registry.py 38.9 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
    "FM9GForCausalLM": ("fm9g", "FM9GForCausalLM"),
73
    "Fairseq2LlamaForCausalLM": ("fairseq2_llama", "Fairseq2LlamaForCausalLM"),
74
75
    "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
    "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
76
    "Gemma3ForCausalLM": ("gemma3", "Gemma3ForCausalLM"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
77
    "Gemma3nForCausalLM": ("gemma3n", "Gemma3nForCausalLM"),
78
    "Qwen3NextForCausalLM": ("qwen3_next", "Qwen3NextForCausalLM"),
79
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
80
    "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
zhuwenwen's avatar
zhuwenwen committed
81
    "Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
82
    "GptOssForCausalLM": ("gpt_oss", "GptOssForCausalLM"),
83
84
85
86
87
88
    "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
    "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
    "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
    "GPTNeoXForCausalLM": ("gpt_neox", "GPTNeoXForCausalLM"),
    "GraniteForCausalLM": ("granite", "GraniteForCausalLM"),
    "GraniteMoeForCausalLM": ("granitemoe", "GraniteMoeForCausalLM"),
89
    "GraniteMoeHybridForCausalLM": ("granitemoehybrid", "GraniteMoeHybridForCausalLM"),   # noqa: E501
90
    "GraniteMoeSharedForCausalLM": ("granitemoeshared", "GraniteMoeSharedForCausalLM"),   # noqa: E501
91
    "GritLM": ("gritlm", "GritLM"),
Michael Goin's avatar
Michael Goin committed
92
    "Grok1ModelForCausalLM": ("grok1", "Grok1ForCausalLM"),
93
94
    "HunYuanMoEV1ForCausalLM": ("hunyuan_v1", "HunYuanMoEV1ForCausalLM"),
    "HunYuanDenseV1ForCausalLM": ("hunyuan_v1", "HunYuanDenseV1ForCausalLM"),
95
    "HCXVisionForCausalLM": ("hyperclovax_vision", "HCXVisionForCausalLM"),
96
97
    "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
    "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"),
98
    "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
99
    "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
100
101
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
102
    "Lfm2ForCausalLM": ("lfm2", "Lfm2ForCausalLM"),
103
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
104
    "Llama4ForCausalLM": ("llama4", "Llama4ForCausalLM"),  # noqa: E501
105
106
    # For decapoda-research/llama-*
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
zhuwenwen's avatar
zhuwenwen committed
107
    "LongcatFlashForCausalLM": ("longcat_flash", "LongcatFlashForCausalLM"),
108
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
109
    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
Dhia Eddine Rhaiem's avatar
Dhia Eddine Rhaiem committed
110
    "FalconH1ForCausalLM":("falcon_h1", "FalconH1ForCausalLM"),
111
    "Mamba2ForCausalLM": ("mamba2", "Mamba2ForCausalLM"),
112
113
    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
    "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
114
115
    "MistralForCausalLM": ("llama", "LlamaForCausalLM"),
    "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
116
    "MotifForCausalLM": ("motif", "MotifForCausalLM"),
117
118
119
    # transformers's mpt class has lower case
    "MptForCausalLM": ("mpt", "MPTForCausalLM"),
    "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
120
    "MiMoForCausalLM": ("mimo", "MiMoForCausalLM"),
121
    "NemotronForCausalLM": ("nemotron", "NemotronForCausalLM"),
Luis Vega's avatar
Luis Vega committed
122
    "NemotronHForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
123
    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
124
    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
125
126
127
128
129
130
131
    "OlmoeForCausalLM": ("olmoe", "OlmoeForCausalLM"),
    "OPTForCausalLM": ("opt", "OPTForCausalLM"),
    "OrionForCausalLM": ("orion", "OrionForCausalLM"),
    "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
    "PhiForCausalLM": ("phi", "PhiForCausalLM"),
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
    "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
132
    "Phi4FlashForCausalLM": ("phi4flash", "Phi4FlashForCausalLM"),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
133
    "Plamo2ForCausalLM": ("plamo2", "Plamo2ForCausalLM"),
134
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
135
136
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
137
138
    "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),
    "Qwen3MoeForCausalLM": ("qwen3_moe", "Qwen3MoeForCausalLM"),
139
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
140
    "SeedOssForCausalLM": ("seed_oss", "SeedOssForCausalLM"),
zhuwenwen's avatar
zhuwenwen committed
141
    "Step3TextForCausalLM": ("step3_text", "Step3TextForCausalLM"),
142
143
144
145
    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
146
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
147
    "TeleFLMForCausalLM": ("teleflm", "TeleFLMForCausalLM"),
148
    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
149
    "Zamba2ForCausalLM": ("zamba2", "Zamba2ForCausalLM"),
150
151
152
    # [Encoder-decoder]
    "BartModel": ("bart", "BartForConditionalGeneration"),
    "BartForConditionalGeneration": ("bart", "BartForConditionalGeneration"),
汪志鹏's avatar
汪志鹏 committed
153
    "MBartForConditionalGeneration": ("bart", "MBartForConditionalGeneration"),
154
155
156
}

