registry.py 49.4 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

zhuwenwen's avatar
zhuwenwen committed
8
import os
9
10
11
import pytest
from packaging.version import Version
from transformers import __version__ as TRANSFORMERS_VERSION
zhuwenwen's avatar
zhuwenwen committed
12
# from ..utils import models_path_prefix
13

zhuwenwen's avatar
zhuwenwen committed
14
models_path_prefix = os.getenv('VLLM_OPTEST_MODELS_PATH') or os.getenv("OPTEST_MODELS_PATH")
15

16
17
from vllm.config import TokenizerMode

zhuwenwen's avatar
zhuwenwen committed
18

19
20
21
22
23
24
25
26
27
28
29
@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."""

30
    tokenizer_mode: TokenizerMode = "auto"
31
32
33
34
35
36
37
38
    """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.
    """

39
40
41
42
43
    min_transformers_version: Optional[str] = None
    """
    The minimum version of HF Transformers that is required to run this model.
    """

44
45
46
47
48
49
50
51
52
53
    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.
    """

54
55
56
57
58
59
60
61
62
63
64
    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."""

65
66
67
    v0_only: bool = False
    """The model is only available with the vLLM V0 engine."""

68
69
70
    hf_overrides: dict[str, Any] = field(default_factory=dict)
    """The ``hf_overrides`` required to load the model."""

71
72
73
74
75
76
    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.
    """

77
78
79
80
81
82
    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.
    """

83
84
85
    def check_transformers_version(
        self,
        *,
86
        on_fail: Literal["error", "skip", "return"],
87
88
        check_min_version: bool = True,
        check_max_version: bool = True,
89
    ) -> Optional[str]:
90
91
92
93
        """
        If the installed transformers version does not meet the requirements,
        perform the given action.
        """
94
95
        if (self.min_transformers_version is None
                and self.max_transformers_version is None):
96
            return None
97
98

        current_version = TRANSFORMERS_VERSION
99
        cur_base_version = Version(current_version).base_version
100
101
102
        min_version = self.min_transformers_version
        max_version = self.max_transformers_version
        msg = f"`transformers=={current_version}` installed, but `transformers"
103
104
        # Only check the base version for the min/max version, otherwise preview
        # models cannot be run because `x.yy.0.dev0`<`x.yy.0`
105
106
        if (check_min_version and min_version
                and Version(cur_base_version) < Version(min_version)):
107
            msg += f">={min_version}` is required to run this model."
108
109
        elif (check_max_version and max_version
              and Version(cur_base_version) > Version(max_version)):
110
111
            msg += f"<={max_version}` is required to run this model."
        else:
112
            return None
113

114
115
116
117
118
        if self.transformers_version_reason:
            msg += f" Reason: {self.transformers_version_reason}"

        if on_fail == "error":
            raise RuntimeError(msg)
119
        elif on_fail == "skip":
120
            pytest.skip(msg)
121

122
123
        return msg

124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
    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)

