pub fn gemm_many_keep(
model: &Arc<CmfModel>,
idxs: &[(usize, usize)],
xs: &[f32],
b: usize,
cols: usize,
out: &mut [f32],
) -> boolExpand description
Several projections of one host activation, read back in one trip:
out = [P0 | P1 | …], each idxs entry (tensor, rows). wgpu only.