registry.py 44.7 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
103
104
105
    use_original_num_layers: bool = False
    """
    If True, use the original number of layers from the model config 
    instead of minimal layers for testing.
    """

106
107
108
    def check_transformers_version(
        self,
        *,
109
        on_fail: Literal["error", "skip", "return"],
110
111
        check_min_version: bool = True,
        check_max_version: bool = True,
112
    ) -> Optional[str]:
113
114
115
116
        """
        If the installed transformers version does not meet the requirements,
        perform the given action.
        """
117
118
119
120
        if (
            self.min_transformers_version is None
            and self.max_transformers_version is None
        ):
121
            return None
122
123

        current_version = TRANSFORMERS_VERSION
124
        cur_base_version = Version(current_version).base_version
125
126
127
        min_version = self.min_transformers_version
        max_version = self.max_transformers_version
        msg = f"`transformers=={current_version}` installed, but `transformers"
128
129
        # Only check the base version for the min/max version, otherwise preview
        # models cannot be run because `x.yy.0.dev0`<`x.yy.0`
130
131
132
133
134
        if (
            check_min_version
            and min_version
            and Version(cur_base_version) < Version(min_version)
        ):
135
            msg += f">={min_version}` is required to run this model."
136
137
138
139
140
        elif (
            check_max_version
            and max_version
            and Version(cur_base_version) > Version(max_version)
        ):
141
142
            msg += f"<={max_version}` is required to run this model."
        else:
143
            return None
144

145
146
147
148
149
        if self.transformers_version_reason:
            msg += f" Reason: {self.transformers_version_reason}"

        if on_fail == "error":
            raise RuntimeError(msg)
150
        elif on_fail == "skip":
151
            pytest.skip(msg)
152

153
154
        return msg

155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
    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)

