KV-Cache Refresh Methods for Long Generation
In this blog, we show how helpful KV cache refreshes can be for long generation from small models, along with efficient ways of finding when to refresh using inference algorithms.
Blog
The MCMC Inference Engine Behind a PPL
Overview of building an MCMC engine with practical implementation details.
Collaborative Filtering Methods for Paper Recommendation Systems
Describing collaborative filtering methods from classical SVD to NeuMF/GraphNeuMF/DMF and an ensemble to recommend papers.