I haven’t been able to reproduce the autotune or Benchmarker elements of the scVI-tools tutorials on a variety of CentOS nodes with several Xeon models and several NVIDIA GPUs (incl. A100 and A40), across about a dozen different fresh conda environments created first by following the installation instructions to the letter, and then also by trying a number of variations when that didn’t work. While the errors differed from case to case, they’re always due to dependency conflicts involving Jax, PyTorch, Tensorflow, and/or Ray.
Does anyone have working conda/pip recipes to share in which autotune and Benchmarker both run without error?