From 8f04d92daf29a5f5ee43a392d61baf6eef5947be Mon Sep 17 00:00:00 2001
From: Paul Donnelly <pauldonnelly@google.com>
Date: Fri, 21 Feb 2020 14:03:08 -0800
Subject: [PATCH] Temporary rollback of QuantizeAndDequantizeV2Grad.

PiperOrigin-RevId: 296507138
Change-Id: Iab04845a3e2a760073aa3c88ef3e10272a2094c1
---
 tensorflow/cc/gradients/array_grad.cc         |  29 ++---
 .../api_def_QuantizeAndDequantizeV2Grad.pbtxt |   8 --
 .../api_def_QuantizeAndDequantizeV2Grad.pbtxt |   3 -
 .../api_def_QuantizeAndDequantizeV2Grad.pbtxt |   4 -
 .../kernels/quantize_and_dequantize_op.cc     | 116 ------------------
 .../core/kernels/quantize_and_dequantize_op.h |  71 -----------
 .../quantize_and_dequantize_op_gpu.cu.cc      |  40 ------
 .../quantize_and_dequantize_op_test.cc        |  48 --------
 tensorflow/core/ops/array_ops.cc              |  32 -----
 .../eager/pywrap_gradient_exclusions.cc       |   5 +-
 tensorflow/python/ops/array_grad.py           |   5 +
 tensorflow/python/ops/array_ops.py            |  17 ---
 .../api/golden/v1/tensorflow.raw_ops.pbtxt    |   4 -
 .../api/golden/v2/tensorflow.raw_ops.pbtxt    |   4 -
 14 files changed, 17 insertions(+), 369 deletions(-)
 delete mode 100644 tensorflow/core/api_def/base_api/api_def_QuantizeAndDequantizeV2Grad.pbtxt
 delete mode 100644 tensorflow/core/api_def/java_api/api_def_QuantizeAndDequantizeV2Grad.pbtxt
 delete mode 100644 tensorflow/core/api_def/python_api/api_def_QuantizeAndDequantizeV2Grad.pbtxt

diff --git a/tensorflow/cc/gradients/array_grad.cc b/tensorflow/cc/gradients/array_grad.cc
index 3c0813bfe23..e9173227aad 100644
--- a/tensorflow/cc/gradients/array_grad.cc
+++ b/tensorflow/cc/gradients/array_grad.cc
@@ -15,12 +15,13 @@ limitations under the License.
 
 #include <vector>
 
-#include "tensorflow/cc/framework/grad_op_registry.h"
-#include "tensorflow/cc/framework/gradients.h"
 #include "tensorflow/cc/ops/array_ops_internal.h"
 #include "tensorflow/cc/ops/standard_ops.h"
 #include "tensorflow/core/lib/strings/strcat.h"
 
+#include "tensorflow/cc/framework/grad_op_registry.h"
+#include "tensorflow/cc/framework/gradients.h"
+
 namespace tensorflow {
 namespace ops {
 namespace {
@@ -89,25 +90,15 @@ Status QuantizeAndDequantizeGrad(const Scope& scope, const Operation& op,
 }
 REGISTER_GRADIENT_OP("QuantizeAndDequantize", QuantizeAndDequantizeGrad);
 
-Status QuantizeAndDequantizeV2GradHelper(const Scope& scope,
-                                         const Operation& op,
-                                         const std::vector<Output>& grad_inputs,
-                                         std::vector<Output>* grad_outputs) {
-  Input input = Shape(scope, op.input(0));
-  Input input_min = op.input(1);
-  Input input_max = op.input(2);
-  int64 axis;
-  TF_RETURN_IF_ERROR(GetNodeAttr(op.node()->attrs(), "axis", &axis));
-  auto qdq_v2_grad = QuantizeAndDequantizeV2Grad(
-      scope, grad_inputs[0], input, input_min, input_max,
-      QuantizeAndDequantizeV2Grad::Axis(axis));
-  grad_outputs->push_back(qdq_v2_grad.input_backprop);
-  grad_outputs->push_back(qdq_v2_grad.input_min_backprop);
-  grad_outputs->push_back(qdq_v2_grad.input_max_backprop);
+Status QuantizeAndDequantizeV2Grad(const Scope& scope, const Operation& op,
+                                   const std::vector<Output>& grad_inputs,
+                                   std::vector<Output>* grad_outputs) {
+  grad_outputs->push_back(Identity(scope, grad_inputs[0]));
+  grad_outputs->push_back(NoGradient());
+  grad_outputs->push_back(NoGradient());
   return scope.status();
 }
-REGISTER_GRADIENT_OP("QuantizeAndDequantizeV2",
-                     QuantizeAndDequantizeV2GradHelper);
+REGISTER_GRADIENT_OP("QuantizeAndDequantizeV2", QuantizeAndDequantizeV2Grad);
 