171
172
173
174

# yapf: disable
_TEXT_GENERATION_EXAMPLE_MODELS = {
    # [Decoder-only]
175
    "ApertusForCausalLM": _HfExamplesInfo("swiss-ai/Apertus-8B-2509",
176
177
                                          min_transformers_version="4.56.0",
                                          trust_remote_code=True),
178
179
180
181
    "AquilaModel": _HfExamplesInfo("BAAI/AquilaChat-7B",
                                   trust_remote_code=True),
    "AquilaForCausalLM": _HfExamplesInfo("BAAI/AquilaChat2-7B",
                                         trust_remote_code=True),
182
    "ArceeForCausalLM": _HfExamplesInfo("arcee-ai/AFM-4.5B-Base"),
183
184
185
186
187
188
    "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),
189
190
    "BailingMoeForCausalLM": _HfExamplesInfo("inclusionAI/Ling-lite-1.5",
                                         trust_remote_code=True),
ant-yy's avatar
ant-yy committed
191
192
    "BailingMoeV2ForCausalLM": _HfExamplesInfo("inclusionAI/Ling-mini-2.0",
                                         trust_remote_code=True),
193
    "BambaForCausalLM": _HfExamplesInfo("ibm-ai-platform/Bamba-9B-v1",
194
                                        min_transformers_version="4.55.3",
195
                                        extras={"tiny": "hmellor/tiny-random-BambaForCausalLM"}),  # noqa: E501
196
197
    "BloomForCausalLM": _HfExamplesInfo("bigscience/bloom-560m",
                                        {"1b": "bigscience/bloomz-1b1"}),
198
    "ChatGLMModel": _HfExamplesInfo("zai-org/chatglm3-6b",
199
                                    trust_remote_code=True,
200
                                    max_transformers_version="4.48"),
201
202
    "ChatGLMForConditionalGeneration": _HfExamplesInfo("thu-coai/ShieldLM-6B-chatglm3",  # noqa: E501
                                                       trust_remote_code=True),
203
204
    "CohereForCausalLM": _HfExamplesInfo("CohereForAI/c4ai-command-r-v01",
                                         trust_remote_code=True),
205
206
    "Cohere2ForCausalLM": _HfExamplesInfo("CohereForAI/c4ai-command-r7b-12-2024", # noqa: E501
                                         trust_remote_code=True),
207
208
209
    "CwmForCausalLM": _HfExamplesInfo("facebook/cwm", # noqa: E501
                                      trust_remote_code=True,
                                      is_available_online=False),
210
    "DbrxForCausalLM": _HfExamplesInfo("databricks/dbrx-instruct"),
211
    "DeciLMForCausalLM": _HfExamplesInfo("nvidia/Llama-3_3-Nemotron-Super-49B-v1", # noqa: E501
212
213
214
215
                                         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
216
217
    "DeepseekV3ForCausalLM": _HfExamplesInfo("deepseek-ai/DeepSeek-V3",  # noqa: E501
                                         trust_remote_code=True),
218
    "DeepseekV32ForCausalLM": _HfExamplesInfo("deepseek-ai/DeepSeek-V3.2-Exp"),
219
    "Ernie4_5ForCausalLM": _HfExamplesInfo("baidu/ERNIE-4.5-0.3B-PT",
220
                                            min_transformers_version="4.54"),
221
    "Ernie4_5_MoeForCausalLM": _HfExamplesInfo("baidu/ERNIE-4.5-21B-A3B-PT",
222
                                               min_transformers_version="4.54"),
223
224
225
226
    "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"),
227
    "Fairseq2LlamaForCausalLM": _HfExamplesInfo("mgleize/fairseq2-dummy-Llama-3.2-1B"),  # noqa: E501
228
    "FalconForCausalLM": _HfExamplesInfo("tiiuae/falcon-7b"),
229
    "FalconH1ForCausalLM":_HfExamplesInfo("tiiuae/Falcon-H1-0.5B-Base"),
230
    "GemmaForCausalLM": _HfExamplesInfo("google/gemma-1.1-2b-it"),
231
    "Gemma2ForCausalLM": _HfExamplesInfo("google/gemma-2-9b"),
232
    "Gemma3ForCausalLM": _HfExamplesInfo("google/gemma-3-1b-it"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
233
    "Gemma3nForCausalLM": _HfExamplesInfo("google/gemma-3n-E2B-it",
Robert Shaw's avatar
Robert Shaw committed
234
                                          min_transformers_version="4.53"),
235
236
    "GlmForCausalLM": _HfExamplesInfo("zai-org/glm-4-9b-chat-hf"),
    "Glm4ForCausalLM": _HfExamplesInfo("zai-org/GLM-4-9B-0414"),
237
238
    "Glm4MoeForCausalLM": _HfExamplesInfo("zai-org/GLM-4.5",
                                          min_transformers_version="4.54"),   # noqa: E501
239
240
241
    "GPT2LMHeadModel": _HfExamplesInfo("openai-community/gpt2",
                                       {"alias": "gpt2"}),
    "GPTBigCodeForCausalLM": _HfExamplesInfo("bigcode/starcoder",
242
