registry.py 42.6 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
        if (self.min_transformers_version is None
                and self.max_transformers_version is None):
119
            return None
120
121

        current_version = TRANSFORMERS_VERSION
122
        cur_base_version = Version(current_version).base_version
123
124
125
        min_version = self.min_transformers_version
        max_version = self.max_transformers_version
        msg = f"`transformers=={current_version}` installed, but `transformers"
126
127
        # Only check the base version for the min/max version, otherwise preview
        # models cannot be run because `x.yy.0.dev0`<`x.yy.0`
128
129
        if (check_min_version and min_version
                and Version(cur_base_version) < Version(min_version)):
130
            msg += f">={min_version}` is required to run this model."
131
132
        elif (check_max_version and max_version
              and Version(cur_base_version) > Version(max_version)):
133
134
            msg += f"<={max_version}` is required to run this model."
        else:
135
            return None
136

137
138
139
140
141
        if self.transformers_version_reason:
            msg += f" Reason: {self.transformers_version_reason}"

        if on_fail == "error":
            raise RuntimeError(msg)
142
        elif on_fail == "skip":
143
            pytest.skip(msg)
144

145
146
        return msg

147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
    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)

163
164
165
166

# yapf: disable
_TEXT_GENERATION_EXAMPLE_MODELS = {
    # [Decoder-only]
167
    "ApertusForCausalLM": _HfExamplesInfo("swiss-ai/Apertus-8B-2509",
168
169
                                          min_transformers_version="4.56.0",
                                          trust_remote_code=True),
170
171
172
173
    "AquilaModel": _HfExamplesInfo("BAAI/AquilaChat-7B",
                                   trust_remote_code=True),
    "AquilaForCausalLM": _HfExamplesInfo("BAAI/AquilaChat2-7B",
                                         trust_remote_code=True),
174
    "ArceeForCausalLM": _HfExamplesInfo("arcee-ai/AFM-4.5B-Base"),
175
176
177
178
179
180
    "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),
181
182
    "BailingMoeForCausalLM": _HfExamplesInfo("inclusionAI/Ling-lite-1.5",
                                         trust_remote_code=True),
ant-yy's avatar
ant-yy committed
183
184
    "BailingMoeV2ForCausalLM": _HfExamplesInfo("inclusionAI/Ling-mini-2.0",
                                         trust_remote_code=True),
185
    "BambaForCausalLM": _HfExamplesInfo("ibm-ai-platform/Bamba-9B-v1",
186
                                        min_transformers_version="4.55.3",
187
                                        extras={"tiny": "hmellor/tiny-random-BambaForCausalLM"}),  # noqa: E501
188
189
    "BloomForCausalLM": _HfExamplesInfo("bigscience/bloom-560m",
                                        {"1b": "bigscience/bloomz-1b1"}),
190
    "ChatGLMModel": _HfExamplesInfo("zai-org/chatglm3-6b",
191
                                    trust_remote_code=True,
192
                                    max_transformers_version="4.48"),
193
194
    "ChatGLMForConditionalGeneration": _HfExamplesInfo("thu-coai/ShieldLM-6B-chatglm3",  # noqa: E501
                                                       trust_remote_code=True),
195
196
    "CohereForCausalLM": _HfExamplesInfo("CohereForAI/c4ai-command-r-v01",
                                         trust_remote_code=True),
197
198
    "Cohere2ForCausalLM": _HfExamplesInfo("CohereForAI/c4ai-command-r7b-12-2024", # noqa: E501
                                         trust_remote_code=True),
199
    "DbrxForCausalLM": _HfExamplesInfo("databricks/dbrx-instruct"),
200
    "DeciLMForCausalLM": _HfExamplesInfo("nvidia/Llama-3_3-Nemotron-Super-49B-v1", # noqa: E501
201
202
203
204
                                         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
205
206
    "DeepseekV3ForCausalLM": _HfExamplesInfo("deepseek-ai/DeepSeek-V3",  # noqa: E501
                                         trust_remote_code=True),
207
    "Ernie4_5ForCausalLM": _HfExamplesInfo("baidu/ERNIE-4.5-0.3B-PT",
208
                                            min_transformers_version="4.54"),
209
    "Ernie4_5_MoeForCausalLM": _HfExamplesInfo("baidu/ERNIE-4.5-21B-A3B-PT",
210
                                               min_transformers_version="4.54"),
211
212
213
214
    "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"),
215
    "Fairseq2LlamaForCausalLM": _HfExamplesInfo("mgleize/fairseq2-dummy-Llama-3.2-1B"),  # noqa: E501
216
    "FalconForCausalLM": _HfExamplesInfo("tiiuae/falcon-7b"),
217
    "FalconH1ForCausalLM":_HfExamplesInfo("tiiuae/Falcon-H1-0.5B-Base"),
218
    "GemmaForCausalLM": _HfExamplesInfo("google/gemma-1.1-2b-it"),
219
    "Gemma2ForCausalLM": _HfExamplesInfo("google/gemma-2-9b"),
