Function code1x16_matmat

inference_lib/src/aqlm/inference_kernels/cuda_kernel.cpp:148–182

}

torch::Tensor code1x16_matmat(
  const torch::Tensor& input,
  const torch::Tensor& codes,
  const torch::Tensor& codebooks,
  const torch::Tensor& scales,
  const std::optional<torch::Tensor>& bias
) {
  bool use_bfloat16 = check_use_bfloat16(input);
  auto input_sizes = input.sizes();
  auto out_features = codes.size(0) * codebooks.size(2);
  auto flat_input = input.reshape({-1, input.size(-1)});
  auto flat_output = torch::empty({flat_input.size(0), out_features},
    torch::TensorOptions()
      .dtype(input.dtype())
      .device(input.device())
  );

  for (int i = 0; i < flat_input.size(0); ++i) {
    auto input_vec = flat_input.index({i});
    auto output_vec = flat_output.index({i});
    code1x16_matvec(
      codes.squeeze(2),
      input_vec,
      output_vec,
      codebooks,
      use_bfloat16
    );
  }
  return scale_bias_unflatten_output(
    flat_output,
    scales,
    bias,
    input_sizes
  );
}

torch::Tensor code1x16_dequant(
  const torch::Tensor& codes,