d881812500 · tduncan/Nvidia-Optimization-Stack-VLLM

docs/models/supported_models.md

+2 -2

@@ -164,9 +164,9 @@ For your model to be compatible with vLLM's tensor parallel and/or pipeline para

164

165 ```python

166

167 -from transformers import PretrainedConfig

167 +from transformers import PreTrainedConfig

168

169 -class MyConfig(PretrainedConfig):

169 +class MyConfig(PreTrainedConfig):

170 base_model_tp_plan = {

171 "layers.*.self_attn.k_proj": "colwise",

172 "layers.*.self_attn.v_proj": "colwise",

tests/config/test_bailing_mtp_config.py

+3 -3

@@ -3,7 +3,7 @@

3

4 from typing import get_args

5

6 -from transformers import PretrainedConfig

6 +from transformers import PreTrainedConfig

7

8 from vllm.config.speculative import MTPModelTypes, SpeculativeConfig

9 from vllm.transformers_utils.model_arch_config_convertor import (

@@ -11,8 +11,8 @@ from vllm.transformers_utils.model_arch_config_convertor import (

11 )

12

13

14 -def _bailing_config() -> PretrainedConfig:

15 - config = PretrainedConfig(

14 +def _bailing_config() -> PreTrainedConfig:

15 + config = PreTrainedConfig(

16 architectures=["BailingMoeV2_5ForCausalLM"],

17 hidden_size=4096,

18 kv_lora_rank=512,

tests/config/test_model_arch_config.py

+8 -8

@@ -10,7 +10,7 @@ from typing import cast

10 from unittest.mock import Mock

11

12 import pytest

13 -from transformers import PretrainedConfig

13 +from transformers import PreTrainedConfig

14 from transformers.models.gemma4.configuration_gemma4 import Gemma4TextConfig

15

16 from vllm.config import ModelConfig, ParallelConfig, SpeculativeConfig

@@ -125,7 +125,7 @@ def _assert_model_config_methods(

125

126 def test_head_size_falls_back_when_head_dim_is_zero():

127 """Regression test for configs that materialize missing head_dim as 0."""

128 - hf_config = PretrainedConfig(

128 + hf_config = PreTrainedConfig(

129 model_type="deepseek_vl_v2",

130 hidden_size=1280,

131 num_attention_heads=10,

@@ -140,7 +140,7 @@ def test_head_size_falls_back_when_head_dim_is_zero():

140

141

142 def test_qk_rope_head_dim_recovery_uses_model_revision(monkeypatch: pytest.MonkeyPatch):

143 - hf_config = PretrainedConfig(

143 + hf_config = PreTrainedConfig(

144 model_type="deepseek_v2",

145 qk_rope_head_dim=128,

146 qk_nope_head_dim=128,

@@ -176,7 +176,7 @@ def test_qk_rope_head_dim_recovery_uses_model_revision(monkeypatch: pytest.Monke

176 ],

177 )

178 def test_num_experts_per_tok_aliases(attribute: str):

179 - hf_config = PretrainedConfig(**{attribute: 4})

179 + hf_config = PreTrainedConfig(**{attribute: 4})

180 convertor = ModelArchConfigConvertorBase(hf_config, hf_config)

181 model_config = cast(

182 ModelConfig,

@@ -191,7 +191,7 @@ def test_num_experts_per_tok_aliases(attribute: str):

191

192

193 def test_num_experts_per_tok_none_is_normalized():

194 - hf_config = PretrainedConfig(top_k_experts=None)

194 + hf_config = PreTrainedConfig(top_k_experts=None)

195 convertor = ModelArchConfigConvertorBase(hf_config, hf_config)

196

197 assert convertor.get_num_experts_per_token() == 0

@@ -207,7 +207,7 @@ def test_legacy_modelopt_config_without_producer_is_normalized():

207 "modelopt_quant_config": {"quant_cfg": {}},

208 }

209 }

210 - hf_config = PretrainedConfig(quantization_config=quantization_config)

210 + hf_config = PreTrainedConfig(quantization_config=quantization_config)

211

212 convertor = ModelArchConfigConvertorBase(hf_config, hf_config)

213

@@ -370,7 +370,7 @@ def test_gemma4_uniform_head_dims_are_homogeneous():

370 assert arch[3] is arch

371

372

373 -class _HeterogeneousConfig(PretrainedConfig):

373 +class _HeterogeneousConfig(PreTrainedConfig):

374 """A heterogeneous config with no convertor of its own.

375

376 Mirrors the parts vLLM uses: per-layer configs are shallow copies with the

@@ -384,7 +384,7 @@ class _HeterogeneousConfig(PretrainedConfig):

384 super().__init__(**kwargs)

385

386 @property

387 - def per_layer_config(self) -> list[PretrainedConfig]:

387 + def per_layer_config(self) -> list[PreTrainedConfig]:

388 layers = []

389 for i in range(self.num_hidden_layers):

390 layer = copy(self)

tests/config/test_multimodal_config.py

+2 -2

@@ -6,7 +6,7 @@ from unittest.mock import MagicMock, patch

6

7 import pytest

8 import torch

9 -from transformers import PretrainedConfig

9 +from transformers import PreTrainedConfig

10

11 from tests.models.utils import build_model_context

12 from vllm.config.ec_transfer import ECRole, ECTransferConfig

@@ -136,7 +136,7 @@ def test_language_model_only_disables_via_supports_multimodal_inputs():

136

137

138 def test_convertor_clears_mm_prefix_when_multimodal_disabled():

139 - hf_config = PretrainedConfig(

139 + hf_config = PreTrainedConfig(

140 model_type="gemma3",

141 architectures=["Gemma3ForConditionalGeneration"],

142 )

tests/config/test_speculative_draft_hf_overrides.py

+6 -6

@@ -15,20 +15,20 @@ import functools

15 from unittest.mock import MagicMock, patch

16

17 import pytest

18 -from transformers import PretrainedConfig

18 +from transformers import PreTrainedConfig

19

20 from vllm.config.parallel import ParallelConfig

21 from vllm.config.speculative import SpeculativeConfig

22

23

24 -def _make_hf_config(**kwargs) -> PretrainedConfig:

24 +def _make_hf_config(**kwargs) -> PreTrainedConfig:

25 defaults = dict(

26 architectures=["LlamaForCausalLM"],

27 model_type="llama",

28 num_hidden_layers=64,

29 )

30 defaults.update(kwargs)

31 - return PretrainedConfig(**defaults)

31 + return PreTrainedConfig(**defaults)

32

33

34 @pytest.mark.cpu_test

@@ -137,7 +137,7 @@ def test_callable_overrides_reach_the_draft_config():

137 """A callable override (config-to-config transform) composes with the

138 architecture-mapping override and is applied to the draft config."""

139

140 - def shrink(hf_config: PretrainedConfig) -> PretrainedConfig:

140 + def shrink(hf_config: PreTrainedConfig) -> PreTrainedConfig:

141 hf_config.num_hidden_layers = 1

142 return hf_config

143

@@ -155,7 +155,7 @@ def test_arch_mapping_applies_before_callable_override():

155 observes (and may adjust) the post-mapping config."""

156 seen_architectures: list[str] = []

157

158 - def record(hf_config: PretrainedConfig) -> PretrainedConfig:

158 + def record(hf_config: PreTrainedConfig) -> PreTrainedConfig:

159 seen_architectures.append(hf_config.architectures[0])

160 return hf_config

161

@@ -202,7 +202,7 @@ def test_inkling_override_exposes_all_mtp_depths():

202 assert out.local_layer_ids == [0, 2, 4]

203

204

205 -def _module_level_shrink(hf_config: PretrainedConfig) -> PretrainedConfig:

205 +def _module_level_shrink(hf_config: PreTrainedConfig) -> PreTrainedConfig:

206 hf_config.num_hidden_layers = 1

207 return hf_config

208

tests/config/test_speculative_draft_max_position_embeddings.py

+4 -4

@@ -17,7 +17,7 @@ import logging

17 from typing import Literal

18

19 import pytest

20 -from transformers import PretrainedConfig

20 +from transformers import PreTrainedConfig

21

22 from vllm.config.model import ModelConfig

23 from vllm.config.parallel import ParallelConfig

@@ -46,7 +46,7 @@ def _override_logged(caplog: pytest.LogCaptureFixture) -> bool:

46

47 @pytest.mark.cpu_test

48 def test_override_raises_smaller_value(vllm_caplog: pytest.LogCaptureFixture):

49 - hf_config = PretrainedConfig(max_position_embeddings=2048)

49 + hf_config = PreTrainedConfig(max_position_embeddings=2048)

50 SpeculativeConfig._maybe_override_draft_max_position_embeddings(

51 hf_config, target_max_model_len=8192

52 )

@@ -56,7 +56,7 @@ def test_override_raises_smaller_value(vllm_caplog: pytest.LogCaptureFixture):

56

57 @pytest.mark.cpu_test

58 def test_override_keeps_sufficient_value(vllm_caplog: pytest.LogCaptureFixture):

59 - hf_config = PretrainedConfig(max_position_embeddings=8192)

59 + hf_config = PreTrainedConfig(max_position_embeddings=8192)

60 SpeculativeConfig._maybe_override_draft_max_position_embeddings(

61 hf_config, target_max_model_len=8192

62 )

@@ -66,7 +66,7 @@ def test_override_keeps_sufficient_value(vllm_caplog: pytest.LogCaptureFixture):

66

67 @pytest.mark.cpu_test

68 def test_override_ignores_missing_attribute(vllm_caplog: pytest.LogCaptureFixture):

69 - hf_config = PretrainedConfig()

69 + hf_config = PreTrainedConfig()

70 hf_config.__dict__.pop("max_position_embeddings", None)

71 SpeculativeConfig._maybe_override_draft_max_position_embeddings(

72 hf_config, target_max_model_len=8192

tests/entrypoints/pooling/scoring/test_jina_ranking_io_processor_unit.py

+2 -2

@@ -7,7 +7,7 @@ from unittest.mock import MagicMock

7

8 import pytest

9 from tokenizers import Tokenizer, models, pre_tokenizers

10 -from transformers import PreTrainedTokenizerFast

10 +from transformers import TokenizersBackend

11

12 from vllm import PoolingParams

13 from vllm.entrypoints.pooling.base.io_processor import PoolingIOProcessor

@@ -70,7 +70,7 @@ def offline_processor_and_context():

70 backend = Tokenizer(models.WordLevel({"[UNK]": 0}, unk_token="[UNK]"))

71 backend.pre_tokenizer = pre_tokenizers.WhitespaceSplit()

72 proc = JinaRankingIOProcessor.__new__(JinaRankingIOProcessor)

73 - proc.tokenizer = PreTrainedTokenizerFast(tokenizer_object=backend)

73 + proc.tokenizer = TokenizersBackend(tokenizer_object=backend)

74 proc.model_config = SimpleNamespace(max_model_len=1024, is_encoder_decoder=False)

75 proc.renderer = SimpleNamespace(

76 default_cmpl_tok_params=TokenizeParams(max_total_tokens=1024)

tests/model_executor/layers/test_pooler_methods.py

+2 -2

@@ -9,7 +9,7 @@ from unittest.mock import patch

9 import numpy as np

10 import pytest

11 import torch

12 -from transformers import PretrainedConfig

12 +from transformers import PreTrainedConfig

13

14 from vllm.config import PoolerConfig, set_current_vllm_config

15 from vllm.model_executor.layers.pooler import PoolingParamsUpdate

@@ -315,7 +315,7 @@ def test_dispatch_seq_cls_honors_token_pooling_type(tok_pooling_type):

315 model_config=SimpleNamespace(

316 pooler_config=pooler_config,

317 head_dtype=None,

318 - hf_config=PretrainedConfig(num_labels=3),

318 + hf_config=PreTrainedConfig(num_labels=3),

319 ),

320 )

321 classifier = torch.nn.Linear(4, 3)

tests/model_executor/test_qwen3_omni.py

+2 -2

@@ -8,7 +8,7 @@ from unittest.mock import Mock, patch

8 import pytest

9 import torch

10 import torch.nn as nn

11 -from transformers import PretrainedConfig

11 +from transformers import PreTrainedConfig

12

13 from vllm.config import ParallelConfig

14 from vllm.multimodal.processing import InputProcessingContext

@@ -52,7 +52,7 @@ def print_input_ids(input_ids):

52 @pytest.fixture

53 def mock_qwen3_omni_config():

54 """Create a mock Qwen3OmniMoeThinker config."""

55 - config = Mock(spec=PretrainedConfig)

55 + config = Mock(spec=PreTrainedConfig)

56 # Token IDs from https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct/blob/main/tokenizer_config.json

57 config.audio_token_id = 151675 # <|audio_pad|>

58 config.video_token_id = 151656 # <|video_pad|>

tests/models/language/pooling/test_jina_embeddings_v5.py

+6 -6

@@ -14,7 +14,7 @@ from types import SimpleNamespace

14 from typing import cast

15

16 import pytest

17 -from transformers import PretrainedConfig

17 +from transformers import PreTrainedConfig

18

19 from vllm.config import ModelConfig

20 from vllm.model_executor.models.config import (

@@ -23,7 +23,7 @@ from vllm.model_executor.models.config import (

23 )

24

25

26 -def _model_config(hf_config: PretrainedConfig) -> ModelConfig:

26 +def _model_config(hf_config: PreTrainedConfig) -> ModelConfig:

27 """Minimal stand-in for ModelConfig; only hf_config is read."""

28 return cast(ModelConfig, SimpleNamespace(hf_config=hf_config))

29

@@ -42,7 +42,7 @@ def test_encoder_backbone_enables_bidirectional_attention():

42 EncoderOnlyAttention; JinaEmbeddingsV5Model then dispatches to the encoder

43 implementation.

44 """

45 - hf_config = PretrainedConfig(is_decoder=False)

45 + hf_config = PreTrainedConfig(is_decoder=False)

46

47 JinaEmbeddingsV5ModelConfig.verify_and_update_model_config(_model_config(hf_config))

48

@@ -55,13 +55,13 @@ def test_supported_decoder_backbone_is_accepted():

55

56 `-small` omits `is_decoder` entirely, so an absent attribute has to be

57 treated as a decoder. The first assertion pins that assumption: if

58 - PretrainedConfig ever gains an `is_decoder=False` default, this fails here

58 + PreTrainedConfig ever gains an `is_decoder=False` default, this fails here

59 rather than silently rejecting a supported checkpoint.

60 """

61 - absent = PretrainedConfig()

61 + absent = PreTrainedConfig()

62 assert not hasattr(absent, "is_decoder")

63

64 JinaEmbeddingsV5ModelConfig.verify_and_update_model_config(_model_config(absent))

65 JinaEmbeddingsV5ModelConfig.verify_and_update_model_config(

66 - _model_config(PretrainedConfig(is_decoder=True))

66 + _model_config(PreTrainedConfig(is_decoder=True))

67 )

tests/models/multimodal/processing/test_audioflamingo3.py

+4 -4

@@ -22,19 +22,19 @@ from unittest.mock import MagicMock

22 import numpy as np

23 import pytest

24 import torch

25 -from transformers import PretrainedConfig

25 +from transformers import PreTrainedConfig

26

27 from tests.models.registry import HF_EXAMPLE_MODELS

28 from vllm.config.multimodal import MultiModalDummyOptions

29

30

31 -class MockAudioFlamingo3Config(PretrainedConfig):

31 +class MockAudioFlamingo3Config(PreTrainedConfig):

32 model_type = "audioflamingo3"

33

34 def __init__(self, **kwargs):

35 super().__init__(**kwargs)

36 - self.audio_config = PretrainedConfig()

37 - self.text_config = PretrainedConfig()

36 + self.audio_config = PreTrainedConfig()

37 + self.text_config = PreTrainedConfig()

38

39

40 class MockAudioFlamingo3Processor:

tests/models/multimodal/processing/test_bailing_moe_v3_vl.py

+2 -2

@@ -8,7 +8,7 @@ import pytest

8 import torch

9 from PIL import Image

10 from tokenizers import Tokenizer, decoders, models, pre_tokenizers

11 -from transformers import PreTrainedTokenizerFast

11 +from transformers import TokenizersBackend

12

13 from vllm.model_executor.models.bailing_moe_v3_vl import (

14 IMAGE_PLACEHOLDER,

@@ -34,7 +34,7 @@ def processor():

34 add_prefix_space=False, use_regex=False

35 )

36 backend.decoder = decoders.ByteLevel()

37 - tokenizer = PreTrainedTokenizerFast(

37 + tokenizer = TokenizersBackend(

38 tokenizer_object=backend,

39 additional_special_tokens=[

40 "<|vision_start|>",

tests/models/multimodal/processing/test_h2ovl.py

+2 -2

@@ -6,7 +6,7 @@ from collections.abc import Mapping

6

7 import pytest

8 from PIL import Image

9 -from transformers import PretrainedConfig

9 +from transformers import PreTrainedConfig

10

11 from vllm.multimodal import MULTIMODAL_REGISTRY

12 from vllm.multimodal.image import rescale_image_size

@@ -17,7 +17,7 @@ from ...utils import build_model_context

17

18

19 def _get_expected_num_patches(

20 - config: PretrainedConfig,

20 + config: PreTrainedConfig,

21 image: Image.Image,

22 num_imgs: int,

23 min_num: int,

tests/models/multimodal/processing/test_internvl.py

+2 -2

@@ -6,7 +6,7 @@ from collections.abc import Mapping

6

7 import pytest

8 from PIL import Image

9 -from transformers import PretrainedConfig

9 +from transformers import PreTrainedConfig

10

11 from vllm.multimodal import MULTIMODAL_REGISTRY

12 from vllm.multimodal.image import rescale_image_size

@@ -17,7 +17,7 @@ from ...utils import build_model_context

17

18

19 def _get_expected_num_patches(

20 - config: PretrainedConfig,

20 + config: PreTrainedConfig,

21 image: Image.Image,

22 num_imgs: int,

23 min_num: int,

tests/models/multimodal/processing/test_nemotron_vl.py

+2 -2

@@ -6,7 +6,7 @@ from collections.abc import Mapping

6

7 import pytest

8 from PIL import Image

9 -from transformers import PretrainedConfig

9 +from transformers import PreTrainedConfig

10

11 from vllm.multimodal import MULTIMODAL_REGISTRY

12 from vllm.multimodal.image import rescale_image_size

@@ -17,7 +17,7 @@ from ...utils import build_model_context

17

18

19 def _get_expected_num_patches(

20 - config: PretrainedConfig,

20 + config: PreTrainedConfig,

21 image: Image.Image,

22 num_imgs: int,

23 min_num: int,

tests/models/registry.py

+2 -2

@@ -7,7 +7,7 @@ from typing import Any, Literal

7

8 import pytest

9 from packaging.version import Version

10 -from transformers import PretrainedConfig

10 +from transformers import PreTrainedConfig

11 from transformers import __version__ as TRANSFORMERS_VERSION

12

13 from vllm.config.model import ModelDType, TokenizerMode

@@ -1233,7 +1233,7 @@ _MULTIMODAL_EXAMPLE_MODELS = {

1233 # NemotronH layers are constructed via `hybrid_override_pattern`

1234 use_original_num_layers=True,

1235 hf_overrides={

1236 - "vision_config": PretrainedConfig(

1236 + "vision_config": PreTrainedConfig(

1237 args={

1238 "min_num_patches": 1,

1239 "max_num_patches": 12,

tests/models/test_adapters.py

+3 -3

@@ -4,7 +4,7 @@

4

5 import pytest

6 import torch

7 -from transformers import Gemma3Config, PretrainedConfig, Qwen2Config

7 +from transformers import Gemma3Config, PreTrainedConfig, Qwen2Config

8

9 from vllm.model_executor.models.adapters import (

10 _create_pooling_model_cls,

@@ -266,14 +266,14 @@ def test_resolve_num_labels_defaults_when_undeclared():

266 config = _composite_config()

267 assert (

268 _resolve_num_labels(config, config.get_text_config())

269 - == PretrainedConfig().num_labels

269 + == PreTrainedConfig().num_labels

270 )

271

272

273 def test_resolve_num_labels_declared_on_outer_config():

274 """Multimodal checkpoints keep id2label/problem_type on the top-level config."""

275 config = _composite_config(outer_labels=20)

276 - assert config.get_text_config().num_labels == PretrainedConfig().num_labels

276 + assert config.get_text_config().num_labels == PreTrainedConfig().num_labels

277 assert _resolve_num_labels(config, config.get_text_config()) == 20

278

279

tests/models/test_qwen3_5_mtp_config.py

+5 -5

@@ -5,7 +5,7 @@

5 from typing import Any

6

7 import pytest

8 -from transformers import AutoConfig, PretrainedConfig

8 +from transformers import AutoConfig, PreTrainedConfig

9

10 from vllm.config.speculative import SpeculativeConfig

11

@@ -15,19 +15,19 @@ _CHECKPOINTS = {

15 }

16

17

18 -def _mtp_config(model_type: str) -> PretrainedConfig:

18 +def _mtp_config(model_type: str) -> PreTrainedConfig:

19 """Create a top-level MTP configuration with mtp_num_hidden_layers."""

20 kwargs: dict[str, Any] = {

21 "model_type": model_type,

22 "architectures": ["SomeArch"],

23 "mtp_num_hidden_layers": 1,

24 }

25 - return PretrainedConfig(**kwargs)

25 + return PreTrainedConfig(**kwargs)

26

27

28 def _multimodal_wrapper_mtp_config(

29 model_type: str, mtp_layers: int = 1

30 -) -> PretrainedConfig:

30 +) -> PreTrainedConfig:

31 """Download a multimodal wrapper config via AutoConfig.from_pretrained

32 and configure mtp_num_hidden_layers in text_config.

33

@@ -36,7 +36,7 @@ def _multimodal_wrapper_mtp_config(

36 - MoE (qwen3_5_moe): Qwen/Qwen3.6-35B-A3B

37 """

38 repo = _CHECKPOINTS[model_type]

39 - config: PretrainedConfig = AutoConfig.from_pretrained(repo)

39 + config: PreTrainedConfig = AutoConfig.from_pretrained(repo)

40 text_config = config.get_text_config()

41 text_config.mtp_num_hidden_layers = mtp_layers

42 return config

tests/models/transformers/test_backend.py

+2 -2

@@ -13,7 +13,7 @@ from typing import Any

13 import pytest

14 import torch

15 import torch.nn as nn

16 -from transformers import AutoConfig, AutoModel, PretrainedConfig

16 +from transformers import AutoConfig, AutoModel, PreTrainedConfig

17

18 from vllm.config import ModelConfig, VllmConfig

19 from vllm.model_executor.models.interfaces import SupportsMultiModal

@@ -524,7 +524,7 @@ def replace_vocab_embeddings(model, **config_kwargs):

524

525 stub = nn.Module()

526 stub.model = model

527 - stub.config = PretrainedConfig(

527 + stub.config = PreTrainedConfig(

528 vocab_size=VOCAB_SIZE, num_positions=NUM_POSITIONS, **config_kwargs

529 )

530 embeddings = Base._vocab_embeddings(stub)

tests/models/utils.py

+3 -3

@@ -8,7 +8,7 @@ from typing import Any

8

9 import torch

10 import torch.nn.functional as F

11 -from transformers import PretrainedConfig

11 +from transformers import PreTrainedConfig

12

13 from vllm.config.model import AttnTypeStr, ModelConfig, ModelDType, RunnerOption

14 from vllm.config.pooler import SequencePoolingType, TokenPoolingType

@@ -446,12 +446,12 @@ def get_vllm_extra_kwargs(model_info: ModelInfo, vllm_extra_kwargs):

446

447

448 def dummy_hf_overrides(

449 - hf_config: PretrainedConfig,

449 + hf_config: PreTrainedConfig,

450 *,

451 model_arch: str = "",

452 exist_overrides: dict[str, Any] | None = None,

453 use_original_num_layers: bool = False,

454 -) -> PretrainedConfig:

454 +) -> PreTrainedConfig:

455 """Dummy HF overrides function used to create dummy model

456 with only minimum nums of layer.

457 """

tests/test_config.py

+10 -10

@@ -62,16 +62,16 @@ def test_nested_rope_validation_patch_preserves_flat_rope_parameters(monkeypatch

62 def original_validate_rope(config, *args, **kwargs):

63 calls.append(config)

64

65 - from transformers import PretrainedConfig

65 + from transformers import PreTrainedConfig

66

67 - monkeypatch.setattr(PretrainedConfig, "validate_rope", original_validate_rope)

67 + monkeypatch.setattr(PreTrainedConfig, "validate_rope", original_validate_rope)

68 _patch_hf_transformers_nested_rope_validation()

69

70 nested_rope_parameters = {

71 "full_attention": {"rope_type": "default"},

72 "original_max_position_embeddings": 32768,

73 }

74 - PretrainedConfig.validate_rope(

74 + PreTrainedConfig.validate_rope(

75 SimpleNamespace(rope_parameters=nested_rope_parameters)

76 )

77 assert nested_rope_parameters == {"full_attention": {"rope_type": "default"}}

@@ -81,7 +81,7 @@ def test_nested_rope_validation_patch_preserves_flat_rope_parameters(monkeypatch

81 "factor": 8.0,

82 "rope_theta": 500000.0,

83 }

84 - PretrainedConfig.validate_rope(

84 + PreTrainedConfig.validate_rope(

85 SimpleNamespace(rope_parameters=flat_rope_parameters)

86 )

87 assert flat_rope_parameters == {

@@ -730,9 +730,9 @@ def test_batch_invariant_breakable_cudagraph(

730 ],

731 )

732 def test_dsa_models_select_matching_mtp(model_type, expected_architecture):

733 - from transformers import PretrainedConfig

733 + from transformers import PreTrainedConfig

734

735 - hf_config = PretrainedConfig(

735 + hf_config = PreTrainedConfig(

736 architectures=["DeepseekV32ForCausalLM"],

737 num_nextn_predict_layers=1,

738 )

@@ -2463,14 +2463,14 @@ def test_get_and_verify_max_len_with_nope_layers(

2463 max_model_len, rope_parameters, expected_max_len

2464 ):

2465 """NoPE layers do not prevent deriving or scaling the context length."""

2466 - from transformers import PretrainedConfig

2466 + from transformers import PreTrainedConfig

2467

2468 from vllm.config.model import _get_and_verify_max_len

2469 from vllm.transformers_utils.model_arch_config_convertor import (

2470 ModelArchConfigConvertorBase,

2471 )

2472

2473 - hf_config = PretrainedConfig(

2473 + hf_config = PreTrainedConfig(

2474 max_position_embeddings=4096,

2475 original_max_position_embeddings=2048,

2476 )

@@ -2514,14 +2514,14 @@ def test_get_and_verify_max_len_yarn_is_already_scaled(

2514 for every YaRN variant, so scaling it again overstates the limit and lets

2515 requests past the end of the cos/sin cache.

2516 """

2517 - from transformers import PretrainedConfig

2517 + from transformers import PreTrainedConfig

2518

2519 from vllm.config.model import _get_and_verify_max_len

2520 from vllm.transformers_utils.model_arch_config_convertor import (

2521 ModelArchConfigConvertorBase,

2522 )

2523

2524 - hf_config = PretrainedConfig(max_position_embeddings=32768)

2524 + hf_config = PreTrainedConfig(max_position_embeddings=32768)

2525 hf_config.rope_parameters = {

2526 "rope_type": rope_type,

2527 "factor": factor,

tests/transformers_utils/test_config.py

+6 -6

@@ -11,7 +11,7 @@ from typing import cast

11 from unittest.mock import MagicMock, patch

12

13 import pytest

14 -from transformers import PretrainedConfig

14 +from transformers import PreTrainedConfig

15

16 from vllm.config.model import ModelConfig

17 from vllm.tokenizers import get_tokenizer

@@ -223,7 +223,7 @@ def test_model_config_generation_fallback_forwards_code_revision():

223 patch.object(

224 config_module,

225 "get_config",

226 - return_value=PretrainedConfig(),

226 + return_value=PreTrainedConfig(),

227 ) as get_config,

228 ):

229 ModelConfig.try_get_generation_config(model_config)

@@ -269,7 +269,7 @@ def test_safetensors_metadata_of_repo_without_safetensors():

269 ],

270 )

271 def test_mrope_num_dims(section_key, mrope_section, expected_num_dims):

272 - config = PretrainedConfig()

272 + config = PreTrainedConfig()

273 config.rope_parameters = {"rope_type": "default", section_key: mrope_section}

274

275 assert uses_mrope(config)

@@ -280,7 +280,7 @@ def test_mrope_num_dims(section_key, mrope_section, expected_num_dims):

280 def test_mrope_num_dims_from_config_attribute(section_name):

281 """Some configs expose the section as an attribute rather than under

282 `rope_parameters`."""

283 - config = PretrainedConfig()

283 + config = PreTrainedConfig()

284 setattr(config, section_name, [16, 16, 16, 16])

285

286 assert uses_mrope(config)

@@ -289,7 +289,7 @@ def test_mrope_num_dims_from_config_attribute(section_name):

289

290 def test_mrope_num_dims_from_nested_rope_parameters():

291 """Sections nested by layer type must be found, not silently defaulted."""

292 - config = PretrainedConfig()

292 + config = PreTrainedConfig()

293 config.rope_parameters = {

294 "full_attention": {"mrope_section": [16, 16, 16, 16]},

295 "linear_attention": {"rope_type": "default"},

@@ -300,4 +300,4 @@ def test_mrope_num_dims_from_nested_rope_parameters():

300

301

302 def test_mrope_num_dims_without_mrope():

303 - assert mrope_num_dims(PretrainedConfig()) == 0

303 + assert mrope_num_dims(PreTrainedConfig()) == 0

tests/transformers_utils/test_config_parser_registry.py

+2 -2

@@ -4,7 +4,7 @@

4 from pathlib import Path

5

6 import pytest

7 -from transformers import PretrainedConfig

7 +from transformers import PreTrainedConfig

8

9 from vllm.transformers_utils.config import get_config_parser, register_config_parser

10 from vllm.transformers_utils.config_parser_base import ConfigParserBase

@@ -19,7 +19,7 @@ class CustomConfigParser(ConfigParserBase):

19 revision: str | None = None,

20 code_revision: str | None = None,

21 **kwargs,

22 - ) -> tuple[dict, PretrainedConfig]:

22 + ) -> tuple[dict, PreTrainedConfig]:

23 raise NotImplementedError

24

25

tests/transformers_utils/test_hf_overrides_model_type.py

+2 -2

@@ -5,12 +5,12 @@

5 import json

6 import tempfile

7

8 -from transformers import PretrainedConfig

8 +from transformers import PreTrainedConfig

9

10 from vllm.transformers_utils.config import _CONFIG_REGISTRY, get_config

11

12

13 -class _TestCustomConfig(PretrainedConfig):

13 +class _TestCustomConfig(PreTrainedConfig):

14 model_type = "test_custom_model"

15

16 def __init__(self, custom_attr=42, **kw):

tests/transformers_utils/test_speculators_override.py

+1 -1

@@ -40,7 +40,7 @@ _SPECULATORS_CONFIG_DICT = {

40 def _patched_get_config_dict(config_dict):

41 """Return a patch context that bypasses HF Hub for get_config_dict."""

42 return patch(

43 - "vllm.transformers_utils.config.PretrainedConfig.get_config_dict",

43 + "vllm.transformers_utils.config.PreTrainedConfig.get_config_dict",

44 return_value=(config_dict, {}),

45 )

46

tests/v1/spec_decode/test_extract_hidden_states.py

+6 -6

@@ -7,7 +7,7 @@ from unittest import mock

7 import numpy as np

8 import pytest

9 import torch

10 -from transformers import CLIPVisionConfig, LlamaConfig, LlavaConfig, PretrainedConfig

10 +from transformers import CLIPVisionConfig, LlamaConfig, LlavaConfig, PreTrainedConfig

11

12 from tests.v1.attention.utils import (

13 BatchSpec,

@@ -336,7 +336,7 @@ def test_propose_different_layer_counts(num_hidden_layers):

336 # ---------------------------------------------------------------------------

337

338

339 -class _DummyVLMConfig(PretrainedConfig):

339 +class _DummyVLMConfig(PreTrainedConfig):

340 """Minimal composite config that mimics VLMs like Kimi-K2.5 or LLaVA.

341

342 The text model's parameters (hidden_size, num_attention_heads, …) live

@@ -345,11 +345,11 @@ class _DummyVLMConfig(PretrainedConfig):

345

346 model_type = "test_vlm"

347

348 - def __init__(self, text_config: PretrainedConfig, **kwargs):

348 + def __init__(self, text_config: PreTrainedConfig, **kwargs):

349 self.text_config = text_config

350 super().__init__(architectures=["LlamaForCausalLM"], **kwargs)

351

352 - def get_text_config(self, decoder: bool = False) -> PretrainedConfig:

352 + def get_text_config(self, decoder: bool = False) -> PreTrainedConfig:

353 del decoder

354 return self.text_config

355

@@ -403,7 +403,7 @@ def test_extract_hidden_states_config_preserves_vlm_text_config():

403 eagle_aux_hidden_state_layer_ids=[1, 2],

404 )

405

406 - # The fix: text_config is still a PretrainedConfig, not a dict.

406 + # The fix: text_config is still a PreTrainedConfig, not a dict.

407 assert isinstance(extract_config.text_config, LlamaConfig)

408

409 extracted = get_hf_text_config(extract_config)

@@ -473,7 +473,7 @@ def test_extract_hidden_states_speculative_config_vlm():

473

474 def test_extract_hidden_states_config_invalid_text_config():

475 """A nested text_config missing required attrs must still be rejected."""

476 - broken_text_config = PretrainedConfig(hidden_size=128)

476 + broken_text_config = PreTrainedConfig(hidden_size=128)

477 vlm_config = _DummyVLMConfig(text_config=broken_text_config)

478

479 extract_config = ExtractHiddenStatesConfig(

tools/pre_commit/check_forbidden_imports.py

+70

@@ -13,6 +13,16 @@ _HF_NAMES = (

13 r"|list_repo_refs|repo_exists"

14 )

15

16 +# Transformers v4 names kept as aliases in v5.

17 +_TRANSFORMERS_LEGACY_NAMES = (

18 + r"PretrainedConfig|PreTrainedTokenizer|PreTrainedTokenizerFast"

19 + r"|BaseImageProcessorFast|\w+ImageProcessorFast"

20 +)

21 +_TRANSFORMERS_LEGACY_MODULES = (

22 + r"tokenization_utils|tokenization_utils_fast|image_processing_utils_fast"

23 + r"|models\.\w+\.\w+_fast"

24 +)

25 +

26

27 @dataclass

28 class ForbiddenImport:

@@ -117,6 +127,26 @@ CHECK_IMPORTS = {

127 allowed_files={"vllm/transformers_utils/repo_utils.py"},

128 allowed_dirs={"examples/"},

129 ),

130 + "transformers legacy names": ForbiddenImport(

131 + pattern=(

132 + r"^\s*from\s+transformers(?:\.[\w.]+)?\s+import\s*\([^)]*\b(?:"

133 + + _TRANSFORMERS_LEGACY_NAMES

134 + + r")\b"

135 + r"|"

136 + r"^\s*from\s+transformers(?:\.[\w.]+)?\s+import\b[^\n]*\b(?:"

137 + + _TRANSFORMERS_LEGACY_NAMES

138 + + r")\b"

139 + r"|"

140 + r"^\s*(?:from|import)\s+transformers\.(?:"

141 + + _TRANSFORMERS_LEGACY_MODULES

142 + + r")\b"

143 + ),

144 + tip=(

145 + "Use the Transformers v5 names, e.g. PreTrainedConfig, PythonBackend, "

146 + "TokenizersBackend, TorchvisionBackend, tokenization_utils_tokenizers, "

147 + "tokenization_utils_sentencepiece and image_processing_backends."

