Add pydocstyle to the ruff rules (#52136)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Signed-off-by: Harry Mellor <hej.mellor@gmail.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Harry Mellor
2026-09-16 16:44:29 +00:00
committed by GitHub
co-authored by Claude Opus 5
parent 804e5377aa
commit 8c1557a79c
2002 changed files with 15024 additions and 14269 deletions
+21 -24
View File
@@ -231,7 +231,6 @@ 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__"):
return
@@ -489,10 +488,11 @@ def register_config_parser(config_format: str):
"""Register a customized vllm config parser.
When a config format is not supported by vllm, you can register a customized
config parser to support it.
Args:
config_format (str): The config parser format name.
Examples:
Args:
config_format (str): The config parser format name.
Examples:
>>> from vllm.transformers_utils.config import (get_config_parser,
register_config_parser)
>>> from vllm.transformers_utils.config_parser_base import ConfigParserBase
@@ -511,6 +511,7 @@ def register_config_parser(config_format: str):
>>>
>>> type(get_config_parser("custom_config_parser"))
<class 'CustomConfigParser'>
""" # noqa: E501
def _wrapper(config_parser_cls):
@@ -538,7 +539,8 @@ def register_config_parser(config_format: str):
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."""
This function sets a default rope_theta if it's missing.
"""
if getattr(config, "rope_parameters", None) is None:
config.rope_parameters = {"rope_type": "default"}
if "rope_theta" not in config.rope_parameters:
@@ -547,8 +549,8 @@ def set_default_rope_theta(config: PretrainedConfig, default_theta: float) -> No
def patch_legacy_rope_type(rope_parameters: dict[str, Any] | None) -> None:
"""Patch legacy RoPE type fields for backwards compatibility with
older custom models which would otherwise fail to load."""
older custom models which would otherwise fail to load.
"""
# No RoPE parameters to patch
if rope_parameters is None:
return
@@ -721,9 +723,7 @@ def is_encoder_decoder(config: PretrainedConfig) -> bool:
def _maybe_update_auto_config_kwargs(kwargs: dict[str, Any], model_type: str):
"""
Update kwargs for AutoConfig initialization based on model_type
"""
"""Update kwargs for AutoConfig initialization based on model_type."""
if model_type in _AUTO_CONFIG_KWARGS_OVERRIDES:
kwargs.update(_AUTO_CONFIG_KWARGS_OVERRIDES[model_type])
return kwargs
@@ -748,8 +748,7 @@ def maybe_override_with_speculators(
hf_token: bool | str | None = None,
**kwargs,
) -> tuple[str, str | None, dict[str, Any] | None]:
"""
Resolve model configuration when speculators are detected.
"""Resolve model configuration when speculators are detected.
Checks if the provided model is a speculators model and if so, extracts
the target model configuration and builds the speculative config.
@@ -764,6 +763,7 @@ def maybe_override_with_speculators(
Returns:
Tuple of (resolved_model, resolved_tokenizer, speculative_config)
"""
kwargs["local_files_only"] = huggingface_hub.constants.HF_HUB_OFFLINE
config_dict, _ = PretrainedConfig.get_config_dict(
@@ -938,8 +938,7 @@ def get_pooling_config(
model: str,
revision: str | None = "main",
) -> dict[str, Any] | None:
"""
This function gets the pooling and normalize
"""This function gets the pooling and normalize
config from the model - only applies to
sentence-transformers models.
@@ -951,6 +950,7 @@ def get_pooling_config(
Returns:
A dictionary containing the pooling type and whether
normalization is used, or None if no pooling configuration is found.
"""
modules_file_name = "modules.json"
@@ -1035,19 +1035,17 @@ def parse_pooling_type(pooling_name: str):
def get_sentence_transformer_tokenizer_config(
model: str | Path, revision: str | None = "main"
) -> dict[str, Any] | None:
"""
Returns the tokenization configuration dictionary for a
"""Returns the tokenization configuration dictionary for a
given Sentence Transformer BERT model.
Parameters:
- model (str|Path): The name of the Sentence Transformer
BERT model.
- revision (str, optional): The revision of the m
odel to use. Defaults to 'main'.
Args:
model (str|Path): The name of the Sentence Transformer BERT model.
revision (str, optional): The revision of the model to use. Defaults to 'main'.
Returns:
- dict: A dictionary containing the configuration parameters
for the Sentence Transformer BERT model.
"""
sentence_transformer_config_files = [
"sentence_bert_config.json",
@@ -1308,9 +1306,7 @@ def get_safetensors_params_metadata(
*,
revision: str | None = None,
) -> dict[str, Any]:
"""
Get the safetensors parameters metadata for remote/local model repository.
"""
"""Get the safetensors parameters metadata for remote/local model repository."""
if (model_path := Path(model)).exists():
return _read_safetensors_metadata_in_dir(model_path)
@@ -1354,6 +1350,7 @@ def checkpoint_has_lm_head(model: str, *, revision: str | None = None) -> bool |
Returns:
`None` if the checkpoint contents could not be determined, for example
because it is not stored as safetensors.
"""
metadata = get_safetensors_params_metadata(model, revision=revision)
if not metadata:
+3 -2
View File
@@ -6,14 +6,14 @@ from transformers import PretrainedConfig
class AXK1Config(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`AXK1Model`].
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
information.
Args:
vocab_size (`int`, *optional*, defaults to 163840):
Vocabulary size of the A.X K1 model. Defines the number of different
@@ -112,6 +112,7 @@ class AXK1Config(PretrainedConfig):
layers during self-attention.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
"""
model_type = "axk1"
+1 -2
View File
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Model configs may be defined in this directory for the following reasons:
"""Model configs may be defined in this directory for the following reasons:
- There is no configuration file defined by HF Hub or Transformers library.
- There is a need to override the existing config to support vLLM.
+1 -2
View File
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
ColPali configuration that extends PaliGemmaConfig with embedding projection
"""ColPali configuration that extends PaliGemmaConfig with embedding projection
fields. This allows ColPali models to be loaded without trust_remote_code
by mapping their custom model_type (colpali) to a standard config class
that vLLM understands.
+1 -2
View File
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
ColQwen3 configuration that extends Qwen3VLConfig with embedding projection
"""ColQwen3 configuration that extends Qwen3VLConfig with embedding projection
fields. This allows ColQwen3 models to be loaded without trust_remote_code
by mapping their custom model_type (colqwen3, ops_colqwen3, etc.) to a
standard config class that vLLM understands.
@@ -1,7 +1,8 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Config definitions for ExtractHiddenStatesModel, to be used with
the extract_hidden_states spec decoding method."""
the extract_hidden_states spec decoding method.
"""
import os
+1 -1
View File
@@ -18,7 +18,7 @@
# 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.
"""Falcon configuration"""
"""Falcon configuration."""
from transformers.configuration_utils import PretrainedConfig
+2 -2
View File
@@ -6,8 +6,7 @@ from transformers.configuration_utils import PretrainedConfig
class HYV3Config(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`HYV3Model`].
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.
@@ -93,6 +92,7 @@ class HYV3Config(PretrainedConfig):
>>> config = HYV3Config()
>>> model = HYV3Model(config)
```
"""
model_type = "hy_v3"
+1
View File
@@ -52,6 +52,7 @@ class HYV4Config(PreTrainedConfig):
below zero disable the clamp.
num_nextn_predict_layers: Number of MTP layers in the checkpoint.
mtp_loss_factor: Training-time MTP loss weight (unused at inference).
