```
# Copy and paste here
```

**Standalone code to reproduce the issue** 
Provide a reproducible test case that is the bare minimum necessary to generate
the problem. If possible, please share a link to Colab/Jupyter/any notebook.

Also, please include a link to a GraphDef or the model if possible.

**Any other info / logs**

Include any logs or source code that would be helpful to diagnose the problem.

TFE_Op* MatMulOp(TFE_Context* ctx, TFE_TensorHandle* a, TFE_TensorHandle* b);

// Return an identity op.
TFE_Op* IdentityOp(TFE_Context* ctx, TFE_TensorHandle* a);

// Return a shape op fetching the shape of `a`.
TFE_Op* ShapeOp(TFE_Context* ctx, TFE_TensorHandle* a);

// Return an allreduce op adding up input tensor `in` from `group_size` workers.

    id: code-to-reproduce
    attributes:
      label: Standalone code to reproduce the issue
      description: Provide a reproducible test case that is the bare minimum necessary to generate the problem. Please share a link to Colab, Jupyter, or any notebook.
      placeholder: Tell us what you see!
      value:
      render: shell
    validations:
      required: true
  - type: textarea
    id: logs
    attributes:

    TAG=$(head -n 1 "$KOKORO_PIPER_DIR/presubmit_request.txt" | cut -d" " -f2)
  else
    TAG="pr-${KOKORO_GITHUB_PULL_REQUEST_NUMBER}"
  fi
fi

# Build for both JAX and TF usage.  We do these in one place because they share
# almost all of the same cache layers
export DOCKER_BUILDKIT=1
for target in jax tf; do
  IMAGE="gcr.io/tensorflow-sigs/build-arm64:$target-$TAG"
  docker pull "$IMAGE" || true

// per device in `underlying_devices`. Implicit copies off of the device throw
// an error.
//
// All component tensors must have the same dtype. Currently they must also have
// the same shape, although this requirement may be relaxed in the future.
//
// `device_name` must not name an existing physical or custom device (see
// the documentation for TFE_RegisterCustomDevice for more information).
//

      TFE_NewOp(context, "VarHandleOp", status), TFE_DeleteOp);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_OpSetAttrType(op.get(), "dtype", type);
  TFE_OpSetAttrShape(op.get(), "shape", dims, num_dims, status);
  TFE_OpSetAttrString(op.get(), "container", "", 0);
  // Use the special GUID for no buffer sharing
  //
  // TODO(allenl): Should we provide a better API for this? AFAIK this is the