140
141
142
143

# yapf: disable
_TEXT_GENERATION_EXAMPLE_MODELS = {
    # [Decoder-only]
144
    "ApertusForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"swiss-ai/Apertus-8B"),
145
146
                                          min_transformers_version="4.56.0",
                                          trust_remote_code=True),
147
    "AquilaModel": _HfExamplesInfo(os.path.join(models_path_prefix,"BAAI/AquilaChat-7B"),
148
                                   trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
149
    "AquilaForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "BAAI/AquilaChat2-7B"),
150
                                         trust_remote_code=True),
151
    "ArceeForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "arcee-ai/AFM-4.5B-Base")),
zhuwenwen's avatar
zhuwenwen committed
152
    "ArcticForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "Snowflake/snowflake-arctic-instruct"),
153
                                         trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
154
    "BaiChuanForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "baichuan-inc/Baichuan-7B"),
155
                                         trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
156
    "BaichuanForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "baichuan-inc/Baichuan2-7B-chat"),
157
                                         trust_remote_code=True),
158
    "BailingMoeForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "inclusionAI/Ling-lite-1.5"),
159
                                         trust_remote_code=True),
160
    "BambaForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "ibm-ai-platform/Bamba-9B-v1"),
161
                                        min_transformers_version="4.56.0",
162
163
164
                                        extras={"tiny": os.path.join(models_path_prefix, "hmellor/tiny-random-BambaForCausalLM")}),  # noqa: E501
    "BloomForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "bigscience/bloom-560m"),
                                        {"1b": os.path.join(models_path_prefix, "bigscience/bloomz-1b1")}),
165
    "ChatGLMModel": _HfExamplesInfo(os.path.join(models_path_prefix, "zai-org/chatglm3-6b"),
166
                                    trust_remote_code=True,
167
                                    max_transformers_version="4.48"),
zhuwenwen's avatar
zhuwenwen committed
168
    "ChatGLMForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "thu-coai/ShieldLM-6B-chatglm3"),  # noqa: E501
169
                                                       trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
170
    "CohereForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "CohereForAI/c4ai-command-r-v01"),
171
                                         trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
172
    "Cohere2ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "CohereForAI/c4ai-command-r7b-12-2024"), # noqa: E501
173
                                         trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
174
175
    "DbrxForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "databricks/dbrx-instruct")),
    "DeciLMForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "nvidia/Llama-3_3-Nemotron-Super-49B-v1"), # noqa: E501
176
                                         trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
177
178
    "DeepseekForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "deepseek-ai/deepseek-llm-7b-chat")),
    "DeepseekV2ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "deepseek-ai/DeepSeek-V2-Lite-Chat"),  # noqa: E501
179
                                         trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
180
    "DeepseekV3ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "deepseek-ai/DeepSeek-V3"),  # noqa: E501
Robert Shaw's avatar
Robert Shaw committed
181
                                         trust_remote_code=True),
182
    "Ernie4_5ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "baidu/ERNIE-4.5-0.3B-PT"),
183
                                            min_transformers_version="4.54"),
184
    "Ernie4_5_MoeForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "baidu/ERNIE-4.5-21B-A3B-PT"),
185
                                               min_transformers_version="4.54"),
186
    "ExaoneForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct"),
187
                                         trust_remote_code=True),
188
    "Exaone4ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "LGAI-EXAONE/EXAONE-4.0-32B"),
189
                                          min_transformers_version="4.54"),
190
191
    "Fairseq2LlamaForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "mgleize/fairseq2-dummy-Llama-3.2-1B")),  # noqa: E501
    "FalconForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "tiiuae/falcon-7b")),
192
193
194
195
196
    "FalconH1ForCausalLM":_HfExamplesInfo(os.path.join(models_path_prefix, "tiiuae/Falcon-H1-0.5B-Base")),
    "GemmaForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "google/gemma-1.1-2b-it")),
    "Gemma2ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "google/gemma-2-9b")),
    "Gemma3ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "google/gemma-3-1b-it")),
    "Gemma3nForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "google/gemma-3n-E2B-it"),
Robert Shaw's avatar
Robert Shaw committed
197
                                          min_transformers_version="4.53"),
198
199
200
    "GlmForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "zai-org/glm-4-9b-chat-hf")),
    "Glm4ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "zai-org/GLM-4-9B-0414")),
    "Glm4MoeForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "zai-org/GLM-4.5"),
201
                                          min_transformers_version="4.54"),   # noqa: E501
202
203
204
205
    "GPT2LMHeadModel": _HfExamplesInfo(os.path.join(models_path_prefix, "openai-community/gpt2"),
                                       {"alias": os.path.join(models_path_prefix, "gpt2")}),
    "GPTBigCodeForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "bigcode/starcoder"),
                                             extras={"tiny": os.path.join(models_path_prefix, "bigcode/tiny_starcoder_py")},  # noqa: E501
206
207
                                             min_transformers_version="4.55.1",
                                             transformers_version_reason="HF model broken in 4.55.0"),  # noqa: E501
208
209
210
211
212
213
214
215
216
217
    "GPTJForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "Milos/slovak-gpt-j-405M"),
                                       {"6b": os.path.join(models_path_prefix, "EleutherAI/gpt-j-6b")}),
    "GPTNeoXForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "EleutherAI/pythia-70m"),
                                          {"1b": os.path.join(models_path_prefix, "EleutherAI/pythia-1.4b")}),
    "GptOssForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "lmsys/gpt-oss-20b-bf16")),
    "GraniteForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "ibm/PowerLM-3b")),
    "GraniteMoeForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "ibm/PowerMoE-3b")),
    "GraniteMoeHybridForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "ibm-granite/granite-4.0-tiny-preview")),  # noqa: E501
    "GraniteMoeSharedForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "ibm-research/moe-7b-1b-active-shared-experts")),  # noqa: E501
    "Grok1ModelForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "hpcai-tech/grok-1"),
Michael Goin's avatar
Michael Goin committed
218
                                             trust_remote_code=True),
219
    "HunYuanMoEV1ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "tencent/Hunyuan-A13B-Instruct"),
220
                                               trust_remote_code=True),
221
    # TODO: Remove is_available_online once their config.json is fixed
222
    "HunYuanDenseV1ForCausalLM":_HfExamplesInfo(os.path.join(models_path_prefix, "tencent/Hunyuan-7B-Instruct-0124"),
223
224
                                                trust_remote_code=True,
                                                is_available_online=False),
225
    "InternLMForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"internlm/internlm-chat-7b"),
226
                                           trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
227
    "InternLM2ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "internlm/internlm2-chat-7b"),
228
                                            trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
229
    "InternLM2VEForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "OpenGVLab/Mono-InternVL-2B"),
230
                                              trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
231
    "InternLM3ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "internlm/internlm3-8b-instruct"),
232
                                            trust_remote_code=True),
233
234
    "JAISLMHeadModel": _HfExamplesInfo(os.path.join(models_path_prefix, "inceptionai/jais-13b-chat")),
    "JambaForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "ai21labs/AI21-Jamba-1.5-Mini"),
235
                                        min_transformers_version="4.56.0",
236
                                        extras={
237
238
                                            "tiny": os.path.join(models_path_prefix, "ai21labs/Jamba-tiny-dev"),
