Function decode_w8a8_quant_act 

pub fn decode_w8a8_quant_act(
    gpu: &dyn GpuBackend,
    scratch: &DecodeW8a8Scratch,
    act_bf16: DevicePtr,
    rows: u32,
    k: u32,
    stream: u64,
) -> Result<()>

Quantize act[rows, k] BF16 ONCE into the shared scratch: FP8 E4M3 bytes + per-token 1×128 FP32 scales, the phantom rows rows..ceil16(rows) zeroed, and the VEC128 scales re-laid-out K-major for cuBLASLt.

Split from the GEMM so a caller with several projections over the SAME activation pays it once — the attention layer’s Q/K/V share normed, so quantizing inside the GEMM helper would run the quantizer and the scale transpose three times per layer per step. Exactly the split dense_ffn_w8a8_prefill makes for gate/up.

A zero scale kills the phantom rows’ CONTRIBUTION, but the FP8 dot product still runs over whatever bytes are there and NaN * 0.0 is NaN — hence the memset of the FP8 bytes and not only of the scales.