220
    "Gemma3ForCausalLM": _HfExamplesInfo("google/gemma-3-1b-it"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
221
    "Gemma3nForCausalLM": _HfExamplesInfo("google/gemma-3n-E2B-it",
Robert Shaw's avatar
Robert Shaw committed
222
                                          min_transformers_version="4.53"),
223
224
    "GlmForCausalLM": _HfExamplesInfo("zai-org/glm-4-9b-chat-hf"),
    "Glm4ForCausalLM": _HfExamplesInfo("zai-org/GLM-4-9B-0414"),
225
226
    "Glm4MoeForCausalLM": _HfExamplesInfo("zai-org/GLM-4.5",
                                          min_transformers_version="4.54"),   # noqa: E501
227
228
229
    "GPT2LMHeadModel": _HfExamplesInfo("openai-community/gpt2",
                                       {"alias": "gpt2"}),
    "GPTBigCodeForCausalLM": _HfExamplesInfo("bigcode/starcoder",
230
                                             extras={"tiny": "bigcode/tiny_starcoder_py"},  # noqa: E501
231
232
                                             min_transformers_version="4.55.1",
                                             transformers_version_reason="HF model broken in 4.55.0"),  # noqa: E501
233
234
235
236
    "GPTJForCausalLM": _HfExamplesInfo("Milos/slovak-gpt-j-405M",
                                       {"6b": "EleutherAI/gpt-j-6b"}),
    "GPTNeoXForCausalLM": _HfExamplesInfo("EleutherAI/pythia-70m",
                                          {"1b": "EleutherAI/pythia-1.4b"}),
237
    "GptOssForCausalLM": _HfExamplesInfo("lmsys/gpt-oss-20b-bf16"),
238
239
    "GraniteForCausalLM": _HfExamplesInfo("ibm/PowerLM-3b"),
    "GraniteMoeForCausalLM": _HfExamplesInfo("ibm/PowerMoE-3b"),
240
241
    "GraniteMoeHybridForCausalLM": _HfExamplesInfo("ibm-granite/granite-4.0-tiny-preview", # noqa: E501
                                                   min_transformers_version="4.55.3"),
242
    "GraniteMoeSharedForCausalLM": _HfExamplesInfo("ibm-research/moe-7b-1b-active-shared-experts"),  # noqa: E501
Michael Goin's avatar
Michael Goin committed
243
244
    "Grok1ModelForCausalLM": _HfExamplesInfo("hpcai-tech/grok-1",
                                             trust_remote_code=True),
245
246
    "HunYuanMoEV1ForCausalLM": _HfExamplesInfo("tencent/Hunyuan-A13B-Instruct",
                                               trust_remote_code=True),
247
    # TODO: Remove is_available_online once their config.json is fixed
248
    "HunYuanDenseV1ForCausalLM":_HfExamplesInfo("tencent/Hunyuan-7B-Instruct-0124",
249
250
                                                trust_remote_code=True,
                                                is_available_online=False),
251
252
253
254
255
256
    "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),
257
258
    "InternLM3ForCausalLM": _HfExamplesInfo("internlm/internlm3-8b-instruct",
                                            trust_remote_code=True),
259
    "JAISLMHeadModel": _HfExamplesInfo("inceptionai/jais-13b-chat"),
260
    "JambaForCausalLM": _HfExamplesInfo("ai21labs/AI21-Jamba-1.5-Mini",
261
                                        min_transformers_version="4.55.3",
262
263
264
265
                                        extras={
                                            "tiny": "ai21labs/Jamba-tiny-dev",
                                            "random": "ai21labs/Jamba-tiny-random",  # noqa: E501
                                        }),
266
267
    "Lfm2ForCausalLM": _HfExamplesInfo("LiquidAI/LFM2-1.2B",
                                       min_transformers_version="4.54"),
268
269
    "LlamaForCausalLM": _HfExamplesInfo("meta-llama/Llama-3.2-1B-Instruct",
                                        extras={"guard": "meta-llama/Llama-Guard-3-1B",  # noqa: E501
270
271
                                                "hermes": "NousResearch/Hermes-3-Llama-3.1-8B", # noqa: E501
                                                "fp8": "RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8"}),  # noqa: E501
272
273
    "LLaMAForCausalLM": _HfExamplesInfo("decapoda-research/llama-7b-hf",
                                        is_available_online=False),
274
275
    "Llama4ForCausalLM": _HfExamplesInfo("meta-llama/Llama-4-Scout-17B-16E-Instruct", # noqa: E501
                                         is_available_online=False),
276
    "MambaForCausalLM": _HfExamplesInfo("state-spaces/mamba-130m-hf"),
277
278
279
280
281
    "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
                                         }),
282
283
284
285
286
    "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),
287
    "MiniMaxForCausalLM": _HfExamplesInfo("MiniMaxAI/MiniMax-Text-01-hf"),
288
    "MiniMaxText01ForCausalLM": _HfExamplesInfo("MiniMaxAI/MiniMax-Text-01",
289
290
                                                trust_remote_code=True,
                                                revision="a59aa9cbc53b9fb8742ca4e9e1531b9802b6fdc3"),  # noqa: E501
291
292
    "MiniMaxM1ForCausalLM": _HfExamplesInfo("MiniMaxAI/MiniMax-M1-40k",