 Status QuantizeAndDequantizeV3Grad(const Scope& scope, const Operation& op,
                                    const std::vector<Output>& grad_inputs,
diff --git a/tensorflow/core/api_def/base_api/api_def_QuantizeAndDequantizeV2Grad.pbtxt b/tensorflow/core/api_def/base_api/api_def_QuantizeAndDequantizeV2Grad.pbtxt
deleted file mode 100644
index 6a7a2f38897..00000000000
--- a/tensorflow/core/api_def/base_api/api_def_QuantizeAndDequantizeV2Grad.pbtxt
+++ /dev/null
@@ -1,8 +0,0 @@
-op {
-  graph_op_name: "QuantizeAndDequantizeV2Grad"
-  summary: "Returns the gradient of `QuantizeAndDequantizeV2`."
-  description: <<END
-Returns a gradient of 1 for inputs that are within the quantization range,
-or 0 otherwise.
-END
-}
diff --git a/tensorflow/core/api_def/java_api/api_def_QuantizeAndDequantizeV2Grad.pbtxt b/tensorflow/core/api_def/java_api/api_def_QuantizeAndDequantizeV2Grad.pbtxt
deleted file mode 100644
index f9f898e9c78..00000000000
--- a/tensorflow/core/api_def/java_api/api_def_QuantizeAndDequantizeV2Grad.pbtxt
+++ /dev/null
@@ -1,3 +0,0 @@
-op {
-  graph_op_name: "QuantizeAndDequantizeV2Grad"
-}
diff --git a/tensorflow/core/api_def/python_api/api_def_QuantizeAndDequantizeV2Grad.pbtxt b/tensorflow/core/api_def/python_api/api_def_QuantizeAndDequantizeV2Grad.pbtxt
deleted file mode 100644
index efa83a9d8e8..00000000000
--- a/tensorflow/core/api_def/python_api/api_def_QuantizeAndDequantizeV2Grad.pbtxt
+++ /dev/null
@@ -1,4 +0,0 @@
-op {
-  graph_op_name: "QuantizeAndDequantizeV2Grad"
-  visibility: HIDDEN
-}
diff --git a/tensorflow/core/kernels/quantize_and_dequantize_op.cc b/tensorflow/core/kernels/quantize_and_dequantize_op.cc
index 9fdf374e455..8f71d09c083 100644
--- a/tensorflow/core/kernels/quantize_and_dequantize_op.cc
+++ b/tensorflow/core/kernels/quantize_and_dequantize_op.cc
@@ -131,75 +131,6 @@ class QuantizeAndDequantizeV2Op : public OpKernel {
   bool narrow_range_;
 };
 
