Idea's on how to get past pin_memory warnings

[xil3]

The majority of the documents it converts work fine, but certain PDFs are presented with constant pin_memory warnings.

dataloader.py:665: UserWarning: 'pin_memory' argument is set as true but no accelerator is found, then device pinned memory won't be used.

At that point, I have to ctrl+c to stop it.

From what I've observed, PDF documents that have an image in them seem to trigger that.

Any ideas?

Here's the snippet of code used to convert:

for file in files:
          tempUploadedFile = NamedTemporaryFile(delete=False)

          try:
              contents = file.file.read()
              with tempUploadedFile as f:
                  f.write(contents)
          except Exception:
              raise HTTPException(status_code=500, detail='Error on uploading the file')
          finally:
              file.file.close()

          temp_file_paths.append(tempUploadedFile.name)

      pipeline_options = PdfPipelineOptions()
      pipeline_options.generate_page_images = True
      pipeline_options.do_ocr = True
      pipeline_options.accelerator_options = AcceleratorOptions(
          num_threads=8, device=AcceleratorDevice.CPU
      )

      doc_converter = DocumentConverter(
          format_options={
              InputFormat.PDF: PdfFormatOption(
                  pipeline_options=pipeline_options, backend=DoclingParseV4DocumentBackend
              )
          }
      )

      print("Starting to convert docs", flush=True)
      conv_results = doc_converter.convert_all(temp_file_paths)
      print("Finished converting docs", flush=True)

      for result in conv_results:
          if result.document:
              # Access the converted document object
              print("Looping through converted results", flush=True)
              converted_docs.append(result.document.export_to_text())
              # Save or process the output as needed
          else:
              print(f"Error converting {result.source_path}: {result.error}", flush=True)

      print("Returning converted docs", flush=True)

      return converted_docs

[xil3]

Tested a few others that had images, and some worked, so it's not just PDFs with images that failed.

[dosubot[bot]]

Hi @xil3! I'm Dosu and I’m helping the docling team.

It's not just images—other factors like file size, page count, PDF encoding, or even subtle corruption can trigger these warnings and hangs. The warning itself comes from PyTorch's DataLoader when it tries to use pinned memory on CPU-only systems, but Docling doesn't set this directly; it's likely coming from a dependency used by one of the models (like OCR or layout) in the pipeline.

Since some image PDFs work and others don't, try comparing the problematic and working files for differences in size, number of pages, or image types. You can also try disabling OCR or image generation in your pipeline options to see if a specific step is causing the issue. If you find a pattern (like only very large or high-resolution PDFs hang), splitting the PDFs or processing them in smaller page ranges can help manage resource usage and avoid hangs see related discussion.

If you want to dig deeper, running with debug logging or isolating the step that triggers the warning can help pinpoint which dependency is at fault. Let me know if you want tips on how to do that.

To reply, just mention @dosu.

---

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

[manikrishna-m]

Hi everyone,

I'm using EasyOCR on CPU, and everything is working fine — the PDF-to-text conversion is accurate. However, I’m seeing this warning message once in the terminal:

UserWarning: 'pin_memory' argument is set as true but no accelerator is found, then device pinned memory won't be used.

I'm a bit puzzled since I'm not using a GPU, and the warning only appears once during execution. Despite the warning, the OCR works as expected — I just want to understand why this is happening and whether it’s something I should worry about.

Has anyone else encountered this or know why this might be happening?

Thanks in advance!

[zaidurrehman]

I turned of OCR (set pipeline_options.do_ocr = False) and avoided these warning messages. There are only logos in my PDFs, so I don't need the information from images in my case. May be this helps in your use cases as well.

[teachamantofish]

I used this code to remove those errors:

# disable pinned memory to avoid the errors
torch.backends.cuda.matmul.allow_tf32 = False  # optional tweak
# monkeypatch dataloader defaults
from torch.utils import data
old_init = data.DataLoader.__init__
def new_init(self, *args, **kwargs):
    kwargs["pin_memory"] = False
    old_init(self, *args, **kwargs)
data.DataLoader.__init__ = new_init

[dolfim-ibm]

This was addressed in #2237.