[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
+33 -33
View File
@@ -17,7 +17,7 @@ import transformers.configuration_utils as hf_configuration_utils
from huggingface_hub import constants
from packaging.version import Version
from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE
from transformers import GenerationConfig, PretrainedConfig
from transformers import GenerationConfig, PreTrainedConfig
from transformers.configuration_utils import ALLOWED_LAYER_TYPES
from transformers.models.auto.image_processing_auto import get_image_processor_config
from transformers.models.auto.modeling_auto import (
@@ -86,7 +86,7 @@ class LazyConfigDict(dict):
return getattr(configs, value)
_CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(
_CONFIG_REGISTRY: dict[str, type[PreTrainedConfig]] = LazyConfigDict(
afmoe="AfmoeConfig",
axk1="AXK1Config",
bagel="BagelConfig",
@@ -189,13 +189,13 @@ _AUTO_CONFIG_KWARGS_OVERRIDES: dict[str, dict[str, Any]] = {
def _register_config_class(
model_type: str, config_class: type[PretrainedConfig]
model_type: str, config_class: type[PreTrainedConfig]
) -> None:
config_class.model_type = model_type
AutoConfig.register(model_type, config_class, exist_ok=True)
def _maybe_register_hf_config(config: PretrainedConfig | None) -> None:
def _maybe_register_hf_config(config: PreTrainedConfig | None) -> None:
if config is None:
return
@@ -228,7 +228,7 @@ def _mistral_patch_hf_hub_constants() -> Iterator[None]:
def _install_hf_config_validator(
name: str, validator: Callable, supersedes: Callable
) -> None:
"""Replace a ``PretrainedConfig`` validator on every ``@strict`` snapshot.
"""Replace a ``PreTrainedConfig`` validator on every ``@strict`` snapshot.
``@strict`` snapshots each ``validate_*`` method into ``__class_validators__``
at class creation and automatic post-``__init__`` validation dispatches off
@@ -237,10 +237,10 @@ def _install_hf_config_validator(
class owns a snapshot, so rewrite them all, matching ``supersedes`` by
identity to leave a genuine per-model override in place.
"""
setattr(PretrainedConfig, name, validator)
setattr(PreTrainedConfig, name, validator)
seen: set[int] = set()
stack = [PretrainedConfig]
stack = [PreTrainedConfig]
while stack:
cls = stack.pop()
if id(cls) in seen:
@@ -261,10 +261,10 @@ def _patch_hf_transformers_validate_rope():
validate_rope() with the ignore_keys parameter work with newer versions of
hf transformers (from v5 onwards)
"""
if hasattr(PretrainedConfig.validate_rope, "__vllm_patched__"):
if hasattr(PreTrainedConfig.validate_rope, "__vllm_patched__"):
return
_original_validate_rope = PretrainedConfig.validate_rope
_original_validate_rope = PreTrainedConfig.validate_rope
@wraps(_original_validate_rope)
def patched_validate_rope(self, *args, **kwargs):
@@ -293,10 +293,10 @@ def _patch_hf_transformers_nested_rope_validation() -> None:
``AttributeError``. The per-layer dicts already carry their own defaults by
the time validation runs, so the shared entries can be dropped.
"""
if hasattr(PretrainedConfig.validate_rope, "__vllm_nested_rope_patched__"):
if hasattr(PreTrainedConfig.validate_rope, "__vllm_nested_rope_patched__"):
return
_original_validate_rope = PretrainedConfig.validate_rope
_original_validate_rope = PreTrainedConfig.validate_rope
@wraps(_original_validate_rope)
def patched_validate_rope(self, *args, **kwargs):
@@ -347,11 +347,11 @@ class HFConfigParser(ConfigParserBase):
revision: str | None = None,
code_revision: str | None = None,
**kwargs,
) -> tuple[dict, PretrainedConfig]:
) -> tuple[dict, PreTrainedConfig]:
kwargs["local_files_only"] = huggingface_hub.constants.HF_HUB_OFFLINE
trust_remote_code |= kwargs.get("trust_remote_code", False)
kwargs = without_trust_remote_code(kwargs)
config_dict, _ = PretrainedConfig.get_config_dict(
config_dict, _ = PreTrainedConfig.get_config_dict(
model,
revision=revision,
code_revision=code_revision,
@@ -374,7 +374,7 @@ class HFConfigParser(ConfigParserBase):
# through and remain unchanged by this elif block
dummy_model_type = f"dummy_{model_type}"
dummy_kwargs = dict(architectures=[""], model_type=dummy_model_type)
dummy_config = PretrainedConfig(**dummy_kwargs)
dummy_config = PreTrainedConfig(**dummy_kwargs)
dummy_model_type = hf_overrides(dummy_config).model_type
model_type = dummy_model_type.removeprefix("dummy_")
@@ -450,7 +450,7 @@ class MistralConfigParser(ConfigParserBase):
revision: str | None = None,
code_revision: str | None = None,
**kwargs,
) -> tuple[dict, PretrainedConfig]:
) -> tuple[dict, PreTrainedConfig]:
# This function loads a params.json config which
# should be used when loading models in mistral format
config_dict = _download_mistral_config_file(model, revision)
@@ -466,7 +466,7 @@ class MistralConfigParser(ConfigParserBase):
# Get missing fields from HF config if available
try:
hf_config_dict, _ = PretrainedConfig.get_config_dict(
hf_config_dict, _ = PreTrainedConfig.get_config_dict(
model,
revision=revision,
code_revision=code_revision,
@@ -540,7 +540,7 @@ def register_config_parser(config_format: str):
... revision: str | None = None,
... code_revision: str | None = None,
... **kwargs,
... ) -> tuple[dict, PretrainedConfig]:
... ) -> tuple[dict, PreTrainedConfig]:
... raise NotImplementedError
>>>
>>> type(get_config_parser("custom_config_parser"))
@@ -571,7 +571,7 @@ def register_config_parser(config_format: str):
return _wrapper
def set_default_rope_theta(config: PretrainedConfig, default_theta: float) -> None:
def set_default_rope_theta(config: PreTrainedConfig, default_theta: float) -> None:
"""Some models may have no rope_theta in their config but still use RoPE.
This function sets a default rope_theta if it's missing."""
if getattr(config, "rope_parameters", None) is None:
@@ -639,7 +639,7 @@ def patch_legacy_rope_type(rope_parameters: dict[str, Any] | None) -> None:
_patch_legacy_rope_type(rope_parameters)
def patch_rope_parameters(config: PretrainedConfig) -> None:
def patch_rope_parameters(config: PreTrainedConfig) -> None:
"""Provide backwards compatibility for RoPE."""
from vllm.config.utils import getattr_iter
@@ -661,7 +661,7 @@ def patch_rope_parameters(config: PretrainedConfig) -> None:
config.validate_rope()
def _iter_rope_parameters(config: PretrainedConfig) -> Iterator[dict[str, Any]]:
def _iter_rope_parameters(config: PreTrainedConfig) -> Iterator[dict[str, Any]]:
"""Yield a config's rope parameters, one dict per layer type if nested."""
rope_parameters = getattr(config, "rope_parameters", None)
if not isinstance(rope_parameters, dict):
@@ -673,7 +673,7 @@ def _iter_rope_parameters(config: PretrainedConfig) -> Iterator[dict[str, Any]]:
yield rope_parameters
def _mrope_section(config: PretrainedConfig) -> Sequence[int] | None:
def _mrope_section(config: PreTrainedConfig) -> Sequence[int] | None:
"""Return the M-RoPE section this config declares, if any.
`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:
return section if isinstance(section, (list, tuple)) else None
def _uses_mrope(config: PretrainedConfig) -> bool:
def _uses_mrope(config: PreTrainedConfig) -> bool:
return _mrope_section(config) is not None
def uses_mrope(config: PretrainedConfig) -> bool:
def uses_mrope(config: PreTrainedConfig) -> bool:
"""Detect if the model with this config uses M-ROPE."""
return (
_uses_mrope(config)
@@ -713,7 +713,7 @@ def uses_mrope(config: PretrainedConfig) -> bool:
)
def thinker_uses_mrope(config: PretrainedConfig) -> bool:
def thinker_uses_mrope(config: PreTrainedConfig) -> bool:
"""Detect if the model contains a thinker config and it uses M-ROPE."""
thinker_config = getattr(config, "thinker_config", None)
if thinker_config is None:
@@ -726,7 +726,7 @@ def thinker_uses_mrope(config: PretrainedConfig) -> bool:
return uses_mrope(thinker_text_config)
def mrope_num_dims(config: PretrainedConfig) -> int:
def mrope_num_dims(config: PreTrainedConfig) -> int:
"""Number of M-RoPE position channels the model consumes.