-// Implementation of QuantizeAndDequantizeV2GradientOp.
-// When back-propagating the error through a quantized layer, the following
-// paper gives evidence that clipped-ReLU is better than non-clipped:
-// "Deep Learning with Low Precision by Half-wave Gaussian Quantization"
-// http://zpascal.net/cvpr2017/Cai_Deep_Learning_With_CVPR_2017_paper.pdf
-template <typename Device, typename T>
-class QuantizeAndDequantizeV2GradientOp : public OpKernel {
- public:
-  explicit QuantizeAndDequantizeV2GradientOp(OpKernelConstruction* ctx)
-      : OpKernel::OpKernel(ctx) {
-    OP_REQUIRES_OK(ctx, ctx->GetAttr("axis", &axis_));
-  }
-
-  void Compute(OpKernelContext* ctx) override {
-    const Tensor& gradient = ctx->input(0);
-    const Tensor& input = ctx->input(1);
-    Tensor* input_backprop = nullptr;
-    OP_REQUIRES_OK(ctx,
-                   ctx->allocate_output(0, input.shape(), &input_backprop));
-
-    OP_REQUIRES(
-        ctx, input.IsSameSize(gradient),
-        errors::InvalidArgument("gradient and input must be the same size"));
-    const int depth = (axis_ == -1) ? 1 : input.dim_size(axis_);
-    const Tensor& input_min_tensor = ctx->input(2);
-    const Tensor& input_max_tensor = ctx->input(3);
-    if (axis_ != -1) {
-      OP_REQUIRES(
-          ctx, input_min_tensor.dim_size(0) == depth,
-          errors::InvalidArgument("min has incorrect size, expected ", depth,
-                                  " was ", input_min_tensor.dim_size(0)));
-      OP_REQUIRES(
-          ctx, input_max_tensor.dim_size(0) == depth,
-          errors::InvalidArgument("max has incorrect size, expected ", depth,
-                                  " was ", input_max_tensor.dim_size(0)));
-    }
-
-    TensorShape min_max_shape(input_min_tensor.shape());
-    Tensor* input_min_backprop;
-    OP_REQUIRES_OK(ctx,
-                   ctx->allocate_output(1, min_max_shape, &input_min_backprop));
-
-    Tensor* input_max_backprop;
-    OP_REQUIRES_OK(ctx,
-                   ctx->allocate_output(2, min_max_shape, &input_max_backprop));
-
-    if (axis_ == -1) {
-      functor::QuantizeAndDequantizeOneScaleGradientFunctor<Device, T> f;
-      f(ctx->eigen_device<Device>(), gradient.template flat<T>(),
-        input.template flat<T>(), input_min_tensor.scalar<T>(),
-        input_max_tensor.scalar<T>(), input_backprop->template flat<T>(),
-        input_min_backprop->template scalar<T>(),
-        input_max_backprop->template scalar<T>());
-    } else {
-      functor::QuantizeAndDequantizePerChannelGradientFunctor<Device, T> f;
-      f(ctx->eigen_device<Device>(),
-        gradient.template flat_inner_outer_dims<T, 3>(axis_ - 1),
-        input.template flat_inner_outer_dims<T, 3>(axis_ - 1),
-        &input_min_tensor, &input_max_tensor,
-        input_backprop->template flat_inner_outer_dims<T, 3>(axis_ - 1),
-        input_min_backprop->template flat<T>(),
-        input_max_backprop->template flat<T>());
-    }
-  }
-
- private:
-  int axis_;
-};
-
 // Simulate quantization precision loss in a float tensor by:
 // 1. Quantize the tensor to fixed point numbers, which should match the target
 //    quantization method when it is used in inference.
@@ -364,43 +295,6 @@ struct QuantizeAndDequantizePerChannelFunctor<CPUDevice, T> {
         input_max_tensor, round_mode, narrow_range, out);
   }
 };
-
-template <typename T>
-struct QuantizeAndDequantizeOneScaleGradientFunctor<CPUDevice, T> {
-  void operator()(const CPUDevice& d, typename TTypes<T>::ConstFlat gradient,
-                  typename TTypes<T>::ConstFlat input,
-                  typename TTypes<T>::ConstScalar input_min_tensor,
-                  typename TTypes<T>::ConstScalar input_max_tensor,
-                  typename TTypes<T>::Flat input_backprop,
-                  typename TTypes<T>::Scalar input_min_backprop,
-                  typename TTypes<T>::Scalar input_max_backprop) {
-    QuantizeAndDequantizeOneScaleGradientImpl<CPUDevice, T>::Compute(
-        d, gradient, input, input_min_tensor, input_max_tensor, input_backprop,
-        input_min_backprop, input_max_backprop);
-  }
-};
-
-template <typename T>
-struct QuantizeAndDequantizePerChannelGradientFunctor<CPUDevice, T> {
-  void operator()(const CPUDevice& d,
-                  typename TTypes<T, 3>::ConstTensor gradient,
-                  typename TTypes<T, 3>::ConstTensor input,
-                  const Tensor* input_min_tensor,
-                  const Tensor* input_max_tensor,
-                  typename TTypes<T, 3>::Tensor input_backprop,
-                  typename TTypes<T>::Flat input_min_backprop,
-                  typename TTypes<T>::Flat input_max_backprop) {
-    QuantizeAndDequantizePerChannelGradientImpl<CPUDevice, T>::Compute(
-        d, gradient, input, input_min_tensor, input_max_tensor, input_backprop,
-        input_min_backprop, input_max_backprop);
-  }
-};
-
-template struct functor::QuantizeAndDequantizeOneScaleGradientFunctor<CPUDevice,
-                                                                      float>;
-template struct functor::QuantizeAndDequantizePerChannelGradientFunctor<
-    CPUDevice, double>;
-
 }  // namespace functor
 
