static_cast<off_t>(offset))};
    if (r > 0) {
      dst += r;
      offset += static_cast<uint64_t>(r);
      n -= r;  // safe as 0 < r <= n so n will never underflow
      read += r;
    } else if (r == 0) {
      TF_SetStatus(status, TF_OUT_OF_RANGE, "Read fewer bytes than requested");
      break;
    } else if (errno == EINTR || errno == EAGAIN) {
      // Retry

      immediate_execution_ctx_.get(), {x.get()}, UseFunction()));
}

TEST_P(CppGradients, TestMatMulGrad) {
  // TODO(vnvo2409): Figure out why `gradient_checker` does not work very
  // well with `MatMul` and remove `TestMatMul*` in
  // `mnist_gradients_test` when done.
  GTEST_SKIP();

  float A_vals[] = {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, 9.0f};
  int64_t A_dims[] = {3, 3};

// if the status is not ok.
TF_CAPI_EXPORT extern int TF_OpIsStateful(const char* op_type,
                                          TF_Status* status);

// Platform specific initialization routine. Very few platforms actually require
// this to be called.
TF_CAPI_EXPORT void TF_InitMain(const char* usage, int* argc, char*** argv);

// Platform-specific implementation to return an unused port. (This should used

      : status_(TF_NewStatus()),
        // If the context's default exector is set to async, re-using that in
        // each thread would cause collectives to deadlock. For consistency we
        // create a new sync executor for every thread.
        //
        // TODO(allenl): We should have an async API that works with the
        // parallel device.
        device_(device),
        executor_(

  Status Shape(const std::vector<int64_t>** shape) const;
  TF_DataType dtype() const { return dtype_; }

  // Sets its output argument to a summary of the values of this tensor on every
  // component device.
  Status SummarizeValue(std::string& summary);

  std::vector<TensorHandlePtr> release_tensors() { return std::move(tensors_); }

  std::vector<TFE_TensorHandle*> tensors() const {

    const TFE_OpAttrs* attributes, int expected_max_outputs,
    TF_Status* status) {
  absl::optional<std::vector<MaybeParallelTensorOwned>> result;
  // TODO(allenl): We should remove "TPU" from these op names at the very least,
  // or consider other ways of packing/unpacking parallel tensors.
  if (operation_name == std::string("TPUReplicatedInput")) {
    // Special-cased operation for packing per-device tensors into one parallel
    // tensor.