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

        break;
      case FetchState::FINISHED:
        return TF_SetStatus(status, TF_OK, "");
    }
  }
  return TF_SetStatus(
      status, TF_INTERNAL,
      "Control flow should never reach the end of RamFileBlockCache::Fetch.");
}

int64_t RamFileBlockCache::Read(const std::string& filename, size_t offset,
                                size_t n, char* buffer, TF_Status* status) {
  if (n == 0) {

  // TODO(srbs): It seems like this is used only for performance optimization
  // and not for correctness. The only downside of keeping this 1 seems to be
  // that the gradient accumulation is unbounded and we will never
  // aggressively aggregate accumulated gradients to recover memory.
  // Revisit and fix.
  return 1;
}

// Consumes references to the tensors in the gradient_tensors list and returns

    EXPECT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
    auto remote_arg =
        tensorflow::TensorHandleFromInterface(tensorflow::unwrap(h1_task2));
    // The input handles should never change since they have been mirrored.
    ASSERT_FALSE(remote_arg->HasLocalMirror(nullptr));
  }

  TFE_TensorHandle* retvals[1];
  int num_retvals = 1;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);

        "operations on them have been supplied.");

  return OkStatus();
}

// Copies a function table from plugin memory space to core memory space.
//
// This has three benefits:
//   * allows having newer plugins than the current core TensorFlow: the
//     additional entries in the plugin's table are just discarded;
//   * allows having older plugins than the current core TensorFlow (though