package(
    # copybara:uncomment default_applicable_licenses = ["//tensorflow:license"],
    licenses = ["notice"],
)

# Currently pybind extension shared objects must use only C API headers since
# the C API has static initializers duplicated in the Python bindings. So we
# need a second rule that omits .cc files, in
# tensorflow/python:_pywrap_parallel_device.
filegroup(
    name = "lib_headers",
    srcs = ["parallel_device_lib.h"],

      immediate_execution_ctx_.get(), {X.get(), Y.get()}, UseFunction()));
}

TEST_P(CppGradients, TestBiasAddGrad) {
  if (UseFunction() && UseMlir()) {
    GTEST_SKIP() << "SetAttrString has not been implemented yet.\n";
  }

  // A
  float A_vals[] = {1.0f, 2.0f, 3.0f, 4.0f};
  int64_t A_dims[] = {2, 2};
  AbstractTensorHandlePtr A;
  {
    AbstractTensorHandle* A_raw;

# Build TensorFlow v2
build --define=tf_api_version=2 --action_env=TF2_BEHAVIOR=1

# Prevent double-compilation of some TF code, ref. b/183279666 (internal)
# > TF's gen_api_init_files has a genrule to run the core TensorFlow code
# > on the host machine. If we don't have --distinct_host_configuration=false,
# > the core TensorFlow code will be built once for the host and once for the
# > target platform.

  const int num_devices = TF_DeviceListCount(devices);
  bool has_gpu0 = false;
  bool has_gpu1 = false;
  for (int i = 0; i < num_devices; ++i) {
    const char* dev = TF_DeviceListName(devices, i, status.get());
    ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());
    string device_name(dev);
    if (device_name.find("GPU:0") != string::npos) {
      has_gpu0 = true;
    }

  EXPECT_EQ(attr_value.i(), 2);

  // Placeholder oper now has a consumer.
  ASSERT_EQ(1, TF_OperationOutputNumConsumers(TF_Output{feed, 0}));
  TF_Input feed_port;
  EXPECT_EQ(1, TF_OperationOutputConsumers(TF_Output{feed, 0}, &feed_port, 1));
  EXPECT_EQ(add, feed_port.oper);
  EXPECT_EQ(0, feed_port.index);

  // The scalar const oper also has a consumer.
  ASSERT_EQ(1, TF_OperationOutputNumConsumers(TF_Output{three, 0}));

libcudnn8=8.9.6.50-1+cuda12.2
libcudnn8-dev=8.9.6.50-1+cuda12.2

# This can be removed once NVIDIA publishes a cuda-12.3.2 Docker image.
# For now it ensures that we install at least version 12.3.107 of PTXAS,
# since 12.3.103 has a bug.

  return tensorflow::errors::Unimplemented(
      "SetAttrFunction has not been implemented yet.");
}
Status SetAttrFunctionName(AbstractOperation* op_, const char* attr_name,
                           const char* value, size_t length,
                           ForwardOperation* forward_op_) {
  return tensorflow::errors::Unimplemented(
      "SetAttrFunctionName has not been implemented "
      "yet.");
}

  Status RegisterFunction(AbstractFunction* func) override {
    return errors::Unimplemented(
        "Registering graph functions has not been implemented yet.");
  }

  Status RemoveFunction(const string& func) override {
    return errors::Unimplemented(
        "GraphContext::RemoveFunction has not been implemented yet.");
  }
  // For LLVM style RTTI.
  static bool classof(const AbstractContext* ptr) {

- We inspect every incoming PR and add certain labels to the PR such as `size:`,
  `comp:` etc.  At this stage we check if the PR is valid and meets certain
  quality requirements. For example, we check if the CLA is signed, PR has
  sufficient description, if applicable unit tests are added, if it is a
  reasonable contribution (meaning it is not a single liner cosmetic PR).

**2. Valid?**

         * It has an added advantage since you can you easily switch to different hardware accelerators (cpu, gpu, tpu) as per the task.
         * All you need is a good internet connection and you are all set.