Cublaslt Grouped Gemm (2024)

If you're building a transformer-based model, a recommender system, or any application that requires many small, independent matrix multiplications, Grouped GEMM should be your default choice. As NVIDIA continues to optimize cuBLASLt for Hopper and future architectures, the performance gap between irregular and regular workloads will only shrink further. For implementation details, refer to the NVIDIA cuBLASLt Developer Guide (CUDA 12.x and later).

In the world of High-Performance Computing (HPC) and Deep Learning (DL), the General Matrix Multiply (GEMM) operation is the undisputed king. From large language models (LLMs) to scientific simulations, performance often hinges on how efficiently you can compute C = α*A*B + β*C . cublaslt grouped gemm

Enter – a modern solution designed to handle the messy, heterogeneous reality of advanced computing. The Problem with Traditional Batched GEMM Imagine training a recommendation system with embedding tables of varying sizes, or running inference on a transformer model with variable sequence lengths. In these scenarios, you might have 1,024 independent GEMM operations, each with different M, N, or K dimensions. If you're building a transformer-based model, a recommender

Traditional cuBLAS offers batched GEMM (e.g., cublas<t>gemmBatched ), which runs a list of independent matrix multiplications. However, it comes with a major limitation: (M, N, K) and data types. In the world of High-Performance Computing (HPC) and

// Allocate and fill matrices...