registry.py 38.7 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
import importlib
8
import os
9
import pickle
10
11
import subprocess
import sys
12
import tempfile
13
from abc import ABC, abstractmethod
14
from collections.abc import Set
15
from dataclasses import dataclass, field
16
from functools import lru_cache
17
from typing import Callable, Optional, TypeVar, Union
18
19

import torch.nn as nn
20
import transformers
21

22
23
from vllm.config import (ModelConfig, ModelImpl, iter_architecture_defaults,
                         try_match_architecture_defaults)
24
from vllm.logger import init_logger
25
26
from vllm.transformers_utils.dynamic_module import (
    try_get_class_from_dynamic_module)
27

28
29
from .interfaces import (has_inner_state, has_noops, is_attention_free,
                         is_hybrid, supports_cross_encoding,
30
31
                         supports_multimodal,
                         supports_multimodal_encoder_tp_data,
32
                         supports_multimodal_raw_input_only, supports_pp,
33
                         supports_transcription, supports_v0_only)
34
35
from .interfaces_base import (get_default_pooling_type, is_pooling_model,
                              is_text_generation_model)
36
37
38

logger = init_logger(__name__)

39
# yapf: disable
40
41
_TEXT_GENERATION_MODELS = {
    # [Decoder-only]
42
    "ApertusForCausalLM": ("apertus", "ApertusForCausalLM"),
43
44
    "AquilaModel": ("llama", "LlamaForCausalLM"),
    "AquilaForCausalLM": ("llama", "LlamaForCausalLM"),  # AquilaChat2
Raghav Ravishankar's avatar
Raghav Ravishankar committed
45
    "ArceeForCausalLM": ("arcee", "ArceeForCausalLM"),
46
    "ArcticForCausalLM": ("arctic", "ArcticForCausalLM"),
47
    "MiniMaxForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
48
    "MiniMaxText01ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
49
    "MiniMaxM1ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
50
51
52
53
    # baichuan-7b, upper case 'C' in the class name
    "BaiChuanForCausalLM": ("baichuan", "BaiChuanForCausalLM"),
    # baichuan-13b, lower case 'c' in the class name
    "BaichuanForCausalLM": ("baichuan", "BaichuanForCausalLM"),
54
    "BailingMoeForCausalLM": ("bailing_moe", "BailingMoeForCausalLM"),
Yu Chin Fabian Lim's avatar
Yu Chin Fabian Lim committed
55
    "BambaForCausalLM": ("bamba", "BambaForCausalLM"),
56
    "BloomForCausalLM": ("bloom", "BloomForCausalLM"),
57
    "ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
58
    "ChatGLMForConditionalGeneration": ("chatglm", "ChatGLMForCausalLM"),
59
    "CohereForCausalLM": ("commandr", "CohereForCausalLM"),
60
    "Cohere2ForCausalLM": ("commandr", "CohereForCausalLM"),
61
    "DbrxForCausalLM": ("dbrx", "DbrxForCausalLM"),
62
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
63
64
    "DeepseekForCausalLM": ("deepseek", "DeepseekForCausalLM"),
    "DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
65
    "DeepseekV3ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
66
    "Dots1ForCausalLM": ("dots1", "Dots1ForCausalLM"),
67
    "Ernie4_5ForCausalLM": ("ernie45", "Ernie4_5ForCausalLM"),
68
    "Ernie4_5_MoeForCausalLM": ("ernie45_moe", "Ernie4_5_MoeForCausalLM"),
69
    "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
70
    "Exaone4ForCausalLM": ("exaone4", "Exaone4ForCausalLM"),
71
    "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
72
    "Fairseq2LlamaForCausalLM": ("fairseq2_llama", "Fairseq2LlamaForCausalLM"),
73
74
    "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
    "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
75
    "Gemma3ForCausalLM": ("gemma3", "Gemma3ForCausalLM"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
76
    "Gemma3nForCausalLM": ("gemma3n", "Gemma3nForCausalLM"),
77
    "Qwen3NextForCausalLM": ("qwen3_next", "Qwen3NextForCausalLM"),
78
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
79
    "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
80
    "Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
81
    "GptOssForCausalLM": ("gpt_oss", "GptOssForCausalLM"),
82
83
84
85
86
87
    "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
    "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