                                             extras={"tiny": "bigcode/tiny_starcoder_py"},  # noqa: E501
243
244
                                             min_transformers_version="4.55.1",
                                             transformers_version_reason="HF model broken in 4.55.0"),  # noqa: E501
245
246
247
248
    "GPTJForCausalLM": _HfExamplesInfo("Milos/slovak-gpt-j-405M",
                                       {"6b": "EleutherAI/gpt-j-6b"}),
    "GPTNeoXForCausalLM": _HfExamplesInfo("EleutherAI/pythia-70m",
                                          {"1b": "EleutherAI/pythia-1.4b"}),
249
    "GptOssForCausalLM": _HfExamplesInfo("lmsys/gpt-oss-20b-bf16"),
250
251
    "GraniteForCausalLM": _HfExamplesInfo("ibm/PowerLM-3b"),
    "GraniteMoeForCausalLM": _HfExamplesInfo("ibm/PowerMoE-3b"),
252
253
    "GraniteMoeHybridForCausalLM": _HfExamplesInfo("ibm-granite/granite-4.0-tiny-preview", # noqa: E501
                                                   min_transformers_version="4.55.3"),
254
    "GraniteMoeSharedForCausalLM": _HfExamplesInfo("ibm-research/moe-7b-1b-active-shared-experts"),  # noqa: E501
Michael Goin's avatar
Michael Goin committed
255
256
    "Grok1ModelForCausalLM": _HfExamplesInfo("hpcai-tech/grok-1",
                                             trust_remote_code=True),
257
258
    "HunYuanMoEV1ForCausalLM": _HfExamplesInfo("tencent/Hunyuan-A13B-Instruct",
                                               trust_remote_code=True),
259
    # TODO: Remove is_available_online once their config.json is fixed
260
    "HunYuanDenseV1ForCausalLM":_HfExamplesInfo("tencent/Hunyuan-7B-Instruct-0124",
261
262
                                                trust_remote_code=True,
                                                is_available_online=False),
263
264
265
266
267
268
    "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),
269
270
    "InternLM3ForCausalLM": _HfExamplesInfo("internlm/internlm3-8b-instruct",
                                            trust_remote_code=True),
271
    "JAISLMHeadModel": _HfExamplesInfo("inceptionai/jais-13b-chat"),
272
    "JambaForCausalLM": _HfExamplesInfo("ai21labs/AI21-Jamba-1.5-Mini",
273
                                        min_transformers_version="4.55.3",
274
275
276
277
                                        extras={
                                            "tiny": "ai21labs/Jamba-tiny-dev",
                                            "random": "ai21labs/Jamba-tiny-random",  # noqa: E501
                                        }),
278
279
    "Lfm2ForCausalLM": _HfExamplesInfo("LiquidAI/LFM2-1.2B",
                                       min_transformers_version="4.54"),
280
281
    "LlamaForCausalLM": _HfExamplesInfo("meta-llama/Llama-3.2-1B-Instruct",
                                        extras={"guard": "meta-llama/Llama-Guard-3-1B",  # noqa: E501
282
283
                                                "hermes": "NousResearch/Hermes-3-Llama-3.1-8B", # noqa: E501
                                                "fp8": "RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8"}),  # noqa: E501
284
285
    "LLaMAForCausalLM": _HfExamplesInfo("decapoda-research/llama-7b-hf",
                                        is_available_online=False),
286
287
    "Llama4ForCausalLM": _HfExamplesInfo("meta-llama/Llama-4-Scout-17B-16E-Instruct", # noqa: E501
                                         is_available_online=False),
XuruiYang's avatar
XuruiYang committed
288
289
    "LongcatFlashForCausalLM": _HfExamplesInfo
                ("meituan-longcat/LongCat-Flash-Chat", trust_remote_code=True),
290
    "MambaForCausalLM": _HfExamplesInfo("state-spaces/mamba-130m-hf"),
291
292
293
294
295
    "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
                                         }),
296
297
298
299
300
    "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),
301
    "MiniMaxForCausalLM": _HfExamplesInfo("MiniMaxAI/MiniMax-Text-01-hf"),
302
    "MiniMaxText01ForCausalLM": _HfExamplesInfo("MiniMaxAI/MiniMax-Text-01",
303
304
                                                trust_remote_code=True,
                                                revision="a59aa9cbc53b9fb8742ca4e9e1531b9802b6fdc3"),  # noqa: E501
305
306
    "MiniMaxM1ForCausalLM": _HfExamplesInfo("MiniMaxAI/MiniMax-M1-40k",
                                            trust_remote_code=True),
307
    "MistralForCausalLM": _HfExamplesInfo("mistralai/Mistral-7B-Instruct-v0.1"),
308
    "MixtralForCausalLM": _HfExamplesInfo("mistralai/Mixtral-8x7B-Instruct-v0.1",  # noqa: E501
309
