" but actual rank is ", rank);
    return;
  }

  if (num_dims == 0) {
    // Output shape is a scalar.
    return;
  }

  // Rank is greater than 0, so fill in the values, if known, and
  // -1 for unknown values.
  for (int i = 0; i < num_dims; ++i) {
    auto dim = ic->Dim(shape, i);
    int64_t value = -1;
    if (ic->ValueKnown(dim)) {

  return op;
}

TF_Operation* LessThan(TF_Output l, TF_Output r, TF_Graph* graph,
                       TF_Status* s) {
  TF_OperationDescription* desc = TF_NewOperation(graph, "Less", "less_than");
  TF_AddInput(desc, l);
  TF_AddInput(desc, r);
  return TF_FinishOperation(desc, s);
}

TF_Operation* RandomUniform(TF_Operation* shape, TF_DataType dtype,

  CHECK(tensorflow::protobuf::TextFormat::ParseFromString(
      VariableAddFunctionSignature(), &def));
  return def.SerializeAsString();
}

// A graph optimization pass that would fail when triggered for more than once.
class GraphErrorInjectionPass : public tensorflow::GraphOptimizationPass {
 public:
  static bool enabled_;
  GraphErrorInjectionPass() {}

  tensorflow::Status Run(

#include "tensorflow/c/eager/c_api.h"
#include "tensorflow/c/eager/c_api_experimental.h"
#include "tensorflow/core/platform/test.h"

// NOTE(allenl): These tests currently go through TFE_Execute and so are
// integration testing rather than purely testing the parallel device. They
// correspond fairly well to the implementation, but testing the C++ directly is
// another option.

namespace tensorflow {
namespace parallel_device {

      auto& indices = iter->second;
      if (std::find(indices.begin(), indices.end(), idx) != indices.end()) {
        return InvalidArgument("TF_Output ", node->name(), ":", idx,
                               " appears more than once in the input list");
      }
      indices.push_back(idx);
    }
  }
  return absl::OkStatus();
}

// Converts `noutputs` and `outputs` into `outputs_tensors` and does various
// checks while doing so.

#include "tensorflow/core/lib/core/status_test_util.h"
#include "tensorflow/core/platform/test.h"

// NOTE(allenl): These tests currently go through TFE_Execute and so are
// integration testing rather than purely testing the parallel device. They
// correspond fairly well to the implementation, but testing the C++ directly is
// another option.

namespace tensorflow {
namespace parallel_device {

  auto posix_file = static_cast<PosixFile*>(file->plugin_file);
  char* dst = buffer;
  int64_t read = 0;

  while (n > 0) {
    // Some platforms, notably macs, throw `EINVAL` if `pread` is asked to read
    // more than fits in a 32-bit integer.
    size_t requested_read_length;
    if (n > INT32_MAX)
      requested_read_length = INT32_MAX;
    else
      requested_read_length = n;

    EXPECT_EQ(got_n, want_n);
    calls++;
    memset(buffer, 'x', got_n);
    TF_SetStatus(status, TF_OK, "");
    return got_n;
  };
  // If block_size, max_bytes, or both are zero, or want_n is larger than
  // max_bytes the cache is a pass-through.
  tf_gcs_filesystem::RamFileBlockCache cache1(1, 0, 0, fetcher);
  tf_gcs_filesystem::RamFileBlockCache cache2(0, 1, 0, fetcher);

  if (!status.ok())
    GTEST_SKIP() << "NewRandomAccessFile() not supported: " << status;

  char scratch[64 /* must be bigger than test_data */] = {0};
  StringPiece result;
  // read at least 1 byte more than test_data
  status = read_file->Read(0, test_data.size() + 1, &result, scratch);
  EXPECT_PRED2(UnimplementedOrReturnsCode, status, Code::OUT_OF_RANGE);
}

  EXPECT_TRUE(absl::StrContains(s, expected))
      << "'" << s << "' does not contain '" << expected << "'";
}

// Returns the GPU device name if there is one (with arbitrary tie breaking if
// there are more than one), or "" otherwise.
string GPUDeviceName(TF_Session* session) {
  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(
      TF_NewStatus(), TF_DeleteStatus);
  TF_Status* s = status.get();