GPU Acceleration¶

rusket supports optional GPU acceleration via CuPy or PyTorch CUDA for models that benefit from large-scale matrix operations (Gramians, Cholesky solves, batch scoring, attention layers).

Installation¶

# CuPy — recommended, fastest
pip install cupy-cuda12x

# Or PyTorch
pip install torch

Global Enable/Disable¶

Instead of passing use_gpu=True to every model, enable CUDA once:

import rusket

rusket.enable_gpu()   # all models now default to GPU
rusket.disable_gpu()  # back to CPU (the default)
rusket.is_gpu_enabled()  # check current state

Per-Model Override¶

An explicit use_gpu on a model always wins over the global setting:

rusket.enable_gpu()

als  = rusket.ALS(factors=128)              # → GPU (from global)
ease = rusket.EASE(use_gpu=False)           # → CPU (explicit override)
bpr  = rusket.BPR(use_gpu=True)            # → GPU (explicit)

rusket.disable_gpu()
svd  = rusket.SVD()                         # → CPU (from global)
knn  = rusket.ItemKNN(use_gpu=True)        # → GPU (explicit override)

Supported Models¶

All 12 recommender models respect the global GPU flag:

Model	GPU-accelerated operations
ALS / eALS	Gramian computation, Cholesky factor solve, batch recommendation scoring
BPR	SGD factor updates, batch recommend
SVD	Factor updates, batch scoring
EASE	Gram matrix inversion (Cholesky)
ItemKNN / UserKNN	Similarity computation and scoring
LightGCN	Graph convolution layers, scoring
FM	Prediction scores
FPMC	Factor updates
SASRec / BERT4Rec	Self-attention forward pass, scoring
NMF	Multiplicative update rules

Checking Availability¶

import rusket

if rusket.check_gpu_available():
    rusket.enable_gpu()
    print("GPU backend active")
else:
    print("No GPU — using Rust CPU backend (still fast!)")

No GPU? No problem. The Rust CPU backend (Rayon multi-threading, SIMD) is already highly optimised. GPU acceleration helps most for very large factor dimensions (≥256) or batch scoring millions of users simultaneously.