                                            "random": os.path.join(models_path_prefix, "ai21labs/Jamba-tiny-random"),  # noqa: E501
239
                                        }),
240
    "Lfm2ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"LiquidAI/LFM2-1.2B"),
241
                                       min_transformers_version="4.54"),
242
    "LlamaForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"meta-llama/Llama-3.2-1B-Instruct"),
243
                                        extras={"guard": "meta-llama/Llama-Guard-3-1B",  # noqa: E501
244
245
                                                "hermes": "NousResearch/Hermes-3-Llama-3.1-8B", # noqa: E501
                                                "fp8": "RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8"}),  # noqa: E501
246
    "LLaMAForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"decapoda-research/llama-7b-hf"),
247
                                        is_available_online=False),
248
    "Llama4ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "meta-llama/Llama-4-Scout-17B-16E-Instruct"), # noqa: E501
249
                                         is_available_online=False),
250
251
252
253
    "MambaForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "state-spaces/mamba-130m-hf")),
    "Mamba2ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "mistralai/Mamba-Codestral-7B-v0.1")),
    "FalconMambaForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "tiiuae/falcon-mamba-7b-instruct")),  # noqa: E501
    "MiniCPMForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "openbmb/MiniCPM-2B-sft-bf16"),
254
                                         trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
255
    "MiniCPM3ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "openbmb/MiniCPM3-4B"),
256
                                         trust_remote_code=True),
257
    "MiniMaxForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "MiniMaxAI/MiniMax-Text-01-hf")),
zhuwenwen's avatar
zhuwenwen committed
258
    "MiniMaxText01ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "MiniMaxAI/MiniMax-Text-01"),
259
260
                                                trust_remote_code=True,
                                                revision="a59aa9cbc53b9fb8742ca4e9e1531b9802b6fdc3"),  # noqa: E501
261
    "MiniMaxM1ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "MiniMaxAI/MiniMax-M1-40k"),
262
                                            trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
263
264
    "MistralForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "mistralai/Mistral-7B-Instruct-v0.1")),
    "MixtralForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "mistralai/Mixtral-8x7B-Instruct-v0.1"),  # noqa: E501
zhuwenwen's avatar
zhuwenwen committed
265
                                          {"tiny": os.path.join(models_path_prefix, "TitanML/tiny-mixtral")}),  # noqa: E501
zhuwenwen's avatar
zhuwenwen committed
266
267
268
269
    "QuantMixtralForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "mistral-community/Mixtral-8x22B-v0.1-AWQ")),  # noqa: E501
    "MptForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "mpt"), is_available_online=False),
    "MPTForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "mosaicml/mpt-7b")),
    "NemotronForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "nvidia/Minitron-8B-Base")),
zhuwenwen's avatar
zhuwenwen committed
270
    "NemotronHForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "nvidia/Nemotron-H-8B-Base-8K"),
Luis Vega's avatar
Luis Vega committed
271
                                            trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
272
    "OlmoForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "allenai/OLMo-1B-hf")),
zhuwenwen's avatar
zhuwenwen committed
273
    "Olmo2ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "allenai/OLMo-2-0425-1B")),
zhuwenwen's avatar
zhuwenwen committed
274
275
276
277
    "OlmoeForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "allenai/OLMoE-1B-7B-0924-Instruct")),
    "OPTForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "facebook/opt-125m"),
                                      {"1b": os.path.join(models_path_prefix, "facebook/opt-iml-max-1.3b")}),
    "OrionForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "OrionStarAI/Orion-14B-Chat"),
278
                                        trust_remote_code=True),
279
280
281
282
    "PersimmonForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "adept/persimmon-8b-chat")),
    "PhiForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "microsoft/phi-2")),
    "Phi3ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "microsoft/Phi-3-mini-4k-instruct")),
    "Phi4FlashForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "microsoft/Phi-4-mini-flash-reasoning"), # noqa: E501
283
284
285
                                        trust_remote_code=True,
                                        v0_only=True,
                                        max_model_len=10240),
286
    "PhiMoEForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "microsoft/Phi-3.5-MoE-instruct"),
287
                                         trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
288
    "Plamo2ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "pfnet/plamo-2-1b"),
289
290
                                         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
291
                                        trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
292
    "QWenLMHeadModel": _HfExamplesInfo(os.path.join(models_path_prefix, "Qwen/Qwen-7B-Chat"),
293
294
                                       max_transformers_version="4.53",
                                       transformers_version_reason="HF model uses remote code that is not compatible with latest Transformers",  # noqa: E501
295
                                       trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
296
    "Qwen2ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"Qwen/Qwen2-0.5B-Instruct"),
297
                                        extras={"2.5": "Qwen/Qwen2.5-0.5B-Instruct"}), # noqa: E501
298
299
300
301
302
    "Qwen2MoeForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"Qwen/Qwen1.5-MoE-A2.7B-Chat")),
    "Qwen3ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"Qwen/Qwen3-8B")),
    "Qwen3MoeForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"Qwen/Qwen3-30B-A3B")),
    "RWForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"tiiuae/falcon-40b")),
    "SeedOssForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"ByteDance-Seed/Seed-OSS-36B-Instruct"), # noqa: E501
303
304
                                          trust_remote_code=True,
                                          is_available_online=False),
305
306
307
308
309
    "SmolLM3ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"HuggingFaceTB/SmolLM3-3B")),
    "StableLMEpochForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"stabilityai/stablelm-zephyr-3b")),  # noqa: E501
    "StableLmForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"stabilityai/stablelm-3b-4e1t")),
    "Starcoder2ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"bigcode/starcoder2-3b")),
    "Step3TextForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"stepfun-ai/step3"),
310
                                            trust_remote_code=True),
311
    "SolarForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"upstage/solar-pro-preview-instruct"),
312
                                        trust_remote_code=True),
313
    "TeleChat2ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "Tele-AI/TeleChat2-3B"),
314
                                            trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
315
    "TeleFLMForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "CofeAI/FLM-2-52B-Instruct-2407"),
316
                                            trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
317
    "XverseForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "xverse/XVERSE-7B-Chat"),
zhuwenwen's avatar
zhuwenwen committed
318
                                         tokenizer=os.path.join(models_path_prefix, "meta-llama/Llama-2-7b"),
319
                                         trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
320
    "Zamba2ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "Zyphra/Zamba2-7B-instruct")),
321
322
323
324
    "Ernie4_5_ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "baidu/ERNIE-4.5-0.3B-PT"),
                                        trust_remote_code=True),
    "Ernie4_5_MoeForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "baidu/ERNIE-4.5-21B-A3B-PT"),
                                        trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
325
    "MiMoForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "XiaomiMiMo/MiMo-7B-RL"),
326
                                        trust_remote_code=True),
327
    "Dots1ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"rednote-hilab/dots.llm1.inst")),
328
    # [Encoder-decoder]
329
330
331
    "BartModel": _HfExamplesInfo(os.path.join(models_path_prefix,"facebook/bart-base")),
    "BartForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix,"facebook/bart-large-cnn")),
    "MBartForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix,"facebook/mbart-large-en-ro"),  # noqa: E501
汪志鹏's avatar
汪志鹏 committed
332
                                                    hf_overrides={"architectures": ["MBartForConditionalGeneration"]}),  # noqa: E501
333
334
335
336
}