_EMBEDDING_MODELS = {
157
    # [Text-only]
158
    "BertModel": ("bert", "BertEmbeddingModel"),
159
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
160
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
161
    "Gemma3TextModel": ("gemma3", "Gemma3Model"),
162
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
163
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
164
    "GritLM": ("gritlm", "GritLM"),
165
166
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
167
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
168
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
169
    "LlamaModel": ("llama", "LlamaForCausalLM"),
170
171
172
173
174
    **{
        # 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"
    },
175
    "MistralModel": ("llama", "LlamaForCausalLM"),
176
    "ModernBertModel": ("modernbert", "ModernBertModel"),
177
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
178
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
179
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
180
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
181
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
182
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
183
184
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
185
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
186
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
187
    # [Multimodal]
Cyrus Leung's avatar
Cyrus Leung committed
188
    "LlavaNextForConditionalGeneration": ("llava_next", "LlavaNextForConditionalGeneration"),  # noqa: E501
189
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
190
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
191
    # Technically Terratorch models work on images, both in
192
193
    # input and output. I am adding it here because it piggy-backs on embedding
    # models for the time being.
194
195
    "PrithviGeoSpatialMAE": ("terratorch", "Terratorch"),
    "Terratorch": ("terratorch", "Terratorch"),
196
197
}

198
199
_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
200
201
202
203
    "GteNewForSequenceClassification": ("bert_with_rope",
                                        "GteNewForSequenceClassification"),
    "ModernBertForSequenceClassification": ("modernbert",
                                            "ModernBertForSequenceClassification"),
204
205
206
207
    "RobertaForSequenceClassification": ("roberta",
                                         "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": ("roberta",
                                            "RobertaForSequenceClassification"),
208
    # [Auto-converted (see adapters.py)]
209
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"), # noqa: E501,
210
211
}

