Notes on local inference, GPU kernels, and Apple silicon. I write these to record what I measured and why a change works, before the change leaves my machine.

• One weight read per verify pass: wider affine row tiles in MLX

  October 18, 2026

  Every qmv_wide tile re-reads the quantized weights. Raising the affine tile cap from 5 to 8 rows removes a full extra read at exactly the row counts a speculative verifier uses.