 #define REGISTER_CPU_KERNEL(T)                                                 \
@@ -408,10 +302,6 @@ template struct functor::QuantizeAndDequantizePerChannelGradientFunctor<
                               .Device(DEVICE_CPU)                              \
                               .TypeConstraint<T>("T"),                         \
                           QuantizeAndDequantizeV2Op<CPUDevice, T>);            \
-  REGISTER_KERNEL_BUILDER(Name("QuantizeAndDequantizeV2Grad")                  \
-                              .Device(DEVICE_CPU)                              \
-                              .TypeConstraint<T>("T"),                         \
-                          QuantizeAndDequantizeV2GradientOp<CPUDevice, T>);    \
   REGISTER_KERNEL_BUILDER(Name("QuantizeAndDequantizeV3")                      \
                               .Device(DEVICE_CPU)                              \
                               .TypeConstraint<T>("T"),                         \
@@ -432,12 +322,6 @@ TF_CALL_double(REGISTER_CPU_KERNEL);
                               .HostMemory("input_max")                         \
                               .TypeConstraint<T>("T"),                         \
                           QuantizeAndDequantizeV2Op<GPUDevice, T>);            \
-  REGISTER_KERNEL_BUILDER(Name("QuantizeAndDequantizeV2Grad")                  \
-                              .Device(DEVICE_GPU)                              \
-                              .HostMemory("input_min")                         \
-                              .HostMemory("input_max")                         \
-                              .TypeConstraint<T>("T"),                         \
-                          QuantizeAndDequantizeV2GradientOp<GPUDevice, T>);    \
   REGISTER_KERNEL_BUILDER(Name("QuantizeAndDequantizeV3")                      \
                               .Device(DEVICE_GPU)                              \
                               .HostMemory("input_min")                         \
diff --git a/tensorflow/core/kernels/quantize_and_dequantize_op.h b/tensorflow/core/kernels/quantize_and_dequantize_op.h
index c286a10a9c6..4dd6e5c839b 100644
--- a/tensorflow/core/kernels/quantize_and_dequantize_op.h
+++ b/tensorflow/core/kernels/quantize_and_dequantize_op.h
@@ -60,28 +60,6 @@ struct QuantizeAndDequantizePerChannelFunctor {
                   typename TTypes<T, 3>::Tensor output);
 };
 
-template <typename Device, typename T>
-struct QuantizeAndDequantizeOneScaleGradientFunctor {
-  void operator()(const Device& d, typename TTypes<T>::ConstFlat gradient,
-                  typename TTypes<T>::ConstFlat input,
-                  typename TTypes<T>::ConstScalar input_min,
-                  typename TTypes<T>::ConstScalar input_max,
-                  typename TTypes<T>::Flat input_backprop,
-                  typename TTypes<T>::Scalar input_min_backprop,
-                  typename TTypes<T>::Scalar input_max_backprop);
-};
-
-template <typename Device, typename T>
-struct QuantizeAndDequantizePerChannelGradientFunctor {
-  void operator()(const Device& d, typename TTypes<T, 3>::ConstTensor gradient,
-                  typename TTypes<T, 3>::ConstTensor input,
-                  const Tensor* input_min_tensor,
-                  const Tensor* input_max_tensor,
-                  typename TTypes<T, 3>::Tensor input_backprop,
-                  typename TTypes<T>::Flat input_min_backprop,
-                  typename TTypes<T>::Flat input_max_backprop);
-};
-
 // The implementation below runs on both CPU and GPU.
 template <typename Device, typename T, typename Func,
           typename Vec = typename TTypes<T>::Vec,
@@ -271,55 +249,6 @@ struct QuantizeAndDequantizePerChannelImpl {
   }
 };
 