                                            trust_remote_code=True),
293
    "MistralForCausalLM": _HfExamplesInfo("mistralai/Mistral-7B-Instruct-v0.1"),
294
    "MixtralForCausalLM": _HfExamplesInfo("mistralai/Mixtral-8x7B-Instruct-v0.1",  # noqa: E501
295
                                          {"tiny": "TitanML/tiny-mixtral"}),  # noqa: E501
296
297
298
    "MotifForCausalLM": _HfExamplesInfo("Motif-Technologies/Motif-2.6B",
                                        trust_remote_code=True,
                                        v0_only=True),
299
300
301
    "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
302
303
    "NemotronHForCausalLM": _HfExamplesInfo("nvidia/Nemotron-H-8B-Base-8K",
                                            trust_remote_code=True),
304
    "OlmoForCausalLM": _HfExamplesInfo("allenai/OLMo-1B-hf"),
305
    "Olmo2ForCausalLM": _HfExamplesInfo("allenai/OLMo-2-0425-1B"),
306
    "Olmo3ForCausalLM": _HfExamplesInfo("shanearora/2025-sep-a-base-model"),
307
    "OlmoeForCausalLM": _HfExamplesInfo("allenai/OLMoE-1B-7B-0924-Instruct"),
308
309
    "OPTForCausalLM": _HfExamplesInfo("facebook/opt-125m",
                                      {"1b": "facebook/opt-iml-max-1.3b"}),
310
311
312
    "OrionForCausalLM": _HfExamplesInfo("OrionStarAI/Orion-14B-Chat",
                                        trust_remote_code=True),
    "PersimmonForCausalLM": _HfExamplesInfo("adept/persimmon-8b-chat"),
313
    "PhiForCausalLM": _HfExamplesInfo("microsoft/phi-2"),
314
    "Phi3ForCausalLM": _HfExamplesInfo("microsoft/Phi-3-mini-4k-instruct"),
315
316
317
318
    "Phi4FlashForCausalLM": _HfExamplesInfo("microsoft/Phi-4-mini-flash-reasoning", # noqa: E501
                                        trust_remote_code=True,
                                        v0_only=True,
                                        max_model_len=10240),
319
320
    "PhiMoEForCausalLM": _HfExamplesInfo("microsoft/Phi-3.5-MoE-instruct",
                                         trust_remote_code=True),
Shinichi Hemmi's avatar
Shinichi Hemmi committed
321
322
    "Plamo2ForCausalLM": _HfExamplesInfo("pfnet/plamo-2-1b",
                                        trust_remote_code=True),
323
    "QWenLMHeadModel": _HfExamplesInfo("Qwen/Qwen-7B-Chat",
324
325
                                       max_transformers_version="4.53",
                                       transformers_version_reason="HF model uses remote code that is not compatible with latest Transformers",  # noqa: E501
326
                                       trust_remote_code=True),
327
328
    "Qwen2ForCausalLM": _HfExamplesInfo("Qwen/Qwen2-0.5B-Instruct",
                                        extras={"2.5": "Qwen/Qwen2.5-0.5B-Instruct"}), # noqa: E501
329
    "Qwen2MoeForCausalLM": _HfExamplesInfo("Qwen/Qwen1.5-MoE-A2.7B-Chat"),
Jee Jee Li's avatar
Jee Jee Li committed
330
331
    "Qwen3ForCausalLM": _HfExamplesInfo("Qwen/Qwen3-8B"),
    "Qwen3MoeForCausalLM": _HfExamplesInfo("Qwen/Qwen3-30B-A3B"),
332
333
    "Qwen3NextForCausalLM": _HfExamplesInfo("Qwen/Qwen3-Next-80B-A3B-Instruct",
                                            min_transformers_version="4.56.2"),
334
    "RWForCausalLM": _HfExamplesInfo("tiiuae/falcon-40b"),
335
336
337
    "SeedOssForCausalLM": _HfExamplesInfo("ByteDance-Seed/Seed-OSS-36B-Instruct", # noqa: E501
                                          trust_remote_code=True,
                                          is_available_online=False),
338
    "SmolLM3ForCausalLM": _HfExamplesInfo("HuggingFaceTB/SmolLM3-3B"),
339
340
    "StableLMEpochForCausalLM": _HfExamplesInfo("stabilityai/stablelm-zephyr-3b"),  # noqa: E501
    "StableLmForCausalLM": _HfExamplesInfo("stabilityai/stablelm-3b-4e1t"),
341
    "Starcoder2ForCausalLM": _HfExamplesInfo("bigcode/starcoder2-3b"),
Song's avatar
Song committed
342
    "Step3TextForCausalLM": _HfExamplesInfo("stepfun-ai/step3",
343
                                            trust_remote_code=True),
344
345
    "SolarForCausalLM": _HfExamplesInfo("upstage/solar-pro-preview-instruct",
                                        trust_remote_code=True),
346
347
    "TeleChat2ForCausalLM": _HfExamplesInfo("Tele-AI/TeleChat2-3B",
                                            trust_remote_code=True),
348
349
    "TeleFLMForCausalLM": _HfExamplesInfo("CofeAI/FLM-2-52B-Instruct-2407",
                                            trust_remote_code=True),
350
    "XverseForCausalLM": _HfExamplesInfo("xverse/XVERSE-7B-Chat",
351
                                         tokenizer="meta-llama/Llama-2-7b",
352
                                         trust_remote_code=True),
353
    "Zamba2ForCausalLM": _HfExamplesInfo("Zyphra/Zamba2-7B-instruct"),
354
355
    "MiMoForCausalLM": _HfExamplesInfo("XiaomiMiMo/MiMo-7B-RL",
                                        trust_remote_code=True),
356
    "Dots1ForCausalLM": _HfExamplesInfo("rednote-hilab/dots.llm1.inst"),
357
358
359
360
}

