ONNX Inference Outputting Gibberish?
#46
by
buckeye17-bah
- opened
Since I've struggled to get SmolDocling working with Transformers in a Windows-based Python environment, I've switched to running it in a Docker container using Docling's own image (ghcr.io/docling-project/docling-serve-cu124). I'm using the code provided on the model card to run it via ONNX. When I input two example documents mentioned by Docling (GazetteFrance.jpg & page_with_table.png), both documents yield gibberish like the following:
A<fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image><fake_token_around_image><image>
Any suggestions on how to fix this? The model I'm running was obtained by running the command docling-tools models download smolvlm. Is it possible this has downloaded an incompatible version?