Random Pytorch Error

[egesnr]

@dosu The code was executing yesterday but now gives a pytorch error:
The code:
import fitz
import pandas as pd
from docling.document_converter import DocumentConverter, PdfFormatOption
from docling.datamodel.base_models import InputFormat
from docling.datamodel.pipeline_options import PdfPipelineOptions, TableFormerMode

from docling.backend.pypdfium2_backend import PyPdfiumDocumentBackend
import os
os.environ["DOCLING_SERVE_ARTIFACTS_PATH"] = "docling_offline"

def docling_table_to_df(table):
grid = table.data.grid
rows = []
for row in grid:
rows.append([cell.text for cell in row])
df = pd.DataFrame(rows)
return df

def extract_pdf_page_tables(name, pdf_path, page_number):
doc = fitz.open(pdf_path)
temp_path = f"tmp_page_{name}_{page_number}.pdf"
new_doc = fitz.open()
new_doc.insert_pdf(doc, from_page=page_number-1, to_page=page_number-1)
new_doc.save(temp_path)
new_doc.close()
doc.close()
pipeline_options = PdfPipelineOptions(
artifacts_path = "docling_offline"
)
pipeline_options.do_ocr = False

#Convert the document
converter = DocumentConverter(
    format_options={
        InputFormat.PDF: PdfFormatOption(
            #backend = PyPdfiumDocumentBackend,
            pipeline_options=pipeline_options)
    }
)

result = converter.convert(temp_path)
docling_doc = result.document
dfs = [docling_table_to_df(t) for t in docling_doc.tables]
return dfs

The error:

RuntimeError Traceback (most recent call last)
Cell In[8], line 3
1 import fitz
2 import pandas as pd
----> 3 from docling.document_converter import DocumentConverter, PdfFormatOption
4 from docling.datamodel.base_models import InputFormat
5 from docling.datamodel.pipeline_options import PdfPipelineOptions, TableFormerMode

RuntimeError: Only a single TORCH_LIBRARY can be used to register the namespace prims; please put all of your definitions in a single TORCH_LIBRARY block. If you were trying to specify implementations, consider using TORCH_LIBRARY_IMPL (which can be duplicated). If you really intended to define operators for a single namespace in a distributed way, you can use TORCH_LIBRARY_FRAGMENT to explicitly indicate this. Previous registration of TORCH_LIBRARY was registered at /data/envs/onprem_models312/lib/python3.12/site-packages/torch/_prims/init.py:37; latest registration was registered at /data/envs/onprem_models312/lib/python3.12/site-packages/torch/_prims/init.py:37

[dosubot[bot]]

This PyTorch error is almost always caused by a conflict or corruption in your Python environment—typically from incompatible versions of torch, transformers, or other dependencies, or from multiple torch installations. Docling itself does not register custom PyTorch operators, so the issue is not in your code or in docling, but in the environment or a third-party package being loaded at import time.

This can happen if, for example, transformers was recently upgraded (especially to >=4.52.0), or if click or torch was upgraded/downgraded, or if you have multiple torch installations (system + venv, or conda + pip). It can also be triggered by custom or third-party PyTorch extensions.

Here's what to do:

1. Run pip freeze and check the versions of torch, transformers, numpy, docling, and click. If transformers is >=4.52.0, downgrade it to 4.44.2 or 4.42.0:

   pip install 'transformers<4.52.0'
   

   This has resolved similar issues for other users (source).

2. Make sure you only have one torch installation (no system/conda conflicts).

3. If you recently upgraded any packages, OS, or Python, try creating a fresh virtual environment and reinstall only the required packages (docling, torch, transformers, numpy, etc.). This often fixes environment corruption (source).

4. If you use custom or third-party PyTorch extensions, try removing or updating them.

Let me know the output of pip freeze and if any packages or the OS were recently changed, and I can help further debug.

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.

[egesnr]

The problem I forget to add offline modelling path so there was a conflict when I restart the kernel problem has fixed