AttributeError when enabling formula enrichment (do_formula_enrichment=True)

[b-g-d]

Bug: AttributeError when enabling formula enrichment

Description

When do_formula_enrichment=True is set in PdfPipelineOptions, Docling fails with:

AttributeError: 'dict' object has no attribute 'model_type'

Error Details

Full Traceback:

File "/.../docling/pipeline/standard_pdf_pipeline.py", line 456, in _init_models
    CodeFormulaModel(
File "/.../docling/models/code_formula_model.py", line 108, in __init__
    self._processor = AutoProcessor.from_pretrained(
File "/.../transformers/tokenization_utils_base.py", line 2419, in _from_pretrained
    if _is_local and _config.model_type not in [
AttributeError: 'dict' object has no attribute 'model_type'

Root Cause

The issue occurs in docling/models/code_formula_model.py line 108:

self._processor = AutoProcessor.from_pretrained(artifacts_path,)

When artifacts_path (a Path object) is passed to AutoProcessor.from_pretrained(), transformers loads the tokenizer config as a dict from JSON, but then tries to access _config.model_type as an object attribute (line 2419 in transformers/tokenization_utils_base.py).

Why it fails:

• Loading from local Path: Config loaded as dict → AttributeError when accessing .model_type
• Loading from model name: Config properly converted to config object → Works correctly

Environment

• Docling version: 2.63.0 (latest)
• Transformers version: 4.57.2
• Python version: 3.12
• OS: macOS (darwin 24.6.0)
• Device: MPS (Apple Silicon)

Steps to Reproduce

from docling.datamodel.base_models import InputFormat
from docling.datamodel.pipeline_options import PdfPipelineOptions, TesseractOcrOptions
from docling.document_converter import DocumentConverter, PdfFormatOption
import os
import subprocess

# Set TESSDATA_PREFIX if needed
if 'TESSDATA_PREFIX' not in os.environ:
    tesseract_prefix = subprocess.run(
        ['brew', '--prefix', 'tesseract'],
        capture_output=True, text=True, check=True
    ).stdout.strip()
    if tesseract_prefix:
        os.environ['TESSDATA_PREFIX'] = f'{tesseract_prefix}/share/tessdata'

# This configuration fails
pipeline_options = PdfPipelineOptions(
    ocr_options=TesseractOcrOptions(lang=['eng'], force_full_page_ocr=True),
    do_formula_enrichment=True,  # ← This triggers the error
)

converter = DocumentConverter(
    format_options={InputFormat.PDF: PdfFormatOption(pipeline_options=pipeline_options)}
)

# Fails here
result = converter.convert('document.pdf')

Expected Behavior

Formula enrichment should initialize successfully and extract formulas from the document.

Actual Behavior

Initialization fails with AttributeError before any document processing occurs.

Workaround

A monkey patch can work around the issue by intercepting AutoProcessor.from_pretrained() and converting Path objects to model names:

from transformers import AutoProcessor
from pathlib import Path

original_from_pretrained = AutoProcessor.from_pretrained

def patched_from_pretrained(model_name_or_path, **kwargs):
    if isinstance(model_name_or_path, Path):
        path_str = str(model_name_or_path)
        if 'CodeFormulaV2' in path_str:
            return original_from_pretrained('docling-project/CodeFormulaV2', **kwargs)
    return original_from_pretrained(model_name_or_path, **kwargs)

AutoProcessor.from_pretrained = patched_from_pretrained

Suggested Fix

In docling/models/code_formula_model.py line 108, change:

# Current (fails):
self._processor = AutoProcessor.from_pretrained(artifacts_path,)

# Suggested fix:
self._processor = AutoProcessor.from_pretrained('docling-project/CodeFormulaV2',)

The transformers library will automatically use the cached model, so there's no need to pass the local path. This avoids the transformers bug while maintaining the same functionality.

Additional Context

• The model docling-project/CodeFormulaV2 loads successfully when using the model name directly
• The issue appears to be a compatibility problem between how Docling passes paths and how transformers 4.57.2 handles local path loading
• Related issues: Bug Report: LaTeX Formula Spacing Issue with do_formula_enrichment=True #2374, Bug: do_formula_enrichment=True produces garbled text (e.g., /C0 apod) and generate_picture_images=True creates empty folders & `` placeholders #2568 (formula enrichment issues, but may be different bugs)

[dosubot]

The AttributeError happens because transformers 4.57.2 loads the config as a plain dict when given a local Path, but expects a config object with a .model_type attribute. This is a known bug when using AutoProcessor.from_pretrained with a local path for the CodeFormula model in Docling. Using the model name (e.g., 'docling-project/CodeFormulaV2') instead of the local path is a safe and effective fix—transformers will use the cached model if available, so you don't lose offline capability or performance, and it avoids the error entirely. This approach is consistent with Hugging Face's recommended usage and is already used in other parts of Docling for model downloading and caching source.

This bug and similar model loading issues have been discussed in several Docling issues and threads, confirming it's a recurring problem with formula enrichment and transformers integration source. The fix does not introduce side effects for offline or cluster deployments, as the model cache is still used and the artifacts_path logic remains compatible with other Docling models source.

If you need robust offline or cluster support, make sure to pre-download all required models and set the DOCLING_SERVE_ARTIFACTS_PATH environment variable to your models directory. For production, mounting a persistent volume with pre-downloaded models is recommended source.

There are no official docs recommending local path usage over model names for transformers; using the model name is the most future-proof and reliable approach for this scenario.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}