                                          {"tiny": "TitanML/tiny-mixtral"}),  # noqa: E501
310
311
312
    "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
313
314
    "NemotronHForCausalLM": _HfExamplesInfo("nvidia/Nemotron-H-8B-Base-8K",
                                            trust_remote_code=True),
315
    "OlmoForCausalLM": _HfExamplesInfo("allenai/OLMo-1B-hf"),
316
    "Olmo2ForCausalLM": _HfExamplesInfo("allenai/OLMo-2-0425-1B"),
317
    "Olmo3ForCausalLM": _HfExamplesInfo("shanearora/2025-sep-a-base-model"),
318
    "OlmoeForCausalLM": _HfExamplesInfo("allenai/OLMoE-1B-7B-0924-Instruct"),
319
320
    "OPTForCausalLM": _HfExamplesInfo("facebook/opt-125m",
                                      {"1b": "facebook/opt-iml-max-1.3b"}),
321
322
323
    "OrionForCausalLM": _HfExamplesInfo("OrionStarAI/Orion-14B-Chat",
                                        trust_remote_code=True),
    "PersimmonForCausalLM": _HfExamplesInfo("adept/persimmon-8b-chat"),
324
    "PhiForCausalLM": _HfExamplesInfo("microsoft/phi-2"),
325
326
327
    "Phi3ForCausalLM": _HfExamplesInfo("microsoft/Phi-3-mini-4k-instruct"),
    "PhiMoEForCausalLM": _HfExamplesInfo("microsoft/Phi-3.5-MoE-instruct",
                                         trust_remote_code=True),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
328
    "Plamo2ForCausalLM": _HfExamplesInfo("pfnet/plamo-2-1b",
329
330
331
                                         max_transformers_version="4.55.4",
                                         transformers_version_reason="HF model uses remote code that is not compatible with latest Transformers",  # noqa: E501
                                         trust_remote_code=True),
332
    "QWenLMHeadModel": _HfExamplesInfo("Qwen/Qwen-7B-Chat",
333
334
                                       max_transformers_version="4.53",
                                       transformers_version_reason="HF model uses remote code that is not compatible with latest Transformers",  # noqa: E501
335
                                       trust_remote_code=True),
336
337
    "Qwen2ForCausalLM": _HfExamplesInfo("Qwen/Qwen2-0.5B-Instruct",
                                        extras={"2.5": "Qwen/Qwen2.5-0.5B-Instruct"}), # noqa: E501
338
    "Qwen2MoeForCausalLM": _HfExamplesInfo("Qwen/Qwen1.5-MoE-A2.7B-Chat"),
Jee Jee Li's avatar
Jee Jee Li committed
339
340
    "Qwen3ForCausalLM": _HfExamplesInfo("Qwen/Qwen3-8B"),
    "Qwen3MoeForCausalLM": _HfExamplesInfo("Qwen/Qwen3-30B-A3B"),
341
    "Qwen3NextForCausalLM": _HfExamplesInfo("Qwen/Qwen3-Next-80B-A3B-Instruct",
342
343
                                            extras={"tiny-random": "tiny-random/qwen3-next-moe"}, # noqa: E501
                                            min_transformers_version="4.56.3"),
344
    "RWForCausalLM": _HfExamplesInfo("tiiuae/falcon-40b"),
345
346
347
    "SeedOssForCausalLM": _HfExamplesInfo("ByteDance-Seed/Seed-OSS-36B-Instruct", # noqa: E501
                                          trust_remote_code=True,
                                          is_available_online=False),
348
    "SmolLM3ForCausalLM": _HfExamplesInfo("HuggingFaceTB/SmolLM3-3B"),
349
350
    "StableLMEpochForCausalLM": _HfExamplesInfo("stabilityai/stablelm-zephyr-3b"),  # noqa: E501
    "StableLmForCausalLM": _HfExamplesInfo("stabilityai/stablelm-3b-4e1t"),
351
    "Starcoder2ForCausalLM": _HfExamplesInfo("bigcode/starcoder2-3b"),
Song's avatar
Song committed
352
    "Step3TextForCausalLM": _HfExamplesInfo("stepfun-ai/step3",
353
                                            trust_remote_code=True),
354
355
    "SolarForCausalLM": _HfExamplesInfo("upstage/solar-pro-preview-instruct",
                                        trust_remote_code=True),
356
357
    "TeleChat2ForCausalLM": _HfExamplesInfo("Tele-AI/TeleChat2-3B",
                                            trust_remote_code=True),
358
359
    "TeleFLMForCausalLM": _HfExamplesInfo("CofeAI/FLM-2-52B-Instruct-2407",
                                            trust_remote_code=True),
360
    "XverseForCausalLM": _HfExamplesInfo("xverse/XVERSE-7B-Chat",
361
                                         tokenizer="meta-llama/Llama-2-7b",
362
                                         trust_remote_code=True),
363
    "Zamba2ForCausalLM": _HfExamplesInfo("Zyphra/Zamba2-7B-instruct"),
364
365
    "MiMoForCausalLM": _HfExamplesInfo("XiaomiMiMo/MiMo-7B-RL",
                                        trust_remote_code=True),
366
    "Dots1ForCausalLM": _HfExamplesInfo("rednote-hilab/dots.llm1.inst"),
367
368
369
370
}

