registry.py 36.2 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
5
6
"""
Whenever you add an architecture to this page, please also update
`tests/models/registry.py` with example HuggingFace models for it.
"""
7
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_raw_input,
                         supports_pp, supports_transcription, supports_v0_only)
32
from .interfaces_base import is_pooling_model, is_text_generation_model
33
34
35

logger = init_logger(__name__)

36
# yapf: disable
37
38
_TEXT_GENERATION_MODELS = {
    # [Decoder-only]
39
40
    "AquilaModel": ("llama", "LlamaForCausalLM"),
    "AquilaForCausalLM": ("llama", "LlamaForCausalLM"),  # AquilaChat2
Raghav Ravishankar's avatar
Raghav Ravishankar committed
41
    "ArceeForCausalLM": ("arcee", "ArceeForCausalLM"),
42
    "ArcticForCausalLM": ("arctic", "ArcticForCausalLM"),
43
    "MiniMaxForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
44
    "MiniMaxText01ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
45
    "MiniMaxM1ForCausalLM": ("minimax_text_01", "MiniMaxText01ForCausalLM"),
46
47
48
49
    # baichuan-7b, upper case 'C' in the class name
    "BaiChuanForCausalLM": ("baichuan", "BaiChuanForCausalLM"),
    # baichuan-13b, lower case 'c' in the class name
    "BaichuanForCausalLM": ("baichuan", "BaichuanForCausalLM"),
50
    "BailingMoeForCausalLM": ("bailing_moe", "BailingMoeForCausalLM"),
Yu Chin Fabian Lim's avatar
Yu Chin Fabian Lim committed
51
    "BambaForCausalLM": ("bamba", "BambaForCausalLM"),
52
    "BloomForCausalLM": ("bloom", "BloomForCausalLM"),
53
    "ChatGLMModel": ("chatglm", "ChatGLMForCausalLM"),
54
    "ChatGLMForConditionalGeneration": ("chatglm", "ChatGLMForCausalLM"),
55
    "CohereForCausalLM": ("commandr", "CohereForCausalLM"),
56
    "Cohere2ForCausalLM": ("commandr", "CohereForCausalLM"),
57
    "DbrxForCausalLM": ("dbrx", "DbrxForCausalLM"),
58
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
59
60
    "DeepseekForCausalLM": ("deepseek", "DeepseekForCausalLM"),
    "DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
61
    "DeepseekV3ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
62
    "Dots1ForCausalLM": ("dots1", "Dots1ForCausalLM"),
63
    "Ernie4_5ForCausalLM": ("ernie45", "Ernie4_5ForCausalLM"),
64
    "Ernie4_5_MoeForCausalLM": ("ernie45_moe", "Ernie4_5_MoeForCausalLM"),
65
    "ExaoneForCausalLM": ("exaone", "ExaoneForCausalLM"),
66
    "Exaone4ForCausalLM": ("exaone4", "Exaone4ForCausalLM"),
67
    "FalconForCausalLM": ("falcon", "FalconForCausalLM"),
68
    "Fairseq2LlamaForCausalLM": ("fairseq2_llama", "Fairseq2LlamaForCausalLM"),
69
70
    "GemmaForCausalLM": ("gemma", "GemmaForCausalLM"),
    "Gemma2ForCausalLM": ("gemma2", "Gemma2ForCausalLM"),
71
    "Gemma3ForCausalLM": ("gemma3", "Gemma3ForCausalLM"),
Robert Shaw's avatar
Robert Shaw committed
72
73
    #TODO(ywang96): Support multimodal gemma3n
    "Gemma3nForConditionalGeneration": ("gemma3n", "Gemma3nForConditionalGeneration"),    # noqa: E501
74
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
75
    "Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
76
    "Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
77
78
79
80
81
82
    "GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
    "GPTBigCodeForCausalLM": ("gpt_bigcode", "GPTBigCodeForCausalLM"),
    "GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
    "GPTNeoXForCausalLM": ("gpt_neox", "GPTNeoXForCausalLM"),
    "GraniteForCausalLM": ("granite", "GraniteForCausalLM"),
    "GraniteMoeForCausalLM": ("granitemoe", "GraniteMoeForCausalLM"),
83
    "GraniteMoeHybridForCausalLM": ("granitemoehybrid", "GraniteMoeHybridForCausalLM"),   # noqa: E501
84
    "GraniteMoeSharedForCausalLM": ("granitemoeshared", "GraniteMoeSharedForCausalLM"),   # noqa: E501
85
    "GritLM": ("gritlm", "GritLM"),
