Stable Diffusion XL inference on Cloud TPU v5e via Diffusers JAX integration
AI Impact Summary
Diffusers now enables serving Stable Diffusion XL (SDXL) using JAX on Cloud TPU v5e, combining FlaxStableDiffusionXLPipeline with XLA-backed parallelism to scale across TPU chips. On a TPU v5e-4 setup, the demo shows four 1024×1024 images produced in about 4 seconds, with actual generation around 2.3 seconds, illustrating strong throughput and latency improvements at significantly lower hardware cost compared to GPU-based inference. The approach relies on JIT compilation with fixed input shapes and uses pmap to spread work across devices; designs must fix batch size and image dimensions to avoid recompilation and preserve performance, or risk slower startup times and potential quality/latency tradeoffs.
Affected Systems
- Date
- Date not specified
- Change type
- capability
- Severity
- info