feat(vlm): track generated tokens and stop reasons for VLM models (#2543)

* feat: add enum StopReason and use it in VlmPrediction

Signed-off-by: ElHachem02 <peterelhachem02@gmail.com>

* add vlm_inference time for api calls and track stop reason

Signed-off-by: ElHachem02 <peterelhachem02@gmail.com>

* fix: rename enum to VlmStopReason

Signed-off-by: ElHachem02 <peterelhachem02@gmail.com>

* Propagate partial success status if page reaches max tokens

Signed-off-by: ElHachem02 <peterelhachem02@gmail.com>

* feat: page with generation stopped by loop detector create partial success status

Signed-off-by: Peter El Hachem <peter.el.hachem@ibm.com>

* Add hint for future improvement

Signed-off-by: Peter El Hachem <peter.el.hachem@ibm.com>

* fix: remove vlm_stop_reason from extracted page data, add UNSPECIFIED state as VlmStopReason to avoid null value

Signed-off-by: Peter El Hachem <peter.el.hachem@ibm.com>

---------

Signed-off-by: ElHachem02 <peterelhachem02@gmail.com>
Signed-off-by: Peter El Hachem <peter.el.hachem@ibm.com>
Co-authored-by: Peter El Hachem <peter.el.hachem@ibm.com>

docling/datamodel/base_models.py

@@ -166,6 +166,13 @@ class DoclingComponentType(str, Enum):
     USER_INPUT = "user_input"
 
 
+class VlmStopReason(str, Enum):
+    LENGTH = "length"  # max tokens reached
+    STOP_SEQUENCE = "stop_sequence"  # Custom stopping criteria met
+    END_OF_SEQUENCE = "end_of_sequence"  # Model generated end-of-text token
+    UNSPECIFIED = "unspecified"  # Defaul none value
+
+
 class ErrorItem(BaseModel):
     component_type: DoclingComponentType
     module_name: str
@@ -208,7 +215,7 @@ class VlmPrediction(BaseModel):
     generated_tokens: list[VlmPredictionToken] = []
     generation_time: float = -1
     num_tokens: Optional[int] = None
-    stop_reason: Optional[str] = None  # todo define an enum for possible stop reasons
+    stop_reason: VlmStopReason = VlmStopReason.UNSPECIFIED
 
 
 class ContainerElement(

docling/datamodel/extraction.py

@@ -4,7 +4,7 @@ from typing import Any, Dict, List, Optional, Type, Union
 
 from pydantic import BaseModel, Field
 
-from docling.datamodel.base_models import ConversionStatus, ErrorItem
+from docling.datamodel.base_models import ConversionStatus, ErrorItem, VlmStopReason
 from docling.datamodel.document import InputDocument
 
 

docling/models/api_vlm_model.py

@@ -3,7 +3,7 @@ from concurrent.futures import ThreadPoolExecutor
 
 from transformers import StoppingCriteria
 
-from docling.datamodel.base_models import Page, VlmPrediction
+from docling.datamodel.base_models import Page, VlmPrediction, VlmStopReason
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options_vlm_model import ApiVlmOptions
 from docling.exceptions import OperationNotAllowed
@@ -59,6 +59,7 @@ class ApiVlmModel(BasePageModel):
                     hi_res_image = hi_res_image.convert("RGB")
 
                 prompt = self.vlm_options.build_prompt(page.parsed_page)
+                stop_reason = VlmStopReason.UNSPECIFIED
 
                 if self.vlm_options.custom_stopping_criteria:
                     # Instantiate any GenerationStopper classes before passing to streaming
@@ -73,6 +74,7 @@ class ApiVlmModel(BasePageModel):
                         # Skip non-GenerationStopper criteria (should have been caught in validation)
 
                     # Streaming path with early abort support
+                    with TimeRecorder(conv_res, "vlm_inference"):
                         page_tags, num_tokens = api_image_request_streaming(
                             image=hi_res_image,
                             prompt=prompt,
@@ -82,9 +84,11 @@ class ApiVlmModel(BasePageModel):
                             generation_stoppers=instantiated_stoppers,
                             **self.params,
                         )
+                        page_tags = self.vlm_options.decode_response(page_tags)
                 else:
                     # Non-streaming fallback (existing behavior)
-                    page_tags, num_tokens = api_image_request(
+                    with TimeRecorder(conv_res, "vlm_inference"):
+                        page_tags, num_tokens, stop_reason = api_image_request(
                             image=hi_res_image,
                             prompt=prompt,
                             url=self.vlm_options.url,
@@ -94,8 +98,9 @@ class ApiVlmModel(BasePageModel):
                         )
 