    "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
    "GPTNeoXForCausalLM": ("gpt_neox", "GPTNeoXForCausalLM"),
    "GraniteForCausalLM": ("granite", "GraniteForCausalLM"),
    "GraniteMoeForCausalLM": ("granitemoe", "GraniteMoeForCausalLM"),
88
    "GraniteMoeHybridForCausalLM": ("granitemoehybrid", "GraniteMoeHybridForCausalLM"),   # noqa: E501
89
    "GraniteMoeSharedForCausalLM": ("granitemoeshared", "GraniteMoeSharedForCausalLM"),   # noqa: E501
90
    "GritLM": ("gritlm", "GritLM"),
Michael Goin's avatar
Michael Goin committed
91
    "Grok1ModelForCausalLM": ("grok1", "Grok1ForCausalLM"),
92
93
    "HunYuanMoEV1ForCausalLM": ("hunyuan_v1", "HunYuanMoEV1ForCausalLM"),
    "HunYuanDenseV1ForCausalLM": ("hunyuan_v1", "HunYuanDenseV1ForCausalLM"),
94
    "HCXVisionForCausalLM": ("hyperclovax_vision", "HCXVisionForCausalLM"),
95
96
    "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
    "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"),
97
    "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
98
    "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
99
100
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
101
    "Lfm2ForCausalLM": ("lfm2", "Lfm2ForCausalLM"),
102
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
103
    "Llama4ForCausalLM": ("llama4", "Llama4ForCausalLM"),  # noqa: E501
104
105
    # For decapoda-research/llama-*
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
106
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
107
    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
Dhia Eddine Rhaiem's avatar
Dhia Eddine Rhaiem committed
108
    "FalconH1ForCausalLM":("falcon_h1", "FalconH1ForCausalLM"),
109
    "Mamba2ForCausalLM": ("mamba2", "Mamba2ForCausalLM"),
110
111
    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
    "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
112
113
    "MistralForCausalLM": ("llama", "LlamaForCausalLM"),
    "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
114
    "MotifForCausalLM": ("motif", "MotifForCausalLM"),
115
116
117
    # transformers's mpt class has lower case
    "MptForCausalLM": ("mpt", "MPTForCausalLM"),
    "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
118
    "MiMoForCausalLM": ("mimo", "MiMoForCausalLM"),
119
    "NemotronForCausalLM": ("nemotron", "NemotronForCausalLM"),
Luis Vega's avatar
Luis Vega committed
120
    "NemotronHForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
121
    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
122
    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
123
124
125
126
127
128
129
    "OlmoeForCausalLM": ("olmoe", "OlmoeForCausalLM"),
    "OPTForCausalLM": ("opt", "OPTForCausalLM"),
    "OrionForCausalLM": ("orion", "OrionForCausalLM"),
    "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
    "PhiForCausalLM": ("phi", "PhiForCausalLM"),
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
    "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
130
    "Phi4FlashForCausalLM": ("phi4flash", "Phi4FlashForCausalLM"),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
131
    "Plamo2ForCausalLM": ("plamo2", "Plamo2ForCausalLM"),
132
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
133
134
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
135
136
    "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),
    "Qwen3MoeForCausalLM": ("qwen3_moe", "Qwen3MoeForCausalLM"),
137
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
138
    "SeedOssForCausalLM": ("seed_oss", "SeedOssForCausalLM"),
Song's avatar
Song committed
139
    "Step3TextForCausalLM": ("step3_text", "Step3TextForCausalLM"),
140
141
142
143
    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
144
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
145
    "TeleFLMForCausalLM": ("teleflm", "TeleFLMForCausalLM"),
146
    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
147
    "Zamba2ForCausalLM": ("zamba2", "Zamba2ForCausalLM"),
148
149
150
    # [Encoder-decoder]
    "BartModel": ("bart", "BartForConditionalGeneration"),
    "BartForConditionalGeneration": ("bart", "BartForConditionalGeneration"),
汪志鹏's avatar
汪志鹏 committed
151
    "MBartForConditionalGeneration": ("bart", "MBartForConditionalGeneration"),
152
153
154
}