148 + ),

149 + ),

150 }

151

152

@@ -247,6 +277,46 @@ def test_regex():

277 f"(expected {should_match}, got {result})"

278 )

279

280 + transformers_cases = [

281 + # Should match

282 + ("from transformers import PretrainedConfig", True),

283 + ("from transformers import AutoConfig, PretrainedConfig", True),

284 + ("from transformers import PreTrainedTokenizer", True),

285 + ("from transformers import PreTrainedTokenizerFast", True),

286 + ("from transformers.models.siglip import SiglipImageProcessorFast", True),

287 + ("from transformers import (\n PreTrainedTokenizerFast,\n)", True),

288 + ("from transformers.configuration_utils import PretrainedConfig", True),

289 + (" from transformers import PretrainedConfig", True),

290 + ("from transformers.tokenization_utils import AddedToken", True),

291 + ("from transformers.tokenization_utils_fast import X", True),

292 + ("from transformers.image_processing_utils_fast import X", True),

293 + (

294 + "from transformers.models.qwen2_vl.image_processing_qwen2_vl_fast import X",

295 + True,

296 + ),

297 + ("import transformers.tokenization_utils_fast", True),

298 + # Should not match

299 + ("from transformers import PreTrainedConfig", False),

300 + ("from transformers import PreTrainedTokenizerBase", False),

301 + ("from transformers import PythonBackend, TokenizersBackend", False),

302 + ("from transformers.tokenization_utils_base import BatchEncoding", False),

303 + ("from transformers.tokenization_utils_tokenizers import X", False),

304 + (

305 + "from transformers.image_processing_backends import TorchvisionBackend",

306 + False,

307 + ),

308 + ("from transformers.models.qwen2_vl import Qwen2VLImageProcessor", False),

309 + ("from vllm.foo import GLM4VImageProcessorFast", False),

310 + ("from transformers import (\n PreTrainedConfig,\n)", False),

311 + ("# from transformers import PretrainedConfig", False),

312 + ]

313 + for i, (content, should_match) in enumerate(transformers_cases):

314 + result = matches("transformers legacy names", content)

315 + assert result == should_match, (

316 + f"transformers case {i} failed: {content!r} "

317 + f"(expected {should_match}, got {result})"

318 + )

319 +

320 print("All regex tests passed.")

321

322

vllm/config/model.py

+12 -12

@@ -55,7 +55,7 @@ from vllm.utils.import_utils import LazyLoader

55 from vllm.v1.attention.backends.registry import AttentionBackendEnum

56

57 if TYPE_CHECKING:

58 - from transformers import PretrainedConfig

58 + from transformers import PreTrainedConfig

59

60 import vllm.model_executor.layers.quantization as me_quant

61 import vllm.model_executor.models as me_models

@@ -64,7 +64,7 @@ if TYPE_CHECKING:

64 from vllm.model_executor.layers.quantization import QuantizationMethods

65 from vllm.v1.sample.logits_processor import LogitsProcessor

66 else:

67 - PretrainedConfig = Any

67 + PreTrainedConfig = Any

68

69 me_quant = LazyLoader(

70 "model_executor", globals(), "vllm.model_executor.layers.quantization"

@@ -105,7 +105,7 @@ PROCESSED_LOGPROBS_MODES: tuple[LogprobsMode, ...] = (

105 "processed_logits",

106 "processed_logprobs",

107 )

108 -HfOverrides = dict[str, Any] | Callable[[PretrainedConfig], PretrainedConfig]

108 +HfOverrides = dict[str, Any] | Callable[[PreTrainedConfig], PreTrainedConfig]

109 ModelImpl = Literal["auto", "vllm", "transformers", "terratorch"]

110 LayerBlockType = Literal["attention", "linear_attention", "mamba"]

111

@@ -183,9 +183,9 @@ class ModelConfig:

183 We must set the global seed because otherwise,

184 different tensor parallel workers would sample different tokens,

185 leading to inconsistent results."""

186 - hf_config: PretrainedConfig = field(init=False)

186 + hf_config: PreTrainedConfig = field(init=False)

187 """The Hugging Face config of the model."""

188 - hf_text_config: PretrainedConfig = field(init=False)

188 + hf_text_config: PreTrainedConfig = field(init=False)

189 """The Hugging Face config of the text model (same as hf_config for text models)."""

190 is_submodel_config: bool = field(default=False, init=False)

191 """Whether this is a submodule view derived by `VllmConfig.with_hf_config`

@@ -483,7 +483,7 @@ class ModelConfig:

483

484 def _update_nested(

485 self,

486 - target: PretrainedConfig | dict[str, Any],

486 + target: PreTrainedConfig | dict[str, Any],

487 updates: dict[str, Any],

488 ) -> None:

489 """Recursively updates a config or dict with nested updates."""

@@ -511,15 +511,15 @@ class ModelConfig:

511

512 def _apply_dict_overrides(

513 self,

514 - config: PretrainedConfig,

514 + config: PreTrainedConfig,

515 overrides: dict[str, Any],

516 ) -> None:

517 """Apply dict overrides, handling both nested configs and dict values."""

518 - from transformers import PretrainedConfig

518 + from transformers import PreTrainedConfig

519

520 for key, value in overrides.items():

521 attr = getattr(config, key, None)

522 - if attr is not None and isinstance(attr, PretrainedConfig):

522 + if attr is not None and isinstance(attr, PreTrainedConfig):

523 # It's a nested config - recursively update it

524 self._update_nested(attr, value)

525 else:

@@ -2393,7 +2393,7 @@ def _resolve_auto_dtype(

2393

2394 def _get_and_verify_dtype(

2395 model_id: str,

2396 - config: PretrainedConfig,

2396 + config: PreTrainedConfig,

2397 dtype: str | torch.dtype,

2398 *,

2399 is_pooling_model: bool,

@@ -2440,7 +2440,7 @@ def _get_and_verify_dtype(

2440

2441

2442 def _get_head_dtype(

2443 - config: PretrainedConfig, dtype: torch.dtype, runner_type: str

2443 + config: PreTrainedConfig, dtype: torch.dtype, runner_type: str

2444 ) -> torch.dtype:

2445 head_dtype: str | torch.dtype | None = getattr(config, "head_dtype", None)

2446

@@ -2464,7 +2464,7 @@ def _get_head_dtype(

2464

2465

2466 def _get_and_verify_max_len(

2467 - hf_config: PretrainedConfig,

2467 + hf_config: PreTrainedConfig,

2468 model_arch_config: ModelArchitectureConfig,

2469 tokenizer_config: dict | None,

2470 max_model_len: int | None,

vllm/config/speculative.py

+9 -9

@@ -22,12 +22,12 @@ from vllm.utils.import_utils import LazyLoader, has_arctic_inference

22 from vllm.v1.attention.backends.registry import AttentionBackendEnum

23

24 if TYPE_CHECKING:

25 - from transformers import PretrainedConfig

25 + from transformers import PreTrainedConfig

26

27 import vllm.model_executor.layers.quantization as me_quant

28 from vllm.config.vllm import VllmConfig

29 else:

30 - PretrainedConfig = Any

30 + PreTrainedConfig = Any

31 VllmConfig = Any

32

33 me_quant = LazyLoader(

@@ -657,7 +657,7 @@ class SpeculativeConfig:

657 return hash_str

658

659 @staticmethod

660 - def hf_config_override(hf_config: PretrainedConfig) -> PretrainedConfig:

660 + def hf_config_override(hf_config: PreTrainedConfig) -> PreTrainedConfig:

661 initial_architecture = hf_config.architectures[0]

662 use_v32_mtp = hf_config.model_type in ("deepseek_v32", "glm_moe_dsa")

663 if hf_config.model_type == "dots3_note":

@@ -1066,16 +1066,16 @@ class SpeculativeConfig:

1066

1067 @staticmethod

1068 def _apply_composed_hf_override(

1069 - target_hf_overrides: Callable[[PretrainedConfig], PretrainedConfig],

1070 - hf_config: PretrainedConfig,

1071 - ) -> PretrainedConfig:

1069 + target_hf_overrides: Callable[[PreTrainedConfig], PreTrainedConfig],

1070 + hf_config: PreTrainedConfig,

1071 + ) -> PreTrainedConfig:

1072 hf_config = SpeculativeConfig.hf_config_override(hf_config)

1073 return target_hf_overrides(hf_config)

1074

1075 @staticmethod

1076 def compose_draft_hf_overrides(

1077 target_hf_overrides: HfOverrides | None,

1078 - ) -> Callable[[PretrainedConfig], PretrainedConfig]:

1078 + ) -> Callable[[PreTrainedConfig], PreTrainedConfig]:

1079 """Build the ``hf_overrides`` for the draft ``ModelConfig``.

1080

1081 Callable overrides on the target are config-to-config transforms

@@ -1667,7 +1667,7 @@ class SpeculativeConfig:

1667

1668 @staticmethod

1669 def _maybe_override_draft_max_position_embeddings(

1670 - draft_hf_config: PretrainedConfig,

1670 + draft_hf_config: PreTrainedConfig,

1671 target_max_model_len: int,

1672 ) -> None:

1673 """Raise an EAGLE draft's max_position_embeddings up to the target's.

@@ -1704,7 +1704,7 @@ class SpeculativeConfig:

1704 def _verify_and_get_draft_tp(

1705 target_parallel_config: ParallelConfig,

1706 speculative_draft_tensor_parallel_size: int | None,

1707 - draft_hf_config: PretrainedConfig,

1707 + draft_hf_config: PreTrainedConfig,

1708 ) -> int:

1709 """Verifies and adjusts the tensor parallel size for a draft model

1710 specified using speculative_draft_tensor_parallel_size.

vllm/config/utils.py

+2 -2

@@ -138,7 +138,7 @@ def getattr_iter(

138 ) -> Any:

139 """A helper function that retrieves an attribute from an object which may

140 have multiple possible names. This is useful when fetching attributes from

141 - arbitrary `transformers.PretrainedConfig` instances.

141 + arbitrary `transformers.PreTrainedConfig` instances.

142

143 In the case where the first name in `names` is the preferred name, and

144 any other names are deprecated aliases, setting `warn=True` will log a

@@ -314,7 +314,7 @@ def normalize_value(x):

314 except Exception:

315 return str(x)

316

317 - # PretrainedConfig (must be before dataclass branch as these are now dataclasses)

317 + # PreTrainedConfig (must be before dataclass branch as these are now dataclasses)

318 if hasattr(x, "to_json_string") and callable(x.to_json_string):

319 try:

320 return x.to_json_string()

vllm/config/vllm.py

+3 -3

@@ -56,12 +56,12 @@ from .watermarking import WatermarkConfig

56 from .weight_transfer import WeightTransferConfig

57

58 if TYPE_CHECKING:

59 - from transformers import PretrainedConfig

59 + from transformers import PreTrainedConfig

60

61 from vllm.model_executor.layers.quantization.base_config import QuantizationConfig

62 from vllm.v1.kv_cache_interface import KVCacheConfig

63 else:

64 - PretrainedConfig = Any

64 + PreTrainedConfig = Any

65

66 QuantizationConfig = Any

67

@@ -939,7 +939,7 @@ class VllmConfig:

939

940 def with_hf_config(

941 self,

942 - hf_config: PretrainedConfig,

942 + hf_config: PreTrainedConfig,

943 architectures: list[str] | None = None,

944 ) -> "VllmConfig":

945 if architectures is not None:

vllm/lora/layers/base.py

+3 -3

@@ -6,7 +6,7 @@ from typing import TYPE_CHECKING, overload

6

7 import torch

8 import torch.nn as nn

9 -from transformers import PretrainedConfig

9 +from transformers import PreTrainedConfig

10

11 from vllm.config.lora import LoRAConfig

12

@@ -73,7 +73,7 @@ class BaseLayerWithLoRA(nn.Module):

73 self,

74 max_loras: int,

75 lora_config: LoRAConfig,

76 - model_config: PretrainedConfig | None = None,

76 + model_config: PreTrainedConfig | None = None,

77 ) -> None:

78 """Initializes lora matrices."""

79 ...

@@ -103,7 +103,7 @@ class BaseLayerWithLoRA(nn.Module):

103 source_layer: nn.Module,

104 lora_config: LoRAConfig,

105 packed_modules_list: list,

106 - model_config: PretrainedConfig | None = None,

106 + model_config: PreTrainedConfig | None = None,

107 ) -> bool:

108 """Returns True if the layer can be replaced by this LoRA layer."""

109 raise NotImplementedError

vllm/lora/layers/base_linear.py

+2 -2

@@ -3,7 +3,7 @@

3

4

5 import torch

6 -from transformers import PretrainedConfig

6 +from transformers import PreTrainedConfig

7

8 from vllm import envs

9 from vllm.config import get_current_vllm_config

@@ -105,7 +105,7 @@ class BaseLinearLayerWithLoRA(BaseLayerWithLoRA):

105 self,

106 max_loras: int,

107 lora_config: LoRAConfig,

108 - model_config: PretrainedConfig | None = None,

108 + model_config: PreTrainedConfig | None = None,

109 ) -> None:

110 self.lora_config = lora_config

111 if isinstance(self.base_layer, ReplicatedLinear):

vllm/lora/layers/classifier.py

+2 -2

@@ -3,7 +3,7 @@

3

4

5 import torch

6 -from transformers import PretrainedConfig

6 +from transformers import PreTrainedConfig

7

8 from vllm.config.lora import LoRAConfig

9 from vllm.platforms import current_platform

@@ -16,7 +16,7 @@ class ClassificationHeadWithLoRA(ReplicatedLinearWithLoRA):

16 self,

17 max_loras: int,

18 lora_config: LoRAConfig,

19 - model_config: PretrainedConfig | None = None,

19 + model_config: PreTrainedConfig | None = None,

20 ) -> None:

21 # Preserve ordinary LoRA A/B support for classification heads.

22 super().create_lora_weights(max_loras, lora_config, model_config)

vllm/lora/layers/column_parallel_linear.py

+12 -12

@@ -4,7 +4,7 @@

4

5 import torch

6 import torch.nn as nn

7 -from transformers import PretrainedConfig

7 +from transformers import PreTrainedConfig

8

9 from vllm.config.lora import LoRAConfig

10 from vllm.distributed import tensor_model_parallel_all_gather

@@ -164,7 +164,7 @@ class ColumnParallelLinearWithLoRA(BaseLinearLayerWithLoRA):

164 source_layer: nn.Module,

165 lora_config: LoRAConfig,

166 packed_modules_list: list,

167 - model_config: PretrainedConfig | None = None,

167 + model_config: PreTrainedConfig | None = None,

168 ) -> bool:

169 if type(source_layer) is maybe_get_oot_by_class(ColumnParallelLinear):

170 return True

@@ -208,7 +208,7 @@ class MergedColumnParallelLinearWithLoRA(ColumnParallelLinearWithLoRA):

208 self,

209 max_loras: int,

210 lora_config: LoRAConfig,

211 - model_config: PretrainedConfig | None = None,

211 + model_config: PreTrainedConfig | None = None,

212 ) -> None:

213 """The main reason for overriding this function is to enhance code

214 maintainability.

@@ -366,7 +366,7 @@ class MergedColumnParallelLinearWithLoRA(ColumnParallelLinearWithLoRA):

366 source_layer: nn.Module,

367 lora_config: LoRAConfig,

368 packed_modules_list: list,

369 - model_config: PretrainedConfig | None = None,

369 + model_config: PreTrainedConfig | None = None,

370 decorate: bool = True,

371 ) -> bool:

372 merged_cls = maybe_get_oot_by_class(MergedColumnParallelLinear)

@@ -442,7 +442,7 @@ class QKVParallelLinearWithLoRA(ColumnParallelLinearWithLoRA):

442 source_layer: nn.Module,

443 lora_config: LoRAConfig,

444 packed_modules_list: list,

445 - model_config: PretrainedConfig | None = None,

445 + model_config: PreTrainedConfig | None = None,

446 ) -> bool:

447 return (

448 type(source_layer) is maybe_get_oot_by_class(QKVParallelLinear)

@@ -488,7 +488,7 @@ class MergedQKVParallelLinearWithLoRA(MergedColumnParallelLinearWithLoRA):

488 self,

489 max_loras: int,

490 lora_config: LoRAConfig,

491 - model_config: PretrainedConfig | None = None,

491 + model_config: PreTrainedConfig | None = None,

492 ) -> None:

493 """The main reason for overloading this function is to handle inconsistent

494 weight dimensions in qkv lora.

@@ -502,7 +502,7 @@ class MergedQKVParallelLinearWithLoRA(MergedColumnParallelLinearWithLoRA):

502 source_layer: nn.Module,

503 lora_config: LoRAConfig,

504 packed_modules_list: list,

505 - model_config: PretrainedConfig | None = None,

505 + model_config: PreTrainedConfig | None = None,

506 decorate: bool = True,

507 ) -> bool:

508 return (

@@ -543,7 +543,7 @@ class ColumnParallelLinearWithShardedLoRA(ColumnParallelLinearWithLoRA):

543 source_layer: nn.Module,

544 lora_config: LoRAConfig,

545 packed_modules_list: list,

546 - model_config: PretrainedConfig | None = None,

546 + model_config: PreTrainedConfig | None = None,

547 decorate: bool = True,

548 ) -> bool:

549 # specifying kwargs so they can be easily accessed in decorator

@@ -585,7 +585,7 @@ class MergedColumnParallelLinearWithShardedLoRA(MergedColumnParallelLinearWithLo

585 source_layer: nn.Module,

586 lora_config: LoRAConfig,

587 packed_modules_list: list,

588 - model_config: PretrainedConfig | None = None,

588 + model_config: PreTrainedConfig | None = None,

589 decorate: bool = True,

590 ) -> bool:

591 # specifying kwargs so they can be easily accessed in decorator

@@ -621,7 +621,7 @@ class QKVParallelLinearWithShardedLoRA(QKVParallelLinearWithLoRA):

621 source_layer: nn.Module,

622 lora_config: LoRAConfig,

623 packed_modules_list: list,

624 - model_config: PretrainedConfig | None = None,

624 + model_config: PreTrainedConfig | None = None,

625 ) -> bool:

626 # specifying kwargs so they can be easily accessed in decorator

627 return super().can_replace_layer(

@@ -669,7 +669,7 @@ class MergedQKVParallelLinearWithShardedLoRA(MergedQKVParallelLinearWithLoRA):

669 source_layer: nn.Module,

670 lora_config: LoRAConfig,

671 packed_modules_list: list,

672 - model_config: PretrainedConfig | None = None,

672 + model_config: PreTrainedConfig | None = None,

673 decorate: bool = True,

674 ) -> bool:

675 # specifying kwargs so they can be easily accessed in decorator

@@ -698,7 +698,7 @@ class MergedColumnParallelLinearVariableSliceWithLoRA(

698 source_layer: nn.Module,

699 lora_config: LoRAConfig,

700 packed_modules_list: list,

701 - model_config: PretrainedConfig | None = None,

701 + model_config: PreTrainedConfig | None = None,

702 decorate: bool = True,

703 ) -> bool:

704 # Support MergedColumnParallelLinear with 3 or more slices

vllm/lora/layers/fused_moe.py

+5 -5

@@ -3,7 +3,7 @@

3

4 import torch

5 import torch.nn as nn

6 -from transformers import PretrainedConfig

6 +from transformers import PreTrainedConfig

7

8 from vllm import envs

9 from vllm.config.lora import LoRAConfig

@@ -245,7 +245,7 @@ class FusedMoEWithLoRA(BaseLayerWithLoRA):

245 self,

246 max_loras: int,

247 lora_config: LoRAConfig,

248 - model_config: PretrainedConfig | None = None,

248 + model_config: PreTrainedConfig | None = None,

249 ) -> None:

250 """Initializes lora matrices."""

251 self._verify_ep_fs(lora_config)

@@ -465,7 +465,7 @@ class FusedMoEWithLoRA(BaseLayerWithLoRA):

465 source_layer: nn.Module,

466 lora_config: LoRAConfig,

467 packed_modules_list: list,

468 - model_config: PretrainedConfig | None = None,

468 + model_config: PreTrainedConfig | None = None,

469 ) -> bool:

470 """Returns True if the layer can be replaced by this LoRA layer."""

471 # source_layer is MoERunner

@@ -511,7 +511,7 @@ class FusedMoE3DWithLoRA(FusedMoEWithLoRA):

511 self,

512 max_loras: int,

513 lora_config: LoRAConfig,

514 - model_config: PretrainedConfig | None = None,

514 + model_config: PreTrainedConfig | None = None,

515 ) -> None:

516 """Initializes lora matrices."""

517 if model_config is None:

@@ -624,7 +624,7 @@ class FusedMoE3DWithLoRA(FusedMoEWithLoRA):

624 source_layer: nn.Module,

625 lora_config: LoRAConfig,

626 packed_modules_list: list,

627 - model_config: PretrainedConfig | None = None,

627 + model_config: PreTrainedConfig | None = None,

628 ) -> bool:

629 """Returns True if the layer can be replaced by this LoRA layer."""

630 # source_layer is MoERunner

vllm/lora/layers/logits_processor.py

+3 -3

@@ -3,7 +3,7 @@

3

4 import torch

5 import torch.nn as nn

6 -from transformers import PretrainedConfig

6 +from transformers import PreTrainedConfig

7

8 from vllm.config.lora import LoRAConfig

9 from vllm.distributed import (

@@ -85,7 +85,7 @@ class LogitsProcessorWithLoRA(BaseLayerWithLoRA):

85 self,

86 max_loras: int,

87 lora_config: LoRAConfig,

88 - model_config: PretrainedConfig | None = None,

88 + model_config: PreTrainedConfig | None = None,

89 ) -> None:

90 # TODO: Verify if this condition can be further relaxed

91 if self.base_layer.vocab_size > 258048:

@@ -221,7 +221,7 @@ class LogitsProcessorWithLoRA(BaseLayerWithLoRA):

221 source_layer: nn.Module,

222 lora_config: LoRAConfig,

223 packed_modules_list: list,

224 - model_config: PretrainedConfig | None = None,

224 + model_config: PreTrainedConfig | None = None,

225 ) -> bool:

226 # Special handling for the LogitsProcessor.

227 return False

vllm/lora/layers/replicated_linear.py

+2 -2

@@ -4,7 +4,7 @@

4

5 import torch

6 import torch.nn as nn

7 -from transformers import PretrainedConfig

7 +from transformers import PreTrainedConfig

8

9 from vllm.config.lora import LoRAConfig

10 from vllm.model_executor.custom_op import maybe_get_oot_by_class

@@ -61,7 +61,7 @@ class ReplicatedLinearWithLoRA(BaseLinearLayerWithLoRA):

61 source_layer: nn.Module,

62 lora_config: LoRAConfig,

63 packed_modules_list: list,

64 - model_config: PretrainedConfig | None = None,

64 + model_config: PreTrainedConfig | None = None,

65 ) -> bool:

66 return isinstance(source_layer, maybe_get_oot_by_class(ReplicatedLinear))

67

vllm/lora/layers/row_parallel_linear.py

+3 -3

@@ -4,7 +4,7 @@

4

5 import torch

6 import torch.nn as nn

7 -from transformers import PretrainedConfig

7 +from transformers import PreTrainedConfig

8

9 from vllm.config.lora import LoRAConfig

10 from vllm.distributed import (

@@ -89,7 +89,7 @@ class RowParallelLinearWithLoRA(BaseLinearLayerWithLoRA):

89 source_layer: nn.Module,

90 lora_config: LoRAConfig,

91 packed_modules_list: list,

92 - model_config: PretrainedConfig | None = None,

92 + model_config: PreTrainedConfig | None = None,

93 ) -> bool:

94 return type(source_layer) is maybe_get_oot_by_class(RowParallelLinear)

95

@@ -165,7 +165,7 @@ class RowParallelLinearWithShardedLoRA(RowParallelLinearWithLoRA):

165 source_layer: nn.Module,

166 lora_config: LoRAConfig,

167 packed_modules_list: list,

168 - model_config: PretrainedConfig | None = None,

168 + model_config: PreTrainedConfig | None = None,

169 ) -> bool:

170 # specifying kwargs so they can be easily accessed in decorator

171 return super().can_replace_layer(

vllm/lora/layers/vocab_parallel_embedding.py

+3 -3

@@ -4,7 +4,7 @@

4 import torch

5 import torch.nn as nn

6 import torch.nn.functional as F

7 -from transformers import PretrainedConfig

7 +from transformers import PreTrainedConfig

8

9 from vllm.config.lora import LoRAConfig

10 from vllm.model_executor.custom_op import maybe_get_oot_by_class

@@ -26,7 +26,7 @@ class VocabParallelEmbeddingWithLoRA(BaseLayerWithLoRA):

26 self,

27 max_loras: int,

28 lora_config: LoRAConfig,

29 - model_config: PretrainedConfig | None = None,

29 + model_config: PreTrainedConfig | None = None,

30 ) -> None:

31 self.lora_a_stacked = torch.zeros(

32 (

@@ -113,7 +113,7 @@ class VocabParallelEmbeddingWithLoRA(BaseLayerWithLoRA):

113 source_layer: nn.Module,

114 lora_config: LoRAConfig,

115 packed_modules_list: list,

116 - model_config: PretrainedConfig | None = None,

116 + model_config: PreTrainedConfig | None = None,

117 ) -> bool:

118 # Accept both the in-tree and any OOT class, because layers built on top of

119 # `VocabParallelEmbedding` (e.g. by the Transformers modelling backend) use the

vllm/lora/utils.py

+4 -4

@@ -6,7 +6,7 @@ from typing import TYPE_CHECKING

6

7 from huggingface_hub.utils import HfHubHTTPError, HFValidationError

8 from torch import nn

9 -from transformers import PretrainedConfig

9 +from transformers import PreTrainedConfig

10

11 from vllm import envs

12 from vllm.config.lora import LoRAConfig

@@ -109,7 +109,7 @@ def from_layer(

109 max_loras: int,

110 lora_config: LoRAConfig,

111 packed_modules_list: list,

112 - model_config: PretrainedConfig | None = None,

112 + model_config: PreTrainedConfig | None = None,

113 ) -> nn.Module:

114 for lora_cls in _all_lora_classes:

115 # specifying kwargs so they can be easily accessed in decorator

@@ -130,7 +130,7 @@ def from_layer_logits_processor(

130 lm_head: "ParallelLMHead",

131 max_loras: int,

132 lora_config: LoRAConfig,

133 - model_config: PretrainedConfig | None = None,

133 + model_config: PreTrainedConfig | None = None,

134 ) -> LogitsProcessorWithLoRA:

135 ret = LogitsProcessorWithLoRA(

136 layer,

@@ -147,7 +147,7 @@ def from_layer_classification(

147 layer: nn.Module,

148 max_loras: int,

149 lora_config: LoRAConfig,

150 - model_config: PretrainedConfig | None = None,

150 + model_config: PreTrainedConfig | None = None,

151 ) -> ClassificationHeadWithLoRA:

152 instance_layer = ClassificationHeadWithLoRA(layer)

153 instance_layer.create_lora_weights(max_loras, lora_config, model_config)

vllm/model_executor/layers/mamba/gdn/base.py

+2 -2

@@ -1,7 +1,7 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 import torch

4 -from transformers import PretrainedConfig

4 +from transformers import PreTrainedConfig

5

6 from vllm.config import (

7 VllmConfig,

@@ -24,7 +24,7 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):

24

25 def __init__(

26 self,

27 - config: PretrainedConfig,

27 + config: PreTrainedConfig,

28 vllm_config: VllmConfig,

29 prefix: str = "",

30 ) -> None:

vllm/model_executor/layers/mamba/linear/bailing_linear_attn.py

+3 -3

@@ -4,7 +4,7 @@ import copy

4

5 import torch

6 import torch.nn.functional as F

7 -from transformers.configuration_utils import PretrainedConfig

7 +from transformers.configuration_utils import PreTrainedConfig

8

9 from vllm.config import (

10 VllmConfig,

@@ -37,7 +37,7 @@ from vllm.triton_utils import tl, triton

37 from vllm.v1.attention.backends.linear_attn import LinearAttentionMetadata

38

39

40 -def _build_rope_parameters(config: PretrainedConfig) -> dict | None:

40 +def _build_rope_parameters(config: PreTrainedConfig) -> dict | None:

41 rope_parameters = copy.deepcopy(getattr(config, "rope_parameters", None)) or {}

42 if "rope_theta" not in rope_parameters and hasattr(config, "rope_theta"):

43 rope_parameters["rope_theta"] = config.rope_theta

@@ -385,7 +385,7 @@ class BailingMoELinearAttention(LinearAttention):

385 # --8<-- [end:bailing_moe_linear_attention]

386 def __init__(

387 self,

388 - config: PretrainedConfig,

388 + config: PreTrainedConfig,

389 vllm_config: VllmConfig,

390 prefix: str = "linear_attn",

391 ):

vllm/model_executor/layers/mamba/linear/base.py

+2 -2

@@ -1,7 +1,7 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 import torch

4 -from transformers import PretrainedConfig

4 +from transformers import PreTrainedConfig

5

6 from vllm.config import (

7 VllmConfig,

@@ -24,7 +24,7 @@ class LinearAttention(PluggableLayer, MambaBase):

24 """Base class for Linear attention layer."""

25

26 def __init__(

27 - self, config: PretrainedConfig, vllm_config: VllmConfig, prefix: str = ""

27 + self, config: PreTrainedConfig, vllm_config: VllmConfig, prefix: str = ""

28 ):

29 super().__init__()

30 self.layer_idx = extract_layer_index(prefix)

vllm/model_executor/layers/pooler/activations.py

+2 -2

@@ -7,7 +7,7 @@ from typing import TypeVar

7 import torch

8 import torch.nn as nn

9 import torch.nn.functional as F

10 -from transformers import PretrainedConfig

10 +from transformers import PreTrainedConfig

11

12 from vllm.config import ModelConfig

13 from vllm.logger import init_logger

@@ -17,7 +17,7 @@ logger = init_logger(__name__)

17

18

19 def get_act_fn(

20 - config: PretrainedConfig,

20 + config: PreTrainedConfig,

21 static_num_labels: bool = True,

22 ) -> "PoolerActivation":

23 # get classification act_fn

vllm/model_executor/layers/quantization/auto_awq.py

+2 -2

@@ -6,7 +6,7 @@ from typing import TYPE_CHECKING, Any, Union

6 import torch

7 from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE

8 from torch.nn import Parameter

9 -from transformers import PretrainedConfig

9 +from transformers import PreTrainedConfig

10

11 import vllm.model_executor.layers.fused_moe # noqa

12 from vllm import _custom_ops as ops

@@ -370,7 +370,7 @@ class AutoAWQConfig(QuantizationConfig):

370 def maybe_update_config(

371 self,

372 model_name: str,

373 - hf_config: PretrainedConfig | None = None,

373 + hf_config: PreTrainedConfig | None = None,

374 revision: str | None = None,

375 ):

376 if self.modules_to_not_convert:

vllm/model_executor/layers/quantization/auto_gptq.py

+2 -2

@@ -6,7 +6,7 @@ from typing import Any

6

7 import torch

8 from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE

9 -from transformers import PretrainedConfig

9 +from transformers import PreTrainedConfig

10

11 import vllm.model_executor.layers.fused_moe # noqa

12 from vllm.logger import init_logger

@@ -273,7 +273,7 @@ class AutoGPTQConfig(QuantizationConfig):

273 def maybe_update_config(

274 self,

275 model_name: str,

276 - hf_config: PretrainedConfig | None = None,

276 + hf_config: PreTrainedConfig | None = None,

277 revision: str | None = None,

278 ):

279 if self.modules_in_block_to_quantize:

vllm/model_executor/layers/quantization/base_config.py

+2 -2

@@ -8,7 +8,7 @@ from typing import TYPE_CHECKING, Any

8 import regex as re

9 import torch

10 from torch import nn

11 -from transformers import PretrainedConfig

11 +from transformers import PreTrainedConfig

12

13 if TYPE_CHECKING:

14 from vllm.model_executor.layers.quantization import QuantizationMethods

@@ -265,7 +265,7 @@ class QuantizationConfig(ABC):

265 def maybe_update_config( # noqa: B027

266 self,

267 model_name: str,

268 - hf_config: PretrainedConfig | None = None,

268 + hf_config: PreTrainedConfig | None = None,

269 revision: str | None = None,

270 ):

271 """Interface to update values after config initialization.

vllm/model_executor/model_loader/tensorizer.py

+2 -2

@@ -18,7 +18,7 @@ import regex as re

18 import torch

19 from torch import nn

20 from torch.utils._python_dispatch import TorchDispatchMode

21 -from transformers import PretrainedConfig

21 +from transformers import PreTrainedConfig

22

23 import vllm.envs as envs

24 from vllm.config import ModelConfig, ParallelConfig, VllmConfig, set_current_vllm_config

@@ -173,7 +173,7 @@ class TensorizerConfig(MutableMapping):

173 deserialization_kwargs: dict[str, Any] | None = None

174 _extra_serialization_attrs: dict[str, Any] | None = field(init=False, default=None)

175 model_class: type[torch.nn.Module] | None = field(init=False, default=None)

176 - hf_config: PretrainedConfig | None = field(init=False, default=None)

176 + hf_config: PreTrainedConfig | None = field(init=False, default=None)

177 dtype: str | torch.dtype | None = field(init=False, default=None)

178 _is_sharded: bool = field(init=False, default=False)

179 _fields: ClassVar[tuple[str, ...]]

vllm/model_executor/models/adapters.py

+3 -3

@@ -284,7 +284,7 @@ def as_embedding_model(cls: type[_T]) -> type[_T]:

284 def _resolve_num_labels(hf_config: Any, text_config: Any) -> int:

285 """Resolve the label count for a sequence classification head.

