I’m running a local DeepSeek model through an inference server and testing the same short classification prompt repeatedly. Most runs return the same label, but occasionally one changes. Temperature is set to zero, which I had assumed would make the output repeatable.

Before blaming the model, I want to check whether I’m actually sending identical requests. My test currently keeps these fixed:

• The model file and quantization.
• The system message, user message, and chat template.
• The output token limit and configured sampling options.

The server also handles other requests, so batching may differ between runs. I haven’t yet compared this with a single-request setup.

What should I log to investigate local model output reproducibility? Could numerical differences or the inference backend affect token selection even without random sampling, and would setting a seed address that kind of variation?

I only need stable classification labels, not identical explanatory prose. I’d like to know what level of repeatability is reasonable to expect and how to build a small test that separates configuration mistakes from backend behavior.