_EMBEDDING_MODELS = {
155
    # [Text-only]
156
    "BertModel": ("bert", "BertEmbeddingModel"),
157
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
158
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
159
    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
160
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
161
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
162
    "GritLM": ("gritlm", "GritLM"),
163
164
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
165
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
166
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
167
    "LlamaModel": ("llama", "LlamaForCausalLM"),
168
169
170
171
172
    **{
        # Multiple models share the same architecture, so we include them all
        k: (mod, arch) for k, (mod, arch) in _TEXT_GENERATION_MODELS.items()
        if arch == "LlamaForCausalLM"
    },
173
    "MistralModel": ("llama", "LlamaForCausalLM"),
174
    "ModernBertModel": ("modernbert", "ModernBertModel"),
175
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
176
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
177
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
178
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
179
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
180
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
181
182
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
183
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
184
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
185
    # [Multimodal]
Cyrus Leung's avatar
Cyrus Leung committed
186
    "LlavaNextForConditionalGeneration": ("llava_next", "LlavaNextForConditionalGeneration"),  # noqa: E501
187
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
188
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
189
190
    # Technically Terratorch models work on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
191
    # models for the time being.
192
193
    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
    "Terratorch": ("terratorch", "Terratorch"),
194
195
}

196
197
_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
198
199
200
201
    "GteNewForSequenceClassification": ("bert_with_rope",
                                        "GteNewForSequenceClassification"),
    "ModernBertForSequenceClassification": ("modernbert",
                                            "ModernBertForSequenceClassification"),
202
203
204
205
    "RobertaForSequenceClassification": ("roberta",
                                         "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": ("roberta",
                                            "RobertaForSequenceClassification"),
206
    # [Auto-converted (see adapters.py)]
207
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"), # noqa: E501,
208
209
}

210
_MULTIMODAL_MODELS = {
211
    # [Decoder-only]
212
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
Jennifer Zhao's avatar
Jennifer Zhao committed
213
    "AyaVisionForConditionalGeneration": ("aya_vision", "AyaVisionForConditionalGeneration"),  # noqa: E501
214
215
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
    "ChameleonForConditionalGeneration": ("chameleon", "ChameleonForConditionalGeneration"),  # noqa: E501
216
    "Cohere2VisionForConditionalGeneration": ("cohere2_vision", "Cohere2VisionForConditionalGeneration"),  # noqa: E501
217
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
218
    "Ernie4_5_VLMoeForConditionalGeneration": ("ernie45_vl", "Ernie4_5_VLMoeForConditionalGeneration"),  # noqa: E501
219
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
220
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
221
    "Gemma3nForConditionalGeneration": ("gemma3n_mm", "Gemma3nForConditionalGeneration"),    # noqa: E501
222
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
223
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),  # noqa: E501
Jee Jee Li's avatar
Jee Jee Li committed
224
    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),  # noqa: E501
225
    "GraniteSpeechForConditionalGeneration": ("granite_speech", "GraniteSpeechForConditionalGeneration"),  # noqa: E501
226
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
227
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
228
    "NemotronH_Nano_VL": ("nano_nemotron_vl", "NemotronH_Nano_VL"),
Lyu Han's avatar
Lyu Han committed
229
    "InternS1ForConditionalGeneration": ("interns1", "InternS1ForConditionalGeneration"),  # noqa: E501
230
    "InternVLForConditionalGeneration": ("interns1", "InternS1ForConditionalGeneration"),  # noqa: E501
231
    "Idefics3ForConditionalGeneration":("idefics3","Idefics3ForConditionalGeneration"),
232
    "SmolVLMForConditionalGeneration": ("smolvlm","SmolVLMForConditionalGeneration"),  # noqa: E501
233
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
234
    "KeyeVL1_5ForConditionalGeneration": ("keye_vl1_5", "KeyeVL1_5ForConditionalGeneration"), # noqa: E501
235
    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
236
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
237
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
238
239
240
241
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
    "LlavaNextForConditionalGeneration": ("llava_next", "LlavaNextForConditionalGeneration"),  # noqa: E501
    "LlavaNextVideoForConditionalGeneration": ("llava_next_video", "LlavaNextVideoForConditionalGeneration"),  # noqa: E501
    "LlavaOnevisionForConditionalGeneration": ("llava_onevision", "LlavaOnevisionForConditionalGeneration"),  # noqa: E501
242
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
243
    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
244
    "MiniMaxVL01ForConditionalGeneration": ("minimax_vl_01", "MiniMaxVL01ForConditionalGeneration"),  # noqa: E501
245
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
246
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
247
    "Mistral3ForConditionalGeneration": ("mistral3", "Mistral3ForConditionalGeneration"),  # noqa: E501
248
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
249
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
250
    "Ovis": ("ovis", "Ovis"),
251
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
252
    "PaliGemmaForConditionalGeneration": ("paligemma", "PaliGemmaForConditionalGeneration"),  # noqa: E501
253
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
254
255
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
    "Phi4MultimodalForCausalLM": ("phi4_multimodal", "Phi4MultimodalForCausalLM"),  # noqa: E501
256
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
257
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),  # noqa: E501
258
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
Roger Wang's avatar
Roger Wang committed
259
    "Qwen2_5_VLForConditionalGeneration": ("qwen2_5_vl", "Qwen2_5_VLForConditionalGeneration"),  # noqa: E501
260
    "Qwen2AudioForConditionalGeneration": ("qwen2_audio", "Qwen2AudioForConditionalGeneration"),  # noqa: E501
261
    "Qwen2_5OmniModel": ("qwen2_5_omni_thinker", "Qwen2_5OmniThinkerForConditionalGeneration"),  # noqa: E501
262
    "Qwen2_5OmniForConditionalGeneration": ("qwen2_5_omni_thinker", "Qwen2_5OmniThinkerForConditionalGeneration"),  # noqa: E501
263
    "UltravoxModel": ("ultravox", "UltravoxModel"),
Song's avatar
Song committed
264
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),  # noqa: E501
汪志鹏's avatar
汪志鹏 committed
265
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),  # noqa: E501
266
    "Tarsier2ForConditionalGeneration": ("qwen2_vl", "Tarsier2ForConditionalGeneration"),  # noqa: E501
Patrick von Platen's avatar
Patrick von Platen committed
267
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),  # noqa: E501
268
    # [Encoder-decoder]
汪志鹏's avatar
汪志鹏 committed
269
    "DonutForConditionalGeneration": ("donut", "DonutForConditionalGeneration"),
270
    "Florence2ForConditionalGeneration": ("florence2", "Florence2ForConditionalGeneration"),  # noqa: E501
271
    "MllamaForConditionalGeneration": ("mllama", "MllamaForConditionalGeneration"),  # noqa: E501
272
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),  # noqa: E501
273
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
274
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
275
}
276
277