_EMBEDDING_EXAMPLE_MODELS = {
    # [Text-only]
361
362
    "BertModel": _HfExamplesInfo("BAAI/bge-base-en-v1.5"),
    "Gemma2Model": _HfExamplesInfo("BAAI/bge-multilingual-gemma2"),  # noqa: E501
363
    "Gemma3TextModel": _HfExamplesInfo("google/embeddinggemma-300m"),
364
    "GritLM": _HfExamplesInfo("parasail-ai/GritLM-7B-vllm"),
365
366
    "GteModel": _HfExamplesInfo("Snowflake/snowflake-arctic-embed-m-v2.0",
                                               trust_remote_code=True),
367
368
    "GteNewModel": _HfExamplesInfo("Alibaba-NLP/gte-base-en-v1.5",
                                   trust_remote_code=True,
369
                                   hf_overrides={"architectures": ["GteNewModel"]}),  # noqa: E501
370
371
    "InternLM2ForRewardModel": _HfExamplesInfo("internlm/internlm2-1_8b-reward",
                                               trust_remote_code=True),
372
    "JambaForSequenceClassification": _HfExamplesInfo("ai21labs/Jamba-tiny-reward-dev"),  # noqa: E501
373
    "LlamaModel": _HfExamplesInfo("llama", is_available_online=False),
374
    "MistralModel": _HfExamplesInfo("intfloat/e5-mistral-7b-instruct"),
375
    "ModernBertModel": _HfExamplesInfo("Alibaba-NLP/gte-modernbert-base",
376
                                trust_remote_code=True),
377
    "NomicBertModel": _HfExamplesInfo("nomic-ai/nomic-embed-text-v2-moe",
378
                                               trust_remote_code=True),  # noqa: E501
379
    "Qwen2Model": _HfExamplesInfo("ssmits/Qwen2-7B-Instruct-embed-base"),
380
381
382
383
384
385
    "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
386
387
388
    "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
389
390
391
392
393
    # [Multimodal]
    "LlavaNextForConditionalGeneration": _HfExamplesInfo("royokong/e5-v"),
    "Phi3VForCausalLM": _HfExamplesInfo("TIGER-Lab/VLM2Vec-Full",
                                         trust_remote_code=True),
    "Qwen2VLForConditionalGeneration": _HfExamplesInfo("MrLight/dse-qwen2-2b-mrl-v1"), # noqa: E501
394
    "PrithviGeoSpatialMAE": _HfExamplesInfo("ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11", # noqa: E501
395
396
397
398
399
400
401
                                            dtype=torch.float16,
                                            enforce_eager=True,
                                            skip_tokenizer_init=True,
                                            # This is to avoid the model
                                            # going OOM in CI
                                            max_num_seqs=32,
                                            ),
402
    "Terratorch": _HfExamplesInfo("ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11", # noqa: E501
403
404
405
406
407
408
                                  dtype=torch.float16,
                                  enforce_eager=True,
                                  skip_tokenizer_init=True,
                                  # This is to avoid the model going OOM in CI
                                  max_num_seqs=32,
                                  ),
409
410
}

