rolv clocks Llama 4 Scout attention

This is a useful benchmark because it moves rolv from giant sparse MoE layers into a smaller, more deployment-relevant attention path where brute-force dense kernels already have less room to lose.

• –The headline result is not the raw 8.3x mean speedup, but that the same canonical hash held across every one of Scout’s 32 attention layers with zero failures.
• –Attention QKV projections are structurally different from the MoE FFN layers rolv highlighted before, so this broadens the claim that the method generalizes across transformer components.
• –The lower gain versus prior 55-82x MoE numbers is believable on its face because the matrix here is much smaller, making it a better stress test of whether the technique still matters outside blockbuster sparse cases.
• –For AI infra readers, this lands as a benchmark story more than a product launch: promising if reproducible, but still best treated as vendor-published performance evidence rather than settled consensus.