Function fp8_act_scale_to_kmajor 

pub fn fp8_act_scale_to_kmajor(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    a_scale: DevicePtr,
    a_scale_kmajor: DevicePtr,
    m: u32,
    m_pad: u32,
    k: u32,
    stream: u64,
) -> Result<()>

Rewrite the quantizer’s row-major [M, K/128] FP32 activation scales into the [K/128, M_pad] cuBLASLt documents for a VEC128 B operand, zero-filling the M..M_pad pad rows.

The index math is pinned on the CPU by spark_runtime::cublaslt::scale_layout (SSOT, with the doc quotes); this is only its launcher. The quantizer’s own output is left in place — the in-tree fp8_gemm_t_blockscaled still reads it directly.