In Trace_ELBO, the num_particles argument allows one to effectively introduce a batch size in estimating the ELBO gradient if num_particles > 1. By default, it's vectorized over the num_particles. Is it possible to also distribute the batch dimension over devices (e.g. when running on multiple GPUs). My particular application is prone to jax OOM errors and would benefit from distribution over jax.pmap.
Hi,
In
Trace_ELBO
, thenum_particles
argument allows one to effectively introduce a batch size in estimating the ELBO gradient ifnum_particles > 1
. By default, it's vectorized over thenum_particles
. Is it possible to also distribute the batch dimension over devices (e.g. when running on multiple GPUs). My particular application is prone to jax OOM errors and would benefit from distribution overjax.pmap
.