Skip to main content

GEMV_MMA_MIN_ROWS

logicaffeine_compile::codegen_cuda::lane::kernels

Constant GEMV_MMA_MIN_ROWS 

Source 
pub const GEMV_MMA_MIN_ROWS: usize = 2;
Expand description

The fewest rows a batch’s Q8_0 projection takes to the tensor cores (lk_gemv_mma); fewer run on the GEMV.