// threadpool of GPU event mgr, as that can trigger more callbacks to be
  // scheduled on that same threadpool, causing a deadlock in cases where the
  // caller of event_mgr->ThenExecute() blocks on the completion of the callback
  // (as in the case of ConstOp kernel creation on GPU, which involves copying a
  // CPU tensor to GPU).
  // Setting a larger thread pool does not help with the Swift caller, as we use

    SINGLE_CASE(kF, TF_ATTR_FLOAT, -1);
    SINGLE_CASE(kB, TF_ATTR_BOOL, -1);
    SINGLE_CASE(kType, TF_ATTR_TYPE, -1);
    SINGLE_CASE(kShape, TF_ATTR_SHAPE,
                attr->shape().unknown_rank() ? -1 : attr->shape().dim_size());
    SINGLE_CASE(kTensor, TF_ATTR_TENSOR, -1);
#undef SINGLE_CASE

    case tensorflow::AttrValue::kList:
      metadata.is_list = 1;
      metadata.list_size = 0;

// explicitly by running a TPUReplicatedOutput operation on the parallel device.
//
// TODO(allenl): There are some use-cases that are only supported by copying to
// host at the moment (e.g. debug print on a tensor, .numpy(), etc.). We either
// need to return something here or address these use-cases one by one.
TFE_TensorHandle* CopyTensorFromParallelDevice(TFE_Context* context,

  }

 private:
  bool first_call_ = true;
};

// After the graph pass is registered, it takes effect globally and can affect
// other test cases. Define a static variable to switch it on and off.
bool GraphErrorInjectionPass::enabled_ = false;

// Test to ensure that a registered graph optimization pass is only executed

  switch (data_type) {
    case TF_DataType::TF_BOOL:
      dtype.code = DLDataTypeCode::kDLBool;
      break;
    case TF_DataType::TF_HALF:
    case TF_DataType::TF_FLOAT:
    case TF_DataType::TF_DOUBLE:
      dtype.code = DLDataTypeCode::kDLFloat;
      break;
    case TF_DataType::TF_INT8:
    case TF_DataType::TF_INT16:
    case TF_DataType::TF_INT32:
    case TF_DataType::TF_INT64:

                                   status.get()),
      TFE_DeleteTensorHandle);
  ASSERT_TRUE(arrived);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  // Base case: two CPU inputs executes fine.
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> matmul(
      MatMulOp(context.get(), hcpu.get(), hcpu.get()), TFE_DeleteOp);
  TFE_TensorHandle* retval;
  int num_retvals = 1;

  if (TF_GetCode(status) != TF_OK) return;

  // unwrap op_ and set step_id only if valid step id value was set.
  // Currently only required for non-TFRT use cases, e.g., EagerOp.
  if (step_id_.has_value()) {
    tensorflow::unwrap(op_.get())->SetStepId(step_id_.value());
  }

  TFE_Execute(op_.get(), unwrapped_results.data(), &real_num_outputs, status);

#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/tfrt/common/async_value_tensor.h"
#include "tensorflow/core/tfrt/common/pjrt_util.h"
#include "tsl/lib/core/status_test_util.h"
#include "tsl/platform/casts.h"
#include "tsl/platform/status_matchers.h"
#include "tsl/protobuf/error_codes.pb.h"

namespace tensorflow {
namespace {

using ::testing::HasSubstr;
using ::testing::NotNull;
using ::tsl::testing::StatusIs;

  FileSystem* fs = nullptr;
  Status s = env_->GetFileSystemForFile(generic_path, &fs);
  if (fs == nullptr || !s.ok())
    GTEST_SKIP() << "No filesystem registered: " << s;

  // First, test some interesting corner cases concerning empty URIs
  if (GetParam().empty()) {
    EXPECT_EQ(fs->TranslateName(""), "");
    EXPECT_EQ(fs->TranslateName("/"), "/");
    EXPECT_EQ(fs->TranslateName("//"), "/");

  ScalarFloatFromTensor(fetchValues[1], &dxy_dyValue);
  EXPECT_EQ(X, dxy_dyValue);

  TF_DeleteTensor(fetchValues[0]);
  TF_DeleteTensor(fetchValues[1]);
}

// REGISTER_OP for CApiAttributesTest test cases.
// Registers two ops, each with a single attribute called 'v'.
// The attribute in one op will have a type 'type', the other
// will have list(type).
#define ATTR_TEST_REGISTER_OP(type)                           \