Autoregressive Video Chunk FPS Calculator

Modern real time video generation models such as Helios operate autoregressively : they produce video one chunk at a time, where each chunk is a fixed number of frames. Within each chunk, the model runs a diffusion denoising loop for a fixed number of steps. Understanding the relationship between chunk size, denoising steps, per step latency, and context encoding overhead is essential for predicting whether a model can meet a real time FPS target. Write a function compute video generation fps that calculates the end to end throughput of an autoregressive video generation pipeline. The function should accept the following parameters: num chunks (int): Total number of video chunks to generate chunk frames (int): Number of frames produced per chunk denoising steps (int): Number of diffusion denoising steps run per chunk time per step ms (float): Wall clock time in milliseconds for a single denoising step context encoding ms (float, default 0.0): Additional overhead in milliseconds per chunk for encoding historical context (compressed history tokens, VAE encoding, etc.) realtime fps threshold (float, default 24.0): The FPS value at or above which generation is considered real time The function should return a dictionary with the following keys: total frames (int): Total number of frames generated across all chunks total time ms (float): Total wall clock generation time in milliseconds total time s (float): Total time in seconds, rounded to 4 decimal places fps (float): Frames per second, rounded to 2 decimal places time per chunk ms (float): Time spent generating a single chunk in milliseconds is realtime (bool): True if fps = realtime fps threshold