Each section entry sizes one channel, so the section length is the channel
@@ -745,10 +745,10 @@ def mrope_num_dims(config: PretrainedConfig) -> int:
return 3
def is_encoder_decoder(config: PretrainedConfig) -> bool:
def is_encoder_decoder(config: PreTrainedConfig) -> bool:
"""Detect if the model with this config is used as an encoder/decoder."""
def _is_encoder_decoder(config: PretrainedConfig) -> bool:
def _is_encoder_decoder(config: PreTrainedConfig) -> bool:
return getattr(config, "is_encoder_decoder", False)
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):
return kwargs
def _maybe_remap_hf_config_attrs(config: PretrainedConfig) -> PretrainedConfig:
def _maybe_remap_hf_config_attrs(config: PreTrainedConfig) -> PreTrainedConfig:
"""Remap config attributes to match the expected names."""
for old_attr, new_attr in _CONFIG_ATTRS_MAPPING.items():
if hasattr(config, old_attr):
@@ -798,7 +798,7 @@ def maybe_override_with_speculators(
"""
kwargs["local_files_only"] = huggingface_hub.constants.HF_HUB_OFFLINE
config_dict, _ = PretrainedConfig.get_config_dict(
config_dict, _ = PreTrainedConfig.get_config_dict(
model,
revision=revision,
token=hf_token,
@@ -840,9 +840,9 @@ def get_config(
code_revision: str | None = None,
config_format: str | ConfigFormat = "auto",
hf_overrides_kw: dict[str, Any] | None = None,
hf_overrides_fn: Callable[[PretrainedConfig], PretrainedConfig] | None = None,
hf_overrides_fn: Callable[[PreTrainedConfig], PreTrainedConfig] | None = None,
**kwargs,
) -> PretrainedConfig:
) -> PreTrainedConfig:
if config_format == "auto":
try:
# First check for Mistral to avoid defaulting to
@@ -953,7 +953,7 @@ def get_config(
# Exhaustively patch RoPE parameters everywhere they might be
patch_rope_parameters(config)
patch_rope_parameters(config.get_text_config())
SubConfigs: TypeAlias = dict[str, PretrainedConfig]
SubConfigs: TypeAlias = dict[str, PreTrainedConfig]
sub_configs: SubConfigs | None = getattr(config, "sub_configs", None)
if sub_configs:
for sub_config in sub_configs:
@@ -1212,7 +1212,7 @@ def get_hf_image_processor_config(
)
def get_hf_text_config(config: PretrainedConfig):
def get_hf_text_config(config: PreTrainedConfig):
"""Get the "sub" config relevant to llm for multi modal models.
No op for pure text models.
"""
@@ -4,7 +4,7 @@
from abc import ABC, abstractmethod
from pathlib import Path
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
class ConfigParserBase(ABC):
@@ -16,5 +16,5 @@ class ConfigParserBase(ABC):
revision: str | None = None,
code_revision: str | None = None,
**kwargs,
) -> tuple[dict, PretrainedConfig]:
) -> tuple[dict, PreTrainedConfig]:
raise NotImplementedError
+4 -4
View File
@@ -2,16 +2,16 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Any
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
class AXK1Config(PretrainedConfig):
class AXK1Config(PreTrainedConfig):
r"""This is the configuration class to store the configuration of a [`AXK1Model`].
It is used to instantiate an A.X model according to the specified arguments,
defining the model architecture. Instantiating a configuration with the defaults
will yield a similar configuration to that of the A.X K1.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control
the model outputs. Read the documentation from [`PretrainedConfig`] for more
Configuration objects inherit from [`PreTrainedConfig`] and can be used to control
the model outputs. Read the documentation from [`PreTrainedConfig`] for more
information.
Args:
+2 -2
View File
@@ -1,10 +1,10 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
class AfmoeConfig(PretrainedConfig):
class AfmoeConfig(PreTrainedConfig):
model_type = "afmoe"
def __init__(
+2 -2
View File
@@ -1,10 +1,10 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from transformers import PretrainedConfig, SiglipVisionConfig
from transformers import PreTrainedConfig, SiglipVisionConfig
from transformers.models.qwen2 import Qwen2Config
class BagelConfig(PretrainedConfig):
class BagelConfig(PreTrainedConfig):
"""Configuration class for BAGEL model."""
model_type = "bagel"
@@ -9,11 +9,11 @@ from __future__ import annotations
from typing import Any
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
def _normalize_mrope_parameters(
config: PretrainedConfig,
config: PreTrainedConfig,
mrope_section: list[int],
) -> None:
"""Add the top-level Bailing M-RoPE split to a text config."""
@@ -50,7 +50,7 @@ def _build_layer_types(
]
class BailingMoeV3TextConfig(PretrainedConfig):
class BailingMoeV3TextConfig(PreTrainedConfig):
model_type = "bailing_hybrid"
base_config_key = "text_config"
keys_to_ignore_at_inference = ["past_key_values"]
@@ -230,7 +230,7 @@ class BailingMoeV3TextConfig(PretrainedConfig):
self.validate_layer_type()
class BailingMoeV3VisionConfig(PretrainedConfig):
class BailingMoeV3VisionConfig(PreTrainedConfig):
model_type = "qwen3_moe_vit"
base_config_key = "vision_config"
@@ -273,7 +273,7 @@ class BailingMoeV3VisionConfig(PretrainedConfig):
self.initializer_range = initializer_range
class BailingMoeV3VLConfig(PretrainedConfig):
class BailingMoeV3VLConfig(PreTrainedConfig):
model_type = "bailing_moe_v3_vl"
keys_to_ignore_at_inference = ["past_key_values"]
sub_configs = {
+2 -2
View File
@@ -3,10 +3,10 @@
# Adapted from
# https://github.com/zai-org/ChatGLM2-6B
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
class ChatGLMConfig(PretrainedConfig):
class ChatGLMConfig(PreTrainedConfig):
model_type = "chatglm"
attribute_map = {
"num_hidden_layers": "num_layers",
@@ -8,10 +8,10 @@ with a pixel shuffle connector and ColBERT-style 128-dim per-token embeddings.
Reference: https://huggingface.co/ModernVBERT/colmodernvbert-merged
"""
from transformers import ModernBertConfig, PretrainedConfig, SiglipVisionConfig
from transformers import ModernBertConfig, PreTrainedConfig, SiglipVisionConfig
class ColModernVBertConfig(PretrainedConfig):
class ColModernVBertConfig(PreTrainedConfig):
model_type = "colmodernvbert"
def __init__(
@@ -1,7 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
from transformers.models.siglip2.configuration_siglip2 import Siglip2VisionConfig
from .nemotron_h import NemotronHConfig
@@ -50,7 +50,7 @@ class Cosmos3EdgeTextConfig(NemotronHConfig):
return config_dict
class Cosmos3EdgeProjectorConfig(PretrainedConfig):
class Cosmos3EdgeProjectorConfig(PreTrainedConfig):
model_type = "cosmos3_edge_projector"
def __init__(
@@ -74,7 +74,7 @@ class Cosmos3EdgeVisionConfig(Siglip2VisionConfig):
model_type = "cosmos3_edge_vision"
class Cosmos3EdgeConfig(PretrainedConfig):
class Cosmos3EdgeConfig(PreTrainedConfig):
model_type = "cosmos3_edge"
keys_to_ignore_at_inference = ["past_key_values"]
has_no_defaults_at_init = True
@@ -2,10 +2,10 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Any
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
class DeepseekV4Config(PretrainedConfig):
class DeepseekV4Config(PreTrainedConfig):
model_type = "deepseek_v4"
def __init__(
@@ -2,10 +2,10 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Any
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
class DeepseekV41Config(PretrainedConfig):
class DeepseekV41Config(PreTrainedConfig):
"""DeepSeek V4.1 config.
