torch.compile and Diffusers: A Hands-On Guide to Peak Performance - Sayak Paul, Hugging Face

This talk covers the integration of torch.compile with the Diffusers library to enhance the performance of diffusion models like Flux-1-Dev. The speaker demonstrates practical techniques for both model authors and users, focusing on making models compiler-friendly and implementing regional compilation to significantly reduce compile times while maintaining runtime performance. Real-world scenarios are discussed, including optimizing resource use on memory-constrained GPUs through CPU offloading and quantization, as well as the efficient swapping of LoRA adapters without triggering recompilation. This session provides valuable insights for anyone involved in building or utilizing diffusion models.