_EMBEDDING_EXAMPLE_MODELS = {
    # [Text-only]
371
372
    "BertModel": _HfExamplesInfo("BAAI/bge-base-en-v1.5"),
    "Gemma2Model": _HfExamplesInfo("BAAI/bge-multilingual-gemma2"),  # noqa: E501
373
    "Gemma3TextModel": _HfExamplesInfo("google/embeddinggemma-300m"),
374
    "GritLM": _HfExamplesInfo("parasail-ai/GritLM-7B-vllm"),
375
376
    "GteModel": _HfExamplesInfo("Snowflake/snowflake-arctic-embed-m-v2.0",
                                               trust_remote_code=True),
377
378
    "GteNewModel": _HfExamplesInfo("Alibaba-NLP/gte-base-en-v1.5",
                                   trust_remote_code=True,
379
                                   hf_overrides={"architectures": ["GteNewModel"]}),  # noqa: E501
380
381
    "InternLM2ForRewardModel": _HfExamplesInfo("internlm/internlm2-1_8b-reward",
                                               trust_remote_code=True),
382
    "JambaForSequenceClassification": _HfExamplesInfo("ai21labs/Jamba-tiny-reward-dev"),  # noqa: E501
383
    "LlamaModel": _HfExamplesInfo("llama", is_available_online=False),
384
    "MistralModel": _HfExamplesInfo("intfloat/e5-mistral-7b-instruct"),
385
    "ModernBertModel": _HfExamplesInfo("Alibaba-NLP/gte-modernbert-base",
386
                                trust_remote_code=True),
387
    "NomicBertModel": _HfExamplesInfo("nomic-ai/nomic-embed-text-v2-moe",
388
                                               trust_remote_code=True),  # noqa: E501
389
    "Qwen2Model": _HfExamplesInfo("ssmits/Qwen2-7B-Instruct-embed-base"),
390
391
392
393
394
395
    "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
396
397
398
    "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
399
    # [Multimodal]
400
    "CLIPModel": _HfExamplesInfo("openai/clip-vit-base-patch32"),
401
402
403
404
    "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
405
    "PrithviGeoSpatialMAE": _HfExamplesInfo("ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11", # noqa: E501
406
407
408
409
410
411
412
                                            dtype=torch.float16,
                                            enforce_eager=True,
                                            skip_tokenizer_init=True,
                                            # This is to avoid the model
                                            # going OOM in CI
                                            max_num_seqs=32,
                                            ),
413
    "Terratorch": _HfExamplesInfo("ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11", # noqa: E501
414
415
416
417
418
419
                                  dtype=torch.float16,
                                  enforce_eager=True,
                                  skip_tokenizer_init=True,
                                  # This is to avoid the model going OOM in CI
                                  max_num_seqs=32,
                                  ),
420
421
}

