[Chore] Use Transformers v5 names and drop redundant processor use_fast (#58550)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Harry Mellor
2026-09-24 14:14:10 +00:00
committed by GitHub
co-authored by Claude Opus 5.5
parent 484c211fe8
commit d881812500
218 changed files with 869 additions and 813 deletions
+3 -3
View File
@@ -3,7 +3,7 @@
from typing import get_args
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.config.speculative import MTPModelTypes, SpeculativeConfig
from vllm.transformers_utils.model_arch_config_convertor import (
@@ -11,8 +11,8 @@ from vllm.transformers_utils.model_arch_config_convertor import (
)
def _bailing_config() -> PretrainedConfig:
config = PretrainedConfig(
def _bailing_config() -> PreTrainedConfig:
config = PreTrainedConfig(
architectures=["BailingMoeV2_5ForCausalLM"],
hidden_size=4096,
kv_lora_rank=512,
+8 -8
View File
@@ -10,7 +10,7 @@ from typing import cast
from unittest.mock import Mock
import pytest
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from transformers.models.gemma4.configuration_gemma4 import Gemma4TextConfig
from vllm.config import ModelConfig, ParallelConfig, SpeculativeConfig
@@ -125,7 +125,7 @@ def _assert_model_config_methods(
def test_head_size_falls_back_when_head_dim_is_zero():
"""Regression test for configs that materialize missing head_dim as 0."""
hf_config = PretrainedConfig(
hf_config = PreTrainedConfig(
model_type="deepseek_vl_v2",
hidden_size=1280,
num_attention_heads=10,
@@ -140,7 +140,7 @@ def test_head_size_falls_back_when_head_dim_is_zero():
def test_qk_rope_head_dim_recovery_uses_model_revision(monkeypatch: pytest.MonkeyPatch):
hf_config = PretrainedConfig(
hf_config = PreTrainedConfig(
model_type="deepseek_v2",
qk_rope_head_dim=128,
qk_nope_head_dim=128,
@@ -176,7 +176,7 @@ def test_qk_rope_head_dim_recovery_uses_model_revision(monkeypatch: pytest.Monke
],
)
def test_num_experts_per_tok_aliases(attribute: str):
hf_config = PretrainedConfig(**{attribute: 4})
hf_config = PreTrainedConfig(**{attribute: 4})
convertor = ModelArchConfigConvertorBase(hf_config, hf_config)
model_config = cast(
ModelConfig,
@@ -191,7 +191,7 @@ def test_num_experts_per_tok_aliases(attribute: str):
def test_num_experts_per_tok_none_is_normalized():
hf_config = PretrainedConfig(top_k_experts=None)
hf_config = PreTrainedConfig(top_k_experts=None)
convertor = ModelArchConfigConvertorBase(hf_config, hf_config)
assert convertor.get_num_experts_per_token() == 0
@@ -207,7 +207,7 @@ def test_legacy_modelopt_config_without_producer_is_normalized():
"modelopt_quant_config": {"quant_cfg": {}},
}
}
hf_config = PretrainedConfig(quantization_config=quantization_config)
hf_config = PreTrainedConfig(quantization_config=quantization_config)
convertor = ModelArchConfigConvertorBase(hf_config, hf_config)
@@ -370,7 +370,7 @@ def test_gemma4_uniform_head_dims_are_homogeneous():
assert arch[3] is arch
class _HeterogeneousConfig(PretrainedConfig):
class _HeterogeneousConfig(PreTrainedConfig):
"""A heterogeneous config with no convertor of its own.