-template <typename Device, typename T>
-struct QuantizeAndDequantizeOneScaleGradientImpl {
-  static void Compute(const Device& d, typename TTypes<T>::ConstFlat gradient,
-                      typename TTypes<T>::ConstFlat input,
-                      typename TTypes<T>::ConstScalar input_min,
-                      typename TTypes<T>::ConstScalar input_max,
-                      typename TTypes<T>::Flat input_backprop,
-                      typename TTypes<T>::Scalar input_min_backprop,
-                      typename TTypes<T>::Scalar input_max_backprop) {
-    const T min_val = input_min();
-    const T max_val = input_max();
-    const auto in_range =
-        (input >= min_val && input <= max_val)
-            .select(input.constant(1.0f), input.constant(0.0f));
-    input_backprop.device(d) = gradient * in_range;
-    input_min_backprop.device(d) = input_min_backprop.constant(0.0f);
-    input_max_backprop.device(d) = input_max_backprop.constant(0.0f);
-  }
-};
-
-template <typename Device, typename T>
-struct QuantizeAndDequantizePerChannelGradientImpl {
-  static void Compute(const Device& d,
-                      typename TTypes<T, 3>::ConstTensor gradient,
-                      typename TTypes<T, 3>::ConstTensor input,
-                      const Tensor* input_min_tensor,
-                      const Tensor* input_max_tensor,
-                      typename TTypes<T, 3>::Tensor input_backprop,
-                      typename TTypes<T>::Flat input_min_backprop,
-                      typename TTypes<T>::Flat input_max_backprop) {
-    using Index = typename tensorflow::TTypes<T>::ConstTensor::Index;
-    auto input_min = input_min_tensor->vec<T>();
-    auto input_max = input_max_tensor->vec<T>();
-    int num_channels = input.dimension(1);
-    for (Index i = 0; i < num_channels; ++i) {
-      const auto gradient_chip = gradient.template chip<1>(i);
-      const auto input_chip = input.template chip<1>(i);
-      const T min_val = input_min(i);
-      const T max_val = input_max(i);
-      const auto in_range =
-          (input_chip >= min_val && input_chip <= max_val)
-              .select(input_chip.constant(1.0f), input_chip.constant(0.0f));
-      input_backprop.template chip<1>(i).device(d) = gradient_chip * in_range;
-    }
-    input_min_backprop.device(d) = input_min_backprop.constant(0.0f);
-    input_max_backprop.device(d) = input_max_backprop.constant(0.0f);
-  }
-};
-
 }  // end of namespace functor
 }  // end of namespace tensorflow
 
diff --git a/tensorflow/core/kernels/quantize_and_dequantize_op_gpu.cu.cc b/tensorflow/core/kernels/quantize_and_dequantize_op_gpu.cu.cc
index 9f074535770..f3bb41071cb 100644
--- a/tensorflow/core/kernels/quantize_and_dequantize_op_gpu.cu.cc
+++ b/tensorflow/core/kernels/quantize_and_dequantize_op_gpu.cu.cc
@@ -53,37 +53,6 @@ struct QuantizeAndDequantizePerChannelFunctor<GPUDevice, T> {
   }
 };
 
-template <typename T>
-struct QuantizeAndDequantizeOneScaleGradientFunctor<GPUDevice, T> {
-  void operator()(const GPUDevice& d, typename TTypes<T>::ConstFlat gradient,
-                  typename TTypes<T>::ConstFlat input,
-                  typename TTypes<T>::ConstScalar input_min_tensor,
-                  typename TTypes<T>::ConstScalar input_max_tensor,
-                  typename TTypes<T>::Flat input_backprop,
-                  typename TTypes<T>::Scalar input_min_backprop,
-                  typename TTypes<T>::Scalar input_max_backprop) {
-    QuantizeAndDequantizeOneScaleGradientImpl<GPUDevice, T>::Compute(
-        d, gradient, input, input_min_tensor, input_max_tensor, input_backprop,
-        input_min_backprop, input_max_backprop);
-  }
-};
-
-template <typename T>
-struct QuantizeAndDequantizePerChannelGradientFunctor<GPUDevice, T> {
-  void operator()(const GPUDevice& d,
-                  typename TTypes<T, 3>::ConstTensor gradient,
-                  typename TTypes<T, 3>::ConstTensor input,
-                  const Tensor* input_min_tensor,
-                  const Tensor* input_max_tensor,
-                  typename TTypes<T, 3>::Tensor input_backprop,
-                  typename TTypes<T>::Flat input_min_backprop,
-                  typename TTypes<T>::Flat input_max_backprop) {
-    QuantizeAndDequantizePerChannelGradientImpl<GPUDevice, T>::Compute(
-        d, gradient, input, input_min_tensor, input_max_tensor, input_backprop,
-        input_min_backprop, input_max_backprop);
-  }
-};
-
 }  // end namespace functor
 
 // Instantiate the GPU implementation for float and double.
@@ -96,15 +65,6 @@ template struct functor::QuantizeAndDequantizePerChannelFunctor<GPUDevice,
 template struct functor::QuantizeAndDequantizePerChannelFunctor<GPUDevice,
                                                                 double>;
 