411
412
413
414
415
_SEQUENCE_CLASSIFICATION_EXAMPLE_MODELS = {
    # [Decoder-only]
    "GPT2ForSequenceClassification": _HfExamplesInfo("nie3e/sentiment-polish-gpt2-small"),  # noqa: E501

    # [Cross-encoder]
416
    "BertForSequenceClassification": _HfExamplesInfo("cross-encoder/ms-marco-MiniLM-L-6-v2"),  # noqa: E501
417
418
419
420
    "GteNewForSequenceClassification": _HfExamplesInfo("Alibaba-NLP/gte-multilingual-reranker-base",  # noqa: E501
                                                       trust_remote_code=True,
                                                       hf_overrides={
                                                           "architectures": ["GteNewForSequenceClassification"]}),# noqa: E501
421
422
423
    "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
424
425
}

426
427
428
429
430
431
432
433
434
435
436
_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
}

437
438
_MULTIMODAL_EXAMPLE_MODELS = {
    # [Decoder-only]
439
    "AriaForConditionalGeneration": _HfExamplesInfo("rhymes-ai/Aria"),
Jennifer Zhao's avatar
Jennifer Zhao committed
440
    "AyaVisionForConditionalGeneration": _HfExamplesInfo("CohereForAI/aya-vision-8b"), # noqa: E501
441
    "Blip2ForConditionalGeneration": _HfExamplesInfo("Salesforce/blip2-opt-2.7b",  # noqa: E501
442
                                                     extras={"6b": "Salesforce/blip2-opt-6.7b"}),  # noqa: E501
443
    "ChameleonForConditionalGeneration": _HfExamplesInfo("facebook/chameleon-7b"),  # noqa: E501
444
    "Cohere2VisionForConditionalGeneration": _HfExamplesInfo("CohereLabs/command-a-vision-07-2025"), # noqa: E501
445
    "DeepseekVLV2ForCausalLM": _HfExamplesInfo("deepseek-ai/deepseek-vl2-tiny",  # noqa: E501
446
447
448
                                                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
449
                                                hf_overrides={"architectures": ["DeepseekVLV2ForCausalLM"]}),  # noqa: E501
450
    "Emu3ForConditionalGeneration": _HfExamplesInfo("BAAI/Emu3-Chat-hf"),
451
452
    "Ernie4_5_VLMoeForConditionalGeneration": _HfExamplesInfo("baidu/ERNIE-4.5-VL-28B-A3B-PT",  # noqa: E501
                                                              trust_remote_code=True),
453
    "FuyuForCausalLM": _HfExamplesInfo("adept/fuyu-8b"),
454
    "Gemma3ForConditionalGeneration": _HfExamplesInfo("google/gemma-3-4b-it"),
Nicolò Lucchesi's avatar
Nicolò Lucchesi committed
455
456
    "Gemma3nForConditionalGeneration": _HfExamplesInfo("google/gemma-3n-E2B-it",    # noqa: E501
                                        min_transformers_version="4.53"),
457
    "GraniteSpeechForConditionalGeneration": _HfExamplesInfo("ibm-granite/granite-speech-3.3-2b"),  # noqa: E501
458
    "GLM4VForCausalLM": _HfExamplesInfo("zai-org/glm-4v-9b",
459
460
                                        trust_remote_code=True,
                                        hf_overrides={"architectures": ["GLM4VForCausalLM"]}),  # noqa: E501
461
    "Glm4vForConditionalGeneration": _HfExamplesInfo("zai-org/GLM-4.1V-9B-Thinking"),  # noqa: E501
462
    "Glm4vMoeForConditionalGeneration": _HfExamplesInfo("zai-org/GLM-4.5V",
463
                                                        min_transformers_version="4.56"),  # noqa: E501
464
    "H2OVLChatModel": _HfExamplesInfo("h2oai/h2ovl-mississippi-800m",
465
                                      trust_remote_code=True,
466
467
468
                                      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
469
470
    "HCXVisionForCausalLM": _HfExamplesInfo("naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B",  # noqa: E501
                                            trust_remote_code=True),
471
472
    "Idefics3ForConditionalGeneration": _HfExamplesInfo("HuggingFaceM4/Idefics3-8B-Llama3",  # noqa: E501
                                                        {"tiny": "HuggingFaceTB/SmolVLM-256M-Instruct"},    # noqa: E501
473
474
                                                        min_transformers_version="4.56",
                                                        transformers_version_reason="HF model broken in 4.55"),  # noqa: E501
475
476
    "InternS1ForConditionalGeneration": _HfExamplesInfo("internlm/Intern-S1",
                                                        trust_remote_code=True),  # noqa: E501
477
    "InternVLChatModel": _HfExamplesInfo("OpenGVLab/InternVL2-1B",
478
                                         extras={"2B": "OpenGVLab/InternVL2-2B",
479
480
481
482
                                                 "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
