Skip to content

FAQ & Troubleshooting

General

What models does Axetract support?

Axetract is optimized for the Qwen family of models, specifically Qwen3-0.6B. However, the BaseLLMClient can be extended to support any transformers or vLLM compatible model.

Can I run this without a GPU?

Yes, using the HuggingFaceClient. However, inference with even small models like 0.6B will be significantly slower on a CPU.

Common Issues

"CUDA Out of Memory"

If you are using vLLM, try reducing gpu_memory_utilization in the config: 

config = {"engine_args": {"gpu_memory_utilization": 0.5}}
pipeline = AXEPipeline.from_config(llm_config=config, use_vllm=True)

Path resolution errors

Ensure you are running the server or scripts from the root of the project, or that the src directory is in your PYTHONPATH.

Low extraction accuracy

  • Ensure your query is specific.
  • If using a schema, ensure the field names are descriptive.
  • Large documents might require a higher max_tokens setting.

LoRA Issues

Adapter not found

Ensure you have an active internet connection to download the adapters from Hugging Face on the first run. If you are in an air-gapped environment, pre-download the models and point to the local paths in the llm_config.