The HF config nests the text model under ``text_config`` and the vision
@@ -34,7 +34,7 @@ class DeepseekV41Config(PretrainedConfig):
for key, value in text_config.items():
if key == "model_type":
continue
# Don't clobber PretrainedConfig properties (e.g. is_encoder_decoder).
# Don't clobber PreTrainedConfig properties (e.g. is_encoder_decoder).
if isinstance(getattr(type(self), key, None), property):
continue
setattr(self, key, value)
@@ -4,10 +4,10 @@
# adapted from https://github.com/deepseek-ai/DeepSeek-VL2/blob/faf18023f24b962b32d9f0a2d89e402a8d383a78/deepseek_vl2/models/modeling_deepseek_vl_v2.py#L115-L268
from huggingface_hub.dataclasses import strict
from transformers import DeepseekV2Config, PretrainedConfig
from transformers import DeepseekV2Config, PreTrainedConfig
class VisionEncoderConfig(PretrainedConfig):
class VisionEncoderConfig(PreTrainedConfig):
model_type: str = "vision"
model_name: str = "vit_so400m_patch14_siglip_384.webli"
@@ -58,7 +58,7 @@ class VisionEncoderConfig(PretrainedConfig):
super().__init__(**kwargs)
class MlpProjectorConfig(PretrainedConfig):
class MlpProjectorConfig(PreTrainedConfig):
model_type = "mlp_projector"
projector_type: str = "downsample_mlp_gelu"
input_dim: int = 1152
@@ -98,7 +98,7 @@ class DeepseekVLV2TextConfig(DeepseekV2Config):
kv_lora_rank: int | None = DeepseekV2Config.kv_lora_rank
class DeepseekVLV2Config(PretrainedConfig):
class DeepseekVLV2Config(PreTrainedConfig):
model_type = "deepseek_vl_v2"
tile_tag: str = "2D"
@@ -2,12 +2,12 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Any
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from transformers.models.gemma4.configuration_gemma4 import Gemma4VisionConfig
def _init_text_config(self: PretrainedConfig, **kwargs: Any) -> None:
PretrainedConfig.__init__(self, **kwargs)
def _init_text_config(self: PreTrainedConfig, **kwargs: Any) -> None:
PreTrainedConfig.__init__(self, **kwargs)
# DiffusionGemma always uses MoE and K=V sharing for full_attention
# layers. The HF reference removed these config fields entirely.
if getattr(self, "num_experts", None):
@@ -15,14 +15,14 @@ def _init_text_config(self: PretrainedConfig, **kwargs: Any) -> None:
self.attention_k_eq_v = True
class DiffusionGemmaTextConfig(PretrainedConfig):
class DiffusionGemmaTextConfig(PreTrainedConfig):
model_type = "diffusion_gemma_text"
def __init__(self, **kwargs: Any):
_init_text_config(self, **kwargs)
class DiffusionGemmaConfig(PretrainedConfig):
class DiffusionGemmaConfig(PreTrainedConfig):
model_type = "diffusion_gemma"
def __init__(
@@ -41,4 +41,4 @@ class DiffusionGemmaConfig(PretrainedConfig):
else:
self.vision_config = vision_config
self.audio_config = None
PretrainedConfig.__init__(self, **kwargs)
PreTrainedConfig.__init__(self, **kwargs)
+2 -2
View File
@@ -2,11 +2,11 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Any
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
from transformers.models.qwen2 import Qwen2Config
class DotsVisionConfig(PretrainedConfig):
class DotsVisionConfig(PreTrainedConfig):
model_type: str = "dots_vit"
def __init__(
+4 -4
View File
@@ -3,22 +3,22 @@
import os
from transformers import AutoConfig, DeepseekV2Config, PretrainedConfig
from transformers import AutoConfig, DeepseekV2Config, PreTrainedConfig
from vllm.transformers_utils.utils import without_trust_remote_code
class EAGLEConfig(PretrainedConfig):
class EAGLEConfig(PreTrainedConfig):
model_type = "eagle"
def __init__(
self,
model: PretrainedConfig | dict | None = None,
model: PreTrainedConfig | dict | None = None,
truncated_vocab_size: int | None = None,
method: str | None = "eagle",
**kwargs,
):
model_config: PretrainedConfig | DeepseekV2Config | None
model_config: PreTrainedConfig | DeepseekV2Config | None
if isinstance(model, dict):
model_config = AutoConfig.for_model(**model)
else:
@@ -5,17 +5,17 @@ the extract_hidden_states spec decoding method."""
import os
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.transformers_utils.utils import without_trust_remote_code
class ExtractHiddenStatesConfig(PretrainedConfig):
class ExtractHiddenStatesConfig(PreTrainedConfig):
model_type = "extract_hidden_states"
def __init__(
self,
model: PretrainedConfig | dict | None = None,
model: PreTrainedConfig | dict | None = None,
method: str | None = "extract_hidden_states",
**kwargs,
):
@@ -24,7 +24,7 @@ class ExtractHiddenStatesConfig(PretrainedConfig):
if isinstance(model, dict):
model_dict = model
source_text_config = None
elif isinstance(model, PretrainedConfig):
elif isinstance(model, PreTrainedConfig):
model_dict = model.to_dict()
text_config = model.get_text_config()
source_text_config = text_config if text_config is not model else None
@@ -40,7 +40,7 @@ class ExtractHiddenStatesConfig(PretrainedConfig):
combined["architectures"] = ["ExtractHiddenStatesModel"]
# to_dict() and kwargs both flatten text_config to a plain dict;
# downstream get_hf_text_config() needs it as a PretrainedConfig
# downstream get_hf_text_config() needs it as a PreTrainedConfig
# for attribute access. Re-insert the original object.
if source_text_config is not None:
combined["text_config"] = source_text_config
+2 -2
View File
@@ -20,10 +20,10 @@
# limitations under the License.
"""Falcon configuration."""
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
class RWConfig(PretrainedConfig):
class RWConfig(PreTrainedConfig):
model_type = "falcon"
keys_to_ignore_at_inference = ["past_key_values"]
attribute_map = {
@@ -3,7 +3,7 @@
from __future__ import annotations
from transformers import CONFIG_MAPPING, PretrainedConfig
from transformers import CONFIG_MAPPING, PreTrainedConfig
# NOTE: Temporary shim for FunAudioChat checkpoints.
# These checkpoints use `model_type="funaudiochat"`, which is not currently
@@ -13,7 +13,7 @@ from transformers import CONFIG_MAPPING, PretrainedConfig
# provides an `auto_map`) and vLLM can rely on `AutoConfig.from_pretrained()`.
class FunAudioChatAudioEncoderConfig(PretrainedConfig):
class FunAudioChatAudioEncoderConfig(PreTrainedConfig):
model_type = "funaudiochat_audio_encoder"
def __init__(
@@ -74,7 +74,7 @@ class FunAudioChatAudioEncoderConfig(PretrainedConfig):
self.pad_token_id = pad_token_id
class FunAudioChatConfig(PretrainedConfig):
class FunAudioChatConfig(PreTrainedConfig):
model_type = "funaudiochat"
attribute_map = {
"audio_token_id": "audio_token_index",
@@ -82,8 +82,8 @@ class FunAudioChatConfig(PretrainedConfig):
def __init__(
self,
audio_config: PretrainedConfig | dict | None = None,
text_config: PretrainedConfig | dict | None = None,
audio_config: PreTrainedConfig | dict | None = None,
text_config: PreTrainedConfig | dict | None = None,
audio_token_index: int = 151646,
ignore_index: int = -100,
hidden_size: int | None = None,
+3 -3
View File
@@ -4,12 +4,12 @@
from copy import copy
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
def gemma4_layer_config(
text_config: PretrainedConfig, layer_idx: int
) -> PretrainedConfig:
text_config: PreTrainedConfig, layer_idx: int
) -> PreTrainedConfig:
"""The Gemma4 text config as it applies to one layer.
Gemma4 uses a larger head dimension on its full attention layers than on its
+6 -6
View File
@@ -1,14 +1,14 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
from vllm.logger import init_logger
logger = init_logger(__name__)
class Glm5NextTextConfig(PretrainedConfig):
class Glm5NextTextConfig(PreTrainedConfig):
model_type = "glm5_next_text"
base_config_key = "text_config"
keys_to_ignore_at_inference = ["past_key_values"]
@@ -271,7 +271,7 @@ class Glm5NextTextConfig(PretrainedConfig):
]
class Glm5NextVisionConfig(PretrainedConfig):
class Glm5NextVisionConfig(PreTrainedConfig):
model_type = "glm5_next_vision"
base_config_key = "vision_config"
@@ -324,7 +324,7 @@ class Glm5NextVisionConfig(PretrainedConfig):
self.swiglu_limit = swiglu_limit
class Glm5NextConfig(PretrainedConfig):
class Glm5NextConfig(PreTrainedConfig):
model_type = "glm5_next"
sub_configs = {
"vision_config": Glm5NextVisionConfig,
@@ -345,7 +345,7 @@ class Glm5NextConfig(PretrainedConfig):
**kwargs,
):
# Init super() first so base-class defaults don't clobber text-config
# values set below (PretrainedConfig has many text-related defaults
# values set below (PreTrainedConfig has many text-related defaults
# that differ from Glm5NextTextConfig).
super().__init__(**kwargs)
@@ -379,7 +379,7 @@ class Glm5NextConfig(PretrainedConfig):
# and must NOT be mirrored onto text_config: ``architectures`` /
# ``torch_dtype`` differ between the top-level config and the text
# sub-config, and mirroring them makes the top-level ``architectures``
# silently read back as None (PretrainedConfig initializes both to None),
# silently read back as None (PreTrainedConfig initializes both to None),
# which then fails model-class resolution ("No model architectures are
# specified").
_UNMIRRORED_KEYS = [
@@ -6,7 +6,7 @@ from typing import Any
import transformers
class Granite4VisionConfig(transformers.PretrainedConfig):
class Granite4VisionConfig(transformers.PreTrainedConfig):
"""Configuration for Granite 4 Vision model.
