registry.py 42.2 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

import pytest
9
import torch
10
11
from packaging.version import Version
from transformers import __version__ as TRANSFORMERS_VERSION
12

13
from vllm.config import ModelDType, TokenizerMode
14

15
16
17
18
19
20
21
22
23
24
25
26

@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."""

27
    tokenizer_mode: TokenizerMode = "auto"
28
29
30
31
32
33
34
35
    """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.
    """

36
37
38
39
40
    min_transformers_version: Optional[str] = None
    """
    The minimum version of HF Transformers that is required to run this model.
    """

41
42
43
44
45
46
47
48
49
50
    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.
    """

51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
    skip_tokenizer_init: bool = False
    """
    If true, skip initialization of tokenizer and detokenizer. 
    """

    dtype: ModelDType = "auto"
    """
    The data type for the model weights and activations.
    """

    enforce_eager: bool = False
    """
    Whether to enforce eager execution. If True, we will
    disable CUDA graph and always execute the model in eager mode.
    If False, we will use CUDA graph and eager execution in hybrid.
    """

68
69
70
71
72
73
74
75
76
77
78
    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."""

79
80
81
    v0_only: bool = False
    """The model is only available with the vLLM V0 engine."""

82
83
84
    hf_overrides: dict[str, Any] = field(default_factory=dict)
    """The ``hf_overrides`` required to load the model."""

85
86
87
88
89
90
    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.
    """

91
92
93
94
95
96
    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.
    """

97
98
99
    max_num_seqs: Optional[int] = None
    """Maximum number of sequences to be processed in a single iteration."""

100
101
102
    def check_transformers_version(
        self,
        *,
103
        on_fail: Literal["error", "skip", "return"],
104
105
        check_min_version: bool = True,
        check_max_version: bool = True,
106
    ) -> Optional[str]:
107
108
109
110
        """
        If the installed transformers version does not meet the requirements,
        perform the given action.
        """
111
112
        if (self.min_transformers_version is None
                and self.max_transformers_version is None):
113
            return None
114
115

        current_version = TRANSFORMERS_VERSION
116
        cur_base_version = Version(current_version).base_version
117
118
119
        min_version = self.min_transformers_version
        max_version = self.max_transformers_version
        msg = f"`transformers=={current_version}` installed, but `transformers"
120
121
        # Only check the base version for the min/max version, otherwise preview
        # models cannot be run because `x.yy.0.dev0`<`x.yy.0`
122
123
        if (check_min_version and min_version
                and Version(cur_base_version) < Version(min_version)):
124
            msg += f">={min_version}` is required to run this model."
125
126
        elif (check_max_version and max_version
              and Version(cur_base_version) > Version(max_version)):
127
128
            msg += f"<={max_version}` is required to run this model."
        else:
129
            return None
130

131
132
133
134
135
        if self.transformers_version_reason:
            msg += f" Reason: {self.transformers_version_reason}"

        if on_fail == "error":
            raise RuntimeError(msg)
136
        elif on_fail == "skip":
137
            pytest.skip(msg)
138

139
140
        return msg

141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
    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)

