Wrote up something I've been working on for a while: profiling Stable Diffusion 3 inference with nsys.

Wrote up something I've been working on for a while: profiling Stable Diffusion 3 inference with nsys. It covers how the SD3 pipeline works end to end (text encoders, scheduler, DiT, VAE), how diffusion inference differs from autoregressive LLM inference, and what the nsys traces reveal at a deeper level about how the CPU and GPU split the work. Full post here if you're into inference internals: https://shivamguptaxia2.substack.com/p/understanding-diffusion-inference