In September 2026 I shared three LinkedIn posts highlighting recent AI‑system contributions: enabling TPU execution for Diffusers via PT/XLA, releasing QwenImage 2.1 with KV‑caching for both text and reference image tokens, and adding a portable Helion attention kernel to 🤗 Kernels that outperforms FlashAttention on H100.
I adopted PT/XLA integration in Diffusers to enable TPU execution, expanding deployment options beyond GPUs.
QwenImage 2.1 is a 7‑billion‑parameter DiT model that supports T2I and TI2I, implements KV‑caching for both text and reference image tokens, and delivers a ~2.5× speedup at the cost of higher memory usage.
The Helion attention kernel, written in portable Python and shared on the Hub, is on average 1.2× faster than PyTorch’s FlashAttention on H100 across 19 tuned shapes.
- Collaborated with the Qwen team for the model release.
- Worked with the Helion team at Meta on the kernel and co‑authored a blog post.
- Special thanks to Dunfan Lu for the partnership.