157
158
159
160

# yapf: disable
_TEXT_GENERATION_EXAMPLE_MODELS = {
    # [Decoder-only]
161
    "ApertusForCausalLM": _HfExamplesInfo("swiss-ai/Apertus-8B-2509",
162
163
                                          min_transformers_version="4.56.0",
                                          trust_remote_code=True),
164
165
166
167
    "AquilaModel": _HfExamplesInfo("BAAI/AquilaChat-7B",
                                   trust_remote_code=True),
    "AquilaForCausalLM": _HfExamplesInfo("BAAI/AquilaChat2-7B",
                                         trust_remote_code=True),
168
    "ArceeForCausalLM": _HfExamplesInfo("arcee-ai/AFM-4.5B-Base"),
169
170
171
172
173
174
    "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),
175
176
    "BailingMoeForCausalLM": _HfExamplesInfo("inclusionAI/Ling-lite-1.5",
                                         trust_remote_code=True),
177
    "BambaForCausalLM": _HfExamplesInfo("ibm-ai-platform/Bamba-9B-v1",
178
                                        min_transformers_version="4.55.3",
179
                                        extras={"tiny": "hmellor/tiny-random-BambaForCausalLM"}),  # noqa: E501
180
181
    "BloomForCausalLM": _HfExamplesInfo("bigscience/bloom-560m",
                                        {"1b": "bigscience/bloomz-1b1"}),
182
    "ChatGLMModel": _HfExamplesInfo("zai-org/chatglm3-6b",
183
                                    trust_remote_code=True,
184
                                    max_transformers_version="4.48"),
185
186
    "ChatGLMForConditionalGeneration": _HfExamplesInfo("thu-coai/ShieldLM-6B-chatglm3",  # noqa: E501
                                                       trust_remote_code=True),
187
188
    "CohereForCausalLM": _HfExamplesInfo("CohereForAI/c4ai-command-r-v01",
                                         trust_remote_code=True),
189
190
    "Cohere2ForCausalLM": _HfExamplesInfo("CohereForAI/c4ai-command-r7b-12-2024", # noqa: E501
                                         trust_remote_code=True),
191
    "DbrxForCausalLM": _HfExamplesInfo("databricks/dbrx-instruct"),
192
    "DeciLMForCausalLM": _HfExamplesInfo("nvidia/Llama-3_3-Nemotron-Super-49B-v1", # noqa: E501
193
194
195
196
                                         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
197
198
    "DeepseekV3ForCausalLM": _HfExamplesInfo("deepseek-ai/DeepSeek-V3",  # noqa: E501
                                         trust_remote_code=True),
199
    "Ernie4_5ForCausalLM": _HfExamplesInfo("baidu/ERNIE-4.5-0.3B-PT",
200
                                            min_transformers_version="4.54"),
201
    "Ernie4_5_MoeForCausalLM": _HfExamplesInfo("baidu/ERNIE-4.5-21B-A3B-PT",
202
                                               min_transformers_version="4.54"),
203
204
205
206
    "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"),
207
    "Fairseq2LlamaForCausalLM": _HfExamplesInfo("mgleize/fairseq2-dummy-Llama-3.2-1B"),  # noqa: E501
208
    "FalconForCausalLM": _HfExamplesInfo("tiiuae/falcon-7b"),
209
    "FalconH1ForCausalLM":_HfExamplesInfo("tiiuae/Falcon-H1-0.5B-Base"),
210
    "GemmaForCausalLM": _HfExamplesInfo("google/gemma-1.1-2b-it"),
211
    "Gemma2ForCausalLM": _HfExamplesInfo("google/gemma-2-9b"),
212
    "Gemma3ForCausalLM": _HfExamplesInfo("google/gemma-3-1b-it"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
213
    "Gemma3nForCausalLM": _HfExamplesInfo("google/gemma-3n-E2B-it",
Robert Shaw's avatar
Robert Shaw committed
214
                                          min_transformers_version="4.53"),
215
216
    "GlmForCausalLM": _HfExamplesInfo("zai-org/glm-4-9b-chat-hf"),
    "Glm4ForCausalLM": _HfExamplesInfo("zai-org/GLM-4-9B-0414"),
217
218
    "Glm4MoeForCausalLM": _HfExamplesInfo("zai-org/GLM-4.5",
                                          min_transformers_version="4.54"),   # noqa: E501
219
220
221
    "GPT2LMHeadModel": _HfExamplesInfo("openai-community/gpt2",
                                       {"alias": "gpt2"}),
    "GPTBigCodeForCausalLM": _HfExamplesInfo("bigcode/starcoder",
222
                                             extras={"tiny": "bigcode/tiny_starcoder_py"},  # noqa: E501
223
224
                                             min_transformers_version="4.55.1",
                                             transformers_version_reason="HF model broken in 4.55.0"),  # noqa: E501
225
226
227
228
    "GPTJForCausalLM": _HfExamplesInfo("Milos/slovak-gpt-j-405M",
                                       {"6b": "EleutherAI/gpt-j-6b"}),
    "GPTNeoXForCausalLM": _HfExamplesInfo("EleutherAI/pythia-70m",
                                          {"1b": "EleutherAI/pythia-1.4b"}),
229
    "GptOssForCausalLM": _HfExamplesInfo("lmsys/gpt-oss-20b-bf16"),
230
231
    "GraniteForCausalLM": _HfExamplesInfo("ibm/PowerLM-3b"),
    "GraniteMoeForCausalLM": _HfExamplesInfo("ibm/PowerMoE-3b"),
232
233
    "GraniteMoeHybridForCausalLM": _HfExamplesInfo("ibm-granite/granite-4.0-tiny-preview", # noqa: E501
                                                   min_transformers_version="4.55.3"),
234
    "GraniteMoeSharedForCausalLM": _HfExamplesInfo("ibm-research/moe-7b-1b-active-shared-experts"),  # noqa: E501
Michael Goin's avatar
Michael Goin committed
235
236
    "Grok1ModelForCausalLM": _HfExamplesInfo("hpcai-tech/grok-1",
                                             trust_remote_code=True),
237
238
    "HunYuanMoEV1ForCausalLM": _HfExamplesInfo("tencent/Hunyuan-A13B-Instruct",
                                               trust_remote_code=True),
239
    # TODO: Remove is_available_online once their config.json is fixed
240
    "HunYuanDenseV1ForCausalLM":_HfExamplesInfo("tencent/Hunyuan-7B-Instruct-0124",
241
242
                                                trust_remote_code=True,
                                                is_available_online=False),
243
244
245
246
247
248
    "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),
249
250
    "InternLM3ForCausalLM": _HfExamplesInfo("internlm/internlm3-8b-instruct",
                                            trust_remote_code=True),
251
    "JAISLMHeadModel": _HfExamplesInfo("inceptionai/jais-13b-chat"),
252
    "JambaForCausalLM": _HfExamplesInfo("ai21labs/AI21-Jamba-1.5-Mini",
253
                                        min_transformers_version="4.55.3",
254
255
256
257
                                        extras={
                                            "tiny": "ai21labs/Jamba-tiny-dev",
                                            "random": "ai21labs/Jamba-tiny-random",  # noqa: E501
                                        }),
258
259
    "Lfm2ForCausalLM": _HfExamplesInfo("LiquidAI/LFM2-1.2B",