_SPECULATIVE_DECODING_MODELS = {
278
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
279
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
280
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
281
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
282
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
283
    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
284
    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
285
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
286
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
287
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
288
    "MedusaModel": ("medusa", "Medusa"),
289
    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
290
291
292
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
293
}
294

295
_TRANSFORMERS_SUPPORTED_MODELS = {
296
297
298
    # Text generation models
    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
    # Multimodal models
299
300
301
302
    "Emu3ForConditionalGeneration": ("transformers", "TransformersForMultimodalLM"),  # noqa: E501
}

_TRANSFORMERS_BACKEND_MODELS = {
303
    "TransformersModel": ("transformers", "TransformersModel"),
304
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
305
    "TransformersForMultimodalLM": ("transformers", "TransformersForMultimodalLM"), # noqa: E501
306
}
307
# yapf: enable
308

309
_VLLM_MODELS = {
310
    **_TEXT_GENERATION_MODELS,
311
    **_EMBEDDING_MODELS,
312
    **_CROSS_ENCODER_MODELS,
313
    **_MULTIMODAL_MODELS,
314
    **_SPECULATIVE_DECODING_MODELS,
315
316
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
317
318
}

319
320
321
322
323
324
325
326
# 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.
_SUBPROCESS_COMMAND = [
    sys.executable, "-m", "vllm.model_executor.models.registry"
]

327
328
_PREVIOUSLY_SUPPORTED_MODELS = {"Phi3SmallForCausalLM": "0.9.2"}

329

330
331
@dataclass(frozen=True)
class _ModelInfo:
332
    architecture: str
333
    is_text_generation_model: bool
334
    is_pooling_model: bool
335
    default_pooling_type: str
336
    supports_cross_encoding: bool
337
    supports_multimodal: bool
338
    supports_multimodal_raw_input_only: bool
339
    supports_multimodal_encoder_tp_data: bool
340
    supports_pp: bool
341
342
    has_inner_state: bool
    is_attention_free: bool
343
    is_hybrid: bool
344
    has_noops: bool
345
    supports_transcription: bool
346
    supports_transcription_only: bool
347
    supports_v0_only: bool
348
349