286

287 - ``PretrainedConfig.num_labels`` is derived from ``id2label``, which always

287 + ``PreTrainedConfig.num_labels`` is derived from ``id2label``, which always

288 carries a default of two entries. Composite configs (such as multimodal

289 checkpoints) declare their label space on the top-level config, so reading

290 ``num_labels`` from ``get_text_config()`` silently returns that default and

@@ -297,9 +297,9 @@ def _resolve_num_labels(hf_config: Any, text_config: Any) -> int:

297 if text_config is hf_config:

298 return hf_config.num_labels

299

300 - from transformers import PretrainedConfig

300 + from transformers import PreTrainedConfig

301

302 - if hf_config.num_labels != PretrainedConfig().num_labels:

302 + if hf_config.num_labels != PreTrainedConfig().num_labels:

303 return hf_config.num_labels

304 return text_config.num_labels

305

vllm/model_executor/models/audioflamingo3.py

+3 -3

@@ -22,7 +22,7 @@ from typing import Annotated, Any, Literal, TypeAlias

22

23 import torch

24 import torch.nn as nn

25 -from transformers import BatchFeature, PretrainedConfig

25 +from transformers import BatchFeature, PreTrainedConfig

26 from transformers.models.audioflamingo3 import (

27 AudioFlamingo3Config,

28 AudioFlamingo3Processor,

@@ -118,7 +118,7 @@ AudioFlamingo3Inputs: TypeAlias = (

118 class AudioFlamingo3Encoder(Qwen2AudioEncoder):

119 def __init__(

120 self,

121 - config: PretrainedConfig,

121 + config: PreTrainedConfig,

122 ):

123 super().__init__(config)

124 self.avg_pooler = nn.AvgPool1d(kernel_size=2, stride=2)

@@ -161,7 +161,7 @@ class AudioFlamingo3Encoder(Qwen2AudioEncoder):

161

162

163 class AudioFlamingo3MultiModalProjector(nn.Module):

164 - def __init__(self, config: PretrainedConfig):

164 + def __init__(self, config: PreTrainedConfig):

165 super().__init__()

166 self.linear_1 = nn.Linear(

167 config.audio_config.hidden_size,

vllm/model_executor/models/bailing_moe.py

+5 -5

@@ -30,7 +30,7 @@ from itertools import islice

30 import torch

31 import torch.nn.functional as F

32 from torch import nn

33 -from transformers.configuration_utils import PretrainedConfig

33 +from transformers.configuration_utils import PreTrainedConfig

34

35 from vllm.compilation.decorators import support_torch_compile

36 from vllm.config import CacheConfig, VllmConfig

@@ -74,7 +74,7 @@ from .utils import (

74 class BailingAttention(nn.Module):

75 def __init__(

76 self,

77 - config: PretrainedConfig,

77 + config: PreTrainedConfig,

78 cache_config: CacheConfig | None = None,

79 quant_config: QuantizationConfig | None = None,

80 reduce_results: bool = True,

@@ -183,7 +183,7 @@ class BailingMLP(nn.Module):

183 def __init__(

184 self,

185 intermediate_size: int,

186 - config: PretrainedConfig,

186 + config: PreTrainedConfig,

187 quant_config: QuantizationConfig | None = None,

188 reduce_results: bool = True,

189 prefix: str = "",

@@ -217,7 +217,7 @@ class BailingMoE(nn.Module):

217 def __init__(

218 self,

219 intermediate_size: int,

220 - config: PretrainedConfig,

220 + config: PreTrainedConfig,

221 quant_config: QuantizationConfig | None = None,

222 reduce_results: bool | None = True,

223 prefix: str = "",

@@ -328,7 +328,7 @@ class BailingMoE(nn.Module):

328 class BailingMoeBlock(nn.Module):

329 def __init__(

330 self,

331 - config: PretrainedConfig,

331 + config: PreTrainedConfig,

332 cache_config: CacheConfig | None = None,

333 quant_config: QuantizationConfig | None = None,

334 prefix: str = "",

vllm/model_executor/models/bailing_moe_linear.py

+4 -4

@@ -5,7 +5,7 @@ from collections.abc import Callable, Iterable

5

6 import torch

7 import torch.nn as nn

8 -from transformers.configuration_utils import PretrainedConfig

8 +from transformers.configuration_utils import PreTrainedConfig

9

10 from vllm.compilation.decorators import support_torch_compile

11 from vllm.config import CacheConfig, VllmConfig

@@ -79,7 +79,7 @@ class BailingMoeV25MLAAttention(nn.Module):

79

80 def __init__(

81 self,

82 - config: PretrainedConfig,

82 + config: PreTrainedConfig,

83 quant_config: QuantizationConfig | None = None,

84 layer_id: int = 0,

85 prefix: str = "attention",

@@ -234,7 +234,7 @@ class BailingMoeV25(nn.Module):

234

235 def __init__(

236 self,

237 - config: PretrainedConfig,

237 + config: PreTrainedConfig,

238 quant_config: QuantizationConfig | None = None,

239 layer_id: int = 0,

240 prefix: str = "",

@@ -349,7 +349,7 @@ class BailingMoeV25DecoderLayer(nn.Module):

349

350 def __init__(

351 self,

352 - config: PretrainedConfig,

352 + config: PreTrainedConfig,

353 vllm_config: VllmConfig,

354 prefix: str = "layer",

355 layer_id: int = 0,

vllm/model_executor/models/bailing_moe_mtp.py

+3 -3

@@ -6,7 +6,7 @@ from collections.abc import Callable, Iterable

6

7 import torch

8 import torch.nn as nn

9 -from transformers.configuration_utils import PretrainedConfig

9 +from transformers.configuration_utils import PreTrainedConfig

10

11 from vllm.compilation.decorators import support_torch_compile

12 from vllm.config import VllmConfig

@@ -36,7 +36,7 @@ from vllm.sequence import IntermediateTensors

36 from .utils import PPMissingLayer, is_pp_missing_parameter, maybe_prefix

37

38

39 -def _get_draft_hf_config(vllm_config: VllmConfig) -> PretrainedConfig:

39 +def _get_draft_hf_config(vllm_config: VllmConfig) -> PreTrainedConfig:

40 speculative_config = vllm_config.speculative_config

41 if speculative_config is not None:

42 draft_model_config = speculative_config.draft_model_config

@@ -48,7 +48,7 @@ def _get_draft_hf_config(vllm_config: VllmConfig) -> PretrainedConfig:

48 class BailingMTPSharedHead(nn.Module):

49 def __init__(

50 self,

51 - config: PretrainedConfig,

51 + config: PreTrainedConfig,

52 prefix: str,

53 vllm_config: VllmConfig,

54 ) -> None:

vllm/model_executor/models/bailing_moe_v3.py

+10 -10

@@ -17,7 +17,7 @@ import regex as re

17 import torch

18 import torch.nn as nn

19 from einops import rearrange

20 -from transformers.configuration_utils import PretrainedConfig

20 +from transformers.configuration_utils import PreTrainedConfig

21

22 from vllm.compilation.decorators import support_torch_compile

23 from vllm.config import CacheConfig, ModelConfig, VllmConfig, get_current_vllm_config

@@ -151,7 +151,7 @@ def _get_kda_state_shape_for_config(

151 return shapes

152

153

154 -def _build_rope_parameters(config: PretrainedConfig) -> dict | None:

154 +def _build_rope_parameters(config: PreTrainedConfig) -> dict | None:

155 rope_parameters = copy.deepcopy(getattr(config, "rope_parameters", None)) or {}

156 if "rope_theta" not in rope_parameters and hasattr(config, "rope_theta"):

157 rope_parameters["rope_theta"] = config.rope_theta

@@ -172,7 +172,7 @@ def _build_rope_parameters(config: PretrainedConfig) -> dict | None:

172

173

174 def _build_mla_rotary_embedding(

175 - config: PretrainedConfig,

175 + config: PreTrainedConfig,

176 head_size: int,

177 ) -> nn.Module:

178 rope_parameters = _build_rope_parameters(config)

@@ -223,7 +223,7 @@ def _is_block_fp8_config(

223

224 def _configure_ling_fp8_quant_config(

225 quant_config: QuantizationConfig | None,

226 - config: PretrainedConfig,

226 + config: PreTrainedConfig,

227 ) -> None:

228 if not _is_block_fp8_config(quant_config):

229 return

@@ -372,7 +372,7 @@ def _pad_block_fp8_mlp_checkpoint_tensor(

372

373 def _maybe_pad_block_fp8_shared_expert_checkpoint_tensor(

374 quant_config: QuantizationConfig | None,

375 - config: PretrainedConfig,

375 + config: PreTrainedConfig,

376 name: str,

377 loaded_weight: torch.Tensor,

378 ) -> torch.Tensor:

@@ -410,7 +410,7 @@ class BailingMoeV3MLP(nn.Module):

410 def __init__(

411 self,

412 intermediate_size: int,

413 - config: PretrainedConfig,

413 + config: PreTrainedConfig,

414 quant_config: QuantizationConfig | None = None,

415 reduce_results: bool = True,

416 prefix: str = "",

@@ -446,7 +446,7 @@ class BailingMoeV3MLP(nn.Module):

446 class BailingMoeV3MLAAttention(nn.Module):

447 def __init__(

448 self,

449 - config: PretrainedConfig,

449 + config: PreTrainedConfig,

450 quant_config: QuantizationConfig | None = None,

451 layer_id: int = 0,

452 prefix: str = "attention",

@@ -625,7 +625,7 @@ class BailingMoeV3KimiDeltaAttention(PluggableLayer, MambaBase):

625

626 def __init__(

627 self,

628 - config: PretrainedConfig,

628 + config: PreTrainedConfig,

629 quant_config: QuantizationConfig | None = None,

630 layer_id: int = 0,

631 prefix: str = "attention",

@@ -1121,7 +1121,7 @@ class BailingMoeV3KimiDeltaAttention(PluggableLayer, MambaBase):

1121 class BailingMoeV3MoE(nn.Module):

1122 def __init__(

1123 self,

1124 - config: PretrainedConfig,

1124 + config: PreTrainedConfig,

1125 quant_config: QuantizationConfig | None = None,

1126 layer_id: int = 0,

1127 prefix: str = "",

@@ -1193,7 +1193,7 @@ class BailingMoeV3MoE(nn.Module):

1193 class BailingMoeV3DecoderLayer(nn.Module):

1194 def __init__(

1195 self,

1196 - config: PretrainedConfig,

1196 + config: PreTrainedConfig,

1197 quant_config: QuantizationConfig | None = None,

1198 layer_id: int = 0,

1199 prefix: str = "layer",

vllm/model_executor/models/bailing_moe_v3_mtp.py

+3 -3

@@ -6,7 +6,7 @@ from collections.abc import Callable, Iterable

6

7 import torch

8 import torch.nn as nn

9 -from transformers.configuration_utils import PretrainedConfig

9 +from transformers.configuration_utils import PreTrainedConfig

10

11 from vllm.compilation.decorators import support_torch_compile

12 from vllm.config import VllmConfig

@@ -47,7 +47,7 @@ from vllm.model_executor.models.utils import (

47 from vllm.sequence import IntermediateTensors

48

49

50 -def _get_draft_hf_config(vllm_config: VllmConfig) -> PretrainedConfig:

50 +def _get_draft_hf_config(vllm_config: VllmConfig) -> PreTrainedConfig:

51 speculative_config = vllm_config.speculative_config

52 if speculative_config is not None:

53 draft_model_config = speculative_config.draft_model_config

@@ -59,7 +59,7 @@ def _get_draft_hf_config(vllm_config: VllmConfig) -> PretrainedConfig:

59 class BailingMoeV3MTPSharedHead(nn.Module):

60 def __init__(

61 self,

62 - config: PretrainedConfig,

62 + config: PreTrainedConfig,

63 prefix: str,

64 vllm_config: VllmConfig,

65 ) -> None:

vllm/model_executor/models/bert_with_rope.py

+3 -3

@@ -4,7 +4,7 @@ from collections.abc import Iterable

4

5 import torch

6 from torch import nn

7 -from transformers import PretrainedConfig

7 +from transformers import PreTrainedConfig

8

9 from vllm.compilation.decorators import support_torch_compile

10 from vllm.config import CacheConfig, VllmConfig

@@ -50,7 +50,7 @@ from .interfaces_base import default_pooling_type

50

51

52 class BertWithRopeEmbedding(nn.Module):

53 - def __init__(self, config: PretrainedConfig):

53 + def __init__(self, config: PreTrainedConfig):

54 super().__init__()

55 if config.position_embedding_type not in ["rope", "rotary"]:

56 raise ValueError(

@@ -347,7 +347,7 @@ class NomicMoE(nn.Module):

347 class BertWithRopeBlock(nn.Module):

348 def __init__(

349 self,

350 - config: PretrainedConfig,

350 + config: PreTrainedConfig,

351 cache_config: CacheConfig | None = None,

352 quant_config: QuantizationConfig | None = None,

353 moe: bool = False,

vllm/model_executor/models/cohere2_vision.py

+5 -5

@@ -8,10 +8,10 @@ from typing import Annotated, Literal

8

9 import torch

10 from torch import nn

11 -from transformers import BatchFeature, PretrainedConfig

11 +from transformers import BatchFeature, PreTrainedConfig

12 from transformers.models.cohere2_vision import Cohere2VisionConfig

13 -from transformers.models.cohere2_vision.image_processing_cohere2_vision_fast import ( # noqa: E501

14 - Cohere2VisionImageProcessorFast,

13 +from transformers.models.cohere2_vision.image_processing_cohere2_vision import ( # noqa: E501

14 + Cohere2VisionImageProcessor,

15 )

16 from transformers.models.cohere2_vision.processing_cohere2_vision import (

17 Cohere2VisionProcessor,

@@ -178,7 +178,7 @@ class Cohere2VisionProcessingInfo(BaseProcessingInfo):

178 """Calculate the number of image patches for a given image.

179 Uses the HF processor to determine the actual number of patches.

180 """

181 - image_processor: Cohere2VisionImageProcessorFast = processor.image_processor

181 + image_processor: Cohere2VisionImageProcessor = processor.image_processor

182

183 return image_processor.get_number_of_image_patches(

184 image_height,

@@ -415,7 +415,7 @@ class Cohere2VisionForConditionalGeneration(

415 )

416

417 def _patch_quant_config(

418 - self, config: PretrainedConfig, quant_config: QuantizationConfig | None

418 + self, config: PreTrainedConfig, quant_config: QuantizationConfig | None

419 ):

420 # the awq models from OpenGVLab missing `modules_to_not_convert`

421 # patch the quant_config to add `modules_to_not_convert` back

vllm/model_executor/models/cohere_asr.py

+2 -2

@@ -8,7 +8,7 @@ from typing import Any, ClassVar

8 import torch

9 import torch.nn.functional as F

10 from torch import nn

11 -from transformers import PretrainedConfig

11 +from transformers import PreTrainedConfig

12

13 from vllm.compilation.decorators import support_torch_compile

14 from vllm.config import CacheConfig, ModelConfig, SpeechToTextConfig, VllmConfig

@@ -1861,7 +1861,7 @@ class CohereASRModel(nn.Module):

1861

1862

1863 class CohereASRProcessingInfo(BaseProcessingInfo):

1864 - def get_hf_config(self) -> PretrainedConfig:

1864 + def get_hf_config(self) -> PreTrainedConfig:

1865 return self.ctx.get_hf_config()

1866

1867 def get_default_tok_params(self) -> TokenizeParams:

vllm/model_executor/models/cohere_compass.py

-1

@@ -1161,7 +1161,6 @@ class CohereCompassProcessingInfo(BaseProcessingInfo):

1161 def get_hf_processor(self, **kwargs: object) -> CohereCompassProcessor:

1162 return self.ctx.get_hf_processor(

1163 CohereCompassProcessor,

1164 - use_fast=kwargs.pop("use_fast", True),

1164 **kwargs,

1165 )

1166

vllm/model_executor/models/colqwen3.py

-1

@@ -61,7 +61,6 @@ class ColQwen3ProcessingInfo(Qwen3VLProcessingInfo):

61 # correctly for the Qwen3-VL backbone.

62 return self.ctx.get_hf_processor(

63 Qwen3VLProcessor,

64 - use_fast=kwargs.pop("use_fast", True),

64 **kwargs,

65 )

66

vllm/model_executor/models/colqwen3_5.py

-1

@@ -58,7 +58,6 @@ class ColQwen3_5ProcessingInfo(Qwen3_5ProcessingInfo):

58 def get_hf_processor(self, **kwargs: object) -> Qwen3VLProcessor:

59 return self.ctx.get_hf_processor(

60 Qwen3VLProcessor,

61 - use_fast=kwargs.pop("use_fast", True),

61 **kwargs,

62 )

63

vllm/model_executor/models/config.py

+2 -2

@@ -6,7 +6,7 @@ from vllm.logger import init_logger

6 from vllm.utils.math_utils import round_up

7

8 if TYPE_CHECKING:

9 - from transformers import PretrainedConfig

9 + from transformers import PreTrainedConfig

10

11 from vllm.config import CacheConfig, ModelConfig, VllmConfig

12 from vllm.config.cache import MambaDType

@@ -681,7 +681,7 @@ class NemotronHForCausalLMConfig(VerifyAndUpdateConfig):

681

682 @classmethod

683 def update_mamba_ssm_cache_dtype(

684 - cls, *, cache_config: "CacheConfig", hf_config: "PretrainedConfig"

684 + cls, *, cache_config: "CacheConfig", hf_config: "PreTrainedConfig"

685 ) -> None:

686 """Update mamba_ssm_cache_dtype for NemotronH models when set to 'auto'

687 (or not explicitly set), to the value specified in the HF config, or to

vllm/model_executor/models/deepseek_mtp.py

+2 -2

@@ -5,7 +5,7 @@ from collections.abc import Callable, Iterable

5

6 import torch

7 import torch.nn as nn

8 -from transformers import PretrainedConfig

8 +from transformers import PreTrainedConfig

9

10 from vllm.compilation.decorators import support_torch_compile

11 from vllm.config import VllmConfig

@@ -51,7 +51,7 @@ from .utils import (

51 class SharedHead(nn.Module):

52 def __init__(

53 self,

54 - config: PretrainedConfig,

54 + config: PreTrainedConfig,

55 prefix: str,

56 quant_config: QuantizationConfig | None = None,

57 *,

vllm/model_executor/models/diffusion_gemma.py

+2 -2

@@ -112,13 +112,13 @@ class DiffusionGemmaProcessingInfo(Gemma4ProcessingInfo):

112 """Processing info for DiffusionGemma.

113

114 Overrides ``get_hf_config`` to accept ``DiffusionGemmaConfig``

115 - (which inherits from ``PretrainedConfig``, not ``Gemma4Config``).

115 + (which inherits from ``PreTrainedConfig``, not ``Gemma4Config``).

116 Supports image and video modalities.

117 """

118

119 def get_hf_config(self):

120 # DiffusionGemmaConfig doesn't inherit from Gemma4Config, so we

121 - # accept any PretrainedConfig here.

121 + # accept any PreTrainedConfig here.

122 return self.ctx.get_hf_config()

123

124 def get_supported_mm_limits(self) -> Mapping[str, int | None]:

vllm/model_executor/models/eagle2_5_vl.py

+3 -3

@@ -8,7 +8,7 @@ from typing import Annotated, Literal, TypeAlias

8

9 import torch

10 import torch.nn as nn

11 -from transformers import PretrainedConfig

11 +from transformers import PreTrainedConfig

12

13 from vllm.config import VllmConfig

14 from vllm.model_executor.layers.quantization import QuantizationConfig

@@ -193,7 +193,7 @@ class Eagle2_5_VLForConditionalGeneration(

193

194 def _init_vision_model(

195 self,

196 - config: PretrainedConfig,

196 + config: PreTrainedConfig,

197 quant_config: QuantizationConfig | None,

198 prefix: str,

199 ):

@@ -220,7 +220,7 @@ class Eagle2_5_VLForConditionalGeneration(

220 prefix=prefix,

221 )

222

223 - def _init_mlp1(self, config: PretrainedConfig) -> nn.Module:

223 + def _init_mlp1(self, config: PreTrainedConfig) -> nn.Module:

224 """Initialize MLP projection layer."""

225 vit_hidden_size = config.vision_config.hidden_size

226 llm_hidden_size = config.text_config.hidden_size

vllm/model_executor/models/ernie45_moe.py

+3 -3

@@ -29,7 +29,7 @@ from typing import Any

29

30 import torch

31 from torch import nn

32 -from transformers import PretrainedConfig

32 +from transformers import PreTrainedConfig

33

34 from vllm.compilation.decorators import support_torch_compile

35 from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config

@@ -119,7 +119,7 @@ class Ernie4_5_MoeMLP(nn.Module):

119 class Ernie4_5_MoeMoE(nn.Module):

120 def __init__(

121 self,

122 - config: PretrainedConfig,

122 + config: PreTrainedConfig,

123 quant_config: QuantizationConfig | None = None,

124 prefix: str = "",

125 enable_eplb: bool = False,

@@ -305,7 +305,7 @@ class Ernie4_5_MoeAttention(nn.Module):

305 class Ernie4_5_MoeDecoderLayer(nn.Module):

306 def __init__(

307 self,

308 - config: PretrainedConfig,

308 + config: PreTrainedConfig,

309 cache_config: CacheConfig | None = None,

310 quant_config: QuantizationConfig | None = None,

311 prefix: str = "",

vllm/model_executor/models/ernie45_vl.py

+1 -1

@@ -881,7 +881,7 @@ class Ernie4_5_VLProcessingInfo(BaseProcessingInfo):

881 return self.ctx.model_config.hf_config

882

883 def get_hf_processor(self, **kwargs: object):

884 - return self.ctx.get_hf_processor(use_fast=True, **kwargs)

884 + return self.ctx.get_hf_processor(**kwargs)

885

886 def get_image_processor(self, **kwargs: object):

887 return self.get_hf_processor(**kwargs).image_processor

vllm/model_executor/models/ernie45_vl_moe.py

+3 -3

@@ -29,7 +29,7 @@ from typing import Any

29

30 import torch

31 from torch import nn

32 -from transformers import PretrainedConfig

32 +from transformers import PreTrainedConfig

33

34 # from vllm.compilation.decorators import support_torch_compile

35 from vllm.config import CacheConfig, VllmConfig

@@ -192,7 +192,7 @@ class Ernie4_5_VLMoeAttention(nn.Module):

192 class Ernie4_5_VLMoeMoE(nn.Module):

193 def __init__(

194 self,

195 - config: PretrainedConfig,

195 + config: PreTrainedConfig,

196 quant_config: QuantizationConfig | None = None,

197 prefix: str = "",

198 ):

@@ -385,7 +385,7 @@ class Ernie4_5_VLMoeMoE(nn.Module):

385 class Ernie4_5_VLMoeDecoderLayer(nn.Module):

386 def __init__(

387 self,

388 - config: PretrainedConfig,

388 + config: PreTrainedConfig,

389 cache_config: CacheConfig | None = None,

390 quant_config: QuantizationConfig | None = None,

391 prefix: str = "",

vllm/model_executor/models/exaone.py

+4 -4

@@ -30,7 +30,7 @@ from itertools import islice

30

31 import torch

32 from torch import nn

33 -from transformers import PretrainedConfig

33 +from transformers import PreTrainedConfig

34

35 from vllm.compilation.decorators import support_torch_compile

36 from vllm.config import CacheConfig, VllmConfig

@@ -104,7 +104,7 @@ class ExaoneGatedMLP(nn.Module):

104 class ExaoneAttention(nn.Module):

105 def __init__(

106 self,

107 - config: PretrainedConfig,

107 + config: PreTrainedConfig,

108 hidden_size: int,

109 num_heads: int,

110 num_kv_heads: int,

@@ -191,7 +191,7 @@ class ExaoneAttention(nn.Module):

191 class ExaoneBlockAttention(nn.Module):

192 def __init__(

193 self,

194 - config: PretrainedConfig,

194 + config: PreTrainedConfig,

195 hidden_size: int,

196 num_heads: int,

197 num_kv_heads: int,

@@ -228,7 +228,7 @@ class ExaoneBlockAttention(nn.Module):

228 class ExaoneDecoderLayer(nn.Module):

229 def __init__(

230 self,

231 - config: PretrainedConfig,

231 + config: PreTrainedConfig,

232 cache_config: CacheConfig | None = None,

233 quant_config: QuantizationConfig | None = None,

234 prefix: str = "",

vllm/model_executor/models/exaone4_5.py

-1

@@ -311,7 +311,6 @@ class Exaone4_5_ProcessingInfo(Qwen2VLProcessingInfo):

311 def get_hf_processor(self, **kwargs: object) -> Exaone4_5_Processor:

312 return self.ctx.get_hf_processor(

313 Exaone4_5_Processor,

314 - use_fast=kwargs.pop("use_fast", True),

314 **kwargs,

315 )

316

vllm/model_executor/models/exaone_moe.py

+3 -3

@@ -20,7 +20,7 @@ from itertools import islice

20

21 import torch

22 from torch import nn

23 -from transformers import PretrainedConfig

23 +from transformers import PreTrainedConfig

24

25 from vllm.compilation.decorators import support_torch_compile

26 from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config

@@ -65,7 +65,7 @@ from .utils import (

65 class ExaoneMoe(nn.Module):

66 def __init__(

67 self,

68 - config: PretrainedConfig,

68 + config: PreTrainedConfig,

69 swiglu_limit: float | None = None,

70 quant_config: QuantizationConfig | None = None,

71 prefix: str = "",

@@ -289,7 +289,7 @@ class ExaoneMoeAttention(nn.Module):

289 class ExaoneMoeDecoderLayer(nn.Module):

290 def __init__(

291 self,

292 - config: PretrainedConfig,

292 + config: PreTrainedConfig,

293 cache_config: CacheConfig | None = None,

294 quant_config: QuantizationConfig | None = None,

295 is_mtp: bool = False,

vllm/model_executor/models/gemma3n_mm.py

+2 -2

@@ -15,7 +15,7 @@ from transformers.models.gemma3n import (

15 Gemma3nTextConfig,

16 Gemma3nVisionConfig,

17 )

18 -from transformers.models.siglip import SiglipImageProcessorFast

18 +from transformers.models.siglip import SiglipImageProcessor

19

20 from vllm.config import ModelConfig, SpeechToTextConfig, VllmConfig

21 from vllm.config.multimodal import MultiModalDummyOptions

@@ -239,7 +239,7 @@ class Gemma3nDummyInputsBuilder(BaseDummyInputsBuilder[Gemma3nProcessingInfo]):

239 processor.feature_extractor

240 )

241 audio_len = audio_feature_extractor.fft_length

242 - image_processor: SiglipImageProcessorFast = processor.image_processor

242 + image_processor: SiglipImageProcessor = processor.image_processor

243 img_width = image_processor.size.get("width", 224)

244 img_height = image_processor.size.get("height", 224)

245

vllm/model_executor/models/glm4_moe_lite_mtp.py

+2 -2

@@ -28,7 +28,7 @@ from collections.abc import Callable, Iterable

28

29 import torch

30 import torch.nn as nn

31 -from transformers import PretrainedConfig

31 +from transformers import PreTrainedConfig

32

33 from vllm.config import VllmConfig

34 from vllm.model_executor.layers.fused_moe import (

@@ -68,7 +68,7 @@ from .utils import get_spec_layer_idx_from_weight_name, maybe_prefix

68 class SharedHead(nn.Module):

69 def __init__(

70 self,

71 - config: PretrainedConfig,

71 + config: PreTrainedConfig,

72 prefix: str,

73 quant_config: QuantizationConfig | None = None,

74 ) -> None:

vllm/model_executor/models/glm4_moe_mtp.py

+3 -3

@@ -29,7 +29,7 @@ from collections.abc import Callable, Iterable

29

30 import torch

31 import torch.nn as nn

32 -from transformers import PretrainedConfig

32 +from transformers import PreTrainedConfig

33

34 from vllm.config import CacheConfig, ParallelConfig, VllmConfig

35 from vllm.model_executor.layers.fused_moe import (

@@ -60,7 +60,7 @@ from .utils import get_spec_layer_idx_from_weight_name, maybe_prefix

60 class SharedHead(nn.Module):

61 def __init__(

62 self,

63 - config: PretrainedConfig,

63 + config: PreTrainedConfig,

64 prefix: str,

65 quant_config: QuantizationConfig | None = None,

66 ) -> None:

@@ -80,7 +80,7 @@ class SharedHead(nn.Module):

80 class Glm4MoeMultiTokenPredictorLayer(nn.Module):

81 def __init__(

82 self,

83 - config: PretrainedConfig,

83 + config: PreTrainedConfig,

84 prefix: str,

85 cache_config: CacheConfig | None = None,

86 quant_config: QuantizationConfig | None = None,

vllm/model_executor/models/granite.py

+2 -2

@@ -36,7 +36,7 @@ from itertools import islice

36 import torch

37 from torch import nn

38 from transformers import GraniteConfig

39 -from transformers.configuration_utils import PretrainedConfig

39 +from transformers.configuration_utils import PreTrainedConfig

40

41 from vllm.compilation.decorators import support_torch_compile

42 from vllm.config import CacheConfig, VllmConfig

@@ -72,7 +72,7 @@ from .utils import (

72

73

74 def granite_layer_attn_params(

75 - config: PretrainedConfig, layer_idx: int

75 + config: PreTrainedConfig, layer_idx: int

76 ) -> tuple[int | None, float, bool]:

77 """Resolve one layer's sliding window, RoPE base and sink usage.

78

vllm/model_executor/models/granite_speech.py

+8 -8

@@ -31,7 +31,7 @@ from typing import Annotated

31 import torch

32 import torch.nn.functional as F

33 from torch import nn

34 -from transformers import BatchFeature, PretrainedConfig

34 +from transformers import BatchFeature, PreTrainedConfig

35

36 from vllm.config import CacheConfig, ModelConfig, SpeechToTextConfig, VllmConfig

37 from vllm.config.multimodal import MultiModalDummyOptions

@@ -229,7 +229,7 @@ class GraniteSpeechDummyInputsBuilder(

229 class GraniteSpeechEncoderProjector(nn.Module):

230 def __init__(

231 self,

232 - config: PretrainedConfig,

232 + config: PreTrainedConfig,

233 cache_config: CacheConfig,

234 quant_config: QuantizationConfig | None = None,

235 prefix: str = "",

@@ -287,7 +287,7 @@ class GraniteSpeechConformerFeedForward(nn.Module):

287

288 def __init__(

289 self,

290 - config: PretrainedConfig,

290 + config: PreTrainedConfig,

291 quant_config: QuantizationConfig | None = None,

292 prefix: str = "",

293 ):

@@ -323,7 +323,7 @@ class GraniteSpeechConformerAttention(nn.Module):

323 for more details.

324 """

325

326 - def __init__(self, config: PretrainedConfig, prefix: str = ""):

326 + def __init__(self, config: PreTrainedConfig, prefix: str = ""):

327 super().__init__()

328

329 inner_dim = config.dim_head * config.num_heads

@@ -429,7 +429,7 @@ class GraniteSpeechConformerConvModule(nn.Module):

429 convolutional layers.

430 """

431

432 - def __init__(self, config: PretrainedConfig, prefix: str = ""):

432 + def __init__(self, config: PreTrainedConfig, prefix: str = ""):

433 super().__init__()

434 inner_dim = config.hidden_dim * config.conv_expansion_factor

435

@@ -460,7 +460,7 @@ class GraniteSpeechConformerBlock(nn.Module):

460 """Conformer block, consisting largely of linear layers,

461 attention, and convolutional layers."""

462

463 - def __init__(self, config: PretrainedConfig, prefix: str = ""):

463 + def __init__(self, config: PreTrainedConfig, prefix: str = ""):

464 super().__init__()

465 self.ff1 = GraniteSpeechConformerFeedForward(config, prefix=f"{prefix}.ff1")

466 self.attn = GraniteSpeechConformerAttention(config, prefix=f"{prefix}.attn")

@@ -486,7 +486,7 @@ class GraniteSpeechCTCEncoder(nn.Module):

486

487 def __init__(

488 self,

489 - config: PretrainedConfig,

489 + config: PreTrainedConfig,

490 prefix: str,

491 quant_config: QuantizationConfig | None = None,

492 ):

@@ -622,7 +622,7 @@ class GraniteSpeechForConditionalGeneration(

622

623 def _build_encoder(

624 self,

625 - config: PretrainedConfig,

625 + config: PreTrainedConfig,

626 quant_config: QuantizationConfig | None,

627 prefix: str,

628 ) -> "GraniteSpeechCTCEncoder":

vllm/model_executor/models/granite_speech_plus.py

+2 -2

@@ -3,7 +3,7 @@

3 """Inference-only IBM Granite Speech Plus model."""

4

5 import torch

6 -from transformers import PretrainedConfig

6 +from transformers import PreTrainedConfig

7

8 from vllm.model_executor.layers.quantization import QuantizationConfig

9 from vllm.multimodal import MULTIMODAL_REGISTRY

@@ -68,7 +68,7 @@ class GraniteSpeechPlusForConditionalGeneration(GraniteSpeechForConditionalGener

68

69 def _build_encoder(

70 self,

71 - config: PretrainedConfig,

71 + config: PreTrainedConfig,

72 quant_config: QuantizationConfig | None,

73 prefix: str,

74 ) -> GraniteSpeechCTCEncoder:

vllm/model_executor/models/granitemoe.py

+2 -2

@@ -29,7 +29,7 @@ from itertools import islice

29

30 import torch

31 from torch import nn

32 -from transformers.configuration_utils import PretrainedConfig

32 +from transformers.configuration_utils import PreTrainedConfig

33

34 from vllm.compilation.decorators import support_torch_compile

35 from vllm.config import CacheConfig, VllmConfig

@@ -141,7 +141,7 @@ class GraniteMoeMoE(nn.Module):

141 class GraniteMoeAttention(nn.Module):

142 def __init__(

143 self,

144 - config: PretrainedConfig,

144 + config: PreTrainedConfig,

145 hidden_size: int,

146 num_heads: int,

147 num_kv_heads: int,

vllm/model_executor/models/h2ovl.py

+2 -2

@@ -10,7 +10,7 @@

10 # --------------------------------------------------------

11

12 import torch

13 -from transformers import PretrainedConfig

13 +from transformers import PreTrainedConfig

14

15 from vllm.model_executor.layers.quantization import QuantizationConfig

16 from vllm.multimodal import MULTIMODAL_REGISTRY

@@ -158,7 +158,7 @@ class H2OVLMultiModalProcessor(BaseInternVLMultiModalProcessor[H2OVLProcessingIn

158 class H2OVLChatModel(InternVLChatModel):

159 def _init_vision_model(

160 self,

161 - config: PretrainedConfig,

161 + config: PreTrainedConfig,

162 quant_config: QuantizationConfig | None,

163 *,

164 prefix: str,

vllm/model_executor/models/hrm_text.py

+4 -4

@@ -25,7 +25,7 @@ from typing import Literal

25

26 import torch

27 from torch import nn

28 -from transformers import PretrainedConfig

28 +from transformers import PreTrainedConfig

29

30 from vllm.compilation.decorators import support_torch_compile

31 from vllm.config import CacheConfig, VllmConfig

@@ -122,7 +122,7 @@ class HrmTextAttention(nn.Module):

122

123 def __init__(

124 self,

125 - config: PretrainedConfig,

125 + config: PreTrainedConfig,

126 layer_idx_in_stack: int,

127 stack_kind: Literal["L", "H"],

128 cache_config: CacheConfig | None = None,

@@ -245,7 +245,7 @@ class HrmTextAttention(nn.Module):

245 class HrmTextDecoderLayer(nn.Module):

246 def __init__(

247 self,

248 - config: PretrainedConfig,

248 + config: PreTrainedConfig,

249 layer_idx_in_stack: int,

250 stack_kind: Literal["L", "H"],

251 cache_config: CacheConfig | None = None,

@@ -310,7 +310,7 @@ class HrmTextStack(nn.Module):

310

311 def __init__(

312 self,

313 - config: PretrainedConfig,

313 + config: PreTrainedConfig,

314 stack_kind: Literal["L", "H"],

315 cache_config: CacheConfig | None = None,

316 quant_config: QuantizationConfig | None = None,

vllm/model_executor/models/hy_v3.py

+3 -3

@@ -31,7 +31,7 @@ from typing import Any

31

32 import torch

33 from torch import nn

34 -from transformers import PretrainedConfig

34 +from transformers import PreTrainedConfig

35

36 from vllm.compilation.decorators import support_torch_compile

37 from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config

@@ -220,7 +220,7 @@ class HYV3MoEFused(nn.Module):

220 class HYV3Attention(nn.Module):

221 def __init__(

222 self,

223 - config: PretrainedConfig,

223 + config: PreTrainedConfig,

224 hidden_size: int,

225 num_heads: int,

226 num_kv_heads: int,

@@ -365,7 +365,7 @@ class HYV3Attention(nn.Module):

365 class HYV3DecoderLayer(nn.Module):

366 def __init__(

367 self,

368 - config: PretrainedConfig,

368 + config: PreTrainedConfig,

369 cache_config: CacheConfig | None = None,

370 quant_config: QuantizationConfig | None = None,

371 prefix: str = "",

vllm/model_executor/models/hy_v3_mtp.py

+5 -5

@@ -29,7 +29,7 @@ from collections.abc import Callable, Iterable

29 import regex as re

30 import torch

31 from torch import nn

32 -from transformers import PretrainedConfig

32 +from transformers import PreTrainedConfig

33

34 from vllm.config import CacheConfig, ModelConfig, VllmConfig

35 from vllm.model_executor.layers.fused_moe import (

@@ -59,7 +59,7 @@ from .utils import (

59 )

60

61

62 -def _is_moe(config: PretrainedConfig) -> bool:

62 +def _is_moe(config: PreTrainedConfig) -> bool:

63 return bool(

64 getattr(config, "num_experts", None)

65 and (

@@ -69,7 +69,7 @@ def _is_moe(config: PretrainedConfig) -> bool:

69 )

70

71

72 -def _get_cla_factor(config: PretrainedConfig) -> int:

72 +def _get_cla_factor(config: PreTrainedConfig) -> int:

73 if not getattr(config, "use_cla", False):

74 return 1

75 return getattr(config, "cla_share_factor", 1)

@@ -78,7 +78,7 @@ def _get_cla_factor(config: PretrainedConfig) -> int:

78 class HYV3SharedHead(nn.Module):

79 def __init__(

80 self,

81 - config: PretrainedConfig,

81 + config: PreTrainedConfig,

82 quant_config: QuantizationConfig | None = None,

83 ) -> None:

84 super().__init__()

@@ -93,7 +93,7 @@ class HYV3SharedHead(nn.Module):

93 class HYV3MultiTokenPredictorLayer(nn.Module):

94 def __init__(

95 self,

96 - config: PretrainedConfig,

96 + config: PreTrainedConfig,

97 prefix: str,

98 model_config: ModelConfig,

99 cache_config: CacheConfig | None = None,

vllm/model_executor/models/intern_vit.py

+8 -8

@@ -13,7 +13,7 @@ from functools import partial

13 import torch

14 import torch.nn as nn

15 import torch.nn.functional as F

16 -from transformers import PretrainedConfig

16 +from transformers import PreTrainedConfig

17

18 from vllm.compilation.decorators import (

19 should_torch_compile_mm_encoder,

@@ -47,7 +47,7 @@ NORM2FN = {

47

48

49 class InternVisionEmbeddings(nn.Module):

50 - def __init__(self, config: PretrainedConfig):

50 + def __init__(self, config: PreTrainedConfig):

51 super().__init__()

52 self.config = config

53 self.embed_dim = config.hidden_size

@@ -119,7 +119,7 @@ class InternParallelAttention(nn.Module):

119

120 def __init__(

121 self,

122 - config: PretrainedConfig,

122 + config: PreTrainedConfig,

123 quant_config: QuantizationConfig | None = None,

124 *,

125 num_dummy_heads: int = 0,

@@ -222,7 +222,7 @@ class InternParallelAttention(nn.Module):

222 class InternMLP(nn.Module):

223 def __init__(

224 self,

225 - config: PretrainedConfig,

225 + config: PreTrainedConfig,

226 quant_config: QuantizationConfig | None = None,

227 prefix: str = "",

228 ) -> None:

@@ -264,7 +264,7 @@ class InternMLP(nn.Module):

264 class InternVisionEncoderLayer(nn.Module):