Michael Goin's avatar
Michael Goin committed
86
    "Grok1ModelForCausalLM": ("grok1", "Grok1ForCausalLM"),
87
88
    "HunYuanMoEV1ForCausalLM": ("hunyuan_v1", "HunYuanMoEV1ForCausalLM"),
    "HunYuanDenseV1ForCausalLM": ("hunyuan_v1", "HunYuanDenseV1ForCausalLM"),
89
    "HCXVisionForCausalLM": ("hyperclovax_vision", "HCXVisionForCausalLM"),
90
91
    "InternLMForCausalLM": ("llama", "LlamaForCausalLM"),
    "InternLM2ForCausalLM": ("internlm2", "InternLM2ForCausalLM"),
92
    "InternLM2VEForCausalLM": ("internlm2_ve", "InternLM2VEForCausalLM"),
93
    "InternLM3ForCausalLM": ("llama", "LlamaForCausalLM"),
94
95
96
    "JAISLMHeadModel": ("jais", "JAISLMHeadModel"),
    "JambaForCausalLM": ("jamba", "JambaForCausalLM"),
    "LlamaForCausalLM": ("llama", "LlamaForCausalLM"),
97
    "Llama4ForCausalLM": ("llama4", "Llama4ForCausalLM"),  # noqa: E501
98
99
    # For decapoda-research/llama-*
    "LLaMAForCausalLM": ("llama", "LlamaForCausalLM"),
100
    "MambaForCausalLM": ("mamba", "MambaForCausalLM"),
101
    "FalconMambaForCausalLM": ("mamba", "MambaForCausalLM"),
Dhia Eddine Rhaiem's avatar
Dhia Eddine Rhaiem committed
102
    "FalconH1ForCausalLM":("falcon_h1", "FalconH1ForCausalLM"),
103
    "Mamba2ForCausalLM": ("mamba2", "Mamba2ForCausalLM"),
104
105
    "MiniCPMForCausalLM": ("minicpm", "MiniCPMForCausalLM"),
    "MiniCPM3ForCausalLM": ("minicpm3", "MiniCPM3ForCausalLM"),
106
107
108
109
110
111
    "MistralForCausalLM": ("llama", "LlamaForCausalLM"),
    "MixtralForCausalLM": ("mixtral", "MixtralForCausalLM"),
    "QuantMixtralForCausalLM": ("mixtral_quant", "MixtralForCausalLM"),
    # transformers's mpt class has lower case
    "MptForCausalLM": ("mpt", "MPTForCausalLM"),
    "MPTForCausalLM": ("mpt", "MPTForCausalLM"),
112
    "MiMoForCausalLM": ("mimo", "MiMoForCausalLM"),
113
    "NemotronForCausalLM": ("nemotron", "NemotronForCausalLM"),
Luis Vega's avatar
Luis Vega committed
114
    "NemotronHForCausalLM": ("nemotron_h", "NemotronHForCausalLM"),
115
    "OlmoForCausalLM": ("olmo", "OlmoForCausalLM"),
116
    "Olmo2ForCausalLM": ("olmo2", "Olmo2ForCausalLM"),
117
118
119
120
121
122
123
    "OlmoeForCausalLM": ("olmoe", "OlmoeForCausalLM"),
    "OPTForCausalLM": ("opt", "OPTForCausalLM"),
    "OrionForCausalLM": ("orion", "OrionForCausalLM"),
    "PersimmonForCausalLM": ("persimmon", "PersimmonForCausalLM"),
    "PhiForCausalLM": ("phi", "PhiForCausalLM"),
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
    "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
124
    "Phi4FlashForCausalLM": ("phi4flash", "Phi4FlashForCausalLM"),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
125
    "Plamo2ForCausalLM": ("plamo2", "Plamo2ForCausalLM"),
126
    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
127
128
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
    "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
129
130
    "Qwen3ForCausalLM": ("qwen3", "Qwen3ForCausalLM"),
    "Qwen3MoeForCausalLM": ("qwen3_moe", "Qwen3MoeForCausalLM"),
131
    "RWForCausalLM": ("falcon", "FalconForCausalLM"),
Song's avatar
Song committed
132
    "Step3TextForCausalLM": ("step3_text", "Step3TextForCausalLM"),
133
134
135
136
    "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"),
    "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"),
    "SolarForCausalLM": ("solar", "SolarForCausalLM"),
137
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
138
    "TeleFLMForCausalLM": ("teleflm", "TeleFLMForCausalLM"),
139
    "XverseForCausalLM": ("llama", "LlamaForCausalLM"),
140
    "Zamba2ForCausalLM": ("zamba2", "Zamba2ForCausalLM"),
141
142
143
    # [Encoder-decoder]
    "BartModel": ("bart", "BartForConditionalGeneration"),
    "BartForConditionalGeneration": ("bart", "BartForConditionalGeneration"),
144
145
146
}

