Expecting numpy.array.tolist() as input parameter on inference for custom handler

deployment
1

From python script, I am doing Torch serve on like:

arr = np.array(np.random.randn(x,y,z))
data = {'arr': arr.tolist()}
response = requests.post(url, json=data)

In the inference method of the custom handler I have:

    def inference(self, ds, *args, **kwargs):
        print("type of ds") ##### <class 'list'>
        print(type(ds)) 
        print("type of ds[0]")
        print(type(ds[0])) ##### <class 'dict'>
        count = 0
        whatineed = None
        for k, v in ds[0].items():
            tk = type(k)
            tv = type(v)
            for kk, vv in v.items():
                print("KK: " + str(kk))  ##### KK: arr
                print("Len of vv: " + str(len(vv))) # 1
                whatineed = vv
            count = count+1

        ... [ other stuff ]

How can I avoid searching for whatineed?

Is there an easier way to package the input array on the request side?