                                       min_transformers_version="4.54"),
260
261
    "LlamaForCausalLM": _HfExamplesInfo("meta-llama/Llama-3.2-1B-Instruct",
                                        extras={"guard": "meta-llama/Llama-Guard-3-1B",  # noqa: E501
262
263
                                                "hermes": "NousResearch/Hermes-3-Llama-3.1-8B", # noqa: E501
                                                "fp8": "RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8"}),  # noqa: E501
264
265
    "LLaMAForCausalLM": _HfExamplesInfo("decapoda-research/llama-7b-hf",
                                        is_available_online=False),
266
267
    "Llama4ForCausalLM": _HfExamplesInfo("meta-llama/Llama-4-Scout-17B-16E-Instruct", # noqa: E501
                                         is_available_online=False),
268
    "MambaForCausalLM": _HfExamplesInfo("state-spaces/mamba-130m-hf"),
269
270
271
272
273
    "Mamba2ForCausalLM": _HfExamplesInfo("mistralai/Mamba-Codestral-7B-v0.1",
                                         min_transformers_version="4.55.3",
                                         extras={
                                            "random": "yujiepan/mamba2-codestral-v0.1-tiny-random", # noqa: E501
                                         }),
274
275
276
277
278
    "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),
279
    "MiniMaxForCausalLM": _HfExamplesInfo("MiniMaxAI/MiniMax-Text-01-hf"),
280
    "MiniMaxText01ForCausalLM": _HfExamplesInfo("MiniMaxAI/MiniMax-Text-01",
281
282
                                                trust_remote_code=True,
                                                revision="a59aa9cbc53b9fb8742ca4e9e1531b9802b6fdc3"),  # noqa: E501
283
284
    "MiniMaxM1ForCausalLM": _HfExamplesInfo("MiniMaxAI/MiniMax-M1-40k",
                                            trust_remote_code=True),
285
    "MistralForCausalLM": _HfExamplesInfo("mistralai/Mistral-7B-Instruct-v0.1"),
286
    "MixtralForCausalLM": _HfExamplesInfo("mistralai/Mixtral-8x7B-Instruct-v0.1",  # noqa: E501
287
                                          {"tiny": "TitanML/tiny-mixtral"}),  # noqa: E501
288
289
290
    "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
291
292
    "NemotronHForCausalLM": _HfExamplesInfo("nvidia/Nemotron-H-8B-Base-8K",
                                            trust_remote_code=True),
293
    "OlmoForCausalLM": _HfExamplesInfo("allenai/OLMo-1B-hf"),
294
    "Olmo2ForCausalLM": _HfExamplesInfo("allenai/OLMo-2-0425-1B"),
295
    "OlmoeForCausalLM": _HfExamplesInfo("allenai/OLMoE-1B-7B-0924-Instruct"),
296
297
    "OPTForCausalLM": _HfExamplesInfo("facebook/opt-125m",
                                      {"1b": "facebook/opt-iml-max-1.3b"}),
298
299
300
    "OrionForCausalLM": _HfExamplesInfo("OrionStarAI/Orion-14B-Chat",
                                        trust_remote_code=True),
    "PersimmonForCausalLM": _HfExamplesInfo("adept/persimmon-8b-chat"),
301
    "PhiForCausalLM": _HfExamplesInfo("microsoft/phi-2"),
302
    "Phi3ForCausalLM": _HfExamplesInfo("microsoft/Phi-3-mini-4k-instruct"),
303
304
305
306
    "Phi4FlashForCausalLM": _HfExamplesInfo("microsoft/Phi-4-mini-flash-reasoning", # noqa: E501
                                        trust_remote_code=True,
                                        v0_only=True,
                                        max_model_len=10240),
307
308
    "PhiMoEForCausalLM": _HfExamplesInfo("microsoft/Phi-3.5-MoE-instruct",
                                         trust_remote_code=True),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
309
310
    "Plamo2ForCausalLM": _HfExamplesInfo("pfnet/plamo-2-1b",
                                        trust_remote_code=True),
311
    "QWenLMHeadModel": _HfExamplesInfo("Qwen/Qwen-7B-Chat",
312
313
                                       max_transformers_version="4.53",
                                       transformers_version_reason="HF model uses remote code that is not compatible with latest Transformers",  # noqa: E501
314
                                       trust_remote_code=True),
315
316
    "Qwen2ForCausalLM": _HfExamplesInfo("Qwen/Qwen2-0.5B-Instruct",
                                        extras={"2.5": "Qwen/Qwen2.5-0.5B-Instruct"}), # noqa: E501
317
    "Qwen2MoeForCausalLM": _HfExamplesInfo("Qwen/Qwen1.5-MoE-A2.7B-Chat"),
Jee Jee Li's avatar
Jee Jee Li committed
318
319
    "Qwen3ForCausalLM": _HfExamplesInfo("Qwen/Qwen3-8B"),
    "Qwen3MoeForCausalLM": _HfExamplesInfo("Qwen/Qwen3-30B-A3B"),
320
    "RWForCausalLM": _HfExamplesInfo("tiiuae/falcon-40b"),
321
322
323
    "SeedOssForCausalLM": _HfExamplesInfo("ByteDance-Seed/Seed-OSS-36B-Instruct", # noqa: E501
                                          trust_remote_code=True,
                                          is_available_online=False),
324
    "SmolLM3ForCausalLM": _HfExamplesInfo("HuggingFaceTB/SmolLM3-3B"),
325
326
    "StableLMEpochForCausalLM": _HfExamplesInfo("stabilityai/stablelm-zephyr-3b"),  # noqa: E501
    "StableLmForCausalLM": _HfExamplesInfo("stabilityai/stablelm-3b-4e1t"),
327
    "Starcoder2ForCausalLM": _HfExamplesInfo("bigcode/starcoder2-3b"),
Song's avatar
Song committed
328
    "Step3TextForCausalLM": _HfExamplesInfo("stepfun-ai/step3",
329
                                            trust_remote_code=True),
330
331
    "SolarForCausalLM": _HfExamplesInfo("upstage/solar-pro-preview-instruct",
                                        trust_remote_code=True),
332
333
    "TeleChat2ForCausalLM": _HfExamplesInfo("Tele-AI/TeleChat2-3B",
                                            trust_remote_code=True),
334
335
    "TeleFLMForCausalLM": _HfExamplesInfo("CofeAI/FLM-2-52B-Instruct-2407",
                                            trust_remote_code=True),
336
    "XverseForCausalLM": _HfExamplesInfo("xverse/XVERSE-7B-Chat",
337
                                         tokenizer="meta-llama/Llama-2-7b",
338
                                         trust_remote_code=True),
339
    "Zamba2ForCausalLM": _HfExamplesInfo("Zyphra/Zamba2-7B-instruct"),
340
341
    "MiMoForCausalLM": _HfExamplesInfo("XiaomiMiMo/MiMo-7B-RL",
                                        trust_remote_code=True),
342
    "Dots1ForCausalLM": _HfExamplesInfo("rednote-hilab/dots.llm1.inst"),
343
344
345
    # [Encoder-decoder]
    "BartModel": _HfExamplesInfo("facebook/bart-base"),
    "BartForConditionalGeneration": _HfExamplesInfo("facebook/bart-large-cnn"),
汪志鹏's avatar
汪志鹏 committed
346
347
    "MBartForConditionalGeneration": _HfExamplesInfo("facebook/mbart-large-en-ro",  # noqa: E501
                                                    hf_overrides={"architectures": ["MBartForConditionalGeneration"]}),  # noqa: E501
348
349
350
351
}

