Large Language Models (LLMs) are revolutionizing how we write code, but they have a hidden inefficiency: they often generate excessive, unnecessary code, a problem researchers call "excess token generation." Imagine an LLM writing a simple function, then needlessly adding extra functions or comments, bloating the code and wasting time. This isn't just about longer code; it impacts LLM performance, making them slower and less efficient. Researchers have introduced CodeFast, a clever solution to curb this code bloat. CodeFast acts like a vigilant editor, predicting when an LLM is about to generate unnecessary code and stopping it in its tracks. It's like having a built-in code optimizer that streamlines the output in real-time. The secret sauce is a component called GenGuard, a lightweight addition to the LLM. GenGuard acts as a gatekeeper, learning to distinguish between essential and superfluous code. During code generation, it constantly monitors the LLM's output, predicting the likelihood of excess code at each step. If the likelihood is high, GenGuard signals the LLM to stop, preventing the unnecessary code from being generated. One interesting design choice in CodeFast is the "line-voting mechanism." Instead of stopping the moment GenGuard flags an issue, CodeFast waits until a complete line of code is generated. It then uses majority voting on GenGuard's predictions for that line to decide whether to terminate the generation, adding an extra layer of robustness and reducing false stops. The results? CodeFast significantly accelerates LLM code generation speed without sacrificing quality, making LLMs more efficient and practical for real-world coding tasks. Experiments on various LLMs and programming languages confirm CodeFast's effectiveness. This approach promises more responsive and efficient AI coding tools, eliminating wasteful processing and boosting developer productivity.