Docling/docling/datamodel/pipeline_options_vlm_model.py
Peter W. J. Staar cfdf4cea25
feat: new vlm-models support (#1570)
* feat: adding new vlm-models support

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* fixed the transformers

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* got microsoft/Phi-4-multimodal-instruct to work

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* working on vlm's

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* refactoring the VLM part

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* all working, now serious refacgtoring necessary

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* refactoring the download_model

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* added the formulate_prompt

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* pixtral 12b runs via MLX and native transformers

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* added the VlmPredictionToken

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* refactoring minimal_vlm_pipeline

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* fixed the MyPy

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* added pipeline_model_specializations file

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* need to get Phi4 working again ...

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* finalising last points for vlms support

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* fixed the pipeline for Phi4

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* streamlining all code

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* reformatted the code

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* fixing the tests

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* added the html backend to the VLM pipeline

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* fixed the static load_from_doctags

Signed-off-by: Peter Staar <taa@zurich.ibm.com>

* restore stable imports

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* use AutoModelForVision2Seq for Pixtral and review example (including rename)

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* remove unused value

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* refactor instances of VLM models

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* skip compare example in CI

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* use lowercase and uppercase only

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* add new minimal_vlm example and refactor pipeline_options_vlm_model for cleaner import

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* rename pipeline_vlm_model_spec

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* move more argument to options and simplify model init

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* add supported_devices

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* remove not-needed function

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* exclude minimal_vlm

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* missing file

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* add message for transformers version

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* rename to specs

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* use module import and remove MLX from non-darwin

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* remove hf_vlm_model and add extra_generation_args

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* use single HF VLM model class

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* remove torch type

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

* add docs for vision models

Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>

---------

Signed-off-by: Peter Staar <taa@zurich.ibm.com>
Signed-off-by: Michele Dolfi <dol@zurich.ibm.com>
Co-authored-by: Michele Dolfi <dol@zurich.ibm.com>
2025-06-02 17:01:06 +02:00

82 lines
2.0 KiB
Python

from enum import Enum
from typing import Any, Dict, List, Literal
from pydantic import AnyUrl, BaseModel
from typing_extensions import deprecated
from docling.datamodel.accelerator_options import AcceleratorDevice
class BaseVlmOptions(BaseModel):
kind: str
prompt: str
class ResponseFormat(str, Enum):
DOCTAGS = "doctags"
MARKDOWN = "markdown"
HTML = "html"
class InferenceFramework(str, Enum):
MLX = "mlx"
TRANSFORMERS = "transformers"
class TransformersModelType(str, Enum):
AUTOMODEL = "automodel"
AUTOMODEL_VISION2SEQ = "automodel-vision2seq"
AUTOMODEL_CAUSALLM = "automodel-causallm"
class InlineVlmOptions(BaseVlmOptions):
kind: Literal["inline_model_options"] = "inline_model_options"
repo_id: str
trust_remote_code: bool = False
load_in_8bit: bool = True
llm_int8_threshold: float = 6.0
quantized: bool = False
inference_framework: InferenceFramework
transformers_model_type: TransformersModelType = TransformersModelType.AUTOMODEL
response_format: ResponseFormat
supported_devices: List[AcceleratorDevice] = [
AcceleratorDevice.CPU,
AcceleratorDevice.CUDA,
AcceleratorDevice.MPS,
]
scale: float = 2.0
temperature: float = 0.0
stop_strings: List[str] = []
extra_generation_config: Dict[str, Any] = {}
use_kv_cache: bool = True
max_new_tokens: int = 4096
@property
def repo_cache_folder(self) -> str:
return self.repo_id.replace("/", "--")
@deprecated("Use InlineVlmOptions instead.")
class HuggingFaceVlmOptions(InlineVlmOptions):
pass
class ApiVlmOptions(BaseVlmOptions):
kind: Literal["api_model_options"] = "api_model_options"
url: AnyUrl = AnyUrl(
"http://localhost:11434/v1/chat/completions"
) # Default to ollama
headers: Dict[str, str] = {}
params: Dict[str, Any] = {}
scale: float = 2.0
timeout: float = 60
concurrency: int = 1
response_format: ResponseFormat