422
423
424
425
426
_SEQUENCE_CLASSIFICATION_EXAMPLE_MODELS = {
    # [Decoder-only]
    "GPT2ForSequenceClassification": _HfExamplesInfo("nie3e/sentiment-polish-gpt2-small"),  # noqa: E501

    # [Cross-encoder]
427
    "BertForSequenceClassification": _HfExamplesInfo("cross-encoder/ms-marco-MiniLM-L-6-v2"),  # noqa: E501
428
    "BertForTokenClassification": _HfExamplesInfo("boltuix/NeuroBERT-NER"),
429
430
431
432
    "GteNewForSequenceClassification": _HfExamplesInfo("Alibaba-NLP/gte-multilingual-reranker-base",  # noqa: E501
                                                       trust_remote_code=True,
                                                       hf_overrides={
                                                           "architectures": ["GteNewForSequenceClassification"]}),# noqa: E501
433
434
435
    "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
436
437
}

438
439
440
441
442
443
444
445
446
447
448
_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
}

449
450
_MULTIMODAL_EXAMPLE_MODELS = {
    # [Decoder-only]
451
    "AriaForConditionalGeneration": _HfExamplesInfo("rhymes-ai/Aria"),
Jennifer Zhao's avatar
Jennifer Zhao committed
452
    "AyaVisionForConditionalGeneration": _HfExamplesInfo("CohereForAI/aya-vision-8b"), # noqa: E501
453
    "Blip2ForConditionalGeneration": _HfExamplesInfo("Salesforce/blip2-opt-2.7b",  # noqa: E501
454
                                                     extras={"6b": "Salesforce/blip2-opt-6.7b"}),  # noqa: E501
455
    "ChameleonForConditionalGeneration": _HfExamplesInfo("facebook/chameleon-7b"),  # noqa: E501
456
    "Cohere2VisionForConditionalGeneration": _HfExamplesInfo("CohereLabs/command-a-vision-07-2025"), # noqa: E501
457
    "DeepseekVLV2ForCausalLM": _HfExamplesInfo("deepseek-ai/deepseek-vl2-tiny",  # noqa: E501
458
459
460
                                                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
461
                                                hf_overrides={"architectures": ["DeepseekVLV2ForCausalLM"]}),  # noqa: E501
Roger Wang's avatar
Roger Wang committed
462
463
    "DotsOCRForCausalLM": _HfExamplesInfo("rednote-hilab/dots.ocr",
                                          trust_remote_code=True),
464
    "Emu3ForConditionalGeneration": _HfExamplesInfo("BAAI/Emu3-Chat-hf"),
465
466
    "Ernie4_5_VLMoeForConditionalGeneration": _HfExamplesInfo("baidu/ERNIE-4.5-VL-28B-A3B-PT",  # noqa: E501
                                                              trust_remote_code=True),
467
    "FuyuForCausalLM": _HfExamplesInfo("adept/fuyu-8b"),
468
    "Gemma3ForConditionalGeneration": _HfExamplesInfo("google/gemma-3-4b-it"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
469
470
    "Gemma3nForConditionalGeneration": _HfExamplesInfo("google/gemma-3n-E2B-it",    # noqa: E501
                                        min_transformers_version="4.53"),
471
    "GraniteSpeechForConditionalGeneration": _HfExamplesInfo("ibm-granite/granite-speech-3.3-2b"),  # noqa: E501
472
    "GLM4VForCausalLM": _HfExamplesInfo("zai-org/glm-4v-9b",
473
474
                                        trust_remote_code=True,
                                        hf_overrides={"architectures": ["GLM4VForCausalLM"]}),  # noqa: E501
475
    "Glm4vForConditionalGeneration": _HfExamplesInfo("zai-org/GLM-4.1V-9B-Thinking"),  # noqa: E501
476
    "Glm4vMoeForConditionalGeneration": _HfExamplesInfo("zai-org/GLM-4.5V",
477
                                                        min_transformers_version="4.56"),  # noqa: E501
478
    "H2OVLChatModel": _HfExamplesInfo("h2oai/h2ovl-mississippi-800m",
479
                                      trust_remote_code=True,
480
481
482
                                      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
483
484
    "HCXVisionForCausalLM": _HfExamplesInfo("naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B",  # noqa: E501
                                            trust_remote_code=True),
485
486
    "Idefics3ForConditionalGeneration": _HfExamplesInfo("HuggingFaceM4/Idefics3-8B-Llama3",  # noqa: E501
                                                        {"tiny": "HuggingFaceTB/SmolVLM-256M-Instruct"},    # noqa: E501
487
488
                                                        min_transformers_version="4.56",
                                                        transformers_version_reason="HF model broken in 4.55"),  # noqa: E501
489
490
    "InternS1ForConditionalGeneration": _HfExamplesInfo("internlm/Intern-S1",
                                                        trust_remote_code=True),  # noqa: E501
491
    "InternVLChatModel": _HfExamplesInfo("OpenGVLab/InternVL2-1B",
492
                                         extras={"2B": "OpenGVLab/InternVL2-2B",
493
494
495
496
                                                 "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
497
                                         trust_remote_code=True),
498
    "InternVLForConditionalGeneration": _HfExamplesInfo("OpenGVLab/InternVL3-1B-hf"),    # noqa: E501
499
500
    "KeyeForConditionalGeneration": _HfExamplesInfo("Kwai-Keye/Keye-VL-8B-Preview", # noqa: E501
                                                    trust_remote_code=True),
501
502
    "KeyeVL1_5ForConditionalGeneration": _HfExamplesInfo("Kwai-Keye/Keye-VL-1_5-8B", # noqa: E501
                                                         trust_remote_code=True),
503
504
    "KimiVLForConditionalGeneration": _HfExamplesInfo("moonshotai/Kimi-VL-A3B-Instruct",  # noqa: E501
                                                      extras={"thinking": "moonshotai/Kimi-VL-A3B-Thinking"},  # noqa: E501
505
                                                      trust_remote_code=True),
506
    "Llama4ForConditionalGeneration": _HfExamplesInfo("meta-llama/Llama-4-Scout-17B-16E-Instruct",   # noqa: E501
507
                                                      max_model_len=10240,
508
                                                      extras={"llama-guard-4": "meta-llama/Llama-Guard-4-12B"},  # noqa: E501
509
                                                      ),
510
    "LlavaForConditionalGeneration": _HfExamplesInfo("llava-hf/llava-1.5-7b-hf",
511
512
                                                     extras={"mistral": "mistral-community/pixtral-12b", # noqa: E501
                                                             "mistral-fp8": "nm-testing/pixtral-12b-FP8-dynamic"}),  # noqa: E501
513
514
515
    "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
516
    "MantisForConditionalGeneration": _HfExamplesInfo("TIGER-Lab/Mantis-8B-siglip-llama3",  # noqa: E501
517
518
                                                      max_transformers_version="4.48",  # noqa: E501
                                                      transformers_version_reason="HF model is not compatible.",  # noqa: E501
519
                                                      hf_overrides={"architectures": ["MantisForConditionalGeneration"]}),  # noqa: E501
520
521
    "MiDashengLMModel": _HfExamplesInfo("mispeech/midashenglm-7b",
