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

* 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>

docling/datamodel/base_models.py

@@ -253,11 +253,18 @@ class Page(BaseModel):
             return []
 
     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]:
         if self._backend is 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 cropbox is None:
                 self._image_cache[scale] = self._backend.get_page_image(scale=scale)

docling/datamodel/pipeline_options_vlm_model.py

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

docling/models/api_vlm_model.py

@@ -48,7 +48,9 @@ class ApiVlmModel(BasePageModel):
                 with TimeRecorder(conv_res, "vlm"):
                     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
                     if hi_res_image:
                         if hi_res_image.mode != "RGB":

docling/models/vlm_models_inline/hf_transformers_model.py

@@ -123,7 +123,9 @@ class HuggingFaceTransformersVlmModel(BasePageModel, HuggingFaceModelDownloadMix
                 with TimeRecorder(conv_res, "vlm"):
                     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
                     prompt = self.formulate_prompt()

docling/models/vlm_models_inline/mlx_model.py

@@ -73,7 +73,9 @@ class HuggingFaceMlxModel(BasePageModel, HuggingFaceModelDownloadMixin):
                 with TimeRecorder(conv_res, f"vlm-mlx-{self.vlm_options.repo_id}"):
                     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:
                         im_width, im_height = hi_res_image.size