_EMBEDDING_EXAMPLE_MODELS = {
    # [Text-only]
337
338
    "BertModel": _HfExamplesInfo(os.path.join(models_path_prefix, "BAAI/bge-base-en-v1.5")),
    "Gemma2Model": _HfExamplesInfo(os.path.join(models_path_prefix, "BAAI/bge-multilingual-gemma2")),  # noqa: E501
339
340
    "GritLM": _HfExamplesInfo(os.path.join(models_path_prefix, "parasail-ai/GritLM-7B-vllm")),
    "GteModel": _HfExamplesInfo(os.path.join(models_path_prefix, "Snowflake/snowflake-arctic-embed-m-v2.0"),
341
                                               trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
342
    "GteNewModel": _HfExamplesInfo(os.path.join(models_path_prefix, "Alibaba-NLP/gte-base-en-v1.5"),
343
                                   trust_remote_code=True,
344
                                   hf_overrides={"architectures": ["GteNewModel"]}),  # noqa: E501
zhuwenwen's avatar
zhuwenwen committed
345
    "InternLM2ForRewardModel": _HfExamplesInfo(os.path.join(models_path_prefix, "internlm/internlm2-1_8b-reward"),
346
                                               trust_remote_code=True),
347
348
349
350
    "JambaForSequenceClassification": _HfExamplesInfo(os.path.join(models_path_prefix, "ai21labs/Jamba-tiny-reward-dev")),  # noqa: E501
    "LlamaModel": _HfExamplesInfo(os.path.join(models_path_prefix, "llama"), is_available_online=False),
    "MistralModel": _HfExamplesInfo(os.path.join(models_path_prefix, "intfloat/e5-mistral-7b-instruct")),
    "ModernBertModel": _HfExamplesInfo(os.path.join(models_path_prefix, "Alibaba-NLP/gte-modernbert-base"),
351
                                trust_remote_code=True),
352
    "NomicBertModel": _HfExamplesInfo(os.path.join(models_path_prefix, "nomic-ai/nomic-embed-text-v2-moe"),
353
                                               trust_remote_code=True),  # noqa: E501
354
355
    "Qwen2Model": _HfExamplesInfo(os.path.join(models_path_prefix, "ssmits/Qwen2-7B-Instruct-embed-base")),
    "Qwen2ForRewardModel": _HfExamplesInfo(os.path.join(models_path_prefix, "Qwen/Qwen2.5-Math-RM-72B"),
356
357
                                           max_transformers_version="4.53",
                                           transformers_version_reason="HF model uses remote code that is not compatible with latest Transformers"),  # noqa: E501
358
    "Qwen2ForProcessRewardModel": _HfExamplesInfo(os.path.join(models_path_prefix, "Qwen/Qwen2.5-Math-PRM-7B"),
359
360
                                                  max_transformers_version="4.53",
                                                  transformers_version_reason="HF model uses remote code that is not compatible with latest Transformers"),  # noqa: E501
361
362
363
    "RobertaModel": _HfExamplesInfo(os.path.join(models_path_prefix, "sentence-transformers/stsb-roberta-base-v2")),  # noqa: E501
    "RobertaForMaskedLM": _HfExamplesInfo(os.path.join(models_path_prefix, "sentence-transformers/all-roberta-large-v1")),  # noqa: E501
    "XLMRobertaModel": _HfExamplesInfo(os.path.join(models_path_prefix, "intfloat/multilingual-e5-small")),  # noqa: E501
364
    # [Multimodal]
zhuwenwen's avatar
zhuwenwen committed
365
366
    "LlavaNextForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "royokong/e5-v")),
    "Phi3VForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "TIGER-Lab/VLM2Vec-Full"),
367
                                         trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
368
    "Qwen2VLForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "MrLight/dse-qwen2-2b-mrl-v1")), # noqa: E501
zhuwenwen's avatar
zhuwenwen committed
369
    "PrithviGeoSpatialMAE": _HfExamplesInfo(os.path.join(models_path_prefix, "ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11"), # noqa: E501
370
                                            is_available_online=False),  # noqa: E501
371
372
}