212
_MULTIMODAL_MODELS = {
213
    # [Decoder-only]
214
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
Jennifer Zhao's avatar
Jennifer Zhao committed
215
    "AyaVisionForConditionalGeneration": ("aya_vision", "AyaVisionForConditionalGeneration"),  # noqa: E501
216
217
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
    "ChameleonForConditionalGeneration": ("chameleon", "ChameleonForConditionalGeneration"),  # noqa: E501
218
    "Cohere2VisionForConditionalGeneration": ("cohere2_vision", "Cohere2VisionForConditionalGeneration"),  # noqa: E501
219
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
220
    "Ernie4_5_VLMoeForConditionalGeneration": ("ernie45_vl", "Ernie4_5_VLMoeForConditionalGeneration"),  # noqa: E501
221
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
222
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
223
    "Gemma3nForConditionalGeneration": ("gemma3n_mm", "Gemma3nForConditionalGeneration"),    # noqa: E501
224
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
225
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),  # noqa: E501
Jee Jee Li's avatar
Jee Jee Li committed
226
    "Glm4vMoeForConditionalGeneration": ("glm4_1v", "Glm4vMoeForConditionalGeneration"),  # noqa: E501
227
    "GraniteSpeechForConditionalGeneration": ("granite_speech", "GraniteSpeechForConditionalGeneration"),  # noqa: E501
228
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
229
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
230
    "NemotronH_Nano_VL": ("nano_nemotron_vl", "NemotronH_Nano_VL"),
Lyu Han's avatar
Lyu Han committed
231
    "InternS1ForConditionalGeneration": ("interns1", "InternS1ForConditionalGeneration"),  # noqa: E501
232
    "InternVLForConditionalGeneration": ("interns1", "InternS1ForConditionalGeneration"),  # noqa: E501
233
    "Idefics3ForConditionalGeneration":("idefics3","Idefics3ForConditionalGeneration"),
234
    "SmolVLMForConditionalGeneration": ("smolvlm","SmolVLMForConditionalGeneration"),  # noqa: E501
235
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
236
    "KeyeVL1_5ForConditionalGeneration": ("keye_vl1_5", "KeyeVL1_5ForConditionalGeneration"), # noqa: E501
237
    "RForConditionalGeneration": ("rvl", "RForConditionalGeneration"),
238
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
239
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
240
241
242
243
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
    "LlavaNextForConditionalGeneration": ("llava_next", "LlavaNextForConditionalGeneration"),  # noqa: E501
    "LlavaNextVideoForConditionalGeneration": ("llava_next_video", "LlavaNextVideoForConditionalGeneration"),  # noqa: E501
    "LlavaOnevisionForConditionalGeneration": ("llava_onevision", "LlavaOnevisionForConditionalGeneration"),  # noqa: E501
244
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
245
    "MiDashengLMModel": ("midashenglm", "MiDashengLMModel"),
246
    "MiniMaxVL01ForConditionalGeneration": ("minimax_vl_01", "MiniMaxVL01ForConditionalGeneration"),  # noqa: E501
247
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
248
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
249
    "Mistral3ForConditionalGeneration": ("mistral3", "Mistral3ForConditionalGeneration"),  # noqa: E501
250
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
251
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
252
    "Ovis": ("ovis", "Ovis"),
253
    "Ovis2_5": ("ovis2_5", "Ovis2_5"),
254
    "PaliGemmaForConditionalGeneration": ("paligemma", "PaliGemmaForConditionalGeneration"),  # noqa: E501
255
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
256
257
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
    "Phi4MultimodalForCausalLM": ("phi4_multimodal", "Phi4MultimodalForCausalLM"),  # noqa: E501
258
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
259
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),  # noqa: E501
260
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
Roger Wang's avatar
Roger Wang committed
261
    "Qwen2_5_VLForConditionalGeneration": ("qwen2_5_vl", "Qwen2_5_VLForConditionalGeneration"),  # noqa: E501
262
    "Qwen2AudioForConditionalGeneration": ("qwen2_audio", "Qwen2AudioForConditionalGeneration"),  # noqa: E501
263
    "Qwen2_5OmniModel": ("qwen2_5_omni_thinker", "Qwen2_5OmniThinkerForConditionalGeneration"),  # noqa: E501
264
    "Qwen2_5OmniForConditionalGeneration": ("qwen2_5_omni_thinker", "Qwen2_5OmniThinkerForConditionalGeneration"),  # noqa: E501
265
    "UltravoxModel": ("ultravox", "UltravoxModel"),
zhuwenwen's avatar
zhuwenwen committed
266
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),  # noqa: E501
汪志鹏's avatar
汪志鹏 committed
267
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),  # noqa: E501
268
    "Tarsier2ForConditionalGeneration": ("qwen2_vl", "Tarsier2ForConditionalGeneration"),  # noqa: E501
Patrick von Platen's avatar
Patrick von Platen committed
269
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),  # noqa: E501
270
    # [Encoder-decoder]
汪志鹏's avatar
汪志鹏 committed
271
    "DonutForConditionalGeneration": ("donut", "DonutForConditionalGeneration"),
272
    "Florence2ForConditionalGeneration": ("florence2", "Florence2ForConditionalGeneration"),  # noqa: E501
273
    "MllamaForConditionalGeneration": ("mllama", "MllamaForConditionalGeneration"),  # noqa: E501
274
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),  # noqa: E501
275
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
276
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
277
}
278
279

