Function cutlass_bf16_proj 

pub fn cutlass_bf16_proj(
    gpu: &dyn GpuBackend,
    derived: &DerivedWeights,
    act: DevicePtr,
    fp8w: &Fp8Weight,
    out: DevicePtr,
    m: u32,
    n: u32,
    k: u32,
    stream: u64,
) -> Result<()>

Route a projection out[M,N] = act[M,K] @ weightᵀ through CUTLASS BF16.

★ A REFERENCE PATH FOR BENCHMARKING, NOT A SHIPPING ONE. Opt-in behind ATLAS_CUTLASS_GEMM=1 and OFF by default; a build without CUTLASS_HOME cannot reach it at all. It exists so a shape can be A/B’d against the industry reference on the same box — if CUTLASS wins a shape, the fix is a faster Atlas kernel, not a promotion. See the module docs on spark_runtime::cutlass for the full rationale (SSOT).