"""
model_type = "hy_v4"
@@ -21,8 +21,7 @@ from transformers.configuration_utils import PretrainedConfig
class HyperCLOVAXConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a
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
@@ -166,6 +165,7 @@ class HyperCLOVAXConfig(PretrainedConfig):
`get_rope`. When provided, takes precedence over `rope_theta`
and `rope_scaling`. If `None`, it is derived from `rope_theta`
and `rope_scaling` automatically.
"""
model_type = "hyperclovax"
+2 -2
View File
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Kimi-K2.5 Model Configuration.
"""Kimi-K2.5 Model Configuration.
This configuration supports video-chunk as an internal modality type.
A video-chunk is the smallest independently processable unit of video.
@@ -73,6 +72,7 @@ class KimiK25Config(PretrainedConfig):
ignore_index: The ignore index for the loss function.
media_placeholder_token_id: The token ID for media placeholders.
pad_token_id: The token ID for padding.
"""
model_type = "kimi_k25"
+4 -3
View File
@@ -6,8 +6,7 @@ from transformers.configuration_utils import 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
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)
@@ -85,7 +84,9 @@ class Lfm2MoeConfig(PretrainedConfig):
>>> # Accessing the model configuration
>>> configuration = model.config
```""" # noqa: E501
```
""" # noqa: E501
model_type = "lfm2_moe"
keys_to_ignore_at_inference = ["past_key_values"]
@@ -24,8 +24,7 @@ class MLPSpeculatorConfig(PretrainedConfig):
scale_input: bool = False,
**kwargs,
):
"""
Initialize an MLPSpeculatorConfig
"""Initialize an MLPSpeculatorConfig.
Args:
vocab_size: int
@@ -52,6 +51,7 @@ class MLPSpeculatorConfig(PretrainedConfig):
scale_input: bool
if True, will scale the initial hidden states from
the base model.
"""
if top_k_tokens_per_head is None:
top_k_tokens_per_head = [5, 4, 3]
@@ -148,5 +148,6 @@ class Moondream3Config(PretrainedConfig):
Args:
decoder: Ignored. Only used for encoder-decoder models.
"""
return self.text_config
+6 -7
View File
@@ -15,7 +15,7 @@
# 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.
"""Nemotron model configuration"""
"""Nemotron model configuration."""
from transformers import PretrainedConfig
from transformers.utils import logging
@@ -24,8 +24,7 @@ logger = logging.get_logger(__name__)
class NemotronConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a
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
@@ -114,7 +113,9 @@ class NemotronConfig(PretrainedConfig):
>>> model = NemotronModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
```
"""
model_type = "nemotron"
keys_to_ignore_at_inference = ["past_key_values"]
@@ -193,9 +194,7 @@ class NemotronConfig(PretrainedConfig):
)
def _rope_parameters_validation(self):
"""
Validate the `rope_parameters` configuration.
"""
"""Validate the `rope_parameters` configuration."""
if self.rope_parameters is None:
return
@@ -15,7 +15,7 @@
# 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.
"""NemotronH model configuration"""
"""NemotronH model configuration."""
import regex as re
from transformers.configuration_utils import PretrainedConfig
@@ -25,12 +25,12 @@ logger = logging.get_logger(__name__)
class NemotronHConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a
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
a configuration with the defaults will yield a similar configuration to
that of the NemotronH-v0.1 model.
Args:
vocab_size (`int`, *optional*, defaults to 131072):
Vocabulary size of the NemotronH model. Defines the number of
@@ -139,6 +139,7 @@ class NemotronHConfig(PretrainedConfig):
Size of chunks for Mamba processing.
rescale_prenorm_residual (`bool`, *optional*, defaults to `True`):
Whether to rescale the pre-normalization residual connections.
"""
model_type = "nemotron_h"
@@ -6,8 +6,7 @@ from transformers.configuration_utils import PretrainedConfig
class OlmoHybridConfig(PretrainedConfig):
r"""
Configuration class for [`OlmoHybridModel`]. It is used to
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
configuration with the defaults will yield a similar
@@ -19,7 +18,7 @@ class OlmoHybridConfig(PretrainedConfig):
can be used to control the model outputs. Read the
documentation from [`PreTrainedConfig`] for more information.
Args:
Args:
vocab_size (`int`, *optional*, defaults to 100352):
Vocabulary size of the OlmoHybrid model. Defines
the number of different tokens that can be
@@ -159,6 +158,7 @@ class OlmoHybridConfig(PretrainedConfig):
>>> model = OlmoHybridModel(configuration)
>>> configuration = model.config
```
"""
model_type = "olmo_hybrid"
+2
View File
@@ -14,6 +14,7 @@ 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).
Args:
hidden_size: Dimension of the hidden representations.
intermediate_size: Dimension of the SwiGLU representations.
@@ -29,6 +30,7 @@ class AIMv2Config(PretrainedConfig):
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`].
"""
model_type: str = "aimv2"
+1 -1
View File
@@ -14,7 +14,7 @@
# 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.
"""Qwen3.5 model configuration"""
"""Qwen3.5 model configuration."""
from transformers.configuration_utils import PretrainedConfig
@@ -14,7 +14,7 @@
# 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.
"""Qwen3.5-MoE model configuration"""
"""Qwen3.5-MoE model configuration."""
from transformers.configuration_utils import PretrainedConfig
@@ -14,7 +14,7 @@
# 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.
"""Qwen3-Next model configuration"""
"""Qwen3-Next model configuration."""
from transformers.configuration_utils import PretrainedConfig
from transformers.utils import logging
@@ -23,8 +23,7 @@ logger = logging.get_logger(__name__)
class Qwen3NextConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`Qwen3NextModel`]. It is used to instantiate a
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).
@@ -157,6 +156,7 @@ class Qwen3NextConfig(PretrainedConfig):
>>> # Accessing the model configuration
>>> configuration = model.config
```
""" # noqa: E501
model_type = "qwen3_next"
+3 -3
View File
@@ -1,6 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Radio vision model configuration"""
"""Radio vision model configuration."""
from typing import Any
@@ -21,8 +21,7 @@ OPENAI_CLIP_STD = (0.26862954, 0.26130258, 0.27577711)
class RadioConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a Radio
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.
@@ -55,6 +54,7 @@ class RadioConfig(PretrainedConfig):
dedicated video patch embedder (3*T*P*P -> hidden) separate from the
image embedder (3*P*P -> hidden). When False, a single embedder with
input size 3*T*P*P is used for both (images are duplicated T times).
"""
model_type = "radio"
@@ -14,8 +14,7 @@ def register_speculator(name):
@register_speculator("eagle3")
def update_eagle3(config_dict: dict, pre_trained_config: dict) -> None:
"""
Apply Eagle-3 specific configuration transformations to the `dict` used to
"""Apply Eagle-3 specific configuration transformations to the `dict` used to
construct the Transformers PreTrainedConfig.
Eagle-3 specific fields:
@@ -28,7 +27,6 @@ def update_eagle3(config_dict: dict, pre_trained_config: dict) -> None:
provide intermediate hidden states that help the drafter make better
predictions. This is the standard field used in Eagle3 checkpoints.
