Numpy matrix multiplication gives slightly different result in multiprocessing pool

Question

I want to compute the numerical gradient of a function in parallel. However, I realized that evaluating my function in a multiprocessing pool gives me a different result tha