libvecflowchamfer.so + libvecflowchamfer_kernels.so, built from the vecflow-chamfer tree in the VecFlow repo. Two-stage multi-vector (ColBERT-style) retrieval pipeline: MaxIVF anchor index (CAGRA-routed) for candidate generation, then a fused chamfer-scoring kernel for full-precision reranking. Document embeddings stay in host RAM and the GPU reads them over C2C (Grace-Hopper) or via managed memory (PCIe-attached GPUs) — the allocator is auto-selected at dataset load.

Depends on libcuvs-vecflow-cuXX of the same version. Install both via:

mamba install -c VecFlow -c rapidsai-nightly -c rapidsai \
              -c conda-forge libvecflow-chamfer-cu12