Mirrors the parts vLLM uses: per-layer configs are shallow copies with the
@@ -384,7 +384,7 @@ class _HeterogeneousConfig(PretrainedConfig):
super().__init__(**kwargs)
@property
def per_layer_config(self) -> list[PretrainedConfig]:
def per_layer_config(self) -> list[PreTrainedConfig]:
layers = []
for i in range(self.num_hidden_layers):
layer = copy(self)
+2 -2
View File
@@ -6,7 +6,7 @@ from unittest.mock import MagicMock, patch
import pytest
import torch
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from tests.models.utils import build_model_context
from vllm.config.ec_transfer import ECRole, ECTransferConfig
@@ -136,7 +136,7 @@ def test_language_model_only_disables_via_supports_multimodal_inputs():
def test_convertor_clears_mm_prefix_when_multimodal_disabled():
hf_config = PretrainedConfig(
hf_config = PreTrainedConfig(
model_type="gemma3",
architectures=["Gemma3ForConditionalGeneration"],
)
@@ -15,20 +15,20 @@ import functools
from unittest.mock import MagicMock, patch
import pytest
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.config.parallel import ParallelConfig
from vllm.config.speculative import SpeculativeConfig
def _make_hf_config(**kwargs) -> PretrainedConfig:
def _make_hf_config(**kwargs) -> PreTrainedConfig:
defaults = dict(
architectures=["LlamaForCausalLM"],
model_type="llama",
num_hidden_layers=64,
)
defaults.update(kwargs)
return PretrainedConfig(**defaults)
return PreTrainedConfig(**defaults)
@pytest.mark.cpu_test
@@ -137,7 +137,7 @@ def test_callable_overrides_reach_the_draft_config():
"""A callable override (config-to-config transform) composes with the
architecture-mapping override and is applied to the draft config."""
def shrink(hf_config: PretrainedConfig) -> PretrainedConfig:
def shrink(hf_config: PreTrainedConfig) -> PreTrainedConfig:
hf_config.num_hidden_layers = 1
return hf_config
@@ -155,7 +155,7 @@ def test_arch_mapping_applies_before_callable_override():
observes (and may adjust) the post-mapping config."""
seen_architectures: list[str] = []
def record(hf_config: PretrainedConfig) -> PretrainedConfig:
def record(hf_config: PreTrainedConfig) -> PreTrainedConfig:
seen_architectures.append(hf_config.architectures[0])
return hf_config
@@ -202,7 +202,7 @@ def test_inkling_override_exposes_all_mtp_depths():
assert out.local_layer_ids == [0, 2, 4]
def _module_level_shrink(hf_config: PretrainedConfig) -> PretrainedConfig:
def _module_level_shrink(hf_config: PreTrainedConfig) -> PreTrainedConfig:
hf_config.num_hidden_layers = 1
return hf_config
@@ -17,7 +17,7 @@ import logging
from typing import Literal
import pytest
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.config.model import ModelConfig
from vllm.config.parallel import ParallelConfig
@@ -46,7 +46,7 @@ def _override_logged(caplog: pytest.LogCaptureFixture) -> bool:
@pytest.mark.cpu_test
def test_override_raises_smaller_value(vllm_caplog: pytest.LogCaptureFixture):
hf_config = PretrainedConfig(max_position_embeddings=2048)
hf_config = PreTrainedConfig(max_position_embeddings=2048)
SpeculativeConfig._maybe_override_draft_max_position_embeddings(
hf_config, target_max_model_len=8192
)
@@ -56,7 +56,7 @@ def test_override_raises_smaller_value(vllm_caplog: pytest.LogCaptureFixture):
@pytest.mark.cpu_test
def test_override_keeps_sufficient_value(vllm_caplog: pytest.LogCaptureFixture):
hf_config = PretrainedConfig(max_position_embeddings=8192)
hf_config = PreTrainedConfig(max_position_embeddings=8192)
SpeculativeConfig._maybe_override_draft_max_position_embeddings(
hf_config, target_max_model_len=8192
)
@@ -66,7 +66,7 @@ def test_override_keeps_sufficient_value(vllm_caplog: pytest.LogCaptureFixture):
@pytest.mark.cpu_test