373
374
_SEQUENCE_CLASSIFICATION_EXAMPLE_MODELS = {
    # [Decoder-only]
375
    "GPT2ForSequenceClassification": _HfExamplesInfo(os.path.join(models_path_prefix, "nie3e/sentiment-polish-gpt2-small")),  # noqa: E501
376
377

    # [Cross-encoder]
378
379
    "BertForSequenceClassification": _HfExamplesInfo(os.path.join(models_path_prefix, "cross-encoder/ms-marco-MiniLM-L-6-v2")),  # noqa: E501
    "GteNewForSequenceClassification": _HfExamplesInfo(os.path.join(models_path_prefix, "Alibaba-NLP/gte-multilingual-reranker-base"),  # noqa: E501
380
381
                                                       trust_remote_code=True,
                                                       hf_overrides={
382
383
384
385
                                                           "architectures": [os.path.join(models_path_prefix, "GteNewForSequenceClassification")]}),# noqa: E501
    "ModernBertForSequenceClassification": _HfExamplesInfo(os.path.join(models_path_prefix, "Alibaba-NLP/gte-reranker-modernbert-base")), # noqa: E501
    "RobertaForSequenceClassification": _HfExamplesInfo(os.path.join(models_path_prefix, "cross-encoder/quora-roberta-base")),  # noqa: E501
    "XLMRobertaForSequenceClassification": _HfExamplesInfo(os.path.join(models_path_prefix, "BAAI/bge-reranker-v2-m3")),  # noqa: E501
386
387
}

