Debug a multiprocessing job that fails with pickling and CUDA errors

import torch, multiprocessing as mp

model = Net().cuda()

def predict(batch):
    return model(batch.cuda()).cpu()

if True:
    with mp.Pool(4) as pool:
        results = pool.map(lambda b: predict(b), batches)

The script fails with Can't pickle local object / Can't pickle <function <lambda>> and, after you remove the lambda, with RuntimeError: Cannot re-initialize CUDA in forked subprocess. Explain both errors and give a correct design.