Search

We present HadaCore, a modified Fast Walsh-Hadamard Transform (FWHT) algorithm optimized for the Tensor Cores present in modern GPU hardware. HadaCore follows the recursive structure of the original FWHT algorithm, achieving the same asymptotic runtime complexity but leveraging a hardware-aware work decomposition that benefits from Tensor Core acceleration. This reduces bottlenecks from compute and data exchange. On Nvidia A100 and H100 GPUs, HadaCore achieves speedups of 1.1-1.4x and 1.0-1.3x, with a peak gain of 3.5x and 3.6x respectively, when compared to the existing state-of-the-art implementation of the original algorithm. We also show that when using FP16 or BF16, our implementation is numerically accurate, enabling comparable accuracy on MMLU benchmarks when used in an end-to-end Llama3 inference run with quantized (FP8) attention.

In this paper, we present a novel method that reduces model inference latency during distributed deployment of Large Language Models (LLMs). Our contribution is an optimized inference deployment scheme that address the current limitations of state-of-the-art quantization kernels when used in conjunction with Tensor Parallel (TP). Our method preserves data locality in GPU memory access patterns and exploits a priori knowledge of TP to reduce global communication. We demonstrate an up to 1.81x speedup over existing methods for Llama-70B and up to 1.78x speedup for IBM WatsonX's Granite-20B MLP layer problem sizes on A100 and H100 NVIDIA DGX Systems for a variety of TP settings.

We propose an implementation of an efficient fused matrix multiplication kernel for W4A16 quantized inference, where we perform dequantization and GEMM in a fused kernel using a SplitK work decomposition. Our implementation shows improvement for the type of skinny matrix-matrix multiplications found in foundation model inference workloads. In particular, this paper surveys the type of matrix multiplication between a skinny activation matrix and a square weight matrix. Our results show an average of 65% speed improvement on A100, and an average of 124% speed improvement on H100 (with a peak of 295%) for a range of matrix dimensions including those found in a llama-style model, where m < n = k.