_EMBEDDING_EXAMPLE_MODELS = {
    # [Text-only]
352
353
    "BertModel": _HfExamplesInfo("BAAI/bge-base-en-v1.5"),
    "Gemma2Model": _HfExamplesInfo("BAAI/bge-multilingual-gemma2"),  # noqa: E501
354
    "Gemma3TextModel": _HfExamplesInfo("google/embeddinggemma-300m"),
355
    "GritLM": _HfExamplesInfo("parasail-ai/GritLM-7B-vllm"),
356
357
    "GteModel": _HfExamplesInfo("Snowflake/snowflake-arctic-embed-m-v2.0",
                                               trust_remote_code=True),
358
359
    "GteNewModel": _HfExamplesInfo("Alibaba-NLP/gte-base-en-v1.5",
                                   trust_remote_code=True,
360
                                   hf_overrides={"architectures": ["GteNewModel"]}),  # noqa: E501
361
362
    "InternLM2ForRewardModel": _HfExamplesInfo("internlm/internlm2-1_8b-reward",
                                               trust_remote_code=True),
363
    "JambaForSequenceClassification": _HfExamplesInfo("ai21labs/Jamba-tiny-reward-dev"),  # noqa: E501
364
    "LlamaModel": _HfExamplesInfo("llama", is_available_online=False),
365
    "MistralModel": _HfExamplesInfo("intfloat/e5-mistral-7b-instruct"),
366
    "ModernBertModel": _HfExamplesInfo("Alibaba-NLP/gte-modernbert-base",
367
                                trust_remote_code=True),
368
    "NomicBertModel": _HfExamplesInfo("nomic-ai/nomic-embed-text-v2-moe",
369
                                               trust_remote_code=True),  # noqa: E501
370
    "Qwen2Model": _HfExamplesInfo("ssmits/Qwen2-7B-Instruct-embed-base"),
371
372
373
374
375
376
    "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
377
378
379
    "RobertaModel": _HfExamplesInfo("sentence-transformers/stsb-roberta-base-v2"),  # noqa: E501
    "RobertaForMaskedLM": _HfExamplesInfo("sentence-transformers/all-roberta-large-v1"),  # noqa: E501
    "XLMRobertaModel": _HfExamplesInfo("intfloat/multilingual-e5-small"),  # noqa: E501
380
381
382
383
384
    # [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
385
    "PrithviGeoSpatialMAE": _HfExamplesInfo("ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11", # noqa: E501
386
387
388
389
390
391
392
                                            dtype=torch.float16,
                                            enforce_eager=True,
                                            skip_tokenizer_init=True,
                                            # This is to avoid the model
                                            # going OOM in CI
                                            max_num_seqs=32,
                                            ),
393
    "Terratorch": _HfExamplesInfo("ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11", # noqa: E501
394
395
396
397
398
399
                                  dtype=torch.float16,
                                  enforce_eager=True,
                                  skip_tokenizer_init=True,
                                  # This is to avoid the model going OOM in CI
                                  max_num_seqs=32,
                                  ),
400
401
}

