mojolearn: GPU machine-learning algorithms in Mojo, portable to Apple Silicon

Résumé fourni par la source

Ports of GPU machine-learning algorithms into Mojo from the CUDA implementations that established them, targeting Apple Silicon via Metal first and CUDA and ROCm from the same source. Gradient-boosted oblivious (symmetric) trees mirror CatBoost; random forests, extremely randomized trees, DBSCAN, PCA, ridge and logistic regression, coordinate descent, support vector classification, kernel density estimation, single-linkage and spectral clustering, isolation forests, Holt-Winters and the scoring metrics mirror cuML; k-means and brute-force k-nearest-neighbors mirror cuVS; ordinary least squares, top-k selection and the MST and Lanczos primitives mirror RAFT. Every derivation is recorded per file in PORTED_MAP.tsv and the license obligations each carries are in NOTICE. Each estimator runs in one of three numeric modes, selected per estimator or per process from Python: FAST promises speed only and its bits move run to run, DETERMINISTIC promises the same bits on every run of one box and says nothing about a second, and IDENTICAL promises both plus the same bits across vendors. All three ship in the one wheel. IDENTICAL pins the seams that make a GPU result depend on its vendor, and where a lane has been measured on three vendors its card is bit-identical across Metal, HIP and CUDA; the certificates are in E1_RESULTS.md, E2_RESULTS.md and E3_RESULTS.md, and the per-lane ledger is IDENTITY_PATHS.md. That measurement does not cover every lane: several have run on one Apple M4 and nowhere else, each such class says so, and FAST makes no cross-vendor claim anywhere.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.