    @staticmethod
350
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
351
        return _ModelInfo(
352
            architecture=model.__name__,
353
            is_text_generation_model=is_text_generation_model(model),
354
            is_pooling_model=is_pooling_model(model),
355
            default_pooling_type=get_default_pooling_type(model),
356
            supports_cross_encoding=supports_cross_encoding(model),
357
            supports_multimodal=supports_multimodal(model),
358
359
            supports_multimodal_raw_input_only=
            supports_multimodal_raw_input_only(model),
360
361
            supports_multimodal_encoder_tp_data=
            supports_multimodal_encoder_tp_data(model),
362
            supports_pp=supports_pp(model),
363
364
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
365
            is_hybrid=is_hybrid(model),
366
            supports_transcription=supports_transcription(model),
367
368
            supports_transcription_only=(supports_transcription(model) and
                                         model.supports_transcription_only),
369
            supports_v0_only=supports_v0_only(model),
370
            has_noops=has_noops(model),
371
        )
372
373


374
class _BaseRegisteredModel(ABC):
375

376
377
378
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
379

380
    @abstractmethod
381
    def load_model_cls(self) -> type[nn.Module]:
382
        raise NotImplementedError
383
384


385
386
387
388
389
390
391
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
392
    model_cls: type[nn.Module]
393
394

    @staticmethod
395
    def from_model_cls(model_cls: type[nn.Module]):
396
397
398
399
400
401
402
403
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

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

404
    def load_model_cls(self) -> type[nn.Module]:
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
        return self.model_cls


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

    # Performed in another process to avoid initializing CUDA
    def inspect_model_cls(self) -> _ModelInfo:
        return _run_in_subprocess(
            lambda: _ModelInfo.from_model_cls(self.load_model_cls()))

421
    def load_model_cls(self) -> type[nn.Module]:
422
423
424
425
426
427
428
429
        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,
430
) -> Optional[type[nn.Module]]:
431
    from vllm.platforms import current_platform
432
    current_platform.verify_model_arch(model_arch)
433
434
435
436
437
438
    try:
        return model.load_model_cls()
    except Exception:
        logger.exception("Error in loading model architecture '%s'",
                         model_arch)
        return None
439
440


441
442
443
444
445
446
447
448
449
450
451
@lru_cache(maxsize=128)
def _try_inspect_model_cls(
    model_arch: str,
    model: _BaseRegisteredModel,
) -> Optional[_ModelInfo]:
    try:
        return model.inspect_model_cls()
    except Exception:
        logger.exception("Error in inspecting model architecture '%s'",
                         model_arch)
        return None
452
453


454
455
456
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
457
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
458

459
    def get_supported_archs(self) -> Set[str]:
460
        return self.models.keys()
461

462
463
464
    def register_model(
        self,
        model_arch: str,
465
        model_cls: Union[type[nn.Module], str],
466
    ) -> None:
467
468
469
        """
        Register an external model to be used in vLLM.

470
        `model_cls` can be either:
471

472
        - A [`torch.nn.Module`][] class directly referencing the model.
473
        - A string in the format `<module>:<class>` which can be used to
474
475
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
476
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
477
        """
478
479
480
481
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

482
        if model_arch in self.models:
483
484
485
            logger.warning(
                "Model architecture %s is already registered, and will be "
                "overwritten by the new model class %s.", model_arch,
486
487
488
489
490
491
492
                model_cls)

        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)
493

494
            model = _LazyRegisteredModel(*split_str)
495
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
496
            model = _RegisteredModel.from_model_cls(model_cls)
497
498
499
500
        else:
            msg = ("`model_cls` should be a string or PyTorch model class, "
                   f"not a {type(model_arch)}")
            raise TypeError(msg)
501

502
        self.models[model_arch] = model
503

504
    def _raise_for_unsupported(self, architectures: list[str]):
505
        all_supported_archs = self.get_supported_archs()
506

507
508
509
510
511
        if any(arch in all_supported_archs for arch in architectures):
            raise ValueError(
                f"Model architectures {architectures} failed "
                "to be inspected. Please check the logs for more details.")

512
513
514
515
516
517
518
519
520
521
        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 "
                    "use this model architecture.")