This config is needed because the granite4_vision model type is not yet
@@ -81,7 +81,7 @@ class Granite4VisionConfig(transformers.PretrainedConfig):
**vision_config
)
else:
self.vision_config = transformers.PretrainedConfig(**vision_config)
self.vision_config = transformers.PreTrainedConfig(**vision_config)
text_model_type = text_config.get("model_type", "granite")
if text_model_type in transformers.CONFIG_MAPPING:
@@ -89,6 +89,6 @@ class Granite4VisionConfig(transformers.PretrainedConfig):
**text_config
)
else:
self.text_config = transformers.PretrainedConfig(**text_config)
self.text_config = transformers.PreTrainedConfig(**text_config)
super().__init__(**kwargs)
+4 -4
View File
@@ -2,16 +2,16 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Any
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
class HYV3Config(PretrainedConfig):
class HYV3Config(PreTrainedConfig):
r"""This is the configuration class to store the configuration of a [`HYV3Model`].
It is used to instantiate a HYV3 model (HY V3 MoE language model) according to
the specified arguments.
Configuration objects inherit from [`PretrainedConfig`] and can be used to
control the model outputs. Read the documentation from [`PretrainedConfig`]
Configuration objects inherit from [`PreTrainedConfig`] and can be used to
control the model outputs. Read the documentation from [`PreTrainedConfig`]
for more information.
Args:
@@ -17,16 +17,16 @@
# limitations under the License.
"""HyperCLOVA X model configuration."""
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
class HyperCLOVAXConfig(PretrainedConfig):
class HyperCLOVAXConfig(PreTrainedConfig):
r"""This is the configuration class to store the configuration of a
[`HyperCLOVAXModel`]. It is used to instantiate a HyperCLOVAX model
according to the specified arguments, defining the model architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used
Configuration objects inherit from [`PreTrainedConfig`] and can be used
to control the model outputs. Read the documentation from
[`PretrainedConfig`] for more information.
[`PreTrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 32000):
+3 -3
View File
@@ -7,10 +7,10 @@ A video-chunk is the smallest independently processable unit of video.
"""
from transformers import DeepseekV3Config
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
class KimiK25VisionConfig(PretrainedConfig):
class KimiK25VisionConfig(PreTrainedConfig):
model_type = "kimi_k25_vision"
def __init__(
@@ -59,7 +59,7 @@ class KimiK25VisionConfig(PretrainedConfig):
self.projector_ln_eps = projector_ln_eps
class KimiK25Config(PretrainedConfig):
class KimiK25Config(PreTrainedConfig):
"""Kimi-K2.5 model configuration.
Kimi-K2.5 extends Kimi-K2 with vision support using video-chunks.
+3 -3
View File
@@ -2,7 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Kimi-K3 multimodal configuration."""
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
from vllm.logger import init_logger
from vllm.transformers_utils.configs.kimi_linear import KimiLinearConfig
@@ -10,7 +10,7 @@ from vllm.transformers_utils.configs.kimi_linear import KimiLinearConfig
logger = init_logger(__name__)
class KimiK3VisionConfig(PretrainedConfig):
class KimiK3VisionConfig(PreTrainedConfig):
model_type = "kimi_k3_vision"
def __init__(
@@ -83,7 +83,7 @@ class KimiK3VisionConfig(PretrainedConfig):
self.intermediate_size = vt_intermediate_size
class KimiK3Config(PretrainedConfig):
class KimiK3Config(PreTrainedConfig):
model_type = "kimi_k3"
def __init__(
@@ -1,10 +1,10 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
class KimiLinearConfig(PretrainedConfig):
class KimiLinearConfig(PreTrainedConfig):
model_type = "kimi_linear"
keys_to_ignore_at_inference = ["past_key_values"]
+2 -2
View File
@@ -3,12 +3,12 @@
# Adapted from https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/blob/main/configuration_kimi_vl.py
from transformers import DeepseekV2Config
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
from vllm.transformers_utils.configs.moonvit import MoonViTConfig
class KimiVLConfig(PretrainedConfig):
class KimiVLConfig(PreTrainedConfig):
model_type = "kimi_vl"
def __init__(
+4 -4
View File
@@ -2,17 +2,17 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Any
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
class Lfm2MoeConfig(PretrainedConfig):
class Lfm2MoeConfig(PreTrainedConfig):
r"""This is the configuration class to store the configuration of a [`Lfm2MoeModel`]. It is used to instantiate a LFM2 Moe
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
defaults will yield a similar configuration to that of the LFM2-8B-A1B model.
e.g. [LiquidAI/LFM2-8B-A1B](https://huggingface.co/LiquidAI/LFM2-8B-A1B)
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PreTrainedConfig`] for more information.
Args:
+2 -2
View File
@@ -3,12 +3,12 @@
import os
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.transformers_utils.utils import without_trust_remote_code
class MedusaConfig(PretrainedConfig):
class MedusaConfig(PreTrainedConfig):
model_type = "medusa"
def __init__(
@@ -22,13 +22,13 @@
# See the License for the specific language governing permissions and
# limitations under the License.
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from transformers.models.qwen2_5_omni.configuration_qwen2_5_omni import (
Qwen2_5OmniTextConfig,
)
class DashengConfig(PretrainedConfig):
class DashengConfig(PreTrainedConfig):
model_type = "midashenglm_dasheng_encoder"
def __init__(
@@ -80,7 +80,7 @@ class DashengConfig(PretrainedConfig):
super().__init__(**kwargs)
class MiDashengLMConfig(PretrainedConfig):
class MiDashengLMConfig(PreTrainedConfig):
model_type = "midashenglm"
def __init__(
@@ -2,10 +2,10 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
class Mimo_VLVisionConfig(PretrainedConfig):
class Mimo_VLVisionConfig(PreTrainedConfig):
model_type = "mimovl"
base_config_key = "vision_config"
@@ -2,10 +2,10 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Any
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
class MiniMaxM3TextConfig(PretrainedConfig):
class MiniMaxM3TextConfig(PreTrainedConfig):
"""Config for the MiniMax M3 text backbone (MiniMaxM3SparseForCausalLM).
Defaults mirror the ``text_config`` of the MiniMax-M3-preview checkpoint.
@@ -120,7 +120,7 @@ class MiniMaxM3MTPConfig(MiniMaxM3TextConfig):
super().__init__(num_hidden_layers=num_hidden_layers, **kwargs)
class MiniMaxM3Config(PretrainedConfig):
class MiniMaxM3Config(PreTrainedConfig):
"""Top-level MiniMax M3 (VL) config.
