Meta's 3.2K-line Triton JFA Kernel Beats FA4 on Blackwell

Meta's PyTorch team introduced Jagged Flash Attention (JFA), a 3.2K-line TLX-based attention kernel for Blackwell GPUs serving its GEM ads model, running 13% faster in forward and 50% faster in backward than FA4.

2026-10-02 ~ 2026-10-02 · 2 related posts