In September 2026 I posted a series of X updates about recent work with 🤗 Diffusers, HF Jobs, and 🤗 Kernels, covering a one‑shot video generation demo, benchmark reporting, TPU execution, KV‑caching speedups in Qwen‑Image 2.1, Helion attention kernels, and the Flash‑BoN wall‑clock evaluation metric. See also my LinkedIn recap linkedin posts september 2026.
Astra can really one‑shot 😨 – I asked it to produce a launch video for the latest Diffusers release and it generated a complete video from a single prompt.
With HF Jobs making lives easier, it's almost a sin for me not to report a benchmark that uses HF Jobs for any PR related to performance, even slightly.
The CLI is quite optimized for the agents to get things up and running quickly.
This summer, I presented how we enable TPU execution in Diffusers through PT/XLA at the AI Systems DevLabs, organized by Google.
On a warmed A100 with 2 refs (each being 1024×1024), 40 steps, DiT time fell 50.57s -> 19.86s (2.55× speedup).
Tradeoff is increase in the memory 32.3G -> 36.0G.