388
389
_AUTOMATIC_CONVERTED_MODELS = {
    # Use as_seq_cls_model for automatic conversion
390
    "GemmaForSequenceClassification": _HfExamplesInfo(os.path.join(models_path_prefix, "BAAI/bge-reranker-v2-gemma"),  # noqa: E501
391
392
393
                                                      hf_overrides={"architectures": ["GemmaForSequenceClassification"], # noqa: E501
                                                                    "classifier_from_token": ["Yes"],  # noqa: E501
                                                                    "method": "no_post_processing"}),  # noqa: E501
394
395
396
    "LlamaForSequenceClassification": _HfExamplesInfo(os.path.join(models_path_prefix, "Skywork/Skywork-Reward-V2-Llama-3.2-1B")),  # noqa: E501
    "Qwen2ForSequenceClassification": _HfExamplesInfo(os.path.join(models_path_prefix, "jason9693/Qwen2.5-1.5B-apeach")),  # noqa: E501
    "Qwen3ForSequenceClassification": _HfExamplesInfo(os.path.join(models_path_prefix, "tomaarsen/Qwen3-Reranker-0.6B-seq-cls")),  # noqa: E501
397
398
}

399
400
_MULTIMODAL_EXAMPLE_MODELS = {
    # [Decoder-only]
401
402
403
404
405
406
407
408
    "AriaForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "rhymes-ai/Aria")),
    "AyaVisionForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "CohereForAI/aya-vision-8b")), # noqa: E501
    "Blip2ForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "Salesforce/blip2-opt-2.7b"),  # noqa: E501
                                                     extras={"6b": os.path.join(models_path_prefix, "Salesforce/blip2-opt-6.7b")}),  # noqa: E501
    "ChameleonForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "facebook/chameleon-7b")),  # noqa: E501
    "Cohere2VisionForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "CohereLabs/command-a-vision-07-2025")), # noqa: E501
    "DeepseekVLV2ForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "deepseek-ai/deepseek-vl2-tiny"),  # noqa: E501
                                                extras={"fork": os.path.join(models_path_prefix, "Isotr0py/deepseek-vl2-tiny")},  # noqa: E501
409
410
                                                max_transformers_version="4.48",  # noqa: E501
                                                transformers_version_reason="HF model is not compatible.",  # noqa: E501
411
                                                hf_overrides={"architectures": ["DeepseekVLV2ForCausalLM"]}),  # noqa: E501
412
    "Emu3ForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "BAAI/Emu3-Chat-hf")),
413
    "Ernie4_5_VLMoeForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "baidu/ERNIE-4.5-VL-28B-A3B-PT"),  # noqa: E501
414
                                                              trust_remote_code=True),
415
416
417
    "FuyuForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "adept/fuyu-8b")),
    "Gemma3ForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "google/gemma-3-4b-it")),
    "Gemma3nForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "google/gemma-3n-E2B-it"),    # noqa: E501
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
418
                                        min_transformers_version="4.53"),
419
420
    "GraniteSpeechForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "ibm-granite/granite-speech-3.3-2b")),  # noqa: E501
    "GLM4VForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "zai-org/glm-4v-9b"),
421
422
                                        trust_remote_code=True,
                                        hf_overrides={"architectures": ["GLM4VForCausalLM"]}),  # noqa: E501
423
424
    "Glm4vForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "zai-org/GLM-4.1V-9B-Thinking")),  # noqa: E501
    "Glm4vMoeForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "zai-org/GLM-4.5V"),
425
                                                        min_transformers_version="4.56"),  # noqa: E501
426
    "H2OVLChatModel": _HfExamplesInfo(os.path.join(models_path_prefix, "h2oai/h2ovl-mississippi-800m"),
427
                                      trust_remote_code=True,
428
                                      extras={"2b": os.path.join(models_path_prefix, "h2oai/h2ovl-mississippi-2b")},  # noqa: E501
429
430
                                      max_transformers_version="4.48",  # noqa: E501
                                      transformers_version_reason="HF model is not compatible."),  # noqa: E501
431
    "HCXVisionForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix,"naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B"),  # noqa: E501
432
                                            trust_remote_code=True),
433
434
    "Idefics3ForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix,"HuggingFaceM4/Idefics3-8B-Llama3"),  # noqa: E501
                                                        {"tiny": os.path.join(models_path_prefix,"HuggingFaceTB/SmolVLM-256M-Instruct")},    # noqa: E501
435
436
                                                        min_transformers_version="4.56",
                                                        transformers_version_reason="HF model broken in 4.55"),  # noqa: E501
437
    "InternS1ForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix,"internlm/Intern-S1"),
438
                                                        trust_remote_code=True),  # noqa: E501
439
440
441
442
443
444
    "InternVLChatModel": _HfExamplesInfo(os.path.join(models_path_prefix,"OpenGVLab/InternVL2-1B"),
                                         extras={"2B": os.path.join(models_path_prefix,"OpenGVLab/InternVL2-2B"),
                                                 "3.0": os.path.join(models_path_prefix,"OpenGVLab/InternVL3-1B"),   # noqa: E501
                                                 "3.5-qwen3": os.path.join(models_path_prefix,"OpenGVLab/InternVL3_5-1B"),   # noqa: E501
                                                 "3.5-qwen3moe": os.path.join(models_path_prefix,"OpenGVLab/InternVL3_5-30B-A3B"),   # noqa: E501
                                                 "3.5-gptoss": os.path.join(models_path_prefix,"OpenGVLab/InternVL3_5-GPT-OSS-20B-A4B-Preview")},  # noqa: E501
Lyu Han's avatar
Lyu Han committed
445
                                         trust_remote_code=True),
446
447
    "InternVLForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix,"OpenGVLab/InternVL3-1B-hf")),    # noqa: E501
    "KeyeForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix,"Kwai-Keye/Keye-VL-8B-Preview"), # noqa: E501
448
                                                    trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
449
450
    "KimiVLForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "moonshotai/Kimi-VL-A3B-Instruct"),  # noqa: E501
                                                      extras={"thinking": os.path.join(models_path_prefix, "moonshotai/Kimi-VL-A3B-Thinking")},  # noqa: E501
451
                                                      trust_remote_code=True),
452
    "Llama4ForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "meta-llama/Llama-4-Scout-17B-16E-Instruct"),   # noqa: E501
453
                                                      max_model_len=10240,
454
                                                      extras={"llama-guard-4": os.path.join(models_path_prefix, "meta-llama/Llama-Guard-4-12B")}, # noqa: E501
455
                                                      ),
zhuwenwen's avatar
zhuwenwen committed
456
457
458
459
460
461
462
    "LlavaForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "llava-hf/llava-1.5-7b-hf"),
                                                     extras={"mistral": os.path.join(models_path_prefix, "mistral-community/pixtral-12b"), # noqa: E501
                                                             "mistral-fp8": os.path.join(models_path_prefix, "nm-testing/pixtral-12b-FP8-dynamic")}),  # noqa: E501
    "LlavaNextForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "llava-hf/llava-v1.6-mistral-7b-hf")),  # noqa: E501
    "LlavaNextVideoForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "llava-hf/LLaVA-NeXT-Video-7B-hf")),  # noqa: E501
    "LlavaOnevisionForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "llava-hf/llava-onevision-qwen2-0.5b-ov-hf")),  # noqa: E501
    "MantisForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "TIGER-Lab/Mantis-8B-siglip-llama3"),  # noqa: E501
463
464
                                                      max_transformers_version="4.48",  # noqa: E501
                                                      transformers_version_reason="HF model is not compatible.",  # noqa: E501
465
                                                      hf_overrides={"architectures": ["MantisForConditionalGeneration"]}),  # noqa: E501
zhuwenwen's avatar
zhuwenwen committed
466
    "MiniCPMO": _HfExamplesInfo(os.path.join(models_path_prefix, "openbmb/MiniCPM-o-2_6"),
467
                                trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
468
    "MiniCPMV": _HfExamplesInfo(os.path.join(models_path_prefix, "openbmb/MiniCPM-Llama3-V-2_5"),
tc-mb's avatar
tc-mb committed
469
                                extras={"2.6": "openbmb/MiniCPM-V-2_6", "4.0": "openbmb/MiniCPM-V-4", "4.5": "openbmb/MiniCPM-V-4_5"},  # noqa: E501
470
                                trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
471
    "MiniMaxVL01ForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "MiniMaxAI/MiniMax-VL-01"), # noqa: E501
472
473
                                              trust_remote_code=True,
                                              v0_only=True),
zhuwenwen's avatar
zhuwenwen committed
474
475
476
    "Mistral3ForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "mistralai/Mistral-Small-3.1-24B-Instruct-2503"),  # noqa: E501
                                                        extras={"fp8": os.path.join(models_path_prefix, "nm-testing/Mistral-Small-3.1-24B-Instruct-2503-FP8-dynamic")}),  # noqa: E501
    "MolmoForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "allenai/Molmo-7B-D-0924"),
477
                                        max_transformers_version="4.48",
478
                                        transformers_version_reason="Incorrectly-detected `tensorflow` import.",  # noqa: E501
zhuwenwen's avatar
zhuwenwen committed
479
                                        extras={"olmo": os.path.join(models_path_prefix, "allenai/Molmo-7B-O-0924")},  # noqa: E501
480
                                        trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
481
    "NVLM_D": _HfExamplesInfo(os.path.join(models_path_prefix, "nvidia/NVLM-D-72B"),
482
                              trust_remote_code=True),
483
484

