MARLIN: Mixed-Precision Auto-Regressive Parallel Inference on Large Language Models

As inference on Large Language Models (LLMs) emerges as an important workload in machine learning applications, weight quantization has become a standard technique for efficient GPU deployment. Quantization not only reduces model size, but has also been shown to yield substantial speedups for single-user inference, due to reduced memory movement, with low accuracy impact. Yet, it remains open whether speedups are achievable also in \emph{batched} settings with multiple parallel clients, which are highly relevant for practical serving. It is unclear whether GPU kernels can be designed to remain practically memory-bound, while supporting the substantially increased compute requirements of batched workloads. This paper resolves this question positively by describing the design of Mixed-precision Auto-Regressive LINear kernels, called MARLIN. Concretely, given a model whose weights are compressed via quantization to, e.g., 4 bits per element, MARLIN shows that batchsizes up to 16-32 can be supported with close to maximum ($4\times$) quantization speedup, and larger batchsizes up to 64-128 with gradually decreasing, but still significant, acceleration. MARLIN accomplishes this via a combination of techniques, such as asynchronous memory access, complex task scheduling and pipelining, and bespoke quantization support. Our experiments show that MARLIN's near-optimal performance on individual LLM layers across different scenarios can also lead to end-to-end LLM inference speedups (of up to $2.8\times$) when integrated with the popular vLLM serving engine. Finally, MARLIN is extensible to further compression techniques, like NVIDIA 2:4 sparsity, leading to additional speedups.

VENOM: A Vectorized N:M Format for Unleashing the Power of Sparse Tensor Cores

The increasing success and scaling of Deep Learning models demands higher computational efficiency and power. Sparsification can lead to both smaller models as well as higher compute efficiency, and accelerated hardware is becoming available. However, exploiting it efficiently requires kernel implementations, pruning algorithms, and storage formats, to utilize hardware support of specialized sparse vector units. An example of those are the NVIDIA's Sparse Tensor Cores (SPTCs), which promise a 2x speedup. However, SPTCs only support the 2:4 format, limiting achievable sparsity ratios to 50%. We present the V:N:M format, which enables the execution of arbitrary N:M ratios on SPTCs. To efficiently exploit the resulting format, we propose Spatha, a high-performance sparse-library for DL routines. We show that Spatha achieves up to 37x speedup over cuBLAS. We also demonstrate a second-order pruning technique that enables sparsification to high sparsity ratios with V:N:M and little to no loss in accuracy in modern transformers.

Reusing Trained Layers of Convolutional Neural Networks to Shorten Hyperparameters Tuning Time

Hyperparameters tuning is a time-consuming approach, particularly when the architecture of the neural network is decided as part of this process. For instance, in convolutional neural networks (CNNs), the selection of the number and the characteristics of the hidden (convolutional) layers may be decided. This implies that the search process involves the training of all these candidate network architectures. This paper describes a proposal to reuse the weights of hidden (convolutional) layers among different trainings to shorten this process. The rationale is that if a set of convolutional layers have been trained to solve a given problem, the weights calculated in this training may be useful when a new convolutional layer is added to the network architecture. This idea has been tested using the CIFAR-10 dataset, testing different CNNs architectures with up to 3 convolutional layers and up to 3 fully connected layers. The experiments compare the training time and the validation loss when reusing and not reusing convolutional layers. They confirm that this strategy reduces the training time while it even increases the accuracy of the resulting neural network. This finding opens up the future possibility of integrating this strategy in existing AutoML methods with the purpose of reducing the total search time.

A Hybrid Approach for Tracking Individual Players in Broadcast Match Videos

Tracking people in a video sequence is a challenging task that has been approached from many perspectives. This task becomes even more complicated when the person to track is a player in a broadcasted sport event, the reasons being the existence of difficulties such as frequent camera movements or switches, total and partial occlusions between players, and blurry frames due to the codification algorithm of the video. This paper introduces a player tracking solution which is both fast and accurate. This allows to track a player precisely in real-time. The approach combines several models that are executed concurrently in a relatively modest hardware, and whose accuracy has been validated against hand-labeled broadcast video sequences. Regarding the accuracy, the tests show that the area under curve (AUC) of our approach is around 0.6, which is similar to generic state of the art solutions. As for performance, our proposal can process high definition videos (1920x1080 px) at 80 fps.