265 def __init__(

266 self,

267 - config: PretrainedConfig,

267 + config: PreTrainedConfig,

268 quant_config: QuantizationConfig | None = None,

269 *,

270 num_dummy_heads: int = 0,

@@ -297,7 +297,7 @@ class InternVisionEncoderLayer(nn.Module):

297

298 def _init_attn(

299 self,

300 - config: PretrainedConfig,

300 + config: PreTrainedConfig,

301 quant_config: QuantizationConfig | None,

302 *,

303 num_dummy_heads: int,

@@ -324,7 +324,7 @@ class InternVisionEncoderLayer(nn.Module):

324 class InternVisionEncoder(nn.Module):

325 def __init__(

326 self,

327 - config: PretrainedConfig,

327 + config: PreTrainedConfig,

328 quant_config: QuantizationConfig | None = None,

329 *,

330 num_hidden_layers_override: int | None = None,

@@ -369,7 +369,7 @@ class InternVisionModel(nn.Module):

369

370 def __init__(

371 self,

372 - config: PretrainedConfig,

372 + config: PreTrainedConfig,

373 quant_config: QuantizationConfig | None = None,

374 *,

375 num_hidden_layers_override: int | None = None,

vllm/model_executor/models/internlm2.py

+2 -2

@@ -8,7 +8,7 @@ from typing import Any

8

9 import torch

10 from torch import nn

11 -from transformers import PretrainedConfig

11 +from transformers import PreTrainedConfig

12

13 from vllm.compilation.decorators import support_torch_compile

14 from vllm.config import CacheConfig, VllmConfig

@@ -195,7 +195,7 @@ class InternLM2Attention(nn.Module):

195 class InternLMDecoderLayer(nn.Module):

196 def __init__(

197 self,

198 - config: PretrainedConfig,

198 + config: PreTrainedConfig,

199 cache_config: CacheConfig | None = None,

200 quant_config: QuantizationConfig | None = None,

201 prefix: str = "",

vllm/model_executor/models/interns1.py

+6 -6

@@ -12,10 +12,10 @@ from typing import Annotated, Literal, TypeAlias, TypedDict

12 import regex as re

13 import torch

14 import torch.nn as nn

15 -from transformers import BatchFeature, InternVLProcessor, PretrainedConfig

15 +from transformers import BatchFeature, InternVLProcessor, PreTrainedConfig

16 from transformers.activations import ACT2FN

17 -from transformers.models.got_ocr2.image_processing_got_ocr2_fast import (

18 - GotOcr2ImageProcessorFast,

17 +from transformers.models.got_ocr2.image_processing_got_ocr2 import (

18 + GotOcr2ImageProcessor,

19 )

20 from transformers.models.internvl.video_processing_internvl import (

21 InternVLVideoProcessor,

@@ -202,7 +202,7 @@ class InternS1ProcessingInfo(BaseProcessingInfo):

202 processor: InternVLProcessor,

203 mm_kwargs: Mapping[str, object],

204 ) -> int:

205 - image_processor: GotOcr2ImageProcessorFast = processor.image_processor

205 + image_processor: GotOcr2ImageProcessor = processor.image_processor

206

207 num_image_patches = image_processor.get_number_of_image_patches(

208 image_height,

@@ -612,7 +612,7 @@ class InternS1ForConditionalGeneration(

612

613 def _init_vision_model(

614 self,

615 - config: PretrainedConfig,

615 + config: PreTrainedConfig,

616 quant_config: QuantizationConfig | None,

617 *,

618 prefix: str,

@@ -625,7 +625,7 @@ class InternS1ForConditionalGeneration(

625 prefix=prefix,

626 )

627

628 - def _init_mlp1(self, config: PretrainedConfig) -> nn.Module:

628 + def _init_mlp1(self, config: PreTrainedConfig) -> nn.Module:

629 return InternS1MultiModalProjector(config)

630

631 def pixel_shuffle(self, x, scale_factor=0.5):

vllm/model_executor/models/interns1_pro.py

+2 -2

@@ -30,7 +30,7 @@ from typing import Any, ClassVar

30

31 import torch

32 from torch import nn

33 -from transformers import AutoProcessor, PretrainedConfig

33 +from transformers import AutoProcessor, PreTrainedConfig

34

35 from vllm.config import CacheConfig, VllmConfig

36 from vllm.distributed import (

@@ -563,7 +563,7 @@ class InternS1ProForConditionalGeneration(

563

564 def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):

565 super(Qwen3VLForConditionalGeneration, self).__init__()

566 - config: PretrainedConfig = vllm_config.model_config.hf_config

566 + config: PreTrainedConfig = vllm_config.model_config.hf_config

567 multimodal_config = vllm_config.model_config.get_multimodal_config()

568

569 self.config = config

vllm/model_executor/models/interns1_vit.py

+8 -8

@@ -11,7 +11,7 @@ from collections.abc import Iterable

11

12 import torch

13 import torch.nn as nn

14 -from transformers import PretrainedConfig

14 +from transformers import PreTrainedConfig

15 from transformers.utils import torch_int

16

17 from vllm.model_executor.layers.activation import get_act_fn

@@ -65,7 +65,7 @@ class InternS1VisionPatchEmbeddings(nn.Module):

65

66

67 class InternS1VisionEmbeddings(nn.Module):

68 - def __init__(self, config: PretrainedConfig):

68 + def __init__(self, config: PreTrainedConfig):

69 super().__init__()

70 self.config = config

71 self.cls_token = nn.Parameter(torch.zeros(1, 1, config.hidden_size))

@@ -166,7 +166,7 @@ class InternSdpaAttention(nn.Module):

166

167 def __init__(

168 self,

169 - config: PretrainedConfig,

169 + config: PreTrainedConfig,

170 *,

171 num_dummy_heads: int = 0,

172 prefix: str = "",

@@ -242,7 +242,7 @@ class InternSdpaAttention(nn.Module):

242 class InternS1VisionMLP(nn.Module):

243 def __init__(

244 self,

245 - config: PretrainedConfig,

245 + config: PreTrainedConfig,

246 quant_config: QuantizationConfig | None = None,

247 prefix: str = "",

248 ) -> None:

@@ -276,7 +276,7 @@ class InternS1VisionMLP(nn.Module):

276 class InternS1VisionLayer(nn.Module):

277 def __init__(

278 self,

279 - config: PretrainedConfig,

279 + config: PreTrainedConfig,

280 quant_config: QuantizationConfig | None = None,

281 *,

282 num_dummy_heads: int = 0,

@@ -311,7 +311,7 @@ class InternS1VisionLayer(nn.Module):

311

312 def _init_attn(

313 self,

314 - config: PretrainedConfig,

314 + config: PreTrainedConfig,

315 quant_config: QuantizationConfig | None,

316 *,

317 num_dummy_heads: int,

@@ -343,7 +343,7 @@ class InternS1VisionLayer(nn.Module):

343 class InternS1VisionEncoder(nn.Module):

344 def __init__(

345 self,

346 - config: PretrainedConfig,

346 + config: PreTrainedConfig,

347 quant_config: QuantizationConfig | None = None,

348 *,

349 num_hidden_layers_override: int | None = None,

@@ -382,7 +382,7 @@ class InternS1VisionEncoder(nn.Module):

382 class InternS1VisionModel(nn.Module):

383 def __init__(

384 self,

385 - config: PretrainedConfig,

385 + config: PreTrainedConfig,

386 quant_config: QuantizationConfig | None = None,

387 *,

388 num_hidden_layers_override: int | None = None,

vllm/model_executor/models/internvl.py

+4 -4

@@ -14,7 +14,7 @@ from typing import Annotated, Any, Literal, TypeAlias, TypedDict, TypeVar

14

15 import torch

16 import torch.nn as nn

17 -from transformers import BatchFeature, PretrainedConfig

17 +from transformers import BatchFeature, PreTrainedConfig

18

19 from vllm.config import VllmConfig

20 from vllm.config.multimodal import MultiModalDummyOptions

@@ -629,7 +629,7 @@ class InternVLChatModel(

629 )

630

631 def _patch_quant_config(

632 - self, config: PretrainedConfig, quant_config: QuantizationConfig | None

632 + self, config: PreTrainedConfig, quant_config: QuantizationConfig | None

633 ):

634 # the awq models from OpenGVLab missing `modules_to_not_convert`

635 # patch the quant_config to add `modules_to_not_convert` back

@@ -643,7 +643,7 @@ class InternVLChatModel(

643

644 def _init_vision_model(

645 self,

646 - config: PretrainedConfig,

646 + config: PreTrainedConfig,

647 quant_config: QuantizationConfig | None,

648 *,

649 prefix: str,

@@ -663,7 +663,7 @@ class InternVLChatModel(

663 prefix=prefix,

664 )

665

666 - def _init_mlp1(self, config: PretrainedConfig) -> nn.Module:

666 + def _init_mlp1(self, config: PreTrainedConfig) -> nn.Module:

667 vit_hidden_size = config.vision_config.hidden_size

668 llm_hidden_size = config.text_config.hidden_size

669

vllm/model_executor/models/iquest_loopcoder.py

+3 -3

@@ -22,7 +22,7 @@ from typing import Any

22

23 import torch

24 from torch import nn

25 -from transformers import PretrainedConfig

25 +from transformers import PreTrainedConfig

26

27 from vllm.compilation.decorators import support_torch_compile

28 from vllm.config import CacheConfig, VllmConfig

@@ -60,7 +60,7 @@ from .utils import (

60 class LoopCoderAttention(nn.Module):

61 def __init__(

62 self,

63 - config: PretrainedConfig,

63 + config: PreTrainedConfig,

64 hidden_size: int,

65 num_heads: int,

66 num_kv_heads: int,

@@ -210,7 +210,7 @@ class LoopCoderAttention(nn.Module):

210 class LoopCoderDecoderLayer(nn.Module):

211 def __init__(

212 self,

213 - config: PretrainedConfig,

213 + config: PreTrainedConfig,

214 cache_config: CacheConfig | None = None,

215 quant_config: QuantizationConfig | None = None,

216 prefix: str = "",

vllm/model_executor/models/keye.py

+15 -15

@@ -10,7 +10,7 @@ import numpy as np

10 import torch

11 import torch.nn as nn

12 from einops import rearrange

13 -from transformers import BaseImageProcessor, PretrainedConfig

13 +from transformers import BaseImageProcessor, PreTrainedConfig

14 from transformers.activations import GELUActivation

15 from transformers.feature_extraction_utils import BatchFeature

16 from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling

@@ -191,7 +191,7 @@ KeyeVideoInputs: TypeAlias = KeyeVideoPixelInputs | KeyeVideoEmbeddingInputs

191

192

193 class KeyeVisionEmbeddings(nn.Module):

194 - def __init__(self, config: PretrainedConfig):

194 + def __init__(self, config: PreTrainedConfig):

195 super().__init__()

196 self.config = config

197 self.embed_dim = config.hidden_size

@@ -349,7 +349,7 @@ class KeyeSiglipAttention(nn.Module):

349

350 def __init__(

351 self,

352 - config: PretrainedConfig,

352 + config: PreTrainedConfig,

353 quant_config: QuantizationConfig | None = None,

354 prefix: str = "",

355 ):

@@ -487,7 +487,7 @@ class SigLIPRotaryEmbedding(nn.Module):

487 class KeyeSiglipEncoderLayer(nn.Module):

488 def __init__(

489 self,

490 - config: PretrainedConfig,

490 + config: PreTrainedConfig,

491 quant_config: QuantizationConfig | None = None,

492 prefix: str = "",

493 ):

@@ -539,7 +539,7 @@ class KeyeSiglipEncoderLayer(nn.Module):

539 class KeyeSiglipEncoder(nn.Module):

540 def __init__(

541 self,

542 - config: PretrainedConfig,

542 + config: PreTrainedConfig,

543 quant_config: QuantizationConfig | None = None,

544 prefix: str = "",

545 ):

@@ -634,7 +634,7 @@ class KeyeSiglipEncoder(nn.Module):

634 class KeyeSiglipVisionTransformer(nn.Module):

635 def __init__(

636 self,

637 - config: PretrainedConfig,

637 + config: PreTrainedConfig,

638 quant_config: QuantizationConfig | None = None,

639 prefix: str = "",

640 ):

@@ -710,7 +710,7 @@ class KeyeSiglipVisionTransformer(nn.Module):

710

711

712 class KeyeSiglipVisionModel(nn.Module):

713 - config_class = PretrainedConfig

713 + config_class = PreTrainedConfig

714 main_input_name = "pixel_values"

715

716 hf_to_vllm_mapper = WeightsMapper(

@@ -724,7 +724,7 @@ class KeyeSiglipVisionModel(nn.Module):

724

725 def __init__(

726 self,

727 - config: PretrainedConfig,

727 + config: PreTrainedConfig,

728 quant_config: QuantizationConfig | None = None,

729 prefix: str = "",

730 ):

@@ -787,8 +787,8 @@ class KeyeSiglipVisionModel(nn.Module):

787 class Projector(nn.Module):

788 def __init__(

789 self,

790 - text_config: PretrainedConfig,

791 - vision_config: PretrainedConfig,

790 + text_config: PreTrainedConfig,

791 + vision_config: PreTrainedConfig,

792 quant_config: QuantizationConfig | None = None,

793 prefix: str = "",

794 ):

@@ -1232,7 +1232,7 @@ class BaseKeyeModule(nn.Module, SupportsMultiModal):

1232

1233 def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):

1234 super().__init__()

1235 - config: PretrainedConfig = vllm_config.model_config.hf_config

1235 + config: PreTrainedConfig = vllm_config.model_config.hf_config

1236 quant_config = vllm_config.quant_config

1237

1238 self.config = config

@@ -1264,8 +1264,8 @@ class BaseKeyeModule(nn.Module, SupportsMultiModal):

1264 @abstractmethod

1265 def _build_projector(

1266 self,

1267 - text_config: PretrainedConfig,

1268 - vision_config: PretrainedConfig,

1267 + text_config: PreTrainedConfig,

1268 + vision_config: PreTrainedConfig,

1269 quant_config: QuantizationConfig | None = None,

1270 prefix: str = "",

1271 ) -> nn.Module:

@@ -1511,8 +1511,8 @@ class KeyeForConditionalGeneration(

1511 ):

1512 def _build_projector(

1513 self,

1514 - text_config: PretrainedConfig,

1515 - vision_config: PretrainedConfig,

1514 + text_config: PreTrainedConfig,

1515 + vision_config: PreTrainedConfig,

1516 quant_config: QuantizationConfig | None = None,

1517 prefix: str = "",

1518 ) -> nn.Module:

vllm/model_executor/models/keye_vl1_5.py

+6 -6

@@ -9,7 +9,7 @@ import numpy as np

9 import torch

10 import torch.nn as nn

11 from einops import rearrange

12 -from transformers import PretrainedConfig

12 +from transformers import PreTrainedConfig

13 from transformers.activations import GELUActivation

14 from transformers.feature_extraction_utils import BatchFeature

15

@@ -197,8 +197,8 @@ KeyeVL1_5VideoInputs: TypeAlias = (

197 class KeyeVL1_5Projector(nn.Module):

198 def __init__(

199 self,

200 - text_config: PretrainedConfig,

201 - vision_config: PretrainedConfig,

200 + text_config: PreTrainedConfig,

201 + vision_config: PreTrainedConfig,

202 quant_config: QuantizationConfig | None = None,

203 prefix: str = "",

204 ):

@@ -509,15 +509,15 @@ class KeyeVL1_5ForConditionalGeneration(

509 ):

510 def _build_projector(

511 self,

512 - text_config: PretrainedConfig,

513 - vision_config: PretrainedConfig,

512 + text_config: PreTrainedConfig,

513 + vision_config: PreTrainedConfig,

514 quant_config: QuantizationConfig | None = None,

515 prefix: str = "",

516 ) -> nn.Module:

517 return KeyeVL1_5Projector(text_config, vision_config, quant_config, prefix)

518

519 def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):

520 - config: PretrainedConfig = vllm_config.model_config.hf_config

520 + config: PreTrainedConfig = vllm_config.model_config.hf_config

521 self.merge_size = config.vision_config.spatial_merge_size

522 super().__init__(vllm_config=vllm_config, prefix=prefix)

523

vllm/model_executor/models/lfm2_siglip2.py

+1 -1

@@ -334,7 +334,7 @@ class Siglip2Encoder(nn.Module):

334 self attention layers. Each layer is a [`Siglip2EncoderLayer`].

335

336 Args:

337 - config: PretrainedConfig

337 + config: PreTrainedConfig

338

339 """

340

vllm/model_executor/models/lfm2_vl.py

+5 -5

@@ -13,8 +13,8 @@ from transformers import BatchFeature

13 from transformers.activations import ACT2FN

14 from transformers.models.lfm2_vl import Lfm2VlProcessor

15 from transformers.models.lfm2_vl.configuration_lfm2_vl import Lfm2VlConfig

16 -from transformers.models.lfm2_vl.image_processing_lfm2_vl_fast import (

17 - Lfm2VlImageProcessorFast,

16 +from transformers.models.lfm2_vl.image_processing_lfm2_vl import (

17 + Lfm2VlImageProcessor,

18 find_closest_aspect_ratio,

19 round_by_factor,

20 )

@@ -104,7 +104,7 @@ class Lfm2VLProcessingInfo(BaseProcessingInfo):

104 def get_hf_processor(self, **kwargs):

105 return self.ctx.get_hf_processor(Lfm2VlProcessor, **kwargs)

106

107 - def get_image_processor(self, **kwargs: object) -> Lfm2VlImageProcessorFast:

107 + def get_image_processor(self, **kwargs: object) -> Lfm2VlImageProcessor:

108 return self.get_hf_processor(**kwargs).image_processor

109

110 def get_default_tok_params(self) -> TokenizeParams:

@@ -213,7 +213,7 @@ class Lfm2VLProcessingInfo(BaseProcessingInfo):

213 processor: Lfm2VlProcessor,

214 mm_kwargs: Mapping[str, object],

215 ) -> tuple[int, int, int]:

216 - image_processor: Lfm2VlImageProcessorFast = processor.image_processor

216 + image_processor: Lfm2VlImageProcessor = processor.image_processor

217

218 mm_kwargs = self.ctx.get_merged_mm_kwargs(mm_kwargs)

219 downsample_factor = mm_kwargs.get(

@@ -336,7 +336,7 @@ class Lfm2VLProcessingInfo(BaseProcessingInfo):

336 processor: Lfm2VlProcessor,

337 mm_kwargs: Mapping[str, object],

338 ) -> tuple[int, int]:

339 - image_processor: Lfm2VlImageProcessorFast = processor.image_processor

339 + image_processor: Lfm2VlImageProcessor = processor.image_processor

340

341 mm_kwargs = self.ctx.get_merged_mm_kwargs(mm_kwargs)

342 downsample_factor = mm_kwargs.get(

vllm/model_executor/models/llava.py

+2 -2

@@ -12,7 +12,7 @@ from transformers import (

12 CLIPVisionConfig,

13 LlavaConfig,

14 PixtralVisionConfig,

15 - PretrainedConfig,

15 + PreTrainedConfig,

16 SiglipVisionConfig,

17 )

18 from transformers.models.llava import LlavaProcessor

@@ -156,7 +156,7 @@ class LlavaMultiModalProjector(nn.Module):

156

157

158 class LlavaLikeConfig(Protocol):

159 - vision_config: PretrainedConfig

159 + vision_config: PreTrainedConfig

160 image_token_index: int

161 vision_feature_select_strategy: str

162 vision_feature_layer: int | list[int]

vllm/model_executor/models/llava_onevision2.py

-1

@@ -1307,7 +1307,6 @@ class LlavaOnevision2ProcessingInfo(BaseProcessingInfo):

1307 # ``_merge_mm_kwargs`` restricts caller ``mm_processor_kwargs`` to known

1308 # processor args and wraps values as hashable for the lru_cache.

1309 merged = _merge_mm_kwargs(model_config, AutoProcessor, **kwargs)

1310 - merged.setdefault("use_fast", True)

1310 return _load_ov2_processor(

1311 model_config.model,

1312 model_config.revision,

vllm/model_executor/models/longcat_flash.py

+2 -2

@@ -39,7 +39,7 @@ from itertools import islice

39

40 import torch

41 from torch import nn

42 -from transformers import PretrainedConfig

42 +from transformers import PreTrainedConfig

43

44 from vllm.compilation.decorators import support_torch_compile

45 from vllm.config import CacheConfig, VllmConfig

@@ -80,7 +80,7 @@ from .utils import (

80 logger = init_logger(__name__)

81

82

83 -class FlashConfig(PretrainedConfig):

83 +class FlashConfig(PreTrainedConfig):

84 """Flash model configuration."""

85

86 moe_intermediate_size: int

vllm/model_executor/models/longcat_flash_mtp.py

+3 -3

@@ -7,7 +7,7 @@ from collections.abc import Iterable

7

8 import torch

9 import torch.nn as nn

10 -from transformers import PretrainedConfig

10 +from transformers import PreTrainedConfig

11

12 from vllm.config import VllmConfig

13 from vllm.model_executor.layers.layernorm import RMSNorm

@@ -30,7 +30,7 @@ from .utils import maybe_prefix

30 class LongCatMultiTokenPredictorLayer(nn.Module):

31 def __init__(

32 self,

33 - config: PretrainedConfig,

33 + config: PreTrainedConfig,

34 prefix: str,

35 vllm_config: VllmConfig,

36 quant_config: QuantizationConfig | None = None,

@@ -353,7 +353,7 @@ class LongCatFlashMTP(nn.Module):

353 return name

354

355 def get_spec_layer_idx_from_weight_name(

356 - self, config: PretrainedConfig, weight_name: str

356 + self, config: PreTrainedConfig, weight_name: str

357 ) -> int | None:

358 if "model.mtp" in weight_name:

359 return config.num_hidden_layers * 2

vllm/model_executor/models/mimo_audio.py

+2 -2

@@ -21,7 +21,7 @@ import torch.nn as nn

21 import torch.nn.functional as F

22 from einops import rearrange, repeat

23 from transformers.activations import ACT2FN

24 -from transformers.configuration_utils import PretrainedConfig

24 +from transformers.configuration_utils import PreTrainedConfig

25 from transformers.modeling_utils import PreTrainedModel

26 from transformers.models.qwen2.configuration_qwen2 import Qwen2Config

27 from transformers.models.qwen2.modeling_qwen2 import Qwen2Model

@@ -518,7 +518,7 @@ class ResidualVectorQuantizer(nn.Module):

518 # ---------------------------------------------------------------------------

519

520

521 -class MiMoAudioTokenizerConfig(PretrainedConfig):

521 +class MiMoAudioTokenizerConfig(PreTrainedConfig):

522 model_type = "mimo_audio_tokenizer"

523

524 def __init__(

vllm/model_executor/models/mimo_mtp.py

+2 -2

@@ -24,7 +24,7 @@ from collections.abc import Iterable

24

25 import torch

26 import torch.nn as nn

27 -from transformers import PretrainedConfig

27 +from transformers import PreTrainedConfig

28

29 from vllm.config import CacheConfig, ModelConfig, VllmConfig

30 from vllm.model_executor.layers.layernorm import RMSNorm

@@ -43,7 +43,7 @@ from .utils import AutoWeightsLoader, WeightsMapper, maybe_prefix

43 class MiMoMultiTokenPredictorLayer(nn.Module):

44 def __init__(

45 self,

46 - config: PretrainedConfig,

46 + config: PreTrainedConfig,

47 prefix: str,

48 model_config: ModelConfig,

49 cache_config: CacheConfig | None = None,

vllm/model_executor/models/mimo_v2_mtp.py

+3 -3

@@ -22,7 +22,7 @@ from collections.abc import Iterable

22

23 import torch

24 import torch.nn as nn

25 -from transformers import PretrainedConfig

25 +from transformers import PreTrainedConfig

26

27 from vllm.config import VllmConfig

28 from vllm.distributed import (

@@ -62,7 +62,7 @@ class MiMoV2MTPLayer(nn.Module):

62

63 def __init__(

64 self,

65 - config: PretrainedConfig,

65 + config: PreTrainedConfig,

66 prefix: str,

67 quant_config: QuantizationConfig | None = None,

68 ) -> None:

@@ -145,7 +145,7 @@ class _MiMoV2MTPLayers(nn.Module):

145

146 def __init__(

147 self,

148 - config: PretrainedConfig,

148 + config: PreTrainedConfig,

149 num_mtp_layers: int,

150 quant_config: QuantizationConfig | None,

151 prefix: str,

vllm/model_executor/models/mimo_v2_omni.py

+2 -2

@@ -10,7 +10,7 @@ import numpy as np

10 import torch

11 import torch.nn as nn

12 import torch.nn.functional as F

13 -from transformers import BatchFeature, PretrainedConfig

13 +from transformers import BatchFeature, PreTrainedConfig

14 from transformers.models.qwen2_vl.image_processing_qwen2_vl import smart_resize

15 from typing_extensions import TypedDict

16

@@ -418,7 +418,7 @@ class MiMoVisionTransformer(nn.Module):

418

419 def __init__(

420 self,

421 - vision_cfg: PretrainedConfig,

421 + vision_cfg: PreTrainedConfig,

422 *,

423 norm_eps: float = 1e-6,

424 quant_config: QuantizationConfig | None = None,

vllm/model_executor/models/minicpm.py

+3 -3

@@ -31,7 +31,7 @@ from typing import Any

31

32 import torch

33 from torch import nn

34 -from transformers import PretrainedConfig

34 +from transformers import PreTrainedConfig

35

36 from vllm.compilation.decorators import support_torch_compile

37 from vllm.config import CacheConfig, VllmConfig

@@ -316,7 +316,7 @@ class MiniCPMAttention(nn.Module):

316 class MiniCPMDecoderLayer(nn.Module):

317 def __init__(

318 self,

319 - config: PretrainedConfig,

319 + config: PreTrainedConfig,

320 cache_config: CacheConfig | None = None,

321 quant_config: QuantizationConfig | None = None,

322 prefix: str = "",

@@ -427,7 +427,7 @@ class MiniCPMModel(nn.Module, EagleModelMixin):

427 def _init_layers(

428 self,

429 prefix: str,

430 - config: PretrainedConfig,

430 + config: PreTrainedConfig,

431 cache_config: CacheConfig | None,

432 quant_config: QuantizationConfig | None,

433 ):

vllm/model_executor/models/minicpm3.py

+3 -3

@@ -27,7 +27,7 @@

27

28 import torch

29 from torch import nn

30 -from transformers import PretrainedConfig

30 +from transformers import PreTrainedConfig

31

32 from vllm.config import CacheConfig, VllmConfig

33 from vllm.distributed import get_tensor_model_parallel_world_size

@@ -52,7 +52,7 @@ from .utils import make_layers

52 class MiniCPM3Attention(nn.Module):

53 def __init__(

54 self,

55 - config: PretrainedConfig,

55 + config: PreTrainedConfig,

56 hidden_size: int,

57 num_heads: int,

58 qk_nope_head_dim: int,

@@ -208,7 +208,7 @@ class MiniCPM3Model(MiniCPMModel):

208 def _init_layers(

209 self,

210 prefix: str,

211 - config: PretrainedConfig,

211 + config: PreTrainedConfig,

212 cache_config: CacheConfig | None,

213 quant_config: QuantizationConfig | None,

214 ):

vllm/model_executor/models/minicpm_eagle.py

+3 -3

@@ -30,7 +30,7 @@ from typing import TYPE_CHECKING

30

31 import torch

32 from torch import nn

33 -from transformers import PretrainedConfig

33 +from transformers import PreTrainedConfig

34

35 from vllm.compilation.decorators import support_torch_compile

36 from vllm.config import CacheConfig, VllmConfig

@@ -68,7 +68,7 @@ else:

68 class EagleMiniCPMDecoderLayer(nn.Module):

69 def __init__(

70 self,

71 - config: PretrainedConfig,

71 + config: PreTrainedConfig,

72 cache_config: CacheConfig | None = None,

73 quant_config: QuantizationConfig | None = None,

74 prefix: str = "",

@@ -189,7 +189,7 @@ class EagleMiniCPMModel(nn.Module):

189 def _init_layers(

190 self,

191 prefix: str,

192 - config: PretrainedConfig,

192 + config: PreTrainedConfig,

193 cache_config: CacheConfig | None,

194 quant_config: QuantizationConfig | None,

195 start_layer: int,

vllm/model_executor/models/minicpmv.py

+9 -9

@@ -36,7 +36,7 @@ import torch

36 import torch.types

37 from torch import nn

38 from torch.nn.init import trunc_normal_

39 -from transformers import BatchFeature, PretrainedConfig

39 +from transformers import BatchFeature, PreTrainedConfig

40 from transformers.dynamic_module_utils import (

41 get_class_from_dynamic_module,

42 resolve_trust_remote_code,

@@ -468,7 +468,7 @@ class Resampler4_5(Resampler2_5):

468 return x

469

470

471 -def get_version_by_config(config: PretrainedConfig) -> tuple[int, ...]:

471 +def get_version_by_config(config: PreTrainedConfig) -> tuple[int, ...]:

472 version_float = getattr(config, "version", None)

473

474 # The old configs do not include version number

@@ -1235,7 +1235,7 @@ class MiniCPMVBaseModel(nn.Module, SupportsMultiModal, SupportsPP):

1235 self.use_data_parallel = multimodal_config.mm_encoder_tp_mode == "data"

1236 super().__init__()

1237 # All MiniCPM-V models disable `tie_word_embeddings` but

1238 - # `PretrainedConfig.tie_word_embeddings` defaults to True; we cannot

1238 + # `PreTrainedConfig.tie_word_embeddings` defaults to True; we cannot

1239 # check `tie_word_embeddings` until vLLM integrate MiniCPM-V model

1240 # and config class

1241 self.config = config

@@ -1425,7 +1425,7 @@ class MiniCPMVBaseModel(nn.Module, SupportsMultiModal, SupportsPP):

1425

1426 def init_vision_module(

1427 self,

1428 - config: PretrainedConfig,

1428 + config: PreTrainedConfig,

1429 quant_config: QuantizationConfig | None,

1430 prefix: str = "",

1431 ) -> nn.Module:

@@ -1957,7 +1957,7 @@ class MiniCPMV2_0(MiniCPMVBaseModel):

1957

1958 def init_vision_module(

1959 self,

1960 - config: PretrainedConfig,

1960 + config: PreTrainedConfig,

1961 quant_config: QuantizationConfig | None,

1962 prefix: str = "",

1963 ) -> nn.Module:

@@ -2058,7 +2058,7 @@ class MiniCPMV2_5(_MiniCPMVEncoderCudaGraphMixin, SupportsLoRA):

2058

2059 def init_vision_module(

2060 self,

2061 - config: PretrainedConfig,

2061 + config: PreTrainedConfig,

2062 quant_config: QuantizationConfig | None,

2063 prefix: str = "",

2064 ) -> nn.Module:

@@ -2154,7 +2154,7 @@ class MiniCPMV2_6(_MiniCPMVEncoderCudaGraphMixin, SupportsLoRA):

2154

2155 def init_vision_module(

2156 self,

2157 - config: PretrainedConfig,

2157 + config: PreTrainedConfig,

2158 quant_config: QuantizationConfig | None = None,

2159 prefix: str = "",

2160 ) -> nn.Module:

@@ -2255,7 +2255,7 @@ class MiniCPMV4_0(_MiniCPMVEncoderCudaGraphMixin, SupportsLoRA):

2255

2256 def init_vision_module(

2257 self,

2258 - config: PretrainedConfig,

2258 + config: PreTrainedConfig,

2259 quant_config: QuantizationConfig | None = None,

2260 prefix: str = "",

2261 ) -> nn.Module:

@@ -2354,7 +2354,7 @@ class MiniCPMV4_5(MiniCPMVBaseModel, SupportsLoRA):

2354

2355 def init_vision_module(

2356 self,

2357 - config: PretrainedConfig,

2357 + config: PreTrainedConfig,

2358 quant_config: QuantizationConfig | None = None,

2359 prefix: str = "",

2360 ) -> nn.Module:

vllm/model_executor/models/minimax_m2.py

+3 -3

@@ -29,7 +29,7 @@ from typing import Any

29

30 import torch

31 from torch import nn

32 -from transformers import PretrainedConfig

32 +from transformers import PreTrainedConfig

33

34 from vllm.compilation.decorators import support_torch_compile

35 from vllm.config import CacheConfig, ModelConfig, VllmConfig

@@ -70,7 +70,7 @@ from .utils import (

70 class MiniMaxM2MoE(nn.Module):

71 def __init__(

72 self,

73 - config: PretrainedConfig,

73 + config: PreTrainedConfig,

74 quant_config: QuantizationConfig | None = None,

75 prefix: str = "",

76 ):

@@ -249,7 +249,7 @@ class MiniMaxM2Attention(nn.Module):

249 class MiniMaxM2DecoderLayer(nn.Module):

250 def __init__(

251 self,

252 - config: PretrainedConfig,

252 + config: PreTrainedConfig,

253 prefix: str,

254 model_config: ModelConfig,

255 cache_config: CacheConfig | None = None,

vllm/model_executor/models/mllama4.py

+2 -4

@@ -25,7 +25,7 @@ from torch import nn

25 from transformers import BatchFeature, Llama4Config, Llama4VisionConfig

26 from transformers.image_utils import SizeDict

27 from transformers.models.llama4 import Llama4Processor

28 -from transformers.models.llama4.image_processing_llama4_fast import (

28 +from transformers.models.llama4.image_processing_llama4 import (

29 find_supported_resolutions,

30 get_best_fit,

31 )

@@ -554,9 +554,7 @@ class Mllama4ProcessingInfo(BaseProcessingInfo):

554 return self.ctx.get_hf_config(Llama4Config)

555

556 def get_hf_processor(self, **kwargs: object) -> Llama4Processor:

557 - return self.ctx.get_hf_processor(

558 - Llama4Processor, use_fast=kwargs.pop("use_fast", True), **kwargs

559 - )

557 + return self.ctx.get_hf_processor(Llama4Processor, **kwargs)

558

559 def get_supported_mm_limits(self) -> Mapping[str, int | None]:

560 # Although vLLM can support more images from an infra capability

vllm/model_executor/models/molmo.py

+6 -6

@@ -16,7 +16,7 @@ from einops import rearrange

16 from transformers import (

17 BaseImageProcessor,

18 BatchFeature,

19 - PretrainedConfig,

19 + PreTrainedConfig,

20 )

21

22 from vllm.compilation.decorators import support_torch_compile

@@ -405,7 +405,7 @@ class MolmoAttention(nn.Module):

405

406 def __init__(

407 self,

408 - config: PretrainedConfig,

408 + config: PreTrainedConfig,

409 cache_config: CacheConfig | None = None,

410 quant_config: QuantizationConfig | None = None,

411 prefix: str = "",

@@ -515,7 +515,7 @@ class LanguageModelMLP(nn.Module):

515

516 def __init__(

517 self,

518 - config: PretrainedConfig,

518 + config: PreTrainedConfig,

519 input_dim: int | None = None,

520 quant_config: QuantizationConfig | None = None,

521 prefix: str = "",

@@ -557,7 +557,7 @@ class ImageProjectorMLP(nn.Module):

557

558 def __init__(

559 self,

560 - config: PretrainedConfig,

560 + config: PreTrainedConfig,

561 input_dim: int | None = None,

562 quant_config: QuantizationConfig | None = None,

563 prefix: str = "",

@@ -598,7 +598,7 @@ class ImageProjectorMLP(nn.Module):

598 class MolmoDecoderLayer(nn.Module):

599 def __init__(

600 self,

601 - config: PretrainedConfig,

601 + config: PreTrainedConfig,

602 cache_config: CacheConfig | None = None,

603 quant_config: QuantizationConfig | None = None,

604 prefix: str = "",

@@ -681,7 +681,7 @@ class MolmoVisionBackbone(nn.Module, SupportsQuant):

681

682 def __init__(

683 self,

684 - config: PretrainedConfig,

684 + config: PreTrainedConfig,

685 vision_config: VisionBackboneConfig,

686 quant_config: QuantizationConfig | None = None,

687 prefix: str = "",

vllm/model_executor/models/molmo2.py

+3 -3

@@ -16,7 +16,7 @@ from transformers import (

16 BaseImageProcessor,

17 BaseVideoProcessor,

18 BatchFeature,

19 - PretrainedConfig,

19 + PreTrainedConfig,

20 ProcessorMixin,

21 )