                         page_tags = self.vlm_options.decode_response(page_tags)
+
                 page.predictions.vlm_response = VlmPrediction(
-                    text=page_tags, num_tokens=num_tokens
+                    text=page_tags, num_tokens=num_tokens, stop_reason=stop_reason
                 )
             return page
 

docling/models/picture_description_api_model.py

@@ -51,7 +51,7 @@ class PictureDescriptionApiModel(PictureDescriptionBaseModel):
         # Note: technically we could make a batch request here,
         # but not all APIs will allow for it. For example, vllm won't allow more than 1.
         def _api_request(image):
-            response, _ = api_image_request(
+            page_tags, _, _ = api_image_request(
                 image=image,
                 prompt=self.options.prompt,
                 url=self.options.url,
@@ -60,7 +60,7 @@ class PictureDescriptionApiModel(PictureDescriptionBaseModel):
                 **self.options.params,
             )
 
-            return response
+            return page_tags
 
         with ThreadPoolExecutor(max_workers=self.concurrency) as executor:
             yield from executor.map(_api_request, images)

docling/models/vlm_models_inline/hf_transformers_model.py

@@ -13,7 +13,7 @@ from transformers import StoppingCriteria, StoppingCriteriaList, StopStringCrite
 from docling.datamodel.accelerator_options import (
     AcceleratorOptions,
 )
-from docling.datamodel.base_models import Page, VlmPrediction
+from docling.datamodel.base_models import Page, VlmPrediction, VlmStopReason
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options_vlm_model import (
     InlineVlmOptions,
@@ -382,4 +382,5 @@ class HuggingFaceTransformersVlmModel(BaseVlmPageModel, HuggingFaceModelDownload
                 text=decoded_text,
                 generation_time=generation_time,
                 num_tokens=num_tokens,
+                stop_reason=VlmStopReason.UNSPECIFIED,
             )

docling/models/vlm_models_inline/mlx_model.py

@@ -13,7 +13,12 @@ from transformers import StoppingCriteria
 from docling.datamodel.accelerator_options import (
     AcceleratorOptions,
 )
-from docling.datamodel.base_models import Page, VlmPrediction, VlmPredictionToken
+from docling.datamodel.base_models import (
+    Page,
+    VlmPrediction,
+    VlmPredictionToken,
+    VlmStopReason,
+)
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options_vlm_model import InlineVlmOptions
 from docling.models.base_model import BaseVlmPageModel
@@ -319,5 +324,6 @@ class HuggingFaceMlxModel(BaseVlmPageModel, HuggingFaceModelDownloadMixin):
                     generation_time=generation_time,
                     generated_tokens=tokens,
                     num_tokens=len(tokens),
+                    stop_reason=VlmStopReason.UNSPECIFIED,
                 )
             _log.debug("MLX model: Released global lock")

docling/models/vlm_models_inline/nuextract_transformers_model.py

@@ -12,7 +12,7 @@ from transformers import AutoModelForImageTextToText, AutoProcessor, GenerationC
 from docling.datamodel.accelerator_options import (
     AcceleratorOptions,
 )
-from docling.datamodel.base_models import VlmPrediction
+from docling.datamodel.base_models import VlmPrediction, VlmStopReason
 from docling.datamodel.pipeline_options_vlm_model import InlineVlmOptions
 from docling.models.base_model import BaseVlmModel
 from docling.models.utils.hf_model_download import (
@@ -284,6 +284,7 @@ class NuExtractTransformersModel(BaseVlmModel, HuggingFaceModelDownloadMixin):
         # Optional logging
         num_tokens = None
         if generated_ids.shape[0] > 0:  # type: ignore
+            # Todo: confirm num tokens is actually from first item, code was already like this
             num_tokens = int(generated_ids[0].shape[0])
             _log.debug(
                 f"Generated {num_tokens} tokens in {generation_time:.2f}s "
@@ -297,4 +298,5 @@ class NuExtractTransformersModel(BaseVlmModel, HuggingFaceModelDownloadMixin):
                 text=decoded_text,
                 generation_time=generation_time,
                 num_tokens=num_tokens,
+                stop_reason=VlmStopReason.UNSPECIFIED,
             )

