QLinearUnpackedDynamicFp16 Class — pytorch Architecture

Source Code

aten/src/ATen/native/quantized/cpu/qlinear_dynamic.cpp lines 884–947

class QLinearUnpackedDynamicFp16 final {
 public:
#ifdef USE_FBGEMM
  static at::Tensor run(
      at::Tensor input,
      const at::Tensor& weight,
      const std::optional<at::Tensor>& bias) {
    // We make a strong guarantee that models using these operators will have
    // the same numerics across different machines. Therefore, we do not provide
    // a fallback path and rather fail loudly if we cannot run FBGEMM.
    TORCH_CHECK(
        fbgemm::fbgemmSupportedCPU(), "Your CPU doesn't support FBGEMM.");

    TORCH_CHECK(
        weight.dim() == 2,
        "The dimension of weight tensor should be equal to 2");

    auto packed_weight = PackedLinearWeightFp16::prepack(weight, bias);
    auto output = packed_weight->apply_dynamic(std::move(input));

    return output;
  }

  static at::Tensor meta(
      at::Tensor input,
      const at::Tensor& weight,
      const std::optional<at::Tensor>& bias) {
    // We make a strong guarantee that models using these operators will have
    // the same numerics across different machines. Therefore, we do not provide
    // a fallback path and rather fail loudly if we cannot run FBGEMM.
    TORCH_CHECK(
        fbgemm::fbgemmSupportedCPU(), "Your CPU doesn't support FBGEMM.");

    TORCH_CHECK(
        weight.dim() == 2,
        "The dimension of weight tensor should be equal to 2");

    auto out_channel = weight.sym_sizes().vec()[0];
    auto out_sizes = input.sym_sizes().vec();
    out_sizes[out_sizes.size() - 1] = out_channel;

    return at::empty_symint(out_sizes, input.options());
  }
#else // USE_FBGEMM
  static at::Tensor run(
      at::Tensor /* input */,
      const at::Tensor& weight,
      const std::optional<at::Tensor>& bias) {
    // We make a strong guarantee that models using these operators will have
    // the same numerics across different machines. Therefore, we do not provide
    // a fallback path and rather fail loudly if we cannot run FBGEMM.
    TORCH_CHECK(
        false, "This PyTorch installation was not built with FBGEMM operators");
  }

  static at::Tensor meta(
      at::Tensor /* input */,
      const at::Tensor& weight,
      const std::optional<at::Tensor>& bias) {
    TORCH_CHECK(
        false, "This PyTorch installation was not built with FBGEMM operators");
  }
#endif // USE_FBGEMM
};