22 from transformers.image_utils import ImageInput

@@ -1305,7 +1305,7 @@ def exif_transpose(

1305

1306 def build_flat_image_bool_length(

1307 image_grids: torch.LongTensor,

1308 - hf_config: PretrainedConfig,

1308 + hf_config: PreTrainedConfig,

1309 image_use_col_tokens: bool = True,

1310 use_single_crop_col_tokens: bool | None = None,

1311 use_single_crop_start_token: bool = True,

@@ -1393,7 +1393,7 @@ def build_flat_image_bool_length(

1393

1394 def build_flat_video_bool_length(

1395 video_grids: torch.LongTensor,

1396 - hf_config: PretrainedConfig,

1396 + hf_config: PreTrainedConfig,

1397 ) -> tuple[torch.LongTensor, torch.LongTensor]:

1398 image_patch_id = hf_config.image_patch_id

1399 frame_start_id = hf_config.frame_start_token_id

vllm/model_executor/models/moss_audio.py

+2 -2

@@ -13,7 +13,7 @@ import regex as re

13 import torch

14 import torch.nn.functional as F

15 from torch import nn

16 -from transformers import BatchFeature, PretrainedConfig, Qwen3Config

16 +from transformers import BatchFeature, PreTrainedConfig, Qwen3Config

17 from transformers.models.whisper import WhisperFeatureExtractor

18 from typing_extensions import TypedDict

19

@@ -263,7 +263,7 @@ class MossAudioEncoderConfig:

263 return cls(**values)

264

265

266 -class MossAudioConfig(PretrainedConfig):

266 +class MossAudioConfig(PreTrainedConfig):

267 model_type = "moss_audio"

268 is_composition = True

269

vllm/model_executor/models/nano_nemotron_vl.py

+2 -2

@@ -16,7 +16,7 @@ from typing import Annotated, Literal, TypeAlias, TypedDict, cast

16

17 import torch

18 import torch.nn as nn

19 -from transformers import BatchFeature, PretrainedConfig

19 +from transformers import BatchFeature, PreTrainedConfig

20

21 from vllm import envs

22 from vllm.config import VllmConfig

@@ -238,7 +238,7 @@ class NanoNemotronVLProcessingInfo(BaseProcessingInfo):

238 return self.ctx.get_mm_config().video_pruning_rate

239

240 @property

241 - def sound_config(self) -> PretrainedConfig | None:

241 + def sound_config(self) -> PreTrainedConfig | None:

242 return getattr(self.get_hf_config(), "sound_config", None)

243

244 def get_default_tok_params(self) -> TokenizeParams:

vllm/model_executor/models/nemotron_parse.py

+3 -3

@@ -17,7 +17,7 @@ from einops import rearrange

17 from transformers import (

18 BartConfig,

19 BatchFeature,

20 - PretrainedConfig,

20 + PreTrainedConfig,

21 )

22

23 from vllm.config import CacheConfig, VllmConfig

@@ -440,7 +440,7 @@ class RadioWithNeck(nn.Module):

440

441 def __init__(

442 self,

443 - config: PretrainedConfig,

443 + config: PreTrainedConfig,

444 quant_config: QuantizationConfig | None = None,

445 prefix: str = "",

446 ):

@@ -480,7 +480,7 @@ class RadioWithNeck(nn.Module):

480

481 def get_vit_model_from_radio_config(

482 self,

483 - hf_config: PretrainedConfig,

483 + hf_config: PreTrainedConfig,

484 quant_config: QuantizationConfig | None = None,

485 ) -> RadioModel:

486 hf_config_vision = hf_config.encoder

vllm/model_executor/models/nemotron_vl.py

+6 -6

@@ -5,7 +5,7 @@ from collections.abc import Iterable

5

6 import torch

7 import torch.nn as nn

8 -from transformers import AutoModel, PretrainedConfig

8 +from transformers import AutoModel, PreTrainedConfig

9

10 from vllm.config import VllmConfig

11 from vllm.model_executor.layers.linear import ReplicatedLinear

@@ -146,7 +146,7 @@ class LlamaNemotronVLChatModel(nn.Module, SupportsMultiModal, SupportsPP, Suppor

146 )

147

148 def _patch_quant_config(

149 - self, config: PretrainedConfig, quant_config: QuantizationConfig | None

149 + self, config: PreTrainedConfig, quant_config: QuantizationConfig | None

150 ):

151 # the awq models from OpenGVLab missing `modules_to_not_convert`

152 # patch the quant_config to add `modules_to_not_convert` back

@@ -160,7 +160,7 @@ class LlamaNemotronVLChatModel(nn.Module, SupportsMultiModal, SupportsPP, Suppor

160

161 def _init_vision_model(

162 self,

163 - config: PretrainedConfig,

163 + config: PreTrainedConfig,

164 quant_config: QuantizationConfig | None,

165 *,

166 prefix: str,

@@ -172,7 +172,7 @@ class LlamaNemotronVLChatModel(nn.Module, SupportsMultiModal, SupportsPP, Suppor

172

173 def _init_mlp1(

174 self,

175 - config: PretrainedConfig,

175 + config: PreTrainedConfig,

176 vit_hidden_size: int | None = None,

177 vision_projection_hidden_size: int | None = None,

178 ) -> nn.Module:

@@ -505,7 +505,7 @@ class LlamaNemotronVLForEmbedding(LlamaNemotronVLChatModel, VllmModelForPooling)

505

506 def _init_vision_model(

507 self,

508 - config: PretrainedConfig,

508 + config: PreTrainedConfig,

509 quant_config,

510 *,

511 prefix: str,

@@ -520,7 +520,7 @@ class LlamaNemotronVLForEmbedding(LlamaNemotronVLChatModel, VllmModelForPooling)

520

521 def _init_mlp1(

522 self,

523 - config: PretrainedConfig,

523 + config: PreTrainedConfig,

524 vit_hidden_size: int | None = None,

525 vision_projection_hidden_size: int | None = None,

526 ) -> nn.Module:

vllm/model_executor/models/nvlm_d.py

+3 -3

@@ -11,7 +11,7 @@ from collections.abc import Mapping

11

12 import torch

13 import torch.nn as nn

14 -from transformers import PretrainedConfig

14 +from transformers import PreTrainedConfig

15

16 from vllm.config.multimodal import MultiModalDummyOptions

17 from vllm.inputs import MultiModalDataDict

@@ -161,7 +161,7 @@ class NVLMMultiModalProcessor(BaseInternVLMultiModalProcessor[NVLMProcessingInfo

161 dummy_inputs=NVLMDummyInputsBuilder,

162 )

163 class NVLM_D_Model(InternVLChatModel):

164 - def _init_mlp1(self, config: PretrainedConfig) -> nn.Module:

164 + def _init_mlp1(self, config: PreTrainedConfig) -> nn.Module:

165 vit_hidden_size = config.vision_config.hidden_size

166 llm_intermediate_size = config.text_config.intermediate_size

167 llm_hidden_size = config.text_config.hidden_size

@@ -179,7 +179,7 @@ class NVLM_D_Model(InternVLChatModel):

179

180 def _init_vision_model(

181 self,

182 - config: PretrainedConfig,

182 + config: PreTrainedConfig,

183 quant_config: QuantizationConfig | None,

184 *,

185 prefix: str,

vllm/model_executor/models/openpangu.py

+8 -8

@@ -26,7 +26,7 @@ from typing import Any

26

27 import torch

28 from torch import nn

29 -from transformers import PretrainedConfig

29 +from transformers import PreTrainedConfig

30

31 from vllm.compilation.decorators import support_torch_compile

32 from vllm.config import CacheConfig, ParallelConfig, VllmConfig

@@ -134,7 +134,7 @@ class OpenPanguMLP(nn.Module):

134 class OpenPanguMoE(nn.Module):

135 def __init__(

136 self,

137 - config: PretrainedConfig,

137 + config: PreTrainedConfig,

138 parallel_config: ParallelConfig,

139 quant_config: QuantizationConfig | None = None,

140 prefix: str = "",

@@ -240,7 +240,7 @@ class OpenPanguMoE(nn.Module):

240 class OpenPanguMLAAttention(nn.Module):

241 def __init__(

242 self,

243 - config: PretrainedConfig,

243 + config: PreTrainedConfig,

244 hidden_size: int,

245 num_heads: int,

246 qk_nope_head_dim: int,

@@ -389,7 +389,7 @@ class OpenPanguMLAAttention(nn.Module):

389 class OpenPanguEmbeddedAttention(nn.Module):

390 def __init__(

391 self,

392 - config: PretrainedConfig,

392 + config: PreTrainedConfig,

393 hidden_size: int,

394 num_heads: int,

395 num_kv_heads: int,

@@ -501,7 +501,7 @@ class OpenPanguEmbeddedAttention(nn.Module):

501

502 def _init_rotary_emb(

503 self,

504 - config: PretrainedConfig,

504 + config: PreTrainedConfig,

505 quant_config: QuantizationConfig | None,

506 ) -> None:

507 is_neox_style = True

@@ -522,7 +522,7 @@ class OpenPanguEmbeddedAttention(nn.Module):

522 class OpenPanguSinkAttention(nn.Module):

523 def __init__(

524 self,

525 - config: PretrainedConfig,

525 + config: PreTrainedConfig,

526 hidden_size: int,

527 num_heads: int,

528 num_kv_heads: int,

@@ -743,7 +743,7 @@ class OpenPanguSinkAttention(nn.Module):

743

744 def _init_rotary_emb(

745 self,

746 - config: PretrainedConfig,

746 + config: PreTrainedConfig,

747 rope_parameters: dict[str, Any] | None,

748 quant_config: QuantizationConfig | None,

749 ) -> None:

@@ -769,7 +769,7 @@ class OpenPanguSinkAttention(nn.Module):

769 class OpenPanguDecoderLayer(nn.Module):

770 def __init__(

771 self,

772 - config: PretrainedConfig,

772 + config: PreTrainedConfig,

773 prefix: str,

774 vllm_config: VllmConfig,

775 ) -> None:

vllm/model_executor/models/openpangu_vl.py

-1

@@ -686,7 +686,6 @@ class OpenPanguVLProcessingInfo(Qwen2_5_VLProcessingInfo):

686 kwargs["fps"] = fps

687

688 return self.ctx.get_hf_processor(

689 - use_fast=kwargs.pop("use_fast", True),

689 **kwargs,

690 )

691

vllm/model_executor/models/orion.py

+2 -2

@@ -13,7 +13,7 @@ from typing import Any

13

14 import torch

15 from torch import nn

16 -from transformers import PretrainedConfig

16 +from transformers import PreTrainedConfig

17

18 from vllm.compilation.decorators import support_torch_compile

19 from vllm.config import CacheConfig, VllmConfig

@@ -163,7 +163,7 @@ class OrionAttention(nn.Module):

163 class OrionDecoderLayer(nn.Module):

164 def __init__(

165 self,

166 - config: PretrainedConfig,

166 + config: PreTrainedConfig,

167 cache_config: CacheConfig | None = None,

168 quant_config: QuantizationConfig | None = None,

169 prefix: str = "",

vllm/model_executor/models/ovis.py

+4 -4

@@ -26,7 +26,7 @@ import torch

26 import torch.nn as nn

27 from torch import Tensor

28 from torch.nn.functional import gumbel_softmax, pad, softmax

29 -from transformers import BatchFeature, PretrainedConfig

29 +from transformers import BatchFeature, PreTrainedConfig

30

31 from vllm.config import VllmConfig

32 from vllm.config.multimodal import MultiModalDummyOptions

@@ -93,7 +93,7 @@ def st_argmax(y_soft: torch.Tensor, dim: int): # straight-through softmax

93 class VisualTokenizer(torch.nn.Module):

94 def __init__(

95 self,

96 - config: PretrainedConfig,

96 + config: PreTrainedConfig,

97 quant_config: QuantizationConfig | None = None,

98 prefix: str = "",

99 ):

@@ -120,7 +120,7 @@ class VisualTokenizer(torch.nn.Module):

120

121 def _init_backbone(

122 self,

123 - config: PretrainedConfig,

123 + config: PreTrainedConfig,

124 quant_config: QuantizationConfig | None = None,

125 prefix: str = "",

126 ) -> nn.Module:

@@ -419,7 +419,7 @@ class Ovis(nn.Module, SupportsMultiModal, SupportsPP):

419 config = vllm_config.model_config.hf_config

420 quant_config = vllm_config.quant_config

421

422 - self.config: PretrainedConfig = config

422 + self.config: PreTrainedConfig = config

423

424 with self._mark_language_model(vllm_config):

425 self.llm = init_vllm_registered_model(

vllm/model_executor/models/ovis2_5.py

+4 -4

@@ -8,7 +8,7 @@ from typing import Annotated, Literal, TypedDict

8

9 import torch

10 import torch.nn as nn

11 -from transformers import BaseImageProcessor, BatchFeature, PretrainedConfig

11 +from transformers import BaseImageProcessor, BatchFeature, PreTrainedConfig

12

13 from vllm.config import VllmConfig

14 from vllm.config.multimodal import MultiModalDummyOptions

@@ -95,7 +95,7 @@ class VisualTokenizer(torch.nn.Module):

95

96 def __init__(

97 self,

98 - config: PretrainedConfig,

98 + config: PreTrainedConfig,

99 visual_vocab_size: int,

100 quant_config: QuantizationConfig | None = None,

101 prefix: str = "",

@@ -121,7 +121,7 @@ class VisualTokenizer(torch.nn.Module):

121

122 def _init_backbone(

123 self,

124 - config: PretrainedConfig,

124 + config: PreTrainedConfig,

125 quant_config: QuantizationConfig | None = None,

126 prefix: str = "",

127 ):

@@ -438,7 +438,7 @@ class Ovis2_5(nn.Module, SupportsMultiModal, SupportsPP):

438 config = vllm_config.model_config.hf_config

439 quant_config = vllm_config.quant_config

440

441 - self.config: PretrainedConfig = config

441 + self.config: PreTrainedConfig = config

442

443 with self._mark_language_model(vllm_config):

444 self.llm = init_vllm_registered_model(

vllm/model_executor/models/paddleocr_vl.py

+7 -7

@@ -23,7 +23,7 @@ import numpy as np

23 import torch

24 import torch.nn as nn

25 from einops import rearrange

26 -from transformers import BaseImageProcessor, BatchFeature, PretrainedConfig

26 +from transformers import BaseImageProcessor, BatchFeature, PreTrainedConfig

27 from transformers.activations import GELUActivation

28 from transformers.image_utils import ChannelDimension

29 from transformers.modeling_outputs import (

@@ -318,8 +318,8 @@ class PaddleOCRVLMultiModalProcessor(

318 class Projector(nn.Module):

319 def __init__(

320 self,

321 - text_config: PretrainedConfig,

322 - vision_config: PretrainedConfig,

321 + text_config: PreTrainedConfig,

322 + vision_config: PreTrainedConfig,

323 prefix: str = "",

324 ):

325 super().__init__()

@@ -392,7 +392,7 @@ class PaddleOCRImagePixelInputs(TensorSchema):

392

393

394 class SiglipVisionEmbeddings(nn.Module):

395 - def __init__(self, config: PretrainedConfig):

395 + def __init__(self, config: PreTrainedConfig):

396 super().__init__()

397 self.config = config

398 self.embed_dim = config.hidden_size

@@ -682,7 +682,7 @@ class SigLIPRotaryEmbedding(nn.Module):

682 class SiglipEncoderLayer(nn.Module):

683 def __init__(

684 self,

685 - config: PretrainedConfig,

685 + config: PreTrainedConfig,

686 quant_config: QuantizationConfig | None = None,

687 prefix: str = "",

688 ):

@@ -735,7 +735,7 @@ class SiglipEncoderLayer(nn.Module):

735 class SiglipEncoder(nn.Module):

736 def __init__(

737 self,

738 - config: PretrainedConfig,

738 + config: PreTrainedConfig,

739 quant_config: QuantizationConfig | None = None,

740 prefix: str = "",

741 ):

@@ -837,7 +837,7 @@ class SiglipEncoder(nn.Module):

837 class SiglipVisionTransformer(nn.Module):

838 def __init__(

839 self,

840 - config: PretrainedConfig,

840 + config: PreTrainedConfig,

841 quant_config: QuantizationConfig | None = None,

842 prefix: str = "",

843 ):

vllm/model_executor/models/parakeet.py

+4 -4

@@ -10,7 +10,7 @@ import numpy as np

10 import torch

11 import torch.nn as nn

12 from transformers import ParakeetEncoder as HFParakeetEncoder

13 -from transformers import PretrainedConfig

13 +from transformers import PreTrainedConfig

14 from transformers.audio_utils import mel_filter_bank

15

16 from vllm.logger import init_logger

@@ -47,7 +47,7 @@ class ParakeetProjection(nn.Module):

47 class ProjectedParakeet(nn.Module):

48 def __init__(

49 self,

50 - config: PretrainedConfig,

50 + config: PreTrainedConfig,

51 *,

52 dtype: torch.dtype,

53 llm_hidden_size: int,

@@ -135,7 +135,7 @@ LOG_ZERO_GUARD_VALUE = 2**-24

135

136

137 class ParakeetExtractor:

138 - def __init__(self, config: PretrainedConfig) -> None:

138 + def __init__(self, config: PreTrainedConfig) -> None:

139 self.config = ExtractorConfig.from_hf_config(config)

140 """`config` is named *exactly* for `._get_subsampling_output_length` below"""

141 self._clip_target_samples = int(

@@ -329,6 +329,6 @@ class ParakeetExtractor:

329 }

330

331 @staticmethod

332 - def audio_length(raw_config: PretrainedConfig, audio_tokens: int) -> int:

332 + def audio_length(raw_config: PreTrainedConfig, audio_tokens: int) -> int:

333 config = ExtractorConfig.from_hf_config(raw_config)

334 return int(audio_tokens * config.subsampling_factor * config.hop_length)

vllm/model_executor/models/phi3v.py

+3 -3

@@ -24,7 +24,7 @@ import torch.nn as nn

24 from transformers import (

25 BatchFeature,

26 CLIPVisionConfig,

27 - PretrainedConfig,

27 + PreTrainedConfig,

28 ProcessorMixin,

29 )

30

@@ -95,7 +95,7 @@ CLIP_VIT_LARGE_PATCH14_336_CONFIG = CLIPVisionConfig(

95

96

97 def _init_img_processor(

98 - hf_config: PretrainedConfig,

98 + hf_config: PreTrainedConfig,

99 quant_config: QuantizationConfig | None,

100 prefix: str = "",

101 ) -> CLIPVisionModel:

@@ -165,7 +165,7 @@ class Phi3HDImageEmbedding(nn.Module):

165

166 def __init__(

167 self,

168 - config: PretrainedConfig,

168 + config: PreTrainedConfig,

169 quant_config: QuantizationConfig | None,

170 prefix: str = "",

171 ) -> None:

vllm/model_executor/models/phi4mm.py

+2 -2

@@ -10,7 +10,7 @@ import torch

10 import torch.nn as nn

11 from transformers import (

12 BatchFeature,

13 - PretrainedConfig,

13 + PreTrainedConfig,

14 ProcessorMixin,

15 SequenceFeatureExtractor,

16 SiglipVisionConfig,

@@ -123,7 +123,7 @@ class Phi4MMImageEncoder(nn.Module):

123

124 def __init__(

125 self,

126 - config: PretrainedConfig,

126 + config: PreTrainedConfig,

127 quant_config: QuantizationConfig | None,

128 prefix: str = "",

129 model_dir: str = "",

vllm/model_executor/models/phi4mm_audio.py

+2 -2

@@ -18,7 +18,7 @@ from torch.distributed.algorithms._checkpoint.checkpoint_wrapper import (

18 CheckpointWrapper,

19 )

20 from torch.distributed.fsdp.fully_sharded_data_parallel import FullyShardedDataParallel

21 -from transformers import PretrainedConfig

21 +from transformers import PreTrainedConfig

22

23 from vllm.model_executor.models.phi4mm_utils import (

24 AbsolutePositionalEncoding,

@@ -1152,7 +1152,7 @@ class WindowQformer(nn.Module):

1152 class AudioEmbedding(nn.Module):

1153 """Image embedding."""

1154

1155 - def __init__(self, config: PretrainedConfig, **kwargs: Any) -> None:

1155 + def __init__(self, config: PreTrainedConfig, **kwargs: Any) -> None:

1156 super().__init__()

1157 self.config = config

1158 # n_embed or hidden_size for text LM

vllm/model_executor/models/phi4siglip.py

+2 -2

@@ -11,7 +11,7 @@ from typing import Annotated, Any, Literal

11

12 import torch

13 import torch.nn as nn

14 -from transformers import BatchFeature, PretrainedConfig, Siglip2VisionConfig

14 +from transformers import BatchFeature, PreTrainedConfig, Siglip2VisionConfig

15

16 from vllm.config import VllmConfig

17 from vllm.config.multimodal import MultiModalDummyOptions

@@ -235,7 +235,7 @@ class Phi4ForCausalLMV(nn.Module, SupportsMultiModal, SupportsPP):

235 def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:

236 super().__init__()

237

238 - config: PretrainedConfig = vllm_config.model_config.hf_config

238 + config: PreTrainedConfig = vllm_config.model_config.hf_config

239 quant_config = vllm_config.quant_config

240 self.config = config

241

vllm/model_executor/models/phimoe.py

+2 -2

@@ -29,7 +29,7 @@ from itertools import islice

29

30 import torch

31 from torch import nn

32 -from transformers.configuration_utils import PretrainedConfig

32 +from transformers.configuration_utils import PreTrainedConfig

33

34 from vllm.compilation.decorators import support_torch_compile

35 from vllm.config import CacheConfig, VllmConfig

@@ -62,7 +62,7 @@ from .utils import (

62 )

63

64

65 -class PhiMoEConfig(PretrainedConfig):

65 +class PhiMoEConfig(PreTrainedConfig):

66 model_type = "phimoe"

67 keys_to_ignore_at_inference = ["past_key_values"]

68

vllm/model_executor/models/plamo3.py

+2 -2

@@ -8,7 +8,7 @@ from typing import TYPE_CHECKING, Any

8

9 import torch

10 from torch import nn

11 -from transformers import PretrainedConfig

11 +from transformers import PreTrainedConfig

12

13 from vllm.compilation.decorators import support_torch_compile

14 from vllm.config import VllmConfig

@@ -54,7 +54,7 @@ from vllm.sequence import IntermediateTensors

54 # Only used for type hinting.

55 if TYPE_CHECKING:

56

57 - class Plamo3Config(PretrainedConfig): # type: ignore

57 + class Plamo3Config(PreTrainedConfig): # type: ignore

58 model_type: str = "plamo3"

59

60 hidden_size: int

vllm/model_executor/models/qianfan_ocr.py

+2 -2

@@ -5,7 +5,7 @@

5 # The model architecture and weights are fully compatible with InternVLChatModel,

6 # only the config model_type / architectures strings differ.

7

8 -from transformers import PretrainedConfig

8 +from transformers import PreTrainedConfig

9

10 from vllm.model_executor.layers.quantization import QuantizationConfig

11 from vllm.model_executor.layers.quantization.fp8 import Fp8Config

@@ -67,7 +67,7 @@ class QianfanOCRForConditionalGeneration(InternVLChatModel):

67 """

68

69 def _patch_quant_config(

70 - self, config: PretrainedConfig, quant_config: QuantizationConfig

70 + self, config: PreTrainedConfig, quant_config: QuantizationConfig

71 ) -> None:

72 super()._patch_quant_config(config, quant_config)

73 # ignore vit layers to preserve model performance

vllm/model_executor/models/qwen2_5_omni_thinker.py

+2 -3

@@ -29,7 +29,7 @@ from typing import Annotated, Any, Literal

29 import numpy as np

30 import torch

31 import torch.nn as nn

32 -from transformers import PretrainedConfig

32 +from transformers import PreTrainedConfig

33 from transformers.feature_extraction_utils import BatchFeature

34 from transformers.models.qwen2_5_omni.configuration_qwen2_5_omni import (

35 Qwen2_5OmniConfig,

@@ -375,7 +375,6 @@ class Qwen2_5OmniThinkerProcessingInfo(

375 def get_hf_processor(self, **kwargs: object) -> Qwen2_5OmniProcessor:

376 return self.ctx.get_hf_processor(

377 Qwen2_5OmniProcessor,

378 - use_fast=kwargs.pop("use_fast", True),

378 **kwargs,

379 )

380

@@ -693,7 +692,7 @@ class Qwen2_5OmniThinkerMultiModalProcessor(

692 @classmethod

693 def omni_get_updates_use_audio_in_video(

694 cls,

696 - thinker_config: PretrainedConfig,

695 + thinker_config: PreTrainedConfig,

696 audio_len: int,

697 video_grid_thw: list[int] | torch.Tensor,

698 video_second_per_grid_t: float,

vllm/model_executor/models/qwen2_5_vl.py

-1

@@ -1183,7 +1183,6 @@ class Qwen2_5_VLProcessingInfo(Qwen2VLProcessingInfo):

1183 def get_hf_processor(self, **kwargs: object) -> Qwen2_5_VLProcessor:

1184 return self.ctx.get_hf_processor(

1185 Qwen2_5_VLProcessor,

1186 - use_fast=kwargs.pop("use_fast", True),

1186 **kwargs,

1187 )

1188

vllm/model_executor/models/qwen2_vl.py

-1

@@ -848,7 +848,6 @@ class Qwen2VLProcessingInfo(BaseProcessingInfo):

848 def get_hf_processor(self, **kwargs: object) -> Qwen2VLProcessor:

849 return self.ctx.get_hf_processor(

850 Qwen2VLProcessor,

851 - use_fast=kwargs.pop("use_fast", True),

851 **kwargs,

852 )

853

vllm/model_executor/models/qwen3_asr.py

-1

@@ -189,7 +189,6 @@ class Qwen3ASRProcessingInfo(BaseProcessingInfo):

189 def get_hf_processor(self, **kwargs: object) -> Qwen3ASRProcessor:

190 processor = self.ctx.get_hf_processor(

191 Qwen3ASRProcessor,

192 - use_fast=kwargs.pop("use_fast", True),

192 **kwargs,

193 )

194 if not hasattr(processor, "audio_token"):

vllm/model_executor/models/qwen3_omni_moe_thinker.py

+2 -3

@@ -30,7 +30,7 @@ import numpy as np

30 import torch

31 import torch.nn as nn

32 import torch.nn.functional as F

33 -from transformers import PretrainedConfig

33 +from transformers import PreTrainedConfig

34 from transformers.feature_extraction_utils import BatchFeature

35 from transformers.models.qwen3_omni_moe.configuration_qwen3_omni_moe import (

36 Qwen3OmniMoeAudioEncoderConfig,

@@ -1169,7 +1169,6 @@ class Qwen3OmniMoeThinkerProcessingInfo(

1169 def get_hf_processor(self, **kwargs: object) -> Qwen3OmniMoeProcessor:

1170 processor = self.ctx.get_hf_processor(

1171 Qwen3OmniMoeProcessor,

1172 - use_fast=kwargs.pop("use_fast", True),

1172 **kwargs,

1173 )

1174 if not hasattr(processor, "audio_token"):

@@ -1368,7 +1367,7 @@ class Qwen3OmniMoeThinkerMultiModalProcessor(

1367

1368 def get_updates_use_audio_in_video(

1369 self,

1371 - thinker_config: PretrainedConfig,

1370 + thinker_config: PreTrainedConfig,

1371 audio_len: int,

1372 video_grid_thw: list[int] | torch.Tensor,

1373 video_second_per_grid_t: float,

vllm/model_executor/models/qwen3_vl.py

-1

@@ -915,7 +915,6 @@ class Qwen3VLProcessingInfo(Qwen2VLProcessingInfo):

915 def get_hf_processor(self, **kwargs: object) -> Qwen3VLProcessor:

916 return self.ctx.get_hf_processor(

917 Qwen3VLProcessor,

918 - use_fast=kwargs.pop("use_fast", True),

918 **kwargs,

919 )

920

vllm/model_executor/models/radio.py

+4 -4

@@ -18,7 +18,7 @@ import torch

18 import torch.nn as nn

19 import torch.nn.functional as F

20 from einops import rearrange

21 -from transformers import PretrainedConfig

21 +from transformers import PreTrainedConfig

22

23 from vllm.model_executor.layers.quantization import QuantizationConfig

24 from vllm.model_executor.model_loader.weight_utils import default_weight_loader

@@ -109,7 +109,7 @@ class ClsToken(nn.Module):

109 class ViTPatchGenerator(nn.Module):

110 def __init__(

111 self,

112 - # config: PretrainedConfig,

112 + # config: PreTrainedConfig,

113 patch_size: int,

114 embed_dim: int,

115 input_dims: input_dim_t,

@@ -535,7 +535,7 @@ class RadioInternVisionModel(nn.Module):

535

536 def __init__(

537 self,

538 - config: PretrainedConfig = None,

538 + config: PreTrainedConfig = None,

539 quant_config: QuantizationConfig | None = None,

540 *,

541 num_hidden_layers_override: int | None = None,

@@ -670,7 +670,7 @@ class RadioModel(nn.Module):

670

671 def __init__(

672 self,

673 - config: PretrainedConfig,

673 + config: PreTrainedConfig,

674 quant_config: QuantizationConfig | None = None,

675 *,

676 num_hidden_layers_override: int | None = None,

vllm/model_executor/models/seed_oss.py

+1 -1

@@ -28,7 +28,7 @@ from itertools import islice

28

29 import torch

30 from torch import nn

31 -from transformers import PretrainedConfig as SeedOssConfig

31 +from transformers import PreTrainedConfig as SeedOssConfig

32

33 from vllm.compilation.decorators import support_torch_compile

34 from vllm.config import CacheConfig, VllmConfig

vllm/model_executor/models/siglip2navit.py

+3 -3

@@ -9,7 +9,7 @@ import torch

9 from torch import nn

10 from torch.nn import functional as F

11 from transformers import Siglip2VisionConfig

12 -from transformers.configuration_utils import PretrainedConfig

12 +from transformers.configuration_utils import PreTrainedConfig

13

14 from vllm.distributed import divide, get_tensor_model_parallel_world_size

15 from vllm.model_executor.layers.activation import get_act_fn

@@ -47,7 +47,7 @@ class VisionRotaryEmbedding(nn.Module):

47

48

49 class Siglip2VisionEmbeddings(nn.Module):

50 - def __init__(self, config: PretrainedConfig):

50 + def __init__(self, config: PreTrainedConfig):

51 super().__init__()

52 self.config = config

53 self.embed_dim = config.hidden_size

@@ -364,7 +364,7 @@ class Siglip2Encoder(nn.Module):

364 self attention layers. Each layer is a [`Siglip2EncoderLayer`].

365

366 Args:

367 - config: PretrainedConfig

367 + config: PreTrainedConfig

368

369 """

370

vllm/model_executor/models/skyworkr1v.py

+4 -4

@@ -12,7 +12,7 @@ from typing import Annotated, Literal, TypeAlias

12

13 import torch

14 import torch.nn as nn

15 -from transformers import PretrainedConfig

15 +from transformers import PreTrainedConfig

16

17 from vllm.config import VllmConfig

18 from vllm.model_executor.layers.linear import ReplicatedLinear

@@ -200,7 +200,7 @@ class SkyworkR1VChatModel(nn.Module, SupportsMultiModal, SupportsPP):

200 )

201

202 def _patch_quant_config(

203 - self, config: PretrainedConfig, quant_config: QuantizationConfig | None

203 + self, config: PreTrainedConfig, quant_config: QuantizationConfig | None

204 ):

205 # the awq models from OpenGVLab missing `modules_to_not_convert`

206 # patch the quant_config to add `modules_to_not_convert` back

@@ -214,7 +214,7 @@ class SkyworkR1VChatModel(nn.Module, SupportsMultiModal, SupportsPP):

214

215 def _init_vision_model(

216 self,

217 - config: PretrainedConfig,

217 + config: PreTrainedConfig,

218 quant_config: QuantizationConfig | None,

219 *,

220 prefix: str,

@@ -236,7 +236,7 @@ class SkyworkR1VChatModel(nn.Module, SupportsMultiModal, SupportsPP):

236

237 def _init_mlp1(

238 self,

239 - config: PretrainedConfig,

239 + config: PreTrainedConfig,

240 quant_config: QuantizationConfig | None,

241 prefix: str = "",

242 ) -> nn.Module:

vllm/model_executor/models/solar.py

+3 -3

@@ -28,7 +28,7 @@ from collections.abc import Iterable

28

29 import torch

30 from torch import nn

31 -from transformers import PretrainedConfig

31 +from transformers import PreTrainedConfig

32

33 from vllm.compilation.decorators import support_torch_compile

34 from vllm.config import CacheConfig, VllmConfig

@@ -102,7 +102,7 @@ class SolarMLP(nn.Module):

102 class SolarAttention(nn.Module):

103 def __init__(

104 self,

105 - config: PretrainedConfig,

105 + config: PreTrainedConfig,

106 hidden_size: int,

107 num_heads: int,

108 num_kv_heads: int,

@@ -185,7 +185,7 @@ class SolarAttention(nn.Module):

185 class SolarDecoderLayer(nn.Module):

186 def __init__(

187 self,

188 - config: PretrainedConfig,

188 + config: PreTrainedConfig,

189 cache_config: CacheConfig | None = None,

190 quant_config: QuantizationConfig | None = None,

191 prefix: str = "",

vllm/model_executor/models/step3p5.py

+1 -1

@@ -49,7 +49,7 @@ from vllm.v1.attention.backend import AttentionType

49 if typing.TYPE_CHECKING:

50 from vllm.transformers_utils.configs.step3p5 import Step3p5Config

51 else:

52 - from transformers import PretrainedConfig as Step3p5Config

52 + from transformers import PreTrainedConfig as Step3p5Config

53

54 from .interfaces import MixtureOfExperts, SupportsPP

55 from .utils import (

vllm/model_executor/models/step3p5_mtp.py

+2 -2

@@ -4,7 +4,7 @@ from collections.abc import Iterable

4

5 import torch

6 import torch.nn as nn

7 -from transformers import PretrainedConfig

7 +from transformers import PreTrainedConfig

8

9 from vllm.config import VllmConfig

10 from vllm.logger import init_logger

@@ -30,7 +30,7 @@ logger = init_logger(__name__)

30 class SharedHead(nn.Module):

31 def __init__(

32 self,

33 - config: PretrainedConfig,

33 + config: PreTrainedConfig,

34 quant_config: QuantizationConfig | None = None,

35 prefix: str = "",

36 ) -> None:

vllm/model_executor/models/transformers/base.py

+3 -3

@@ -29,7 +29,7 @@ import torch

29 import transformers

30 from packaging.version import Version

31 from torch import nn

32 -from transformers import AutoModel, PretrainedConfig

32 +from transformers import AutoModel, PreTrainedConfig

33 from transformers.conversion_mapping import (

34 WeightRenaming,

35 get_model_conversion_mapping,

@@ -433,9 +433,9 @@ class Base(

433 def _vocab_embeddings(self) -> set[nn.Embedding]:

434 """The `nn.Embedding`s in `self.model` which hold a vocab table."""