522
523
524
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
            f"Supported architectures: {all_supported_archs}")
525

526
    def _try_load_model_cls(self,
527
                            model_arch: str) -> Optional[type[nn.Module]]:
528
529
        if model_arch not in self.models:
            return None
530

531
        return _try_load_model_cls(model_arch, self.models[model_arch])
532

533
    def _try_inspect_model_cls(self, model_arch: str) -> Optional[_ModelInfo]:
534
535
        if model_arch not in self.models:
            return None
536

537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
        return _try_inspect_model_cls(model_arch, self.models[model_arch])

    def _try_resolve_transformers(
        self,
        architecture: str,
        model_config: ModelConfig,
    ) -> Optional[str]:
        if architecture in _TRANSFORMERS_BACKEND_MODELS:
            return architecture

        auto_map: dict[str, str] = getattr(model_config.hf_config, "auto_map",
                                           None) or dict()

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

        model_module = getattr(transformers, architecture, None)

        if model_module is None:
            for name, module in auto_map.items():
                if name.startswith("AutoModel"):
                    model_module = try_get_class_from_dynamic_module(
                        module,
                        model_config.model,
                        revision=model_config.revision,
                        warn_on_fail=True,
                    )
                    if model_module is not None:
                        break
            else:
                if model_config.model_impl != ModelImpl.TRANSFORMERS:
                    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 "
                    "'auto_map' (relevant if the model is custom).")

        if not model_module.is_backend_compatible():
            if model_config.model_impl != ModelImpl.TRANSFORMERS:
594
                return None
595

596
597
598
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
                "is not compatible with vLLM.")
599

600
        return model_config._get_transformers_backend_cls()
601

602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
    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
627

628
629
    def inspect_model_cls(
        self,
630
        architectures: Union[str, list[str]],
631
        model_config: ModelConfig,
632
    ) -> tuple[_ModelInfo, str]:
633
634
        if isinstance(architectures, str):
            architectures = [architectures]
635
636
        if not architectures:
            raise ValueError("No model architectures are specified")
637
638
639
640
641
642
643
644
645

        # Require transformers impl
        if model_config.model_impl == ModelImpl.TRANSFORMERS:
            arch = self._try_resolve_transformers(architectures[0],
                                                  model_config)
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)
646
647
648
        elif model_config.model_impl == ModelImpl.TERRATORCH:
            model_info = self._try_inspect_model_cls("Terratorch")
            return (model_info, "Terratorch")
649

650
651
652
653
654
655
656
657
658
659
660
661
662
        # Fallback to transformers impl (after resolving convert_type)
        if (all(arch not in self.models for arch in architectures)
                and model_config.model_impl == ModelImpl.AUTO
                and getattr(model_config, "convert_type", "none") == "none"):
            arch = self._try_resolve_transformers(architectures[0],
                                                  model_config)
            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)
663
            model_info = self._try_inspect_model_cls(normalized_arch)
664
            if model_info is not None:
665
                return (model_info, arch)
666

667
668
669
        # Fallback to transformers impl (before resolving runner_type)
        if (all(arch not in self.models for arch in architectures)
                and model_config.model_impl == ModelImpl.AUTO):
670
671
672
673
674
675
676
            arch = self._try_resolve_transformers(architectures[0],
                                                  model_config)
            if arch is not None:
                model_info = self._try_inspect_model_cls(arch)
                if model_info is not None:
                    return (model_info, arch)

677
        return self._raise_for_unsupported(architectures)
678

679
680
    def resolve_model_cls(
        self,
681
        architectures: Union[str, list[str]],
682
        model_config: ModelConfig,
683
    ) -> tuple[type[nn.Module], str]:
684
685
        if isinstance(architectures, str):
            architectures = [architectures]
686
687
        if not architectures:
            raise ValueError("No model architectures are specified")
688
689
690
691
692
693
694
695
696

        # Require transformers impl
        if model_config.model_impl == ModelImpl.TRANSFORMERS:
            arch = self._try_resolve_transformers(architectures[0],
                                                  model_config)
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)
697
698
699
700
701
        elif model_config.model_impl == ModelImpl.TERRATORCH:
            arch = "Terratorch"
            model_cls = self._try_load_model_cls(arch)
            if model_cls is not None:
                return (model_cls, arch)
702

703
704
705
706
707
708
709
710
711
712
713
714
715
        # Fallback to transformers impl (after resolving convert_type)
        if (all(arch not in self.models for arch in architectures)
                and model_config.model_impl == ModelImpl.AUTO
                and getattr(model_config, "convert_type", "none") == "none"):
            arch = self._try_resolve_transformers(architectures[0],
                                                  model_config)
            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)
716
            model_cls = self._try_load_model_cls(normalized_arch)
717
718
            if model_cls is not None:
                return (model_cls, arch)
719

