Why does FFCV use extra memory on main GPU in distributed mode?

[numpee]

When training on ImageNet with FFCV, I've noticed that additional processes will spawn on GPU:0 that use up some memory. The additional processes spawn at the first iteration, if I haven't previously cached the dataset onto RAM.
Below is a screenshot of my nvidia-smi output when loading data via FFCV, loading a model to GPU, but not actually running training (no forward or backward pass through the model):

In the figure above, processes 49846, 49847, and 49848 on GPU:0 are spawned at the first iteration.

Based on the behavior, I assume the caching operation uses some GPU memory. Is this correct? Also, is there a way to disable this?
The issue is that I want to use a specific batch size, but will run into OOM because of the additional memory used.

FFCV version: 0.0.2

[GuillaumeLeclerc]

This should be fixed in the next release. Have you tried v0.0.3 ?

[numpee]

Thanks for the reply, @GuillaumeLeclerc.
I just updated to 0.0.3rc1 (seems like this is the latest release on PyPI?) and the issue still persists.
I've tested on 2 separate machines.
Perhaps PyPI hasn't been updated with 0.0.3?

Edit 1: I just noticed that this was mention in this pull request. Will try to install from the v0.0.4 branch

Edit 2: Ok I manually installed v0.0.4 from the git repo. Seems to be fixed now - sorry for the hassle 😢