CUDA Memory error in precompute cache
Opened this issue · 1 comments
mcodoni commented
Hi,
I'm trying to precompute the cache for 2019 pretrained model with a google collab GPU (12 GB) and I always get a CUDA alllocating memory error
How much memory we need to run the script?
regards
Maurizio
norairn commented
I've spent some time playing around with the script (06_precompute_cache.py) options to work around the CUDA memory allocation error. I ended up with -g 0 -c 1000000 -b 100
, which means limit the search to 1 million top rows in batches of 100.
I had to sign up for Google Colab Pro and use GPU+High RAM runtime type, otherwise it was too unstable.
Anyway, here is the link to the zipped cache file: cache.zip. Copy it to the 2019 vectors folder and it will be picked up automatically.