// the take
A Python library for collaborative filtering on implicit feedback data — think play counts, clicks, and purchases rather than star ratings. It implements ALS, BPR, and LMF with Cython/OpenMP for CPU parallelism and CUDA kernels for GPU training. Aimed at ML engineers who need a production-grade recommender without building the math from scratch.
ALS GPU training via custom CUDA kernels is genuinely fast — the benchmarks against Spark show meaningful wall-clock wins on large datasets. The approximate nearest-neighbour integrations (Annoy, Faiss, NMSLIB) are first-class, not bolted on. The API is stable and consistent across models: fit, recommend, similar_items work the same way regardless of which algorithm you pick. Prebuilt wheels for x86_64 Linux/Windows/macOS with GPU support included mean you rarely need a working CUDA build environment just to install.
GPU support requires CUDA 13 and RMM, which is a surprisingly heavy dependency chain for what is otherwise a lightweight library — if your GPU is even one major version behind you are out of luck. The library is fundamentally offline-batch: there is no incremental update path, so adding new users or items means retraining from scratch. The examples folder has two scripts and a notebook from 2016-era datasets (Last.fm, MovieLens), which is thin for a library at this maturity. No built-in support for feature-enriched models (side information on users or items), so if you need that you will be gluing in LightFM or building it yourself.
We read the README, the directory tree and a key source file in July 2026. Nobody pays to be listed. How we judge →