Function fp8_gemm_act_weight_t_blkscaled_ldc 

pub fn fp8_gemm_act_weight_t_blkscaled_ldc(
    act_fp8: u64,
    act_scale: u64,
    weight_fp8: u64,
    weight_block_scale: u64,
    out: u64,
    m: u32,
    n: u32,
    k: u32,
    ldc: u32,
    stream: u64,
) -> Result<()>

fp8_gemm_act_weight_t_blkscaled with an explicit output ROW PITCH.

WHY (#927, the 5..16-row decode projections). cuBLASLt’s D operand is a column-major [N, M] layout with leading dimension ldc, which is exactly a row-major [M, N] whose rows are ldc BF16 elements apart — so one parameter is the whole difference between a contiguous [M, N] output and writing straight into a strided slot. The multi-seq decode QKV buffer is [n, per_seq_qkv] with Q at 0, K at q_proj_bytes and V after it, so ldc = per_seq_qkv / 2 puts each row’s n outputs in its own sequence’s slot with the gaps left alone — the same thing the _strided GEMV entry points do, without a staging buffer or a scatter kernel.

⚠ WRITE EXTENT. The library writes n elements of EACH of the m columns, i.e. the last byte touched is at element (m - 1) * ldc + n. m here is the caller’s PADDED row count, so the phantom rows are written too; callers must bound that extent against their buffer — see spark_model::layers::ops::strided_out_extent_elems, which is the SSOT for the arithmetic and is unit-tested on the CPU.

ldc >= n is required by the library (a leading dimension shorter than the column is rejected); it is checked here so the failure names itself instead of arriving as a cuBLAS status code.