_EMBEDDING_MODELS = {
147
    # [Text-only]
148
    "BertModel": ("bert", "BertEmbeddingModel"),
149
    "DeciLMForCausalLM": ("nemotron_nas", "DeciLMForCausalLM"),
150
    "Gemma2Model": ("gemma2", "Gemma2ForCausalLM"),
151
    "GlmForCausalLM": ("glm", "GlmForCausalLM"),
152
    "GPT2ForSequenceClassification": ("gpt2", "GPT2ForSequenceClassification"),
153
    "GritLM": ("gritlm", "GritLM"),
154
155
    "GteModel": ("bert_with_rope", "SnowflakeGteNewModel"),
    "GteNewModel": ("bert_with_rope", "GteNewModel"),
156
    "InternLM2ForRewardModel": ("internlm2", "InternLM2ForRewardModel"),
157
    "JambaForSequenceClassification": ("jamba", "JambaForSequenceClassification"),  # noqa: E501
158
    "LlamaModel": ("llama", "LlamaForCausalLM"),
159
160
161
162
163
    **{
        # 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"
    },
164
    "MistralModel": ("llama", "LlamaForCausalLM"),
165
    "ModernBertModel": ("modernbert", "ModernBertModel"),
166
    "NomicBertModel": ("bert_with_rope", "NomicBertModel"),
167
    "Phi3ForCausalLM": ("phi3", "Phi3ForCausalLM"),
168
    "Qwen2Model": ("qwen2", "Qwen2ForCausalLM"),
169
    "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
170
    "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
171
    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
172
173
    "RobertaForMaskedLM": ("roberta", "RobertaEmbeddingModel"),
    "RobertaModel": ("roberta", "RobertaEmbeddingModel"),
174
    "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
175
    "XLMRobertaModel": ("roberta", "RobertaEmbeddingModel"),
176
    # [Multimodal]
Cyrus Leung's avatar
Cyrus Leung committed
177
    "LlavaNextForConditionalGeneration": ("llava_next", "LlavaNextForConditionalGeneration"),  # noqa: E501
178
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
179
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
180
181
182
183
    # Technically PrithviGeoSpatialMAE is a model that works on images, both in
    # input and output. I am adding it here because it piggy-backs on embedding
    # models for the time being.
    "PrithviGeoSpatialMAE": ("prithvi_geospatial_mae", "PrithviGeoSpatialMAE"),
184
185
}

186
187
188
189
190
191
_CROSS_ENCODER_MODELS = {
    "BertForSequenceClassification": ("bert", "BertForSequenceClassification"),
    "RobertaForSequenceClassification": ("roberta",
                                         "RobertaForSequenceClassification"),
    "XLMRobertaForSequenceClassification": ("roberta",
                                            "RobertaForSequenceClassification"),
xsank's avatar
xsank committed
192
193
    "ModernBertForSequenceClassification": ("modernbert",
                                            "ModernBertForSequenceClassification"),
194
    # [Auto-converted (see adapters.py)]
195
    "JinaVLForRanking": ("jina_vl", "JinaVLForSequenceClassification"), # noqa: E501,
196
197
}

