Function cublas_bf16_proj_dense_f32_out 

pub fn cublas_bf16_proj_dense_f32_out(
    act: DevicePtr,
    weight_bf16: DevicePtr,
    out: DevicePtr,
    m: u32,
    n: u32,
    k: u32,
    stream: u64,
) -> Result<()>

cublas_bf16_proj_dense writing an FP32 output buffer.

For consumers whose downstream kernel reads FP32 — GLM’s DSA indexer wq_b — which otherwise cannot use a batched GEMM at all and degrade to one M=1 GEMV per row.