def test_override_ignores_missing_attribute(vllm_caplog: pytest.LogCaptureFixture):
hf_config = PretrainedConfig()
hf_config = PreTrainedConfig()
hf_config.__dict__.pop("max_position_embeddings", None)
SpeculativeConfig._maybe_override_draft_max_position_embeddings(
hf_config, target_max_model_len=8192
@@ -7,7 +7,7 @@ from unittest.mock import MagicMock
import pytest
from tokenizers import Tokenizer, models, pre_tokenizers
from transformers import PreTrainedTokenizerFast
from transformers import TokenizersBackend
from vllm import PoolingParams
from vllm.entrypoints.pooling.base.io_processor import PoolingIOProcessor
@@ -70,7 +70,7 @@ def offline_processor_and_context():
backend = Tokenizer(models.WordLevel({"[UNK]": 0}, unk_token="[UNK]"))
backend.pre_tokenizer = pre_tokenizers.WhitespaceSplit()
proc = JinaRankingIOProcessor.__new__(JinaRankingIOProcessor)
proc.tokenizer = PreTrainedTokenizerFast(tokenizer_object=backend)
proc.tokenizer = TokenizersBackend(tokenizer_object=backend)
proc.model_config = SimpleNamespace(max_model_len=1024, is_encoder_decoder=False)
proc.renderer = SimpleNamespace(
default_cmpl_tok_params=TokenizeParams(max_total_tokens=1024)
@@ -9,7 +9,7 @@ from unittest.mock import patch
import numpy as np
import pytest
import torch
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.config import PoolerConfig, set_current_vllm_config
from vllm.model_executor.layers.pooler import PoolingParamsUpdate
@@ -315,7 +315,7 @@ def test_dispatch_seq_cls_honors_token_pooling_type(tok_pooling_type):
model_config=SimpleNamespace(
pooler_config=pooler_config,
head_dtype=None,
hf_config=PretrainedConfig(num_labels=3),
hf_config=PreTrainedConfig(num_labels=3),
),
)
classifier = torch.nn.Linear(4, 3)
+2 -2
View File
@@ -8,7 +8,7 @@ from unittest.mock import Mock, patch
import pytest
import torch
import torch.nn as nn
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.config import ParallelConfig
from vllm.multimodal.processing import InputProcessingContext
@@ -52,7 +52,7 @@ def print_input_ids(input_ids):
@pytest.fixture
def mock_qwen3_omni_config():
"""Create a mock Qwen3OmniMoeThinker config."""
config = Mock(spec=PretrainedConfig)
config = Mock(spec=PreTrainedConfig)
# Token IDs from https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct/blob/main/tokenizer_config.json
config.audio_token_id = 151675 # <|audio_pad|>
config.video_token_id = 151656 # <|video_pad|>
@@ -14,7 +14,7 @@ from types import SimpleNamespace
from typing import cast
import pytest
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.config import ModelConfig
from vllm.model_executor.models.config import (
@@ -23,7 +23,7 @@ from vllm.model_executor.models.config import (
)
def _model_config(hf_config: PretrainedConfig) -> ModelConfig:
def _model_config(hf_config: PreTrainedConfig) -> ModelConfig:
"""Minimal stand-in for ModelConfig; only hf_config is read."""
return cast(ModelConfig, SimpleNamespace(hf_config=hf_config))
@@ -42,7 +42,7 @@ def test_encoder_backbone_enables_bidirectional_attention():
EncoderOnlyAttention; JinaEmbeddingsV5Model then dispatches to the encoder
implementation.
"""
hf_config = PretrainedConfig(is_decoder=False)
hf_config = PreTrainedConfig(is_decoder=False)
JinaEmbeddingsV5ModelConfig.verify_and_update_model_config(_model_config(hf_config))
@@ -55,13 +55,13 @@ def test_supported_decoder_backbone_is_accepted():
`-small` omits `is_decoder` entirely, so an absent attribute has to be
treated as a decoder. The first assertion pins that assumption: if
PretrainedConfig ever gains an `is_decoder=False` default, this fails here
PreTrainedConfig ever gains an `is_decoder=False` default, this fails here
rather than silently rejecting a supported checkpoint.