198
_MULTIMODAL_MODELS = {
199
    # [Decoder-only]
200
    "AriaForConditionalGeneration": ("aria", "AriaForConditionalGeneration"),
Jennifer Zhao's avatar
Jennifer Zhao committed
201
    "AyaVisionForConditionalGeneration": ("aya_vision", "AyaVisionForConditionalGeneration"),  # noqa: E501
202
203
    "Blip2ForConditionalGeneration": ("blip2", "Blip2ForConditionalGeneration"),
    "ChameleonForConditionalGeneration": ("chameleon", "ChameleonForConditionalGeneration"),  # noqa: E501
204
    "DeepseekVLV2ForCausalLM": ("deepseek_vl2", "DeepseekVLV2ForCausalLM"),
205
    "FuyuForCausalLM": ("fuyu", "FuyuForCausalLM"),
206
    "Gemma3ForConditionalGeneration": ("gemma3_mm", "Gemma3ForConditionalGeneration"),  # noqa: E501
207
    "GLM4VForCausalLM": ("glm4v", "GLM4VForCausalLM"),
208
    "Glm4vForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),  # noqa: E501
Yuxuan Zhang's avatar
Yuxuan Zhang committed
209
    "Glm4v_moeForConditionalGeneration": ("glm4_1v", "Glm4vForConditionalGeneration"),  # noqa: E501
210
    "GraniteSpeechForConditionalGeneration": ("granite_speech", "GraniteSpeechForConditionalGeneration"),  # noqa: E501
211
    "H2OVLChatModel": ("h2ovl", "H2OVLChatModel"),
212
    "InternVLChatModel": ("internvl", "InternVLChatModel"),
Lyu Han's avatar
Lyu Han committed
213
    "InternS1ForConditionalGeneration": ("interns1", "InternS1ForConditionalGeneration"),  # noqa: E501
214
    "Idefics3ForConditionalGeneration":("idefics3","Idefics3ForConditionalGeneration"),
215
    "SmolVLMForConditionalGeneration": ("smolvlm","SmolVLMForConditionalGeneration"),  # noqa: E501
216
    "KeyeForConditionalGeneration": ("keye", "KeyeForConditionalGeneration"),
217
    "KimiVLForConditionalGeneration": ("kimi_vl", "KimiVLForConditionalGeneration"),  # noqa: E501
218
    "Llama_Nemotron_Nano_VL": ("nemotron_vl", "LlamaNemotronVLChatModel"),
219
220
221
222
    "LlavaForConditionalGeneration": ("llava", "LlavaForConditionalGeneration"),
    "LlavaNextForConditionalGeneration": ("llava_next", "LlavaNextForConditionalGeneration"),  # noqa: E501
    "LlavaNextVideoForConditionalGeneration": ("llava_next_video", "LlavaNextVideoForConditionalGeneration"),  # noqa: E501
    "LlavaOnevisionForConditionalGeneration": ("llava_onevision", "LlavaOnevisionForConditionalGeneration"),  # noqa: E501
223
    "MantisForConditionalGeneration": ("llava", "MantisForConditionalGeneration"),  # noqa: E501
224
    "MiniMaxVL01ForConditionalGeneration": ("minimax_vl_01", "MiniMaxVL01ForConditionalGeneration"),  # noqa: E501
225
    "MiniCPMO": ("minicpmo", "MiniCPMO"),