"""
pre_trained_config["draft_vocab_size"] = config_dict.get("draft_vocab_size")
if config_dict.get("target_hidden_size") is not None:
pre_trained_config["target_hidden_size"] = config_dict["target_hidden_size"]
@@ -54,8 +52,7 @@ def update_eagle3(config_dict: dict, pre_trained_config: dict) -> None:
@register_speculator("peagle")
def update_peagle(config_dict: dict, pre_trained_config: dict) -> None:
"""
Apply PEagle (Parallel Eagle) specific configuration transformations to
"""Apply PEagle (Parallel Eagle) specific configuration transformations to
the `dict` used to construct the Transformers PreTrainedConfig.
PEagle specific fields:
@@ -92,8 +89,7 @@ def update_peagle(config_dict: dict, pre_trained_config: dict) -> None:
@register_speculator("dflash")
def update_dflash(config_dict: dict, pre_trained_config: dict) -> None:
"""
Apply DFlash specific configuration transformations to the `dict` used to
"""Apply DFlash specific configuration transformations to the `dict` used to
construct the Transformers PreTrainedConfig.
DFlash specific fields:
@@ -132,8 +128,7 @@ def update_dflash(config_dict: dict, pre_trained_config: dict) -> None:
@register_speculator("dflash2")
def update_dflash2(config_dict: dict, pre_trained_config: dict) -> None:
"""
Apply DFlash2 specific configuration transformations to the `dict` used
"""Apply DFlash2 specific configuration transformations to the `dict` used
to construct the Transformers PreTrainedConfig.
DFlash2 extends DFlash with a grouped convolution layer and a
@@ -156,8 +151,7 @@ def update_dflash2(config_dict: dict, pre_trained_config: dict) -> None:
@register_speculator("dspark")
def update_dspark(config_dict: dict, pre_trained_config: dict) -> None:
"""
Apply DSpark specific configuration transformations to the `dict` used to
"""Apply DSpark specific configuration transformations to the `dict` used to
construct the Transformers PreTrainedConfig.
DSpark extends DFlash with a Markov logit-bias head, reusing the same
@@ -47,9 +47,7 @@ class SpeculatorsConfig(PretrainedConfig):
def extract_transformers_pre_trained_config(
cls, config_dict: dict[str, Any]
) -> dict[str, Any]:
"""
Extract standard Transformers PreTrainedConfig config from speculators config.
"""
"""Extract the Transformers `PreTrainedConfig` from a speculators config."""
speculators_model_type = config_dict.get("speculators_model_type")
if speculators_model_type not in SUPPORTED_SPECULATORS_TYPES:
raise ValueError(
@@ -99,8 +97,7 @@ class SpeculatorsConfig(PretrainedConfig):
def build_vllm_speculative_config(
cls, config_dict: dict[str, Any]
) -> dict[str, Any]:
"""
Build vLLM-compatible speculative configuration from speculators format.
"""Build vLLM-compatible speculative configuration from speculators format.
This method extracts and transforms speculative configuration from the
speculators format into the structure expected by vLLM.
@@ -110,6 +107,7 @@ class SpeculatorsConfig(PretrainedConfig):
Returns:
Dictionary with vLLM-compatible speculative configuration
"""
# Extract speculators configuration
spec_config = config_dict["speculators_config"]
+2 -2
View File
@@ -8,8 +8,7 @@ import transformers
class UltravoxConfig(transformers.PretrainedConfig):
r"""
This is the configuration class to store the configuration of a
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.
@@ -41,6 +40,7 @@ class UltravoxConfig(transformers.PretrainedConfig):
Whether to apply layer normalization at the middle of the
projector or at the end. Versions v0.4.1 and below
use `False`, but v0.5 and above use `True`.
"""
model_type = "ultravox"
+1 -2
View File
@@ -29,8 +29,7 @@ def try_get_class_from_dynamic_module(
warn_on_fail: bool = True,
**kwargs,
) -> type | None:
"""
As `transformers.dynamic_module_utils.get_class_from_dynamic_module`,
"""As `transformers.dynamic_module_utils.get_class_from_dynamic_module`,
but ignoring any errors.
"""
try:
+8 -9
View File
@@ -37,7 +37,8 @@ def _transformers_v4_compatibility_import():
subset of `ProcessorChatTemplateKwargs` as defined in Transformers v4.
In Transformers v5 these were merged into `ProcessorChatTemplateKwargs` and
`ChatTemplateLoadKwargs` was removed. For backward compatibility, we add an alias
for `ChatTemplateLoadKwargs` if it doesn't exist."""
for `ChatTemplateLoadKwargs` if it doesn't exist.
"""
old_import = getattr(processing_utils, "ChatTemplateLoadKwargs", None)
new_import = getattr(processing_utils, "ProcessorChatTemplateKwargs", None)
if old_import is None and new_import is not None:
@@ -52,7 +53,8 @@ def _transformers_v4_compatibility_init() -> Any:
processor instance before calling the original `ProcessorMixin.__init__`.
This can be removed if `Molmo2ForConditionalGeneration` is upstreamed to
Transformers."""
Transformers.
"""
if hasattr(ProcessorMixin.__init__, "_vllm_patched"):
return
@@ -80,9 +82,7 @@ _V = TypeVar("_V", bound=BaseVideoProcessor, default=BaseVideoProcessor)
class HashableDict(dict):
"""
A dictionary that can be hashed by lru_cache.
"""
"""A dictionary that can be hashed by lru_cache."""
# NOTE: pythonic dict is not hashable,
# we override on it directly for simplicity
@@ -91,9 +91,7 @@ class HashableDict(dict):
class HashableList(list):
"""
A list that can be hashed by lru_cache.
"""
"""A list that can be hashed by lru_cache."""
def __hash__(self) -> int: # type: ignore[override]
return hash(tuple(self))
@@ -399,7 +397,8 @@ def get_feature_extractor(
**kwargs: Any,
):
"""Load an audio feature extractor for the given model name
via HuggingFace."""
via HuggingFace.
"""
try:
processor_name = convert_model_repo_to_path(processor_name)
feature_extractor = AutoFeatureExtractor.from_pretrained(
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Multi-modal processors may be defined in this directory for the following
"""Multi-modal processors may be defined in this directory for the following
reasons:
- There is no processing file defined by HF Hub or Transformers library.
+4 -11
View File
@@ -18,8 +18,7 @@ class BagelProcessorKwargs(ProcessingKwargs, total=False): # type: ignore[call-
class BagelProcessor(ProcessorMixin):
"""
Constructs a BAGEL processor which wraps a
"""Constructs a BAGEL processor which wraps a
SigLIP image processor and a Qwen2 tokenizer.
"""
@@ -36,9 +35,7 @@ class BagelProcessor(ProcessorMixin):
images: ImageInput = None,
**kwargs: Unpack[BagelProcessorKwargs],
):
"""
Main method to prepare for the model one or several sequences(s) and image(s).
"""
"""Prepare one or more sequences and images for the model."""
output_kwargs = self._merge_kwargs(
BagelProcessorKwargs,
tokenizer_init_kwargs=self.tokenizer.init_kwargs,
@@ -62,15 +59,11 @@ class BagelProcessor(ProcessorMixin):
return BatchFeature(data={**pixel_values, **text_inputs})
def batch_decode(self, *args, **kwargs):
"""
This method forwards all its arguments to Qwen2TokenizerFast's batch_decode.
