Function fp8_gemm_act_weight_t_blkscaled 

pub fn fp8_gemm_act_weight_t_blkscaled(
    act_fp8: u64,
    act_scale: u64,
    weight_fp8: u64,
    weight_block_scale: u64,
    out: u64,
    m: u32,
    n: u32,
    k: u32,
    stream: u64,
) -> Result<()>

Native FP8 (E4M3) out[M,N] = act[M,K] @ weight[N,K]ᵀ → BF16, with the weight per-128×128-block FP32-scaled (matches Atlas’s Fp8Weight.row_scale layout exactly) and the activation per-[token,128-of-K] FP32-scaled. ~1.8× the bf16 path (152 vs 85 TFLOPS on GB10).

⚠ SCALE-TENSOR LAYOUTS — the two operands do NOT agree, and getting this wrong is silent (see super::scale_layout for the doc quotes, the H100 measurement that caught it, and the index math):

• weight_block_scale (A, BLK128x128_32F) is K-major, L4 × ⌈N/128⌉ — the checkpoint’s row-major [N/128, K/128] grid as-is, valid while ⌈K/128⌉ is a multiple of 4 (scale_layout::blk128x128_stride_ok).
• act_scale (B, VEC128_32F) is N-major, M × ⌈K/128⌉ with the TOKEN index contiguous — i.e. [K/128, M], the TRANSPOSE of what per_token_group_quant_fp8 writes. Callers adapt it with the fp8_act_scale_to_kmajor kernel; passing the quantizer’s buffer straight through permutes the scales and costs ~8% relative RMS at M≈1200.

m must already include the caller’s pad (the docs require the matmul’s M and N to be multiples of 4), and act_fp8/act_scale must cover it.

The output is CONTIGUOUS [M, N]; fp8_gemm_act_weight_t_blkscaled_ldc is the same GEMM with a caller-chosen output row pitch.