-template struct functor::QuantizeAndDequantizeOneScaleGradientFunctor<GPUDevice,
-                                                                      float>;
-template struct functor::QuantizeAndDequantizeOneScaleGradientFunctor<GPUDevice,
-                                                                      double>;
-template struct functor::QuantizeAndDequantizePerChannelGradientFunctor<
-    GPUDevice, float>;
-template struct functor::QuantizeAndDequantizePerChannelGradientFunctor<
-    GPUDevice, double>;
-
 }  // end namespace tensorflow
 
 #endif  // GOOGLE_CUDA || TENSORFLOW_USE_ROCM
diff --git a/tensorflow/core/kernels/quantize_and_dequantize_op_test.cc b/tensorflow/core/kernels/quantize_and_dequantize_op_test.cc
index 9f8aa8e6f77..90764b0feb2 100644
--- a/tensorflow/core/kernels/quantize_and_dequantize_op_test.cc
+++ b/tensorflow/core/kernels/quantize_and_dequantize_op_test.cc
@@ -362,54 +362,6 @@ TEST_P(ParameterizedQuantizeAndDequantizeTest,
   }
 }
 
-// Convert a 1D tensor with signed 8 bits and round_mode half_up.
-TEST_P(ParameterizedQuantizeAndDequantizeTest, GradientV2_op) {
-  const int axis = GetParam();
-  TF_ASSERT_OK(NodeDefBuilder("qdq_v2_grad_op", "QuantizeAndDequantizeV2Grad")
-                   .Input(FakeInput(DT_FLOAT))
-                   .Input(FakeInput(DT_FLOAT))
-                   .Input(FakeInput(DT_FLOAT))
-                   .Input(FakeInput(DT_FLOAT))
-                   .Attr("axis", axis)
-                   .Finalize(node_def()));
-  TF_ASSERT_OK(InitOp());
-  const std::vector<int64> dims = {2, 3, 4, 5};
-  // Input gradient. (repeating 11 values multiplied by (slice_idx + 1))
-  auto gradients = ScalePerSliceAlongAxis<float>(
-      dims, axis, {1, -2, -3, 4, 5, 6, -7, -8, -9, -10, 11});
-  AddInputFromArray<float>(TensorShape(dims), gradients);
-  // Forward op inputs. (repeating 7 values multiplied by (slice_idx + 1)).
-  auto inputs = ScalePerSliceAlongAxis<float>(
-      dims, axis, {-1, -0.5, 0, 0.3, 0.8, 0.55, 0.6});
-  AddInputFromArray<float>(TensorShape(dims), inputs);
-  const int num_slices = (axis == -1) ? 1 : dims[axis];
-  const TensorShape range_shape =
-      (axis == -1) ? TensorShape({}) : TensorShape({num_slices});
-  std::vector<float> input_min_values(num_slices), input_max_values(num_slices);
-  for (int i = 0; i < num_slices; ++i) {
-    input_max_values[i] = 0.8f + i * 0.4f;
-    input_min_values[i] = -input_max_values[i];
-  }
-  AddInputFromArray<float>(range_shape, input_min_values);
-  AddInputFromArray<float>(range_shape, input_max_values);
-  std::vector<float> expected_vals(inputs.size());
-  int minor_size = 1;
-  for (int i = axis + 1; i < dims.size(); ++i) {
-    minor_size *= dims[i];
-  }
-  for (int i = 0; i < inputs.size(); ++i) {
-    int slice_idx = (i / minor_size) % num_slices;
-    expected_vals[i] = ((inputs[i] >= input_min_values[slice_idx]) &&
-                        (inputs[i] <= input_max_values[slice_idx]))
-                           ? gradients[i]
-                           : 0;
-  }
-  TF_ASSERT_OK(RunOpKernel());
-  Tensor expected(allocator(), DT_FLOAT, TensorShape(dims));
-  test::FillValues<float>(&expected, expected_vals);
-  test::ExpectTensorNear<float>(expected, *GetOutput(0), 1e-5);
-}
-
 // Instantiate parameterized tests for axis = -1, 1, 3.
 INSTANTIATE_TEST_SUITE_P(All, ParameterizedQuantizeAndDequantizeTest,
                          ::testing::Values(-1, 1, 3));
diff --git a/tensorflow/core/ops/array_ops.cc b/tensorflow/core/ops/array_ops.cc
index 1d4526f506e..65c9510a1f2 100644
--- a/tensorflow/core/ops/array_ops.cc
+++ b/tensorflow/core/ops/array_ops.cc
@@ -2800,38 +2800,6 @@ REGISTER_OP("QuantizeAndDequantizeV2")
       return Status::OK();
     });
 
