Metropolis may be faster per sample but it will usually take orders of magnitude more samples to converge than NUTS (which is the default). Can you try using NUTS instead? You may have some other difficulties if the problem isn’t identifiable but in general it should work better.