                            trust_remote_code=True),
522
523
    "MiniCPMO": _HfExamplesInfo("openbmb/MiniCPM-o-2_6",
                                trust_remote_code=True),
524
    "MiniCPMV": _HfExamplesInfo("openbmb/MiniCPM-Llama3-V-2_5",
tc-mb's avatar
tc-mb committed
525
                                extras={"2.6": "openbmb/MiniCPM-V-2_6", "4.0": "openbmb/MiniCPM-V-4", "4.5": "openbmb/MiniCPM-V-4_5"},  # noqa: E501
526
                                trust_remote_code=True),
527
    "MiniMaxVL01ForConditionalGeneration": _HfExamplesInfo("MiniMaxAI/MiniMax-VL-01", # noqa: E501
528
529
                                              trust_remote_code=True,
                                              v0_only=True),
530
531
    "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
532
    "MolmoForCausalLM": _HfExamplesInfo("allenai/Molmo-7B-D-0924",
533
                                        max_transformers_version="4.48",
534
                                        transformers_version_reason="Incorrectly-detected `tensorflow` import.",  # noqa: E501
535
                                        extras={"olmo": "allenai/Molmo-7B-O-0924"},  # noqa: E501
536
537
538
                                        trust_remote_code=True),
    "NVLM_D": _HfExamplesInfo("nvidia/NVLM-D-72B",
                              trust_remote_code=True),
539
540
    "Llama_Nemotron_Nano_VL" : _HfExamplesInfo("nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1", # noqa: E501
                                                     trust_remote_code=True),
541
    "NemotronH_Nano_VL_V2": _HfExamplesInfo("nano_vl_dummy",
542
543
                                          is_available_online=False,
                                          trust_remote_code=True),
544
    "Ovis": _HfExamplesInfo("AIDC-AI/Ovis2-1B", trust_remote_code=True,
545
546
                            max_transformers_version="4.53",
                            transformers_version_reason="HF model is not compatible",  # noqa: E501
547
548
                            extras={"1.6-llama": "AIDC-AI/Ovis1.6-Llama3.2-3B",
                                    "1.6-gemma": "AIDC-AI/Ovis1.6-Gemma2-9B"}),  # noqa: E501
549
    "Ovis2_5": _HfExamplesInfo("AIDC-AI/Ovis2.5-2B",
550
                               trust_remote_code=True),
551
552
    "PaliGemmaForConditionalGeneration": _HfExamplesInfo("google/paligemma-3b-mix-224",  # noqa: E501
                                                         extras={"v2": "google/paligemma2-3b-ft-docci-448"}),  # noqa: E501
553
554
    "Phi3VForCausalLM": _HfExamplesInfo("microsoft/Phi-3-vision-128k-instruct",
                                        trust_remote_code=True,
555
556
                                        max_transformers_version="4.48",
                                        transformers_version_reason="Use of deprecated imports which have been removed.",  # noqa: E501
557
                                        extras={"phi3.5": "microsoft/Phi-3.5-vision-instruct"}),  # noqa: E501
558
559
    "Phi4MMForCausalLM": _HfExamplesInfo("microsoft/Phi-4-multimodal-instruct",
                                        trust_remote_code=True),
560
561
    "Phi4MultimodalForCausalLM": _HfExamplesInfo("microsoft/Phi-4-multimodal-instruct",  # noqa: E501
                                                 revision="refs/pr/70"),
562
    "PixtralForConditionalGeneration": _HfExamplesInfo("mistralai/Pixtral-12B-2409",  # noqa: E501
563
                                                       tokenizer_mode="mistral"),
564
565
566
567
    "QwenVLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen-VL",
                                                      extras={"chat": "Qwen/Qwen-VL-Chat"},  # noqa: E501
                                                      trust_remote_code=True,
                                                      hf_overrides={"architectures": ["QwenVLForConditionalGeneration"]}),  # noqa: E501
568
569
    "Qwen2AudioForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2-Audio-7B-Instruct"),  # noqa: E501
    "Qwen2VLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2-VL-2B-Instruct"),  # noqa: E501
570
571
    "Qwen2_5_VLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2.5-VL-3B-Instruct", # noqa: E501
                                                          max_model_len=4096),
572
573
    "Qwen2_5OmniModel": _HfExamplesInfo("Qwen/Qwen2.5-Omni-3B"),
    "Qwen2_5OmniForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2.5-Omni-7B-AWQ"),  # noqa: E501
574
575
    "Qwen3VLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen3-VL-4B-Instruct", # noqa: E501
                                                        max_model_len=4096,
576
577
                                                        min_transformers_version="4.57",
                                                        is_available_online=False),
578
    "Qwen3VLMoeForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen3-VL-30B-A3B-Instruct", # noqa: E501
579
580
581
                                                          max_model_len=4096,
                                                          min_transformers_version="4.57",
                                                          is_available_online=False),
582
583
    "RForConditionalGeneration": _HfExamplesInfo("YannQi/R-4B",
                                                 trust_remote_code=True),
584
585
    "SkyworkR1VChatModel": _HfExamplesInfo("Skywork/Skywork-R1V-38B",
                                           trust_remote_code=True),
586
    "SmolVLMForConditionalGeneration": _HfExamplesInfo("HuggingFaceTB/SmolVLM2-2.2B-Instruct",  # noqa: E501
587
588
                                                       min_transformers_version="4.56",
                                                       transformers_version_reason="HF model broken in 4.55"),  # noqa: E501
Song's avatar
Song committed
589
    "Step3VLForConditionalGeneration": _HfExamplesInfo("stepfun-ai/step3",
590
                                                        trust_remote_code=True),
591
    "UltravoxModel": _HfExamplesInfo("fixie-ai/ultravox-v0_5-llama-3_2-1b",  # noqa: E501
592
                                     trust_remote_code=True),
593
    "TarsierForConditionalGeneration": _HfExamplesInfo("omni-research/Tarsier-7b"),  # noqa: E501
594
595
    "Tarsier2ForConditionalGeneration": _HfExamplesInfo("omni-research/Tarsier2-Recap-7b",  # noqa: E501
                                                        hf_overrides={"architectures": ["Tarsier2ForConditionalGeneration"]}),  # noqa: E501
596
597
598
599
600
601
    "VoxtralForConditionalGeneration": _HfExamplesInfo(
        "mistralai/Voxtral-Mini-3B-2507",
        min_transformers_version="4.54",
        # disable this temporarily until we support HF format
        is_available_online=False,
    ),
602
    # [Encoder-decoder]
603
    "WhisperForConditionalGeneration": _HfExamplesInfo("openai/whisper-large-v3"),  # noqa: E501
604
605
    # [Cross-encoder]
    "JinaVLForRanking": _HfExamplesInfo("jinaai/jina-reranker-m0"),   # noqa: E501
606
607
}

