PyTorch blog: optimizing Jagged Flash Attention on Blackwell with TLX

On October 2, 2026, the PyTorch Blog published a post on using TLX to optimize Jagged Flash Attention (JFA) on NVIDIA Blackwell (B200). The post identifies JFA as the attention kernel behind Meta's Generative Ads Model (GEM), placing the optimization work in the context of that model's attention implementation. According to the post, the write-up uses TLX to describe the current state of the work and lays out a path toward SOTA FA4.

The piece opens with a TL;DR that summarizes the subject of the work, the target hardware platform, and the optimization framework, before moving into the main body. The opening line states that the post presents work on Jagged Flash Attention (JFA) — the attention kernel behind Meta's Generative Ads Model (GEM) — on NVIDIA Blackwell (B200), built with the TLX approach.

As reported so far, the post is a progress update rather than a finished result. The available material does not include performance figures, benchmark scores, version numbers, or a schedule for completing the path to SOTA FA4. It also does not state what prompted the optimization effort or how it compares with prior implementations of JFA. The current state of the story is simply the publication of the PyTorch Blog post itself, with the description of the JFA optimization work on Blackwell and the route toward SOTA FA4 resting on that blog's own account.