} else {
    return ToOperation(iter->second);
  }
}

TF_Operation* TF_GraphNextOperation(TF_Graph* graph, size_t* pos) {
  if (*pos == 0) {
    // Advance past the first sentinel nodes in every graph (the source & sink).
    *pos += 2;
  } else {
    // Advance to the next node.
    *pos += 1;
  }

  mutex_lock l(graph->mu);

                              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};

//
// Each filesystem implementation can be provided by DSOs, so we provide the
// `--dsos` flag to specify a list of shared objects to be loaded in order.
// If the flag is not used, no shared objects are loaded.
//
// Every filesystem provides support for accessing URIs of form
// `[<scheme>://]<path>` where `<scheme>` is optional (if missing, we are
// accessing local paths). This test suite tests exactly one scheme for each

  // 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

      : 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_(

    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.