"""
"""Forward all arguments to `Qwen2TokenizerFast.batch_decode`."""
return self.tokenizer.batch_decode(*args, **kwargs)
def decode(self, *args, **kwargs):
"""
This method forwards all its arguments to Qwen2TokenizerFast's decode.
"""
"""This method forwards all its arguments to Qwen2TokenizerFast's decode."""
return self.tokenizer.decode(*args, **kwargs)
@property
@@ -228,7 +228,7 @@ class FilterbankFeatures(nn.Module):
return self.fb
def splice_frames(self, x, frame_splicing):
"""Stacks frames together across feature dim
"""Stacks frames together across feature dim.
input is batch_size, feature_dim, num_frames
output is batch_size, feature_dim*frame_splicing, num_frames
@@ -537,7 +537,8 @@ class CohereASRFeatureExtractor(SequenceFeatureExtractor):
class CohereASRProcessor(ProcessorMixin):
"""HF-compatible processor combining CohereASRFeatureExtractor and a
tokenizer."""
tokenizer.
"""
feature_extractor_class = "CohereASRFeatureExtractor"
tokenizer_class = "AutoTokenizer"
@@ -70,8 +70,7 @@ class Cosmos3EdgeImageProcessor(Siglip2ImageProcessor):
resample: PILImageResampling | InterpolationMode | int | None = None,
**kwargs,
) -> torch.Tensor:
"""
Rescales the image so that the following conditions are met:
"""Rescales the image so that the following conditions are met:
1. Both dimensions (height and width) are divisible by 'factor'.
2. The total number of pixels is within the range ['min_pixels', 'max_pixels'].
@@ -322,6 +321,7 @@ class Cosmos3EdgeProcessor(Qwen3VLProcessor):
normalize, and patch video inputs.
chat_template: An optional Jinja template used to render chat messages
into a tokenizable string.
"""
attributes = ["image_processor", "tokenizer", "video_processor"]
@@ -234,9 +234,7 @@ class DeepseekOCRProcessor(ProcessorMixin):
images: list[Image.Image],
crop_mode: bool = CROP_MODE,
):
"""
Args:
"""Args:
prompt (str): the formatted prompt;
images (List[ImageType]): the list of images;
crop_mode (bool): if True, then crop the image;
@@ -248,8 +246,8 @@ class DeepseekOCRProcessor(ProcessorMixin):
- pixel_values (torch.FloatTensor): [n_patches, 3, H, W]
- image_id (int): the id of the image token
- num_image_tokens (List[int]): the number of image tokens
"""
"""
if prompt is None or images is None:
raise ValueError("prompt and images must be used at the same time.")
@@ -309,7 +307,6 @@ class DeepseekOCRProcessor(ProcessorMixin):
cropping: bool = True,
):
"""Tokenize text with <image> tags."""
num_image_tags = conversation.count(self.image_token)
if num_image_tags != len(images):
raise ValueError(
@@ -186,9 +186,7 @@ class DeepseekVLV2Processor(ProcessorMixin):
inference_mode: bool = True,
**kwargs: Any,
):
"""
Args:
"""Args:
prompt (str): the formatted prompt;
images (list[ImageType]): the list of images;
inference_mode (bool): if True, then remove the last eos token;
@@ -201,8 +199,8 @@ class DeepseekVLV2Processor(ProcessorMixin):
- pixel_values (torch.FloatTensor): [n_patches, 3, H, W]
- image_id (int): the id of the image token
- num_image_tokens (list[int]): the number of image tokens
"""
"""
if prompt is None or images is None:
raise ValueError("prompt and images must be used at the same time.")
@@ -276,9 +274,7 @@ class DeepseekVLV2Processor(ProcessorMixin):
inference_mode: bool = True,
**kwargs: Any,
):
"""
Args:
"""Args:
text (str): the formatted prompt;
images (list[ImageType]): the list of images;
inference_mode (bool): if True, then remove the last eos token;
@@ -290,8 +286,8 @@ class DeepseekVLV2Processor(ProcessorMixin):
- images (torch.FloatTensor): [n_images, 3, H, W]
- image_id (int): the id of the image token
- num_image_tokens (list[int]): the number of image tokens
"""
"""
prepare = self.process_one(
prompt=text,
images=images,
@@ -68,8 +68,7 @@ class KaldifeatFbank:
class FireRedASR2FeatureExtractor(SequenceFeatureExtractor):
r"""
Constructs a FireRedASR2 feature extractor.
r"""Constructs a FireRedASR2 feature extractor.
This feature extractor inherits from [`~feature_extraction_sequence_
utils.SequenceFeatureExtractor`] which contains most of the main
@@ -99,6 +98,7 @@ class FireRedASR2FeatureExtractor(SequenceFeatureExtractor):
Dithering has similar effect as `spectrogram(mel_floor=...)`. It reduces
the high log_mel_fbank values for signals with hard-zero sections,
when VAD cutoff is present in the signal.
"""
model_input_names = ["input_features"]
@@ -230,9 +230,8 @@ class FireRedASR2FeatureExtractor(SequenceFeatureExtractor):
class FireRedASR2Processor(ProcessorMixin):
r"""
Constructs a FireRedASR2 processor which wraps a FireRedASR2 feature extractor and
a FireRedASR2 tokenizer into a single processor.
r"""Constructs a FireRedASR2 processor which wraps a FireRedASR2 feature extractor
and a FireRedASR2 tokenizer into a single processor.
[`FireRedASR2Processor`] offers all the functionalities of
[`FireRedASR2FeatureExtractor`] and [`Qwen2Tokenizer`]. See the
@@ -246,6 +245,7 @@ class FireRedASR2Processor(ProcessorMixin):
tokenizer (`Qwen2Tokenizer`):
An instance of [`Qwen2Tokenizer`]. The tokenizer is a required
input.
"""
feature_extractor_class = "FireRedASR2FeatureExtractor"
@@ -271,8 +271,7 @@ class FireRedASR2Processor(ProcessorMixin):
)
def __call__(self, *args, **kwargs):
"""
Forwards the `audio` argument to FireRedASR2FeatureExtractor's
"""Forwards the `audio` argument to FireRedASR2FeatureExtractor's
[`~FireRedASR2FeatureExtractor.__call__`] and the `text` argument to
[`~Qwen2Tokenizer.__call__`]. Please refer to the docstring of the
above two methods for more information.
+6 -10
View File
@@ -21,10 +21,7 @@ logger = init_logger(__name__)
def apply_cmvn(inputs, cmvn): # noqa
"""
Apply CMVN with mvn data
"""
"""Apply CMVN with mvn data."""
device = inputs.device
# dtype = inputs.dtype
frame, dim = inputs.shape
@@ -219,8 +216,7 @@ class WavFrontend(nn.Module):
class FunASRFeatureExtractor(SequenceFeatureExtractor):
r"""
Constructs a FunASR feature extractor.
r"""Constructs a FunASR feature extractor.
This feature extractor inherits from [`~feature_extraction_sequence_
utils.SequenceFeatureExtractor`] which contains most of the main
@@ -255,6 +251,7 @@ class FunASRFeatureExtractor(SequenceFeatureExtractor):
Dithering has similar effect as `spectrogram(mel_floor=...)`. It reduces
the high log_mel_fbank values for signals with hard-zero sections,
when VAD cutoff is present in the signal.