402
403
404
405
406
_SEQUENCE_CLASSIFICATION_EXAMPLE_MODELS = {
    # [Decoder-only]
    "GPT2ForSequenceClassification": _HfExamplesInfo("nie3e/sentiment-polish-gpt2-small"),  # noqa: E501

    # [Cross-encoder]
407
    "BertForSequenceClassification": _HfExamplesInfo("cross-encoder/ms-marco-MiniLM-L-6-v2"),  # noqa: E501
408
409
410
411
    "GteNewForSequenceClassification": _HfExamplesInfo("Alibaba-NLP/gte-multilingual-reranker-base",  # noqa: E501
                                                       trust_remote_code=True,
                                                       hf_overrides={
                                                           "architectures": ["GteNewForSequenceClassification"]}),# noqa: E501
412
413
414
    "ModernBertForSequenceClassification": _HfExamplesInfo("Alibaba-NLP/gte-reranker-modernbert-base"), # noqa: E501
    "RobertaForSequenceClassification": _HfExamplesInfo("cross-encoder/quora-roberta-base"),  # noqa: E501
    "XLMRobertaForSequenceClassification": _HfExamplesInfo("BAAI/bge-reranker-v2-m3"),  # noqa: E501
415
416
}

417
418
419
420
421
422
423
424
425
426
427
_AUTOMATIC_CONVERTED_MODELS = {
    # Use as_seq_cls_model for automatic conversion
    "GemmaForSequenceClassification": _HfExamplesInfo("BAAI/bge-reranker-v2-gemma",  # noqa: E501
                                                      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
}

