Comment on: Engineering findings: K/V norm disparity + MSE > Prod + outlier mixed precision

this is great work, thanks for sharing. the K/V norm disparity data across models is something we hadn't quantified — 182x ratio on Qwen2.5-1.5B is wild. that directly informs the head_dim=128 quality gap we've been chasing. the MSE vs Prod finding for keys is interesting too. we dropped QJL early on (MSE-only approach) and buun's CUDA experiments independently confirmed it — your GPT-2 numbers showing Prod at +300% PPL vs MSE at +7.6% add more evidence that QJL variance gets amplified through softmax. dynamic outlier detection with per-layer RMS thresholding is a smarter approach than the fixed 32/96 split. getting Qwen2.5-1.5B to 3.6-bit avg at +2.1% PPL vs +78% uniform is a massive improvement. we'll take a closer look at your repo. appreciate the contribution.