## Known Bugs

*   Using XLA:GPU with CUDA 9 and CUDA 9.1 results in garbage results and/or
    `CUDA_ILLEGAL_ADDRESS` failures.

    Google discovered in mid-December 2017 that the PTX-to-SASS compiler in CUDA
    9 and CUDA 9.1 sometimes does not properly compute the carry bit when
    decomposing 64-bit address calculations with large offsets (e.g. `load [x +
    large_constant]`) into 32-bit arithmetic in SASS.

  } else {
    const void* address =
        mmap(nullptr, st.st_size, PROT_READ, MAP_PRIVATE, fd, 0);
    if (address == MAP_FAILED) {
      TF_SetStatusFromIOError(status, errno, path);
    } else {
      region->plugin_memory_region =
          new tf_read_only_memory_region::PosixMemoryRegion{
              address, static_cast<uint64_t>(st.st_size)};
      TF_SetStatus(status, TF_OK, "");

  std::vector<Tensor> all_input_tensors;
  // Update the vector with information from `input_tensors` if provided.
  if (input_tensors != nullptr) {
    // Note that we take the address of the elements in `all_input_tensors`
    // below. Allocate enough space so that no reallocation happens, which will
    // make the pointers invalid.
    all_input_tensors.reserve(num_inputs);