Function cublas_fp8_rowwise_proj 

pub fn cublas_fp8_rowwise_proj(
    gpu: &dyn GpuBackend,
    derived: &DerivedWeights,
    act_bf16: DevicePtr,
    act_fp8_scratch: DevicePtr,
    act_scale_scratch: DevicePtr,
    fp8w: &Fp8Weight,
    out: DevicePtr,
    m: u32,
    n: u32,
    k: u32,
    stream: u64,
) -> Result<()>

2026-09-25: Route a projection through row-wise FP8 cuBLASLt. The weight pair comes from requant_weight_rowwise_fp8_cached; the activation is quantized per token on every call.

cuBLASLt is handed ceil16(m) rows, so act_fp8_scratch must hold ceil16(m) * k bytes, act_scale_scratch ceil16(m) f32 and out ceil16(m) * n BF16 elements. The pad rows of both scratch buffers are zeroed here, and the pad output rows are written.