"""
absent = PretrainedConfig()
absent = PreTrainedConfig()
assert not hasattr(absent, "is_decoder")
JinaEmbeddingsV5ModelConfig.verify_and_update_model_config(_model_config(absent))
JinaEmbeddingsV5ModelConfig.verify_and_update_model_config(
_model_config(PretrainedConfig(is_decoder=True))
_model_config(PreTrainedConfig(is_decoder=True))
)
@@ -22,19 +22,19 @@ from unittest.mock import MagicMock
import numpy as np
import pytest
import torch
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from tests.models.registry import HF_EXAMPLE_MODELS
from vllm.config.multimodal import MultiModalDummyOptions
class MockAudioFlamingo3Config(PretrainedConfig):
class MockAudioFlamingo3Config(PreTrainedConfig):
model_type = "audioflamingo3"
def __init__(self, **kwargs):
super().__init__(**kwargs)
self.audio_config = PretrainedConfig()
self.text_config = PretrainedConfig()
self.audio_config = PreTrainedConfig()
self.text_config = PreTrainedConfig()
class MockAudioFlamingo3Processor:
@@ -8,7 +8,7 @@ import pytest
import torch
from PIL import Image
from tokenizers import Tokenizer, decoders, models, pre_tokenizers
from transformers import PreTrainedTokenizerFast
from transformers import TokenizersBackend
from vllm.model_executor.models.bailing_moe_v3_vl import (
IMAGE_PLACEHOLDER,
@@ -34,7 +34,7 @@ def processor():
add_prefix_space=False, use_regex=False
)
backend.decoder = decoders.ByteLevel()
tokenizer = PreTrainedTokenizerFast(
tokenizer = TokenizersBackend(
tokenizer_object=backend,
additional_special_tokens=[
"<|vision_start|>",
@@ -6,7 +6,7 @@ from collections.abc import Mapping
import pytest
from PIL import Image
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.image import rescale_image_size
@@ -17,7 +17,7 @@ from ...utils import build_model_context
def _get_expected_num_patches(
config: PretrainedConfig,
config: PreTrainedConfig,
image: Image.Image,
num_imgs: int,
min_num: int,
@@ -6,7 +6,7 @@ from collections.abc import Mapping
import pytest
from PIL import Image
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.image import rescale_image_size
@@ -17,7 +17,7 @@ from ...utils import build_model_context
def _get_expected_num_patches(
config: PretrainedConfig,
config: PreTrainedConfig,
image: Image.Image,
num_imgs: int,
min_num: int,
@@ -6,7 +6,7 @@ from collections.abc import Mapping
import pytest
from PIL import Image
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.multimodal import MULTIMODAL_REGISTRY
from vllm.multimodal.image import rescale_image_size
@@ -17,7 +17,7 @@ from ...utils import build_model_context
def _get_expected_num_patches(
config: PretrainedConfig,
config: PreTrainedConfig,
image: Image.Image,
num_imgs: int,
min_num: int,
+2 -2
View File
@@ -7,7 +7,7 @@ from typing import Any, Literal
import pytest
from packaging.version import Version
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from transformers import __version__ as TRANSFORMERS_VERSION
from vllm.config.model import ModelDType, TokenizerMode
@@ -1233,7 +1233,7 @@ _MULTIMODAL_EXAMPLE_MODELS = {
# NemotronH layers are constructed via `hybrid_override_pattern`
use_original_num_layers=True,
hf_overrides={
"vision_config": PretrainedConfig(
"vision_config": PreTrainedConfig(
args={
"min_num_patches": 1,
"max_num_patches": 12,
+3 -3
View File
@@ -4,7 +4,7 @@
import pytest
import torch
from transformers import Gemma3Config, PretrainedConfig, Qwen2Config
from transformers import Gemma3Config, PreTrainedConfig, Qwen2Config
from vllm.model_executor.models.adapters import (
_create_pooling_model_cls,
@@ -266,14 +266,14 @@ def test_resolve_num_labels_defaults_when_undeclared():
config = _composite_config()
assert (
_resolve_num_labels(config, config.get_text_config())
== PretrainedConfig().num_labels
== PreTrainedConfig().num_labels
)
def test_resolve_num_labels_declared_on_outer_config():
"""Multimodal checkpoints keep id2label/problem_type on the top-level config."""