-REGISTER_OP("QuantizeAndDequantizeV2Grad")
-    .Input("gradients: T")
-    .Input("input: T")
-    .Input("input_min: T")
-    .Input("input_max: T")
-    .Output("input_backprop: T")
-    .Output("input_min_backprop: T")
-    .Output("input_max_backprop: T")
-    .Attr("T: {bfloat16, half, float, double}")
-    .Attr("axis: int = -1")
-    .SetShapeFn([](InferenceContext* c) {
-      int axis;
-      TF_RETURN_IF_ERROR(c->GetAttr("axis", &axis));
-      const int minmax_rank = (axis == -1) ? 0 : 1;
-      ShapeHandle minmax;
-      TF_RETURN_IF_ERROR(c->WithRank(c->input(2), minmax_rank, &minmax));
-      TF_RETURN_IF_ERROR(c->Merge(c->input(3), minmax, &minmax));
-      if (axis != -1) {
-        ShapeHandle input;
-        TF_RETURN_IF_ERROR(c->WithRankAtLeast(c->input(0), axis + 1, &input));
-        DimensionHandle depth;
-        TF_RETURN_IF_ERROR(
-            c->Merge(c->Dim(minmax, 0), c->Dim(input, axis), &depth));
-      }
-      ShapeHandle inputs;
-      TF_RETURN_IF_ERROR(c->Merge(c->input(0), c->input(1), &inputs));
-      c->set_output(0, inputs);
-      c->set_output(1, minmax);
-      c->set_output(2, minmax);
-      return Status::OK();
-    });
-
 REGISTER_OP("QuantizeAndDequantizeV3")
     .Input("input: T")
     .Input("input_min: T")
diff --git a/tensorflow/python/eager/pywrap_gradient_exclusions.cc b/tensorflow/python/eager/pywrap_gradient_exclusions.cc
index c483847b969..afae0b57ee7 100644
--- a/tensorflow/python/eager/pywrap_gradient_exclusions.cc
+++ b/tensorflow/python/eager/pywrap_gradient_exclusions.cc
@@ -224,7 +224,7 @@ absl::optional<tensorflow::gtl::FlatSet<int>> OpGradientUnusedInputIndices(
       {"PreventGradient"},
       {"Qr"},
       {"QuantizeAndDequantize"},
-      {"QuantizeAndDequantizeV2Grad", 1, {3}},
+      {"QuantizeAndDequantizeV2"},
       {"QuantizeAndDequantizeV3"},
       {"QueueClose"},
       {"QueueDequeue"},
@@ -410,7 +410,7 @@ absl::optional<tensorflow::gtl::FlatSet<int>> OpGradientUnusedInputIndices(
 
 absl::optional<tensorflow::gtl::FlatSet<int>> OpGradientUnusedOutputIndices(
     const tensorflow::string &op_name) {
-  static std::array<OpIndexInfo, 470> a = {{
+  static std::array<OpIndexInfo, 469> a = {{
       {"Abs"},
       {"AccumulateNV2"},
       {"Acos"},
@@ -652,7 +652,6 @@ absl::optional<tensorflow::gtl::FlatSet<int>> OpGradientUnusedOutputIndices(
       {"Prod"},
       {"QuantizeAndDequantize"},
       {"QuantizeAndDequantizeV2"},
-      {"QuantizeAndDequantizeV2Grad"},
       {"QuantizeAndDequantizeV3"},
       {"QueueClose"},
       {"QueueEnqueue"},
diff --git a/tensorflow/python/ops/array_grad.py b/tensorflow/python/ops/array_grad.py
index 73edd9e2b62..e1ed4443017 100644
--- a/tensorflow/python/ops/array_grad.py
+++ b/tensorflow/python/ops/array_grad.py
@@ -959,6 +959,11 @@ def _QuantizeAndDequantizeGrad(_, grad):
   return grad
 