483
                                         trust_remote_code=True),
484
    "InternVLForConditionalGeneration": _HfExamplesInfo("OpenGVLab/InternVL3-1B-hf"),    # noqa: E501
485
486
    "KeyeForConditionalGeneration": _HfExamplesInfo("Kwai-Keye/Keye-VL-8B-Preview", # noqa: E501
                                                    trust_remote_code=True),
487
488
    "KeyeVL1_5ForConditionalGeneration": _HfExamplesInfo("Kwai-Keye/Keye-VL-1_5-8B", # noqa: E501
                                                         trust_remote_code=True),
489
490
    "KimiVLForConditionalGeneration": _HfExamplesInfo("moonshotai/Kimi-VL-A3B-Instruct",  # noqa: E501
                                                      extras={"thinking": "moonshotai/Kimi-VL-A3B-Thinking"},  # noqa: E501
491
                                                      trust_remote_code=True),
492
    "Llama4ForConditionalGeneration": _HfExamplesInfo("meta-llama/Llama-4-Scout-17B-16E-Instruct",   # noqa: E501
493
                                                      max_model_len=10240,
494
                                                      extras={"llama-guard-4": "meta-llama/Llama-Guard-4-12B"},  # noqa: E501
495
                                                      ),
496
    "LlavaForConditionalGeneration": _HfExamplesInfo("llava-hf/llava-1.5-7b-hf",
497
498
                                                     extras={"mistral": "mistral-community/pixtral-12b", # noqa: E501
                                                             "mistral-fp8": "nm-testing/pixtral-12b-FP8-dynamic"}),  # noqa: E501
499
500
501
    "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
502
    "MantisForConditionalGeneration": _HfExamplesInfo("TIGER-Lab/Mantis-8B-siglip-llama3",  # noqa: E501
503
504
                                                      max_transformers_version="4.48",  # noqa: E501
                                                      transformers_version_reason="HF model is not compatible.",  # noqa: E501
505
                                                      hf_overrides={"architectures": ["MantisForConditionalGeneration"]}),  # noqa: E501
506
507
    "MiDashengLMModel": _HfExamplesInfo("mispeech/midashenglm-7b",
                            trust_remote_code=True),
508
509
    "MiniCPMO": _HfExamplesInfo("openbmb/MiniCPM-o-2_6",
                                trust_remote_code=True),
510
    "MiniCPMV": _HfExamplesInfo("openbmb/MiniCPM-Llama3-V-2_5",
tc-mb's avatar
tc-mb committed
511
                                extras={"2.6": "openbmb/MiniCPM-V-2_6", "4.0": "openbmb/MiniCPM-V-4", "4.5": "openbmb/MiniCPM-V-4_5"},  # noqa: E501
512
                                trust_remote_code=True),
513
    "MiniMaxVL01ForConditionalGeneration": _HfExamplesInfo("MiniMaxAI/MiniMax-VL-01", # noqa: E501
514
515
                                              trust_remote_code=True,
                                              v0_only=True),
516
517
    "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
518
    "MolmoForCausalLM": _HfExamplesInfo("allenai/Molmo-7B-D-0924",
519
                                        max_transformers_version="4.48",
520
                                        transformers_version_reason="Incorrectly-detected `tensorflow` import.",  # noqa: E501
521
                                        extras={"olmo": "allenai/Molmo-7B-O-0924"},  # noqa: E501
522
523
524
                                        trust_remote_code=True),
    "NVLM_D": _HfExamplesInfo("nvidia/NVLM-D-72B",
                              trust_remote_code=True),
525
526
    "Llama_Nemotron_Nano_VL" : _HfExamplesInfo("nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1", # noqa: E501
                                                     trust_remote_code=True),
527
528
529
    "NemotronH_Nano_VL": _HfExamplesInfo("nano_vl_dummy",
                                          is_available_online=False,
                                          trust_remote_code=True),
530
    "Ovis": _HfExamplesInfo("AIDC-AI/Ovis2-1B", trust_remote_code=True,
531
532
                            max_transformers_version="4.53",
                            transformers_version_reason="HF model is not compatible",  # noqa: E501
533
534
                            extras={"1.6-llama": "AIDC-AI/Ovis1.6-Llama3.2-3B",
                                    "1.6-gemma": "AIDC-AI/Ovis1.6-Gemma2-9B"}),  # noqa: E501
535
    "Ovis2_5": _HfExamplesInfo("AIDC-AI/Ovis2.5-2B",
536
                               trust_remote_code=True),
537
538
