Tokenflood is a tool designed for load testing instruction-tuned large language models (LLMs). It allows users to create various load profiles by specifying parameters such as prompt lengths, prefix lengths, output lengths, and request rates, all without needing specific prompt or response data. This flexibility helps in analyzing how latency varies across different providers, hardware setups, quantizations, and prompt configurations. Built on top of litellm, Tokenflood supports all providers compatible with that framework. The tool is particularly beneficial for assessing latency patterns before deploying models in production, especially with hosted LLM providers. It highlights that latency can significantly increase during peak business hours, as demonstrated with OpenAI's model, which showed a drop of 500-1000 ms once US business hours began. Users are warned about potential high costs when using pay-per-token services, emphasizing the importance of budget management during testing. For installation, users need to set up vllm and serve a small model before initializing Tokenflood. The initial setup generates configuration files that define test parameters. A run suite can consist of multiple phases with varying request rates, allowing for comprehensive testing. Additionally, the observation spec enables continuous monitoring of an endpoint over a specified duration, sending requests at defined intervals. This structured approach helps users optimize LLM performance while controlling costs and understanding the impact of different configurations on latency and throughput.