 
+@ops.RegisterGradient("QuantizeAndDequantizeV2")
+def _QuantizeAndDequantizeV2Grad(_, grad):
+  return [grad, None, None]
+
+
 @ops.RegisterGradient("QuantizeAndDequantizeV3")
 def _QuantizeAndDequantizeV3Grad(_, grad):
   # Only propagate the gradient for the unquantized input.
diff --git a/tensorflow/python/ops/array_ops.py b/tensorflow/python/ops/array_ops.py
index 4356874b4dc..4b805a64d36 100644
--- a/tensorflow/python/ops/array_ops.py
+++ b/tensorflow/python/ops/array_ops.py
@@ -3551,23 +3551,6 @@ def _FakeQuantWithMinMaxVarsPerChannelGradient(op, grad):
       narrow_range=op.get_attr("narrow_range"))
 
 
-@ops.RegisterGradient("QuantizeAndDequantizeV2")
-def _QuantizeAndDequantizeV2Grad(op, grad):
-  """Gradient for QuantizeAndDequantizeV2 op."""
-  return quantize_and_dequantize_v2_grad(
-      grad,
-      op.inputs[0],
-      op.inputs[1],
-      op.inputs[2],
-      axis=op.get_attr("axis"))
-
-
-@ops.RegisterGradient("QuantizeAndDequantizeV2Grad")
-def _QuantizeAndDequantizeV2GradGrad(op, grad):
-  """Gradient for QuantizeAndDequantizeV2Grad op."""
-  return _QuantizeAndDequantizeV2Grad(op, grad)
-
-
 @tf_export("required_space_to_batch_paddings")
 def required_space_to_batch_paddings(input_shape,
                                      block_shape,
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.raw_ops.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.raw_ops.pbtxt
index 78a2bb6f8cb..853f67c12de 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.raw_ops.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.raw_ops.pbtxt
@@ -2784,10 +2784,6 @@ tf_module {
     name: "QuantizeAndDequantizeV2"
     argspec: "args=[\'input\', \'input_min\', \'input_max\', \'signed_input\', \'num_bits\', \'range_given\', \'round_mode\', \'narrow_range\', \'axis\', \'name\'], varargs=None, keywords=None, defaults=[\'True\', \'8\', \'False\', \'HALF_TO_EVEN\', \'False\', \'-1\', \'None\'], "
   }
-  member_method {
-    name: "QuantizeAndDequantizeV2Grad"
-    argspec: "args=[\'gradients\', \'input\', \'input_min\', \'input_max\', \'axis\', \'name\'], varargs=None, keywords=None, defaults=[\'-1\', \'None\'], "
-  }
   member_method {
     name: "QuantizeAndDequantizeV3"
     argspec: "args=[\'input\', \'input_min\', \'input_max\', \'num_bits\', \'signed_input\', \'range_given\', \'narrow_range\', \'axis\', \'name\'], varargs=None, keywords=None, defaults=[\'True\', \'True\', \'False\', \'-1\', \'None\'], "
diff --git a/tensorflow/tools/api/golden/v2/tensorflow.raw_ops.pbtxt b/tensorflow/tools/api/golden/v2/tensorflow.raw_ops.pbtxt
index 78a2bb6f8cb..853f67c12de 100644
--- a/tensorflow/tools/api/golden/v2/tensorflow.raw_ops.pbtxt
+++ b/tensorflow/tools/api/golden/v2/tensorflow.raw_ops.pbtxt
@@ -2784,10 +2784,6 @@ tf_module {
     name: "QuantizeAndDequantizeV2"
     argspec: "args=[\'input\', \'input_min\', \'input_max\', \'signed_input\', \'num_bits\', \'range_given\', \'round_mode\', \'narrow_range\', \'axis\', \'name\'], varargs=None, keywords=None, defaults=[\'True\', \'8\', \'False\', \'HALF_TO_EVEN\', \'False\', \'-1\', \'None\'], "
   }
-  member_method {
-    name: "QuantizeAndDequantizeV2Grad"
-    argspec: "args=[\'gradients\', \'input\', \'input_min\', \'input_max\', \'axis\', \'name\'], varargs=None, keywords=None, defaults=[\'-1\', \'None\'], "
-  }
   member_method {
     name: "QuantizeAndDequantizeV3"
     argspec: "args=[\'input\', \'input_min\', \'input_max\', \'num_bits\', \'signed_input\', \'range_given\', \'narrow_range\', \'axis\', \'name\'], varargs=None, keywords=None, defaults=[\'True\', \'True\', \'False\', \'-1\', \'None\'], "