It feels like a bit of an aside, but (for the specific benefit of @louis-mandel and @stevebronder as well as everyone else generically) our interest is focused on the fact that because Louis’ team has facilitated a backend in NumPyro, running multiple chains on each of many GPUs becomes a lot easier than it would be otherwise (and doesn’t require any changes to the Stan files). That should, we hope (and believe), make it possible to implement large-scale (aka using multiple GPUs) SMC samplers on GPUs. Our hope (and anticipation) is that we can then replace the MCMC in Stan with an SMC sampler and thereby offer substantial speed-ups relative to Stan. We’re excited about this prospect and see the work that Louis’ team has done as an important stepping stone towards an exciting future!
3 Likes