kolergy
kolergy
Those fixes makes a lot of sense, I'm surprised it was not merged
it seems that the ingest.py dose not return the Vram it use. I get out of memory errors if I repeatedly run it and I have 24Gb of Vram.
I'm not sure to have the competences but I'm trying to find where it comes from, it seems to be deeper than the ingest maybe in the vector store call....
@mingyuwanggithub The documents are all loaded, then split into chunks then embedding are generated all without using the GPU. The VRAM usage seems to come from the Duckdb, which to...