Function cutlass_scaled_nvfp4_mm

lightx2v_kernel/python/lightx2v_kernel/gemm.py:4–8

(mat_a, mat_b, scales_a, scales_b, alpha, bias=None)

def _fp8_f16_accum_meta(mat_a, mat_b, scales_a, scales_b, out_dtype, bias=None):
    del scales_a, scales_b, bias
    return torch.empty((mat_a.shape[0], mat_b.shape[1]), dtype=out_dtype, device=mat_a.device)


FP8_F16_ACCUM_MM_AVAILABLE = hasattr(torch.ops.lightx2v_kernel, "cutlass_scaled_fp8_mm_f16_accum_sm120")
if FP8_F16_ACCUM_MM_AVAILABLE:
    _fp8_f16_accum_op = torch.ops.lightx2v_kernel.cutlass_scaled_fp8_mm_f16_accum_sm120.default