"""
model_input_names = ["input_features"]
@@ -369,8 +366,7 @@ class FunASRFeatureExtractor(SequenceFeatureExtractor):
class FunASRProcessor(ProcessorMixin):
r"""
Constructs a FunASR processor which wraps a FunASR feature extractor and
r"""Constructs a FunASR processor which wraps a FunASR feature extractor and
a FunASR tokenizer into a single processor.
[`FunASRProcessor`] offers all the functionalities of
@@ -385,6 +381,7 @@ class FunASRProcessor(ProcessorMixin):
tokenizer (`Qwen2Tokenizer`):
An instance of [`Qwen2Tokenizer`]. The tokenizer is a required
input.
"""
feature_extractor_class = "FunASRFeatureExtractor"
@@ -410,8 +407,7 @@ class FunASRProcessor(ProcessorMixin):
)
def __call__(self, *args, **kwargs):
"""
Forwards the `audio` argument to FunASRFeatureExtractor's
"""Forwards the `audio` argument to FunASRFeatureExtractor's
[`~FunASRFeatureExtractor.__call__`] and the `text` argument to
[`~Qwen2Tokenizer.__call__`]. Please refer to the docstring of the
above two methods for more information.
+1 -2
View File
@@ -10,8 +10,7 @@ from transformers.processing_utils import ProcessorMixin
class GLM4VImageProcessorFast(BaseImageProcessorFast):
"""
Port of https://huggingface.co/zai-org/glm-4v-9b/blob/main/tokenization_chatglm.py#L177
"""Port of https://huggingface.co/zai-org/glm-4v-9b/blob/main/tokenization_chatglm.py#L177
to HF Transformers.
"""
@@ -261,7 +261,8 @@ def _resize_or_pad(
allow_upscale: bool = False,
) -> torch.Tensor:
"""Resize onto the aligned canvas, or keep the aspect ratio and
zero-pad the right/bottom sides (``resize_mode="pad"``)."""
zero-pad the right/bottom sides (``resize_mode="pad"``).
"""
height, width = stacked_images.shape[-2:]
if resize_mode == "resize":
@@ -307,7 +308,8 @@ def _pixel_budget(
) -> tuple[int, int]:
"""(min_pixels, max_pixels) from the token bounds of
``processor_config.json``; one vision token covers
``temporal_patch_size * (patch_size * merge_size) ** 2`` pixels."""
``temporal_patch_size * (patch_size * merge_size) ** 2`` pixels.
"""
if min_image_tokens is None or max_image_tokens is None:
raise ValueError(
"min_image_tokens and max_image_tokens must be provided by "
@@ -358,8 +358,7 @@ class InternVLVideoProcessor(BaseVideoProcessor):
class InternVLProcessor(ProcessorMixin):
"""
This model doesn't define its own HF processor,
"""This model doesn't define its own HF processor,
so we implement our own one here.
The code to insert image tokens is based on:
@@ -91,6 +91,7 @@ def get_image_size_for_max_num_patches(
`tuple[int, int]`: Height and width (in pixels) that are multiples of
`patch_size * pixel_shuffle_scale` and respect both the maximum and
optional minimum patch-count constraints.
"""
def get_scaled_image_size(scale, original_size, patch_size, pixel_shuffle_scale):
@@ -177,9 +178,11 @@ def prepare_image_tensor(
The tensor is converted to floating point if needed.
scale (`float`, *optional*, defaults to `VISION_SCALE`):
Scalar multiplier applied before normalization.
Returns:
`torch.Tensor`: Normalized tensor with the same shape as the input and
dtype `torch.float32`.
"""
if not torch.is_floating_point(image):
image = image.float()
@@ -209,6 +212,7 @@ def patchify_vision(image: torch.Tensor, patch_size: int) -> torch.Tensor:
Raises:
ValueError: If `height` or `width` is not divisible by `patch_size`.
"""
num_images, height, width, channels = image.shape
if height % patch_size or width % patch_size:
@@ -265,6 +269,7 @@ def process_vision_for_patches(
/ patch_size, channels * patch_size**2)` and `dims_virtual` encodes
effective `(images, height, width)` dimensions after optional pixel
shuffling.
"""
# Add batch dim if single image
if images.dim() == 3:
@@ -28,8 +28,7 @@ class KimiK25Processor(ProcessorMixin):
return_tensors: str | TensorType | None = None,
**kwargs,
) -> BatchFeature:
"""
Args:
"""Args:
text: The text to be field to the model.
vision_chunks: List of `VisionChunk` items to be processed.
For image: `VisionChunkImage` with
@@ -44,6 +43,7 @@ class KimiK25Processor(ProcessorMixin):
Returned when `vision_chunks` is not `None`.
- **grid_thws** -- list of image 3D grid in LLM.
Returned when `vision_chunks` is not `None`.
"""
if vision_chunks is not None:
mm_inputs = self.image_processor.preprocess(
@@ -1071,6 +1071,8 @@ class MiMoOmniProcessor(ProcessorMixin):
tuples (pre-decoded).
audio: list of ``str`` (path/url/base64), ``bytes``,
``(waveform_1D, sample_rate)`` tuples, or ``np.ndarray``.
video_audio: audio track extracted from ``videos``, in the same
formats accepted by ``audio``.
return_tensors: Passed to :class:`BatchFeature`.
Returns:
@@ -1079,6 +1081,7 @@ class MiMoOmniProcessor(ProcessorMixin):
- ``pixel_values`` + ``image_grid_thw``
- ``pixel_values_videos`` + ``video_grid_thw`` + ``second_per_grid_ts``
- ``audio_features``
"""
if isinstance(text, list):
text = text[0] if len(text) == 1 else "\n".join(text)
+7 -15
View File
@@ -14,9 +14,7 @@
# 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.
"""
Processor class for MiniCPMO.
"""
"""Processor class for MiniCPMO."""
import math
from typing import Literal, TypeAlias
@@ -35,8 +33,7 @@ MiniCPMOBatchFeature: TypeAlias = BatchFeature
class MiniCPMOProcessor(ProcessorMixin):
r"""
Constructs a MiniCPMV processor which wraps a MiniCPMV image
r"""Constructs a MiniCPMV processor which wraps a MiniCPMV image
processor and a MiniCPMV tokenizer into a single processor.
[`MiniCPMVProcessor`] offers all the functionalities of
@@ -49,6 +46,7 @@ class MiniCPMOProcessor(ProcessorMixin):
The image processor is a required input.
tokenizer ([`LlamaTokenizerWrapper`], *optional*):
The tokenizer is a required input.
"""
attributes = ["image_processor", "feature_extractor", "tokenizer"]
@@ -253,8 +251,7 @@ class MiniCPMOProcessor(ProcessorMixin):
# Copied from transformers.models.clip.processing_clip.CLIPProcessor.batch_decode
# with CLIP->Llama
def batch_decode(self, *args, **kwargs):
"""
This method forwards all its arguments to LlamaTokenizerFast's
"""This method forwards all its arguments to LlamaTokenizerFast's
[`~PythonBackend.batch_decode`]. Please refer to the
docstring of this method for more information.
"""
@@ -274,8 +271,7 @@ class MiniCPMOProcessor(ProcessorMixin):
# Copied from transformers.models.clip.processing_clip.CLIPProcessor.decode
# with CLIP->Llama
def decode(self, *args, **kwargs):
"""
This method forwards all its arguments to LlamaTokenizerFast's
"""This method forwards all its arguments to LlamaTokenizerFast's
[`~PythonBackend.decode`]. Please refer to the docstring
of this method for more information.