226
    "MiniCPMV": ("minicpmv", "MiniCPMV"),
227
    "Mistral3ForConditionalGeneration": ("mistral3", "Mistral3ForConditionalGeneration"),  # noqa: E501
228
    "MolmoForCausalLM": ("molmo", "MolmoForCausalLM"),
229
    "NVLM_D": ("nvlm_d", "NVLM_D_Model"),
230
    "Ovis": ("ovis", "Ovis"),
231
    "PaliGemmaForConditionalGeneration": ("paligemma", "PaliGemmaForConditionalGeneration"),  # noqa: E501
232
    "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
233
234
    "Phi4MMForCausalLM": ("phi4mm", "Phi4MMForCausalLM"),
    "Phi4MultimodalForCausalLM": ("phi4_multimodal", "Phi4MultimodalForCausalLM"),  # noqa: E501
235
    "PixtralForConditionalGeneration": ("pixtral", "PixtralForConditionalGeneration"),  # noqa: E501
236
    "QwenVLForConditionalGeneration": ("qwen_vl", "QwenVLForConditionalGeneration"),  # noqa: E501
237
    "Qwen2VLForConditionalGeneration": ("qwen2_vl", "Qwen2VLForConditionalGeneration"),  # noqa: E501
Roger Wang's avatar
Roger Wang committed
238
    "Qwen2_5_VLForConditionalGeneration": ("qwen2_5_vl", "Qwen2_5_VLForConditionalGeneration"),  # noqa: E501
239
    "Qwen2AudioForConditionalGeneration": ("qwen2_audio", "Qwen2AudioForConditionalGeneration"),  # noqa: E501
240
    "Qwen2_5OmniModel": ("qwen2_5_omni_thinker", "Qwen2_5OmniThinkerForConditionalGeneration"),  # noqa: E501
241
    "Qwen2_5OmniForConditionalGeneration": ("qwen2_5_omni_thinker", "Qwen2_5OmniThinkerForConditionalGeneration"),  # noqa: E501
242
    "UltravoxModel": ("ultravox", "UltravoxModel"),
Song's avatar
Song committed
243
    "Step3VLForConditionalGeneration": ("step3_vl", "Step3VLForConditionalGeneration"),  # noqa: E501
汪志鹏's avatar
汪志鹏 committed
244
    "TarsierForConditionalGeneration": ("tarsier", "TarsierForConditionalGeneration"),  # noqa: E501
245
    "Tarsier2ForConditionalGeneration": ("qwen2_vl", "Tarsier2ForConditionalGeneration"),  # noqa: E501
Patrick von Platen's avatar
Patrick von Platen committed
246
    "VoxtralForConditionalGeneration": ("voxtral", "VoxtralForConditionalGeneration"),  # noqa: E501
247
    # [Encoder-decoder]
248
    "Florence2ForConditionalGeneration": ("florence2", "Florence2ForConditionalGeneration"),  # noqa: E501
249
    "MllamaForConditionalGeneration": ("mllama", "MllamaForConditionalGeneration"),  # noqa: E501
250
    "Llama4ForConditionalGeneration": ("mllama4", "Llama4ForConditionalGeneration"),  # noqa: E501
251
    "SkyworkR1VChatModel": ("skyworkr1v", "SkyworkR1VChatModel"),
252
    "WhisperForConditionalGeneration": ("whisper", "WhisperForConditionalGeneration"),  # noqa: E501
253
}
254
255

