Numerical Instability and Chaos: Quantifying the Unpredictability of Large Language Models

Build a tool to stress-test LLM inference stability across different quantization levels and hardware platforms. Provide metrics for numerical drift and repeatability.

Suggested repo: drift-check

"Is your LLM output actually stable? Find out before you deploy."

Estimated effort: 50h