input_ = fused_func_op_.getArgument(0);
  bias_ = fused_func_op_.getArgument(2);

  weight_ = fused_func_op_.getArgument(1);
  weight_type_ = mlir::cast<RankedTensorType>(weight_.getType());

  if (weight_type_.getRank() != 2) {
    return fused_func_op_.emitError() << "The weight tensor was not of rank 2";
  }

  if (weight_type_.getDimSize(1) % num_gates_ != 0) {

                         METHOD_STATIC_RANGE_WEIGHT_ONLY_INT8,
                     "weight_only", "Post-training weight-only quantizaiton"))};

  Option<bool> enable_per_channel_quantization_{
      *this, "enable-per-channel-quantization", llvm::cl::init(false),
      llvm::cl::desc("Whether enable per-channel quantized weights.")};
};

// Apply constant transformations for the op_set.

              input.getDefiningOp())) {
        // Tensors with derived scale are biases, and handled in propagation.
        if (tensor_property.use_derived_scale) continue;
        // For weights, use quantization scale inferred from the values.
        if (failed(processConstantOp(op, input.getDefiningOp(), index,
                                     tensor_property, rewriter))) {
          return failure();

    LOG(INFO) << quantized_tensor->name()->str() << " "
              << float_tensor->name()->str();
    if (ExpectEqualTensor(quantized_tensor, float_tensor)) {
      if (quantized && quantized_tensor->name()->str().find("weights")) {
        // If tensor is quantized, data type and buffer contents can be
        // different between float and quantized tensors. So do those tests
        // separately in the test body without checking them here.

    }
    storage_type = mlir::cast<mlir::IntegerType>(raw_elem_type);
  }

  // TFlite uses narrow-range [u]int8 for constant buffers of quantized weights.
  // Since we don't know which ones are weights, we represent this optimization
  // as a change in the storage bounds for the type for all constants of this
  // type.
  const int bitwidth = storage_type.getIntOrFloatBitWidth();

						Port:   &networking.PortSelector{Number: 443},
						Subset: "prod",
					},
					Weight: 75,
				},
				{
					Destination: &networking.Destination{
						Host:   "example-canary.com",
						Port:   &networking.PortSelector{Number: 443},
						Subset: "canary",
					},
					Weight: 25,
				},
			},
		},
	}

	for _, tt := range cases {
		t.Run(tt.name, func(t *testing.T) {

  OpSet op_set_;

  Option<bool> enable_per_channel_quantization_{
      *this, "enable-per-channel-quantization", llvm::cl::init(false),
      llvm::cl::desc("Whether enable per-channel quantized weights.")};
};

// If the weight is applicable to dynamic range quantization, insert Quantize
// and Dequantize ops with per-tensor scale.
class PrepareDRQQuantizableOp : public OpRewritePattern<arith::ConstantOp> {
 public:

  SPARSE = 1,
  DENSE = 2,
}

table LSHProjectionOptions {
  type: LSHProjectionType;
}

table SVDFOptions {
  rank:int;
  fused_activation_function:ActivationFunctionType;
  // For weights-only quantization, use asymmetric quantization for non
  // constant inputs at evaluation time.
  asymmetric_quantize_inputs:bool;
}

// An implementation of TensorFlow RNNCell.
table RNNOptions {

          &q_builder, input_model, quantized_type, use_updated_hybrid_scheme,
          ::tflite::optimize::QuantizerType::OLD_QUANTIZER) != kTfLiteOk) {
    return absl::InvalidArgumentError(
        "Quantize weights transformation failed.");
  }
  const uint8_t* q_buffer = q_builder.GetBufferPointer();
  *result =
      std::string(reinterpret_cast<const char*>(q_buffer), q_builder.GetSize());

  return absl::OkStatus();
}

inline constexpr std::array<int64_t, 4> kNchwToNhwcPermutation = {0, 2, 3, 1};

// Permutation from the OIHW (== (output features, input features, height,
// width)) tensor format to HWIO. This is commonly used to transpose convolution
// weights represented as OIHW format to HWIO, which is more desirable for
// certain downstream optimization passes (e.g. XLA).