428
429
_MULTIMODAL_EXAMPLE_MODELS = {
    # [Decoder-only]
430
    "AriaForConditionalGeneration": _HfExamplesInfo("rhymes-ai/Aria"),
Jennifer Zhao's avatar
Jennifer Zhao committed
431
    "AyaVisionForConditionalGeneration": _HfExamplesInfo("CohereForAI/aya-vision-8b"), # noqa: E501
432
    "Blip2ForConditionalGeneration": _HfExamplesInfo("Salesforce/blip2-opt-2.7b",  # noqa: E501
433
                                                     extras={"6b": "Salesforce/blip2-opt-6.7b"}),  # noqa: E501
434
    "ChameleonForConditionalGeneration": _HfExamplesInfo("facebook/chameleon-7b"),  # noqa: E501
435
    "Cohere2VisionForConditionalGeneration": _HfExamplesInfo("CohereLabs/command-a-vision-07-2025"), # noqa: E501
436
    "DeepseekVLV2ForCausalLM": _HfExamplesInfo("deepseek-ai/deepseek-vl2-tiny",  # noqa: E501
437
438
439
                                                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
440
                                                hf_overrides={"architectures": ["DeepseekVLV2ForCausalLM"]}),  # noqa: E501
441
    "Emu3ForConditionalGeneration": _HfExamplesInfo("BAAI/Emu3-Chat-hf"),
442
443
    "Ernie4_5_VLMoeForConditionalGeneration": _HfExamplesInfo("baidu/ERNIE-4.5-VL-28B-A3B-PT",  # noqa: E501
                                                              trust_remote_code=True),
444
    "FuyuForCausalLM": _HfExamplesInfo("adept/fuyu-8b"),
445
    "Gemma3ForConditionalGeneration": _HfExamplesInfo("google/gemma-3-4b-it"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
446
447
    "Gemma3nForConditionalGeneration": _HfExamplesInfo("google/gemma-3n-E2B-it",    # noqa: E501
                                        min_transformers_version="4.53"),
448
    "GraniteSpeechForConditionalGeneration": _HfExamplesInfo("ibm-granite/granite-speech-3.3-2b"),  # noqa: E501
449
    "GLM4VForCausalLM": _HfExamplesInfo("zai-org/glm-4v-9b",
450
451
                                        trust_remote_code=True,
                                        hf_overrides={"architectures": ["GLM4VForCausalLM"]}),  # noqa: E501
452
    "Glm4vForConditionalGeneration": _HfExamplesInfo("zai-org/GLM-4.1V-9B-Thinking"),  # noqa: E501
453
    "Glm4vMoeForConditionalGeneration": _HfExamplesInfo("zai-org/GLM-4.5V",
454
                                                        min_transformers_version="4.56"),  # noqa: E501
455
    "H2OVLChatModel": _HfExamplesInfo("h2oai/h2ovl-mississippi-800m",
456
                                      trust_remote_code=True,
457
458
459
                                      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
460
461
    "HCXVisionForCausalLM": _HfExamplesInfo("naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B",  # noqa: E501
                                            trust_remote_code=True),
462
463
    "Idefics3ForConditionalGeneration": _HfExamplesInfo("HuggingFaceM4/Idefics3-8B-Llama3",  # noqa: E501
                                                        {"tiny": "HuggingFaceTB/SmolVLM-256M-Instruct"},    # noqa: E501
464
465
                                                        min_transformers_version="4.56",
                                                        transformers_version_reason="HF model broken in 4.55"),  # noqa: E501
466
467
    "InternS1ForConditionalGeneration": _HfExamplesInfo("internlm/Intern-S1",
                                                        trust_remote_code=True),  # noqa: E501
468
    "InternVLChatModel": _HfExamplesInfo("OpenGVLab/InternVL2-1B",
469
                                         extras={"2B": "OpenGVLab/InternVL2-2B",
470
471
472
473
                                                 "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
474
                                         trust_remote_code=True),
475
    "InternVLForConditionalGeneration": _HfExamplesInfo("OpenGVLab/InternVL3-1B-hf"),    # noqa: E501
476
477
    "KeyeForConditionalGeneration": _HfExamplesInfo("Kwai-Keye/Keye-VL-8B-Preview", # noqa: E501
                                                    trust_remote_code=True),
478
479
    "KeyeVL1_5ForConditionalGeneration": _HfExamplesInfo("Kwai-Keye/Keye-VL-1_5-8B", # noqa: E501
                                                         trust_remote_code=True),
480
481
    "KimiVLForConditionalGeneration": _HfExamplesInfo("moonshotai/Kimi-VL-A3B-Instruct",  # noqa: E501
                                                      extras={"thinking": "moonshotai/Kimi-VL-A3B-Thinking"},  # noqa: E501
482
                                                      trust_remote_code=True),
483
    "Llama4ForConditionalGeneration": _HfExamplesInfo("meta-llama/Llama-4-Scout-17B-16E-Instruct",   # noqa: E501
484
485
486
                                                      max_model_len=10240,
                                                      extras={"llama-guard-4": "meta-llama/Llama-Guard-4-12B"}, # noqa: E501
                                                      ),
487
    "LlavaForConditionalGeneration": _HfExamplesInfo("llava-hf/llava-1.5-7b-hf",
488
489
                                                     extras={"mistral": "mistral-community/pixtral-12b", # noqa: E501
                                                             "mistral-fp8": "nm-testing/pixtral-12b-FP8-dynamic"}),  # noqa: E501
490
491
492
    "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
493
    "MantisForConditionalGeneration": _HfExamplesInfo("TIGER-Lab/Mantis-8B-siglip-llama3",  # noqa: E501
494
495
                                                      max_transformers_version="4.48",  # noqa: E501
                                                      transformers_version_reason="HF model is not compatible.",  # noqa: E501
496
                                                      hf_overrides={"architectures": ["MantisForConditionalGeneration"]}),  # noqa: E501
497
498
    "MiDashengLMModel": _HfExamplesInfo("mispeech/midashenglm-7b",
                            trust_remote_code=True),
499
500
    "MiniCPMO": _HfExamplesInfo("openbmb/MiniCPM-o-2_6",
                                trust_remote_code=True),
501
    "MiniCPMV": _HfExamplesInfo("openbmb/MiniCPM-Llama3-V-2_5",
tc-mb's avatar
tc-mb committed
502
                                extras={"2.6": "openbmb/MiniCPM-V-2_6", "4.0": "openbmb/MiniCPM-V-4", "4.5": "openbmb/MiniCPM-V-4_5"},  # noqa: E501
