
        model_config: ModelConfig,
1005
    ) -> str | None:
1006
1007
1008
        if architecture in _TRANSFORMERS_BACKEND_MODELS:
            return architecture

1009
1010
1011
        auto_map: dict[str, str] = (
            getattr(model_config.hf_config, "auto_map", None) or dict()
        )
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027

        # 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,
1028
                        trust_remote_code=model_config.trust_remote_code,
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
                        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,
1041
                        trust_remote_code=model_config.trust_remote_code,
1042
1043
1044
1045
1046
                        warn_on_fail=True,
                    )
                    if model_module is not None:
                        break
            else:
1047
                if model_config.model_impl != "transformers":
1048
1049
1050
1051
1052
1053
1054
                    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 "
1055
1056
                    "'auto_map' (relevant if the model is custom)."
                )
1057
1058

        if not model_module.is_backend_compatible():
1059
            if model_config.model_impl != "transformers":
1060
                return None
1061

1062
1063
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
1064
1065
                "is not compatible with vLLM."
            )
1066

1067
        return model_config._get_transformers_backend_cls()
1068

1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
    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
1094

1095
1096
    def inspect_model_cls(
        self,
1097
        architectures: str | list[str],
1098
        model_config: ModelConfig,
1099
    ) -> tuple[_ModelInfo, str]:
1100
1101
        if isinstance(architectures, str):
            architectures = [architectures]
1102
1103
        if not architectures:
            raise ValueError("No model architectures are specified")
1104
1105

        # Require transformers impl
1106
        if model_config.model_impl == "transformers":
1107
            arch = self._try_resolve_transformers(architectures[0], model_config)
1108
1109
1110
1111
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)
1112
        elif model_config.model_impl == "terratorch":
1113
1114
            model_info = self._try_inspect_model_cls("Terratorch")
            return (model_info, "Terratorch")
1115

1116
        # Fallback to transformers impl (after resolving convert_type)
1117
1118
1119
1120
1121
1122
        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)
1123
1124
1125
1126
1127
1128
1129
            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)
1130
            model_info = self._try_inspect_model_cls(normalized_arch)
1131
            if model_info is not None:
1132
                return (model_info, arch)
1133

1134
        # Fallback to transformers impl (before resolving runner_type)
1135
1136
1137
1138
1139
        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)
1140
1141
1142
1143
1144
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

1145
        return self._raise_for_unsupported(architectures)
1146

1147
1148
    def resolve_model_cls(
        self,
1149
        architectures: str | list[str],
1150
        model_config: ModelConfig,
1151
    ) -> tuple[type[nn.Module], str]:
1152
1153
        if isinstance(architectures, str):
            architectures = [architectures]
1154
1155
        if not architectures:
            raise ValueError("No model architectures are specified")
1156
1157

        # Require transformers impl
1158
        if model_config.model_impl == "transformers":
1159
            arch = self._try_resolve_transformers(architectures[0], model_config)
1160
1161
1162
1163
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)
1164
        elif model_config.model_impl == "terratorch":
1165
1166
1167
1168
            arch = "Terratorch"
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
1169

1170
        # Fallback to transformers impl (after resolving convert_type)
1171
1172
1173
1174
1175
1176
        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)
1177
1178
1179
1180
1181
1182
1183
            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)
1184
            model_cls = self._try_load_model_cls(normalized_arch)
1185
1186
            if model_cls is not None:
                return (model_cls, arch)
1187

1188
        # Fallback to transformers impl (before resolving runner_type)
1189
1190
1191
1192
1193
        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)
1194
1195
1196
1197
1198
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

1199
        return self._raise_for_unsupported(architectures)
1200

1201
1202
    def is_text_generation_model(
        self,
1203
        architectures: str | list[str],
1204
        model_config: ModelConfig,
1205
    ) -> bool:
1206
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1207
        return model_cls.is_text_generation_model
1208

1209
    def is_pooling_model(
1210
        self,
1211
        architectures: str | list[str],
1212
        model_config: ModelConfig,
1213
    ) -> bool:
1214
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1215
        return model_cls.is_pooling_model
1216
1217
1218

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

1225
    def is_multimodal_raw_input_only_model(
1226
        self,
1227
        architectures: str | list[str],
1228
        model_config: ModelConfig,
1229
    ) -> bool:
1230
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1231
        return model_cls.supports_multimodal_raw_input_only
1232

1233
1234
    def is_pp_supported_model(
        self,
1235
        architectures: str | list[str],
1236
        model_config: ModelConfig,
1237
    ) -> bool:
1238
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1239
        return model_cls.supports_pp
1240

1241
1242
    def model_has_inner_state(
        self,
1243
        architectures: str | list[str],
1244
        model_config: ModelConfig,
1245
    ) -> bool:
1246
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1247
        return model_cls.has_inner_state
1248

1249
1250
    def is_attention_free_model(
        self,
1251
        architectures: str | list[str],
1252
        model_config: ModelConfig,
1253
    ) -> bool:
1254
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1255
        return model_cls.is_attention_free
1256

1257
1258
    def is_hybrid_model(
        self,
1259
        architectures: str | list[str],
1260
        model_config: ModelConfig,
1261
    ) -> bool:
1262
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1263
1264
        return model_cls.is_hybrid

1265
1266
    def is_noops_model(
        self,
1267
        architectures: str | list[str],
1268
        model_config: ModelConfig,
1269
    ) -> bool:
1270
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1271
1272
        return model_cls.has_noops

1273
1274
    def is_transcription_model(
        self,
1275
        architectures: str | list[str],
1276
        model_config: ModelConfig,
1277
    ) -> bool:
1278
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1279
1280
        return model_cls.supports_transcription

1281
1282
    def is_transcription_only_model(
        self,
1283
        architectures: str | list[str],
1284
        model_config: ModelConfig,
1285
    ) -> bool:
1286
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
1287
1288
        return model_cls.supports_transcription_only

1289

1290
1291
1292
1293
1294
1295
1296
1297
1298
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()
    }
)
1299
1300
1301
1302
1303

_T = TypeVar("_T")


def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
1304
1305
1306
1307
1308
    # 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")

1309
        # `cloudpickle` allows pickling lambda functions directly
1310
        import cloudpickle
1311

1312
        input_bytes = cloudpickle.dumps((fn, output_filepath))
1313
1314
1315

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
1316
1317
1318
        returned = subprocess.run(
            _SUBPROCESS_COMMAND, input=input_bytes, capture_output=True
        )
1319
1320
1321
1322
1323
1324

        # check if the subprocess is successful
        try:
            returned.check_returncode()
        except Exception as e:
            # wrap raised exception to provide more information
1325
1326
1327
            raise RuntimeError(
                f"Error raised in subprocess:\n{returned.stderr.decode()}"
            ) from e
1328

1329
        with open(output_filepath, "rb") as f:
1330
1331
1332
1333
1334
1335
            return pickle.load(f)


def _run() -> None:
    # Setup plugins
    from vllm.plugins import load_general_plugins
1336

1337
1338
1339
1340
1341
    load_general_plugins()

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

    result = fn()
1342
1343
1344

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
1345
1346
1347


if __name__ == "__main__":
1348
    _run()