Holds the text backbone as ``text_config`` so that
+6 -6
View File
@@ -2,7 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Any
from transformers import PretrainedConfig, WhisperConfig
from transformers import PreTrainedConfig, WhisperConfig
from vllm.logger import init_logger
@@ -12,7 +12,7 @@ logger = init_logger(__name__)
def adapt_config_dict(
config_dict: dict[str, Any],
defaults: dict[str, Any],
) -> PretrainedConfig:
) -> PreTrainedConfig:
config_dict = _remap_general_mistral_args(config_dict)
config_dict = _remap_mistral_sliding_window(config_dict)
@@ -86,7 +86,7 @@ def adapt_config_dict(
for k, v in defaults.items():
config_dict.setdefault(k, v)
config = PretrainedConfig.from_dict(config_dict)
config = PreTrainedConfig.from_dict(config_dict)
logger.debug("Initialized config %s", config)
@@ -103,8 +103,8 @@ def _remap_mistral_vision_args(config: dict) -> dict:
config = {
"model_type": "pixtral",
"architectures": ["PixtralForConditionalGeneration"],
"text_config": PretrainedConfig.from_dict(config),
"vision_config": PretrainedConfig.from_dict(vision_config),
"text_config": PreTrainedConfig.from_dict(config),
"vision_config": PreTrainedConfig.from_dict(vision_config),
}
if quant_config:
config["quantization_config"] = quant_config
@@ -248,7 +248,7 @@ def _remap_mistral_audio_args(config: dict) -> dict:
config = {
"model_type": "voxtral",
"architectures": [architecture],
"text_config": PretrainedConfig.from_dict(config),
"text_config": PreTrainedConfig.from_dict(config),
"audio_config": WhisperConfig(
num_mel_bins=encoder_args["audio_encoding_args"]["num_mel_bins"],
window_size=encoder_args["audio_encoding_args"]["window_size"],
@@ -2,10 +2,10 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
class MLPSpeculatorConfig(PretrainedConfig):
class MLPSpeculatorConfig(PreTrainedConfig):
model_type = "mlp_speculator"
attribute_map = {
@@ -2,10 +2,10 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Configuration for Moondream3 model."""
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
class Moondream3VisionConfig(PretrainedConfig):
class Moondream3VisionConfig(PreTrainedConfig):
"""Vision encoder configuration for Moondream3."""
model_type = "moondream3_vision"
@@ -43,7 +43,7 @@ class Moondream3VisionConfig(PretrainedConfig):
self.image_size = crop_size
class Moondream3TextConfig(PretrainedConfig):
class Moondream3TextConfig(PreTrainedConfig):
"""Text decoder configuration for Moondream3."""
model_type = "moondream3_text"
@@ -101,7 +101,7 @@ class Moondream3TextConfig(PretrainedConfig):
self.num_experts_per_tok = self.moe_experts_per_token
class Moondream3Config(PretrainedConfig):
class Moondream3Config(PreTrainedConfig):
"""Combined configuration for Moondream3 multimodal model."""
model_type = "moondream3"
+2 -2
View File
@@ -1,10 +1,10 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# Adapted from https://huggingface.co/moonshotai/Kimi-VL-A3B-Instruct/blob/main/configuration_kimi_vl.py
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
class MoonViTConfig(PretrainedConfig):
class MoonViTConfig(PreTrainedConfig):
model_type = "moonvit"
def __init__(
@@ -3,11 +3,11 @@
from typing import Any
from transformers import PretrainedConfig, Qwen3Config
from transformers import PreTrainedConfig, Qwen3Config
from transformers.models.whisper.configuration_whisper import WhisperConfig
class MossTranscribeDiarizeConfig(PretrainedConfig):
class MossTranscribeDiarizeConfig(PreTrainedConfig):
"""Configuration for MOSS-Transcribe-Diarize."""
model_type = "moss_transcribe_diarize"
@@ -15,7 +15,7 @@ and vision configs are consumed by the native multimodal serving path.
from __future__ import annotations
from transformers import Qwen3Config
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
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]:
]
class MuseGlimmerTextConfig(PretrainedConfig):
class MuseGlimmerTextConfig(PreTrainedConfig):
model_type = "muse_glimmer_text"
keys_to_ignore_at_inference = ["past_key_values"]
@@ -127,7 +127,7 @@ class MuseGlimmerTextConfig(PretrainedConfig):
)
class MuseGlimmerVisionConfig(PretrainedConfig):
class MuseGlimmerVisionConfig(PreTrainedConfig):
model_type = "muse_glimmer_vision"
def __init__(
@@ -177,7 +177,7 @@ class MuseGlimmerVisionConfig(PretrainedConfig):
self.layer_types = layer_types
class MuseGlimmerConfig(PretrainedConfig):
class MuseGlimmerConfig(PreTrainedConfig):
model_type = "muse_glimmer"
sub_configs = {
"text_config": MuseGlimmerTextConfig,
+4 -4
View File
@@ -17,22 +17,22 @@
# limitations under the License.
"""Nemotron model configuration."""
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from transformers.utils import logging
logger = logging.get_logger(__name__)
class NemotronConfig(PretrainedConfig):
class NemotronConfig(PreTrainedConfig):
r"""This is the configuration class to store the configuration of a
[`NemotronModel`]. It is used to instantiate a Nemotron model
according to the specified arguments, defining the model architecture.
Instantiating a configuration with the defaults will yield a similar
configuration to that of the Nemotron-8B.
Configuration objects inherit from [`PretrainedConfig`] and can be
Configuration objects inherit from [`PreTrainedConfig`] and can be
used to control the model outputs. Read the documentation from
[`PretrainedConfig`] for more information.
[`PreTrainedConfig`] for more information.
Args:
@@ -18,13 +18,13 @@
"""NemotronH model configuration."""
import regex as re
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
from transformers.utils import logging
logger = logging.get_logger(__name__)
class NemotronHConfig(PretrainedConfig):
class NemotronHConfig(PreTrainedConfig):
r"""This is the configuration class to store the configuration of a
[`NemotronHModel`]. It is used to instantiate a NemotronH model according
to the specified arguments, defining the model architecture. Instantiating
@@ -2,10 +2,10 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
class OlmoHybridConfig(PretrainedConfig):
class OlmoHybridConfig(PreTrainedConfig):
r"""Configuration class for [`OlmoHybridModel`]. It is used to
instantiate an OLMo Hybrid model according to the specified
arguments, defining the model architecture. Instantiating a
+2 -2
View File
@@ -9,10 +9,10 @@ configuration without executing Hugging Face remote code.
from typing import Any
from transformers import LlamaConfig, PretrainedConfig
from transformers import LlamaConfig, PreTrainedConfig
class OpenVLAConfig(PretrainedConfig):
class OpenVLAConfig(PreTrainedConfig):
"""Configuration class for OpenVLA models."""
model_type = "openvla"
+17 -17
View File
@@ -7,10 +7,10 @@
# Ovis Config with AimV2 config registration removed for Transformers compatibility
from typing import Any
from transformers import AutoConfig, PretrainedConfig
from transformers import AutoConfig, PreTrainedConfig
class AIMv2Config(PretrainedConfig):
class AIMv2Config(PreTrainedConfig):
"""This is the configuration class to store the configuration of an [`AIMv2Model`].
Instantiating a configuration with the defaults will yield a similar configuration
to that of the [apple/aimv2-large-patch14-224](https://huggingface.co/apple/aimv2-large-patch14-224).
@@ -29,7 +29,7 @@ class AIMv2Config(PretrainedConfig):
projection_dropout: Dropout ratio for the projection layer after the attention.
qkv_bias: Whether to add a bias to the queries, keys and values.
use_bias: Whether to add a bias in the feed-forward and projection layers.
kwargs: Keyword arguments for the [`PretrainedConfig`].
kwargs: Keyword arguments for the [`PreTrainedConfig`].