    "Llama_Nemotron_Nano_VL" : _HfExamplesInfo(os.path.join(models_path_prefix, "nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1"), # noqa: E501
485
                                                     trust_remote_code=True),
486
    "Ovis": _HfExamplesInfo(os.path.join(models_path_prefix, "AIDC-AI/Ovis2-1B"), trust_remote_code=True,
487
488
                            max_transformers_version="4.53",
                            transformers_version_reason="HF model is not compatible",  # noqa: E501
489
490
                            extras={"1.6-llama": os.path.join(models_path_prefix, "AIDC-AI/Ovis1.6-Llama3.2-3B"),
                                    "1.6-gemma": os.path.join(models_path_prefix, "AIDC-AI/Ovis1.6-Gemma2-9B")}),  # noqa: E501
491
    "Ovis2_5": _HfExamplesInfo(os.path.join(models_path_prefix, "AIDC-AI/Ovis2.5-2B"),
492
                               trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
493
494
    "PaliGemmaForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "google/paligemma-3b-mix-224"),  # noqa: E501
                                                         extras={"v2": os.path.join(models_path_prefix, "google/paligemma2-3b-ft-docci-448")}),  # noqa: E501
495
    "Phi3VForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "microsoft/Phi-3-vision-128k-instruct"),
496
                                        trust_remote_code=True,
497
498
                                        max_transformers_version="4.48",
                                        transformers_version_reason="Use of deprecated imports which have been removed.",  # noqa: E501
499
500
                                        extras={"phi3.5": os.path.join(models_path_prefix, "microsoft/Phi-3.5-vision-instruct")}),  # noqa: E501
    "Phi4MMForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "microsoft/Phi-4-multimodal-instruct"),
501
                                        trust_remote_code=True),
502
    "Phi4MultimodalForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "microsoft/Phi-4-multimodal-instruct"),  # noqa: E501
503
                                                 revision="refs/pr/70"),
zhuwenwen's avatar
zhuwenwen committed
504
    "PixtralForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "mistralai/Pixtral-12B-2409"),  # noqa: E501
505
                                                       tokenizer_mode="mistral"),
zhuwenwen's avatar
zhuwenwen committed
506
507
    "QwenVLForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "Qwen/Qwen-VL"),
                                                      extras={"chat": os.path.join(models_path_prefix, "Qwen/Qwen-VL-Chat")},  # noqa: E501
508
509
                                                      trust_remote_code=True,
                                                      hf_overrides={"architectures": ["QwenVLForConditionalGeneration"]}),  # noqa: E501
zhuwenwen's avatar
zhuwenwen committed
510
511
    "Qwen2AudioForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "Qwen/Qwen2-Audio-7B-Instruct")),  # noqa: E501
    "Qwen2VLForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "Qwen/Qwen2-VL-2B-Instruct")),  # noqa: E501
512
    "Qwen2_5_VLForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "Qwen/Qwen2.5-VL-3B-Instruct"), # noqa: E501
513
                                                          max_model_len=4096),
zhuwenwen's avatar
zhuwenwen committed
514
515
    "Qwen2_5OmniModel": _HfExamplesInfo(os.path.join(models_path_prefix, "Qwen/Qwen2.5-Omni-3B")),
    "Qwen2_5OmniForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "Qwen/Qwen2.5-Omni-7B-AWQ")),  # noqa: E501
516
    "RForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "YannQi/R-4B"),
517
                                                 trust_remote_code=True),
518
    "SkyworkR1VChatModel": _HfExamplesInfo(os.path.join(models_path_prefix, "Skywork/Skywork-R1V-38B"),
519
                                           trust_remote_code=True),
520
    "SmolVLMForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "HuggingFaceTB/SmolVLM2-2.2B-Instruct"),  # noqa: E501
521
522
                                                       min_transformers_version="4.56",
                                                       transformers_version_reason="HF model broken in 4.55"),  # noqa: E501
zhuwenwen's avatar
zhuwenwen committed
523
    "Step3VLForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "stepfun-ai/step3"),
524
                                                        trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
525
    "UltravoxModel": _HfExamplesInfo(os.path.join(models_path_prefix, "fixie-ai/ultravox-v0_5-llama-3_2-1b"),  # noqa: E501
526
                                     trust_remote_code=True),
527
528
    "TarsierForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "omni-research/Tarsier-7b")),  # noqa: E501
    "Tarsier2ForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "omni-research/Tarsier2-Recap-7b"),  # noqa: E501
529
                                                        hf_overrides={"architectures": ["Tarsier2ForConditionalGeneration"]}),  # noqa: E501
530
    "VoxtralForConditionalGeneration": _HfExamplesInfo(
531
        os.path.join(models_path_prefix, "mistralai/Voxtral-Mini-3B-2507"),
532
533
534
535
        min_transformers_version="4.54",
        # disable this temporarily until we support HF format
        is_available_online=False,
    ),
536
    # [Encoder-decoder]
537
    "DonutForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "naver-clova-ix/donut-base-finetuned-docvqa"),  # noqa: E501
汪志鹏's avatar
汪志鹏 committed
538
                                                    hf_overrides={"architectures": ["DonutForConditionalGeneration"], "model_type": "donut"},  # noqa: E501
539
                                                    extras={"dolphin": os.path.join(models_path_prefix, "ByteDance/Dolphin")}),  # noqa: E501
540
541
    # Florence-2 uses BartFastTokenizer which can't be loaded from AutoTokenizer
    # Therefore, we borrow the BartTokenizer from the original Bart model
zhuwenwen's avatar
zhuwenwen committed
542
543
    "Florence2ForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix,"microsoft/Florence-2-base"),  # noqa: E501
                                                         tokenizer=os.path.join(models_path_prefix,"Isotr0py/Florence-2-tokenizer"),  # noqa: E501
544
                                                         trust_remote_code=True),  # noqa: E501
zhuwenwen's avatar
zhuwenwen committed
545
546
    "MllamaForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "meta-llama/Llama-3.2-11B-Vision-Instruct")),  # noqa: E501
    "WhisperForConditionalGeneration": _HfExamplesInfo(os.path.join(models_path_prefix, "openai/whisper-large-v3")),  # noqa: E501
547
    # [Cross-encoder]
548
    "JinaVLForRanking": _HfExamplesInfo(os.path.join(models_path_prefix, "jinaai/jina-reranker-m0")),   # noqa: E501
549
550
}

