#pragma once

#include <torch/extension.h>

#include <optional>

torch::Tensor smallm_nvfp4_linear_cuda(
    const torch::Tensor& input,
    const torch::Tensor& packed_weight,
    const torch::Tensor& weight_block_scales,
    const torch::Tensor& weight_tensor_scale,
    const std::optional<torch::Tensor>& bias);