"""
@@ -70,7 +70,7 @@ class AIMv2Config(PretrainedConfig):
# ----------------------------------------------------------------------
# Visual Tokenizer Configuration
# ----------------------------------------------------------------------
class BaseVisualTokenizerConfig(PretrainedConfig):
class BaseVisualTokenizerConfig(PreTrainedConfig):
def __init__(
self,
vocab_size=16384,
@@ -78,7 +78,7 @@ class BaseVisualTokenizerConfig(PretrainedConfig):
tau=1.0,
depths=None,
drop_cls_token=False,
backbone_config: PretrainedConfig | dict | None = None,
backbone_config: PreTrainedConfig | dict | None = None,
hidden_stride: int = 1,
**kwargs,
):
@@ -92,10 +92,10 @@ class BaseVisualTokenizerConfig(PretrainedConfig):
self.backbone_kwargs = dict[str, Any]()
self.drop_cls_token = drop_cls_token
if backbone_config is not None:
assert isinstance(backbone_config, (PretrainedConfig, dict)), (
f"expect `backbone_config` to be instance of PretrainedConfig or dict, but got {type(backbone_config)} type"
assert isinstance(backbone_config, (PreTrainedConfig, dict)), (
f"expect `backbone_config` to be instance of PreTrainedConfig or dict, but got {type(backbone_config)} type"
)
if not isinstance(backbone_config, PretrainedConfig):
if not isinstance(backbone_config, PreTrainedConfig):
model_type = backbone_config["model_type"]
if model_type != "aimv2":
backbone_config.pop("model_type")
@@ -139,13 +139,13 @@ AutoConfig.register("aimv2_visual_tokenizer", Aimv2VisualTokenizerConfig)
# ----------------------------------------------------------------------
# Ovis Configuration
# ----------------------------------------------------------------------
class OvisConfig(PretrainedConfig):
class OvisConfig(PreTrainedConfig):
model_type = "ovis"
def __init__(
self,
llm_config: PretrainedConfig | dict | None = None,
visual_tokenizer_config: PretrainedConfig | dict | None = None,
llm_config: PreTrainedConfig | dict | None = None,
visual_tokenizer_config: PreTrainedConfig | dict | None = None,
multimodal_max_length=8192,
hidden_size=None,
conversation_formatter_class=None,
@@ -155,10 +155,10 @@ class OvisConfig(PretrainedConfig):
):
super().__init__(**kwargs)
if llm_config is not None:
assert isinstance(llm_config, (PretrainedConfig, dict)), (
f"expect `llm_config` to be instance of PretrainedConfig or dict, but got {type(llm_config)} type"
assert isinstance(llm_config, (PreTrainedConfig, dict)), (
f"expect `llm_config` to be instance of PreTrainedConfig or dict, but got {type(llm_config)} type"
)
if not isinstance(llm_config, PretrainedConfig):
if not isinstance(llm_config, PreTrainedConfig):
model_type = llm_config["model_type"]
llm_config.pop("model_type")
llm_config = AutoConfig.for_model(model_type, **llm_config)
@@ -166,10 +166,10 @@ class OvisConfig(PretrainedConfig):
# map llm_config to text_config
self.text_config = llm_config
if visual_tokenizer_config is not None:
assert isinstance(visual_tokenizer_config, (PretrainedConfig, dict)), (
f"expect `visual_tokenizer_config` to be instance of PretrainedConfig or dict, but got {type(visual_tokenizer_config)} type"
assert isinstance(visual_tokenizer_config, (PreTrainedConfig, dict)), (
f"expect `visual_tokenizer_config` to be instance of PreTrainedConfig or dict, but got {type(visual_tokenizer_config)} type"
)
if not isinstance(visual_tokenizer_config, PretrainedConfig):
if not isinstance(visual_tokenizer_config, PreTrainedConfig):
model_type = visual_tokenizer_config["model_type"]
visual_tokenizer_config.pop("model_type")
visual_tokenizer_config = AutoConfig.for_model(
+5 -5
View File
@@ -2,7 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from dataclasses import dataclass
from transformers import ParakeetEncoderConfig, PretrainedConfig
from transformers import ParakeetEncoderConfig, PreTrainedConfig
class ParakeetConfig(ParakeetEncoderConfig):
@@ -24,9 +24,9 @@ class ParakeetConfig(ParakeetEncoderConfig):
@staticmethod
def from_hf_config(
config: PretrainedConfig, *, llm_hidden_size: int, max_model_len: int
config: PreTrainedConfig, *, llm_hidden_size: int, max_model_len: int
) -> "ParakeetConfig":
assert isinstance(config, PretrainedConfig)
assert isinstance(config, PreTrainedConfig)
return ParakeetConfig(
**config.to_dict(),
scale_input=False,
@@ -55,8 +55,8 @@ class ExtractorConfig:
padding_value: float = 0.0
@classmethod
def from_hf_config(cls, config: PretrainedConfig) -> "ExtractorConfig":
assert isinstance(config, PretrainedConfig)
def from_hf_config(cls, config: PreTrainedConfig) -> "ExtractorConfig":
assert isinstance(config, PreTrainedConfig)
defaults = ("hop_length", "win_length", "preemphasis", "n_fft", "padding_value")
optional_kwargs = {
name: getattr(config, name) for name in defaults if hasattr(config, name)
@@ -3,11 +3,11 @@
from typing import Any
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from transformers.models.auto import CONFIG_MAPPING
class QianfanOCRVisionConfig(PretrainedConfig):
class QianfanOCRVisionConfig(PreTrainedConfig):
model_type = "qianfan_ocr_vision"
def __init__(
@@ -53,7 +53,7 @@ class QianfanOCRVisionConfig(PretrainedConfig):
self.use_mean_pooling = use_mean_pooling
class QianfanOCRConfig(PretrainedConfig):
class QianfanOCRConfig(PreTrainedConfig):
model_type = "qianfan_ocr"
def __init__(
+6 -6
View File
@@ -16,10 +16,10 @@
# limitations under the License.
"""Qwen3.5 model configuration."""
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
class Qwen3_5TextConfig(PretrainedConfig):
class Qwen3_5TextConfig(PreTrainedConfig):
model_type = "qwen3_5_text"
keys_to_ignore_at_inference = ["past_key_values"]
@@ -108,7 +108,7 @@ class Qwen3_5TextConfig(PretrainedConfig):
self.linear_num_value_heads = linear_num_value_heads
super().__init__(**kwargs)
# Set these AFTER super().__init__() because transformers v4's
# PretrainedConfig.__init__ has these as explicit params with different
# PreTrainedConfig.__init__ has these as explicit params with different
# defaults (e.g. tie_word_embeddings=True) that would overwrite our values.
self.pad_token_id = pad_token_id
self.bos_token_id = bos_token_id
@@ -116,7 +116,7 @@ class Qwen3_5TextConfig(PretrainedConfig):
self.tie_word_embeddings = tie_word_embeddings
class Qwen3_5VisionConfig(PretrainedConfig):
class Qwen3_5VisionConfig(PreTrainedConfig):
model_type = "qwen3_5"
base_config_key = "vision_config"
@@ -152,7 +152,7 @@ class Qwen3_5VisionConfig(PretrainedConfig):
self.initializer_range = initializer_range
class Qwen3_5Config(PretrainedConfig):
class Qwen3_5Config(PreTrainedConfig):
model_type = "qwen3_5"
sub_configs = {
"vision_config": Qwen3_5VisionConfig,
@@ -186,7 +186,7 @@ class Qwen3_5Config(PretrainedConfig):
self.vision_start_token_id = vision_start_token_id
self.vision_end_token_id = vision_end_token_id
super().__init__(**kwargs)
# Set after super().__init__() to avoid v4 PretrainedConfig overwrite
# Set after super().__init__() to avoid v4 PreTrainedConfig overwrite
self.tie_word_embeddings = tie_word_embeddings
@@ -16,10 +16,10 @@
# limitations under the License.
"""Qwen3.5-MoE model configuration."""
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
class Qwen3_5MoeTextConfig(PretrainedConfig):
class Qwen3_5MoeTextConfig(PreTrainedConfig):
model_type = "qwen3_5_moe_text"
keys_to_ignore_at_inference = ["past_key_values"]
@@ -120,7 +120,7 @@ class Qwen3_5MoeTextConfig(PretrainedConfig):
self.router_aux_loss_coef = router_aux_loss_coef
super().__init__(**kwargs)
# Set these AFTER super().__init__() because transformers v4's
# PretrainedConfig.__init__ has these as explicit params with different
# PreTrainedConfig.__init__ has these as explicit params with different
# defaults (e.g. tie_word_embeddings=True) that would overwrite our values.
self.pad_token_id = pad_token_id
self.bos_token_id = bos_token_id
@@ -128,7 +128,7 @@ class Qwen3_5MoeTextConfig(PretrainedConfig):
self.tie_word_embeddings = tie_word_embeddings
class Qwen3_5MoeVisionConfig(PretrainedConfig):
class Qwen3_5MoeVisionConfig(PreTrainedConfig):
model_type = "qwen3_5_moe"
base_config_key = "vision_config"
@@ -164,7 +164,7 @@ class Qwen3_5MoeVisionConfig(PretrainedConfig):
self.initializer_range = initializer_range
class Qwen3_5MoeConfig(PretrainedConfig):
class Qwen3_5MoeConfig(PreTrainedConfig):
model_type = "qwen3_5_moe"
sub_configs = {
"vision_config": Qwen3_5MoeVisionConfig,
@@ -198,7 +198,7 @@ class Qwen3_5MoeConfig(PretrainedConfig):
self.vision_start_token_id = vision_start_token_id
self.vision_end_token_id = vision_end_token_id
super().__init__(**kwargs)
# Set after super().__init__() to avoid v4 PretrainedConfig overwrite
# Set after super().__init__() to avoid v4 PreTrainedConfig overwrite
self.tie_word_embeddings = tie_word_embeddings
+14 -14
View File
@@ -17,14 +17,14 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
from transformers.modeling_rope_utils import rope_config_validation
from transformers.utils import logging
logger = logging.get_logger(__name__)
class Qwen3ASRAudioEncoderConfig(PretrainedConfig):
class Qwen3ASRAudioEncoderConfig(PreTrainedConfig):
r"""
This is the configuration class to store the configuration of a [`Qwen3ASRAudioEncoder`]. It is used to instantiate a
Qwen3-ASR audio encoder according to the specified arguments, defining the model architecture. Instantiating a
@@ -33,8 +33,8 @@ class Qwen3ASRAudioEncoderConfig(PretrainedConfig):
e.g. [Qwen/Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B)
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PreTrainedConfig`] for more information.
