[BUG: Output] marker with llm is not working properly

[ashunaveed]

Got it

[ocg2347]

If you do not use gemini, I advice you to look at this:

https://github.com/datalab-to/marker/issues/785#issuecomment-3035236316

[alexdmccabe]

I'm seeing this too. I checked, and it happens in 1.8.1 and 1.8.0, but it does not happen in 1.7.5.

I'm on Windows 11, using the marker CLI in WSL 2, RHEL 9.5, Python 3.11. However, LLMs are running natively on Windows in Ollama.

Ollama produces a HTTP 500 error, but using Ollama's OpenAI API produces a HTTP 400 error, instead.

I also think this might be the same as https://github.com/datalab-to/marker/issues/621 and possibly https://github.com/datalab-to/marker/issues/606.

Edit: Using the Ollama API fails either way, but the Ollama OpenAI API only fails if there are images in the PDF.

LLMTableProcessor running: 0it [00:00, ?it/s]2025-07-15 18:09:48,585 [ERROR] marker: OpenAI inference failed: Error code: 400 - {'error': {'message': 'invalid image input', 'type': 'invalid_request_error', 'param': None, 'code': None}}

Edit 2: The Ollama OpenAI API fails in marker 1.7.5, too.

So, that just doesn't work (which is a secondary concern, I think), but the Ollama API stopped working with images after 1.7.5.

Edit 3: It's not just Ollama's OpenAI API - using LM Studio's OpenAI API with marker produces this error on marker 1.7.5, 1.8.0, and 1.8.1:

LLMTableProcessor running: 0it [00:00, ?it/s]2025-07-15 23:04:09,561 [ERROR] marker: OpenAI inference failed: Error code: 400 - {'error': "'url' field must be a base64 encoded image."}