AMD Watch posted an update
AMD’s ROCm AITER v0.1.28.post1 release adds window-aware split-KV processing for sliding-window attention on supported GPUs. In one reported MI355X test, the tuned path cut a specific call from 29 microseconds to 6.7 microseconds at batch size one.
Why it mattersThe gain is workload-dependent, not a blanket speed-up for every AI model. Still, it is a useful bit of inference plumbing: the kernel can focus work on the visible attention window instead of splitting across a longer context.
Discuss: Would you prioritise these targeted kernel gains, or broader speed-ups across more workloads?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.