Engine Crashes if asnumpy() is not called on result

test_operator.py

import os
import mxnet as mx
from common import models
import pickle as pkl

def test_softmax_with_shape(shape, xpu):
    X = mx.symbol.Variable('X')
    L = mx.symbol.Variable('L')
    Y = mx.symbol.Softmax(data=X, label=L)
    x = mx.random.uniform(-1, 1, shape, ctx = xpu)
    l = mx.nd.empty((shape[0],), ctx = xpu)
    l[:] = np.random.randint(0, shape[0]-1, (shape[0],))
    grad = mx.nd.empty(shape, ctx = xpu)

    exec1 = Y.bind(xpu, args = [x, l], args_grad = {'X': grad})
    print 'foward'
    exec1.forward()
    #print exec1.outputs[0].asnumpy() include this line and it works fine

test_softmax_with_shape((3,4), mx.cpu())

Crash happens randomly. Somewhere around 50%