automl/promptolution

Supports multiple state-of-the-art prompt optimization algorithms (CAPO, EvoPrompt, OPRO) with a unified LLM backend spanning API-based models, local inference via vLLM/transformers, and cluster deployments. Built-in response caching, parallelized inference, and detailed token tracking enable cost-efficient, reproducible large-scale experiments. Decomposes optimization into modular components—Task, Predictor, LLM, and Optimizer—allowing researchers to customize any stage without rigid abstractions.

114 stars.