config = _composite_config(outer_labels=20)
assert config.get_text_config().num_labels == PretrainedConfig().num_labels
assert config.get_text_config().num_labels == PreTrainedConfig().num_labels
assert _resolve_num_labels(config, config.get_text_config()) == 20
+5 -5
View File
@@ -5,7 +5,7 @@
from typing import Any
import pytest
from transformers import AutoConfig, PretrainedConfig
from transformers import AutoConfig, PreTrainedConfig
from vllm.config.speculative import SpeculativeConfig
@@ -15,19 +15,19 @@ _CHECKPOINTS = {
}
def _mtp_config(model_type: str) -> PretrainedConfig:
def _mtp_config(model_type: str) -> PreTrainedConfig:
"""Create a top-level MTP configuration with mtp_num_hidden_layers."""
kwargs: dict[str, Any] = {
"model_type": model_type,
"architectures": ["SomeArch"],
"mtp_num_hidden_layers": 1,
}
return PretrainedConfig(**kwargs)
return PreTrainedConfig(**kwargs)
def _multimodal_wrapper_mtp_config(
model_type: str, mtp_layers: int = 1
) -> PretrainedConfig:
) -> PreTrainedConfig:
"""Download a multimodal wrapper config via AutoConfig.from_pretrained
and configure mtp_num_hidden_layers in text_config.
@@ -36,7 +36,7 @@ def _multimodal_wrapper_mtp_config(
- MoE (qwen3_5_moe): Qwen/Qwen3.6-35B-A3B
"""
repo = _CHECKPOINTS[model_type]
config: PretrainedConfig = AutoConfig.from_pretrained(repo)
config: PreTrainedConfig = AutoConfig.from_pretrained(repo)
text_config = config.get_text_config()
text_config.mtp_num_hidden_layers = mtp_layers
return config
+2 -2
View File
@@ -13,7 +13,7 @@ from typing import Any
import pytest
import torch
import torch.nn as nn
from transformers import AutoConfig, AutoModel, PretrainedConfig
from transformers import AutoConfig, AutoModel, PreTrainedConfig
from vllm.config import ModelConfig, VllmConfig
from vllm.model_executor.models.interfaces import SupportsMultiModal
@@ -524,7 +524,7 @@ def replace_vocab_embeddings(model, **config_kwargs):
stub = nn.Module()
stub.model = model
stub.config = PretrainedConfig(
stub.config = PreTrainedConfig(
vocab_size=VOCAB_SIZE, num_positions=NUM_POSITIONS, **config_kwargs
)
embeddings = Base._vocab_embeddings(stub)
+3 -3
View File
@@ -8,7 +8,7 @@ from typing import Any
import torch
import torch.nn.functional as F
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.config.model import AttnTypeStr, ModelConfig, ModelDType, RunnerOption
from vllm.config.pooler import SequencePoolingType, TokenPoolingType
@@ -446,12 +446,12 @@ def get_vllm_extra_kwargs(model_info: ModelInfo, vllm_extra_kwargs):
def dummy_hf_overrides(
hf_config: PretrainedConfig,
hf_config: PreTrainedConfig,
*,
model_arch: str = "",
exist_overrides: dict[str, Any] | None = None,
use_original_num_layers: bool = False,
) -> PretrainedConfig:
) -> PreTrainedConfig:
"""Dummy HF overrides function used to create dummy model
with only minimum nums of layer.
