Accelerating AI Inference for 3D Creation on Roblox

Implemented CUDA Graphs and KV caching to optimize AI inference for 3D mesh generation, reducing token generation time from 60.5ms to 7.8ms and full object generation from 31 seconds to 4 seconds. This was achieved by profiling the inference pipeline to identify and reduce idle time between operations.