Describe the proposed new feature or enhancement:
Current implementation uses faiss-cpu. Try looking into the possibility of using native faiss to use GPU indexes.
While at it, use their API for direct Tensor operation (no numpy/CPU bottleneck). See facebookresearch/faiss#561 (comment).
Describe the proposed new feature or enhancement:
Current implementation uses
faiss-cpu. Try looking into the possibility of using nativefaissto use GPU indexes.While at it, use their API for direct
Tensoroperation (nonumpy/CPU bottleneck). See facebookresearch/faiss#561 (comment).