Josh Watzman b50ff4f4e4 Reduce peak memory usage when changing models ...

A few tweaks to reduce peak memory usage, the biggest being that if we
aren't using the checkpoint cache, we shouldn't duplicate the model
state dict just to immediately throw it away.

On my machine with 16GB of RAM, this change means I can typically change
models, whereas before it would typically OOM.

2022-10-27 22:01:06 +01:00

10 KiB

Raw Blame History

View Raw