    "PaliGemmaForConditionalGeneration": _HfExamplesInfo("google/paligemma-3b-mix-224",  # noqa: E501
                                                         extras={"v2": "google/paligemma2-3b-ft-docci-448"}),  # noqa: E501
539
540
    "Phi3VForCausalLM": _HfExamplesInfo("microsoft/Phi-3-vision-128k-instruct",
                                        trust_remote_code=True,
541
542
                                        max_transformers_version="4.48",
                                        transformers_version_reason="Use of deprecated imports which have been removed.",  # noqa: E501
543
                                        extras={"phi3.5": "microsoft/Phi-3.5-vision-instruct"}),  # noqa: E501
544
545
    "Phi4MMForCausalLM": _HfExamplesInfo("microsoft/Phi-4-multimodal-instruct",
                                        trust_remote_code=True),
546
547
    "Phi4MultimodalForCausalLM": _HfExamplesInfo("microsoft/Phi-4-multimodal-instruct",  # noqa: E501
                                                 revision="refs/pr/70"),
548
    "PixtralForConditionalGeneration": _HfExamplesInfo("mistralai/Pixtral-12B-2409",  # noqa: E501
549
                                                       tokenizer_mode="mistral"),
550
551
552
553
    "QwenVLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen-VL",
                                                      extras={"chat": "Qwen/Qwen-VL-Chat"},  # noqa: E501
                                                      trust_remote_code=True,
                                                      hf_overrides={"architectures": ["QwenVLForConditionalGeneration"]}),  # noqa: E501
554
555
    "Qwen2AudioForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2-Audio-7B-Instruct"),  # noqa: E501
    "Qwen2VLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2-VL-2B-Instruct"),  # noqa: E501
556
557
    "Qwen2_5_VLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2.5-VL-3B-Instruct", # noqa: E501
                                                          max_model_len=4096),
558
559
    "Qwen2_5OmniModel": _HfExamplesInfo("Qwen/Qwen2.5-Omni-3B"),
    "Qwen2_5OmniForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2.5-Omni-7B-AWQ"),  # noqa: E501
560
561
562
563
564
565
    "Qwen3VLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen3-VL-4B-Instruct", # noqa: E501
                                                        max_model_len=4096,
                                                        min_transformers_version="4.57"),  # noqa: E501
    "Qwen3VLMoeForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen3-VL-30B-A3B-Instruct", # noqa: E501
                                                        max_model_len=4096,
                                                        min_transformers_version="4.57"),
566
567
    "RForConditionalGeneration": _HfExamplesInfo("YannQi/R-4B",
                                                 trust_remote_code=True),
568
569
    "SkyworkR1VChatModel": _HfExamplesInfo("Skywork/Skywork-R1V-38B",
                                           trust_remote_code=True),
570
    "SmolVLMForConditionalGeneration": _HfExamplesInfo("HuggingFaceTB/SmolVLM2-2.2B-Instruct",  # noqa: E501
571
572
                                                       min_transformers_version="4.56",
                                                       transformers_version_reason="HF model broken in 4.55"),  # noqa: E501
Song's avatar
Song committed
573
    "Step3VLForConditionalGeneration": _HfExamplesInfo("stepfun-ai/step3",
574
                                                        trust_remote_code=True),
575
    "UltravoxModel": _HfExamplesInfo("fixie-ai/ultravox-v0_5-llama-3_2-1b",  # noqa: E501
576
                                     trust_remote_code=True),
577
    "TarsierForConditionalGeneration": _HfExamplesInfo("omni-research/Tarsier-7b"),  # noqa: E501
578
579
    "Tarsier2ForConditionalGeneration": _HfExamplesInfo("omni-research/Tarsier2-Recap-7b",  # noqa: E501
                                                        hf_overrides={"architectures": ["Tarsier2ForConditionalGeneration"]}),  # noqa: E501
580
581
582
583
584
585
    "VoxtralForConditionalGeneration": _HfExamplesInfo(
        "mistralai/Voxtral-Mini-3B-2507",
        min_transformers_version="4.54",
        # disable this temporarily until we support HF format
        is_available_online=False,
    ),
586
    # [Encoder-decoder]
587
    "WhisperForConditionalGeneration": _HfExamplesInfo("openai/whisper-large-v3"),  # noqa: E501
588
589
    # [Cross-encoder]
    "JinaVLForRanking": _HfExamplesInfo("jinaai/jina-reranker-m0"),   # noqa: E501
590
591
}

