nvfp4_gemm_bf16_act_weight_t

spark_runtime::cutlass

Function nvfp4_gemm_bf16_act_weight_t 

Source 
pub fn nvfp4_gemm_bf16_act_weight_t(
    act: u64,
    weight_packed_t: u64,
    weight_scale_t: u64,
    weight_scale_2: f32,
    out: u64,
    m: u32,
    n: u32,
    k: u32,
    stream: u64,
) -> Result<()>
Expand description

Native CUTLASS NVFP4 dense projection: out[M,N] = quant_nvfp4(act[M,K]) @ weight_t[N,K]^T -> BF16.

weight_packed_t and weight_scale_t are Atlas’s transposed NVFP4 prefill layout: packed data [K/2,N], scales [K/16,N]. The wrapper repacks activation and scale tensors into CUTLASS’s SM120 blockscaled layouts in the shared CUTLASS workspace before dispatch.