720
721
722
        # Fallback to transformers impl (before resolving runner_type)
        if (all(arch not in self.models for arch in architectures)
                and model_config.model_impl == ModelImpl.AUTO):
723
724
725
726
727
728
729
            arch = self._try_resolve_transformers(architectures[0],
                                                  model_config)
            if arch is not None:
                model_cls = self._try_load_model_cls(arch)
                if model_cls is not None:
                    return (model_cls, arch)

730
        return self._raise_for_unsupported(architectures)
731

732
733
    def is_text_generation_model(
        self,
734
        architectures: Union[str, list[str]],
735
        model_config: ModelConfig,
736
    ) -> bool:
737
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
738
        return model_cls.is_text_generation_model
739

740
    def is_pooling_model(
741
        self,
742
        architectures: Union[str, list[str]],
743
        model_config: ModelConfig,
744
    ) -> bool:
745
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
746
        return model_cls.is_pooling_model
747

748
749
    def is_cross_encoder_model(
        self,
750
        architectures: Union[str, list[str]],
751
        model_config: ModelConfig,
752
    ) -> bool:
753
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
754
        return model_cls.supports_cross_encoding
755

756
757
    def is_multimodal_model(
        self,
758
        architectures: Union[str, list[str]],
759
        model_config: ModelConfig,
760
    ) -> bool:
761
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
762
        return model_cls.supports_multimodal
763

764
    def is_multimodal_raw_input_only_model(
765
766
        self,
        architectures: Union[str, list[str]],
767
        model_config: ModelConfig,
768
    ) -> bool:
769
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
770
        return model_cls.supports_multimodal_raw_input_only
771

772
773
    def is_pp_supported_model(
        self,
774
        architectures: Union[str, list[str]],
775
        model_config: ModelConfig,
776
    ) -> bool:
777
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
778
        return model_cls.supports_pp
779

780
781
    def model_has_inner_state(
        self,
782
        architectures: Union[str, list[str]],
783
        model_config: ModelConfig,
784
    ) -> bool:
785
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
786
        return model_cls.has_inner_state
787

788
789
    def is_attention_free_model(
        self,
790
        architectures: Union[str, list[str]],
791
        model_config: ModelConfig,
792
    ) -> bool:
793
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
794
        return model_cls.is_attention_free
795

796
797
    def is_hybrid_model(
        self,
798
        architectures: Union[str, list[str]],
799
        model_config: ModelConfig,
800
    ) -> bool:
801
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
802
803
        return model_cls.is_hybrid

804
805
    def is_noops_model(
        self,
806
        architectures: Union[str, list[str]],
807
        model_config: ModelConfig,
808
    ) -> bool:
809
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
810
811
        return model_cls.has_noops

812
813
    def is_transcription_model(
        self,
814
        architectures: Union[str, list[str]],
815
        model_config: ModelConfig,
816
    ) -> bool:
817
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
818
819
        return model_cls.supports_transcription

820
821
822
    def is_transcription_only_model(
        self,
        architectures: Union[str, list[str]],
823
        model_config: ModelConfig,
824
    ) -> bool:
825
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
826
827
        return model_cls.supports_transcription_only

828
829
    def is_v1_compatible(
        self,
830
        architectures: Union[str, list[str]],
831
        model_config: ModelConfig,
832
    ) -> bool:
833
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
834
835
        return not model_cls.supports_v0_only

836
837

ModelRegistry = _ModelRegistry({
838
839
    model_arch:
    _LazyRegisteredModel(
840
841
842
843
844
845
846
847
848
849
        module_name=f"vllm.model_executor.models.{mod_relname}",
        class_name=cls_name,
    )
    for model_arch, (mod_relname, cls_name) in _VLLM_MODELS.items()
})

_T = TypeVar("_T")


def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
850
851
852
853
854
    # 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")

855
        # `cloudpickle` allows pickling lambda functions directly
856
        import cloudpickle
857
        input_bytes = cloudpickle.dumps((fn, output_filepath))
858
859
860

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
861
862
863
        returned = subprocess.run(_SUBPROCESS_COMMAND,
                                  input=input_bytes,
                                  capture_output=True)
864
865
866
867
868
869
870
871
872

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

873
        with open(output_filepath, "rb") as f:
874
875
876
877
878
879
880
881
882
883
884
            return pickle.load(f)


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

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

    result = fn()
885
886
887

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
888
889
890


if __name__ == "__main__":
891
    _run()