_SPECULATIVE_DECODING_MODELS = {
280
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
281
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
282
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
283
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
284
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
285
    "LlamaForCausalLMEagle3": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
286
    "EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
287
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
288
    "ErnieMTPModel": ("ernie_mtp", "ErnieMTP"),
zhuwenwen's avatar
zhuwenwen committed
289
    "LongCatFlashMTPModel": ("longcat_flash_mtp", "LongCatFlashMTP"),
zhuwenwen's avatar
zhuwenwen committed
290
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
291
    "MedusaModel": ("medusa", "Medusa"),
292
    "Qwen3NextMTP": ("qwen3_next_mtp", "Qwen3NextMTP"),
293
294
295
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
296
}
297

298
_TRANSFORMERS_SUPPORTED_MODELS = {
299
300
301
    # Text generation models
    "SmolLM3ForCausalLM": ("transformers", "TransformersForCausalLM"),
    # Multimodal models
302
303
304
305
    "Emu3ForConditionalGeneration": ("transformers", "TransformersForMultimodalLM"),  # noqa: E501
}

_TRANSFORMERS_BACKEND_MODELS = {
306
    "TransformersModel": ("transformers", "TransformersModel"),
307
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
308
    "TransformersForMultimodalLM": ("transformers", "TransformersForMultimodalLM"), # noqa: E501
309
}
310
# yapf: enable
311

312
_VLLM_MODELS = {
313
    **_TEXT_GENERATION_MODELS,
314
    **_EMBEDDING_MODELS,
315
    **_CROSS_ENCODER_MODELS,
316
    **_MULTIMODAL_MODELS,
317
    **_SPECULATIVE_DECODING_MODELS,
318
319
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
320
321
}

322
323
324
325
326
327
328
329
# 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"
]

330
331
_PREVIOUSLY_SUPPORTED_MODELS = {"Phi3SmallForCausalLM": "0.9.2"}

332

333
334
@dataclass(frozen=True)
class _ModelInfo:
335
    architecture: str
336
    is_text_generation_model: bool
337
    is_pooling_model: bool
338
    default_pooling_type: str
339
    supports_cross_encoding: bool
340
    supports_multimodal: bool
341
    supports_multimodal_raw_input_only: bool
342
    supports_multimodal_encoder_tp_data: bool
343
    supports_pp: bool
344
345
    has_inner_state: bool
    is_attention_free: bool
346
    is_hybrid: bool
347
    has_noops: bool
348
    supports_transcription: bool
349
    supports_transcription_only: bool
350
    supports_v0_only: bool
351
352

    @staticmethod
353
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
354
        return _ModelInfo(
355
            architecture=model.__name__,
356
            is_text_generation_model=is_text_generation_model(model),
357
            is_pooling_model=is_pooling_model(model),
358
            default_pooling_type=get_default_pooling_type(model),
359
            supports_cross_encoding=supports_cross_encoding(model),
360
            supports_multimodal=supports_multimodal(model),
361
362
            supports_multimodal_raw_input_only=
            supports_multimodal_raw_input_only(model),
363
364
            supports_multimodal_encoder_tp_data=
            supports_multimodal_encoder_tp_data(model),
365
            supports_pp=supports_pp(model),
366
367
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
368
            is_hybrid=is_hybrid(model),
369
            supports_transcription=supports_transcription(model),
370
371
            supports_transcription_only=(supports_transcription(model) and
                                         model.supports_transcription_only),
372
            supports_v0_only=supports_v0_only(model),
373
            has_noops=has_noops(model),
374
        )
375
376


377
class _BaseRegisteredModel(ABC):
378

379
380
381
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
382

383
    @abstractmethod
384
    def load_model_cls(self) -> type[nn.Module]:
385
        raise NotImplementedError
386
387


388
389
390
391
392
393
394
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
395
    model_cls: type[nn.Module]
396
397

    @staticmethod
398
    def from_model_cls(model_cls: type[nn.Module]):
399
400
401
402
403
404
405
406
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