_SPECULATIVE_DECODING_MODELS = {
256
    "MiMoMTPModel": ("mimo_mtp", "MiMoMTP"),
257
    "EagleLlamaForCausalLM": ("llama_eagle", "EagleLlamaForCausalLM"),
zhiweiz's avatar
zhiweiz committed
258
    "EagleLlama4ForCausalLM": ("llama4_eagle", "EagleLlama4ForCausalLM"),
259
    "EagleMiniCPMForCausalLM": ("minicpm_eagle", "EagleMiniCPMForCausalLM"),
260
    "Eagle3LlamaForCausalLM": ("llama_eagle3", "Eagle3LlamaForCausalLM"),
261
    "DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
262
    "Glm4MoeMTPModel": ("glm4_moe_mtp", "Glm4MoeMTP"),
263
    "MedusaModel": ("medusa", "Medusa"),
264
265
266
    # Temporarily disabled.
    # # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": ("mlp_speculator", "MLPSpeculator"),
267
}
268

269
270
271
272
273
_TRANSFORMERS_SUPPORTED_MODELS = {
    "Emu3ForConditionalGeneration": ("transformers", "TransformersForMultimodalLM"),  # noqa: E501
}

_TRANSFORMERS_BACKEND_MODELS = {
274
    "TransformersModel": ("transformers", "TransformersModel"),
275
    "TransformersForCausalLM": ("transformers", "TransformersForCausalLM"),
276
    "TransformersForMultimodalLM": ("transformers", "TransformersForMultimodalLM"), # noqa: E501
277
}
278
# yapf: enable
279

280
_VLLM_MODELS = {
281
    **_TEXT_GENERATION_MODELS,
282
    **_EMBEDDING_MODELS,
283
    **_CROSS_ENCODER_MODELS,
284
    **_MULTIMODAL_MODELS,
285
    **_SPECULATIVE_DECODING_MODELS,
286
287
    **_TRANSFORMERS_SUPPORTED_MODELS,
    **_TRANSFORMERS_BACKEND_MODELS,
288
289
}

290
291
292
293
294
295
296
297
# 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"
]

298
299
_PREVIOUSLY_SUPPORTED_MODELS = {"Phi3SmallForCausalLM": "0.9.2"}

300

301
302
@dataclass(frozen=True)
class _ModelInfo:
303
    architecture: str
304
    is_text_generation_model: bool
305
    is_pooling_model: bool
306
    supports_cross_encoding: bool
307
    supports_multimodal: bool
308
    supports_multimodal_raw_input: bool
309
    supports_pp: bool
310
311
    has_inner_state: bool
    is_attention_free: bool
312
    is_hybrid: bool
313
    has_noops: bool
314
    supports_transcription: bool
315
    supports_transcription_only: bool
316
    supports_v0_only: bool
317
318

    @staticmethod
319
    def from_model_cls(model: type[nn.Module]) -> "_ModelInfo":
320
        return _ModelInfo(
321
            architecture=model.__name__,
322
            is_text_generation_model=is_text_generation_model(model),
323
            is_pooling_model=is_pooling_model(model),
324
            supports_cross_encoding=supports_cross_encoding(model),
325
            supports_multimodal=supports_multimodal(model),
326
            supports_multimodal_raw_input=supports_multimodal_raw_input(model),
327
            supports_pp=supports_pp(model),
328
329
            has_inner_state=has_inner_state(model),
            is_attention_free=is_attention_free(model),
330
            is_hybrid=is_hybrid(model),
331
            supports_transcription=supports_transcription(model),
332
333
            supports_transcription_only=(supports_transcription(model) and
                                         model.supports_transcription_only),
334
            supports_v0_only=supports_v0_only(model),
335
            has_noops=has_noops(model),
336
        )
337
338


339
class _BaseRegisteredModel(ABC):
340

341
342
343
    @abstractmethod
    def inspect_model_cls(self) -> _ModelInfo:
        raise NotImplementedError
344

345
    @abstractmethod
346
    def load_model_cls(self) -> type[nn.Module]:
347
        raise NotImplementedError
348
349


350
351
352
353
354
355
356
@dataclass(frozen=True)
class _RegisteredModel(_BaseRegisteredModel):
    """
    Represents a model that has already been imported in the main process.
    """

    interfaces: _ModelInfo
357
    model_cls: type[nn.Module]
358
359

    @staticmethod
360
    def from_model_cls(model_cls: type[nn.Module]):
361
362
363
364
365
366
367
368
        return _RegisteredModel(
            interfaces=_ModelInfo.from_model_cls(model_cls),
            model_cls=model_cls,
        )

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

369
    def load_model_cls(self) -> type[nn.Module]:
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
        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()))

386
    def load_model_cls(self) -> type[nn.Module]:
387
388
389
390
391
392
393
394
        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,
395
) -> Optional[type[nn.Module]]:
396
    from vllm.platforms import current_platform
397
    current_platform.verify_model_arch(model_arch)
398
399
400
401
402
403
    try:
        return model.load_model_cls()
    except Exception:
        logger.exception("Error in loading model architecture '%s'",
                         model_arch)
        return None
404
405


406
407
408
409
410
411
412
413
414
415
416
@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
417
418


419
420
421
@dataclass
class _ModelRegistry:
    # Keyed by model_arch
422
    models: dict[str, _BaseRegisteredModel] = field(default_factory=dict)
423

424
    def get_supported_archs(self) -> Set[str]:
425
        return self.models.keys()
426

427
428
429
    def register_model(
        self,
        model_arch: str,
430
        model_cls: Union[type[nn.Module], str],
431
    ) -> None:
432
433
434
        """
        Register an external model to be used in vLLM.

