-
after dropping the locking and self-referencing backend<->cell logger, the code needs a serious cleanup. currently gpu memory measurement functions are a duplicate - should refactor that.
-
try to switch to the pytorch counter instead of peak monitor thread. The 2 functions are: torch.cuda.max_memory_allocated() torch.cuda.reset_max_memory_allocated Unfortunately there can be only one counter - so it'd be tricky to measure different scopes (cell-level and the whole experiment)