Shmoosh: a faster packed-attention path
An opt-in TF32 mode makes Shmoosh’s packed-attention path faster while keeping the fidelity policy explicit. Here are the measurements and the remaining tradeoffs.
Lynn Cole / 标签
A thread through the work.
An opt-in TF32 mode makes Shmoosh’s packed-attention path faster while keeping the fidelity policy explicit. Here are the measurements and the remaining tradeoffs.
Shmoosh’s early speed gains are modest, but its attention policies expose useful control over image quality and performance.