CUDAExecutionProvider may silently fall back to CPU - recommend using tuple format

[kelunyang]

Summary

When using --device cuda, ONNX Runtime may silently fall back to CPU because the provider is specified as a simple string list instead of tuple format with device configuration.

Background

I'm a high school teacher from Taiwan. I was using Claude (AI assistant) to write a Google Colab notebook for OCR-ing pre-war Japanese official documents (戰前日文公文). While processing a 1400-page PDF, I noticed that GPU VRAM usage was suspiciously low (0 MB) despite specifying --device cuda. After investigation, Claude found the issue in the provider configuration.

Environment

• Google Colab (Tesla T4)
• onnxruntime-gpu 1.24.2
• CUDA 12.8
• Python 3.12

Problem

Current code in deim.py (line 38) and parseq.py (line 32):

providers = ['CUDAExecutionProvider','CPUExecutionProvider']

This format can cause ONNX Runtime to silently fall back to CPU without any warning.

Suggested Fix

Use tuple format with explicit device configuration:
providers = [('CUDAExecutionProvider', {'device_id': 0}), 'CPUExecutionProvider']

How We Discovered This

1. Ran OCR with --device cuda
2. Checked GPU memory with nvidia-smi → showed 0 MB used
3. Applied the fix above → GPU memory increased to 183 MB
4. Confirmed with session.get_providers() that CUDAExecutionProvider was actually active

References

- https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html

Thank you for developing this excellent tool for historical document recognition!

---