Late Interaction Kernels

CUDA MaxSim kernels from v0.4.5 of H Company's Late Interaction Kernels, packaged for Hugging Face Kernels. The kernel algorithms are H Company's work, distributed under the Apache 2.0 license included here.

Usage

from kernels import get_kernel

kernel = get_kernel("sentence-transformers/late-interaction-kernels", version=0, trust_remote_code=True)

scores = kernel.layers.MaxSim()(queries, documents, query_mask, document_mask)
paired_scores = kernel.layers.MaxSimPairwise()(queries, paired_documents, query_mask, paired_document_mask)

MaxSim computes scores between all query-document combinations. MaxSimPairwise scores matching query-document pairs. Both support token masks and autograd.