NeoAnd/Docling
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: Maximum image size for Vlm models (#1802)

Browse Source 
* Image scale moved to base vlm options.
Added max_size image limit (options and vlm models).

* DCO Remediation Commit for Shkarupa Alex <shkarupa.alex@gmail.com>

I, Shkarupa Alex <shkarupa.alex@gmail.com>, hereby add my Signed-off-by to this commit: e93602a0d02fdb6f6dea1f65686cffcc4c616011

Signed-off-by: Shkarupa Alex <shkarupa.alex@gmail.com>

---------

Signed-off-by: Shkarupa Alex <shkarupa.alex@gmail.com>
This commit is contained in:
Shkarupa Alex 2025-06-18 13:57:37 +03:00 committed by GitHub
parent dbab30e92c
commit 215b540f6c
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
5 changed files with 19 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
docling/datamodel/base_models.py
View File

@ -253,11 +253,18 @@ class Page(BaseModel):
return [] return []
def get_image( def get_image(
self, scale: float = 1.0, cropbox: Optional[BoundingBox] = None self,
scale: float = 1.0,
max_size: Optional[int] = None,
cropbox: Optional[BoundingBox] = None,
) -> Optional[Image]: ) -> Optional[Image]:
if self._backend is None: if self._backend is None:
return self._image_cache.get(scale, None) return self._image_cache.get(scale, None)
if max_size:
assert self.size is not None
scale = min(scale, max_size / max(self.size.as_tuple()))
if scale not in self._image_cache: if scale not in self._image_cache:
if cropbox is None: if cropbox is None:
self._image_cache[scale] = self._backend.get_page_image(scale=scale) self._image_cache[scale] = self._backend.get_page_image(scale=scale)

5
docling/datamodel/pipeline_options_vlm_model.py
View File

@ -10,6 +10,8 @@ from docling.datamodel.accelerator_options import AcceleratorDevice
class BaseVlmOptions(BaseModel): class BaseVlmOptions(BaseModel):
kind: str kind: str
prompt: str prompt: str
scale: float = 2.0
max_size: Optional[int] = None
class ResponseFormat(str, Enum): class ResponseFormat(str, Enum):
@ -49,8 +51,6 @@ class InlineVlmOptions(BaseVlmOptions):
AcceleratorDevice.MPS, AcceleratorDevice.MPS,
] ]
scale: float = 2.0
temperature: float = 0.0 temperature: float = 0.0
stop_strings: List[str] = [] stop_strings: List[str] = []
extra_generation_config: Dict[str, Any] = {} extra_generation_config: Dict[str, Any] = {}
@ -76,7 +76,6 @@ class ApiVlmOptions(BaseVlmOptions):
) # Default to ollama ) # Default to ollama
headers: Dict[str, str] = {} headers: Dict[str, str] = {}
params: Dict[str, Any] = {} params: Dict[str, Any] = {}
scale: float = 2.0
timeout: float = 60 timeout: float = 60
concurrency: int = 1 concurrency: int = 1
response_format: ResponseFormat response_format: ResponseFormat

4
docling/models/api_vlm_model.py
View File

@ -48,7 +48,9 @@ class ApiVlmModel(BasePageModel):
with TimeRecorder(conv_res, "vlm"): with TimeRecorder(conv_res, "vlm"):
assert page.size is not None assert page.size is not None
hi_res_image = page.get_image(scale=self.vlm_options.scale) hi_res_image = page.get_image(
scale=self.vlm_options.scale, max_size=self.vlm_options.max_size
)
assert hi_res_image is not None assert hi_res_image is not None
if hi_res_image: if hi_res_image:
if hi_res_image.mode != "RGB": if hi_res_image.mode != "RGB":

4
docling/models/vlm_models_inline/hf_transformers_model.py
View File

@ -123,7 +123,9 @@ class HuggingFaceTransformersVlmModel(BasePageModel, HuggingFaceModelDownloadMix
with TimeRecorder(conv_res, "vlm"): with TimeRecorder(conv_res, "vlm"):
assert page.size is not None assert page.size is not None
hi_res_image = page.get_image(scale=self.vlm_options.scale) hi_res_image = page.get_image(
scale=self.vlm_options.scale, max_size=self.vlm_options.max_size
)
# Define prompt structure # Define prompt structure
prompt = self.formulate_prompt() prompt = self.formulate_prompt()

4
docling/models/vlm_models_inline/mlx_model.py
View File

@ -73,7 +73,9 @@ class HuggingFaceMlxModel(BasePageModel, HuggingFaceModelDownloadMixin):
with TimeRecorder(conv_res, f"vlm-mlx-{self.vlm_options.repo_id}"): with TimeRecorder(conv_res, f"vlm-mlx-{self.vlm_options.repo_id}"):
assert page.size is not None assert page.size is not None
hi_res_image = page.get_image(scale=self.vlm_options.scale) hi_res_image = page.get_image(
scale=self.vlm_options.scale, max_size=self.vlm_options.max_size
)
if hi_res_image is not None: if hi_res_image is not None:
im_width, im_height = hi_res_image.size im_width, im_height = hi_res_image.size