You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
it seems that during indexing facebook/opt-2.7b, there are two problems:
there's a memory usage spike of 5GB each time one of the archives is read (makes sense, this is probably the pickle file being deserialized)
memory usage permanently rises by 200-600MB (!!!) each time the model is indexed
the pickle spike is a little weird because in theory we're intercepting tensor creation stuff with the pickle patch/subclass thingey, so maybe its reading the file to a memory buffer somewhere for speed. in that case we might be able to use a streaming capable pickle module, but i do wonder what kind of effect that will have on one big tree-like object such as an nn.Module
the 500ish MB is a bit more concerning, no clue what could be causing that (caching?)
The text was updated successfully, but these errors were encountered:
it seems that during indexing
facebook/opt-2.7b
, there are two problems:the pickle spike is a little weird because in theory we're intercepting tensor creation stuff with the pickle patch/subclass thingey, so maybe its reading the file to a memory buffer somewhere for speed. in that case we might be able to use a streaming capable pickle module, but i do wonder what kind of effect that will have on one big tree-like object such as an
nn.Module
the 500ish MB is a bit more concerning, no clue what could be causing that (caching?)
The text was updated successfully, but these errors were encountered: