registry.py 40.1 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

4
from collections.abc import Mapping, Set
5
from dataclasses import dataclass, field
6
from typing import Any, Literal, Optional
7
8
9
10

import pytest
from packaging.version import Version
from transformers import __version__ as TRANSFORMERS_VERSION
11

12
13
from vllm.config import TokenizerMode

14
15
16
17
18
19
20
21
22
23
24
25

@dataclass(frozen=True)
class _HfExamplesInfo:
    default: str
    """The default model to use for testing this architecture."""

    extras: Mapping[str, str] = field(default_factory=dict)
    """Extra models to use for testing this architecture."""

    tokenizer: Optional[str] = None
    """Set the tokenizer to load for this architecture."""

26
    tokenizer_mode: TokenizerMode = "auto"
27
28
29
30
31
32
33
34
    """Set the tokenizer type for this architecture."""

    speculative_model: Optional[str] = None
    """
    The default model to use for testing this architecture, which is only used
    for speculative decoding.
    """

35
36
37
38
39
    min_transformers_version: Optional[str] = None
    """
    The minimum version of HF Transformers that is required to run this model.
    """

40
41
42
43
44
45
46
47
48
49
    max_transformers_version: Optional[str] = None
    """
    The maximum version of HF Transformers that this model runs on.
    """

    transformers_version_reason: Optional[str] = None
    """
    The reason for the minimum/maximum version requirement.
    """

50
51
52
53
54
55
56
57
58
59
60
    is_available_online: bool = True
    """
    Set this to ``False`` if the name of this architecture no longer exists on
    the HF repo. To maintain backwards compatibility, we have not removed them
    from the main model registry, so without this flag the registry tests will
    fail.
    """

    trust_remote_code: bool = False
    """The ``trust_remote_code`` level required to load the model."""

61
62
63
    v0_only: bool = False
    """The model is only available with the vLLM V0 engine."""

64
65
66
    hf_overrides: dict[str, Any] = field(default_factory=dict)
    """The ``hf_overrides`` required to load the model."""

67
68
69
70
71
72
    max_model_len: Optional[int] = None
    """
    The maximum model length to use for this model. Some models default to a
    length that is too large to fit into memory in CI.
    """

73
74
75
76
77
78
    revision: Optional[str] = None
    """
    The specific revision (commit hash, tag, or branch) to use for the model.
    If not specified, the default revision will be used.
    """

79
80
81
    def check_transformers_version(
        self,
        *,
82
        on_fail: Literal["error", "skip", "return"],
83
84
        check_min_version: bool = True,
        check_max_version: bool = True,
85
    ) -> Optional[str]:
86
87
88
89
        """
        If the installed transformers version does not meet the requirements,
        perform the given action.
        """
90
91
        if (self.min_transformers_version is None
                and self.max_transformers_version is None):
92
            return None
93
94

        current_version = TRANSFORMERS_VERSION
95
        cur_base_version = Version(current_version).base_version
96
97
98
        min_version = self.min_transformers_version
        max_version = self.max_transformers_version
        msg = f"`transformers=={current_version}` installed, but `transformers"
99
100
        # Only check the base version for the min/max version, otherwise preview
        # models cannot be run because `x.yy.0.dev0`<`x.yy.0`
101
102
        if (check_min_version and min_version
                and Version(cur_base_version) < Version(min_version)):
103
            msg += f">={min_version}` is required to run this model."
104
105
        elif (check_max_version and max_version
              and Version(cur_base_version) > Version(max_version)):
106
107
            msg += f"<={max_version}` is required to run this model."
        else:
108
            return None
109

110
111
112
113
114
        if self.transformers_version_reason:
            msg += f" Reason: {self.transformers_version_reason}"

        if on_fail == "error":
            raise RuntimeError(msg)
115
        elif on_fail == "skip":
116
            pytest.skip(msg)
117

118
119
        return msg

120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
    def check_available_online(
        self,
        *,
        on_fail: Literal["error", "skip"],
    ) -> None:
        """
        If the model is not available online, perform the given action.
        """
        if not self.is_available_online:
            msg = "Model is not available online"

            if on_fail == "error":
                raise RuntimeError(msg)
            else:
                pytest.skip(msg)

136
137
138
139
140
141
142
143

# yapf: disable
_TEXT_GENERATION_EXAMPLE_MODELS = {
    # [Decoder-only]
    "AquilaModel": _HfExamplesInfo("BAAI/AquilaChat-7B",
                                   trust_remote_code=True),
    "AquilaForCausalLM": _HfExamplesInfo("BAAI/AquilaChat2-7B",
                                         trust_remote_code=True),
144
    "ArceeForCausalLM": _HfExamplesInfo("arcee-ai/AFM-4.5B-Base"),
145
146
147
148
149
150
    "ArcticForCausalLM": _HfExamplesInfo("Snowflake/snowflake-arctic-instruct",
                                         trust_remote_code=True),
    "BaiChuanForCausalLM": _HfExamplesInfo("baichuan-inc/Baichuan-7B",
                                         trust_remote_code=True),
    "BaichuanForCausalLM": _HfExamplesInfo("baichuan-inc/Baichuan2-7B-chat",
                                         trust_remote_code=True),
151
152
    "BailingMoeForCausalLM": _HfExamplesInfo("inclusionAI/Ling-lite-1.5",
                                         trust_remote_code=True),
153
    "BambaForCausalLM": _HfExamplesInfo("ibm-ai-platform/Bamba-9B-v1",
154
                                        min_transformers_version="4.56.0",
155
                                        extras={"tiny": "hmellor/tiny-random-BambaForCausalLM"}),  # noqa: E501
156
157
    "BloomForCausalLM": _HfExamplesInfo("bigscience/bloom-560m",
                                        {"1b": "bigscience/bloomz-1b1"}),
158
    "ChatGLMModel": _HfExamplesInfo("zai-org/chatglm3-6b",
159
                                    trust_remote_code=True,
160
                                    max_transformers_version="4.48"),
161
162
    "ChatGLMForConditionalGeneration": _HfExamplesInfo("thu-coai/ShieldLM-6B-chatglm3",  # noqa: E501
                                                       trust_remote_code=True),
163
164
    "CohereForCausalLM": _HfExamplesInfo("CohereForAI/c4ai-command-r-v01",
                                         trust_remote_code=True),
165
166
    "Cohere2ForCausalLM": _HfExamplesInfo("CohereForAI/c4ai-command-r7b-12-2024", # noqa: E501
                                         trust_remote_code=True),
167
    "DbrxForCausalLM": _HfExamplesInfo("databricks/dbrx-instruct"),
168
    "DeciLMForCausalLM": _HfExamplesInfo("nvidia/Llama-3_3-Nemotron-Super-49B-v1", # noqa: E501
169
170
171
172
                                         trust_remote_code=True),
    "DeepseekForCausalLM": _HfExamplesInfo("deepseek-ai/deepseek-llm-7b-chat"),
    "DeepseekV2ForCausalLM": _HfExamplesInfo("deepseek-ai/DeepSeek-V2-Lite-Chat",  # noqa: E501
                                         trust_remote_code=True),
Robert Shaw's avatar
Robert Shaw committed
173
174
    "DeepseekV3ForCausalLM": _HfExamplesInfo("deepseek-ai/DeepSeek-V3",  # noqa: E501
                                         trust_remote_code=True),
175
    "Ernie4_5ForCausalLM": _HfExamplesInfo("baidu/ERNIE-4.5-0.3B-PT",
176
                                            min_transformers_version="4.54"),
177
    "Ernie4_5_MoeForCausalLM": _HfExamplesInfo("baidu/ERNIE-4.5-21B-A3B-PT",
178
                                               min_transformers_version="4.54"),
179
180
181
182
    "ExaoneForCausalLM": _HfExamplesInfo("LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct",
                                         trust_remote_code=True),
    "Exaone4ForCausalLM": _HfExamplesInfo("LGAI-EXAONE/EXAONE-4.0-32B",
                                          min_transformers_version="4.54"),
183
    "Fairseq2LlamaForCausalLM": _HfExamplesInfo("mgleize/fairseq2-dummy-Llama-3.2-1B"),  # noqa: E501
184
    "FalconForCausalLM": _HfExamplesInfo("tiiuae/falcon-7b"),
185
    "FalconH1ForCausalLM":_HfExamplesInfo("tiiuae/Falcon-H1-0.5B-Base"),
186
    "GemmaForCausalLM": _HfExamplesInfo("google/gemma-1.1-2b-it"),
187
    "Gemma2ForCausalLM": _HfExamplesInfo("google/gemma-2-9b"),
188
    "Gemma3ForCausalLM": _HfExamplesInfo("google/gemma-3-1b-it"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
189
    "Gemma3nForCausalLM": _HfExamplesInfo("google/gemma-3n-E2B-it",
Robert Shaw's avatar
Robert Shaw committed
190
                                          min_transformers_version="4.53"),
191
192
    "GlmForCausalLM": _HfExamplesInfo("zai-org/glm-4-9b-chat-hf"),
    "Glm4ForCausalLM": _HfExamplesInfo("zai-org/GLM-4-9B-0414"),
193
194
    "Glm4MoeForCausalLM": _HfExamplesInfo("zai-org/GLM-4.5",
                                          min_transformers_version="4.54"),   # noqa: E501
195
196
197
    "GPT2LMHeadModel": _HfExamplesInfo("openai-community/gpt2",
                                       {"alias": "gpt2"}),
    "GPTBigCodeForCausalLM": _HfExamplesInfo("bigcode/starcoder",
198
                                             extras={"tiny": "bigcode/tiny_starcoder_py"},  # noqa: E501
199
200
                                             min_transformers_version="4.55.1",
                                             transformers_version_reason="HF model broken in 4.55.0"),  # noqa: E501
201
202
203
204
    "GPTJForCausalLM": _HfExamplesInfo("Milos/slovak-gpt-j-405M",
                                       {"6b": "EleutherAI/gpt-j-6b"}),
    "GPTNeoXForCausalLM": _HfExamplesInfo("EleutherAI/pythia-70m",
                                          {"1b": "EleutherAI/pythia-1.4b"}),
205
    "GptOssForCausalLM": _HfExamplesInfo("lmsys/gpt-oss-20b-bf16"),
206
207
    "GraniteForCausalLM": _HfExamplesInfo("ibm/PowerLM-3b"),
    "GraniteMoeForCausalLM": _HfExamplesInfo("ibm/PowerMoE-3b"),
208
    "GraniteMoeHybridForCausalLM": _HfExamplesInfo("ibm-granite/granite-4.0-tiny-preview"),  # noqa: E501
209
    "GraniteMoeSharedForCausalLM": _HfExamplesInfo("ibm-research/moe-7b-1b-active-shared-experts"),  # noqa: E501
Michael Goin's avatar
Michael Goin committed
210
211
    "Grok1ModelForCausalLM": _HfExamplesInfo("hpcai-tech/grok-1",
                                             trust_remote_code=True),
212
213
    "HunYuanMoEV1ForCausalLM": _HfExamplesInfo("tencent/Hunyuan-A13B-Instruct",
                                               trust_remote_code=True),
214
    # TODO: Remove is_available_online once their config.json is fixed
215
    "HunYuanDenseV1ForCausalLM":_HfExamplesInfo("tencent/Hunyuan-7B-Instruct-0124",
216
217
                                                trust_remote_code=True,
                                                is_available_online=False),
218
219
220
221
222
223
    "InternLMForCausalLM": _HfExamplesInfo("internlm/internlm-chat-7b",
                                           trust_remote_code=True),
    "InternLM2ForCausalLM": _HfExamplesInfo("internlm/internlm2-chat-7b",
                                            trust_remote_code=True),
    "InternLM2VEForCausalLM": _HfExamplesInfo("OpenGVLab/Mono-InternVL-2B",
                                              trust_remote_code=True),
224
225
    "InternLM3ForCausalLM": _HfExamplesInfo("internlm/internlm3-8b-instruct",
                                            trust_remote_code=True),
226
    "JAISLMHeadModel": _HfExamplesInfo("inceptionai/jais-13b-chat"),
227
    "JambaForCausalLM": _HfExamplesInfo("ai21labs/AI21-Jamba-1.5-Mini",
228
                                        min_transformers_version="4.56.0",
229
230
231
232
                                        extras={
                                            "tiny": "ai21labs/Jamba-tiny-dev",
                                            "random": "ai21labs/Jamba-tiny-random",  # noqa: E501
                                        }),
233
234
    "Lfm2ForCausalLM": _HfExamplesInfo("LiquidAI/LFM2-1.2B",
                                       min_transformers_version="4.54"),
235
236
    "LlamaForCausalLM": _HfExamplesInfo("meta-llama/Llama-3.2-1B-Instruct",
                                        extras={"guard": "meta-llama/Llama-Guard-3-1B",  # noqa: E501
237
238
                                                "hermes": "NousResearch/Hermes-3-Llama-3.1-8B", # noqa: E501
                                                "fp8": "RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8"}),  # noqa: E501
239
240
    "LLaMAForCausalLM": _HfExamplesInfo("decapoda-research/llama-7b-hf",
                                        is_available_online=False),
241
242
    "Llama4ForCausalLM": _HfExamplesInfo("meta-llama/Llama-4-Scout-17B-16E-Instruct", # noqa: E501
                                         is_available_online=False),
243
    "MambaForCausalLM": _HfExamplesInfo("state-spaces/mamba-130m-hf"),
244
    "Mamba2ForCausalLM": _HfExamplesInfo("mistralai/Mamba-Codestral-7B-v0.1"),
245
246
247
248
249
    "FalconMambaForCausalLM": _HfExamplesInfo("tiiuae/falcon-mamba-7b-instruct"),  # noqa: E501
    "MiniCPMForCausalLM": _HfExamplesInfo("openbmb/MiniCPM-2B-sft-bf16",
                                         trust_remote_code=True),
    "MiniCPM3ForCausalLM": _HfExamplesInfo("openbmb/MiniCPM3-4B",
                                         trust_remote_code=True),
250
    "MiniMaxForCausalLM": _HfExamplesInfo("MiniMaxAI/MiniMax-Text-01-hf"),
251
    "MiniMaxText01ForCausalLM": _HfExamplesInfo("MiniMaxAI/MiniMax-Text-01",
252
253
                                                trust_remote_code=True,
                                                revision="a59aa9cbc53b9fb8742ca4e9e1531b9802b6fdc3"),  # noqa: E501
254
255
    "MiniMaxM1ForCausalLM": _HfExamplesInfo("MiniMaxAI/MiniMax-M1-40k",
                                            trust_remote_code=True),
256
    "MistralForCausalLM": _HfExamplesInfo("mistralai/Mistral-7B-Instruct-v0.1"),
257
    "MixtralForCausalLM": _HfExamplesInfo("mistralai/Mixtral-8x7B-Instruct-v0.1",  # noqa: E501
258
                                          {"tiny": "TitanML/tiny-mixtral"}),  # noqa: E501
259
260
261
262
    "QuantMixtralForCausalLM": _HfExamplesInfo("mistral-community/Mixtral-8x22B-v0.1-AWQ"),  # noqa: E501
    "MptForCausalLM": _HfExamplesInfo("mpt", is_available_online=False),
    "MPTForCausalLM": _HfExamplesInfo("mosaicml/mpt-7b"),
    "NemotronForCausalLM": _HfExamplesInfo("nvidia/Minitron-8B-Base"),
Luis Vega's avatar
Luis Vega committed
263
264
    "NemotronHForCausalLM": _HfExamplesInfo("nvidia/Nemotron-H-8B-Base-8K",
                                            trust_remote_code=True),
265
    "OlmoForCausalLM": _HfExamplesInfo("allenai/OLMo-1B-hf"),
266
    "Olmo2ForCausalLM": _HfExamplesInfo("allenai/OLMo-2-0425-1B"),
267
    "OlmoeForCausalLM": _HfExamplesInfo("allenai/OLMoE-1B-7B-0924-Instruct"),
268
269
    "OPTForCausalLM": _HfExamplesInfo("facebook/opt-125m",
                                      {"1b": "facebook/opt-iml-max-1.3b"}),
270
271
272
    "OrionForCausalLM": _HfExamplesInfo("OrionStarAI/Orion-14B-Chat",
                                        trust_remote_code=True),
    "PersimmonForCausalLM": _HfExamplesInfo("adept/persimmon-8b-chat"),
273
    "PhiForCausalLM": _HfExamplesInfo("microsoft/phi-2"),
274
    "Phi3ForCausalLM": _HfExamplesInfo("microsoft/Phi-3-mini-4k-instruct"),
275
276
277
278
    "Phi4FlashForCausalLM": _HfExamplesInfo("microsoft/Phi-4-mini-flash-reasoning", # noqa: E501
                                        trust_remote_code=True,
                                        v0_only=True,
                                        max_model_len=10240),
279
280
    "PhiMoEForCausalLM": _HfExamplesInfo("microsoft/Phi-3.5-MoE-instruct",
                                         trust_remote_code=True),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
281
    "Plamo2ForCausalLM": _HfExamplesInfo("pfnet/plamo-2-1b",
282
283
                                         max_transformers_version="4.53",
                                         transformers_version_reason="vLLM impl inherits PreTrainedModel and clashes with get_input_embeddings",  # noqa: E501
Shinichi Hemmi's avatar
Shinichi Hemmi committed
284
                                        trust_remote_code=True),
285
    "QWenLMHeadModel": _HfExamplesInfo("Qwen/Qwen-7B-Chat",
286
287
                                       max_transformers_version="4.53",
                                       transformers_version_reason="HF model uses remote code that is not compatible with latest Transformers",  # noqa: E501
288
                                       trust_remote_code=True),
289
290
    "Qwen2ForCausalLM": _HfExamplesInfo("Qwen/Qwen2-0.5B-Instruct",
                                        extras={"2.5": "Qwen/Qwen2.5-0.5B-Instruct"}), # noqa: E501
291
    "Qwen2MoeForCausalLM": _HfExamplesInfo("Qwen/Qwen1.5-MoE-A2.7B-Chat"),
Jee Jee Li's avatar
Jee Jee Li committed
292
293
    "Qwen3ForCausalLM": _HfExamplesInfo("Qwen/Qwen3-8B"),
    "Qwen3MoeForCausalLM": _HfExamplesInfo("Qwen/Qwen3-30B-A3B"),
294
    "RWForCausalLM": _HfExamplesInfo("tiiuae/falcon-40b"),
295
296
297
    "SeedOssForCausalLM": _HfExamplesInfo("ByteDance-Seed/Seed-OSS-36B-Instruct", # noqa: E501
                                          trust_remote_code=True,
                                          is_available_online=False),
298
    "SmolLM3ForCausalLM": _HfExamplesInfo("HuggingFaceTB/SmolLM3-3B"),
299
300
    "StableLMEpochForCausalLM": _HfExamplesInfo("stabilityai/stablelm-zephyr-3b"),  # noqa: E501
    "StableLmForCausalLM": _HfExamplesInfo("stabilityai/stablelm-3b-4e1t"),
301
    "Starcoder2ForCausalLM": _HfExamplesInfo("bigcode/starcoder2-3b"),
Song's avatar
Song committed
302
    "Step3TextForCausalLM": _HfExamplesInfo("stepfun-ai/step3",
303
                                            trust_remote_code=True),
304
305
    "SolarForCausalLM": _HfExamplesInfo("upstage/solar-pro-preview-instruct",
                                        trust_remote_code=True),
306
307
    "TeleChat2ForCausalLM": _HfExamplesInfo("Tele-AI/TeleChat2-3B",
                                            trust_remote_code=True),
308
309
    "TeleFLMForCausalLM": _HfExamplesInfo("CofeAI/FLM-2-52B-Instruct-2407",
                                            trust_remote_code=True),
310
    "XverseForCausalLM": _HfExamplesInfo("xverse/XVERSE-7B-Chat",
311
                                         tokenizer="meta-llama/Llama-2-7b",
312
                                         trust_remote_code=True),
313
    "Zamba2ForCausalLM": _HfExamplesInfo("Zyphra/Zamba2-7B-instruct"),
314
315
    "MiMoForCausalLM": _HfExamplesInfo("XiaomiMiMo/MiMo-7B-RL",
                                        trust_remote_code=True),
316
    "Dots1ForCausalLM": _HfExamplesInfo("rednote-hilab/dots.llm1.inst"),
317
318
319
    # [Encoder-decoder]
    "BartModel": _HfExamplesInfo("facebook/bart-base"),
    "BartForConditionalGeneration": _HfExamplesInfo("facebook/bart-large-cnn"),
汪志鹏's avatar
汪志鹏 committed
320
321
    "MBartForConditionalGeneration": _HfExamplesInfo("facebook/mbart-large-en-ro",  # noqa: E501
                                                    hf_overrides={"architectures": ["MBartForConditionalGeneration"]}),  # noqa: E501
322
323
324
325
}

_EMBEDDING_EXAMPLE_MODELS = {
    # [Text-only]
326
327
    "BertModel": _HfExamplesInfo("BAAI/bge-base-en-v1.5", v0_only=True),
    "Gemma2Model": _HfExamplesInfo("BAAI/bge-multilingual-gemma2", v0_only=True),  # noqa: E501
328
    "GritLM": _HfExamplesInfo("parasail-ai/GritLM-7B-vllm"),
329
330
    "GteModel": _HfExamplesInfo("Snowflake/snowflake-arctic-embed-m-v2.0",
                                               trust_remote_code=True),
331
332
    "GteNewModel": _HfExamplesInfo("Alibaba-NLP/gte-base-en-v1.5",
                                   trust_remote_code=True,
333
                                   hf_overrides={"architectures": ["GteNewModel"]}),  # noqa: E501
334
335
    "InternLM2ForRewardModel": _HfExamplesInfo("internlm/internlm2-1_8b-reward",
                                               trust_remote_code=True),
336
    "JambaForSequenceClassification": _HfExamplesInfo("ai21labs/Jamba-tiny-reward-dev"),  # noqa: E501
337
    "LlamaModel": _HfExamplesInfo("llama", is_available_online=False),
338
    "MistralModel": _HfExamplesInfo("intfloat/e5-mistral-7b-instruct"),
339
    "ModernBertModel": _HfExamplesInfo("Alibaba-NLP/gte-modernbert-base",
340
                                trust_remote_code=True, v0_only=True),
341
    "NomicBertModel": _HfExamplesInfo("nomic-ai/nomic-embed-text-v2-moe",
342
                                               trust_remote_code=True, v0_only=True),  # noqa: E501
343
    "Qwen2Model": _HfExamplesInfo("ssmits/Qwen2-7B-Instruct-embed-base"),
344
345
346
347
348
349
    "Qwen2ForRewardModel": _HfExamplesInfo("Qwen/Qwen2.5-Math-RM-72B",
                                           max_transformers_version="4.53",
                                           transformers_version_reason="HF model uses remote code that is not compatible with latest Transformers"),  # noqa: E501
    "Qwen2ForProcessRewardModel": _HfExamplesInfo("Qwen/Qwen2.5-Math-PRM-7B",
                                                  max_transformers_version="4.53",
                                                  transformers_version_reason="HF model uses remote code that is not compatible with latest Transformers"),  # noqa: E501
350
351
352
    "RobertaModel": _HfExamplesInfo("sentence-transformers/stsb-roberta-base-v2", v0_only=True),  # noqa: E501
    "RobertaForMaskedLM": _HfExamplesInfo("sentence-transformers/all-roberta-large-v1", v0_only=True),  # noqa: E501
    "XLMRobertaModel": _HfExamplesInfo("intfloat/multilingual-e5-small", v0_only=True),  # noqa: E501
353
354
355
356
357
    # [Multimodal]
    "LlavaNextForConditionalGeneration": _HfExamplesInfo("royokong/e5-v"),
    "Phi3VForCausalLM": _HfExamplesInfo("TIGER-Lab/VLM2Vec-Full",
                                         trust_remote_code=True),
    "Qwen2VLForConditionalGeneration": _HfExamplesInfo("MrLight/dse-qwen2-2b-mrl-v1"), # noqa: E501
358
359
    "PrithviGeoSpatialMAE": _HfExamplesInfo("ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11", # noqa: E501
                                            is_available_online=False),  # noqa: E501
360
361
}

362
363
364
365
366
_SEQUENCE_CLASSIFICATION_EXAMPLE_MODELS = {
    # [Decoder-only]
    "GPT2ForSequenceClassification": _HfExamplesInfo("nie3e/sentiment-polish-gpt2-small"),  # noqa: E501

    # [Cross-encoder]
367
    "BertForSequenceClassification": _HfExamplesInfo("cross-encoder/ms-marco-MiniLM-L-6-v2", v0_only=True),  # noqa: E501
368
    "ModernBertForSequenceClassification": _HfExamplesInfo("Alibaba-NLP/gte-reranker-modernbert-base", v0_only=True), # noqa: E501
369
370
    "RobertaForSequenceClassification": _HfExamplesInfo("cross-encoder/quora-roberta-base", v0_only=True),  # noqa: E501
    "XLMRobertaForSequenceClassification": _HfExamplesInfo("BAAI/bge-reranker-v2-m3", v0_only=True),  # noqa: E501
371
372
}

373
374
375
376
377
378
379
380
381
382
383
384
_AUTOMATIC_CONVERTED_MODELS = {
    # Use as_seq_cls_model for automatic conversion
    "GemmaForSequenceClassification": _HfExamplesInfo("BAAI/bge-reranker-v2-gemma",  # noqa: E501
                                                      v0_only=True,
                                                      hf_overrides={"architectures": ["GemmaForSequenceClassification"], # noqa: E501
                                                                    "classifier_from_token": ["Yes"],  # noqa: E501
                                                                    "method": "no_post_processing"}),  # noqa: E501
    "LlamaForSequenceClassification": _HfExamplesInfo("Skywork/Skywork-Reward-V2-Llama-3.2-1B"),  # noqa: E501
    "Qwen2ForSequenceClassification": _HfExamplesInfo("jason9693/Qwen2.5-1.5B-apeach"),  # noqa: E501
    "Qwen3ForSequenceClassification": _HfExamplesInfo("tomaarsen/Qwen3-Reranker-0.6B-seq-cls"),  # noqa: E501
}

385
386
_MULTIMODAL_EXAMPLE_MODELS = {
    # [Decoder-only]
387
    "AriaForConditionalGeneration": _HfExamplesInfo("rhymes-ai/Aria"),
Jennifer Zhao's avatar
Jennifer Zhao committed
388
    "AyaVisionForConditionalGeneration": _HfExamplesInfo("CohereForAI/aya-vision-8b"), # noqa: E501
389
    "Blip2ForConditionalGeneration": _HfExamplesInfo("Salesforce/blip2-opt-2.7b",  # noqa: E501
390
                                                     extras={"6b": "Salesforce/blip2-opt-6.7b"}),  # noqa: E501
391
    "ChameleonForConditionalGeneration": _HfExamplesInfo("facebook/chameleon-7b"),  # noqa: E501
392
    "Cohere2VisionForConditionalGeneration": _HfExamplesInfo("CohereLabs/command-a-vision-07-2025"), # noqa: E501
393
    "DeepseekVLV2ForCausalLM": _HfExamplesInfo("deepseek-ai/deepseek-vl2-tiny",  # noqa: E501
394
395
396
                                                extras={"fork": "Isotr0py/deepseek-vl2-tiny"},  # noqa: E501
                                                max_transformers_version="4.48",  # noqa: E501
                                                transformers_version_reason="HF model is not compatible.",  # noqa: E501
397
                                                hf_overrides={"architectures": ["DeepseekVLV2ForCausalLM"]}),  # noqa: E501
398
    "Emu3ForConditionalGeneration": _HfExamplesInfo("BAAI/Emu3-Chat-hf"),
399
400
    "Ernie4_5_VLMoeForConditionalGeneration": _HfExamplesInfo("baidu/ERNIE-4.5-VL-28B-A3B-PT",  # noqa: E501
                                                              trust_remote_code=True),
401
    "FuyuForCausalLM": _HfExamplesInfo("adept/fuyu-8b"),
402
    "Gemma3ForConditionalGeneration": _HfExamplesInfo("google/gemma-3-4b-it"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
403
404
    "Gemma3nForConditionalGeneration": _HfExamplesInfo("google/gemma-3n-E2B-it",    # noqa: E501
                                        min_transformers_version="4.53"),
405
    "GraniteSpeechForConditionalGeneration": _HfExamplesInfo("ibm-granite/granite-speech-3.3-2b"),  # noqa: E501
406
    "GLM4VForCausalLM": _HfExamplesInfo("zai-org/glm-4v-9b",
407
408
                                        trust_remote_code=True,
                                        hf_overrides={"architectures": ["GLM4VForCausalLM"]}),  # noqa: E501
409
    "Glm4vForConditionalGeneration": _HfExamplesInfo("zai-org/GLM-4.1V-9B-Thinking"),  # noqa: E501
410
    "Glm4vMoeForConditionalGeneration": _HfExamplesInfo("zai-org/GLM-4.5V",
411
                                                        min_transformers_version="4.56"),  # noqa: E501
412
    "H2OVLChatModel": _HfExamplesInfo("h2oai/h2ovl-mississippi-800m",
413
                                      trust_remote_code=True,
414
415
416
                                      extras={"2b": "h2oai/h2ovl-mississippi-2b"},  # noqa: E501
                                      max_transformers_version="4.48",  # noqa: E501
                                      transformers_version_reason="HF model is not compatible."),  # noqa: E501
417
418
    "HCXVisionForCausalLM": _HfExamplesInfo("naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B",  # noqa: E501
                                            trust_remote_code=True),
419
420
    "Idefics3ForConditionalGeneration": _HfExamplesInfo("HuggingFaceM4/Idefics3-8B-Llama3",  # noqa: E501
                                                        {"tiny": "HuggingFaceTB/SmolVLM-256M-Instruct"},    # noqa: E501
421
422
                                                        min_transformers_version="4.56",
                                                        transformers_version_reason="HF model broken in 4.55"),  # noqa: E501
423
424
    "InternS1ForConditionalGeneration": _HfExamplesInfo("internlm/Intern-S1",
                                                        trust_remote_code=True),  # noqa: E501
425
    "InternVLChatModel": _HfExamplesInfo("OpenGVLab/InternVL2-1B",
426
                                         extras={"2B": "OpenGVLab/InternVL2-2B",
427
428
429
430
                                                 "3.0": "OpenGVLab/InternVL3-1B",   # noqa: E501
                                                 "3.5-qwen3": "OpenGVLab/InternVL3_5-1B",   # noqa: E501
                                                 "3.5-qwen3moe": "OpenGVLab/InternVL3_5-30B-A3B",   # noqa: E501
                                                 "3.5-gptoss": "OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview"},  # noqa: E501
431
                                         trust_remote_code=True),
432
    "InternVLForConditionalGeneration": _HfExamplesInfo("OpenGVLab/InternVL3-1B-hf"),    # noqa: E501
433
434
    "KeyeForConditionalGeneration": _HfExamplesInfo("Kwai-Keye/Keye-VL-8B-Preview", # noqa: E501
                                                    trust_remote_code=True),
435
436
    "KimiVLForConditionalGeneration": _HfExamplesInfo("moonshotai/Kimi-VL-A3B-Instruct",  # noqa: E501
                                                      extras={"thinking": "moonshotai/Kimi-VL-A3B-Thinking"},  # noqa: E501
437
                                                      trust_remote_code=True),
438
    "Llama4ForConditionalGeneration": _HfExamplesInfo("meta-llama/Llama-4-Scout-17B-16E-Instruct",   # noqa: E501
439
440
441
                                                      max_model_len=10240,
                                                      extras={"llama-guard-4": "meta-llama/Llama-Guard-4-12B"}, # noqa: E501
                                                      ),
442
    "LlavaForConditionalGeneration": _HfExamplesInfo("llava-hf/llava-1.5-7b-hf",
443
444
                                                     extras={"mistral": "mistral-community/pixtral-12b", # noqa: E501
                                                             "mistral-fp8": "nm-testing/pixtral-12b-FP8-dynamic"}),  # noqa: E501
445
446
447
    "LlavaNextForConditionalGeneration": _HfExamplesInfo("llava-hf/llava-v1.6-mistral-7b-hf"),  # noqa: E501
    "LlavaNextVideoForConditionalGeneration": _HfExamplesInfo("llava-hf/LLaVA-NeXT-Video-7B-hf"),  # noqa: E501
    "LlavaOnevisionForConditionalGeneration": _HfExamplesInfo("llava-hf/llava-onevision-qwen2-0.5b-ov-hf"),  # noqa: E501
448
    "MantisForConditionalGeneration": _HfExamplesInfo("TIGER-Lab/Mantis-8B-siglip-llama3",  # noqa: E501
449
450
                                                      max_transformers_version="4.48",  # noqa: E501
                                                      transformers_version_reason="HF model is not compatible.",  # noqa: E501
451
                                                      hf_overrides={"architectures": ["MantisForConditionalGeneration"]}),  # noqa: E501
452
453
    "MiniCPMO": _HfExamplesInfo("openbmb/MiniCPM-o-2_6",
                                trust_remote_code=True),
454
    "MiniCPMV": _HfExamplesInfo("openbmb/MiniCPM-Llama3-V-2_5",
tc-mb's avatar
tc-mb committed
455
                                extras={"2.6": "openbmb/MiniCPM-V-2_6", "4.0": "openbmb/MiniCPM-V-4", "4.5": "openbmb/MiniCPM-V-4_5"},  # noqa: E501