592

593
594
595
_SPECULATIVE_DECODING_EXAMPLE_MODELS = {
    "MedusaModel": _HfExamplesInfo("JackFram/llama-68m",
                                   speculative_model="abhigoyal/vllm-medusa-llama-68m-random"),  # noqa: E501
596
597
598
599
    # 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
600
601
602
    "DeepSeekMTPModel": _HfExamplesInfo("luccafong/deepseek_mtp_main_random",
                                        speculative_model="luccafong/deepseek_mtp_draft_random",  # noqa: E501
                                        trust_remote_code=True),
603
604
605
    "EagleDeepSeekMTPModel": _HfExamplesInfo("eagle618/deepseek-v3-random",
                                        speculative_model="eagle618/eagle-deepseek-v3-random",  # noqa: E501
                                        trust_remote_code=True),
606
    "EagleLlamaForCausalLM": _HfExamplesInfo("meta-llama/Meta-Llama-3-8B-Instruct", # noqa: E501
607
608
                                             trust_remote_code=True,
                                             speculative_model="yuhuili/EAGLE-LLaMA3-Instruct-8B",
609
610
                                             tokenizer="meta-llama/Meta-Llama-3-8B-Instruct"), # noqa: E501
    "Eagle3LlamaForCausalLM": _HfExamplesInfo("meta-llama/Llama-3.1-8B-Instruct",  # noqa: E501
611
                                            trust_remote_code=True,
612
613
614
615
616
                                            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
617
618
                                            trust_remote_code=True,
                                            speculative_model="AngelSlim/Qwen3-8B_eagle3",   # noqa: E501
619
620
                                            tokenizer="Qwen/Qwen3-8B",
                                            use_original_num_layers=True),
zhiweiz's avatar
zhiweiz committed
621
622
623
624
625
    "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
626
627
628
629
630
    "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"),
631
632
633
    "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
634
635
    "Glm4MoeMTPModel": _HfExamplesInfo("zai-org/GLM-4.5",
                                        speculative_model="zai-org/GLM-4.5",
Yuxuan Zhang's avatar
Yuxuan Zhang committed
636
637
                                        min_transformers_version="4.54",
                                        is_available_online=False),
638
639
    "MiMoMTPModel": _HfExamplesInfo("XiaomiMiMo/MiMo-7B-RL",
                                    trust_remote_code=True,
640
641
642
                                    speculative_model="XiaomiMiMo/MiMo-7B-RL"),
    "Qwen3NextMTP": _HfExamplesInfo("Qwen/Qwen3-Next-80B-A3B-Instruct",
                                     min_transformers_version="4.56.2"),
643
644
}

645
_TRANSFORMERS_BACKEND_MODELS = {
646
    "TransformersModel": _HfExamplesInfo("Qwen/Qwen3-Embedding-0.6B"),
647
    "TransformersForCausalLM": _HfExamplesInfo("hmellor/Ilama-3.2-1B", trust_remote_code=True),  # noqa: E501
648
    "TransformersForMultimodalLM": _HfExamplesInfo("BAAI/Emu3-Chat-hf"),
649
650
}

651
652
653
_EXAMPLE_MODELS = {
    **_TEXT_GENERATION_EXAMPLE_MODELS,
    **_EMBEDDING_EXAMPLE_MODELS,
654
    **_SEQUENCE_CLASSIFICATION_EXAMPLE_MODELS,
655
656
    **_MULTIMODAL_EXAMPLE_MODELS,
    **_SPECULATIVE_DECODING_EXAMPLE_MODELS,
657
    **_TRANSFORMERS_BACKEND_MODELS,
658
659
660
661
662
663
664
665
666
}


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

        self.hf_models = hf_models

667
    def get_supported_archs(self) -> Set[str]:
668
669
670
671
672
        return self.hf_models.keys()

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

673
674
675
676
677
    def find_hf_info(self, model_id: str) -> _HfExamplesInfo:
        for info in self.hf_models.values():
            if info.default == model_id:
                return info

678
679
680
681
682
        # Fallback to extras
        for info in self.hf_models.values():
            if any(extra == model_id for extra in info.extras.values()):
                return info

683
684
        raise ValueError(f"No example model defined for {model_id}")

685

Patrick von Platen's avatar
Patrick von Platen committed
686
HF_EXAMPLE_MODELS = HfExampleModels(_EXAMPLE_MODELS)
687
AUTO_EXAMPLE_MODELS = HfExampleModels(_AUTOMATIC_CONVERTED_MODELS)