Args:
num_mel_bins (`int`, *optional*, defaults to 128):
@@ -130,15 +130,15 @@ class Qwen3ASRAudioEncoderConfig(PretrainedConfig):
self.downsample_hidden_size = downsample_hidden_size
class Qwen3ASRTextConfig(PretrainedConfig):
class Qwen3ASRTextConfig(PreTrainedConfig):
r"""
This is the configuration class to store the configuration of a [`Qwen3ASRTextModel`]. It is used to instantiate a
Qwen3-ASR model according to the specified arguments, defining the model architecture. Instantiating a configuration
with the defaults will yield a similar configuration to that of
Qwen3-ASR-1.7B [Qwen/Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B)
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PreTrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 151936):
@@ -285,7 +285,7 @@ class Qwen3ASRTextConfig(PretrainedConfig):
super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
class Qwen3ASRThinkerConfig(PretrainedConfig):
class Qwen3ASRThinkerConfig(PreTrainedConfig):
r"""
This is the configuration class to store the configuration of a [`Qwen3ASRThinker`]. It is used to instantiate a
Qwen3-ASR-Thinker model according to the specified arguments, defining the model architecture. Instantiating a
@@ -294,8 +294,8 @@ class Qwen3ASRThinkerConfig(PretrainedConfig):
e.g. [Qwen/Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B)
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PreTrainedConfig`] for more information.
Args:
audio_config (`dict`, *optional*):
@@ -365,7 +365,7 @@ class Qwen3ASRThinkerConfig(PretrainedConfig):
self.audio_token_id = audio_token_id
class Qwen3ASRConfig(PretrainedConfig):
class Qwen3ASRConfig(PreTrainedConfig):
"""
This is the configuration class to store the configuration of a [`Qwen3ASRForConditionalGeneration`]. It is used to instantiate a Qwen3ASR
model according to the specified sub-models configurations, defining the model architecture.
@@ -373,8 +373,8 @@ class Qwen3ASRConfig(PretrainedConfig):
Instantiating a configuration with the defaults will yield a similar configuration to that of the
[Qwen/Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PreTrainedConfig`] for more information.
Args:
thinker_config (`dict`, *optional*): Configuration of the underlying thinker sub-model.
@@ -437,7 +437,7 @@ class Qwen3ASRConfig(PretrainedConfig):
self.timestamp_segment_time = timestamp_segment_time
super().__init__(**kwargs)
def get_text_config(self, decoder=False) -> "PretrainedConfig":
def get_text_config(self, decoder=False) -> "PreTrainedConfig":
"""
Returns the config that is meant to be used with text IO. On most models, it is the original config instance
itself. On specific composite models, it is under a set of valid names.
@@ -16,20 +16,20 @@
# limitations under the License.
"""Qwen3-Next model configuration."""
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
from transformers.utils import logging
logger = logging.get_logger(__name__)
class Qwen3NextConfig(PretrainedConfig):
class Qwen3NextConfig(PreTrainedConfig):
r"""This is the configuration class to store the configuration of a [`Qwen3NextModel`]. It is used to instantiate a
Qwen3-Next model according to the specified arguments, defining the model architecture.
Instantiating a configuration with the defaults will yield a similar configuration to that of
Qwen3-Next-80B-A3B-Instruct [Qwen/Qwen3-Next-80B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct).
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Configuration objects inherit from [`PreTrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PreTrainedConfig`] for more information.
Args:
+2 -2
View File
@@ -4,7 +4,7 @@
from typing import Any
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
from transformers.utils import logging
logger = logging.get_logger(__name__)
@@ -20,7 +20,7 @@ OPENAI_CLIP_MEAN = (0.48145466, 0.4578275, 0.40821073)
OPENAI_CLIP_STD = (0.26862954, 0.26130258, 0.27577711)
class RadioConfig(PretrainedConfig):
class RadioConfig(PreTrainedConfig):
r"""This is the configuration class to store the configuration of a Radio
vision model. It is used to instantiate a Radio model according to the
specified arguments, defining the model architecture.
@@ -4,7 +4,7 @@ import os
from dataclasses import fields
from typing import Any
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm.transformers_utils.configs.speculators.algos import (
SUPPORTED_SPECULATORS_TYPES,
@@ -12,13 +12,13 @@ from vllm.transformers_utils.configs.speculators.algos import (
from vllm.transformers_utils.utils import without_trust_remote_code
class SpeculatorsConfig(PretrainedConfig):
class SpeculatorsConfig(PreTrainedConfig):
model_type = "speculators"
def __init__(self, **kwargs):
# super().__init__ performs some validation before setting all kwargs as
# attributes, so we set them first to be safe
pre_trained_config_fields = {f.name for f in fields(PretrainedConfig)}
pre_trained_config_fields = {f.name for f in fields(PreTrainedConfig)}
super_kwargs = dict()
for key, value in kwargs.items():
if key == "model_type":
+4 -4
View File
@@ -2,10 +2,10 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Any
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
class Step3VisionEncoderConfig(PretrainedConfig):
class Step3VisionEncoderConfig(PreTrainedConfig):
model_type = "step3_vision_encoder"
def __init__(
@@ -35,7 +35,7 @@ class Step3VisionEncoderConfig(PretrainedConfig):
super().__init__(**kwargs)
class Step3TextConfig(PretrainedConfig):
class Step3TextConfig(PreTrainedConfig):
model_type = "step3_text"
architectures = ["Step3TextForCausalLM"]
@@ -146,7 +146,7 @@ class Step3TextConfig(PretrainedConfig):
super().__init__(**kwargs)
class Step3VLConfig(PretrainedConfig):
class Step3VLConfig(PreTrainedConfig):
model_type = "step3_vl"
def __init__(
+2 -2
View File
@@ -2,10 +2,10 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Any
from transformers.configuration_utils import PretrainedConfig
from transformers.configuration_utils import PreTrainedConfig
class Step3p5Config(PretrainedConfig):
class Step3p5Config(PreTrainedConfig):
model_type = "step3p5"
def __init__(
+5 -5
View File
@@ -7,14 +7,14 @@ from typing import Any
import transformers
class UltravoxConfig(transformers.PretrainedConfig):
class UltravoxConfig(transformers.PreTrainedConfig):
r"""This is the configuration class to store the configuration of a
[`UltravoxForConditionalGeneration`]. It is used to instantiate an
Ultravox model according to the specified arguments, defining the model
architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to
control the model outputs. Read the documentation from [`PretrainedConfig`]
Configuration objects inherit from [`PreTrainedConfig`] and can be used to
control the model outputs. Read the documentation from [`PreTrainedConfig`]
for more information.
Args:
@@ -74,7 +74,7 @@ class UltravoxConfig(transformers.PretrainedConfig):
self.num_projector_layers = num_projector_layers
# N.B. May set the wrapped_model_config below.
self.wrapped_model_config: transformers.PretrainedConfig
self.wrapped_model_config: transformers.PreTrainedConfig
self.text_model_id = text_model_id
if text_model_id is None:
text_config = text_config or {}
@@ -113,7 +113,7 @@ class UltravoxConfig(transformers.PretrainedConfig):
return super().__setattr__(key, value)
@property
def text_config(self) -> transformers.PretrainedConfig:
def text_config(self) -> transformers.PreTrainedConfig:
# When Ultravox wraps a multi-modal model (e.g. Gemma), we instantiate
# the full model, but the text config is the text config of the inner
# model.
@@ -5,7 +5,7 @@ from typing import final
import torch
from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE
from transformers import PretrainedConfig
from transformers import PreTrainedConfig
from vllm import envs
from vllm.config.model_arch import (
@@ -26,8 +26,8 @@ logger = init_logger(__name__)
class ModelArchConfigConvertorBase:
def __init__(
self,
hf_config: PretrainedConfig,
hf_text_config: PretrainedConfig,
hf_config: PreTrainedConfig,
hf_text_config: PreTrainedConfig,
revision: str | None = None,
):
self.hf_config = hf_config
@@ -36,7 +36,7 @@ class ModelArchConfigConvertorBase:
def get_per_layer_hf_configs(
self,
) -> list[tuple[PretrainedConfig, PretrainedConfig]] | None:
) -> list[tuple[PreTrainedConfig, PreTrainedConfig]] | None:
"""`(hf_config, hf_text_config)` per layer, or `None` if homogeneous.