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

407
    def load_model_cls(self) -> type[nn.Module]:
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
        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()))

424
    def load_model_cls(self) -> type[nn.Module]:
425
426
427
428
429
430
431
432
        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,
433
) -> Optional[type[nn.Module]]:
434
    from vllm.platforms import current_platform
435
    current_platform.verify_model_arch(model_arch)
436
437
438
439
440
441
    try:
        return model.load_model_cls()
    except Exception:
        logger.exception("Error in loading model architecture '%s'",
                         model_arch)
        return None
442
443


444
445
446
447
448
449
450
451
452
453
454
@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
455
456


457
458
459
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
460
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
461

462
    def get_supported_archs(self) -> Set[str]:
463
        return self.models.keys()
464

465
466
467
    def register_model(
        self,
        model_arch: str,
468
        model_cls: Union[type[nn.Module], str],
469
    ) -> None:
470
471
472
        """
        Register an external model to be used in vLLM.

473
        `model_cls` can be either:
474

475
        - A [`torch.nn.Module`][] class directly referencing the model.
476
        - A string in the format `<module>:<class>` which can be used to
477
478
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
479
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
480
        """
481
482
483
484
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

485
        if model_arch in self.models:
486
487
488
            logger.warning(
                "Model architecture %s is already registered, and will be "
                "overwritten by the new model class %s.", model_arch,
489
490
491
492
493
494
495
                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)
496

497
            model = _LazyRegisteredModel(*split_str)
498
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
499
            model = _RegisteredModel.from_model_cls(model_cls)
500
501
502
503
        else:
            msg = ("`model_cls` should be a string or PyTorch model class, "
                   f"not a {type(model_arch)}")
            raise TypeError(msg)
504

505
        self.models[model_arch] = model
506

507
    def _raise_for_unsupported(self, architectures: list[str]):
508
        all_supported_archs = self.get_supported_archs()
509

510
511
512
513
514
        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.")

515
516
517
518
519
520
521
522
523
524
        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.")

525
526
527
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
            f"Supported architectures: {all_supported_archs}")
528

529
    def _try_load_model_cls(self,
530
                            model_arch: str) -> Optional[type[nn.Module]]:
531
532
        if model_arch not in self.models:
            return None
533

534
        return _try_load_model_cls(model_arch, self.models[model_arch])
535

536
    def _try_inspect_model_cls(self, model_arch: str) -> Optional[_ModelInfo]:
537
538
        if model_arch not in self.models:
            return None
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
594
595
596
        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:
597
                return None
598

599
600
601
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
                "is not compatible with vLLM.")
602

603
        return model_config._get_transformers_backend_cls()
604

605
    def _normalize_arch(
606
        self,
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
        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
622

623
624
625
626
627
            # 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
628

629
        return architecture
630

631
632
    def inspect_model_cls(
        self,
633
        architectures: Union[str, list[str]],
634
        model_config: ModelConfig,
635
    ) -> tuple[_ModelInfo, str]:
636
637
        if isinstance(architectures, str):
            architectures = [architectures]
638
639
        if not architectures:
            raise ValueError("No model architectures are specified")
640
641
642
643
644
645
646
647
648

        # 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)
649
650
651
        elif model_config.model_impl == ModelImpl.TERRATORCH:
            model_info = self._try_inspect_model_cls("Terratorch")
            return (model_info, "Terratorch")
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)
663

664
        for arch in architectures:
665
            normalized_arch = self._normalize_arch(arch, model_config)
666
            model_info = self._try_inspect_model_cls(normalized_arch)
667
            if model_info is not None:
668
                return (model_info, arch)
669

670
671
672
        # 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):
673
674
675
676
677
678
679
            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)

680
        return self._raise_for_unsupported(architectures)
681

682
683
    def resolve_model_cls(
        self,
684
        architectures: Union[str, list[str]],
685
        model_config: ModelConfig,
686
    ) -> tuple[type[nn.Module], str]:
687
688
        if isinstance(architectures, str):
            architectures = [architectures]
689
690
        if not architectures:
            raise ValueError("No model architectures are specified")
691
692
693
694
695
696
697
698
699