"""
@@ -551,15 +547,11 @@ class MelSpectrogramFeatures(torch.nn.Module):
)
def __call__(self, audio: torch.Tensor) -> torch.Tensor:
"""
audio: Tensor([num_channels, num_samples])
"""
"""audio: Tensor([num_channels, num_samples])."""
return super().__call__(audio)
def forward(self, audio: torch.Tensor) -> torch.Tensor:
"""
audio: Tensor([num_channels, num_samples])
"""
"""audio: Tensor([num_channels, num_samples])."""
mel: torch.Tensor = self.mel_spec(audio)
features = torch.log(torch.clip(mel, min=1e-5))
return features
@@ -13,9 +13,7 @@
# 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.
"""
Processor class for MiniCPMV.
"""
"""Processor class for MiniCPMV."""
from typing import TypeAlias
@@ -36,8 +34,7 @@ MiniCPMVBatchFeature: TypeAlias = BatchFeature
class MiniCPMVProcessor(ProcessorMixin):
r"""
Constructs a MiniCPMV processor which wraps a MiniCPMV image
r"""Constructs a MiniCPMV processor which wraps a MiniCPMV image
processor and a MiniCPMV tokenizer into a single processor.
[`MiniCPMVProcessor`] offers all the functionalities of
@@ -50,6 +47,7 @@ class MiniCPMVProcessor(ProcessorMixin):
The image processor is a required input.
tokenizer ([`LlamaTokenizerWrapper`], *optional*):
The tokenizer is a required input.
"""
attributes = ["image_processor", "tokenizer"]
@@ -93,8 +91,7 @@ class MiniCPMVProcessor(ProcessorMixin):
# Copied from transformers.models.clip.processing_clip.CLIPProcessor
# .batch_decode with CLIP->Llama
def batch_decode(self, *args, **kwargs):
"""
This method forwards all its arguments to LlamaTokenizerFast's
"""This method forwards all its arguments to LlamaTokenizerFast's
[`~PythonBackend.batch_decode`]. Please refer to the
docstring of this method for more information.
"""
@@ -126,8 +123,7 @@ class MiniCPMVProcessor(ProcessorMixin):
# Copied from transformers.models.clip.processing_clip.CLIPProcessor
# .decode with CLIP->Llama
def decode(self, *args, **kwargs):
"""
This method forwards all its arguments to LlamaTokenizerFast's
"""This method forwards all its arguments to LlamaTokenizerFast's
[`~PythonBackend.decode`]. Please refer to the docstring
of this method for more information.
"""
@@ -67,8 +67,7 @@ def select_tiling(
class Moondream3Processor(ProcessorMixin):
"""
Constructs a Moondream3 processor which handles image preprocessing
"""Constructs a Moondream3 processor which handles image preprocessing
and tokenization for the Moondream3 multimodal model.
Args:
@@ -78,6 +77,7 @@ class Moondream3Processor(ProcessorMixin):
max_crops: Maximum number of crops per image.
overlap_margin: Margin for overlapping crops in patches.
patch_size: Size of each patch.
"""
attributes = ["tokenizer"]
@@ -189,8 +189,7 @@ class Moondream3Processor(ProcessorMixin):
pretrained_model_name_or_path,
**kwargs,
):
"""
Load the processor, using a separate tokenizer repo.
"""Load the processor, using a separate tokenizer repo.
The moondream3 model uses a custom tokenizer from 'moondream/starmie-v1'
instead of having tokenizer files in the model repo.
@@ -294,8 +293,7 @@ class Moondream3Processor(ProcessorMixin):
| list[PreTokenizedInput] = None,
**kwargs: Unpack[Moondream3ProcessorKwargs],
) -> BatchFeature:
"""
Process images and text for Moondream3 model.
"""Process images and text for Moondream3 model.
Args:
images: Input images (PIL Image, numpy array, or list thereof).
@@ -304,6 +302,7 @@ class Moondream3Processor(ProcessorMixin):
Returns:
BatchFeature with processed inputs.
"""
output_kwargs = self._merge_kwargs(
Moondream3ProcessorKwargs,
@@ -428,8 +427,7 @@ class Moondream3Processor(ProcessorMixin):
convert_to_rgb: bool = True,
return_tensors: str = "pt",
) -> tuple[torch.Tensor, tuple[int, int]]:
"""
Preprocess an image using overlap-and-resize cropping strategy.
"""Preprocess an image using overlap-and-resize cropping strategy.
Args:
image: Input PIL image, numpy array, or torch tensor.
@@ -442,6 +440,7 @@ class Moondream3Processor(ProcessorMixin):
Returns:
Tuple of (pixel_values tensor, tiling tuple).
"""
image = self._to_pil_image(image)
if convert_to_rgb:
@@ -292,8 +292,7 @@ class DynamicResolutionImageTiler:
self,
target_num_tokens_post_shuffle: int,
) -> tuple[int, int]:
"""
TODO: optimize this so it squeezes closer to target number of tokens.
"""TODO: optimize this so it squeezes closer to target number of tokens.
Calculate image dimensions that produce approximately `target` tokens after
pixel_shuffle.
@@ -318,6 +317,7 @@ class DynamicResolutionImageTiler:
... height // PATCH_SIZE
... ) // 2**2 == 8100 # tokens post-shuffle
>>> assert tiler._get_num_embeddings(width=width, height=height) == 8100
"""
side_pixels = (
math.isqrt(target_num_tokens_post_shuffle)
@@ -386,6 +386,7 @@ class DynamicResolutionImageTiler:
num_tokens_available: Number of tokens available for this media
Returns:
DynamicResolutionParams for the media
"""
current_num_tokens_available = num_tokens_available
assert isinstance(media, Image.Image), (
@@ -473,6 +474,7 @@ class DynamicResolutionImageTiler:
num_tokens_available: Total number of tokens available across all media
Returns:
List of ImageTilingParams for each media item
"""
num_tokens_available = (
num_tokens_available
@@ -571,8 +573,7 @@ class DynamicResolutionImageTiler:
class BaseNanoNemotronVLProcessor(ABC):
"""
This model doesn't define its own HF processor,
"""This model doesn't define its own HF processor,
so we implement our own one here.
The code to insert image tokens is based on:
@@ -761,8 +762,7 @@ class BaseNanoNemotronVLProcessor(ABC):
class NanoNemotronVLProcessor(BaseNanoNemotronVLProcessor):
"""
HF Processor with extended video processing logic.
"""HF Processor with extended video processing logic.
Code for video processing is adapted from video example:
https://huggingface.co/OpenGVLab/InternVL3-1B#inference-with-transformers
"""
@@ -1125,8 +1125,7 @@ class NanoNemotronVLProcessor(BaseNanoNemotronVLProcessor):
img_context_token_ids: list[int],
video_temporal_patch_size: int = 1,
) -> PromptUpdateDetails:
"""
Build prompt replacement for a video.
"""Build prompt replacement for a video.
The replacement returned is not actually used to replace the placeholder
tokens - it's just used to make sure we allocate the correct number
of tokens.
@@ -191,8 +191,7 @@ class LlamaNemotronNanoVLImageProcessor(InternVLImageProcessor):
class LlamaNemotronNanoVLProcessor(InternVLProcessor):
"""
This model doesn't define its own HF processor,
"""This model doesn't define its own HF processor,
so we implement our own one here.
The image processor is given by:
@@ -287,8 +286,7 @@ class LlamaNemotronVLEmbedImageProcessor(InternVLImageProcessor):
class LlamaNemotronVLEmbedProcessor(InternVLProcessor):
"""
Processor for LlamaNemotronVL embedding model.
"""Processor for LlamaNemotronVL embedding model.
Inherits from NemotronVLProcessor and specializes it for embedding tasks:
- Uses SigLIP transform with normalization instead of base transform
+15 -13
View File
@@ -50,10 +50,10 @@ class OvisProcessorKwargs(ProcessingKwargs, total=False): # type: ignore[call-a
class OvisProcessor(ProcessorMixin):
r"""
Constructs an Ovis processor which wraps an Ovis image processor and a Qwen2 tokenizer into a single processor.
r"""Constructs an Ovis processor which wraps an Ovis image processor and a Qwen2 tokenizer into a single processor.
[`OvisProcessor`] offers all the functionalities of [`Qwen2VLImageProcessor`] and [`Qwen2TokenizerFast`]. See the
[`~OvisProcessor.__call__`] and [`~OvisProcessor.decode`] for more information.
Args:
image_processor ([`Qwen2VLImageProcessor`], *optional*):
The image processor is a required input.
@@ -61,6 +61,7 @@ class OvisProcessor(ProcessorMixin):
The tokenizer is a required input.
chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages
in a chat into a tokenizable string.
"""
attributes = ["image_processor", "tokenizer"]
@@ -107,12 +108,12 @@ class OvisProcessor(ProcessorMixin):
| list[PreTokenizedInput] = None,
**kwargs: Unpack[OvisProcessorKwargs],
) -> BatchFeature:
"""
Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
"""Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
and `kwargs` arguments to Qwen2TokenizerFast's [`~Qwen2TokenizerFast.__call__`] if `text` is not `None` to encode
the text. To prepare the vision inputs, this method forwards the `vision_infos` and `kwrags` arguments to
Qwen2VLImageProcessor's [`~Qwen2VLImageProcessor.__call__`] if `vision_infos` is not `None`.
Args:
Args:
images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `list[PIL.Image.Image]`, `list[np.ndarray]`, `list[torch.Tensor]`):
The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
tensor. Both channels-first and channels-last formats are supported.
@@ -129,7 +130,8 @@ class OvisProcessor(ProcessorMixin):
- `'pt'`: Return PyTorch `torch.Tensor` objects.
- `'np'`: Return NumPy `np.ndarray` objects.
- `'jax'`: Return JAX `jnp.ndarray` objects.
Returns:
Returns:
[`BatchFeature`]: A [`BatchFeature`] with the following fields:
- **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
- **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
@@ -140,8 +142,8 @@ class OvisProcessor(ProcessorMixin):
- **image_grid_thw** -- List of image 3D grid in LLM. Returned when `images` is not `None`.
- **video_grid_thw** -- List of video 3D grid in LLM. Returned when `videos` is not `None`.
- **second_per_grid_ts** -- List of video seconds per time grid. Returned when `videos` is not `None`.
"""
"""
max_partition = kwargs.pop("max_partition", 9)
covering_threshold = kwargs.pop("covering_threshold", 0.9)
@@ -416,28 +418,28 @@ class OvisProcessor(ProcessorMixin):
return torch.tensor(pixel_values), image_placeholders, torch.tensor(grid)
def batch_decode(self, *args, **kwargs):
"""
This method forwards all its arguments to Qwen2TokenizerFast's [`~PythonBackend.batch_decode`]. Please
"""This method forwards all its arguments to Qwen2TokenizerFast's [`~PythonBackend.batch_decode`]. Please
refer to the docstring of this method for more information.
"""
return self.tokenizer.batch_decode(*args, **kwargs)
def decode(self, *args, **kwargs):
"""
This method forwards all its arguments to Qwen2TokenizerFast's [`~PythonBackend.decode`]. Please refer to
"""This method forwards all its arguments to Qwen2TokenizerFast's [`~PythonBackend.decode`]. Please refer to
the docstring of this method for more information.
"""
return self.tokenizer.decode(*args, **kwargs)
def post_process_image_text_to_text(self, generated_outputs):
"""
Post-process the output of the model to decode the text.
"""Post-process the output of the model to decode the text.
Args:
generated_outputs (`torch.Tensor` or `np.ndarray`):
The output of the model `generate` function. The output is expected to be a tensor of shape `(batch_size, sequence_length)`
or `(sequence_length,)`.
Returns:
`list[str]`: The decoded text.
"""
return self.tokenizer.batch_decode(
generated_outputs,
+31 -33
View File
@@ -33,13 +33,13 @@ class Ovis2_5ProcessorKwargs(ProcessingKwargs, total=False): # type: ignore[cal
class Ovis2_5Processor(ProcessorMixin):
r"""
Constructs an Ovis processor which wraps an Ovis image processor
r"""Constructs an Ovis processor which wraps an Ovis image processor
and a Qwen2 tokenizer into a single processor.
[`OvisProcessor`] offers all the functionalities of
[`Qwen2VLImageProcessor`] and [`Qwen2TokenizerFast`].
See the [`~OvisProcessor.__call__`] and [`~OvisProcessor.decode`]
for more information.
Args:
image_processor ([`Qwen2VLImageProcessor`], *optional*):
The image processor is a required input.
@@ -48,6 +48,7 @@ class Ovis2_5Processor(ProcessorMixin):
chat_template (`str`, *optional*): A Jinja template which will
be used to convert lists of messages in a chat into
a tokenizable string.
"""
attributes = ["image_processor", "tokenizer"]
@@ -110,43 +111,39 @@ class Ovis2_5Processor(ProcessorMixin):
| list[PreTokenizedInput] = None,
**kwargs: Unpack[Ovis2_5ProcessorKwargs],
) -> BatchFeature:
"""
Main method to prepare for the model one or several sequences(s)
"""Main method to prepare for the model one or several sequences(s)
and image(s). This method forwards the `text`and `kwargs` arguments
to Qwen2TokenizerFast's [`~Qwen2TokenizerFast.__call__`] if `text`
is not `None` to encode the text. To prepare the vision inputs,
this method forwards the `vision_infos` and `kwrags` arguments to
Qwen2VLImageProcessor's [`~Qwen2VLImageProcessor.__call__`]
if `vision_infos` is not `None`.
Args:
images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`,
`list[PIL.Image.Image]`, `list[np.ndarray]`,
`list[torch.Tensor]`):
The image or batch of images to be prepared.
Each image can be a PIL image, NumPy array or PyTorch
tensor. Both channels-first and channels-last formats
are supported.
text (`str`, `list[str]`, `list[list[str]]`):
The sequence or batch of sequences to be encoded.
Each sequence can be a string or a list of strings
(pretokenized string). If the sequences are provided as
list of strings (pretokenized), you must set
`is_split_into_words=True` (to lift the ambiguity with
a batch of sequences).
videos (`np.ndarray`, `torch.Tensor`, `list[np.ndarray]`,
`list[torch.Tensor]`):
The image or batch of videos to be prepared. Each video
can be a 4D NumPy array or PyTorch tensor, or a nested
list of 3D frames. Both channels-first and channels-last
formats are supported.
return_tensors (`str` or [`~utils.TensorType`], *optional*):
If set, will return tensors of a particular framework.
Acceptable values are:
- `'tf'`: Return TensorFlow `tf.constant` objects.