This is the only place that decides whether a checkpoint is heterogeneous
@@ -219,7 +219,7 @@ class ModelArchConfigConvertorBase:
@classmethod
def get_torch_dtype(
cls,
hf_config: PretrainedConfig,
hf_config: PreTrainedConfig,
model_id: str,
revision: str | None,
config_format: str | ConfigFormat,
@@ -257,7 +257,7 @@ class ModelArchConfigConvertorBase:
return config_dtype
def _normalize_quantization_config(self, config: PretrainedConfig):
def _normalize_quantization_config(self, config: PreTrainedConfig):
quant_cfg = getattr(config, "quantization_config", None)
if quant_cfg is None:
# compressed-tensors uses a "compression_config" key
@@ -585,8 +585,8 @@ class DeepSeekMTPModelArchConfigConvertor(ModelArchConfigConvertorBase):
class DeepseekV4ModelArchConfigConvertor(ModelArchConfigConvertorBase):
def __init__(
self,
hf_config: PretrainedConfig,
hf_text_config: PretrainedConfig,
hf_config: PreTrainedConfig,
hf_text_config: PreTrainedConfig,
revision: str | None = None,
):
# DeepSeek-V4-Flash-Vision-Exp ships the same architectures/model_type
@@ -622,7 +622,7 @@ class MimoMTPModelArchConfigConvertor(ModelArchConfigConvertorBase):
def _strip_mimo_v2_attention_chunk_size(
hf_config: PretrainedConfig, hf_text_config: PretrainedConfig
hf_config: PreTrainedConfig, hf_text_config: PreTrainedConfig
) -> None:
# MiMo-V2-Flash's config.json sets `attention_chunk_size=128` but the
# architecture does not actually use chunked local attention. Leaving it
@@ -635,8 +635,8 @@ def _strip_mimo_v2_attention_chunk_size(
class MimoV2ModelArchConfigConvertor(ModelArchConfigConvertorBase):
def __init__(
self,
hf_config: PretrainedConfig,
hf_text_config: PretrainedConfig,
hf_config: PreTrainedConfig,
hf_text_config: PreTrainedConfig,
revision: str | None = None,
):
if getattr(hf_config, "vision_config", None):
@@ -648,8 +648,8 @@ class MimoV2ModelArchConfigConvertor(ModelArchConfigConvertorBase):
class MimoV2MTPModelArchConfigConvertor(ModelArchConfigConvertorBase):
def __init__(
self,
hf_config: PretrainedConfig,
hf_text_config: PretrainedConfig,
hf_config: PreTrainedConfig,
hf_text_config: PreTrainedConfig,
revision: str | None = None,
):
super().__init__(hf_config, hf_text_config, revision)
@@ -742,7 +742,7 @@ class Gemma4ModelArchConfigConvertor(ModelArchConfigConvertorBase):
def get_per_layer_hf_configs(
self,
) -> list[tuple[PretrainedConfig, PretrainedConfig]] | None:
) -> list[tuple[PreTrainedConfig, PreTrainedConfig]] | None:
# Gemma4 uses a larger head dimension, and sometimes more KV heads, on its
# full attention layers than on its sliding ones. Transformers >= 5.15.0
# says so in the config; before that the values are flat attributes picked
@@ -763,7 +763,7 @@ class Gemma4ModelArchConfigConvertor(ModelArchConfigConvertorBase):
class MossAudioModelArchConfigConvertor(ModelArchConfigConvertorBase):
def _language_config(self) -> PretrainedConfig:
def _language_config(self) -> PreTrainedConfig:
return self.hf_config.language_config
def get_num_hidden_layers(self) -> int:
+2 -2
View File
@@ -4,12 +4,12 @@
# Adapted from
# https://github.com/zai-org/CogAgent
from transformers import PythonBackend
from transformers.image_processing_utils_fast import BaseImageProcessorFast
from transformers.image_processing_backends import TorchvisionBackend
from transformers.image_utils import PILImageResampling
from transformers.processing_utils import ProcessorMixin
class GLM4VImageProcessorFast(BaseImageProcessorFast):
class GLM4VImageProcessorFast(TorchvisionBackend):
"""Port of https://huggingface.co/zai-org/glm-4v-9b/blob/main/tokenization_chatglm.py#L177
to HF Transformers.
"""
@@ -7,12 +7,12 @@ import math
import numpy as np
import torch
from torchvision.transforms.v2 import functional as tvF
from transformers.image_processing_utils import BatchFeature
from transformers.image_processing_utils_fast import (
BaseImageProcessorFast,
from transformers.image_processing_backends import (
TorchvisionBackend,
group_images_by_shape,
reorder_images,
)
from transformers.image_processing_utils import BatchFeature
from transformers.image_utils import (
OPENAI_CLIP_MEAN,
OPENAI_CLIP_STD,
@@ -327,7 +327,7 @@ class Glm5NextImageProcessorKwargs(ImagesKwargs, total=False): # type: ignore[c
max_image_tokens: int | None
class Glm5NextImageProcessor(BaseImageProcessorFast):
class Glm5NextImageProcessor(TorchvisionBackend):
"""Fast torchvision image processor for GLM-5.3-Flash.
``patch_expand_factor`` multiplies into the ``smart_resize`` spatial
@@ -759,7 +759,7 @@ class Glm5NextProcessor(ProcessorMixin):
attributes = ["image_processor", "tokenizer", "video_processor"]
image_processor_class = "AutoImageProcessor"
video_processor_class = "AutoVideoProcessor"
tokenizer_class = ("PreTrainedTokenizer", "PreTrainedTokenizerFast")
tokenizer_class = ("PythonBackend", "TokenizersBackend")
def __init__(
self,
@@ -18,8 +18,8 @@ import math
import torch
from torchvision.transforms import InterpolationMode
from transformers import AutoTokenizer, BatchFeature
from transformers.image_processing_utils_fast import (
BaseImageProcessorFast,
from transformers.image_processing_backends import (
TorchvisionBackend,
group_images_by_shape,
reorder_images,
)
@@ -156,7 +156,7 @@ class MiniMaxM3VLImageProcessorKwargs(ImagesKwargs, total=False): # type: ignor
max_long_side_pixel: int
class MiniMaxM3VLImageProcessor(BaseImageProcessorFast):
class MiniMaxM3VLImageProcessor(TorchvisionBackend):
do_resize = True
resample = PILImageResampling.BICUBIC
# required by base-class validation, not used as the resize bound
@@ -437,7 +437,7 @@ class MuseGlimmerProcessor(ProcessorMixin):
attributes = ["image_processor", "video_processor", "tokenizer"]
image_processor_class = "AutoImageProcessor"
video_processor_class = "AutoVideoProcessor"
tokenizer_class = "PreTrainedTokenizerFast"
tokenizer_class = "TokenizersBackend"
def __init__(
self,
@@ -20,7 +20,7 @@ import numpy.typing as npt
import regex as re
import torch
from PIL import Image
from transformers import BatchFeature, PretrainedConfig, TensorType
from transformers import BatchFeature, PreTrainedConfig, TensorType
from vllm.model_executor.models.parakeet import ParakeetExtractor
from vllm.multimodal.inputs import AudioItem
@@ -584,7 +584,7 @@ class BaseNanoNemotronVLProcessor(ABC):
def __init__(
self,
config: PretrainedConfig,
config: PreTrainedConfig,
tokenizer: HfTokenizer,
*args,
max_model_len: int,
@@ -623,7 +623,7 @@ class BaseNanoNemotronVLProcessor(ABC):
self.dtype: torch.dtype = getattr(config, "dtype", torch.float32)
@staticmethod
def use_dynamic_resolution(config: PretrainedConfig) -> bool:
def use_dynamic_resolution(config: PreTrainedConfig) -> bool:
return "min_num_patches" in config.vision_config.args
@property
@@ -771,7 +771,7 @@ class NanoNemotronVLProcessor(BaseNanoNemotronVLProcessor):
def __init__(
self,
config: PretrainedConfig,
config: PreTrainedConfig,
tokenizer: HfTokenizer,
*,
max_model_len: int,