551

552
_SPECULATIVE_DECODING_EXAMPLE_MODELS = {
zhuwenwen's avatar
zhuwenwen committed
553
554
    "MedusaModel": _HfExamplesInfo(os.path.join(models_path_prefix, "JackFram/llama-68m"),
                                   speculative_model=os.path.join(models_path_prefix, "abhigoyal/vllm-medusa-llama-68m-random")),  # noqa: E501
555
556
557
558
    # 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
zhuwenwen's avatar
zhuwenwen committed
559
560
    "DeepSeekMTPModel": _HfExamplesInfo(os.path.join(models_path_prefix, "luccafong/deepseek_mtp_main_random"),
                                        speculative_model=os.path.join(models_path_prefix, "luccafong/deepseek_mtp_draft_random"),  # noqa: E501
561
                                        trust_remote_code=True),
562
    "EagleDeepSeekMTPModel": _HfExamplesInfo(os.path.join(models_path_prefix, "eagle618/deepseek-v3-random"),
563
564
                                        speculative_model="eagle618/eagle-deepseek-v3-random",  # noqa: E501
                                        trust_remote_code=True),
zhuwenwen's avatar
zhuwenwen committed
565
    "EagleLlamaForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "yuhuili/EAGLE-LLaMA3-Instruct-8B"),
566
                                             trust_remote_code=True,
zhuwenwen's avatar
zhuwenwen committed
567
568
                                             speculative_model=os.path.join(models_path_prefix, "yuhuili/EAGLE-LLaMA3-Instruct-8B"),
                                             tokenizer=os.path.join(models_path_prefix, "meta-llama/Meta-Llama-3-8B-Instruct")),  # noqa: E501
569
    "Eagle3LlamaForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "yuhuili/EAGLE3-LLaMA3.1-Instruct-8B"),  # noqa: E501
570
                                            trust_remote_code=True,
571
572
                                            speculative_model=os.path.join(models_path_prefix, "yuhuili/EAGLE3-LLaMA3.1-Instruct-8B"),
                                            tokenizer=os.path.join(models_path_prefix, "meta-llama/Llama-3.1-8B-Instruct")),
573
574
575
576
577
    # 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
578
    "EagleLlama4ForCausalLM": _HfExamplesInfo(
579
        os.path.join(models_path_prefix, "morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct"),
zhiweiz's avatar
zhiweiz committed
580
        trust_remote_code=True,
581
582
583
        speculative_model=os.path.join(models_path_prefix, "morgendave/EAGLE-Llama-4-Scout-17B-16E-Instruct"),
        tokenizer=os.path.join(models_path_prefix, "meta-llama/Llama-4-Scout-17B-16E-Instruct")),  # noqa: E501
    "EagleMiniCPMForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "openbmb/MiniCPM-1B-sft-bf16"),
584
585
                                            trust_remote_code=True,
                                            is_available_online=False,
586
587
                                            speculative_model=os.path.join(models_path_prefix, "openbmb/MiniCPM-2B-sft-bf16"),
                                            tokenizer=os.path.join(models_path_prefix, "openbmb/MiniCPM-2B-sft-bf16")),
588
    "ErnieMTPModel": _HfExamplesInfo(os.path.join(models_path_prefix, "baidu/ERNIE-4.5-21B-A3B-PT"),
589
                                    trust_remote_code=True,
590
                                    speculative_model=os.path.join(models_path_prefix, "baidu/ERNIE-4.5-21B-A3B-PT")),
591
592
    "Glm4MoeMTPModel": _HfExamplesInfo(os.path.join(models_path_prefix, "zai-org/GLM-4.5"),
                                        speculative_model=os.path.join(models_path_prefix, "zai-org/GLM-4.5"),
zhuwenwen's avatar
zhuwenwen committed
593
594
                                        min_transformers_version="4.54",
                                        is_available_online=False),
595
    "MiMoMTPModel": _HfExamplesInfo(os.path.join(models_path_prefix, "XiaomiMiMo/MiMo-7B-RL"),
596
                                    trust_remote_code=True,
597
                                    speculative_model=os.path.join(models_path_prefix, "XiaomiMiMo/MiMo-7B-RL"))
598
599
}

600
_TRANSFORMERS_BACKEND_MODELS = {
601
    "TransformersModel": _HfExamplesInfo(os.path.join(models_path_prefix, "Qwen/Qwen3-Embedding-0.6B")),
602
    "TransformersForCausalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "hmellor/Ilama-3.2-1B"), trust_remote_code=True),  # noqa: E501
603
    "TransformersForMultimodalLM": _HfExamplesInfo(os.path.join(models_path_prefix, "BAAI/Emu3-Chat-hf")),
604
605
}

606
607
608
_EXAMPLE_MODELS = {
    **_TEXT_GENERATION_EXAMPLE_MODELS,
    **_EMBEDDING_EXAMPLE_MODELS,
609
    **_SEQUENCE_CLASSIFICATION_EXAMPLE_MODELS,
610
611
    **_MULTIMODAL_EXAMPLE_MODELS,
    **_SPECULATIVE_DECODING_EXAMPLE_MODELS,
612
    **_TRANSFORMERS_BACKEND_MODELS,
613
614
615
616
617
618
619
620
621
}


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

        self.hf_models = hf_models

622
    def get_supported_archs(self) -> Set[str]:
623
624
625
626
627
        return self.hf_models.keys()

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

628
629
630
631
632
    def find_hf_info(self, model_id: str) -> _HfExamplesInfo:
        for info in self.hf_models.values():
            if info.default == model_id:
                return info

633
634
635
636
637
        # Fallback to extras
        for info in self.hf_models.values():
            if any(extra == model_id for extra in info.extras.values()):
                return info

638
639
        raise ValueError(f"No example model defined for {model_id}")

640

Patrick von Platen's avatar
Patrick von Platen committed
641
HF_EXAMPLE_MODELS = HfExampleModels(_EXAMPLE_MODELS)
642
AUTO_EXAMPLE_MODELS = HfExampleModels(_AUTOMATIC_CONVERTED_MODELS)