from_pyro fails when MCMC is run on GPU

[chrhck]

When MCMC is performed on a GPU, from_pyro fails, as the sites returned by
trace = pyro.poutine.trace(self.model).get_trace(*self._args, **self._kwargs) can be tensors that still live on the GPU.
This will cause this line to break:

arviz/arviz/data/io_pyro.py

Line 231 in 1f3b8fd

vals = utils.one_de(vals)

Since there is currently no easy way of transferring a pyro MCMC object to the cpu, a workaround is adding a .cpu()to this line:

arviz/arviz/data/io_pyro.py

Line 111 in 1f3b8fd

name: site["value"]

[ahartikainen]

That sounds ok solution.

Is there a way to see if the trace is on cpu or gpu? And should we unpack data from gpu in one step or one-by-one (when numpy array is creates)

[fonnesbeck]

Calling the cpu() method on data that's already on the cpu() will just be ignored, so no need to check. I will submit a PR for this now.