435
        `model_cls` can be either:
436

437
        - A [`torch.nn.Module`][] class directly referencing the model.
438
        - A string in the format `<module>:<class>` which can be used to
439
440
          lazily import the model. This is useful to avoid initializing CUDA
          when importing the model and thus the related error
441
          `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
442
        """
443
444
445
446
        if not isinstance(model_arch, str):
            msg = f"`model_arch` should be a string, not a {type(model_arch)}"
            raise TypeError(msg)

447
        if model_arch in self.models:
448
449
450
            logger.warning(
                "Model architecture %s is already registered, and will be "
                "overwritten by the new model class %s.", model_arch,
451
452
453
454
455
456
457
                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)
458

459
            model = _LazyRegisteredModel(*split_str)
460
        elif isinstance(model_cls, type) and issubclass(model_cls, nn.Module):
461
            model = _RegisteredModel.from_model_cls(model_cls)
462
463
464
465
        else:
            msg = ("`model_cls` should be a string or PyTorch model class, "
                   f"not a {type(model_arch)}")
            raise TypeError(msg)
466

467
        self.models[model_arch] = model
468

469
    def _raise_for_unsupported(self, architectures: list[str]):
470
        all_supported_archs = self.get_supported_archs()
471

472
473
474
475
476
        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.")

477
478
479
480
481
482
483
484
485
486
        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.")

487
488
489
        raise ValueError(
            f"Model architectures {architectures} are not supported for now. "
            f"Supported architectures: {all_supported_archs}")
490

491
    def _try_load_model_cls(self,
492
                            model_arch: str) -> Optional[type[nn.Module]]:
493
494
        if model_arch not in self.models:
            return None
495

496
        return _try_load_model_cls(model_arch, self.models[model_arch])
497

498
    def _try_inspect_model_cls(self, model_arch: str) -> Optional[_ModelInfo]:
499
500
        if model_arch not in self.models:
            return None
501

502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
        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:
559
                return None
560

561
562
563
            raise ValueError(
                f"The Transformers implementation of {architecture!r} "
                "is not compatible with vLLM.")
564

565
        return model_config._get_transformers_backend_cls()
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
    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
592

593
594
    def inspect_model_cls(
        self,
595
        architectures: Union[str, list[str]],
596
        model_config: ModelConfig,
597
    ) -> tuple[_ModelInfo, str]:
598
599
        if isinstance(architectures, str):
            architectures = [architectures]
600
601
        if not architectures:
            raise ValueError("No model architectures are specified")
602
603
604
605
606
607
608
609
610
611

        # 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)

612
613
614
615
616
617
618
619
620
621
622
623
624
        # 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)
625
            model_info = self._try_inspect_model_cls(normalized_arch)
626
            if model_info is not None:
627
                return (model_info, arch)
628

629
630
631
        # 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):
632
633
634
635
636
637
638
            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)

639
        return self._raise_for_unsupported(architectures)
640

641
642
    def resolve_model_cls(
        self,
643
        architectures: Union[str, list[str]],
644
        model_config: ModelConfig,
645
    ) -> tuple[type[nn.Module], str]:
646
647
        if isinstance(architectures, str):
            architectures = [architectures]
648
649
        if not architectures:
            raise ValueError("No model architectures are specified")
650
651
652
653
654
655
656
657
658
659

        # 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)

660
661
662
663
664
665
666
667
668
669
670
671
672
        # 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)
673
            model_cls = self._try_load_model_cls(normalized_arch)
674
675
            if model_cls is not None:
                return (model_cls, arch)
676

