// These tests should not be confused with `[*]_grad_test` which compare the
// result of `gradient_checker` and `[*]_grad`. The tests here test the
// functionality of `gradient_checker` by comparing the result with expected
// manual user-provided gradients.
class GradientCheckerTest
    : public ::testing::TestWithParam<std::tuple<const char*, bool, bool>> {
 protected:
  void SetUp() override {

    placeholder: e.g., 3.9 validations: required: false

-   type: input id: Bazel attributes: label: Bazel version description: if
    compiling from source placeholder: validations: required: false

-   type: input id: Compiler attributes: label: GCC/Compiler version
    description: if compiling from source placeholder: validations: required:
    false

-   type: input id: Cuda attributes: label: CUDA/cuDNN version description:

  // `gradient_checker` only works with model that returns only 1 tensor.
  // Although, `ops::SparseSoftmaxCrossEntropyWithLogits` returns 2 tensors, the
  // second tensor isn't needed for computing gradient so we could safely drop
  // it.
  outputs[0] = loss;
  backprop->Unref();
  return absl::OkStatus();
}

Status BiasAddModel(AbstractContext* ctx,

  // Returns size of specified dimension
  //
  // -1 indicates an unknown axis length; this is unreachable for most standard
  // ImmediateExecutionTensorHandles, but comes up for example when computing
  // the shape of a parallel tensor with component shapes differing across
  // devices.
  virtual Status Dim(int dim_index, int64_t* dim) const = 0;

  // Returns the device which created the handle.

  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);
  auto* arg1 = TF_AddFunctionParameter(graph_ctx, TF_FLOAT, {-1, nullptr}, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  // Create a first "Add" computing `arg0 + arg1`.
  TF_AbstractTensor* add_output1;
  {
    // Build an abstract operation, inputs and output.
    auto* add_op = TF_NewAbstractOp(graph_ctx);
    TF_AbstractOpSetOpType(add_op, "Add", s);

  std::string shape_string;
  if (Shape(&shape).ok()) {
    shape_string = shape.DebugString();
  } else {
    shape_string = "<error computing shape>";
  }
  std::string value_string;
  if (!SummarizeValue(value_string).ok()) {
    value_string = "<error computing value>";
  }
  if (value_string.length() > 100) {
    // The default NumPy-style output can be distractingly long in error
    // messages.

  TF_RETURN_IF_ERROR(
      ops::IdentityN(ctx, inputs, absl::MakeSpan(temp_outputs), "IdentityN"));
  // Although, `ops::IdentityN` returns 2 tensors, the first tensor isn't needed
  // for computing gradient so we could safely drop it.
  outputs[0] = temp_outputs[1];
  temp_outputs[0]->Unref();
  return absl::OkStatus();
}

class CppGradients

in TensorFlow ops can be recognized as security issues only if they are
reachable and exploitable through production-grade, benign models.

### Compilation

Compiling models via the recommended entry points described in
[XLA](https://www.tensorflow.org/xla) and
[JAX](https://jax.readthedocs.io/en/latest/jax-101/02-jitting.html)

#include "tensorflow/c/eager/tape.h"
#include "tensorflow/core/common_runtime/eager/attr_builder.h"

namespace tensorflow {
namespace gradients {

// =============== Experimental C++ API for computing gradients ===============

// Sample gradient function:
//
// class AddGradientFunction : public GradientFunction {
//  public:
//   Status Compute(Context* ctx,

  # VMs.
  export CLOUDSDK_PYTHON=$(which python3.11)
  gcloud auth activate-service-account --key-file="${GOOGLE_APPLICATION_CREDENTIALS}"
fi
set -x

# When cross-compiling with RBE, we need to copy the macOS sysroot to be
# inside the TensorFlow root directory. We then define them as a filegroup
# target inside "tensorflow/tools/toolchains/cross_compile/cc" so that Bazel