456
                                trust_remote_code=True),
457
    "MiniMaxVL01ForConditionalGeneration": _HfExamplesInfo("MiniMaxAI/MiniMax-VL-01", # noqa: E501
458
459
                                              trust_remote_code=True,
                                              v0_only=True),
460
461
    "Mistral3ForConditionalGeneration": _HfExamplesInfo("mistralai/Mistral-Small-3.1-24B-Instruct-2503",  # noqa: E501
                                                        extras={"fp8": "nm-testing/Mistral-Small-3.1-24B-Instruct-2503-FP8-dynamic"}),  # noqa: E501
462
    "MolmoForCausalLM": _HfExamplesInfo("allenai/Molmo-7B-D-0924",
463
                                        max_transformers_version="4.48",
464
                                        transformers_version_reason="Incorrectly-detected `tensorflow` import.",  # noqa: E501
465
                                        extras={"olmo": "allenai/Molmo-7B-O-0924"},  # noqa: E501
466
467
468
                                        trust_remote_code=True),
    "NVLM_D": _HfExamplesInfo("nvidia/NVLM-D-72B",
                              trust_remote_code=True),
469
470
    "Llama_Nemotron_Nano_VL" : _HfExamplesInfo("nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1", # noqa: E501
                                                     trust_remote_code=True),
471
    "Ovis": _HfExamplesInfo("AIDC-AI/Ovis2-1B", trust_remote_code=True,
472
473
                            max_transformers_version="4.53",
                            transformers_version_reason="HF model is not compatible",  # noqa: E501
474
475
                            extras={"1.6-llama": "AIDC-AI/Ovis1.6-Llama3.2-3B",
                                    "1.6-gemma": "AIDC-AI/Ovis1.6-Gemma2-9B"}),  # noqa: E501
476
    "Ovis2_5": _HfExamplesInfo("AIDC-AI/Ovis2.5-2B",
477
                               trust_remote_code=True),
478
479
    "PaliGemmaForConditionalGeneration": _HfExamplesInfo("google/paligemma-3b-mix-224",  # noqa: E501
                                                         extras={"v2": "google/paligemma2-3b-ft-docci-448"}),  # noqa: E501
480
481
    "Phi3VForCausalLM": _HfExamplesInfo("microsoft/Phi-3-vision-128k-instruct",
                                        trust_remote_code=True,
482
483
                                        max_transformers_version="4.48",
                                        transformers_version_reason="Use of deprecated imports which have been removed.",  # noqa: E501
484
                                        extras={"phi3.5": "microsoft/Phi-3.5-vision-instruct"}),  # noqa: E501
485
486
    "Phi4MMForCausalLM": _HfExamplesInfo("microsoft/Phi-4-multimodal-instruct",
                                        trust_remote_code=True),
487
488
    "Phi4MultimodalForCausalLM": _HfExamplesInfo("microsoft/Phi-4-multimodal-instruct",  # noqa: E501
                                                 revision="refs/pr/70"),
489
    "PixtralForConditionalGeneration": _HfExamplesInfo("mistralai/Pixtral-12B-2409",  # noqa: E501
490
                                                       tokenizer_mode="mistral"),
491
492
493
494
    "QwenVLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen-VL",
                                                      extras={"chat": "Qwen/Qwen-VL-Chat"},  # noqa: E501
                                                      trust_remote_code=True,
                                                      hf_overrides={"architectures": ["QwenVLForConditionalGeneration"]}),  # noqa: E501
495
496
    "Qwen2AudioForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2-Audio-7B-Instruct"),  # noqa: E501
    "Qwen2VLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2-VL-2B-Instruct"),  # noqa: E501
497
498
    "Qwen2_5_VLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2.5-VL-3B-Instruct", # noqa: E501
                                                          max_model_len=4096),
499
500
    "Qwen2_5OmniModel": _HfExamplesInfo("Qwen/Qwen2.5-Omni-3B"),
    "Qwen2_5OmniForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2.5-Omni-7B-AWQ"),  # noqa: E501
501
502
    "RForConditionalGeneration": _HfExamplesInfo("YannQi/R-4B",
                                                 trust_remote_code=True),
503
504
    "SkyworkR1VChatModel": _HfExamplesInfo("Skywork/Skywork-R1V-38B",
                                           trust_remote_code=True),
505
    "SmolVLMForConditionalGeneration": _HfExamplesInfo("HuggingFaceTB/SmolVLM2-2.2B-Instruct",  # noqa: E501
506
507
                                                       min_transformers_version="4.56",
                                                       transformers_version_reason="HF model broken in 4.55"),  # noqa: E501
Song's avatar
Song committed
508
    "Step3VLForConditionalGeneration": _HfExamplesInfo("stepfun-ai/step3",
509
                                                        trust_remote_code=True),
510
    "UltravoxModel": _HfExamplesInfo("fixie-ai/ultravox-v0_5-llama-3_2-1b",  # noqa: E501
511
                                     trust_remote_code=True),
512
    "TarsierForConditionalGeneration": _HfExamplesInfo("omni-research/Tarsier-7b"),  # noqa: E501
513
514
    "Tarsier2ForConditionalGeneration": _HfExamplesInfo("omni-research/Tarsier2-Recap-7b",  # noqa: E501
                                                        hf_overrides={"architectures": ["Tarsier2ForConditionalGeneration"]}),  # noqa: E501
515
516
517
518
519
520
    "VoxtralForConditionalGeneration": _HfExamplesInfo(
        "mistralai/Voxtral-Mini-3B-2507",
        min_transformers_version="4.54",
        # disable this temporarily until we support HF format
        is_available_online=False,
    ),
521
    # [Encoder-decoder]
汪志鹏's avatar
汪志鹏 committed
522
523
524
    "DonutForConditionalGeneration": _HfExamplesInfo("naver-clova-ix/donut-base-finetuned-docvqa",  # noqa: E501
                                                    hf_overrides={"architectures": ["DonutForConditionalGeneration"], "model_type": "donut"},  # noqa: E501
                                                    extras={"dolphin": "ByteDance/Dolphin"}),  # noqa: E501
525
526
527
    # Florence-2 uses BartFastTokenizer which can't be loaded from AutoTokenizer
    # Therefore, we borrow the BartTokenizer from the original Bart model
    "Florence2ForConditionalGeneration": _HfExamplesInfo("microsoft/Florence-2-base",  # noqa: E501
528
529
                                                         tokenizer="Isotr0py/Florence-2-tokenizer",  # noqa: E501
                                                         trust_remote_code=True),  # noqa: E501
530
    "MllamaForConditionalGeneration": _HfExamplesInfo("meta-llama/Llama-3.2-11B-Vision-Instruct"),  # noqa: E501
531
    "WhisperForConditionalGeneration": _HfExamplesInfo("openai/whisper-large-v3"),  # noqa: E501
532
533
    # [Cross-encoder]
    "JinaVLForRanking": _HfExamplesInfo("jinaai/jina-reranker-m0"),   # noqa: E501
534
535
}

536

537
538
539
_SPECULATIVE_DECODING_EXAMPLE_MODELS = {
    "MedusaModel": _HfExamplesInfo("JackFram/llama-68m",
                                   speculative_model="abhigoyal/vllm-medusa-llama-68m-random"),  # noqa: E501
540
541
542
543
    # Temporarily disabled.
    # TODO(woosuk): Re-enable this once the MLP Speculator is supported in V1.
    # "MLPSpeculatorPreTrainedModel": _HfExamplesInfo("JackFram/llama-160m",
    #                                                 speculative_model="ibm-ai-platform/llama-160m-accelerator"),  # noqa: E501
544
545
546
    "DeepSeekMTPModel": _HfExamplesInfo("luccafong/deepseek_mtp_main_random",
                                        speculative_model="luccafong/deepseek_mtp_draft_random",  # noqa: E501
                                        trust_remote_code=True),
547
548
549
    "EagleDeepSeekMTPModel": _HfExamplesInfo("eagle618/deepseek-v3-random",
                                        speculative_model="eagle618/eagle-deepseek-v3-random",  # noqa: E501
                                        trust_remote_code=True),
550
551
552
553
    "EagleLlamaForCausalLM": _HfExamplesInfo("yuhuili/EAGLE-LLaMA3-Instruct-8B",
                                             trust_remote_code=True,
                                             speculative_model="yuhuili/EAGLE-LLaMA3-Instruct-8B",
                                             tokenizer="meta-llama/Meta-Llama-3-8B-Instruct"),  # noqa: E501
554
555
556
557
    "Eagle3LlamaForCausalLM": _HfExamplesInfo("yuhuili/EAGLE3-LLaMA3.1-Instruct-8B",  # noqa: E501
                                            trust_remote_code=True,
                                            speculative_model="yuhuili/EAGLE3-LLaMA3.1-Instruct-8B",
                                            tokenizer="meta-llama/Llama-3.1-8B-Instruct"),
558
559
560
561
562
    # TODO: Re-enable this once tests/models/test_initialization.py is fixed, see PR #22333 #22611   # noqa: E501
    # "LlamaForCausalLMEagle3": _HfExamplesInfo("AngelSlim/Qwen3-8B_eagle3",  # noqa: E501
    #                                         trust_remote_code=True,
    #                                         speculative_model="AngelSlim/Qwen3-8B_eagle3",   # noqa: E501
    #                                         tokenizer="Qwen/Qwen3-8B"),
zhiweiz's avatar
zhiweiz committed
563
564
565
566
567
    "EagleLlama4ForCausalLM": _HfExamplesInfo(
        "morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct",
        trust_remote_code=True,
        speculative_model="morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct",
        tokenizer="meta-llama/Llama-4-Scout-17B-16E-Instruct"),  # noqa: E501
568
569
570
571
572
    "EagleMiniCPMForCausalLM": _HfExamplesInfo("openbmb/MiniCPM-1B-sft-bf16",
                                            trust_remote_code=True,
                                            is_available_online=False,
                                            speculative_model="openbmb/MiniCPM-2B-sft-bf16",
                                            tokenizer="openbmb/MiniCPM-2B-sft-bf16"),
573
574
575
    "ErnieMTPModel": _HfExamplesInfo("baidu/ERNIE-4.5-21B-A3B-PT",
                                    trust_remote_code=True,
                                    speculative_model="baidu/ERNIE-4.5-21B-A3B-PT"),
Yuxuan Zhang's avatar
Yuxuan Zhang committed
576
577
    "Glm4MoeMTPModel": _HfExamplesInfo("zai-org/GLM-4.5",
                                        speculative_model="zai-org/GLM-4.5",
Yuxuan Zhang's avatar
Yuxuan Zhang committed
578
579
                                        min_transformers_version="4.54",
                                        is_available_online=False),
580
581
582
    "MiMoMTPModel": _HfExamplesInfo("XiaomiMiMo/MiMo-7B-RL",
                                    trust_remote_code=True,
                                    speculative_model="XiaomiMiMo/MiMo-7B-RL")
583
584
}

585
_TRANSFORMERS_BACKEND_MODELS = {
586
    "TransformersModel": _HfExamplesInfo("Qwen/Qwen3-Embedding-0.6B"),
587
    "TransformersForCausalLM": _HfExamplesInfo("hmellor/Ilama-3.2-1B", trust_remote_code=True),  # noqa: E501
588
    "TransformersForMultimodalLM": _HfExamplesInfo("BAAI/Emu3-Chat-hf"),
589
590
}

591
592
593
_EXAMPLE_MODELS = {
    **_TEXT_GENERATION_EXAMPLE_MODELS,
    **_EMBEDDING_EXAMPLE_MODELS,
594
    **_SEQUENCE_CLASSIFICATION_EXAMPLE_MODELS,
595
596
    **_MULTIMODAL_EXAMPLE_MODELS,
    **_SPECULATIVE_DECODING_EXAMPLE_MODELS,
597
    **_TRANSFORMERS_BACKEND_MODELS,
598
599
600
601
602
603
604
605
606
}


class HfExampleModels:
    def __init__(self, hf_models: Mapping[str, _HfExamplesInfo]) -> None:
        super().__init__()

        self.hf_models = hf_models

607
    def get_supported_archs(self) -> Set[str]:
608
609
610
611
612
        return self.hf_models.keys()

    def get_hf_info(self, model_arch: str) -> _HfExamplesInfo:
        return self.hf_models[model_arch]

613
614
615
616
617
    def find_hf_info(self, model_id: str) -> _HfExamplesInfo:
        for info in self.hf_models.values():
            if info.default == model_id:
                return info

618
619
620
621
622
        # Fallback to extras
        for info in self.hf_models.values():
            if any(extra == model_id for extra in info.extras.values()):
                return info

623
624
        raise ValueError(f"No example model defined for {model_id}")

625

Patrick von Platen's avatar
Patrick von Platen committed
626
HF_EXAMPLE_MODELS = HfExampleModels(_EXAMPLE_MODELS)
627
AUTO_EXAMPLE_MODELS = HfExampleModels(_AUTOMATIC_CONVERTED_MODELS)