I am going to be training item embeddings where there are 8-9 figures of items, and that will lead to memory issues to have all those embeddings loaded into memory all at once.

Since a training step only updates a fraction of the embeddings at a time, I am hoping there’s some library or framework that has the vectors stores on the disk until the training step.

This comes pretty close, https://github.com/plasticityai/magnitude as the vectors are stored to sqlite database, but there’s no way to train the vectors.