608

609
610
611
_SPECULATIVE_DECODING_EXAMPLE_MODELS = {
    "MedusaModel": _HfExamplesInfo("JackFram/llama-68m",
                                   speculative_model="abhigoyal/vllm-medusa-llama-68m-random"),  # noqa: E501
612
613
614
615
    # 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
616
617
618
    "DeepSeekMTPModel": _HfExamplesInfo("luccafong/deepseek_mtp_main_random",
                                        speculative_model="luccafong/deepseek_mtp_draft_random",  # noqa: E501
                                        trust_remote_code=True),
619
620
621
    "EagleDeepSeekMTPModel": _HfExamplesInfo("eagle618/deepseek-v3-random",
                                        speculative_model="eagle618/eagle-deepseek-v3-random",  # noqa: E501
                                        trust_remote_code=True),
622
    "EagleLlamaForCausalLM": _HfExamplesInfo("meta-llama/Meta-Llama-3-8B-Instruct", # noqa: E501
623
624
                                             trust_remote_code=True,
                                             speculative_model="yuhuili/EAGLE-LLaMA3-Instruct-8B",
625
626
                                             tokenizer="meta-llama/Meta-Llama-3-8B-Instruct"), # noqa: E501
    "Eagle3LlamaForCausalLM": _HfExamplesInfo("meta-llama/Llama-3.1-8B-Instruct",  # noqa: E501
627
                                            trust_remote_code=True,
628
629
630
631
632
                                            speculative_model="yuhuili/EAGLE3-LLaMA3.1-Instruct-8B", # noqa: E501
                                            tokenizer="meta-llama/Llama-3.1-8B-Instruct",
                                            use_original_num_layers=True,
                                            max_model_len=10240),
    "LlamaForCausalLMEagle3": _HfExamplesInfo("Qwen/Qwen3-8B",  # noqa: E501
633
634
                                            trust_remote_code=True,
                                            speculative_model="AngelSlim/Qwen3-8B_eagle3",   # noqa: E501
635
636
                                            tokenizer="Qwen/Qwen3-8B",
                                            use_original_num_layers=True),
zhiweiz's avatar
zhiweiz committed
637
638
639
640
641
    "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
642
643
644
645
646
    "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"),
647
648
649
    "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
650
651
    "Glm4MoeMTPModel": _HfExamplesInfo("zai-org/GLM-4.5",
                                        speculative_model="zai-org/GLM-4.5",
652
                                        min_transformers_version="4.56",
Yuxuan Zhang's avatar
Yuxuan Zhang committed
653
                                        is_available_online=False),
XuruiYang's avatar
XuruiYang committed
654
655
656
657
    "LongCatFlashMTPModel": _HfExamplesInfo(
        "meituan-longcat/LongCat-Flash-Chat",
        trust_remote_code=True,
        speculative_model="meituan-longcat/LongCat-Flash-Chat"),
658
659
    "MiMoMTPModel": _HfExamplesInfo("XiaomiMiMo/MiMo-7B-RL",
                                    trust_remote_code=True,
660
                                    speculative_model="XiaomiMiMo/MiMo-7B-RL"),
661
662
663
    "Eagle3Qwen2_5vlForCausalLM": _HfExamplesInfo(
        "Qwen/Qwen2.5-VL-7B-Instruct",
        speculative_model="Rayzl/qwen2.5-vl-7b-eagle3-sgl"),
664
    "Qwen3NextMTP": _HfExamplesInfo("Qwen/Qwen3-Next-80B-A3B-Instruct",
665
                                     min_transformers_version="4.56.3"),
666
667
}

668
_TRANSFORMERS_BACKEND_MODELS = {
669
670
    "TransformersEmbeddingModel": _HfExamplesInfo("BAAI/bge-base-en-v1.5", min_transformers_version="4.57.0.dev0"),  # noqa: E501
    "TransformersForSequenceClassification": _HfExamplesInfo("papluca/xlm-roberta-base-language-detection", min_transformers_version="4.57.0.dev0"),  # noqa: E501
671
    "TransformersForCausalLM": _HfExamplesInfo("hmellor/Ilama-3.2-1B", trust_remote_code=True),  # noqa: E501
672
    "TransformersForMultimodalLM": _HfExamplesInfo("BAAI/Emu3-Chat-hf"),
673
674
675
676
    "TransformersMoEForCausalLM": _HfExamplesInfo("allenai/OLMoE-1B-7B-0924", min_transformers_version="4.57.0.dev0"),  # noqa: E501
    "TransformersMoEForMultimodalLM": _HfExamplesInfo("Qwen/Qwen3-VL-30B-A3B-Instruct", min_transformers_version="4.57.0.dev0"),  # noqa: E501
    "TransformersMoEEmbeddingModel": _HfExamplesInfo("Qwen/Qwen3-30B-A3B", min_transformers_version="4.57.0.dev0"),  # noqa: E501
    "TransformersMoEForSequenceClassification": _HfExamplesInfo("Qwen/Qwen3-30B-A3B", min_transformers_version="4.57.0.dev0"),  # noqa: E501
677
678
}

679
680
681
_EXAMPLE_MODELS = {
    **_TEXT_GENERATION_EXAMPLE_MODELS,
    **_EMBEDDING_EXAMPLE_MODELS,
682
    **_SEQUENCE_CLASSIFICATION_EXAMPLE_MODELS,
683
684
    **_MULTIMODAL_EXAMPLE_MODELS,
    **_SPECULATIVE_DECODING_EXAMPLE_MODELS,
685
    **_TRANSFORMERS_BACKEND_MODELS,
686
687
688
689
690
691
692
693
694
}


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

        self.hf_models = hf_models

695
    def get_supported_archs(self) -> Set[str]:
696
697
698
        return self.hf_models.keys()

    def get_hf_info(self, model_arch: str) -> _HfExamplesInfo:
699
700
701
702
703
        try:
            return self.hf_models[model_arch]
        except KeyError:
            raise ValueError(f"No example model defined for {model_arch}; "
                             f"please update this file.") from None
704

705
706
707
708
709
    def find_hf_info(self, model_id: str) -> _HfExamplesInfo:
        for info in self.hf_models.values():
            if info.default == model_id:
                return info

710
711
712
713
714
        # Fallback to extras
        for info in self.hf_models.values():
            if any(extra == model_id for extra in info.extras.values()):
                return info

715
716
        raise ValueError(f"No example model defined for {model_id}; "
                         f"please update this file.")
717

718

Patrick von Platen's avatar
Patrick von Platen committed
719
HF_EXAMPLE_MODELS = HfExampleModels(_EXAMPLE_MODELS)
720
AUTO_EXAMPLE_MODELS = HfExampleModels(_AUTOMATIC_CONVERTED_MODELS)