"""
+10 -10
View File
@@ -62,16 +62,16 @@ def test_nested_rope_validation_patch_preserves_flat_rope_parameters(monkeypatch
def original_validate_rope(config, *args, **kwargs):
calls.append(config)
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
monkeypatch.setattr(PretrainedConfig, "validate_rope", original_validate_rope)
monkeypatch.setattr(PreTrainedConfig, "validate_rope", original_validate_rope)
_patch_hf_transformers_nested_rope_validation()
nested_rope_parameters = {
"full_attention": {"rope_type": "default"},
"original_max_position_embeddings": 32768,
}
PretrainedConfig.validate_rope(
PreTrainedConfig.validate_rope(
SimpleNamespace(rope_parameters=nested_rope_parameters)
)
assert nested_rope_parameters == {"full_attention": {"rope_type": "default"}}
@@ -81,7 +81,7 @@ def test_nested_rope_validation_patch_preserves_flat_rope_parameters(monkeypatch
"factor": 8.0,
"rope_theta": 500000.0,
}
PretrainedConfig.validate_rope(
PreTrainedConfig.validate_rope(
SimpleNamespace(rope_parameters=flat_rope_parameters)
)
assert flat_rope_parameters == {
@@ -730,9 +730,9 @@ def test_batch_invariant_breakable_cudagraph(
],
)
def test_dsa_models_select_matching_mtp(model_type, expected_architecture):
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
hf_config = PretrainedConfig(
hf_config = PreTrainedConfig(
architectures=["DeepseekV32ForCausalLM"],
num_nextn_predict_layers=1,
)
@@ -2463,14 +2463,14 @@ def test_get_and_verify_max_len_with_nope_layers(
max_model_len, rope_parameters, expected_max_len
):
"""NoPE layers do not prevent deriving or scaling the context length."""
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.config.model import _get_and_verify_max_len
from vllm.transformers_utils.model_arch_config_convertor import (
ModelArchConfigConvertorBase,
)
hf_config = PretrainedConfig(
hf_config = PreTrainedConfig(
max_position_embeddings=4096,
original_max_position_embeddings=2048,
)
@@ -2514,14 +2514,14 @@ def test_get_and_verify_max_len_yarn_is_already_scaled(
for every YaRN variant, so scaling it again overstates the limit and lets
requests past the end of the cos/sin cache.
"""
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.config.model import _get_and_verify_max_len
from vllm.transformers_utils.model_arch_config_convertor import (
ModelArchConfigConvertorBase,
)
hf_config = PretrainedConfig(max_position_embeddings=32768)
hf_config = PreTrainedConfig(max_position_embeddings=32768)
hf_config.rope_parameters = {
"rope_type": rope_type,
"factor": factor,
+6 -6
View File
@@ -11,7 +11,7 @@ from typing import cast
from unittest.mock import MagicMock, patch
import pytest
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.config.model import ModelConfig
from vllm.tokenizers import get_tokenizer
@@ -223,7 +223,7 @@ def test_model_config_generation_fallback_forwards_code_revision():
patch.object(
config_module,
"get_config",
return_value=PretrainedConfig(),
return_value=PreTrainedConfig(),
) as get_config,
):
ModelConfig.try_get_generation_config(model_config)
@@ -269,7 +269,7 @@ def test_safetensors_metadata_of_repo_without_safetensors():
],
)
def test_mrope_num_dims(section_key, mrope_section, expected_num_dims):
config = PretrainedConfig()
config = PreTrainedConfig()
config.rope_parameters = {"rope_type": "default", section_key: mrope_section}
assert uses_mrope(config)
@@ -280,7 +280,7 @@ def test_mrope_num_dims(section_key, mrope_section, expected_num_dims):
def test_mrope_num_dims_from_config_attribute(section_name):
"""Some configs expose the section as an attribute rather than under
`rope_parameters`."""
config = PretrainedConfig()
config = PreTrainedConfig()
setattr(config, section_name, [16, 16, 16, 16])
assert uses_mrope(config)
@@ -289,7 +289,7 @@ def test_mrope_num_dims_from_config_attribute(section_name):
def test_mrope_num_dims_from_nested_rope_parameters():
"""Sections nested by layer type must be found, not silently defaulted."""
