Recommendation Systems (RecSys) comprise a large footprint of production-deployed AI today. The neural network-based recommender systems differ from deep learning models in other domains in using high-cardinality categorical sparse features that require large embedding tables to be trained. In this talk we introduce TorchRec, a PyTorch domain library for Recommendation Systems. This new library provides common sparsity and parallelism primitives, enabling researchers to build state-of-the-art personalization models and deploy them in production. In this talk we cover the building blocks of the TorchRec library including modeling primitives such as embedding bags and jagged tensors, optimized recommender system kernels powered by FBGEMM, a flexible sharder that supports a veriety of strategies for partitioning embedding tables, a planner that automatically generates optimized and performant sharding plans, support for GPU inference and common modeling modules for building recommender system models. TorchRec library is currently used to train large-scale recommender models at Meta. We will present how TorchRec helped Meta’s recommender system platform to transition from CPU asynchronous training to accelerator-based full-sync training.

The k-means clustering algorithm is widely employed for unsupervised learning. The algorithm takes as input a multidimensional dataset of points and number of clusters/centroids, k, where each point is assigned to one of the clusters. For exact k-means clustering, the algorithm must compute the same result as Lloyd's algorithm, which is well-known to be computationally expensive due to the large number of distance comparisons between each point and the k centroids. Several algorithms have been proposed for k-means clustering that avoid distance calculations but produce an exact result. However, these algorithms have all been designed for execution using the CPU, and no published works have examined using the GPU to accelerate k-means while simultaneously avoiding distance calculations. This paper examines the state-of-the-art Yinyang algorithm that avoids distance calculations as executed on the GPU. Since Lloyd's algorithm is well-suited to a GPU execution, it is not clear whether the Yinyang algorithm will obtain significant performance gains on GPU hardware. In this context, this paper: (i) proposes the first GPU-accelerated Yinyang algorithm in the literature; (ii) advances several optimizations to GPU kernels; (iii) contrasts and evaluates different degrees of distance calculation pruning; and, (iv) compares the performance of our GPU-accelerated Yinyang algorithm to four reference implementations. Our GPU algorithm achieves a speedup over the multi-core CPU Yinyang algorithm of up to 8× on real-world datasets.

In this paper, we consider a trajectory planning problem where an autonomous vehicle aims to rendezvous with another cooperating vehicle in minimum time. The first vehicle has kinematic constraints, consequently feasible trajectories must have a maximum curvature less than a specified limit. Rendezvous is said to occur at the instant that the two vehicles are collocated with the same heading. We propose a technique to construct a trajectory, composed of piecewise quadratic Bézier curves, that satisfies the vehicle motion constraints and achieves rendezvous in minimum time. The methodology begins by finding safe flight corridors, which are constructed from sequences of triangles using constrained Delaunay triangulation of the feasible space; the triangles define the bounds of Bézier curves. We formulate the necessary constraints for continuity and feasibility as functions of the control points that define the Bézier curves, and the resulting optimization problem is solved using a nonlinear programming solver. The techniques developed were tested using simulated scenarios, and we present the results which highlight the efficacy of the proposed solution approach. Furthermore, the algorithm was implemented and tested in a field test and those results are presented.