Banana (@banana_baeee)

Finally broke the 3k token per second input/prompt processing barrier for Qwen 3.5 27B on Spark/GB10 thanks to FlashQLA! Results and steps to reproduce up on @LottoLabs LocalMaxxing here: <a target="_blank" href="https://www.localmaxxing.com/runs/cmouqgx9q00jtld01ajgiran7" color="blue">localmaxxing...