435

436 - def vocab_sizes(config: PretrainedConfig):

436 + def vocab_sizes(config: PreTrainedConfig):

437 for key, value in vars(config).items():

438 - if isinstance(value, PretrainedConfig):

438 + if isinstance(value, PreTrainedConfig):

439 yield from vocab_sizes(value)

440 elif "vocab_size" in key and isinstance(value, int):

441 yield value

vllm/model_executor/models/utils.py

+2 -2

@@ -32,7 +32,7 @@ from vllm.utils.torch_utils import (

32 )

33

34 if TYPE_CHECKING:

35 - from transformers import PretrainedConfig

35 + from transformers import PreTrainedConfig

36 from transformers.conversion_mapping import WeightRenaming

37

38 from vllm.config.model import ModelConfig

@@ -610,7 +610,7 @@ def init_vllm_registered_model(

610 vllm_config: VllmConfig,

611 *,

612 prefix: str = "",

613 - hf_config: "PretrainedConfig | None" = None,

613 + hf_config: "PreTrainedConfig | None" = None,

614 architectures: list[str] | None = None,

615 ) -> nn.Module:

616 """Helper function to initialize an inner model registered to vLLM,

vllm/model_executor/models/vision.py

+3 -3

@@ -9,7 +9,7 @@ from typing import Final, Generic, Literal, Protocol, TypeAlias, TypeVar

9

10 import torch

11 import torch.nn as nn

12 -from transformers import PretrainedConfig

12 +from transformers import PreTrainedConfig

13

14 from vllm.config import ModelConfig, MultiModalConfig, get_current_vllm_config_or_none

15 from vllm.distributed import (

@@ -25,7 +25,7 @@ from vllm.v1.attention.backends.registry import AttentionBackendEnum

25

26 logger = init_logger(__name__)

27

28 -_C = TypeVar("_C", bound=PretrainedConfig)

28 +_C = TypeVar("_C", bound=PreTrainedConfig)

29

30

31 class _RootConfig(Protocol[_C]):

@@ -62,7 +62,7 @@ class VisionEncoderInfo(ABC, Generic[_C]):

62

63

64 class VisionLanguageConfig(Protocol):

65 - vision_config: Final[PretrainedConfig]

65 + vision_config: Final[PreTrainedConfig]

66

67

68 def get_vision_encoder_info(hf_config: VisionLanguageConfig) -> VisionEncoderInfo:

vllm/models/dots3_note/nvidia/vision.py

+2 -2

@@ -9,7 +9,7 @@ import torch

9 import torch.nn.functional as F

10 from torch import nn

11 from torch.nn import LayerNorm

12 -from transformers.configuration_utils import PretrainedConfig

12 +from transformers.configuration_utils import PreTrainedConfig

13 from transformers.modeling_utils import PreTrainedModel

14

15 from vllm.third_party.deep_gemm import per_block_cast_to_fp8

@@ -27,7 +27,7 @@ from .vision_attention import (

27 from .vision_moe import note_vision_fused_moe_fp8

28

29

30 -class DotsMoEVitConfig(PretrainedConfig):

30 +class DotsMoEVitConfig(PreTrainedConfig):

31 model_type: str = "dots_moe_vit"

32

33 def __init__(

vllm/models/hy_v4/amd/model.py

+3 -3

@@ -30,7 +30,7 @@ from itertools import islice

30 import regex as re

31 import torch

32 from torch import nn

33 -from transformers import PretrainedConfig

33 +from transformers import PreTrainedConfig

34

35 from vllm.compilation.decorators import support_torch_compile

36 from vllm.config import CacheConfig, VllmConfig

@@ -75,7 +75,7 @@ from ..nvidia.moe import HYV4FeedForward, HYV4MoEFused

75 logger = init_logger(__name__)

76

77

78 -def _normalize_hyv4_config(config: PretrainedConfig) -> PretrainedConfig:

78 +def _normalize_hyv4_config(config: PreTrainedConfig) -> PreTrainedConfig:

79 """Populate the aliases consumed by the shared MoE implementation."""

80 config.router_scaling_factor = config.routed_scaling_factor

81 config.num_experts = config.n_routed_experts

@@ -95,7 +95,7 @@ class HYV4DecoderLayer(nn.Module):

95

96 def __init__(

97 self,

98 - config: PretrainedConfig,

98 + config: PreTrainedConfig,

99 vllm_config: VllmConfig,

100 cache_config: CacheConfig | None = None,

101 quant_config: QuantizationConfig | None = None,

vllm/models/hy_v4/nvidia/attention.py

+3 -3

@@ -18,7 +18,7 @@ from typing import cast

18 import regex as re

19 import torch

20 from torch import nn

21 -from transformers import DeepseekV2Config, DeepseekV3Config, PretrainedConfig

21 +from transformers import DeepseekV2Config, DeepseekV3Config, PreTrainedConfig

22

23 from vllm.compilation.breakable_cudagraph import eager_break_during_capture

24 from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config

@@ -51,7 +51,7 @@ _SPARSE_LAYER_TYPES = ("sparse_attention", "sparse", "deepseek_sparse_attention"

51 _WEIGHT_LAYER_INDEX_RE = re.compile(r"(?:^|\.)layers\.(\d+)(?:\.|$)")

52

53

54 -def compute_skip_topk_layers(config: PretrainedConfig) -> set[int]:

54 +def compute_skip_topk_layers(config: PreTrainedConfig) -> set[int]:

55 """Return the backbone layers that reuse a previous layer's top-k indices.

56

57 A "shared" indexer layer performs sparse attention with the indices computed

@@ -280,7 +280,7 @@ class HYV4MLAAttention(nn.Module):

280 def __init__(

281 self,

282 vllm_config: VllmConfig,

283 - config: PretrainedConfig,

283 + config: PreTrainedConfig,

284 hidden_size: int,

285 num_heads: int,

286 qk_nope_head_dim: int,

vllm/models/hy_v4/nvidia/hc.py

+5 -5

@@ -17,7 +17,7 @@ requires restructuring the decoder-layer forward scheduling.

17

18 import torch

19 from torch import nn

20 -from transformers import PretrainedConfig

20 +from transformers import PreTrainedConfig

21

22 from vllm.model_executor.layers.hpc import HpcIHCHead, HpcIHCPost, HpcIHCPre

23 from vllm.model_executor.layers.linear import ReplicatedLinear

@@ -40,7 +40,7 @@ class HYV4HCPreLayer(nn.Module):

40

41 def __init__(

42 self,

43 - config: PretrainedConfig,

43 + config: PreTrainedConfig,

44 hidden_dim: int,

45 hc_mult: int = 4,

46 magnitude: float = 2.0,

@@ -163,7 +163,7 @@ class HYV4HCPostLayer(nn.Module):

163 y[n, i, d] = post[n, i] * x[n, d] + residual[n, i, d]

164 """

165

166 - def __init__(self, config: PretrainedConfig):

166 + def __init__(self, config: PreTrainedConfig):

167 super().__init__()

168 self.config = config

169

@@ -213,7 +213,7 @@ class HYV4HCHeadLayer(nn.Module):

213

214 def __init__(

215 self,

216 - config: PretrainedConfig,

216 + config: PreTrainedConfig,

217 hidden_size: int,

218 hc_mult: int = 4,

219 hc_eps: float = 1e-6,

@@ -301,7 +301,7 @@ class HYV4HCLayer(nn.Module):

301

302 def __init__(

303 self,

304 - config: PretrainedConfig,

304 + config: PreTrainedConfig,

305 layer_idx: int,

306 init_std: float = 6e-3,

307 base_noise_std: float = 0.0,

vllm/models/hy_v4/nvidia/model.py

+3 -3

@@ -30,7 +30,7 @@ from itertools import islice

30 import regex as re

31 import torch

32 from torch import nn

33 -from transformers import PretrainedConfig

33 +from transformers import PreTrainedConfig

34

35 from vllm.config import CacheConfig, VllmConfig

36 from vllm.distributed import (

@@ -74,7 +74,7 @@ from .moe import HYV4FeedForward, HYV4MoEFused

74 logger = init_logger(__name__)

75

76

77 -def _normalize_hyv4_config(config: PretrainedConfig) -> PretrainedConfig:

77 +def _normalize_hyv4_config(config: PreTrainedConfig) -> PreTrainedConfig:

78 """Populate the aliases consumed by the shared MoE implementation."""

79 config.router_scaling_factor = config.routed_scaling_factor

80 config.num_experts = config.n_routed_experts

@@ -94,7 +94,7 @@ class HYV4DecoderLayer(nn.Module):

94

95 def __init__(

96 self,

97 - config: PretrainedConfig,

97 + config: PreTrainedConfig,

98 vllm_config: VllmConfig,

99 cache_config: CacheConfig | None = None,

100 quant_config: QuantizationConfig | None = None,

vllm/models/hy_v4/nvidia/moe.py

+2 -2

@@ -4,7 +4,7 @@

4

5 import torch

6 from torch import nn

7 -from transformers import PretrainedConfig

7 +from transformers import PreTrainedConfig

8

9 from vllm.config import get_current_vllm_config

10 from vllm.distributed import get_ep_group, get_tensor_model_parallel_world_size

@@ -87,7 +87,7 @@ class HYV4MoEFused(nn.Module):

87

88 def __init__(

89 self,

90 - config: PretrainedConfig,

90 + config: PreTrainedConfig,

91 quant_config: QuantizationConfig | None = None,

92 prefix: str = "",

93 enable_eplb: bool = False,

vllm/models/hy_v4/nvidia/mtp.py

+5 -5

@@ -8,7 +8,7 @@ from collections.abc import Callable, Iterable

8

9 import torch

10 from torch import nn

11 -from transformers import PretrainedConfig

11 +from transformers import PreTrainedConfig

12

13 from vllm.config import CacheConfig, ModelConfig, VllmConfig

14 from vllm.distributed import (

@@ -59,7 +59,7 @@ _MTP_QUANT_EXCLUSION_ATTRS = ("ignored_layers", "exclude_modules")

59

60

61 def _get_spec_layer_idx_from_weight_name(

62 - config: PretrainedConfig, weight_name: str

62 + config: PreTrainedConfig, weight_name: str

63 ) -> int | None:

64 """Return the MTP layer index a checkpoint weight belongs to, or None.

65

@@ -147,7 +147,7 @@ def _prepare_mtp_fp8_expert_scale(

147

148

149 def _create_mtp_quant_config(

150 - hf_config: PretrainedConfig,

150 + hf_config: PreTrainedConfig,

151 backbone_quant_config: QuantizationConfig | None = None,

152 ) -> QuantizationConfig | None:

153 """Create the quantization config for the MTP layers.

@@ -303,7 +303,7 @@ class HYV4SharedHead(nn.Module):

303

304 def __init__(

305 self,

306 - config: PretrainedConfig,

306 + config: PreTrainedConfig,

307 quant_config: QuantizationConfig | None = None,

308 ) -> None:

309 super().__init__()

@@ -329,7 +329,7 @@ class HYV4MultiTokenPredictorLayer(nn.Module):

329

330 def __init__(

331 self,

332 - config: PretrainedConfig,

332 + config: PreTrainedConfig,

333 prefix: str,

334 vllm_config: VllmConfig,

335 model_config: ModelConfig,

vllm/models/inkling/configs.py

+6 -6

@@ -8,10 +8,10 @@ from dataclasses import dataclass, field

8 from typing import Any, ClassVar, Literal, cast

9

10 import torch

11 -from transformers.configuration_utils import PretrainedConfig

11 +from transformers.configuration_utils import PreTrainedConfig

12

13

14 -class InklingModelConfig(PretrainedConfig):

14 +class InklingModelConfig(PreTrainedConfig):

15 model_type = "inkling_model"

16 keys_to_ignore_at_inference = ["past_key_values"]

17

@@ -215,7 +215,7 @@ class InklingModelConfig(PretrainedConfig):

215 return TMLConvCacheParams(shape=shape, layers=self.conv_layer_ids, dtype=dtype)

216

217

218 -class InklingAudioConfig(PretrainedConfig):

218 +class InklingAudioConfig(PreTrainedConfig):

219 model_type = "inkling_audio_model"

220

221 def __init__(

@@ -255,7 +255,7 @@ class InklingAudioConfig(PretrainedConfig):

255 super().__init__(**kwargs)

256

257

258 -class InklingVisionConfig(PretrainedConfig):

258 +class InklingVisionConfig(PreTrainedConfig):

259 model_type = "inkling_vision_model"

260

261 def __init__(

@@ -295,10 +295,10 @@ class InklingVisionConfig(PretrainedConfig):

295 super().__init__(**kwargs)

296

297

298 -class InklingMMConfig(PretrainedConfig):

298 +class InklingMMConfig(PreTrainedConfig):

299 model_type = "inkling_mm_model"

300 keys_to_ignore_at_inference = ["past_key_values"]

301 - sub_configs: ClassVar[dict[str, type[PretrainedConfig]]] = {

301 + sub_configs: ClassVar[dict[str, type[PreTrainedConfig]]] = {

302 "text_config": InklingModelConfig,

303 "audio_config": InklingAudioConfig,

304 "vision_config": InklingVisionConfig,

vllm/models/minimax_m3/amd/model.py

+12 -12

@@ -21,7 +21,7 @@ from collections.abc import Iterable

21

22 import torch

23 from torch import nn

24 -from transformers import PretrainedConfig

24 +from transformers import PreTrainedConfig

25

26 from vllm import _custom_ops as ops

27 from vllm import envs

@@ -134,7 +134,7 @@ from vllm.v1.kv_cache_interface import (

134 logger = init_logger(__name__)

135

136

137 -def _sparse_attention_layer_ids(config: PretrainedConfig) -> set[int]:

137 +def _sparse_attention_layer_ids(config: PreTrainedConfig) -> set[int]:

138 """Layer ids whose attention runs the extra sparse "index" branch."""

139 cfg = getattr(config, "sparse_attention_config", None)

140 if not cfg:

@@ -145,7 +145,7 @@ def _sparse_attention_layer_ids(config: PretrainedConfig) -> set[int]:

145 return {i for i, f in enumerate(freq) if f != 0}

146

147

148 -def _sparse_attention_layer_ordinals(config: PretrainedConfig) -> dict[int, int]:

148 +def _sparse_attention_layer_ordinals(config: PreTrainedConfig) -> dict[int, int]:

149 """Map each sparse-attention layer id to its ordinal among sparse layers."""

150 return {

151 lid: ordinal

@@ -153,7 +153,7 @@ def _sparse_attention_layer_ordinals(config: PretrainedConfig) -> dict[int, int]

153 }

154

155

156 -def _should_skip_index_topk(config: PretrainedConfig, layer_id: int) -> bool:

156 +def _should_skip_index_topk(config: PreTrainedConfig, layer_id: int) -> bool:

157 """ATOM ``index_topk_freq`` (cross-layer index sharing).

158

159 Only 1 of every ``index_topk_freq`` sparse-attention layers recomputes the

@@ -176,7 +176,7 @@ def _should_skip_index_topk(config: PretrainedConfig, layer_id: int) -> bool:

176 return max(ordinal - offset, 0) % freq != 0

177

178

179 -def _is_moe_layer(config: PretrainedConfig, layer_id: int) -> bool:

179 +def _is_moe_layer(config: PreTrainedConfig, layer_id: int) -> bool:

180 """Whether this layer's MLP is a sparse MoE block (vs a dense MLP)."""

181 moe_layer_freq = getattr(config, "moe_layer_freq", None)

182 if moe_layer_freq is None:

@@ -184,7 +184,7 @@ def _is_moe_layer(config: PretrainedConfig, layer_id: int) -> bool:

184 return moe_layer_freq[layer_id] != 0

185

186

187 -def _build_rotary_emb(config: PretrainedConfig, head_dim: int):

187 +def _build_rotary_emb(config: PreTrainedConfig, head_dim: int):

188 """Build the (partial NeoX) RoPE, honoring an optional ``rope_scaling`` config.

189

190 Without scaling the cos/sin cache is sized to ``max_position_embeddings``

@@ -262,7 +262,7 @@ class MiniMaxM3MLP(nn.Module):

262

263 def __init__(

264 self,

265 - config: PretrainedConfig,

265 + config: PreTrainedConfig,

266 intermediate_size: int,

267 quant_config: QuantizationConfig | None = None,

268 reduce_results: bool = True,

@@ -338,7 +338,7 @@ class MiniMaxM3MoE(nn.Module):

338

339 def __init__(

340 self,

341 - config: PretrainedConfig,

341 + config: PreTrainedConfig,

342 layer_id: int,

343 quant_config: QuantizationConfig | None = None,

344 prefix: str = "",

@@ -488,7 +488,7 @@ class MiniMaxM3Attention(nn.Module):

488

489 def __init__(

490 self,

491 - config: PretrainedConfig,

491 + config: PreTrainedConfig,

492 layer_id: int,

493 quant_config: QuantizationConfig | None = None,

494 prefix: str = "",

@@ -623,7 +623,7 @@ class MiniMaxM3SparseAttention(nn.Module, AttentionLayerBase):

623

624 def __init__(

625 self,

626 - config: PretrainedConfig,

626 + config: PreTrainedConfig,

627 layer_id: int,

628 quant_config: QuantizationConfig | None = None,

629 prefix: str = "",

@@ -1264,7 +1264,7 @@ class MiniMaxM3SparseAttention(nn.Module, AttentionLayerBase):

1264 class MiniMaxM3DecoderLayer(nn.Module):

1265 def __init__(

1266 self,

1267 - config: PretrainedConfig,

1267 + config: PreTrainedConfig,

1268 prefix: str,

1269 cache_config: CacheConfig | None = None,

1270 quant_config: QuantizationConfig | None = None,

@@ -1738,7 +1738,7 @@ class MiniMaxM3SparseForConditionalGeneration(

1738 with self._mark_tower_model(vllm_config, {"image", "video"}):

1739 vision_config = config.vision_config

1740 self.vision_tower = MiniMaxVLVisionModel(

1741 - config=PretrainedConfig.from_dict(vision_config),

1741 + config=PreTrainedConfig.from_dict(vision_config),

1742 text_hidden_size=text_hidden_size,

1743 projector_hidden_size=projector_hidden_size,

1744 quant_config=self.quant_config,

vllm/models/minimax_m3/common/vision_tower.py

+6 -6

@@ -6,7 +6,7 @@ import numpy as np

6 import torch

7 import torch.nn as nn

8 from einops import rearrange

9 -from transformers import PretrainedConfig

9 +from transformers import PreTrainedConfig

10

11 from vllm.distributed import parallel_state

12 from vllm.distributed import utils as dist_utils

@@ -44,7 +44,7 @@ class MiniMaxVLPatchEmbed(nn.Module):

44 and projects each to a hidden-size embedding.

45 """

46

47 - def __init__(self, config: PretrainedConfig) -> None:

47 + def __init__(self, config: PreTrainedConfig) -> None:

48 super().__init__()

49 compression = config.img_token_compression_config

50 temporal_patch_size = compression.get("temporal_patch_size", 2)

@@ -232,7 +232,7 @@ class MiniMaxVLEncoderLayer(nn.Module):

232

233 def __init__(

234 self,

235 - config: PretrainedConfig,

235 + config: PreTrainedConfig,

236 quant_config: QuantizationConfig | None = None,

237 prefix: str = "",

238 ) -> None:

@@ -295,7 +295,7 @@ class MiniMaxVLEncoderLayer(nn.Module):

295 class MiniMaxVLEncoder(nn.Module):

296 def __init__(

297 self,

298 - config: PretrainedConfig,

298 + config: PreTrainedConfig,

299 num_hidden_layers_override: int | None = None,

300 quant_config: QuantizationConfig | None = None,

301 prefix: str = "",

@@ -349,7 +349,7 @@ class MiniMaxVLVisionTransformer(nn.Module):

349

350 def __init__(

351 self,

352 - config: PretrainedConfig,

352 + config: PreTrainedConfig,

353 num_hidden_layers_override: int | None = None,

354 require_post_norm: bool | None = None,

355 quant_config: QuantizationConfig | None = None,

@@ -679,7 +679,7 @@ class MiniMaxVLVisionModel(nn.Module):

679

680 def __init__(

681 self,

682 - config: PretrainedConfig,

682 + config: PreTrainedConfig,

683 text_hidden_size: int,

684 projector_hidden_size: int | None = None,

685 quant_config: QuantizationConfig | None = None,

vllm/models/minimax_m3/nvidia/model.py

+8 -8

@@ -16,7 +16,7 @@ from collections.abc import Iterable

16

17 import torch

18 from torch import nn

19 -from transformers import PretrainedConfig

19 +from transformers import PreTrainedConfig

20

21 from vllm import _custom_ops as ops

22 from vllm.compilation.breakable_cudagraph import eager_break_during_capture

@@ -92,7 +92,7 @@ from vllm.v1.kv_cache_interface import (

92 )

93

94

95 -def _sparse_attention_layer_ids(config: PretrainedConfig) -> set[int]:

95 +def _sparse_attention_layer_ids(config: PreTrainedConfig) -> set[int]:

96 """Layer ids whose attention runs the extra sparse "index" branch."""

97 cfg = getattr(config, "sparse_attention_config", None)

98 if not cfg:

@@ -103,7 +103,7 @@ def _sparse_attention_layer_ids(config: PretrainedConfig) -> set[int]:

103 return {i for i, f in enumerate(freq) if f != 0}

104

105

106 -def _is_moe_layer(config: PretrainedConfig, layer_id: int) -> bool:

106 +def _is_moe_layer(config: PreTrainedConfig, layer_id: int) -> bool:

107 """Whether this layer's MLP is a sparse MoE block (vs a dense MLP)."""

108 moe_layer_freq = getattr(config, "moe_layer_freq", None)

109 if moe_layer_freq is None:

@@ -147,7 +147,7 @@ class MiniMaxM3MLP(nn.Module):

147

148 def __init__(

149 self,

150 - config: PretrainedConfig,

150 + config: PreTrainedConfig,

151 intermediate_size: int,

152 quant_config: QuantizationConfig | None = None,

153 reduce_results: bool = True,

@@ -194,7 +194,7 @@ class MiniMaxM3MoE(nn.Module):

194

195 def __init__(

196 self,

197 - config: PretrainedConfig,

197 + config: PreTrainedConfig,

198 layer_id: int,

199 quant_config: QuantizationConfig | None = None,

200 reduce_results: bool = True,

@@ -290,7 +290,7 @@ class MiniMaxM3Attention(nn.Module):

290

291 def __init__(

292 self,

293 - config: PretrainedConfig,

293 + config: PreTrainedConfig,

294 layer_id: int,

295 quant_config: QuantizationConfig | None = None,

296 prefix: str = "",

@@ -403,7 +403,7 @@ class MiniMaxM3SparseAttention(nn.Module, AttentionLayerBase):

403

404 def __init__(

405 self,

406 - config: PretrainedConfig,

406 + config: PreTrainedConfig,

407 layer_id: int,

408 quant_config: QuantizationConfig | None = None,

409 prefix: str = "",

@@ -1130,7 +1130,7 @@ class MiniMaxM3SparseForConditionalGeneration(

1130 with self._mark_tower_model(vllm_config, {"image", "video"}):

1131 vision_config = config.vision_config

1132 self.vision_tower = MiniMaxVLVisionModel(

1133 - config=PretrainedConfig.from_dict(vision_config),

1133 + config=PreTrainedConfig.from_dict(vision_config),

1134 text_hidden_size=text_hidden_size,

1135 projector_hidden_size=projector_hidden_size,

1136 quant_config=self.quant_config,

vllm/models/qwen4_exp/config.py

+3 -3

@@ -4,7 +4,7 @@

4

5 from typing import Any, ClassVar, cast

6

7 -from transformers import PretrainedConfig

7 +from transformers import PreTrainedConfig

8 from transformers.models.qwen3_vl.configuration_qwen3_vl import (

9 Qwen3VLVisionConfig,

10 )

@@ -201,9 +201,9 @@ class Qwen4ExpTextConfig(Qwen3NextConfig):

201 return max(int(self.ngram_size) - 1, 0)

202

203

204 -class Qwen4ExpConfig(PretrainedConfig):

204 +class Qwen4ExpConfig(PreTrainedConfig):

205 model_type = "qwen4_exp"

206 - sub_configs: ClassVar[dict[str, type[PretrainedConfig]]] = {

206 + sub_configs: ClassVar[dict[str, type[PreTrainedConfig]]] = {

207 "vision_config": Qwen4ExpVisionConfig,

208 "text_config": Qwen4ExpTextConfig,

209 }

vllm/multimodal/processing/context.py

+8 -8

@@ -26,14 +26,14 @@ from vllm.utils.jsontree import JSONTree, json_map_leaves

26 from vllm.utils.mistral import is_mistral_tokenizer

27

28 if TYPE_CHECKING:

29 - from transformers.configuration_utils import PretrainedConfig

29 + from transformers.configuration_utils import PreTrainedConfig

30 from transformers.feature_extraction_utils import BatchFeature

31 from transformers.processing_utils import ProcessorMixin

32

33 from vllm.config import ModelConfig

34 from vllm.renderers import TokenizeParams

35 else:

36 - PretrainedConfig = object

36 + PreTrainedConfig = object

37 BatchFeature = object

38 ProcessorMixin = object

39

@@ -123,7 +123,7 @@ class TimingContext:

123

124

125 _T = TypeVar("_T")

126 -_C = TypeVar("_C", bound=PretrainedConfig, default=PretrainedConfig)

126 +_C = TypeVar("_C", bound=PreTrainedConfig, default=PreTrainedConfig)

127 _P = TypeVar("_P", bound=ProcessorMixin, default=ProcessorMixin)

128

129

@@ -148,7 +148,7 @@ class InputProcessingContext:

148 return self.tokenizer

149

150 @overload

151 - def get_hf_config(self, /) -> PretrainedConfig: ...

151 + def get_hf_config(self, /) -> PreTrainedConfig: ...

152

153 @overload

154 def get_hf_config(

@@ -163,7 +163,7 @@ class InputProcessingContext:

163 /,

164 ) -> Any:

165 """Get the HuggingFace configuration

166 - (`transformers.PretrainedConfig`) of the model,

166 + (`transformers.PreTrainedConfig`) of the model,

167 additionally checking its type.

168

169 Raises:

@@ -171,9 +171,9 @@ class InputProcessingContext:

171

172 """

173 if typ is None:

174 - from transformers.configuration_utils import PretrainedConfig

174 + from transformers.configuration_utils import PreTrainedConfig

175

176 - typ = PretrainedConfig

176 + typ = PreTrainedConfig

177

178 hf_config = self.model_config.hf_config

179 if not isinstance(hf_config, typ):

@@ -371,7 +371,7 @@ class BaseProcessingInfo:

371 def get_tokenizer(self) -> TokenizerLike:

372 return self.ctx.get_tokenizer()

373

374 - def get_hf_config(self) -> PretrainedConfig:

374 + def get_hf_config(self) -> PreTrainedConfig:

375 return self.ctx.get_hf_config()

376

377 def get_hf_processor(self, **kwargs: object) -> ProcessorMixin:

vllm/transformers_utils/config.py

+33 -33

@@ -17,7 +17,7 @@ import transformers.configuration_utils as hf_configuration_utils

17 from huggingface_hub import constants

18 from packaging.version import Version

19 from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE

20 -from transformers import GenerationConfig, PretrainedConfig

20 +from transformers import GenerationConfig, PreTrainedConfig

21 from transformers.configuration_utils import ALLOWED_LAYER_TYPES

22 from transformers.models.auto.image_processing_auto import get_image_processor_config

23 from transformers.models.auto.modeling_auto import (

@@ -86,7 +86,7 @@ class LazyConfigDict(dict):

86 return getattr(configs, value)

87

88

89 -_CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(

89 +_CONFIG_REGISTRY: dict[str, type[PreTrainedConfig]] = LazyConfigDict(

90 afmoe="AfmoeConfig",

91 axk1="AXK1Config",

92 bagel="BagelConfig",

@@ -189,13 +189,13 @@ _AUTO_CONFIG_KWARGS_OVERRIDES: dict[str, dict[str, Any]] = {

189

190

191 def _register_config_class(

192 - model_type: str, config_class: type[PretrainedConfig]

192 + model_type: str, config_class: type[PreTrainedConfig]

193 ) -> None:

194 config_class.model_type = model_type

195 AutoConfig.register(model_type, config_class, exist_ok=True)

196

197

198 -def _maybe_register_hf_config(config: PretrainedConfig | None) -> None:

198 +def _maybe_register_hf_config(config: PreTrainedConfig | None) -> None:

199 if config is None:

200 return

201

@@ -228,7 +228,7 @@ def _mistral_patch_hf_hub_constants() -> Iterator[None]:

228 def _install_hf_config_validator(

229 name: str, validator: Callable, supersedes: Callable

230 ) -> None:

231 - """Replace a ``PretrainedConfig`` validator on every ``@strict`` snapshot.

231 + """Replace a ``PreTrainedConfig`` validator on every ``@strict`` snapshot.

232

233 ``@strict`` snapshots each ``validate_*`` method into ``__class_validators__``

234 at class creation and automatic post-``__init__`` validation dispatches off

@@ -237,10 +237,10 @@ def _install_hf_config_validator(

237 class owns a snapshot, so rewrite them all, matching ``supersedes`` by

238 identity to leave a genuine per-model override in place.

239 """

240 - setattr(PretrainedConfig, name, validator)

240 + setattr(PreTrainedConfig, name, validator)

241

242 seen: set[int] = set()

243 - stack = [PretrainedConfig]

243 + stack = [PreTrainedConfig]

244 while stack:

245 cls = stack.pop()

246 if id(cls) in seen:

@@ -261,10 +261,10 @@ def _patch_hf_transformers_validate_rope():

261 validate_rope() with the ignore_keys parameter work with newer versions of

262 hf transformers (from v5 onwards)

263 """

264 - if hasattr(PretrainedConfig.validate_rope, "__vllm_patched__"):

264 + if hasattr(PreTrainedConfig.validate_rope, "__vllm_patched__"):

265 return

266

267 - _original_validate_rope = PretrainedConfig.validate_rope

267 + _original_validate_rope = PreTrainedConfig.validate_rope

268

269 @wraps(_original_validate_rope)

270 def patched_validate_rope(self, *args, **kwargs):

@@ -293,10 +293,10 @@ def _patch_hf_transformers_nested_rope_validation() -> None:

293 ``AttributeError``. The per-layer dicts already carry their own defaults by

294 the time validation runs, so the shared entries can be dropped.

295 """

296 - if hasattr(PretrainedConfig.validate_rope, "__vllm_nested_rope_patched__"):

296 + if hasattr(PreTrainedConfig.validate_rope, "__vllm_nested_rope_patched__"):

297 return

298

299 - _original_validate_rope = PretrainedConfig.validate_rope

299 + _original_validate_rope = PreTrainedConfig.validate_rope

300

301 @wraps(_original_validate_rope)

302 def patched_validate_rope(self, *args, **kwargs):

@@ -347,11 +347,11 @@ class HFConfigParser(ConfigParserBase):

347 revision: str | None = None,

348 code_revision: str | None = None,

349 **kwargs,

350 - ) -> tuple[dict, PretrainedConfig]:

350 + ) -> tuple[dict, PreTrainedConfig]:

351 kwargs["local_files_only"] = huggingface_hub.constants.HF_HUB_OFFLINE

352 trust_remote_code |= kwargs.get("trust_remote_code", False)

353 kwargs = without_trust_remote_code(kwargs)

354 - config_dict, _ = PretrainedConfig.get_config_dict(

354 + config_dict, _ = PreTrainedConfig.get_config_dict(

355 model,

356 revision=revision,

357 code_revision=code_revision,

@@ -374,7 +374,7 @@ class HFConfigParser(ConfigParserBase):

374 # through and remain unchanged by this elif block

375 dummy_model_type = f"dummy_{model_type}"

376 dummy_kwargs = dict(architectures=[""], model_type=dummy_model_type)

377 - dummy_config = PretrainedConfig(**dummy_kwargs)

377 + dummy_config = PreTrainedConfig(**dummy_kwargs)

378 dummy_model_type = hf_overrides(dummy_config).model_type

379 model_type = dummy_model_type.removeprefix("dummy_")

380

@@ -450,7 +450,7 @@ class MistralConfigParser(ConfigParserBase):

450 revision: str | None = None,

451 code_revision: str | None = None,

452 **kwargs,

453 - ) -> tuple[dict, PretrainedConfig]:

453 + ) -> tuple[dict, PreTrainedConfig]:

454 # This function loads a params.json config which

455 # should be used when loading models in mistral format

456 config_dict = _download_mistral_config_file(model, revision)

@@ -466,7 +466,7 @@ class MistralConfigParser(ConfigParserBase):

466

467 # Get missing fields from HF config if available

468 try:

469 - hf_config_dict, _ = PretrainedConfig.get_config_dict(

469 + hf_config_dict, _ = PreTrainedConfig.get_config_dict(

470 model,

471 revision=revision,

472 code_revision=code_revision,

@@ -540,7 +540,7 @@ def register_config_parser(config_format: str):

540 ... revision: str | None = None,

541 ... code_revision: str | None = None,

542 ... **kwargs,

543 - ... ) -> tuple[dict, PretrainedConfig]:

543 + ... ) -> tuple[dict, PreTrainedConfig]:

544 ... raise NotImplementedError

545 >>>

546 >>> type(get_config_parser("custom_config_parser"))

@@ -571,7 +571,7 @@ def register_config_parser(config_format: str):

571 return _wrapper

572

573

574 -def set_default_rope_theta(config: PretrainedConfig, default_theta: float) -> None:

574 +def set_default_rope_theta(config: PreTrainedConfig, default_theta: float) -> None:

575 """Some models may have no rope_theta in their config but still use RoPE.

576 This function sets a default rope_theta if it's missing."""

577 if getattr(config, "rope_parameters", None) is None:

@@ -639,7 +639,7 @@ def patch_legacy_rope_type(rope_parameters: dict[str, Any] | None) -> None:

639 _patch_legacy_rope_type(rope_parameters)

640

641

642 -def patch_rope_parameters(config: PretrainedConfig) -> None:

642 +def patch_rope_parameters(config: PreTrainedConfig) -> None:

643 """Provide backwards compatibility for RoPE."""

644 from vllm.config.utils import getattr_iter

645

@@ -661,7 +661,7 @@ def patch_rope_parameters(config: PretrainedConfig) -> None:

661 config.validate_rope()

662

663

664 -def _iter_rope_parameters(config: PretrainedConfig) -> Iterator[dict[str, Any]]:

664 +def _iter_rope_parameters(config: PreTrainedConfig) -> Iterator[dict[str, Any]]:

665 """Yield a config's rope parameters, one dict per layer type if nested."""

666 rope_parameters = getattr(config, "rope_parameters", None)

667 if not isinstance(rope_parameters, dict):

@@ -673,7 +673,7 @@ def _iter_rope_parameters(config: PretrainedConfig) -> Iterator[dict[str, Any]]:

673 yield rope_parameters

674

675

676 -def _mrope_section(config: PretrainedConfig) -> Sequence[int] | None:

676 +def _mrope_section(config: PreTrainedConfig) -> Sequence[int] | None:

677 """Return the M-RoPE section this config declares, if any.

678

679 `xdrope_section` is the legacy name HunYuan-VL checkpoints use for the same

@@ -700,11 +700,11 @@ def _mrope_section(config: PretrainedConfig) -> Sequence[int] | None:

700 return section if isinstance(section, (list, tuple)) else None

701

702

703 -def _uses_mrope(config: PretrainedConfig) -> bool:

703 +def _uses_mrope(config: PreTrainedConfig) -> bool:

704 return _mrope_section(config) is not None

705

706

707 -def uses_mrope(config: PretrainedConfig) -> bool:

707 +def uses_mrope(config: PreTrainedConfig) -> bool:

708 """Detect if the model with this config uses M-ROPE."""

709 return (

710 _uses_mrope(config)

@@ -713,7 +713,7 @@ def uses_mrope(config: PretrainedConfig) -> bool:

713 )

714

715

716 -def thinker_uses_mrope(config: PretrainedConfig) -> bool:

716 +def thinker_uses_mrope(config: PreTrainedConfig) -> bool:

717 """Detect if the model contains a thinker config and it uses M-ROPE."""

718 thinker_config = getattr(config, "thinker_config", None)

719 if thinker_config is None:

@@ -726,7 +726,7 @@ def thinker_uses_mrope(config: PretrainedConfig) -> bool:

726 return uses_mrope(thinker_text_config)

727

728

729 -def mrope_num_dims(config: PretrainedConfig) -> int:

729 +def mrope_num_dims(config: PreTrainedConfig) -> int:

730 """Number of M-RoPE position channels the model consumes.

731

732 Each section entry sizes one channel, so the section length is the channel

@@ -745,10 +745,10 @@ def mrope_num_dims(config: PretrainedConfig) -> int:

745 return 3

746

747

748 -def is_encoder_decoder(config: PretrainedConfig) -> bool:

748 +def is_encoder_decoder(config: PreTrainedConfig) -> bool:

749 """Detect if the model with this config is used as an encoder/decoder."""

750

751 - def _is_encoder_decoder(config: PretrainedConfig) -> bool:

751 + def _is_encoder_decoder(config: PreTrainedConfig) -> bool:

752 return getattr(config, "is_encoder_decoder", False)

753

754 return _is_encoder_decoder(config) or _is_encoder_decoder(config.get_text_config())

@@ -761,7 +761,7 @@ def _maybe_update_auto_config_kwargs(kwargs: dict[str, Any], model_type: str):

761 return kwargs

762

763

764 -def _maybe_remap_hf_config_attrs(config: PretrainedConfig) -> PretrainedConfig:

764 +def _maybe_remap_hf_config_attrs(config: PreTrainedConfig) -> PreTrainedConfig:

765 """Remap config attributes to match the expected names."""

766 for old_attr, new_attr in _CONFIG_ATTRS_MAPPING.items():

767 if hasattr(config, old_attr):

@@ -798,7 +798,7 @@ def maybe_override_with_speculators(

798

799 """

800 kwargs["local_files_only"] = huggingface_hub.constants.HF_HUB_OFFLINE

801 - config_dict, _ = PretrainedConfig.get_config_dict(

801 + config_dict, _ = PreTrainedConfig.get_config_dict(

802 model,

803 revision=revision,

804 token=hf_token,

@@ -840,9 +840,9 @@ def get_config(

840 code_revision: str | None = None,

841 config_format: str | ConfigFormat = "auto",

842 hf_overrides_kw: dict[str, Any] | None = None,

843 - hf_overrides_fn: Callable[[PretrainedConfig], PretrainedConfig] | None = None,

843 + hf_overrides_fn: Callable[[PreTrainedConfig], PreTrainedConfig] | None = None,

844 **kwargs,

845 -) -> PretrainedConfig:

845 +) -> PreTrainedConfig:

846 if config_format == "auto":

847 try:

848 # First check for Mistral to avoid defaulting to

@@ -953,7 +953,7 @@ def get_config(

953 # Exhaustively patch RoPE parameters everywhere they might be

954 patch_rope_parameters(config)

955 patch_rope_parameters(config.get_text_config())

956 - SubConfigs: TypeAlias = dict[str, PretrainedConfig]

956 + SubConfigs: TypeAlias = dict[str, PreTrainedConfig]

957 sub_configs: SubConfigs | None = getattr(config, "sub_configs", None)

958 if sub_configs:

959 for sub_config in sub_configs:

@@ -1212,7 +1212,7 @@ def get_hf_image_processor_config(

1212 )

1213

1214

1215 -def get_hf_text_config(config: PretrainedConfig):

1215 +def get_hf_text_config(config: PreTrainedConfig):

1216 """Get the "sub" config relevant to llm for multi modal models.

1217 No op for pure text models.

1218 """

vllm/transformers_utils/config_parser_base.py

+2 -2

@@ -4,7 +4,7 @@

4 from abc import ABC, abstractmethod

5 from pathlib import Path

6

7 -from transformers import PretrainedConfig

7 +from transformers import PreTrainedConfig

8

9

10 class ConfigParserBase(ABC):

@@ -16,5 +16,5 @@ class ConfigParserBase(ABC):

16 revision: str | None = None,

17 code_revision: str | None = None,

18 **kwargs,

19 - ) -> tuple[dict, PretrainedConfig]:

19 + ) -> tuple[dict, PreTrainedConfig]:

20 raise NotImplementedError

vllm/transformers_utils/configs/AXK1.py

+4 -4

@@ -2,16 +2,16 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 from typing import Any

4

5 -from transformers import PretrainedConfig

5 +from transformers import PreTrainedConfig

6

7

8 -class AXK1Config(PretrainedConfig):

8 +class AXK1Config(PreTrainedConfig):

9 r"""This is the configuration class to store the configuration of a [`AXK1Model`].

10 It is used to instantiate an A.X model according to the specified arguments,

11 defining the model architecture. Instantiating a configuration with the defaults

12 will yield a similar configuration to that of the A.X K1.

13 - Configuration objects inherit from [`PretrainedConfig`] and can be used to control

14 - the model outputs. Read the documentation from [`PretrainedConfig`] for more

13 + Configuration objects inherit from [`PreTrainedConfig`] and can be used to control

14 + the model outputs. Read the documentation from [`PreTrainedConfig`] for more

15 information.

16

17 Args:

vllm/transformers_utils/configs/afmoe.py

+2 -2

@@ -1,10 +1,10 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4 -from transformers.configuration_utils import PretrainedConfig

4 +from transformers.configuration_utils import PreTrainedConfig

5

6

7 -class AfmoeConfig(PretrainedConfig):

7 +class AfmoeConfig(PreTrainedConfig):

8 model_type = "afmoe"

9

10 def __init__(

vllm/transformers_utils/configs/bagel.py

+2 -2

@@ -1,10 +1,10 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 -from transformers import PretrainedConfig, SiglipVisionConfig

3 +from transformers import PreTrainedConfig, SiglipVisionConfig

4 from transformers.models.qwen2 import Qwen2Config

5

6

7 -class BagelConfig(PretrainedConfig):

7 +class BagelConfig(PreTrainedConfig):

8 """Configuration class for BAGEL model."""

9

10 model_type = "bagel"

vllm/transformers_utils/configs/bailing_moe_v3_vl.py

+5 -5

@@ -9,11 +9,11 @@ from __future__ import annotations

9

10 from typing import Any

11

12 -from transformers.configuration_utils import PretrainedConfig

12 +from transformers.configuration_utils import PreTrainedConfig

13

14

15 def _normalize_mrope_parameters(

16 - config: PretrainedConfig,

16 + config: PreTrainedConfig,

17 mrope_section: list[int],

18 ) -> None:

19 """Add the top-level Bailing M-RoPE split to a text config."""

@@ -50,7 +50,7 @@ def _build_layer_types(

50 ]

51

52

53 -class BailingMoeV3TextConfig(PretrainedConfig):

53 +class BailingMoeV3TextConfig(PreTrainedConfig):

54 model_type = "bailing_hybrid"

55 base_config_key = "text_config"

56 keys_to_ignore_at_inference = ["past_key_values"]

@@ -230,7 +230,7 @@ class BailingMoeV3TextConfig(PretrainedConfig):

230 self.validate_layer_type()

231

232

233 -class BailingMoeV3VisionConfig(PretrainedConfig):

233 +class BailingMoeV3VisionConfig(PreTrainedConfig):

234 model_type = "qwen3_moe_vit"

235 base_config_key = "vision_config"

236

@@ -273,7 +273,7 @@ class BailingMoeV3VisionConfig(PretrainedConfig):

273 self.initializer_range = initializer_range

274

275

276 -class BailingMoeV3VLConfig(PretrainedConfig):

276 +class BailingMoeV3VLConfig(PreTrainedConfig):

277 model_type = "bailing_moe_v3_vl"

278 keys_to_ignore_at_inference = ["past_key_values"]

279 sub_configs = {

vllm/transformers_utils/configs/chatglm.py

+2 -2

@@ -3,10 +3,10 @@

3

4 # Adapted from

5 # https://github.com/zai-org/ChatGLM2-6B

6 -from transformers import PretrainedConfig

6 +from transformers import PreTrainedConfig

7

8

9 -class ChatGLMConfig(PretrainedConfig):

9 +class ChatGLMConfig(PreTrainedConfig):

10 model_type = "chatglm"

11 attribute_map = {

12 "num_hidden_layers": "num_layers",

vllm/transformers_utils/configs/colmodernvbert.py

+2 -2

@@ -8,10 +8,10 @@ with a pixel shuffle connector and ColBERT-style 128-dim per-token embeddings.

8 Reference: https://huggingface.co/ModernVBERT/colmodernvbert-merged

9 """

10

11 -from transformers import ModernBertConfig, PretrainedConfig, SiglipVisionConfig

11 +from transformers import ModernBertConfig, PreTrainedConfig, SiglipVisionConfig

12

13

14 -class ColModernVBertConfig(PretrainedConfig):

14 +class ColModernVBertConfig(PreTrainedConfig):

15 model_type = "colmodernvbert"

16

17 def __init__(

vllm/transformers_utils/configs/cosmos3_edge.py

+3 -3

@@ -1,7 +1,7 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4 -from transformers.configuration_utils import PretrainedConfig

4 +from transformers.configuration_utils import PreTrainedConfig

5 from transformers.models.siglip2.configuration_siglip2 import Siglip2VisionConfig

6

7 from .nemotron_h import NemotronHConfig

@@ -50,7 +50,7 @@ class Cosmos3EdgeTextConfig(NemotronHConfig):

50 return config_dict

51

52

53 -class Cosmos3EdgeProjectorConfig(PretrainedConfig):

53 +class Cosmos3EdgeProjectorConfig(PreTrainedConfig):

54 model_type = "cosmos3_edge_projector"

55

56 def __init__(

@@ -74,7 +74,7 @@ class Cosmos3EdgeVisionConfig(Siglip2VisionConfig):

74 model_type = "cosmos3_edge_vision"

75

76

77 -class Cosmos3EdgeConfig(PretrainedConfig):

77 +class Cosmos3EdgeConfig(PreTrainedConfig):

78 model_type = "cosmos3_edge"

79 keys_to_ignore_at_inference = ["past_key_values"]

80 has_no_defaults_at_init = True

vllm/transformers_utils/configs/deepseek_v4.py

+2 -2

@@ -2,10 +2,10 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 from typing import Any

4

5 -from transformers import PretrainedConfig

5 +from transformers import PreTrainedConfig

6

7

8 -class DeepseekV4Config(PretrainedConfig):

8 +class DeepseekV4Config(PreTrainedConfig):

9 model_type = "deepseek_v4"

10

11 def __init__(

vllm/transformers_utils/configs/deepseek_v41.py

+3 -3

@@ -2,10 +2,10 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 from typing import Any

4

5 -from transformers import PretrainedConfig

5 +from transformers import PreTrainedConfig

6

7

8 -class DeepseekV41Config(PretrainedConfig):

8 +class DeepseekV41Config(PreTrainedConfig):

9 """DeepSeek V4.1 config.

10

11 The HF config nests the text model under ``text_config`` and the vision

@@ -34,7 +34,7 @@ class DeepseekV41Config(PretrainedConfig):

34 for key, value in text_config.items():

35 if key == "model_type":

36 continue

37 - # Don't clobber PretrainedConfig properties (e.g. is_encoder_decoder).

37 + # Don't clobber PreTrainedConfig properties (e.g. is_encoder_decoder).

38 if isinstance(getattr(type(self), key, None), property):

39 continue

40 setattr(self, key, value)

vllm/transformers_utils/configs/deepseek_vl2.py

+4 -4

@@ -4,10 +4,10 @@

4 # adapted from https://github.com/deepseek-ai/DeepSeek-VL2/blob/faf18023f24b962b32d9f0a2d89e402a8d383a78/deepseek_vl2/models/modeling_deepseek_vl_v2.py#L115-L268

5

6 from huggingface_hub.dataclasses import strict

7 -from transformers import DeepseekV2Config, PretrainedConfig

7 +from transformers import DeepseekV2Config, PreTrainedConfig

8

9

10 -class VisionEncoderConfig(PretrainedConfig):

10 +class VisionEncoderConfig(PreTrainedConfig):

11 model_type: str = "vision"

12

13 model_name: str = "vit_so400m_patch14_siglip_384.webli"

@@ -58,7 +58,7 @@ class VisionEncoderConfig(PretrainedConfig):

58 super().__init__(**kwargs)

59

60

61 -class MlpProjectorConfig(PretrainedConfig):

61 +class MlpProjectorConfig(PreTrainedConfig):

62 model_type = "mlp_projector"

63 projector_type: str = "downsample_mlp_gelu"

64 input_dim: int = 1152

@@ -98,7 +98,7 @@ class DeepseekVLV2TextConfig(DeepseekV2Config):

98 kv_lora_rank: int | None = DeepseekV2Config.kv_lora_rank

99

100

101 -class DeepseekVLV2Config(PretrainedConfig):

101 +class DeepseekVLV2Config(PreTrainedConfig):

102 model_type = "deepseek_vl_v2"

103

104 tile_tag: str = "2D"

vllm/transformers_utils/configs/diffusion_gemma.py

+6 -6

@@ -2,12 +2,12 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 from typing import Any

4

5 -from transformers import PretrainedConfig

5 +from transformers import PreTrainedConfig

6 from transformers.models.gemma4.configuration_gemma4 import Gemma4VisionConfig

7

8

9 -def _init_text_config(self: PretrainedConfig, **kwargs: Any) -> None:

10 - PretrainedConfig.__init__(self, **kwargs)

9 +def _init_text_config(self: PreTrainedConfig, **kwargs: Any) -> None:

10 + PreTrainedConfig.__init__(self, **kwargs)

11 # DiffusionGemma always uses MoE and K=V sharing for full_attention

12 # layers. The HF reference removed these config fields entirely.

13 if getattr(self, "num_experts", None):

@@ -15,14 +15,14 @@ def _init_text_config(self: PretrainedConfig, **kwargs: Any) -> None:

15 self.attention_k_eq_v = True

16

17

18 -class DiffusionGemmaTextConfig(PretrainedConfig):

18 +class DiffusionGemmaTextConfig(PreTrainedConfig):

19 model_type = "diffusion_gemma_text"

20

21 def __init__(self, **kwargs: Any):

22 _init_text_config(self, **kwargs)

23

24

25 -class DiffusionGemmaConfig(PretrainedConfig):

25 +class DiffusionGemmaConfig(PreTrainedConfig):

26 model_type = "diffusion_gemma"

27

28 def __init__(

@@ -41,4 +41,4 @@ class DiffusionGemmaConfig(PretrainedConfig):

41 else:

42 self.vision_config = vision_config

43 self.audio_config = None

44 - PretrainedConfig.__init__(self, **kwargs)

44 + PreTrainedConfig.__init__(self, **kwargs)

vllm/transformers_utils/configs/dotsocr.py

+2 -2

@@ -2,11 +2,11 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 from typing import Any

4

5 -from transformers.configuration_utils import PretrainedConfig

5 +from transformers.configuration_utils import PreTrainedConfig

6 from transformers.models.qwen2 import Qwen2Config

7

8

9 -class DotsVisionConfig(PretrainedConfig):

9 +class DotsVisionConfig(PreTrainedConfig):

10 model_type: str = "dots_vit"

11

12 def __init__(

vllm/transformers_utils/configs/eagle.py

+4 -4

@@ -3,22 +3,22 @@

3

4 import os

5

6 -from transformers import AutoConfig, DeepseekV2Config, PretrainedConfig

6 +from transformers import AutoConfig, DeepseekV2Config, PreTrainedConfig

7

8 from vllm.transformers_utils.utils import without_trust_remote_code

9

10

11 -class EAGLEConfig(PretrainedConfig):

11 +class EAGLEConfig(PreTrainedConfig):

12 model_type = "eagle"

13

14 def __init__(

15 self,

16 - model: PretrainedConfig | dict | None = None,

16 + model: PreTrainedConfig | dict | None = None,

17 truncated_vocab_size: int | None = None,

18 method: str | None = "eagle",

19 **kwargs,

20 ):

21 - model_config: PretrainedConfig | DeepseekV2Config | None

21 + model_config: PreTrainedConfig | DeepseekV2Config | None

22 if isinstance(model, dict):

23 model_config = AutoConfig.for_model(**model)

24 else:

vllm/transformers_utils/configs/extract_hidden_states.py

+5 -5

@@ -5,17 +5,17 @@ the extract_hidden_states spec decoding method."""

5

6 import os

7

8 -from transformers import PretrainedConfig

8 +from transformers import PreTrainedConfig

9

10 from vllm.transformers_utils.utils import without_trust_remote_code

11

12

13 -class ExtractHiddenStatesConfig(PretrainedConfig):

13 +class ExtractHiddenStatesConfig(PreTrainedConfig):

14 model_type = "extract_hidden_states"

15

16 def __init__(

17 self,

18 - model: PretrainedConfig | dict | None = None,

18 + model: PreTrainedConfig | dict | None = None,

19 method: str | None = "extract_hidden_states",

20 **kwargs,

21 ):

@@ -24,7 +24,7 @@ class ExtractHiddenStatesConfig(PretrainedConfig):

24 if isinstance(model, dict):

25 model_dict = model

26 source_text_config = None

27 - elif isinstance(model, PretrainedConfig):

27 + elif isinstance(model, PreTrainedConfig):

28 model_dict = model.to_dict()

29 text_config = model.get_text_config()

30 source_text_config = text_config if text_config is not model else None

@@ -40,7 +40,7 @@ class ExtractHiddenStatesConfig(PretrainedConfig):

40 combined["architectures"] = ["ExtractHiddenStatesModel"]

41

42 # to_dict() and kwargs both flatten text_config to a plain dict;

43 - # downstream get_hf_text_config() needs it as a PretrainedConfig

43 + # downstream get_hf_text_config() needs it as a PreTrainedConfig

44 # for attribute access. Re-insert the original object.

45 if source_text_config is not None:

46 combined["text_config"] = source_text_config

vllm/transformers_utils/configs/falcon.py

+2 -2

@@ -20,10 +20,10 @@

20 # limitations under the License.

21 """Falcon configuration."""

22

23 -from transformers.configuration_utils import PretrainedConfig

23 +from transformers.configuration_utils import PreTrainedConfig

24

25

26 -class RWConfig(PretrainedConfig):

26 +class RWConfig(PreTrainedConfig):

27 model_type = "falcon"

28 keys_to_ignore_at_inference = ["past_key_values"]

29 attribute_map = {

vllm/transformers_utils/configs/funaudiochat.py

+5 -5

@@ -3,7 +3,7 @@

3

4 from __future__ import annotations

5

6 -from transformers import CONFIG_MAPPING, PretrainedConfig

6 +from transformers import CONFIG_MAPPING, PreTrainedConfig

7

8 # NOTE: Temporary shim for FunAudioChat checkpoints.

9 # These checkpoints use `model_type="funaudiochat"`, which is not currently

@@ -13,7 +13,7 @@ from transformers import CONFIG_MAPPING, PretrainedConfig

13 # provides an `auto_map`) and vLLM can rely on `AutoConfig.from_pretrained()`.

14

15

16 -class FunAudioChatAudioEncoderConfig(PretrainedConfig):

16 +class FunAudioChatAudioEncoderConfig(PreTrainedConfig):

17 model_type = "funaudiochat_audio_encoder"

18

19 def __init__(

@@ -74,7 +74,7 @@ class FunAudioChatAudioEncoderConfig(PretrainedConfig):

74 self.pad_token_id = pad_token_id

75

76

77 -class FunAudioChatConfig(PretrainedConfig):

77 +class FunAudioChatConfig(PreTrainedConfig):

78 model_type = "funaudiochat"

79 attribute_map = {

80 "audio_token_id": "audio_token_index",

@@ -82,8 +82,8 @@ class FunAudioChatConfig(PretrainedConfig):

82

83 def __init__(

84 self,

85 - audio_config: PretrainedConfig | dict | None = None,

86 - text_config: PretrainedConfig | dict | None = None,

85 + audio_config: PreTrainedConfig | dict | None = None,

86 + text_config: PreTrainedConfig | dict | None = None,

87 audio_token_index: int = 151646,

88 ignore_index: int = -100,

89 hidden_size: int | None = None,

vllm/transformers_utils/configs/gemma4.py

+3 -3

@@ -4,12 +4,12 @@

4

5 from copy import copy

6

7 -from transformers import PretrainedConfig

7 +from transformers import PreTrainedConfig

8

9

10 def gemma4_layer_config(

11 - text_config: PretrainedConfig, layer_idx: int

12 -) -> PretrainedConfig:

11 + text_config: PreTrainedConfig, layer_idx: int

12 +) -> PreTrainedConfig:

13 """The Gemma4 text config as it applies to one layer.

14

15 Gemma4 uses a larger head dimension on its full attention layers than on its

vllm/transformers_utils/configs/glm5_next.py

+6 -6

@@ -1,14 +1,14 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4 -from transformers.configuration_utils import PretrainedConfig

4 +from transformers.configuration_utils import PreTrainedConfig

5

6 from vllm.logger import init_logger

7

8 logger = init_logger(__name__)

9

10

11 -class Glm5NextTextConfig(PretrainedConfig):

11 +class Glm5NextTextConfig(PreTrainedConfig):

12 model_type = "glm5_next_text"

13 base_config_key = "text_config"

14 keys_to_ignore_at_inference = ["past_key_values"]

@@ -271,7 +271,7 @@ class Glm5NextTextConfig(PretrainedConfig):

271 ]

272

273

274 -class Glm5NextVisionConfig(PretrainedConfig):

274 +class Glm5NextVisionConfig(PreTrainedConfig):

275 model_type = "glm5_next_vision"

276 base_config_key = "vision_config"

277

@@ -324,7 +324,7 @@ class Glm5NextVisionConfig(PretrainedConfig):

324 self.swiglu_limit = swiglu_limit

325

326

327 -class Glm5NextConfig(PretrainedConfig):

327 +class Glm5NextConfig(PreTrainedConfig):

328 model_type = "glm5_next"

329 sub_configs = {

330 "vision_config": Glm5NextVisionConfig,

@@ -345,7 +345,7 @@ class Glm5NextConfig(PretrainedConfig):

345 **kwargs,

346 ):

347 # Init super() first so base-class defaults don't clobber text-config

348 - # values set below (PretrainedConfig has many text-related defaults

348 + # values set below (PreTrainedConfig has many text-related defaults

349 # that differ from Glm5NextTextConfig).

350 super().__init__(**kwargs)

351

@@ -379,7 +379,7 @@ class Glm5NextConfig(PretrainedConfig):

379 # and must NOT be mirrored onto text_config: ``architectures`` /

380 # ``torch_dtype`` differ between the top-level config and the text

381 # sub-config, and mirroring them makes the top-level ``architectures``

382 - # silently read back as None (PretrainedConfig initializes both to None),

382 + # silently read back as None (PreTrainedConfig initializes both to None),

383 # which then fails model-class resolution ("No model architectures are

384 # specified").

385 _UNMIRRORED_KEYS = [

vllm/transformers_utils/configs/granite4_vision.py

+3 -3

@@ -6,7 +6,7 @@ from typing import Any

6 import transformers

7

8

9 -class Granite4VisionConfig(transformers.PretrainedConfig):

9 +class Granite4VisionConfig(transformers.PreTrainedConfig):

10 """Configuration for Granite 4 Vision model.

11

12 This config is needed because the granite4_vision model type is not yet

@@ -81,7 +81,7 @@ class Granite4VisionConfig(transformers.PretrainedConfig):

81 **vision_config

82 )

83 else:

84 - self.vision_config = transformers.PretrainedConfig(**vision_config)

84 + self.vision_config = transformers.PreTrainedConfig(**vision_config)

85

86 text_model_type = text_config.get("model_type", "granite")

87 if text_model_type in transformers.CONFIG_MAPPING:

@@ -89,6 +89,6 @@ class Granite4VisionConfig(transformers.PretrainedConfig):

89 **text_config

90 )

91 else:

92 - self.text_config = transformers.PretrainedConfig(**text_config)

92 + self.text_config = transformers.PreTrainedConfig(**text_config)

93

94 super().__init__(**kwargs)

vllm/transformers_utils/configs/hy_v3.py

+4 -4

@@ -2,16 +2,16 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 from typing import Any

4

5 -from transformers.configuration_utils import PretrainedConfig

5 +from transformers.configuration_utils import PreTrainedConfig

6

7

8 -class HYV3Config(PretrainedConfig):

8 +class HYV3Config(PreTrainedConfig):

9 r"""This is the configuration class to store the configuration of a [`HYV3Model`].

10 It is used to instantiate a HYV3 model (HY V3 MoE language model) according to

11 the specified arguments.

12

13 - Configuration objects inherit from [`PretrainedConfig`] and can be used to

14 - control the model outputs. Read the documentation from [`PretrainedConfig`]

13 + Configuration objects inherit from [`PreTrainedConfig`] and can be used to

14 + control the model outputs. Read the documentation from [`PreTrainedConfig`]

15 for more information.

16

17 Args:

vllm/transformers_utils/configs/hyperclovax.py

+4 -4

@@ -17,16 +17,16 @@

17 # limitations under the License.

18 """HyperCLOVA X model configuration."""

19

20 -from transformers.configuration_utils import PretrainedConfig

20 +from transformers.configuration_utils import PreTrainedConfig

21

22

23 -class HyperCLOVAXConfig(PretrainedConfig):

23 +class HyperCLOVAXConfig(PreTrainedConfig):

24 r"""This is the configuration class to store the configuration of a

25 [`HyperCLOVAXModel`]. It is used to instantiate a HyperCLOVAX model

26 according to the specified arguments, defining the model architecture.

27 - Configuration objects inherit from [`PretrainedConfig`] and can be used

27 + Configuration objects inherit from [`PreTrainedConfig`] and can be used

28 to control the model outputs. Read the documentation from

29 - [`PretrainedConfig`] for more information.

29 + [`PreTrainedConfig`] for more information.

30

31 Args:

32 vocab_size (`int`, *optional*, defaults to 32000):

vllm/transformers_utils/configs/kimi_k25.py

+3 -3

@@ -7,10 +7,10 @@ A video-chunk is the smallest independently processable unit of video.

7 """

8

9 from transformers import DeepseekV3Config

10 -from transformers.configuration_utils import PretrainedConfig

10 +from transformers.configuration_utils import PreTrainedConfig

11

12

13 -class KimiK25VisionConfig(PretrainedConfig):

13 +class KimiK25VisionConfig(PreTrainedConfig):

14 model_type = "kimi_k25_vision"

15

16 def __init__(

@@ -59,7 +59,7 @@ class KimiK25VisionConfig(PretrainedConfig):

59 self.projector_ln_eps = projector_ln_eps

60

61

62 -class KimiK25Config(PretrainedConfig):

62 +class KimiK25Config(PreTrainedConfig):

63 """Kimi-K2.5 model configuration.

64

65 Kimi-K2.5 extends Kimi-K2 with vision support using video-chunks.

vllm/transformers_utils/configs/kimi_k3.py

+3 -3

@@ -2,7 +2,7 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 """Kimi-K3 multimodal configuration."""

4

5 -from transformers.configuration_utils import PretrainedConfig

5 +from transformers.configuration_utils import PreTrainedConfig

6

7 from vllm.logger import init_logger

8 from vllm.transformers_utils.configs.kimi_linear import KimiLinearConfig

@@ -10,7 +10,7 @@ from vllm.transformers_utils.configs.kimi_linear import KimiLinearConfig

10 logger = init_logger(__name__)

11

12

13 -class KimiK3VisionConfig(PretrainedConfig):

13 +class KimiK3VisionConfig(PreTrainedConfig):

14 model_type = "kimi_k3_vision"

15

16 def __init__(

@@ -83,7 +83,7 @@ class KimiK3VisionConfig(PretrainedConfig):

83 self.intermediate_size = vt_intermediate_size

84

85

86 -class KimiK3Config(PretrainedConfig):

86 +class KimiK3Config(PreTrainedConfig):

87 model_type = "kimi_k3"

88

89 def __init__(

vllm/transformers_utils/configs/kimi_linear.py

+2 -2

@@ -1,10 +1,10 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4 -from transformers.configuration_utils import PretrainedConfig

4 +from transformers.configuration_utils import PreTrainedConfig

5

6

7 -class KimiLinearConfig(PretrainedConfig):

7 +class KimiLinearConfig(PreTrainedConfig):

8 model_type = "kimi_linear"

9 keys_to_ignore_at_inference = ["past_key_values"]

10

vllm/transformers_utils/configs/kimi_vl.py

+2 -2

@@ -3,12 +3,12 @@

3 # Adapted from https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/blob/main/configuration_kimi_vl.py

4

5 from transformers import DeepseekV2Config

6 -from transformers.configuration_utils import PretrainedConfig

6 +from transformers.configuration_utils import PreTrainedConfig

7

8 from vllm.transformers_utils.configs.moonvit import MoonViTConfig

9

10

11 -class KimiVLConfig(PretrainedConfig):

11 +class KimiVLConfig(PreTrainedConfig):

12 model_type = "kimi_vl"

13

14 def __init__(

vllm/transformers_utils/configs/lfm2_moe.py

+4 -4

@@ -2,17 +2,17 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 from typing import Any

4

5 -from transformers.configuration_utils import PretrainedConfig

5 +from transformers.configuration_utils import PreTrainedConfig

6

7

8 -class Lfm2MoeConfig(PretrainedConfig):

8 +class Lfm2MoeConfig(PreTrainedConfig):

9 r"""This is the configuration class to store the configuration of a [`Lfm2MoeModel`]. It is used to instantiate a LFM2 Moe

10 model according to the specified arguments, defining the model architecture. Instantiating a configuration with the

11 defaults will yield a similar configuration to that of the LFM2-8B-A1B model.

12 e.g. [LiquidAI/LFM2-8B-A1B](https://huggingface.co/LiquidAI/LFM2-8B-A1B)

13

14 - Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the

15 - documentation from [`PretrainedConfig`] for more information.

14 + Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the

15 + documentation from [`PreTrainedConfig`] for more information.

16

17

18 Args:

vllm/transformers_utils/configs/medusa.py

+2 -2

@@ -3,12 +3,12 @@

3

4 import os

5

6 -from transformers import PretrainedConfig

6 +from transformers import PreTrainedConfig

7

8 from vllm.transformers_utils.utils import without_trust_remote_code

9

10

11 -class MedusaConfig(PretrainedConfig):

11 +class MedusaConfig(PreTrainedConfig):

12 model_type = "medusa"

13

14 def __init__(

vllm/transformers_utils/configs/midashenglm.py

+3 -3

@@ -22,13 +22,13 @@

22 # See the License for the specific language governing permissions and

23 # limitations under the License.

24

25 -from transformers import PretrainedConfig

25 +from transformers import PreTrainedConfig

26 from transformers.models.qwen2_5_omni.configuration_qwen2_5_omni import (

27 Qwen2_5OmniTextConfig,

28 )

29

30

31 -class DashengConfig(PretrainedConfig):

31 +class DashengConfig(PreTrainedConfig):

32 model_type = "midashenglm_dasheng_encoder"

33

34 def __init__(

@@ -80,7 +80,7 @@ class DashengConfig(PretrainedConfig):

80 super().__init__(**kwargs)

81

82

83 -class MiDashengLMConfig(PretrainedConfig):

83 +class MiDashengLMConfig(PreTrainedConfig):

84 model_type = "midashenglm"

85

86 def __init__(

vllm/transformers_utils/configs/mimo_v2_omni.py

+2 -2

@@ -2,10 +2,10 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4

5 -from transformers import PretrainedConfig

5 +from transformers import PreTrainedConfig

6

7

8 -class Mimo_VLVisionConfig(PretrainedConfig):

8 +class Mimo_VLVisionConfig(PreTrainedConfig):

9 model_type = "mimovl"

10 base_config_key = "vision_config"

11

vllm/transformers_utils/configs/minimax_m3.py

+3 -3

@@ -2,10 +2,10 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 from typing import Any

4

5 -from transformers import PretrainedConfig

5 +from transformers import PreTrainedConfig

6

7

8 -class MiniMaxM3TextConfig(PretrainedConfig):

8 +class MiniMaxM3TextConfig(PreTrainedConfig):

9 """Config for the MiniMax M3 text backbone (MiniMaxM3SparseForCausalLM).

10

11 Defaults mirror the ``text_config`` of the MiniMax-M3-preview checkpoint.

@@ -120,7 +120,7 @@ class MiniMaxM3MTPConfig(MiniMaxM3TextConfig):

120 super().__init__(num_hidden_layers=num_hidden_layers, **kwargs)

121

122

123 -class MiniMaxM3Config(PretrainedConfig):

123 +class MiniMaxM3Config(PreTrainedConfig):

124 """Top-level MiniMax M3 (VL) config.

125

126 Holds the text backbone as ``text_config`` so that

vllm/transformers_utils/configs/mistral.py

+6 -6

@@ -2,7 +2,7 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 from typing import Any

4

5 -from transformers import PretrainedConfig, WhisperConfig

5 +from transformers import PreTrainedConfig, WhisperConfig

6

7 from vllm.logger import init_logger

8

@@ -12,7 +12,7 @@ logger = init_logger(__name__)

12 def adapt_config_dict(

13 config_dict: dict[str, Any],

14 defaults: dict[str, Any],

15 -) -> PretrainedConfig:

15 +) -> PreTrainedConfig:

16 config_dict = _remap_general_mistral_args(config_dict)

17 config_dict = _remap_mistral_sliding_window(config_dict)

18

@@ -86,7 +86,7 @@ def adapt_config_dict(

86 for k, v in defaults.items():

87 config_dict.setdefault(k, v)

88

89 - config = PretrainedConfig.from_dict(config_dict)

89 + config = PreTrainedConfig.from_dict(config_dict)

90

91 logger.debug("Initialized config %s", config)

92

@@ -103,8 +103,8 @@ def _remap_mistral_vision_args(config: dict) -> dict:

103 config = {

104 "model_type": "pixtral",

105 "architectures": ["PixtralForConditionalGeneration"],

106 - "text_config": PretrainedConfig.from_dict(config),

107 - "vision_config": PretrainedConfig.from_dict(vision_config),

106 + "text_config": PreTrainedConfig.from_dict(config),

107 + "vision_config": PreTrainedConfig.from_dict(vision_config),

108 }

109 if quant_config:

110 config["quantization_config"] = quant_config

@@ -248,7 +248,7 @@ def _remap_mistral_audio_args(config: dict) -> dict:

248 config = {

249 "model_type": "voxtral",

250 "architectures": [architecture],

251 - "text_config": PretrainedConfig.from_dict(config),

251 + "text_config": PreTrainedConfig.from_dict(config),

252 "audio_config": WhisperConfig(

253 num_mel_bins=encoder_args["audio_encoding_args"]["num_mel_bins"],

254 window_size=encoder_args["audio_encoding_args"]["window_size"],

vllm/transformers_utils/configs/mlp_speculator.py

+2 -2

@@ -2,10 +2,10 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4

5 -from transformers import PretrainedConfig

5 +from transformers import PreTrainedConfig

6

7

8 -class MLPSpeculatorConfig(PretrainedConfig):

8 +class MLPSpeculatorConfig(PreTrainedConfig):

9 model_type = "mlp_speculator"

10

11 attribute_map = {

vllm/transformers_utils/configs/moondream3.py

+4 -4

@@ -2,10 +2,10 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 """Configuration for Moondream3 model."""

4

5 -from transformers import PretrainedConfig

5 +from transformers import PreTrainedConfig

6

7

8 -class Moondream3VisionConfig(PretrainedConfig):

8 +class Moondream3VisionConfig(PreTrainedConfig):

9 """Vision encoder configuration for Moondream3."""

10

11 model_type = "moondream3_vision"

@@ -43,7 +43,7 @@ class Moondream3VisionConfig(PretrainedConfig):

43 self.image_size = crop_size

44

45

46 -class Moondream3TextConfig(PretrainedConfig):

46 +class Moondream3TextConfig(PreTrainedConfig):

47 """Text decoder configuration for Moondream3."""

48

49 model_type = "moondream3_text"

@@ -101,7 +101,7 @@ class Moondream3TextConfig(PretrainedConfig):

101 self.num_experts_per_tok = self.moe_experts_per_token

102

103

104 -class Moondream3Config(PretrainedConfig):

104 +class Moondream3Config(PreTrainedConfig):

105 """Combined configuration for Moondream3 multimodal model."""

106

107 model_type = "moondream3"

vllm/transformers_utils/configs/moonvit.py

+2 -2

@@ -1,10 +1,10 @@

1 # SPDX-License-Identifier: Apache-2.0

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 # Adapted from https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/blob/main/configuration_kimi_vl.py

4 -from transformers.configuration_utils import PretrainedConfig

4 +from transformers.configuration_utils import PreTrainedConfig

5

6

7 -class MoonViTConfig(PretrainedConfig):

7 +class MoonViTConfig(PreTrainedConfig):

8 model_type = "moonvit"

9

10 def __init__(

vllm/transformers_utils/configs/moss_transcribe_diarize.py

+2 -2

@@ -3,11 +3,11 @@

3

4 from typing import Any

5

6 -from transformers import PretrainedConfig, Qwen3Config

6 +from transformers import PreTrainedConfig, Qwen3Config

7 from transformers.models.whisper.configuration_whisper import WhisperConfig

8

9

10 -class MossTranscribeDiarizeConfig(PretrainedConfig):

10 +class MossTranscribeDiarizeConfig(PreTrainedConfig):

11 """Configuration for MOSS-Transcribe-Diarize."""

12

13 model_type = "moss_transcribe_diarize"

vllm/transformers_utils/configs/muse_glimmer.py

+4 -4

@@ -15,7 +15,7 @@ and vision configs are consumed by the native multimodal serving path.

15 from __future__ import annotations

16

17 from transformers import Qwen3Config

18 -from transformers.configuration_utils import PretrainedConfig

18 +from transformers.configuration_utils import PreTrainedConfig

19

20

21 def _default_no_rope_layers(num_hidden_layers: int) -> list[int]:

@@ -27,7 +27,7 @@ def _default_no_rope_layers(num_hidden_layers: int) -> list[int]:

27 ]

28

29

30 -class MuseGlimmerTextConfig(PretrainedConfig):

30 +class MuseGlimmerTextConfig(PreTrainedConfig):

31 model_type = "muse_glimmer_text"

32 keys_to_ignore_at_inference = ["past_key_values"]

33

@@ -127,7 +127,7 @@ class MuseGlimmerTextConfig(PretrainedConfig):

127 )

128

129

130 -class MuseGlimmerVisionConfig(PretrainedConfig):

130 +class MuseGlimmerVisionConfig(PreTrainedConfig):

131 model_type = "muse_glimmer_vision"

132

133 def __init__(

@@ -177,7 +177,7 @@ class MuseGlimmerVisionConfig(PretrainedConfig):

177 self.layer_types = layer_types

178

179

180 -class MuseGlimmerConfig(PretrainedConfig):

180 +class MuseGlimmerConfig(PreTrainedConfig):

181 model_type = "muse_glimmer"

182 sub_configs = {

183 "text_config": MuseGlimmerTextConfig,

vllm/transformers_utils/configs/nemotron.py

+4 -4

@@ -17,22 +17,22 @@

17 # limitations under the License.