503
                                trust_remote_code=True),
504
    "MiniMaxVL01ForConditionalGeneration": _HfExamplesInfo("MiniMaxAI/MiniMax-VL-01", # noqa: E501
505
506
                                              trust_remote_code=True,
                                              v0_only=True),
507
508
    "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
509
    "MolmoForCausalLM": _HfExamplesInfo("allenai/Molmo-7B-D-0924",
510
                                        max_transformers_version="4.48",
511
                                        transformers_version_reason="Incorrectly-detected `tensorflow` import.",  # noqa: E501
512
                                        extras={"olmo": "allenai/Molmo-7B-O-0924"},  # noqa: E501
513
514
515
                                        trust_remote_code=True),
    "NVLM_D": _HfExamplesInfo("nvidia/NVLM-D-72B",
                              trust_remote_code=True),
516
517
    "Llama_Nemotron_Nano_VL" : _HfExamplesInfo("nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1", # noqa: E501
                                                     trust_remote_code=True),
518
    "Ovis": _HfExamplesInfo("AIDC-AI/Ovis2-1B", trust_remote_code=True,
519
520
                            max_transformers_version="4.53",
                            transformers_version_reason="HF model is not compatible",  # noqa: E501
521
522
                            extras={"1.6-llama": "AIDC-AI/Ovis1.6-Llama3.2-3B",
                                    "1.6-gemma": "AIDC-AI/Ovis1.6-Gemma2-9B"}),  # noqa: E501
523
    "Ovis2_5": _HfExamplesInfo("AIDC-AI/Ovis2.5-2B",
524
                               trust_remote_code=True),
525
526
    "PaliGemmaForConditionalGeneration": _HfExamplesInfo("google/paligemma-3b-mix-224",  # noqa: E501
                                                         extras={"v2": "google/paligemma2-3b-ft-docci-448"}),  # noqa: E501
527
528
    "Phi3VForCausalLM": _HfExamplesInfo("microsoft/Phi-3-vision-128k-instruct",
                                        trust_remote_code=True,
529
530
                                        max_transformers_version="4.48",
                                        transformers_version_reason="Use of deprecated imports which have been removed.",  # noqa: E501
531
                                        extras={"phi3.5": "microsoft/Phi-3.5-vision-instruct"}),  # noqa: E501
532
533
    "Phi4MMForCausalLM": _HfExamplesInfo("microsoft/Phi-4-multimodal-instruct",
                                        trust_remote_code=True),
534
535
    "Phi4MultimodalForCausalLM": _HfExamplesInfo("microsoft/Phi-4-multimodal-instruct",  # noqa: E501
                                                 revision="refs/pr/70"),
536
    "PixtralForConditionalGeneration": _HfExamplesInfo("mistralai/Pixtral-12B-2409",  # noqa: E501
537
                                                       tokenizer_mode="mistral"),
538
539
540
541
    "QwenVLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen-VL",
                                                      extras={"chat": "Qwen/Qwen-VL-Chat"},  # noqa: E501
                                                      trust_remote_code=True,
                                                      hf_overrides={"architectures": ["QwenVLForConditionalGeneration"]}),  # noqa: E501
542
543
    "Qwen2AudioForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2-Audio-7B-Instruct"),  # noqa: E501
    "Qwen2VLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2-VL-2B-Instruct"),  # noqa: E501
544
545
    "Qwen2_5_VLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2.5-VL-3B-Instruct", # noqa: E501
                                                          max_model_len=4096),
546
547
    "Qwen2_5OmniModel": _HfExamplesInfo("Qwen/Qwen2.5-Omni-3B"),
    "Qwen2_5OmniForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2.5-Omni-7B-AWQ"),  # noqa: E501
548
549
    "RForConditionalGeneration": _HfExamplesInfo("YannQi/R-4B",
                                                 trust_remote_code=True),
550
551
    "SkyworkR1VChatModel": _HfExamplesInfo("Skywork/Skywork-R1V-38B",
                                           trust_remote_code=True),
552
    "SmolVLMForConditionalGeneration": _HfExamplesInfo("HuggingFaceTB/SmolVLM2-2.2B-Instruct",  # noqa: E501
553
554
                                                       min_transformers_version="4.56",
                                                       transformers_version_reason="HF model broken in 4.55"),  # noqa: E501
Song's avatar
Song committed
555
    "Step3VLForConditionalGeneration": _HfExamplesInfo("stepfun-ai/step3",
556
                                                        trust_remote_code=True),
557
    "UltravoxModel": _HfExamplesInfo("fixie-ai/ultravox-v0_5-llama-3_2-1b",  # noqa: E501
558
                                     trust_remote_code=True),
559
    "TarsierForConditionalGeneration": _HfExamplesInfo("omni-research/Tarsier-7b"),  # noqa: E501
560
561
    "Tarsier2ForConditionalGeneration": _HfExamplesInfo("omni-research/Tarsier2-Recap-7b",  # noqa: E501
                                                        hf_overrides={"architectures": ["Tarsier2ForConditionalGeneration"]}),  # noqa: E501
562
563
564
565
566
567
    "VoxtralForConditionalGeneration": _HfExamplesInfo(
        "mistralai/Voxtral-Mini-3B-2507",
        min_transformers_version="4.54",
        # disable this temporarily until we support HF format
        is_available_online=False,
    ),
568
    # [Encoder-decoder]
汪志鹏's avatar
汪志鹏 committed
569
570
571
    "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
572
573
574
    # 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
575
576
                                                         tokenizer="Isotr0py/Florence-2-tokenizer",  # noqa: E501
                                                         trust_remote_code=True),  # noqa: E501
577
    "MllamaForConditionalGeneration": _HfExamplesInfo("meta-llama/Llama-3.2-11B-Vision-Instruct"),  # noqa: E501
578
    "WhisperForConditionalGeneration": _HfExamplesInfo("openai/whisper-large-v3"),  # noqa: E501
579
580
    # [Cross-encoder]
    "JinaVLForRanking": _HfExamplesInfo("jinaai/jina-reranker-m0"),   # noqa: E501
581
582
}

