CUDA OOM with small PDF file

Question

I try to convert a small PDF file (vietnamese) to markdown. It works perfectly on the web version but fails when I switch to use local install with GPU (on Google Colab).

Environment

• docstrange version: 1.1.8
• GPU: Tesla T4 / 14.74 GiB VRAM
• CUDA version: 12.6
• PyTorch version: 2.9.0+cu126

Description

Getting CUDA out of memory error when processing a small PDF file (801KB, 17 pages).

Error Message

ERROR:docstrange.pipeline.nanonets_processor:Nanonets OCR extraction failed: 
CUDA out of memory. Tried to allocate 122.64 GiB. GPU 0 has a total capacity 
of 14.74 GiB of which 8.54 GiB is free.

PDF Details

• File size: 801 KB
• Pages: 17
• Page dimensions: 1675 x 2353 pixels
• Type: Scanned images

Code to Reproduce

!apt-get install poppler-utils

!pip install docstrange -q


from docstrange import DocumentExtractor

extractor = DocumentExtractor(gpu=True)
result = extractor.extract("sample.pdf")
markdown = result.extract_markdown()

[linear]

PLG-322

[EricBorland]

I also have the same issue. I wonder if there's a guide or recommended hardware specifications to run in local...