From v1.5.3 to v1.6.3, GPU memory usage has increased when adding vectors to an IVF index, and then never querying. In v1.5.3 GPU memory use was constant. Now it grows linearly with number of vectors.

wv.key_to_index, wv.index_to_key and wv.vectors are all empty. FastTextKeyedVectors is a child of KeyedVectors where the add_vector/s methods are implemented. add_vector does a few checks then calls ...