config = PretrainedConfig()
config = PreTrainedConfig()
config.rope_parameters = {
"full_attention": {"mrope_section": [16, 16, 16, 16]},
"linear_attention": {"rope_type": "default"},
@@ -300,4 +300,4 @@ def test_mrope_num_dims_from_nested_rope_parameters():
def test_mrope_num_dims_without_mrope():
assert mrope_num_dims(PretrainedConfig()) == 0
assert mrope_num_dims(PreTrainedConfig()) == 0
@@ -4,7 +4,7 @@
from pathlib import Path
import pytest
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.transformers_utils.config import get_config_parser, register_config_parser
from vllm.transformers_utils.config_parser_base import ConfigParserBase
@@ -19,7 +19,7 @@ class CustomConfigParser(ConfigParserBase):
revision: str | None = None,
code_revision: str | None = None,
**kwargs,
) -> tuple[dict, PretrainedConfig]:
) -> tuple[dict, PreTrainedConfig]:
raise NotImplementedError
@@ -5,12 +5,12 @@
import json
import tempfile
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.transformers_utils.config import _CONFIG_REGISTRY, get_config
class _TestCustomConfig(PretrainedConfig):
class _TestCustomConfig(PreTrainedConfig):
model_type = "test_custom_model"
def __init__(self, custom_attr=42, **kw):
@@ -40,7 +40,7 @@ _SPECULATORS_CONFIG_DICT = {
def _patched_get_config_dict(config_dict):
"""Return a patch context that bypasses HF Hub for get_config_dict."""
return patch(
"vllm.transformers_utils.config.PretrainedConfig.get_config_dict",
"vllm.transformers_utils.config.PreTrainedConfig.get_config_dict",
return_value=(config_dict, {}),
)
@@ -7,7 +7,7 @@ from unittest import mock
import numpy as np
import pytest
import torch
from transformers import CLIPVisionConfig, LlamaConfig, LlavaConfig, PretrainedConfig
from transformers import CLIPVisionConfig, LlamaConfig, LlavaConfig, PreTrainedConfig
from tests.v1.attention.utils import (
BatchSpec,
@@ -336,7 +336,7 @@ def test_propose_different_layer_counts(num_hidden_layers):
# ---------------------------------------------------------------------------
class _DummyVLMConfig(PretrainedConfig):
class _DummyVLMConfig(PreTrainedConfig):
"""Minimal composite config that mimics VLMs like Kimi-K2.5 or LLaVA.
The text model's parameters (hidden_size, num_attention_heads, …) live
@@ -345,11 +345,11 @@ class _DummyVLMConfig(PretrainedConfig):
model_type = "test_vlm"
def __init__(self, text_config: PretrainedConfig, **kwargs):
def __init__(self, text_config: PreTrainedConfig, **kwargs):
self.text_config = text_config
super().__init__(architectures=["LlamaForCausalLM"], **kwargs)
def get_text_config(self, decoder: bool = False) -> PretrainedConfig:
def get_text_config(self, decoder: bool = False) -> PreTrainedConfig:
del decoder
return self.text_config
@@ -403,7 +403,7 @@ def test_extract_hidden_states_config_preserves_vlm_text_config():
eagle_aux_hidden_state_layer_ids=[1, 2],
)
# The fix: text_config is still a PretrainedConfig, not a dict.
# The fix: text_config is still a PreTrainedConfig, not a dict.
assert isinstance(extract_config.text_config, LlamaConfig)
extracted = get_hf_text_config(extract_config)
@@ -473,7 +473,7 @@ def test_extract_hidden_states_speculative_config_vlm():
def test_extract_hidden_states_config_invalid_text_config():
"""A nested text_config missing required attrs must still be rejected."""
broken_text_config = PretrainedConfig(hidden_size=128)
broken_text_config = PreTrainedConfig(hidden_size=128)
vlm_config = _DummyVLMConfig(text_config=broken_text_config)
extract_config = ExtractHiddenStatesConfig(