583

584
585
586
_SPECULATIVE_DECODING_EXAMPLE_MODELS = {
    "MedusaModel": _HfExamplesInfo("JackFram/llama-68m",
                                   speculative_model="abhigoyal/vllm-medusa-llama-68m-random"),  # noqa: E501
587
588
589
590
    # 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
591
592
593
    "DeepSeekMTPModel": _HfExamplesInfo("luccafong/deepseek_mtp_main_random",
                                        speculative_model="luccafong/deepseek_mtp_draft_random",  # noqa: E501
                                        trust_remote_code=True),
594
595
596
    "EagleDeepSeekMTPModel": _HfExamplesInfo("eagle618/deepseek-v3-random",
                                        speculative_model="eagle618/eagle-deepseek-v3-random",  # noqa: E501
                                        trust_remote_code=True),
597
598
599
600
    "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
601
602
603
604
    "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"),
605
606
607
608
609
    # 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
610
611
612
613
614
    "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
615
616
617
618
619
    "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"),
620
621
622
    "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
623
624
    "Glm4MoeMTPModel": _HfExamplesInfo("zai-org/GLM-4.5",
                                        speculative_model="zai-org/GLM-4.5",
Yuxuan Zhang's avatar
Yuxuan Zhang committed
625
626
                                        min_transformers_version="4.54",
                                        is_available_online=False),
627
628
629
    "MiMoMTPModel": _HfExamplesInfo("XiaomiMiMo/MiMo-7B-RL",
                                    trust_remote_code=True,
                                    speculative_model="XiaomiMiMo/MiMo-7B-RL")
630
631
}

632
_TRANSFORMERS_BACKEND_MODELS = {
633
    "TransformersModel": _HfExamplesInfo("Qwen/Qwen3-Embedding-0.6B"),
634
    "TransformersForCausalLM": _HfExamplesInfo("hmellor/Ilama-3.2-1B", trust_remote_code=True),  # noqa: E501
635
    "TransformersForMultimodalLM": _HfExamplesInfo("BAAI/Emu3-Chat-hf"),
636
637
}

638
639
640
_EXAMPLE_MODELS = {
    **_TEXT_GENERATION_EXAMPLE_MODELS,
    **_EMBEDDING_EXAMPLE_MODELS,
641
    **_SEQUENCE_CLASSIFICATION_EXAMPLE_MODELS,
642
643
    **_MULTIMODAL_EXAMPLE_MODELS,
    **_SPECULATIVE_DECODING_EXAMPLE_MODELS,
644
    **_TRANSFORMERS_BACKEND_MODELS,
645
646
647
648
649
650
651
652
653
}


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

        self.hf_models = hf_models

654
    def get_supported_archs(self) -> Set[str]:
655
656
657
658
659
        return self.hf_models.keys()

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

660
661
662
663
664
    def find_hf_info(self, model_id: str) -> _HfExamplesInfo:
        for info in self.hf_models.values():
            if info.default == model_id:
                return info

665
666
667
668
669
        # Fallback to extras
        for info in self.hf_models.values():
            if any(extra == model_id for extra in info.extras.values()):
                return info

670
671
        raise ValueError(f"No example model defined for {model_id}")

672

Patrick von Platen's avatar
Patrick von Platen committed
673
HF_EXAMPLE_MODELS = HfExampleModels(_EXAMPLE_MODELS)
674
AUTO_EXAMPLE_MODELS = HfExampleModels(_AUTOMATIC_CONVERTED_MODELS)