        # 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)
700
701
702
703
704
        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)
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)
716

717
        for arch in architectures:
718
            normalized_arch = self._normalize_arch(arch, model_config)
719
            model_cls = self._try_load_model_cls(normalized_arch)
720
721
            if model_cls is not None:
                return (model_cls, arch)
722

723
724
725
        # 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):
726
727
728
729
730
731
732
            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)

733
        return self._raise_for_unsupported(architectures)
734

735
736
    def is_text_generation_model(
        self,
737
        architectures: Union[str, list[str]],
738
        model_config: ModelConfig,
739
    ) -> bool:
740
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
741
        return model_cls.is_text_generation_model
742

743
    def is_pooling_model(
744
        self,
745
        architectures: Union[str, list[str]],
746
        model_config: ModelConfig,
747
    ) -> bool:
748
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
749
        return model_cls.is_pooling_model
750

751
752
    def is_cross_encoder_model(
        self,
753
        architectures: Union[str, list[str]],
754
        model_config: ModelConfig,
755
    ) -> bool:
756
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
757
        return model_cls.supports_cross_encoding
758

759
760
    def is_multimodal_model(
        self,
761
        architectures: Union[str, list[str]],
762
        model_config: ModelConfig,
763
    ) -> bool:
764
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
765
        return model_cls.supports_multimodal
766

767
    def is_multimodal_raw_input_only_model(
768
769
        self,
        architectures: Union[str, list[str]],
770
        model_config: ModelConfig,
771
    ) -> bool:
772
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
773
        return model_cls.supports_multimodal_raw_input_only
774

775
776
    def is_pp_supported_model(
        self,
777
        architectures: Union[str, list[str]],
778
        model_config: ModelConfig,
779
    ) -> bool:
780
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
781
        return model_cls.supports_pp
782

783
784
    def model_has_inner_state(
        self,
785
        architectures: Union[str, list[str]],
786
        model_config: ModelConfig,
787
    ) -> bool:
788
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
789
        return model_cls.has_inner_state
790

791
792
    def is_attention_free_model(
        self,
793
        architectures: Union[str, list[str]],
794
        model_config: ModelConfig,
795
    ) -> bool:
796
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
797
        return model_cls.is_attention_free
798

799
800
    def is_hybrid_model(
        self,
801
        architectures: Union[str, list[str]],
802
        model_config: ModelConfig,
803
    ) -> bool:
804
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
805
806
        return model_cls.is_hybrid

807
808
    def is_noops_model(
        self,
809
        architectures: Union[str, list[str]],
810
        model_config: ModelConfig,
811
    ) -> bool:
812
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
813
814
        return model_cls.has_noops

815
816
    def is_transcription_model(
        self,
817
        architectures: Union[str, list[str]],
818
        model_config: ModelConfig,
819
    ) -> bool:
820
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
821
822
        return model_cls.supports_transcription

823
824
825
    def is_transcription_only_model(
        self,
        architectures: Union[str, list[str]],
826
        model_config: ModelConfig,
827
    ) -> bool:
828
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
829
830
        return model_cls.supports_transcription_only

831
832
    def is_v1_compatible(
        self,
833
        architectures: Union[str, list[str]],
834
        model_config: ModelConfig,
835
    ) -> bool:
836
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
837
838
        return not model_cls.supports_v0_only

839
840

ModelRegistry = _ModelRegistry({
841
842
    model_arch:
    _LazyRegisteredModel(
843
844
845
846
847
848
849
850
851
852
        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:
853
854
855
856
857
    # 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")

858
        # `cloudpickle` allows pickling lambda functions directly
859
        import cloudpickle
860
        input_bytes = cloudpickle.dumps((fn, output_filepath))
861
862
863

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
864
865
866
        returned = subprocess.run(_SUBPROCESS_COMMAND,
                                  input=input_bytes,
                                  capture_output=True)
867
868
869
870
871
872
873
874
875

        # 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

876
        with open(output_filepath, "rb") as f:
877
878
879
880
881
882
883
884
885
886
887
            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()
888
889
890

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
891
892
893


if __name__ == "__main__":
894
    _run()