Accelerating SDXL Inference with JAX and TPU v5e

Hugging Face Diffusers now supports JAX, enabling accelerated Stable Diffusion XL (SDXL) inference on Cloud TPU v5e environments.

By leveraging JAX's JIT (Just-In-Time) compilation feature, TPU binaries optimized for image generation tasks with static shapes can be created, achieving very fast inference speeds.

Additionally, pmap based on the XLA compiler allows tasks to be distributed in parallel across multiple TPU chips. For example, using 8 chips can generate 8 images within the same time it takes to generate a single image on one chip, ensuring high throughput.

Cloud TPU v5e offers less than half the cost of TPU v4 while providing optimized performance for training and inference of large-scale AI models, enabling efficient deployment of large models like SDXL.