QLinearDynamicFp16 Class — pytorch Architecture

Source Code

aten/src/ATen/native/quantized/cpu/qlinear_dynamic.cpp lines 850–882

class QLinearDynamicFp16 final {
 public:
#ifdef USE_FBGEMM
  static at::Tensor run(
      at::Tensor input,
      const c10::intrusive_ptr<LinearPackedParamsBase>& packed_weight) {
    // 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.");

    auto output = packed_weight->apply_dynamic(std::move(input));

    // Call the relu operator here until fp16 linear dynamic in FBGEMM
    // supports it natively.
    if (ReluFused) {
      output.relu_();
    }
    return output;
  }
#else // USE_FBGEMM
  static at::Tensor run(
      at::Tensor /* input */,
      const c10::intrusive_ptr<LinearPackedParamsBase>& /* packed_weight */) {
    // 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");
  }
#endif // USE_FBGEMM
};