18 """Nemotron model configuration."""

19

20 -from transformers import PretrainedConfig

20 +from transformers import PreTrainedConfig

21 from transformers.utils import logging

22

23 logger = logging.get_logger(__name__)

24

25

26 -class NemotronConfig(PretrainedConfig):

26 +class NemotronConfig(PreTrainedConfig):

27 r"""This is the configuration class to store the configuration of a

28 [`NemotronModel`]. It is used to instantiate a Nemotron model

29 according to the specified arguments, defining the model architecture.

30 Instantiating a configuration with the defaults will yield a similar

31 configuration to that of the Nemotron-8B.

32

33 - Configuration objects inherit from [`PretrainedConfig`] and can be

33 + Configuration objects inherit from [`PreTrainedConfig`] and can be

34 used to control the model outputs. Read the documentation from

35 - [`PretrainedConfig`] for more information.

35 + [`PreTrainedConfig`] for more information.

36

37

38 Args:

vllm/transformers_utils/configs/nemotron_h.py

+2 -2

@@ -18,13 +18,13 @@

18 """NemotronH model configuration."""

19

20 import regex as re

21 -from transformers.configuration_utils import PretrainedConfig

21 +from transformers.configuration_utils import PreTrainedConfig

22 from transformers.utils import logging

23

24 logger = logging.get_logger(__name__)

25

26

27 -class NemotronHConfig(PretrainedConfig):

27 +class NemotronHConfig(PreTrainedConfig):

28 r"""This is the configuration class to store the configuration of a

29 [`NemotronHModel`]. It is used to instantiate a NemotronH model according

30 to the specified arguments, defining the model architecture. Instantiating

vllm/transformers_utils/configs/olmo_hybrid.py

+2 -2

@@ -2,10 +2,10 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3

4

5 -from transformers.configuration_utils import PretrainedConfig

5 +from transformers.configuration_utils import PreTrainedConfig

6

7

8 -class OlmoHybridConfig(PretrainedConfig):

8 +class OlmoHybridConfig(PreTrainedConfig):

9 r"""Configuration class for [`OlmoHybridModel`]. It is used to

10 instantiate an OLMo Hybrid model according to the specified

11 arguments, defining the model architecture. Instantiating a

vllm/transformers_utils/configs/openvla.py

+2 -2

@@ -9,10 +9,10 @@ configuration without executing Hugging Face remote code.

9

10 from typing import Any

11

12 -from transformers import LlamaConfig, PretrainedConfig

12 +from transformers import LlamaConfig, PreTrainedConfig

13

14

15 -class OpenVLAConfig(PretrainedConfig):

15 +class OpenVLAConfig(PreTrainedConfig):

16 """Configuration class for OpenVLA models."""

17

18 model_type = "openvla"

vllm/transformers_utils/configs/ovis.py

+17 -17

@@ -7,10 +7,10 @@

7 # Ovis Config with AimV2 config registration removed for Transformers compatibility

8 from typing import Any

9

10 -from transformers import AutoConfig, PretrainedConfig

10 +from transformers import AutoConfig, PreTrainedConfig

11

12

13 -class AIMv2Config(PretrainedConfig):

13 +class AIMv2Config(PreTrainedConfig):

14 """This is the configuration class to store the configuration of an [`AIMv2Model`].

15 Instantiating a configuration with the defaults will yield a similar configuration

16 to that of the [apple/aimv2-large-patch14-224](https://huggingface.co/apple/aimv2-large-patch14-224).

@@ -29,7 +29,7 @@ class AIMv2Config(PretrainedConfig):

29 projection_dropout: Dropout ratio for the projection layer after the attention.

30 qkv_bias: Whether to add a bias to the queries, keys and values.

31 use_bias: Whether to add a bias in the feed-forward and projection layers.

32 - kwargs: Keyword arguments for the [`PretrainedConfig`].

32 + kwargs: Keyword arguments for the [`PreTrainedConfig`].

33

34 """

35

@@ -70,7 +70,7 @@ class AIMv2Config(PretrainedConfig):

70 # ----------------------------------------------------------------------

71 # Visual Tokenizer Configuration

72 # ----------------------------------------------------------------------

73 -class BaseVisualTokenizerConfig(PretrainedConfig):

73 +class BaseVisualTokenizerConfig(PreTrainedConfig):

74 def __init__(

75 self,

76 vocab_size=16384,

@@ -78,7 +78,7 @@ class BaseVisualTokenizerConfig(PretrainedConfig):

78 tau=1.0,

79 depths=None,

80 drop_cls_token=False,

81 - backbone_config: PretrainedConfig | dict | None = None,

81 + backbone_config: PreTrainedConfig | dict | None = None,

82 hidden_stride: int = 1,

83 **kwargs,

84 ):

@@ -92,10 +92,10 @@ class BaseVisualTokenizerConfig(PretrainedConfig):

92 self.backbone_kwargs = dict[str, Any]()

93 self.drop_cls_token = drop_cls_token

94 if backbone_config is not None:

95 - assert isinstance(backbone_config, (PretrainedConfig, dict)), (

96 - f"expect `backbone_config` to be instance of PretrainedConfig or dict, but got {type(backbone_config)} type"

95 + assert isinstance(backbone_config, (PreTrainedConfig, dict)), (

96 + f"expect `backbone_config` to be instance of PreTrainedConfig or dict, but got {type(backbone_config)} type"

97 )

98 - if not isinstance(backbone_config, PretrainedConfig):

98 + if not isinstance(backbone_config, PreTrainedConfig):

99 model_type = backbone_config["model_type"]

100 if model_type != "aimv2":

101 backbone_config.pop("model_type")

@@ -139,13 +139,13 @@ AutoConfig.register("aimv2_visual_tokenizer", Aimv2VisualTokenizerConfig)

139 # ----------------------------------------------------------------------

140 # Ovis Configuration

141 # ----------------------------------------------------------------------

142 -class OvisConfig(PretrainedConfig):

142 +class OvisConfig(PreTrainedConfig):

143 model_type = "ovis"

144

145 def __init__(

146 self,

147 - llm_config: PretrainedConfig | dict | None = None,

148 - visual_tokenizer_config: PretrainedConfig | dict | None = None,

147 + llm_config: PreTrainedConfig | dict | None = None,

148 + visual_tokenizer_config: PreTrainedConfig | dict | None = None,

149 multimodal_max_length=8192,

150 hidden_size=None,

151 conversation_formatter_class=None,

@@ -155,10 +155,10 @@ class OvisConfig(PretrainedConfig):

155 ):

156 super().__init__(**kwargs)

157 if llm_config is not None:

158 - assert isinstance(llm_config, (PretrainedConfig, dict)), (

159 - f"expect `llm_config` to be instance of PretrainedConfig or dict, but got {type(llm_config)} type"

158 + assert isinstance(llm_config, (PreTrainedConfig, dict)), (

159 + f"expect `llm_config` to be instance of PreTrainedConfig or dict, but got {type(llm_config)} type"

160 )

161 - if not isinstance(llm_config, PretrainedConfig):

161 + if not isinstance(llm_config, PreTrainedConfig):

162 model_type = llm_config["model_type"]

163 llm_config.pop("model_type")

164 llm_config = AutoConfig.for_model(model_type, **llm_config)

@@ -166,10 +166,10 @@ class OvisConfig(PretrainedConfig):

166 # map llm_config to text_config

167 self.text_config = llm_config

168 if visual_tokenizer_config is not None:

169 - assert isinstance(visual_tokenizer_config, (PretrainedConfig, dict)), (

170 - f"expect `visual_tokenizer_config` to be instance of PretrainedConfig or dict, but got {type(visual_tokenizer_config)} type"

169 + assert isinstance(visual_tokenizer_config, (PreTrainedConfig, dict)), (

170 + f"expect `visual_tokenizer_config` to be instance of PreTrainedConfig or dict, but got {type(visual_tokenizer_config)} type"

171 )

172 - if not isinstance(visual_tokenizer_config, PretrainedConfig):

172 + if not isinstance(visual_tokenizer_config, PreTrainedConfig):

173 model_type = visual_tokenizer_config["model_type"]

174 visual_tokenizer_config.pop("model_type")

175 visual_tokenizer_config = AutoConfig.for_model(

vllm/transformers_utils/configs/parakeet.py

+5 -5

@@ -2,7 +2,7 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 from dataclasses import dataclass

4

5 -from transformers import ParakeetEncoderConfig, PretrainedConfig

5 +from transformers import ParakeetEncoderConfig, PreTrainedConfig

6

7

8 class ParakeetConfig(ParakeetEncoderConfig):

@@ -24,9 +24,9 @@ class ParakeetConfig(ParakeetEncoderConfig):

24

25 @staticmethod

26 def from_hf_config(

27 - config: PretrainedConfig, *, llm_hidden_size: int, max_model_len: int

27 + config: PreTrainedConfig, *, llm_hidden_size: int, max_model_len: int

28 ) -> "ParakeetConfig":

29 - assert isinstance(config, PretrainedConfig)

29 + assert isinstance(config, PreTrainedConfig)

30 return ParakeetConfig(

31 **config.to_dict(),

32 scale_input=False,

@@ -55,8 +55,8 @@ class ExtractorConfig:

55 padding_value: float = 0.0

56

57 @classmethod

58 - def from_hf_config(cls, config: PretrainedConfig) -> "ExtractorConfig":

59 - assert isinstance(config, PretrainedConfig)

58 + def from_hf_config(cls, config: PreTrainedConfig) -> "ExtractorConfig":

59 + assert isinstance(config, PreTrainedConfig)

60 defaults = ("hop_length", "win_length", "preemphasis", "n_fft", "padding_value")

61 optional_kwargs = {

62 name: getattr(config, name) for name in defaults if hasattr(config, name)

vllm/transformers_utils/configs/qianfan_ocr.py

+3 -3

@@ -3,11 +3,11 @@

3

4 from typing import Any

5

6 -from transformers import PretrainedConfig

6 +from transformers import PreTrainedConfig

7 from transformers.models.auto import CONFIG_MAPPING

8

9

10 -class QianfanOCRVisionConfig(PretrainedConfig):

10 +class QianfanOCRVisionConfig(PreTrainedConfig):

11 model_type = "qianfan_ocr_vision"

12

13 def __init__(

@@ -53,7 +53,7 @@ class QianfanOCRVisionConfig(PretrainedConfig):

53 self.use_mean_pooling = use_mean_pooling

54

55

56 -class QianfanOCRConfig(PretrainedConfig):

56 +class QianfanOCRConfig(PreTrainedConfig):

57 model_type = "qianfan_ocr"

58

59 def __init__(

vllm/transformers_utils/configs/qwen3_5.py

+6 -6

@@ -16,10 +16,10 @@

16 # limitations under the License.

17 """Qwen3.5 model configuration."""

18

19 -from transformers.configuration_utils import PretrainedConfig

19 +from transformers.configuration_utils import PreTrainedConfig

20

21

22 -class Qwen3_5TextConfig(PretrainedConfig):

22 +class Qwen3_5TextConfig(PreTrainedConfig):

23 model_type = "qwen3_5_text"

24 keys_to_ignore_at_inference = ["past_key_values"]

25

@@ -108,7 +108,7 @@ class Qwen3_5TextConfig(PretrainedConfig):

108 self.linear_num_value_heads = linear_num_value_heads

109 super().__init__(**kwargs)

110 # Set these AFTER super().__init__() because transformers v4's

111 - # PretrainedConfig.__init__ has these as explicit params with different

111 + # PreTrainedConfig.__init__ has these as explicit params with different

112 # defaults (e.g. tie_word_embeddings=True) that would overwrite our values.

113 self.pad_token_id = pad_token_id

114 self.bos_token_id = bos_token_id

@@ -116,7 +116,7 @@ class Qwen3_5TextConfig(PretrainedConfig):

116 self.tie_word_embeddings = tie_word_embeddings

117

118

119 -class Qwen3_5VisionConfig(PretrainedConfig):

119 +class Qwen3_5VisionConfig(PreTrainedConfig):

120 model_type = "qwen3_5"

121 base_config_key = "vision_config"

122

@@ -152,7 +152,7 @@ class Qwen3_5VisionConfig(PretrainedConfig):

152 self.initializer_range = initializer_range

153

154

155 -class Qwen3_5Config(PretrainedConfig):

155 +class Qwen3_5Config(PreTrainedConfig):

156 model_type = "qwen3_5"

157 sub_configs = {

158 "vision_config": Qwen3_5VisionConfig,

@@ -186,7 +186,7 @@ class Qwen3_5Config(PretrainedConfig):

186 self.vision_start_token_id = vision_start_token_id

187 self.vision_end_token_id = vision_end_token_id

188 super().__init__(**kwargs)

189 - # Set after super().__init__() to avoid v4 PretrainedConfig overwrite

189 + # Set after super().__init__() to avoid v4 PreTrainedConfig overwrite

190 self.tie_word_embeddings = tie_word_embeddings

191

192

vllm/transformers_utils/configs/qwen3_5_moe.py

+6 -6

@@ -16,10 +16,10 @@

16 # limitations under the License.

17 """Qwen3.5-MoE model configuration."""

18

19 -from transformers.configuration_utils import PretrainedConfig

19 +from transformers.configuration_utils import PreTrainedConfig

20

21

22 -class Qwen3_5MoeTextConfig(PretrainedConfig):

22 +class Qwen3_5MoeTextConfig(PreTrainedConfig):

23 model_type = "qwen3_5_moe_text"

24 keys_to_ignore_at_inference = ["past_key_values"]

25

@@ -120,7 +120,7 @@ class Qwen3_5MoeTextConfig(PretrainedConfig):

120 self.router_aux_loss_coef = router_aux_loss_coef

121 super().__init__(**kwargs)

122 # Set these AFTER super().__init__() because transformers v4's

123 - # PretrainedConfig.__init__ has these as explicit params with different

123 + # PreTrainedConfig.__init__ has these as explicit params with different

124 # defaults (e.g. tie_word_embeddings=True) that would overwrite our values.

125 self.pad_token_id = pad_token_id

126 self.bos_token_id = bos_token_id

@@ -128,7 +128,7 @@ class Qwen3_5MoeTextConfig(PretrainedConfig):

128 self.tie_word_embeddings = tie_word_embeddings

129

130

131 -class Qwen3_5MoeVisionConfig(PretrainedConfig):

131 +class Qwen3_5MoeVisionConfig(PreTrainedConfig):

132 model_type = "qwen3_5_moe"

133 base_config_key = "vision_config"

134

@@ -164,7 +164,7 @@ class Qwen3_5MoeVisionConfig(PretrainedConfig):

164 self.initializer_range = initializer_range

165

166

167 -class Qwen3_5MoeConfig(PretrainedConfig):

167 +class Qwen3_5MoeConfig(PreTrainedConfig):

168 model_type = "qwen3_5_moe"

169 sub_configs = {

170 "vision_config": Qwen3_5MoeVisionConfig,

@@ -198,7 +198,7 @@ class Qwen3_5MoeConfig(PretrainedConfig):

198 self.vision_start_token_id = vision_start_token_id

199 self.vision_end_token_id = vision_end_token_id

200 super().__init__(**kwargs)

201 - # Set after super().__init__() to avoid v4 PretrainedConfig overwrite

201 + # Set after super().__init__() to avoid v4 PreTrainedConfig overwrite

202 self.tie_word_embeddings = tie_word_embeddings

203

204

vllm/transformers_utils/configs/qwen3_asr.py

+14 -14

@@ -17,14 +17,14 @@

17 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

18 # See the License for the specific language governing permissions and

19 # limitations under the License.

20 -from transformers.configuration_utils import PretrainedConfig

20 +from transformers.configuration_utils import PreTrainedConfig

21 from transformers.modeling_rope_utils import rope_config_validation

22 from transformers.utils import logging

23

24 logger = logging.get_logger(__name__)

25

26

27 -class Qwen3ASRAudioEncoderConfig(PretrainedConfig):

27 +class Qwen3ASRAudioEncoderConfig(PreTrainedConfig):

28 r"""

29 This is the configuration class to store the configuration of a [`Qwen3ASRAudioEncoder`]. It is used to instantiate a

30 Qwen3-ASR audio encoder according to the specified arguments, defining the model architecture. Instantiating a

@@ -33,8 +33,8 @@ class Qwen3ASRAudioEncoderConfig(PretrainedConfig):

33

34 e.g. [Qwen/Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B)

35

36 - Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the

37 - documentation from [`PretrainedConfig`] for more information.

36 + Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the

37 + documentation from [`PreTrainedConfig`] for more information.

38

39 Args:

40 num_mel_bins (`int`, *optional*, defaults to 128):

@@ -130,15 +130,15 @@ class Qwen3ASRAudioEncoderConfig(PretrainedConfig):

130 self.downsample_hidden_size = downsample_hidden_size

131

132

133 -class Qwen3ASRTextConfig(PretrainedConfig):

133 +class Qwen3ASRTextConfig(PreTrainedConfig):

134 r"""

135 This is the configuration class to store the configuration of a [`Qwen3ASRTextModel`]. It is used to instantiate a

136 Qwen3-ASR model according to the specified arguments, defining the model architecture. Instantiating a configuration

137 with the defaults will yield a similar configuration to that of

138 Qwen3-ASR-1.7B [Qwen/Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B)

139

140 - Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the

141 - documentation from [`PretrainedConfig`] for more information.

140 + Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the

141 + documentation from [`PreTrainedConfig`] for more information.

142

143 Args:

144 vocab_size (`int`, *optional*, defaults to 151936):

@@ -285,7 +285,7 @@ class Qwen3ASRTextConfig(PretrainedConfig):

285 super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)

286

287

288 -class Qwen3ASRThinkerConfig(PretrainedConfig):

288 +class Qwen3ASRThinkerConfig(PreTrainedConfig):

289 r"""

290 This is the configuration class to store the configuration of a [`Qwen3ASRThinker`]. It is used to instantiate a

291 Qwen3-ASR-Thinker model according to the specified arguments, defining the model architecture. Instantiating a

@@ -294,8 +294,8 @@ class Qwen3ASRThinkerConfig(PretrainedConfig):

294

295 e.g. [Qwen/Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B)

296

297 - Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the

298 - documentation from [`PretrainedConfig`] for more information.

297 + Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the

298 + documentation from [`PreTrainedConfig`] for more information.

299

300 Args:

301 audio_config (`dict`, *optional*):

@@ -365,7 +365,7 @@ class Qwen3ASRThinkerConfig(PretrainedConfig):

365 self.audio_token_id = audio_token_id

366

367

368 -class Qwen3ASRConfig(PretrainedConfig):

368 +class Qwen3ASRConfig(PreTrainedConfig):

369 """

370 This is the configuration class to store the configuration of a [`Qwen3ASRForConditionalGeneration`]. It is used to instantiate a Qwen3ASR

371 model according to the specified sub-models configurations, defining the model architecture.

@@ -373,8 +373,8 @@ class Qwen3ASRConfig(PretrainedConfig):

373 Instantiating a configuration with the defaults will yield a similar configuration to that of the

374 [Qwen/Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B) architecture.

375

376 - Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the

377 - documentation from [`PretrainedConfig`] for more information.

376 + Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the

377 + documentation from [`PreTrainedConfig`] for more information.

378

379 Args:

380 thinker_config (`dict`, *optional*): Configuration of the underlying thinker sub-model.

@@ -437,7 +437,7 @@ class Qwen3ASRConfig(PretrainedConfig):

437 self.timestamp_segment_time = timestamp_segment_time

438 super().__init__(**kwargs)

439

440 - def get_text_config(self, decoder=False) -> "PretrainedConfig":

440 + def get_text_config(self, decoder=False) -> "PreTrainedConfig":

441 """

442 Returns the config that is meant to be used with text IO. On most models, it is the original config instance

443 itself. On specific composite models, it is under a set of valid names.

vllm/transformers_utils/configs/qwen3_next.py

+4 -4

@@ -16,20 +16,20 @@

16 # limitations under the License.

17 """Qwen3-Next model configuration."""

18

19 -from transformers.configuration_utils import PretrainedConfig

19 +from transformers.configuration_utils import PreTrainedConfig

20 from transformers.utils import logging

21

22 logger = logging.get_logger(__name__)

23

24

25 -class Qwen3NextConfig(PretrainedConfig):

25 +class Qwen3NextConfig(PreTrainedConfig):

26 r"""This is the configuration class to store the configuration of a [`Qwen3NextModel`]. It is used to instantiate a

27 Qwen3-Next model according to the specified arguments, defining the model architecture.

28 Instantiating a configuration with the defaults will yield a similar configuration to that of

29 Qwen3-Next-80B-A3B-Instruct [Qwen/Qwen3-Next-80B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct).

30

31 - Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the

32 - documentation from [`PretrainedConfig`] for more information.

31 + Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the

32 + documentation from [`PreTrainedConfig`] for more information.

33

34

35 Args:

vllm/transformers_utils/configs/radio.py

+2 -2

@@ -4,7 +4,7 @@

4

5 from typing import Any

6

7 -from transformers.configuration_utils import PretrainedConfig

7 +from transformers.configuration_utils import PreTrainedConfig

8 from transformers.utils import logging

9

10 logger = logging.get_logger(__name__)

@@ -20,7 +20,7 @@ OPENAI_CLIP_MEAN = (0.48145466, 0.4578275, 0.40821073)

20 OPENAI_CLIP_STD = (0.26862954, 0.26130258, 0.27577711)

21

22

23 -class RadioConfig(PretrainedConfig):

23 +class RadioConfig(PreTrainedConfig):

24 r"""This is the configuration class to store the configuration of a Radio

25 vision model. It is used to instantiate a Radio model according to the

26 specified arguments, defining the model architecture.

vllm/transformers_utils/configs/speculators/base.py

+3 -3

@@ -4,7 +4,7 @@ import os

4 from dataclasses import fields

5 from typing import Any

6

7 -from transformers import PretrainedConfig

7 +from transformers import PreTrainedConfig

8

9 from vllm.transformers_utils.configs.speculators.algos import (

10 SUPPORTED_SPECULATORS_TYPES,

@@ -12,13 +12,13 @@ from vllm.transformers_utils.configs.speculators.algos import (

12 from vllm.transformers_utils.utils import without_trust_remote_code

13

14

15 -class SpeculatorsConfig(PretrainedConfig):

15 +class SpeculatorsConfig(PreTrainedConfig):

16 model_type = "speculators"

17

18 def __init__(self, **kwargs):

19 # super().__init__ performs some validation before setting all kwargs as

20 # attributes, so we set them first to be safe

21 - pre_trained_config_fields = {f.name for f in fields(PretrainedConfig)}

21 + pre_trained_config_fields = {f.name for f in fields(PreTrainedConfig)}

22 super_kwargs = dict()

23 for key, value in kwargs.items():

24 if key == "model_type":

vllm/transformers_utils/configs/step3_vl.py

+4 -4

@@ -2,10 +2,10 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 from typing import Any

4

5 -from transformers.configuration_utils import PretrainedConfig

5 +from transformers.configuration_utils import PreTrainedConfig

6

7

8 -class Step3VisionEncoderConfig(PretrainedConfig):

8 +class Step3VisionEncoderConfig(PreTrainedConfig):

9 model_type = "step3_vision_encoder"

10

11 def __init__(

@@ -35,7 +35,7 @@ class Step3VisionEncoderConfig(PretrainedConfig):

35 super().__init__(**kwargs)

36

37

38 -class Step3TextConfig(PretrainedConfig):

38 +class Step3TextConfig(PreTrainedConfig):

39 model_type = "step3_text"

40 architectures = ["Step3TextForCausalLM"]

41

@@ -146,7 +146,7 @@ class Step3TextConfig(PretrainedConfig):

146 super().__init__(**kwargs)

147

148

149 -class Step3VLConfig(PretrainedConfig):

149 +class Step3VLConfig(PreTrainedConfig):

150 model_type = "step3_vl"

151

152 def __init__(

vllm/transformers_utils/configs/step3p5.py

+2 -2

@@ -2,10 +2,10 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 from typing import Any

4

5 -from transformers.configuration_utils import PretrainedConfig

5 +from transformers.configuration_utils import PreTrainedConfig

6

7

8 -class Step3p5Config(PretrainedConfig):

8 +class Step3p5Config(PreTrainedConfig):

9 model_type = "step3p5"

10

11 def __init__(

vllm/transformers_utils/configs/ultravox.py

+5 -5

@@ -7,14 +7,14 @@ from typing import Any

7 import transformers

8

9

10 -class UltravoxConfig(transformers.PretrainedConfig):

10 +class UltravoxConfig(transformers.PreTrainedConfig):

11 r"""This is the configuration class to store the configuration of a

12 [`UltravoxForConditionalGeneration`]. It is used to instantiate an

13 Ultravox model according to the specified arguments, defining the model

14 architecture.

15

16 - Configuration objects inherit from [`PretrainedConfig`] and can be used to

17 - control the model outputs. Read the documentation from [`PretrainedConfig`]

16 + Configuration objects inherit from [`PreTrainedConfig`] and can be used to

17 + control the model outputs. Read the documentation from [`PreTrainedConfig`]

18 for more information.

19

20 Args:

@@ -74,7 +74,7 @@ class UltravoxConfig(transformers.PretrainedConfig):

74 self.num_projector_layers = num_projector_layers

75

76 # N.B. May set the wrapped_model_config below.

77 - self.wrapped_model_config: transformers.PretrainedConfig

77 + self.wrapped_model_config: transformers.PreTrainedConfig

78 self.text_model_id = text_model_id

79 if text_model_id is None:

80 text_config = text_config or {}

@@ -113,7 +113,7 @@ class UltravoxConfig(transformers.PretrainedConfig):

113 return super().__setattr__(key, value)

114

115 @property

116 - def text_config(self) -> transformers.PretrainedConfig:

116 + def text_config(self) -> transformers.PreTrainedConfig:

117 # When Ultravox wraps a multi-modal model (e.g. Gemma), we instantiate

118 # the full model, but the text config is the text config of the inner

119 # model.

vllm/transformers_utils/model_arch_config_convertor.py

+15 -15

@@ -5,7 +5,7 @@ from typing import final

5

6 import torch

7 from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE

8 -from transformers import PretrainedConfig

8 +from transformers import PreTrainedConfig

9

10 from vllm import envs

11 from vllm.config.model_arch import (

@@ -26,8 +26,8 @@ logger = init_logger(__name__)

26 class ModelArchConfigConvertorBase:

27 def __init__(

28 self,

29 - hf_config: PretrainedConfig,

30 - hf_text_config: PretrainedConfig,

29 + hf_config: PreTrainedConfig,

30 + hf_text_config: PreTrainedConfig,

31 revision: str | None = None,

32 ):

33 self.hf_config = hf_config

@@ -36,7 +36,7 @@ class ModelArchConfigConvertorBase:

36

37 def get_per_layer_hf_configs(

38 self,

39 - ) -> list[tuple[PretrainedConfig, PretrainedConfig]] | None:

39 + ) -> list[tuple[PreTrainedConfig, PreTrainedConfig]] | None:

40 """`(hf_config, hf_text_config)` per layer, or `None` if homogeneous.

41

42 This is the only place that decides whether a checkpoint is heterogeneous

@@ -219,7 +219,7 @@ class ModelArchConfigConvertorBase:

219 @classmethod

220 def get_torch_dtype(

221 cls,

222 - hf_config: PretrainedConfig,

222 + hf_config: PreTrainedConfig,

223 model_id: str,

224 revision: str | None,

225 config_format: str | ConfigFormat,

@@ -257,7 +257,7 @@ class ModelArchConfigConvertorBase:

257

258 return config_dtype

259

260 - def _normalize_quantization_config(self, config: PretrainedConfig):

260 + def _normalize_quantization_config(self, config: PreTrainedConfig):

261 quant_cfg = getattr(config, "quantization_config", None)

262 if quant_cfg is None:

263 # compressed-tensors uses a "compression_config" key

@@ -585,8 +585,8 @@ class DeepSeekMTPModelArchConfigConvertor(ModelArchConfigConvertorBase):

585 class DeepseekV4ModelArchConfigConvertor(ModelArchConfigConvertorBase):

586 def __init__(

587 self,

588 - hf_config: PretrainedConfig,

589 - hf_text_config: PretrainedConfig,

588 + hf_config: PreTrainedConfig,

589 + hf_text_config: PreTrainedConfig,

590 revision: str | None = None,

591 ):

592 # DeepSeek-V4-Flash-Vision-Exp ships the same architectures/model_type

@@ -622,7 +622,7 @@ class MimoMTPModelArchConfigConvertor(ModelArchConfigConvertorBase):

622

623

624 def _strip_mimo_v2_attention_chunk_size(

625 - hf_config: PretrainedConfig, hf_text_config: PretrainedConfig

625 + hf_config: PreTrainedConfig, hf_text_config: PreTrainedConfig

626 ) -> None:

627 # MiMo-V2-Flash's config.json sets `attention_chunk_size=128` but the

628 # architecture does not actually use chunked local attention. Leaving it

@@ -635,8 +635,8 @@ def _strip_mimo_v2_attention_chunk_size(

635 class MimoV2ModelArchConfigConvertor(ModelArchConfigConvertorBase):

636 def __init__(

637 self,

638 - hf_config: PretrainedConfig,

639 - hf_text_config: PretrainedConfig,

638 + hf_config: PreTrainedConfig,

639 + hf_text_config: PreTrainedConfig,

640 revision: str | None = None,

641 ):

642 if getattr(hf_config, "vision_config", None):

@@ -648,8 +648,8 @@ class MimoV2ModelArchConfigConvertor(ModelArchConfigConvertorBase):

648 class MimoV2MTPModelArchConfigConvertor(ModelArchConfigConvertorBase):

649 def __init__(

650 self,

651 - hf_config: PretrainedConfig,

652 - hf_text_config: PretrainedConfig,

651 + hf_config: PreTrainedConfig,

652 + hf_text_config: PreTrainedConfig,

653 revision: str | None = None,

654 ):

655 super().__init__(hf_config, hf_text_config, revision)

@@ -742,7 +742,7 @@ class Gemma4ModelArchConfigConvertor(ModelArchConfigConvertorBase):

742

743 def get_per_layer_hf_configs(

744 self,

745 - ) -> list[tuple[PretrainedConfig, PretrainedConfig]] | None:

745 + ) -> list[tuple[PreTrainedConfig, PreTrainedConfig]] | None:

746 # Gemma4 uses a larger head dimension, and sometimes more KV heads, on its

747 # full attention layers than on its sliding ones. Transformers >= 5.15.0

748 # says so in the config; before that the values are flat attributes picked

@@ -763,7 +763,7 @@ class Gemma4ModelArchConfigConvertor(ModelArchConfigConvertorBase):

763

764

765 class MossAudioModelArchConfigConvertor(ModelArchConfigConvertorBase):

766 - def _language_config(self) -> PretrainedConfig:

766 + def _language_config(self) -> PreTrainedConfig:

767 return self.hf_config.language_config

768

769 def get_num_hidden_layers(self) -> int:

vllm/transformers_utils/processors/glm4v.py

+2 -2

@@ -4,12 +4,12 @@

4 # Adapted from

5 # https://github.com/zai-org/CogAgent

6 from transformers import PythonBackend

7 -from transformers.image_processing_utils_fast import BaseImageProcessorFast

7 +from transformers.image_processing_backends import TorchvisionBackend

8 from transformers.image_utils import PILImageResampling

9 from transformers.processing_utils import ProcessorMixin

10

11

12 -class GLM4VImageProcessorFast(BaseImageProcessorFast):

12 +class GLM4VImageProcessorFast(TorchvisionBackend):

13 """Port of https://huggingface.co/zai-org/glm-4v-9b/blob/main/tokenization_chatglm.py#L177

14 to HF Transformers.

15 """

vllm/transformers_utils/processors/glm5next.py

+5 -5

@@ -7,12 +7,12 @@ import math

7 import numpy as np

8 import torch

9 from torchvision.transforms.v2 import functional as tvF

10 -from transformers.image_processing_utils import BatchFeature

11 -from transformers.image_processing_utils_fast import (

12 - BaseImageProcessorFast,

10 +from transformers.image_processing_backends import (

11 + TorchvisionBackend,

12 group_images_by_shape,

13 reorder_images,

14 )

15 +from transformers.image_processing_utils import BatchFeature

16 from transformers.image_utils import (

17 OPENAI_CLIP_MEAN,

18 OPENAI_CLIP_STD,

@@ -327,7 +327,7 @@ class Glm5NextImageProcessorKwargs(ImagesKwargs, total=False): # type: ignore[c

327 max_image_tokens: int | None

328

329

330 -class Glm5NextImageProcessor(BaseImageProcessorFast):

330 +class Glm5NextImageProcessor(TorchvisionBackend):

331 """Fast torchvision image processor for GLM-5.3-Flash.

332

333 ``patch_expand_factor`` multiplies into the ``smart_resize`` spatial

@@ -759,7 +759,7 @@ class Glm5NextProcessor(ProcessorMixin):

759 attributes = ["image_processor", "tokenizer", "video_processor"]

760 image_processor_class = "AutoImageProcessor"

761 video_processor_class = "AutoVideoProcessor"

762 - tokenizer_class = ("PreTrainedTokenizer", "PreTrainedTokenizerFast")

762 + tokenizer_class = ("PythonBackend", "TokenizersBackend")

763

764 def __init__(

765 self,

vllm/transformers_utils/processors/minimax_m3.py

+3 -3

@@ -18,8 +18,8 @@ import math

18 import torch

19 from torchvision.transforms import InterpolationMode

20 from transformers import AutoTokenizer, BatchFeature

21 -from transformers.image_processing_utils_fast import (

22 - BaseImageProcessorFast,

21 +from transformers.image_processing_backends import (

22 + TorchvisionBackend,

23 group_images_by_shape,

24 reorder_images,

25 )

@@ -156,7 +156,7 @@ class MiniMaxM3VLImageProcessorKwargs(ImagesKwargs, total=False): # type: ignor

156 max_long_side_pixel: int

157

158

159 -class MiniMaxM3VLImageProcessor(BaseImageProcessorFast):

159 +class MiniMaxM3VLImageProcessor(TorchvisionBackend):

160 do_resize = True

161 resample = PILImageResampling.BICUBIC

162 # required by base-class validation, not used as the resize bound

vllm/transformers_utils/processors/muse_glimmer.py

+1 -1

@@ -437,7 +437,7 @@ class MuseGlimmerProcessor(ProcessorMixin):

437 attributes = ["image_processor", "video_processor", "tokenizer"]

438 image_processor_class = "AutoImageProcessor"

439 video_processor_class = "AutoVideoProcessor"

440 - tokenizer_class = "PreTrainedTokenizerFast"

440 + tokenizer_class = "TokenizersBackend"

441

442 def __init__(

443 self,

vllm/transformers_utils/processors/nano_nemotron_vl.py

+4 -4

@@ -20,7 +20,7 @@ import numpy.typing as npt

20 import regex as re

21 import torch

22 from PIL import Image

23 -from transformers import BatchFeature, PretrainedConfig, TensorType

23 +from transformers import BatchFeature, PreTrainedConfig, TensorType

24

25 from vllm.model_executor.models.parakeet import ParakeetExtractor

26 from vllm.multimodal.inputs import AudioItem

@@ -584,7 +584,7 @@ class BaseNanoNemotronVLProcessor(ABC):

584

585 def __init__(

586 self,

587 - config: PretrainedConfig,

587 + config: PreTrainedConfig,

588 tokenizer: HfTokenizer,

589 *args,

590 max_model_len: int,

@@ -623,7 +623,7 @@ class BaseNanoNemotronVLProcessor(ABC):

623 self.dtype: torch.dtype = getattr(config, "dtype", torch.float32)

624

625 @staticmethod

626 - def use_dynamic_resolution(config: PretrainedConfig) -> bool:

626 + def use_dynamic_resolution(config: PreTrainedConfig) -> bool:

627 return "min_num_patches" in config.vision_config.args

628

629 @property

@@ -771,7 +771,7 @@ class NanoNemotronVLProcessor(BaseNanoNemotronVLProcessor):

771

772 def __init__(

773 self,

774 - config: PretrainedConfig,

774 + config: PreTrainedConfig,

775 tokenizer: HfTokenizer,

776 *,

777 max_model_len: int,

[Chore] Use Transformers v5 names and drop redundant processor `use_fast` (#58550)