cublas_scale_layout_kmajor

metrale_model_layers::layers::ops

Function cublas_scale_layout_kmajor 

Source 
pub fn cublas_scale_layout_kmajor() -> bool
Expand description

2026-09-25: Which VEC128 activation-scale layout the cuBLASLt block-scaled arm passes, from METRALE_CUBLAS_SCALE_LAYOUT, read once per process.

  • true (unset, or any value but rowmajor): [K/128, ceil16(M)], tokens contiguous, the layout metrale_gpu_runtime::cublaslt::scale_layout quotes from the cuBLAS manual.
  • false (rowmajor): the quantizer’s [M, K/128] passed as is. It is a measurement control: measured 2026-09-11 on H100, it gave rel_rms 7.7e-2 against the in-tree kernel.