677
678
679
        # 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):
680
681
682
683
684
685
686
            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)

687
        return self._raise_for_unsupported(architectures)
688

689
690
    def is_text_generation_model(
        self,
691
        architectures: Union[str, list[str]],
692
        model_config: ModelConfig,
693
    ) -> bool:
694
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
695
        return model_cls.is_text_generation_model
696

697
    def is_pooling_model(
698
        self,
699
        architectures: Union[str, list[str]],
700
        model_config: ModelConfig,
701
    ) -> bool:
702
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
703
        return model_cls.is_pooling_model
704

705
706
    def is_cross_encoder_model(
        self,
707
        architectures: Union[str, list[str]],
708
        model_config: ModelConfig,
709
    ) -> bool:
710
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
711
        return model_cls.supports_cross_encoding
712

713
714
    def is_multimodal_model(
        self,
715
        architectures: Union[str, list[str]],
716
        model_config: ModelConfig,
717
    ) -> bool:
718
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
719
        return model_cls.supports_multimodal
720

721
722
723
    def supports_multimodal_raw_input(
        self,
        architectures: Union[str, list[str]],
724
        model_config: ModelConfig,
725
    ) -> bool:
726
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
727
728
        return model_cls.supports_multimodal_raw_input

729
730
    def is_pp_supported_model(
        self,
731
        architectures: Union[str, list[str]],
732
        model_config: ModelConfig,
733
    ) -> bool:
734
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
735
        return model_cls.supports_pp
736

737
738
    def model_has_inner_state(
        self,
739
        architectures: Union[str, list[str]],
740
        model_config: ModelConfig,
741
    ) -> bool:
742
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
743
        return model_cls.has_inner_state
744

745
746
    def is_attention_free_model(
        self,
747
        architectures: Union[str, list[str]],
748
        model_config: ModelConfig,
749
    ) -> bool:
750
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
751
        return model_cls.is_attention_free
752

753
754
    def is_hybrid_model(
        self,
755
        architectures: Union[str, list[str]],
756
        model_config: ModelConfig,
757
    ) -> bool:
758
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
759
760
        return model_cls.is_hybrid

761
762
    def is_noops_model(
        self,
763
        architectures: Union[str, list[str]],
764
        model_config: ModelConfig,
765
    ) -> bool:
766
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
767
768
        return model_cls.has_noops

769
770
    def is_transcription_model(
        self,
771
        architectures: Union[str, list[str]],
772
        model_config: ModelConfig,
773
    ) -> bool:
774
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
775
776
        return model_cls.supports_transcription

777
778
779
    def is_transcription_only_model(
        self,
        architectures: Union[str, list[str]],
780
        model_config: ModelConfig,
781
    ) -> bool:
782
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
783
784
        return model_cls.supports_transcription_only

785
786
    def is_v1_compatible(
        self,
787
        architectures: Union[str, list[str]],
788
        model_config: ModelConfig,
789
    ) -> bool:
790
        model_cls, _ = self.inspect_model_cls(architectures, model_config)
791
792
        return not model_cls.supports_v0_only

793
794

ModelRegistry = _ModelRegistry({
795
796
    model_arch:
    _LazyRegisteredModel(
797
798
799
800
801
802
803
804
805
806
        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:
807
808
809
810
811
    # 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")

812
        # `cloudpickle` allows pickling lambda functions directly
813
        import cloudpickle
814
        input_bytes = cloudpickle.dumps((fn, output_filepath))
815
816
817

        # cannot use `sys.executable __file__` here because the script
        # contains relative imports
818
819
820
        returned = subprocess.run(_SUBPROCESS_COMMAND,
                                  input=input_bytes,
                                  capture_output=True)
821
822
823
824
825
826
827
828
829

        # 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

830
        with open(output_filepath, "rb") as f:
831
832
833
834
835
836
837
838
839
840
841
            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()
842
843
844

    with open(output_file, "wb") as f:
        f.write(pickle.dumps(result))
845
846
847


if __name__ == "__main__":
848
    _run()