Inquiry for Format or Interface of cuVS Index Transfering for CPU-GPU Workflow

When using cuVS without a GPU, both the build and search processes throw errors. Is this related to the RMM (RAPIDS Memory Manager) resource management, or does cuVS require a GPU to run? If cuVS is GPU-dependent, the limited GPU memory may prevent the construction of large datasets, such as being unable to build an index for a dataset of 10M 1024-dimensional vectors. However, if I use compression methods like IVFPQ, the search phase would require less memory as it doesn't need to store the original vectors, possibly allowing the search to be executed on the GPU. I would like to try using other tools to build the index on a CPU platform and then use the GPU platform for search. In this case, what format should I use to pass the built index into ivf_pq or cagra? Could you provide the format specification or interface details for loading the index?

The issue names cuVS, ivf_pq, and cagra but no files or tests. Start by reviewing the existing index-loading interfaces and GPU/RMM requirements for those components. Done means documenting whether CPU-built indexes can be loaded for GPU search and, if supported, the accepted transfer format.