docling/models/vlm_models_inline/vllm_model.py

@@ -9,7 +9,12 @@ import numpy as np
 from PIL.Image import Image
 
 from docling.datamodel.accelerator_options import AcceleratorOptions
-from docling.datamodel.base_models import Page, VlmPrediction, VlmPredictionToken
+from docling.datamodel.base_models import (
+    Page,
+    VlmPrediction,
+    VlmPredictionToken,
+    VlmStopReason,
+)
 from docling.datamodel.document import ConversionResult
 from docling.datamodel.pipeline_options_vlm_model import (
     InlineVlmOptions,
@@ -311,13 +316,18 @@ class VllmVlmModel(BaseVlmPageModel, HuggingFaceModelDownloadMixin):
         # Emit predictions
         for output in outputs:
             text = output.outputs[0].text if output.outputs else ""
-            stop_reason = output.outputs[0].stop_reason if output.outputs else ""
+            stop_reason = (
-            generated_tokens = [
+                VlmStopReason.END_OF_SEQUENCE
-                VlmPredictionToken(token=int(p)) for p in output.outputs[0].token_ids
+                if output.outputs[0].stop_reason
-            ]
+                else VlmStopReason.LENGTH
+            )
+            generated_tokens = (
+                [VlmPredictionToken(token=int(t)) for t in output.outputs[0].token_ids]
+                if self.vlm_options.track_generated_tokens
+                else []
+            )
             num_tokens = len(generated_tokens)
             decoded_text = self.vlm_options.decode_response(text)
-            if self.vlm_options.track_generated_tokens:
             yield VlmPrediction(
                 text=decoded_text,
                 generation_time=generation_time,
@@ -325,10 +335,3 @@ class VllmVlmModel(BaseVlmPageModel, HuggingFaceModelDownloadMixin):
                 stop_reason=stop_reason,
                 generated_tokens=generated_tokens,
             )
-            else:
-                yield VlmPrediction(
-                    text=decoded_text,
-                    generation_time=generation_time,
-                    num_tokens=num_tokens,
-                    stop_reason=stop_reason,
-                )

docling/pipeline/extraction_vlm_pipeline.py

@@ -8,7 +8,7 @@ from pydantic import BaseModel
 
 from docling.backend.abstract_backend import PaginatedDocumentBackend
 from docling.backend.pdf_backend import PdfDocumentBackend
-from docling.datamodel.base_models import ConversionStatus, ErrorItem
+from docling.datamodel.base_models import ConversionStatus, ErrorItem, VlmStopReason
 from docling.datamodel.document import InputDocument
 from docling.datamodel.extraction import (
     ExtractedPageData,
@@ -83,6 +83,12 @@ class ExtractionVlmPipeline(BaseExtractionPipeline):
                         # Parse the extracted text as JSON if possible, otherwise use as-is
                         extracted_text = predictions[0].text
                         extracted_data = None
+                        vlm_stop_reason: VlmStopReason = predictions[0].stop_reason
+                        if (
+                            vlm_stop_reason == VlmStopReason.LENGTH
+                            or vlm_stop_reason == VlmStopReason.STOP_SEQUENCE
+                        ):
+                            ext_res.status = ConversionStatus.PARTIAL_SUCCESS
 
                         try:
                             extracted_data = json.loads(extracted_text)
@@ -128,7 +134,11 @@ class ExtractionVlmPipeline(BaseExtractionPipeline):
     def _determine_status(self, ext_res: ExtractionResult) -> ConversionStatus:
         """Determine the status based on extraction results."""
         if ext_res.pages and not any(page.errors for page in ext_res.pages):
-            return ConversionStatus.SUCCESS
+            return (
+                ConversionStatus.PARTIAL_SUCCESS
+                if ext_res.status == ConversionStatus.PARTIAL_SUCCESS
+                else ConversionStatus.SUCCESS
+            )
         else:
             return ConversionStatus.FAILURE
 

docling/utils/api_image_request.py

@@ -8,7 +8,7 @@ import requests
 from PIL import Image
 from pydantic import AnyUrl
 
-from docling.datamodel.base_models import OpenAiApiResponse
+from docling.datamodel.base_models import OpenAiApiResponse, VlmStopReason
 from docling.models.utils.generation_utils import GenerationStopper
 
 _log = logging.getLogger(__name__)
@@ -21,7 +21,7 @@ def api_image_request(
     timeout: float = 20,
     headers: Optional[dict[str, str]] = None,
     **params,
-) -> Tuple[str, Optional[int]]:
+) -> Tuple[str, Optional[int], VlmStopReason]:
     img_io = BytesIO()
     image.save(img_io, "PNG")
     image_base64 = base64.b64encode(img_io.getvalue()).decode("utf-8")
@@ -61,7 +61,13 @@ def api_image_request(
     api_resp = OpenAiApiResponse.model_validate_json(r.text)
     generated_text = api_resp.choices[0].message.content.strip()
     num_tokens = api_resp.usage.total_tokens
-    return generated_text, num_tokens
+    stop_reason = (
+        VlmStopReason.LENGTH
+        if api_resp.choices[0].finish_reason == "length"
+        else VlmStopReason.END_OF_SEQUENCE
+    )
+
+    return generated_text, num_tokens, stop_reason
 
 
 def api_image_request_streaming(