- `'pt'`: Return PyTorch `torch.Tensor` objects.
- `'np'`: Return NumPy `np.ndarray` objects.
- `'jax'`: Return JAX `jnp.ndarray` objects.
Returns:
Args:
images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, or a list):
The image or batch of images to be prepared. Each image can
be a PIL image, NumPy array or PyTorch tensor. Both
channels-first and channels-last formats are supported.
videos (`np.ndarray`, `torch.Tensor`, or a list of either):
The video or batch of videos to be prepared. Each video can
be a 4D NumPy array or PyTorch tensor, or a nested list of
3D frames. Both channels-first and channels-last formats are
supported.
text (`str`, `list[str]`, `list[list[str]]`):
The sequence or batch of sequences to be encoded. Each
sequence can be a string or a list of strings (pretokenized
string). If the sequences are provided as list of strings
(pretokenized), you must set `is_split_into_words=True` (to
lift the ambiguity with a batch of sequences).
return_tensors (`str` or [`~utils.TensorType`], *optional*):
If set, will return tensors of a particular framework.
Acceptable values are:
- `'tf'`: Return TensorFlow `tf.constant` objects.
- `'pt'`: Return PyTorch `torch.Tensor` objects.
- `'np'`: Return NumPy `np.ndarray` objects.
- `'jax'`: Return JAX `jnp.ndarray` objects.
Returns:
[`BatchFeature`]: A [`BatchFeature`] with the following fields:
- **input_ids** -- list of token ids to be fed to a model.
Returned when `text` is not `None`.
@@ -164,6 +161,7 @@ class Ovis2_5Processor(ProcessorMixin):
when `videos` is not `None`.
- **second_per_grid_ts** -- list of video seconds per time grid.
Returned when `videos` is not `None`.
"""
output_kwargs = self._merge_kwargs(
Ovis2_5ProcessorKwargs,
@@ -11,8 +11,7 @@ from vllm.tokenizers.mistral import MistralTokenizer
class MistralCommonImageProcessor(ImageProcessingMixin):
"""
Provide a HF-compatible interface for
"""Provide a HF-compatible interface for
`mistral_common.tokens.tokenizers.multimodal.ImageEncoder`.
"""
@@ -18,8 +18,7 @@ from vllm.tokenizers.mistral import MistralTokenizer
class MistralCommonFeatureExtractor(SequenceFeatureExtractor):
"""
Provide a HF-compatible interface for
"""Provide a HF-compatible interface for
`mistral_common.tokens.tokenizers.multimodal.AudioEncoder`.
"""
+9 -9
View File
@@ -66,6 +66,7 @@ def resolve_revision(
Returns:
The resolved revision, or `revision` unchanged if it cannot be resolved
(local path, ModelScope, or any Hub error).
"""
if Path(repo_id).exists() or envs.VLLM_USE_MODELSCOPE:
return revision
@@ -106,6 +107,7 @@ def with_retry(
retry_delay: Seconds to wait after the first failure, doubled each time.
fatal_errors: Exceptions that are a definitive answer rather than a
transient failure. These are raised immediately, without logging.
"""
for attempt in range(max_retries):
try:
@@ -353,8 +355,7 @@ def get_hf_file_bytes(
def try_get_local_file(
model: str | Path, file_name: str, revision: str | None = "main"
) -> Path | Any | None:
"""
Try to get a local file from the HuggingFace repository.
"""Try to get a local file from the HuggingFace repository.
The possible return values are:
@@ -385,20 +386,19 @@ def try_get_local_file(
def get_hf_file_to_dict(
file_name: str, model: str | Path, revision: str | None = "main"
):
"""
Downloads a file from the Hugging Face Hub and returns
"""Downloads a file from the Hugging Face Hub and returns
its contents as a dictionary.
Parameters:
- file_name (str): The name of the file to download.
- model (str): The name of the model on the Hugging Face Hub.
- revision (str): The specific version of the model.
Args:
file_name (str): The name of the file to download.
model (str): The name of the model on the Hugging Face Hub.
revision (str): The specific version of the model.
Returns:
- config_dict (dict): A dictionary containing
the contents of the downloaded file.
"""
"""
file_path = try_get_local_file(model=model, file_name=file_name, revision=revision)
if file_path is None:
+5 -6
View File
@@ -25,14 +25,14 @@ except ImportError:
def list_safetensors(path: str = "") -> list[str]:
"""
List full file names from object path and filter by allow pattern.
"""List full file names from object path and filter by allow pattern.
Args:
path: The object storage path to list from.
Returns:
list[str]: List of full object storage paths allowed by the pattern
"""
return runai_list_safetensors(path)
@@ -42,8 +42,7 @@ def is_runai_obj_uri(model_or_path: str) -> bool:
class ObjectStorageModel:
"""
A class representing an ObjectStorage model mirrored into a
"""A class representing an ObjectStorage model mirrored into a
temporary directory.
Attributes:
@@ -51,6 +50,7 @@ class ObjectStorageModel:
Methods:
pull_files(): Pull model from object storage to the temporary directory.
"""
def __init__(self, url: str) -> None:
@@ -86,8 +86,7 @@ class ObjectStorageModel:
allow_pattern: list[str] | None = None,
ignore_pattern: list[str] | None = None,
) -> None:
"""
Pull files from object storage into the temporary directory.
"""Pull files from object storage into the temporary directory.
Args:
model_path: The object storage path of the model.
+4 -4
View File
@@ -36,8 +36,7 @@ def glob(
path: str = "",
allow_pattern: list[str] | None = None,
) -> list[str]:
"""
List full file names from S3 path and filter by allow pattern.
"""List full file names from S3 path and filter by allow pattern.
Args:
s3: S3 client to use.
@@ -46,6 +45,7 @@ def glob(
Returns:
list[str]: List of full S3 paths allowed by the pattern
"""
if s3 is None:
s3 = boto3.client("s3")
@@ -61,8 +61,7 @@ def list_files(
allow_pattern: list[str] | None = None,
ignore_pattern: list[str] | None = None,
) -> tuple[str, str, list[str]]:
"""
List files from S3 path and filter by pattern.
"""List files from S3 path and filter by pattern.
Args:
s3: S3 client to use.
@@ -77,6 +76,7 @@ def list_files(
and the prefix as a dir like string
- The third element is a list of files allowed or
disallowed by pattern
"""
parts = path.removeprefix("s3://").split("/")
prefix = "/".join(parts[1:])
+4 -4
View File
@@ -80,16 +80,15 @@ def _maybe_space_split_dict(path: str | PathLike) -> dict[str, str]:
@cache
def maybe_model_redirect(model: str) -> str:
"""
Use model_redirect to redirect the model name to a local folder.
"""Use model_redirect to redirect the model name to a local folder.
Args:
model: hf model name
Returns:
maybe redirect to a local folder
"""
"""
model_redirect_path = envs.VLLM_MODEL_REDIRECT_PATH
if not model_redirect_path:
@@ -117,7 +116,8 @@ def parse_safetensors_file_metadata(path: str | PathLike) -> dict[str, Any]:
def convert_model_repo_to_path(model_repo: str) -> str:
"""When VLLM_USE_MODELSCOPE is True convert a model
repository string to a Path str."""
repository string to a Path str.
"""
if not envs.VLLM_USE_MODELSCOPE or Path(model_repo).exists():
return model_repo
from modelscope.utils.file_utils import get_model_cache_root