using a job called "copybara".

**5. Copy to Google Internal codebase and run internal CI**

-   Once the PR is in the Google codebase, we make sure it integrates well with
    its dependencies and the rest of the system.
-   Rarely, If the tests fail at this stage, we cannot merge the code.
-   If needed, we may come to you to make some changes. At times, it may not be

//
// Below here we do the gradient algorithm. It works as follows:
//
// First we filter the tape to just the subset of operations we want to
// differentiate. In the process of doing so we count how many times each Tensor
// is used as an input to an op (so we know when we're done computing gradients
// for that Tensor). We also count, for each tape entry, how many of its output

        "//conditions:default": [],
    }),
    tags = [
        "no_cuda_asan",  # TODO(b/181771536)
        "no_windows",  # TODO(b/155444728)
    ],
    # We must ensure that the dependencies can be dynamically linked since
    # the shared library must be able to use core:framework.
    deps = [
        ":c_api",
        ":c_api_internal",
        ":c_test_util",

    }
    if (op_type_.empty()) {
      return errors::FailedPrecondition(
          "GraphOperation::Reset must be called before calling SetOpName.");
    }
    // TODO(b/145674566): We use Graph::NewName to get a unique name here but
    // this may not be consistent with python's naming policy.
    mutex_lock l(g_->mu);
    op_.reset(new TF_OperationDescription(g_, op_type_.c_str(),

  gpu_options->set_allow_growth(gpu_memory_allow_growth);

  (*config.mutable_device_count())["CPU"] = num_cpu_devices;

  // TODO(b/113217601): This is needed for EagerContext::runner_ to use a
  // threadpool, so that we avoid the possibility of running the runner_ in the
  // 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

  // There are two ways to represent compact row-major data
  // 1) nullptr indicates tensor is compact and row-majored.
  // 2) fill in the strides array as the real case for compact row-major data.
  // Here we choose option 2, since some frameworks didn't handle the strides
  // argument properly.
  dlm_tensor->dl_tensor.strides = stride_arr->data();

  dlm_tensor->dl_tensor.byte_offset =

        tensorflow::unwrap(handles[i]);
    if (tensorflow::CustomDeviceTensorHandle::classof(unwrapped_handle)) {
      // One of the inputs we're trying to pack is on a custom device. We'll let
      // the first custom device we see handle all of the packing.
      auto* custom_device_handle =
          tensorflow::down_cast<tensorflow::CustomDeviceTensorHandle*>(
              unwrapped_handle);

      TF_SetStatusFromIOError(status, errno, dst);
      return;
    }
  } else if (S_ISDIR(st.st_mode)) {
    TF_SetStatus(status, TF_FAILED_PRECONDITION, "target path is a directory");
    return;
  }

  // We cannot rename directories yet, so prevent this.
  if (stat(src, &st) != 0) {
    TF_SetStatusFromIOError(status, errno, src);
    return;
  } else if (S_ISDIR(st.st_mode)) {

  // 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

to a banned CUDA dependency when '--define framework_shared_object=false' is set.
This means that a CUDA target was probably included via an is_static condition,
used when targeting platforms like Windows where we build statically instead
of dynamically. Here's the output from bazel query:
EOF
  cat $BATS_TEST_TMPDIR/out
  [[ ! -s $BATS_TEST_TMPDIR/out ]]
}

@test "All tensorflow.org/code links point to real files" {